diff options
Diffstat (limited to 'media/libwebp/dsp')
63 files changed, 18181 insertions, 329 deletions
diff --git a/media/libwebp/dsp/alpha_processing.c b/media/libwebp/dsp/alpha_processing.c index 6ff1352ae2..8c5e90210f 100644 --- a/media/libwebp/dsp/alpha_processing.c +++ b/media/libwebp/dsp/alpha_processing.c @@ -157,7 +157,8 @@ void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse) { } } -void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha, +void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse) { int x; for (x = 0; x < width; ++x) { @@ -178,7 +179,8 @@ void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha, #undef MFIX void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse); -void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha, +void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse); //------------------------------------------------------------------------------ @@ -193,8 +195,8 @@ void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows, } } -void WebPMultRows(uint8_t* ptr, int stride, - const uint8_t* alpha, int alpha_stride, +void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride, + const uint8_t* WEBP_RESTRICT alpha, int alpha_stride, int width, int num_rows, int inverse) { int n; for (n = 0; n < num_rows; ++n) { @@ -290,9 +292,9 @@ static void ApplyAlphaMultiply_16b_C(uint8_t* rgba4444, } #if !WEBP_NEON_OMIT_C_CODE -static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride, +static int DispatchAlpha_C(const uint8_t* WEBP_RESTRICT alpha, int alpha_stride, int width, int height, - uint8_t* dst, int dst_stride) { + uint8_t* WEBP_RESTRICT dst, int dst_stride) { uint32_t alpha_mask = 0xff; int i, j; @@ -309,9 +311,10 @@ static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride, return (alpha_mask != 0xff); } -static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint32_t* dst, int dst_stride) { +static void DispatchAlphaToGreen_C(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT dst, + int dst_stride) { int i, j; for (j = 0; j < height; ++j) { for (i = 0; i < width; ++i) { @@ -322,9 +325,9 @@ static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride, } } -static int ExtractAlpha_C(const uint8_t* argb, int argb_stride, +static int ExtractAlpha_C(const uint8_t* WEBP_RESTRICT argb, int argb_stride, int width, int height, - uint8_t* alpha, int alpha_stride) { + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { uint8_t alpha_mask = 0xff; int i, j; @@ -340,7 +343,8 @@ static int ExtractAlpha_C(const uint8_t* argb, int argb_stride, return (alpha_mask == 0xff); } -static void ExtractGreen_C(const uint32_t* argb, uint8_t* alpha, int size) { +static void ExtractGreen_C(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size) { int i; for (i = 0; i < size; ++i) alpha[i] = argb[i] >> 8; } @@ -359,6 +363,11 @@ static int HasAlpha32b_C(const uint8_t* src, int length) { return 0; } +static void AlphaReplace_C(uint32_t* src, int length, uint32_t color) { + int x; + for (x = 0; x < length; ++x) if ((src[x] >> 24) == 0) src[x] = color; +} + //------------------------------------------------------------------------------ // Simple channel manipulations. @@ -367,8 +376,11 @@ static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) { } #ifdef WORDS_BIGENDIAN -static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g, - const uint8_t* b, int len, uint32_t* out) { +static void PackARGB_C(const uint8_t* WEBP_RESTRICT a, + const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, uint32_t* WEBP_RESTRICT out) { int i; for (i = 0; i < len; ++i) { out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]); @@ -376,8 +388,10 @@ static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g, } #endif -static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b, - int len, int step, uint32_t* out) { +static void PackRGB_C(const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, int step, uint32_t* WEBP_RESTRICT out) { int i, offset = 0; for (i = 0; i < len; ++i) { out[i] = MakeARGB32(0xff, r[offset], g[offset], b[offset]); @@ -387,19 +401,26 @@ static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b, void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int); void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int); -int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int); -void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int); -int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int); -void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size); +int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int, + uint8_t* WEBP_RESTRICT, int); +void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT, int, int, int, + uint32_t* WEBP_RESTRICT, int); +int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int, + uint8_t* WEBP_RESTRICT, int); +void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size); #ifdef WORDS_BIGENDIAN void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, const uint8_t* g, const uint8_t* b, int, uint32_t*); #endif -void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b, - int len, int step, uint32_t* out); +void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, int step, uint32_t* WEBP_RESTRICT out); int (*WebPHasAlpha8b)(const uint8_t* src, int length); int (*WebPHasAlpha32b)(const uint8_t* src, int length); +void (*WebPAlphaReplace)(uint32_t* src, int length, uint32_t color); //------------------------------------------------------------------------------ // Init function @@ -428,13 +449,14 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) { WebPHasAlpha8b = HasAlpha8b_C; WebPHasAlpha32b = HasAlpha32b_C; + WebPAlphaReplace = AlphaReplace_C; // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitAlphaProcessingSSE2(); -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitAlphaProcessingSSE41(); } @@ -448,7 +470,7 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPInitAlphaProcessingNEON(); @@ -469,4 +491,5 @@ WEBP_DSP_INIT_FUNC(WebPInitAlphaProcessing) { assert(WebPPackRGB != NULL); assert(WebPHasAlpha8b != NULL); assert(WebPHasAlpha32b != NULL); + assert(WebPAlphaReplace != NULL); } diff --git a/media/libwebp/dsp/alpha_processing_mips_dsp_r2.c b/media/libwebp/dsp/alpha_processing_mips_dsp_r2.c new file mode 100644 index 0000000000..ab597e68bb --- /dev/null +++ b/media/libwebp/dsp/alpha_processing_mips_dsp_r2.c @@ -0,0 +1,228 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Utilities for processing transparent channel. +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +static int DispatchAlpha_MIPSdspR2(const uint8_t* alpha, int alpha_stride, + int width, int height, + uint8_t* dst, int dst_stride) { + uint32_t alpha_mask = 0xffffffff; + int i, j, temp0; + + for (j = 0; j < height; ++j) { + uint8_t* pdst = dst; + const uint8_t* palpha = alpha; + for (i = 0; i < (width >> 2); ++i) { + int temp1, temp2, temp3; + + __asm__ volatile ( + "ulw %[temp0], 0(%[palpha]) \n\t" + "addiu %[palpha], %[palpha], 4 \n\t" + "addiu %[pdst], %[pdst], 16 \n\t" + "srl %[temp1], %[temp0], 8 \n\t" + "srl %[temp2], %[temp0], 16 \n\t" + "srl %[temp3], %[temp0], 24 \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + "sb %[temp0], -16(%[pdst]) \n\t" + "sb %[temp1], -12(%[pdst]) \n\t" + "sb %[temp2], -8(%[pdst]) \n\t" + "sb %[temp3], -4(%[pdst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [palpha]"+r"(palpha), [pdst]"+r"(pdst), + [alpha_mask]"+r"(alpha_mask) + : + : "memory" + ); + } + + for (i = 0; i < (width & 3); ++i) { + __asm__ volatile ( + "lbu %[temp0], 0(%[palpha]) \n\t" + "addiu %[palpha], %[palpha], 1 \n\t" + "sb %[temp0], 0(%[pdst]) \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + "addiu %[pdst], %[pdst], 4 \n\t" + : [temp0]"=&r"(temp0), [palpha]"+r"(palpha), [pdst]"+r"(pdst), + [alpha_mask]"+r"(alpha_mask) + : + : "memory" + ); + } + alpha += alpha_stride; + dst += dst_stride; + } + + __asm__ volatile ( + "ext %[temp0], %[alpha_mask], 0, 16 \n\t" + "srl %[alpha_mask], %[alpha_mask], 16 \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + "ext %[temp0], %[alpha_mask], 0, 8 \n\t" + "srl %[alpha_mask], %[alpha_mask], 8 \n\t" + "and %[alpha_mask], %[alpha_mask], %[temp0] \n\t" + : [temp0]"=&r"(temp0), [alpha_mask]"+r"(alpha_mask) + : + ); + + return (alpha_mask != 0xff); +} + +static void MultARGBRow_MIPSdspR2(uint32_t* const ptr, int width, + int inverse) { + int x; + const uint32_t c_00ffffff = 0x00ffffffu; + const uint32_t c_ff000000 = 0xff000000u; + const uint32_t c_8000000 = 0x00800000u; + const uint32_t c_8000080 = 0x00800080u; + for (x = 0; x < width; ++x) { + const uint32_t argb = ptr[x]; + if (argb < 0xff000000u) { // alpha < 255 + if (argb <= 0x00ffffffu) { // alpha == 0 + ptr[x] = 0; + } else { + int temp0, temp1, temp2, temp3, alpha; + __asm__ volatile ( + "srl %[alpha], %[argb], 24 \n\t" + "replv.qb %[temp0], %[alpha] \n\t" + "and %[temp0], %[temp0], %[c_00ffffff] \n\t" + "beqz %[inverse], 0f \n\t" + "divu $zero, %[c_ff000000], %[alpha] \n\t" + "mflo %[temp0] \n\t" + "0: \n\t" + "andi %[temp1], %[argb], 0xff \n\t" + "ext %[temp2], %[argb], 8, 8 \n\t" + "ext %[temp3], %[argb], 16, 8 \n\t" + "mul %[temp1], %[temp1], %[temp0] \n\t" + "mul %[temp2], %[temp2], %[temp0] \n\t" + "mul %[temp3], %[temp3], %[temp0] \n\t" + "precrq.ph.w %[temp1], %[temp2], %[temp1] \n\t" + "addu %[temp3], %[temp3], %[c_8000000] \n\t" + "addu %[temp1], %[temp1], %[c_8000080] \n\t" + "precrq.ph.w %[temp3], %[argb], %[temp3] \n\t" + "precrq.qb.ph %[temp1], %[temp3], %[temp1] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [alpha]"=&r"(alpha) + : [inverse]"r"(inverse), [c_00ffffff]"r"(c_00ffffff), + [c_8000000]"r"(c_8000000), [c_8000080]"r"(c_8000080), + [c_ff000000]"r"(c_ff000000), [argb]"r"(argb) + : "memory", "hi", "lo" + ); + ptr[x] = temp1; + } + } + } +} + +#ifdef WORDS_BIGENDIAN +static void PackARGB_MIPSdspR2(const uint8_t* a, const uint8_t* r, + const uint8_t* g, const uint8_t* b, int len, + uint32_t* out) { + int temp0, temp1, temp2, temp3, offset; + const int rest = len & 1; + const uint32_t* const loop_end = out + len - rest; + const int step = 4; + __asm__ volatile ( + "xor %[offset], %[offset], %[offset] \n\t" + "beq %[loop_end], %[out], 0f \n\t" + "2: \n\t" + "lbux %[temp0], %[offset](%[a]) \n\t" + "lbux %[temp1], %[offset](%[r]) \n\t" + "lbux %[temp2], %[offset](%[g]) \n\t" + "lbux %[temp3], %[offset](%[b]) \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "ins %[temp3], %[temp2], 16, 16 \n\t" + "addiu %[out], %[out], 4 \n\t" + "precr.qb.ph %[temp0], %[temp1], %[temp3] \n\t" + "sw %[temp0], -4(%[out]) \n\t" + "addu %[offset], %[offset], %[step] \n\t" + "bne %[loop_end], %[out], 2b \n\t" + "0: \n\t" + "beq %[rest], $zero, 1f \n\t" + "lbux %[temp0], %[offset](%[a]) \n\t" + "lbux %[temp1], %[offset](%[r]) \n\t" + "lbux %[temp2], %[offset](%[g]) \n\t" + "lbux %[temp3], %[offset](%[b]) \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "ins %[temp3], %[temp2], 16, 16 \n\t" + "precr.qb.ph %[temp0], %[temp1], %[temp3] \n\t" + "sw %[temp0], 0(%[out]) \n\t" + "1: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [offset]"=&r"(offset), [out]"+&r"(out) + : [a]"r"(a), [r]"r"(r), [g]"r"(g), [b]"r"(b), [step]"r"(step), + [loop_end]"r"(loop_end), [rest]"r"(rest) + : "memory" + ); +} +#endif // WORDS_BIGENDIAN + +static void PackRGB_MIPSdspR2(const uint8_t* r, const uint8_t* g, + const uint8_t* b, int len, int step, + uint32_t* out) { + int temp0, temp1, temp2, offset; + const int rest = len & 1; + const int a = 0xff; + const uint32_t* const loop_end = out + len - rest; + __asm__ volatile ( + "xor %[offset], %[offset], %[offset] \n\t" + "beq %[loop_end], %[out], 0f \n\t" + "2: \n\t" + "lbux %[temp0], %[offset](%[r]) \n\t" + "lbux %[temp1], %[offset](%[g]) \n\t" + "lbux %[temp2], %[offset](%[b]) \n\t" + "ins %[temp0], %[a], 16, 16 \n\t" + "ins %[temp2], %[temp1], 16, 16 \n\t" + "addiu %[out], %[out], 4 \n\t" + "precr.qb.ph %[temp0], %[temp0], %[temp2] \n\t" + "sw %[temp0], -4(%[out]) \n\t" + "addu %[offset], %[offset], %[step] \n\t" + "bne %[loop_end], %[out], 2b \n\t" + "0: \n\t" + "beq %[rest], $zero, 1f \n\t" + "lbux %[temp0], %[offset](%[r]) \n\t" + "lbux %[temp1], %[offset](%[g]) \n\t" + "lbux %[temp2], %[offset](%[b]) \n\t" + "ins %[temp0], %[a], 16, 16 \n\t" + "ins %[temp2], %[temp1], 16, 16 \n\t" + "precr.qb.ph %[temp0], %[temp0], %[temp2] \n\t" + "sw %[temp0], 0(%[out]) \n\t" + "1: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [offset]"=&r"(offset), [out]"+&r"(out) + : [a]"r"(a), [r]"r"(r), [g]"r"(g), [b]"r"(b), [step]"r"(step), + [loop_end]"r"(loop_end), [rest]"r"(rest) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitAlphaProcessingMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingMIPSdspR2(void) { + WebPDispatchAlpha = DispatchAlpha_MIPSdspR2; + WebPMultARGBRow = MultARGBRow_MIPSdspR2; +#ifdef WORDS_BIGENDIAN + WebPPackARGB = PackARGB_MIPSdspR2; +#endif + WebPPackRGB = PackRGB_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/alpha_processing_neon.c b/media/libwebp/dsp/alpha_processing_neon.c index 53dfce2b36..c900279a35 100644 --- a/media/libwebp/dsp/alpha_processing_neon.c +++ b/media/libwebp/dsp/alpha_processing_neon.c @@ -80,9 +80,9 @@ static void ApplyAlphaMultiply_NEON(uint8_t* rgba, int alpha_first, //------------------------------------------------------------------------------ -static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint8_t* dst, int dst_stride) { +static int DispatchAlpha_NEON(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint8_t* WEBP_RESTRICT dst, int dst_stride) { uint32_t alpha_mask = 0xffffffffu; uint8x8_t mask8 = vdup_n_u8(0xff); uint32_t tmp[2]; @@ -112,9 +112,10 @@ static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride, return (alpha_mask != 0xffffffffu); } -static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint32_t* dst, int dst_stride) { +static void DispatchAlphaToGreen_NEON(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT dst, + int dst_stride) { int i, j; uint8x8x4_t greens; // leave A/R/B channels zero'd. greens.val[0] = vdup_n_u8(0); @@ -131,9 +132,9 @@ static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride, } } -static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride, +static int ExtractAlpha_NEON(const uint8_t* WEBP_RESTRICT argb, int argb_stride, int width, int height, - uint8_t* alpha, int alpha_stride) { + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { uint32_t alpha_mask = 0xffffffffu; uint8x8_t mask8 = vdup_n_u8(0xff); uint32_t tmp[2]; @@ -161,8 +162,8 @@ static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride, return (alpha_mask == 0xffffffffu); } -static void ExtractGreen_NEON(const uint32_t* argb, - uint8_t* alpha, int size) { +static void ExtractGreen_NEON(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size) { int i; for (i = 0; i + 16 <= size; i += 16) { const uint8x16x4_t rgbX = vld4q_u8((const uint8_t*)(argb + i)); diff --git a/media/libwebp/dsp/alpha_processing_sse2.c b/media/libwebp/dsp/alpha_processing_sse2.c index 9a3bc4485a..56d9ee5e98 100644 --- a/media/libwebp/dsp/alpha_processing_sse2.c +++ b/media/libwebp/dsp/alpha_processing_sse2.c @@ -18,9 +18,9 @@ //------------------------------------------------------------------------------ -static int DispatchAlpha_SSE2(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint8_t* dst, int dst_stride) { +static int DispatchAlpha_SSE2(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint8_t* WEBP_RESTRICT dst, int dst_stride) { // alpha_and stores an 'and' operation of all the alpha[] values. The final // value is not 0xff if any of the alpha[] is not equal to 0xff. uint32_t alpha_and = 0xff; @@ -72,9 +72,10 @@ static int DispatchAlpha_SSE2(const uint8_t* alpha, int alpha_stride, return (alpha_and != 0xff); } -static void DispatchAlphaToGreen_SSE2(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint32_t* dst, int dst_stride) { +static void DispatchAlphaToGreen_SSE2(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT dst, + int dst_stride) { int i, j; const __m128i zero = _mm_setzero_si128(); const int limit = width & ~15; @@ -98,9 +99,9 @@ static void DispatchAlphaToGreen_SSE2(const uint8_t* alpha, int alpha_stride, } } -static int ExtractAlpha_SSE2(const uint8_t* argb, int argb_stride, +static int ExtractAlpha_SSE2(const uint8_t* WEBP_RESTRICT argb, int argb_stride, int width, int height, - uint8_t* alpha, int alpha_stride) { + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { // alpha_and stores an 'and' operation of all the alpha[] values. The final // value is not 0xff if any of the alpha[] is not equal to 0xff. uint32_t alpha_and = 0xff; @@ -214,7 +215,7 @@ static void ApplyAlphaMultiply_SSE2(uint8_t* rgba, int alpha_first, // Alpha detection static int HasAlpha8b_SSE2(const uint8_t* src, int length) { - const __m128i all_0xff = _mm_set1_epi8(0xff); + const __m128i all_0xff = _mm_set1_epi8((char)0xff); int i = 0; for (; i + 16 <= length; i += 16) { const __m128i v = _mm_loadu_si128((const __m128i*)(src + i)); @@ -228,7 +229,7 @@ static int HasAlpha8b_SSE2(const uint8_t* src, int length) { static int HasAlpha32b_SSE2(const uint8_t* src, int length) { const __m128i alpha_mask = _mm_set1_epi32(0xff); - const __m128i all_0xff = _mm_set1_epi8(0xff); + const __m128i all_0xff = _mm_set1_epi8((char)0xff); int i = 0; // We don't know if we can access the last 3 bytes after the last alpha // value 'src[4 * length - 4]' (because we don't know if alpha is the first @@ -265,6 +266,27 @@ static int HasAlpha32b_SSE2(const uint8_t* src, int length) { return 0; } +static void AlphaReplace_SSE2(uint32_t* src, int length, uint32_t color) { + const __m128i m_color = _mm_set1_epi32(color); + const __m128i zero = _mm_setzero_si128(); + int i = 0; + for (; i + 8 <= length; i += 8) { + const __m128i a0 = _mm_loadu_si128((const __m128i*)(src + i + 0)); + const __m128i a1 = _mm_loadu_si128((const __m128i*)(src + i + 4)); + const __m128i b0 = _mm_srai_epi32(a0, 24); + const __m128i b1 = _mm_srai_epi32(a1, 24); + const __m128i c0 = _mm_cmpeq_epi32(b0, zero); + const __m128i c1 = _mm_cmpeq_epi32(b1, zero); + const __m128i d0 = _mm_and_si128(c0, m_color); + const __m128i d1 = _mm_and_si128(c1, m_color); + const __m128i e0 = _mm_andnot_si128(c0, a0); + const __m128i e1 = _mm_andnot_si128(c1, a1); + _mm_storeu_si128((__m128i*)(src + i + 0), _mm_or_si128(d0, e0)); + _mm_storeu_si128((__m128i*)(src + i + 4), _mm_or_si128(d1, e1)); + } + for (; i < length; ++i) if ((src[i] >> 24) == 0) src[i] = color; +} + // ----------------------------------------------------------------------------- // Apply alpha value to rows @@ -296,7 +318,8 @@ static void MultARGBRow_SSE2(uint32_t* const ptr, int width, int inverse) { if (width > 0) WebPMultARGBRow_C(ptr + x, width, inverse); } -static void MultRow_SSE2(uint8_t* const ptr, const uint8_t* const alpha, +static void MultRow_SSE2(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse) { int x = 0; if (!inverse) { @@ -334,6 +357,7 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingSSE2(void) { WebPHasAlpha8b = HasAlpha8b_SSE2; WebPHasAlpha32b = HasAlpha32b_SSE2; + WebPAlphaReplace = AlphaReplace_SSE2; } #else // !WEBP_USE_SSE2 diff --git a/media/libwebp/dsp/alpha_processing_sse41.c b/media/libwebp/dsp/alpha_processing_sse41.c index e33c1aba4d..307d200f1f 100644 --- a/media/libwebp/dsp/alpha_processing_sse41.c +++ b/media/libwebp/dsp/alpha_processing_sse41.c @@ -19,9 +19,9 @@ //------------------------------------------------------------------------------ -static int ExtractAlpha_SSE41(const uint8_t* argb, int argb_stride, - int width, int height, - uint8_t* alpha, int alpha_stride) { +static int ExtractAlpha_SSE41(const uint8_t* WEBP_RESTRICT argb, + int argb_stride, int width, int height, + uint8_t* WEBP_RESTRICT alpha, int alpha_stride) { // alpha_and stores an 'and' operation of all the alpha[] values. The final // value is not 0xff if any of the alpha[] is not equal to 0xff. uint32_t alpha_and = 0xff; diff --git a/media/libwebp/dsp/cost.c b/media/libwebp/dsp/cost.c new file mode 100644 index 0000000000..bf112c7f0c --- /dev/null +++ b/media/libwebp/dsp/cost.c @@ -0,0 +1,411 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "../dsp/dsp.h" +#include "../enc/cost_enc.h" + +//------------------------------------------------------------------------------ +// Boolean-cost cost table + +const uint16_t VP8EntropyCost[256] = { + 1792, 1792, 1792, 1536, 1536, 1408, 1366, 1280, 1280, 1216, + 1178, 1152, 1110, 1076, 1061, 1024, 1024, 992, 968, 951, + 939, 911, 896, 878, 871, 854, 838, 820, 811, 794, + 786, 768, 768, 752, 740, 732, 720, 709, 704, 690, + 683, 672, 666, 655, 647, 640, 631, 622, 615, 607, + 598, 592, 586, 576, 572, 564, 559, 555, 547, 541, + 534, 528, 522, 512, 512, 504, 500, 494, 488, 483, + 477, 473, 467, 461, 458, 452, 448, 443, 438, 434, + 427, 424, 419, 415, 410, 406, 403, 399, 394, 390, + 384, 384, 377, 374, 370, 366, 362, 359, 355, 351, + 347, 342, 342, 336, 333, 330, 326, 323, 320, 316, + 312, 308, 305, 302, 299, 296, 293, 288, 287, 283, + 280, 277, 274, 272, 268, 266, 262, 256, 256, 256, + 251, 248, 245, 242, 240, 237, 234, 232, 228, 226, + 223, 221, 218, 216, 214, 211, 208, 205, 203, 201, + 198, 196, 192, 191, 188, 187, 183, 181, 179, 176, + 175, 171, 171, 168, 165, 163, 160, 159, 156, 154, + 152, 150, 148, 146, 144, 142, 139, 138, 135, 133, + 131, 128, 128, 125, 123, 121, 119, 117, 115, 113, + 111, 110, 107, 105, 103, 102, 100, 98, 96, 94, + 92, 91, 89, 86, 86, 83, 82, 80, 77, 76, + 74, 73, 71, 69, 67, 66, 64, 63, 61, 59, + 57, 55, 54, 52, 51, 49, 47, 46, 44, 43, + 41, 40, 38, 36, 35, 33, 32, 30, 29, 27, + 25, 24, 22, 21, 19, 18, 16, 15, 13, 12, + 10, 9, 7, 6, 4, 3 +}; + +//------------------------------------------------------------------------------ +// Level cost tables + +// fixed costs for coding levels, deduce from the coding tree. +// This is only the part that doesn't depend on the probability state. +const uint16_t VP8LevelFixedCosts[MAX_LEVEL + 1] = { + 0, 256, 256, 256, 256, 432, 618, 630, + 731, 640, 640, 828, 901, 948, 1021, 1101, + 1174, 1221, 1294, 1042, 1085, 1115, 1158, 1202, + 1245, 1275, 1318, 1337, 1380, 1410, 1453, 1497, + 1540, 1570, 1613, 1280, 1295, 1317, 1332, 1358, + 1373, 1395, 1410, 1454, 1469, 1491, 1506, 1532, + 1547, 1569, 1584, 1601, 1616, 1638, 1653, 1679, + 1694, 1716, 1731, 1775, 1790, 1812, 1827, 1853, + 1868, 1890, 1905, 1727, 1733, 1742, 1748, 1759, + 1765, 1774, 1780, 1800, 1806, 1815, 1821, 1832, + 1838, 1847, 1853, 1878, 1884, 1893, 1899, 1910, + 1916, 1925, 1931, 1951, 1957, 1966, 1972, 1983, + 1989, 1998, 2004, 2027, 2033, 2042, 2048, 2059, + 2065, 2074, 2080, 2100, 2106, 2115, 2121, 2132, + 2138, 2147, 2153, 2178, 2184, 2193, 2199, 2210, + 2216, 2225, 2231, 2251, 2257, 2266, 2272, 2283, + 2289, 2298, 2304, 2168, 2174, 2183, 2189, 2200, + 2206, 2215, 2221, 2241, 2247, 2256, 2262, 2273, + 2279, 2288, 2294, 2319, 2325, 2334, 2340, 2351, + 2357, 2366, 2372, 2392, 2398, 2407, 2413, 2424, + 2430, 2439, 2445, 2468, 2474, 2483, 2489, 2500, + 2506, 2515, 2521, 2541, 2547, 2556, 2562, 2573, + 2579, 2588, 2594, 2619, 2625, 2634, 2640, 2651, + 2657, 2666, 2672, 2692, 2698, 2707, 2713, 2724, + 2730, 2739, 2745, 2540, 2546, 2555, 2561, 2572, + 2578, 2587, 2593, 2613, 2619, 2628, 2634, 2645, + 2651, 2660, 2666, 2691, 2697, 2706, 2712, 2723, + 2729, 2738, 2744, 2764, 2770, 2779, 2785, 2796, + 2802, 2811, 2817, 2840, 2846, 2855, 2861, 2872, + 2878, 2887, 2893, 2913, 2919, 2928, 2934, 2945, + 2951, 2960, 2966, 2991, 2997, 3006, 3012, 3023, + 3029, 3038, 3044, 3064, 3070, 3079, 3085, 3096, + 3102, 3111, 3117, 2981, 2987, 2996, 3002, 3013, + 3019, 3028, 3034, 3054, 3060, 3069, 3075, 3086, + 3092, 3101, 3107, 3132, 3138, 3147, 3153, 3164, + 3170, 3179, 3185, 3205, 3211, 3220, 3226, 3237, + 3243, 3252, 3258, 3281, 3287, 3296, 3302, 3313, + 3319, 3328, 3334, 3354, 3360, 3369, 3375, 3386, + 3392, 3401, 3407, 3432, 3438, 3447, 3453, 3464, + 3470, 3479, 3485, 3505, 3511, 3520, 3526, 3537, + 3543, 3552, 3558, 2816, 2822, 2831, 2837, 2848, + 2854, 2863, 2869, 2889, 2895, 2904, 2910, 2921, + 2927, 2936, 2942, 2967, 2973, 2982, 2988, 2999, + 3005, 3014, 3020, 3040, 3046, 3055, 3061, 3072, + 3078, 3087, 3093, 3116, 3122, 3131, 3137, 3148, + 3154, 3163, 3169, 3189, 3195, 3204, 3210, 3221, + 3227, 3236, 3242, 3267, 3273, 3282, 3288, 3299, + 3305, 3314, 3320, 3340, 3346, 3355, 3361, 3372, + 3378, 3387, 3393, 3257, 3263, 3272, 3278, 3289, + 3295, 3304, 3310, 3330, 3336, 3345, 3351, 3362, + 3368, 3377, 3383, 3408, 3414, 3423, 3429, 3440, + 3446, 3455, 3461, 3481, 3487, 3496, 3502, 3513, + 3519, 3528, 3534, 3557, 3563, 3572, 3578, 3589, + 3595, 3604, 3610, 3630, 3636, 3645, 3651, 3662, + 3668, 3677, 3683, 3708, 3714, 3723, 3729, 3740, + 3746, 3755, 3761, 3781, 3787, 3796, 3802, 3813, + 3819, 3828, 3834, 3629, 3635, 3644, 3650, 3661, + 3667, 3676, 3682, 3702, 3708, 3717, 3723, 3734, + 3740, 3749, 3755, 3780, 3786, 3795, 3801, 3812, + 3818, 3827, 3833, 3853, 3859, 3868, 3874, 3885, + 3891, 3900, 3906, 3929, 3935, 3944, 3950, 3961, + 3967, 3976, 3982, 4002, 4008, 4017, 4023, 4034, + 4040, 4049, 4055, 4080, 4086, 4095, 4101, 4112, + 4118, 4127, 4133, 4153, 4159, 4168, 4174, 4185, + 4191, 4200, 4206, 4070, 4076, 4085, 4091, 4102, + 4108, 4117, 4123, 4143, 4149, 4158, 4164, 4175, + 4181, 4190, 4196, 4221, 4227, 4236, 4242, 4253, + 4259, 4268, 4274, 4294, 4300, 4309, 4315, 4326, + 4332, 4341, 4347, 4370, 4376, 4385, 4391, 4402, + 4408, 4417, 4423, 4443, 4449, 4458, 4464, 4475, + 4481, 4490, 4496, 4521, 4527, 4536, 4542, 4553, + 4559, 4568, 4574, 4594, 4600, 4609, 4615, 4626, + 4632, 4641, 4647, 3515, 3521, 3530, 3536, 3547, + 3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620, + 3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698, + 3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771, + 3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847, + 3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920, + 3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998, + 4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071, + 4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988, + 3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061, + 4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139, + 4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212, + 4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288, + 4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361, + 4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439, + 4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512, + 4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360, + 4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433, + 4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511, + 4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584, + 4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660, + 4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733, + 4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811, + 4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884, + 4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801, + 4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874, + 4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952, + 4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025, + 5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101, + 5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174, + 5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252, + 5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325, + 5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636, + 4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709, + 4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787, + 4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860, + 4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936, + 4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009, + 5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087, + 5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160, + 5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077, + 5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150, + 5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228, + 5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301, + 5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377, + 5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450, + 5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528, + 5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601, + 5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449, + 5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522, + 5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600, + 5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673, + 5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749, + 5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822, + 5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900, + 5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973, + 5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890, + 5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963, + 5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041, + 6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114, + 6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190, + 6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263, + 6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341, + 6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414, + 6420, 6429, 6435, 3515, 3521, 3530, 3536, 3547, + 3553, 3562, 3568, 3588, 3594, 3603, 3609, 3620, + 3626, 3635, 3641, 3666, 3672, 3681, 3687, 3698, + 3704, 3713, 3719, 3739, 3745, 3754, 3760, 3771, + 3777, 3786, 3792, 3815, 3821, 3830, 3836, 3847, + 3853, 3862, 3868, 3888, 3894, 3903, 3909, 3920, + 3926, 3935, 3941, 3966, 3972, 3981, 3987, 3998, + 4004, 4013, 4019, 4039, 4045, 4054, 4060, 4071, + 4077, 4086, 4092, 3956, 3962, 3971, 3977, 3988, + 3994, 4003, 4009, 4029, 4035, 4044, 4050, 4061, + 4067, 4076, 4082, 4107, 4113, 4122, 4128, 4139, + 4145, 4154, 4160, 4180, 4186, 4195, 4201, 4212, + 4218, 4227, 4233, 4256, 4262, 4271, 4277, 4288, + 4294, 4303, 4309, 4329, 4335, 4344, 4350, 4361, + 4367, 4376, 4382, 4407, 4413, 4422, 4428, 4439, + 4445, 4454, 4460, 4480, 4486, 4495, 4501, 4512, + 4518, 4527, 4533, 4328, 4334, 4343, 4349, 4360, + 4366, 4375, 4381, 4401, 4407, 4416, 4422, 4433, + 4439, 4448, 4454, 4479, 4485, 4494, 4500, 4511, + 4517, 4526, 4532, 4552, 4558, 4567, 4573, 4584, + 4590, 4599, 4605, 4628, 4634, 4643, 4649, 4660, + 4666, 4675, 4681, 4701, 4707, 4716, 4722, 4733, + 4739, 4748, 4754, 4779, 4785, 4794, 4800, 4811, + 4817, 4826, 4832, 4852, 4858, 4867, 4873, 4884, + 4890, 4899, 4905, 4769, 4775, 4784, 4790, 4801, + 4807, 4816, 4822, 4842, 4848, 4857, 4863, 4874, + 4880, 4889, 4895, 4920, 4926, 4935, 4941, 4952, + 4958, 4967, 4973, 4993, 4999, 5008, 5014, 5025, + 5031, 5040, 5046, 5069, 5075, 5084, 5090, 5101, + 5107, 5116, 5122, 5142, 5148, 5157, 5163, 5174, + 5180, 5189, 5195, 5220, 5226, 5235, 5241, 5252, + 5258, 5267, 5273, 5293, 5299, 5308, 5314, 5325, + 5331, 5340, 5346, 4604, 4610, 4619, 4625, 4636, + 4642, 4651, 4657, 4677, 4683, 4692, 4698, 4709, + 4715, 4724, 4730, 4755, 4761, 4770, 4776, 4787, + 4793, 4802, 4808, 4828, 4834, 4843, 4849, 4860, + 4866, 4875, 4881, 4904, 4910, 4919, 4925, 4936, + 4942, 4951, 4957, 4977, 4983, 4992, 4998, 5009, + 5015, 5024, 5030, 5055, 5061, 5070, 5076, 5087, + 5093, 5102, 5108, 5128, 5134, 5143, 5149, 5160, + 5166, 5175, 5181, 5045, 5051, 5060, 5066, 5077, + 5083, 5092, 5098, 5118, 5124, 5133, 5139, 5150, + 5156, 5165, 5171, 5196, 5202, 5211, 5217, 5228, + 5234, 5243, 5249, 5269, 5275, 5284, 5290, 5301, + 5307, 5316, 5322, 5345, 5351, 5360, 5366, 5377, + 5383, 5392, 5398, 5418, 5424, 5433, 5439, 5450, + 5456, 5465, 5471, 5496, 5502, 5511, 5517, 5528, + 5534, 5543, 5549, 5569, 5575, 5584, 5590, 5601, + 5607, 5616, 5622, 5417, 5423, 5432, 5438, 5449, + 5455, 5464, 5470, 5490, 5496, 5505, 5511, 5522, + 5528, 5537, 5543, 5568, 5574, 5583, 5589, 5600, + 5606, 5615, 5621, 5641, 5647, 5656, 5662, 5673, + 5679, 5688, 5694, 5717, 5723, 5732, 5738, 5749, + 5755, 5764, 5770, 5790, 5796, 5805, 5811, 5822, + 5828, 5837, 5843, 5868, 5874, 5883, 5889, 5900, + 5906, 5915, 5921, 5941, 5947, 5956, 5962, 5973, + 5979, 5988, 5994, 5858, 5864, 5873, 5879, 5890, + 5896, 5905, 5911, 5931, 5937, 5946, 5952, 5963, + 5969, 5978, 5984, 6009, 6015, 6024, 6030, 6041, + 6047, 6056, 6062, 6082, 6088, 6097, 6103, 6114, + 6120, 6129, 6135, 6158, 6164, 6173, 6179, 6190, + 6196, 6205, 6211, 6231, 6237, 6246, 6252, 6263, + 6269, 6278, 6284, 6309, 6315, 6324, 6330, 6341, + 6347, 6356, 6362, 6382, 6388, 6397, 6403, 6414, + 6420, 6429, 6435, 5303, 5309, 5318, 5324, 5335, + 5341, 5350, 5356, 5376, 5382, 5391, 5397, 5408, + 5414, 5423, 5429, 5454, 5460, 5469, 5475, 5486, + 5492, 5501, 5507, 5527, 5533, 5542, 5548, 5559, + 5565, 5574, 5580, 5603, 5609, 5618, 5624, 5635, + 5641, 5650, 5656, 5676, 5682, 5691, 5697, 5708, + 5714, 5723, 5729, 5754, 5760, 5769, 5775, 5786, + 5792, 5801, 5807, 5827, 5833, 5842, 5848, 5859, + 5865, 5874, 5880, 5744, 5750, 5759, 5765, 5776, + 5782, 5791, 5797, 5817, 5823, 5832, 5838, 5849, + 5855, 5864, 5870, 5895, 5901, 5910, 5916, 5927, + 5933, 5942, 5948, 5968, 5974, 5983, 5989, 6000, + 6006, 6015, 6021, 6044, 6050, 6059, 6065, 6076, + 6082, 6091, 6097, 6117, 6123, 6132, 6138, 6149, + 6155, 6164, 6170, 6195, 6201, 6210, 6216, 6227, + 6233, 6242, 6248, 6268, 6274, 6283, 6289, 6300, + 6306, 6315, 6321, 6116, 6122, 6131, 6137, 6148, + 6154, 6163, 6169, 6189, 6195, 6204, 6210, 6221, + 6227, 6236, 6242, 6267, 6273, 6282, 6288, 6299, + 6305, 6314, 6320, 6340, 6346, 6355, 6361, 6372, + 6378, 6387, 6393, 6416, 6422, 6431, 6437, 6448, + 6454, 6463, 6469, 6489, 6495, 6504, 6510, 6521, + 6527, 6536, 6542, 6567, 6573, 6582, 6588, 6599, + 6605, 6614, 6620, 6640, 6646, 6655, 6661, 6672, + 6678, 6687, 6693, 6557, 6563, 6572, 6578, 6589, + 6595, 6604, 6610, 6630, 6636, 6645, 6651, 6662, + 6668, 6677, 6683, 6708, 6714, 6723, 6729, 6740, + 6746, 6755, 6761, 6781, 6787, 6796, 6802, 6813, + 6819, 6828, 6834, 6857, 6863, 6872, 6878, 6889, + 6895, 6904, 6910, 6930, 6936, 6945, 6951, 6962, + 6968, 6977, 6983, 7008, 7014, 7023, 7029, 7040, + 7046, 7055, 7061, 7081, 7087, 7096, 7102, 7113, + 7119, 7128, 7134, 6392, 6398, 6407, 6413, 6424, + 6430, 6439, 6445, 6465, 6471, 6480, 6486, 6497, + 6503, 6512, 6518, 6543, 6549, 6558, 6564, 6575, + 6581, 6590, 6596, 6616, 6622, 6631, 6637, 6648, + 6654, 6663, 6669, 6692, 6698, 6707, 6713, 6724, + 6730, 6739, 6745, 6765, 6771, 6780, 6786, 6797, + 6803, 6812, 6818, 6843, 6849, 6858, 6864, 6875, + 6881, 6890, 6896, 6916, 6922, 6931, 6937, 6948, + 6954, 6963, 6969, 6833, 6839, 6848, 6854, 6865, + 6871, 6880, 6886, 6906, 6912, 6921, 6927, 6938, + 6944, 6953, 6959, 6984, 6990, 6999, 7005, 7016, + 7022, 7031, 7037, 7057, 7063, 7072, 7078, 7089, + 7095, 7104, 7110, 7133, 7139, 7148, 7154, 7165, + 7171, 7180, 7186, 7206, 7212, 7221, 7227, 7238, + 7244, 7253, 7259, 7284, 7290, 7299, 7305, 7316, + 7322, 7331, 7337, 7357, 7363, 7372, 7378, 7389, + 7395, 7404, 7410, 7205, 7211, 7220, 7226, 7237, + 7243, 7252, 7258, 7278, 7284, 7293, 7299, 7310, + 7316, 7325, 7331, 7356, 7362, 7371, 7377, 7388, + 7394, 7403, 7409, 7429, 7435, 7444, 7450, 7461, + 7467, 7476, 7482, 7505, 7511, 7520, 7526, 7537, + 7543, 7552, 7558, 7578, 7584, 7593, 7599, 7610, + 7616, 7625, 7631, 7656, 7662, 7671, 7677, 7688, + 7694, 7703, 7709, 7729, 7735, 7744, 7750, 7761 +}; + +//------------------------------------------------------------------------------ +// Tables for level coding + +const uint8_t VP8EncBands[16 + 1] = { + 0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7, + 0 // sentinel +}; + +//------------------------------------------------------------------------------ +// Mode costs + +static int GetResidualCost_C(int ctx0, const VP8Residual* const res) { + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + const int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + for (; n < res->last; ++n) { + const int v = abs(res->coeffs[n]); + const int ctx = (v >= 2) ? 2 : v; + cost += VP8LevelCost(t, v); + t = costs[n + 1][ctx]; + } + // Last coefficient is always non-zero + { + const int v = abs(res->coeffs[n]); + assert(v != 0); + cost += VP8LevelCost(t, v); + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = (v == 1) ? 1 : 2; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +static void SetResidualCoeffs_C(const int16_t* const coeffs, + VP8Residual* const res) { + int n; + res->last = -1; + assert(res->first == 0 || coeffs[0] == 0); + for (n = 15; n >= 0; --n) { + if (coeffs[n]) { + res->last = n; + break; + } + } + res->coeffs = coeffs; +} + +//------------------------------------------------------------------------------ +// init function + +VP8GetResidualCostFunc VP8GetResidualCost; +VP8SetResidualCoeffsFunc VP8SetResidualCoeffs; + +extern void VP8EncDspCostInitMIPS32(void); +extern void VP8EncDspCostInitMIPSdspR2(void); +extern void VP8EncDspCostInitSSE2(void); +extern void VP8EncDspCostInitNEON(void); + +WEBP_DSP_INIT_FUNC(VP8EncDspCostInit) { + VP8GetResidualCost = GetResidualCost_C; + VP8SetResidualCoeffs = SetResidualCoeffs_C; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + VP8EncDspCostInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8EncDspCostInitMIPSdspR2(); + } +#endif +#if defined(WEBP_HAVE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8EncDspCostInitSSE2(); + } +#endif +#if defined(WEBP_HAVE_NEON) + if (VP8GetCPUInfo(kNEON)) { + VP8EncDspCostInitNEON(); + } +#endif + } +} + +//------------------------------------------------------------------------------ diff --git a/media/libwebp/dsp/cost_mips32.c b/media/libwebp/dsp/cost_mips32.c new file mode 100644 index 0000000000..4e97e8a756 --- /dev/null +++ b/media/libwebp/dsp/cost_mips32.c @@ -0,0 +1,154 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Author: Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "../enc/cost_enc.h" + +static int GetResidualCost_MIPS32(int ctx0, const VP8Residual* const res) { + int temp0, temp1; + int v_reg, ctx_reg; + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + const int16_t* res_coeffs = res->coeffs; + const int res_last = res->last; + const int const_max_level = MAX_VARIABLE_LEVEL; + const int const_2 = 2; + const uint16_t** p_costs = &costs[n][0]; + const size_t inc_p_costs = NUM_CTX * sizeof(*p_costs); + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "subu %[temp1], %[res_last], %[n] \n\t" + "sll %[temp0], %[n], 1 \n\t" + "blez %[temp1], 2f \n\t" + " addu %[res_coeffs], %[res_coeffs], %[temp0] \n\t" + "1: \n\t" + "lh %[v_reg], 0(%[res_coeffs]) \n\t" + "addiu %[n], %[n], 1 \n\t" + "negu %[temp0], %[v_reg] \n\t" + "slti %[temp1], %[v_reg], 0 \n\t" + "movn %[v_reg], %[temp0], %[temp1] \n\t" + "sltiu %[temp0], %[v_reg], 2 \n\t" + "move %[ctx_reg], %[v_reg] \n\t" + "movz %[ctx_reg], %[const_2], %[temp0] \n\t" + "sll %[temp1], %[v_reg], 1 \n\t" + "addu %[temp1], %[temp1], %[VP8LevelFixedCosts] \n\t" + "lhu %[temp1], 0(%[temp1]) \n\t" + "slt %[temp0], %[v_reg], %[const_max_level] \n\t" + "movz %[v_reg], %[const_max_level], %[temp0] \n\t" + "addu %[cost], %[cost], %[temp1] \n\t" + "sll %[v_reg], %[v_reg], 1 \n\t" + "sll %[ctx_reg], %[ctx_reg], 2 \n\t" + "addu %[v_reg], %[v_reg], %[t] \n\t" + "lhu %[temp0], 0(%[v_reg]) \n\t" + "addu %[p_costs], %[p_costs], %[inc_p_costs] \n\t" + "addu %[t], %[p_costs], %[ctx_reg] \n\t" + "addu %[cost], %[cost], %[temp0] \n\t" + "addiu %[res_coeffs], %[res_coeffs], 2 \n\t" + "bne %[n], %[res_last], 1b \n\t" + " lw %[t], 0(%[t]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [cost]"+&r"(cost), [t]"+&r"(t), [n]"+&r"(n), [v_reg]"=&r"(v_reg), + [ctx_reg]"=&r"(ctx_reg), [p_costs]"+&r"(p_costs), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [res_coeffs]"+&r"(res_coeffs) + : [const_2]"r"(const_2), [const_max_level]"r"(const_max_level), + [VP8LevelFixedCosts]"r"(VP8LevelFixedCosts), [res_last]"r"(res_last), + [inc_p_costs]"r"(inc_p_costs) + : "memory" + ); + + // Last coefficient is always non-zero + { + const int v = abs(res->coeffs[n]); + assert(v != 0); + cost += VP8LevelCost(t, v); + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = (v == 1) ? 1 : 2; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +static void SetResidualCoeffs_MIPS32(const int16_t* const coeffs, + VP8Residual* const res) { + const int16_t* p_coeffs = (int16_t*)coeffs; + int temp0, temp1, temp2, n, n1; + assert(res->first == 0 || coeffs[0] == 0); + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "addiu %[p_coeffs], %[p_coeffs], 28 \n\t" + "li %[n], 15 \n\t" + "li %[temp2], -1 \n\t" + "0: \n\t" + "ulw %[temp0], 0(%[p_coeffs]) \n\t" + "beqz %[temp0], 1f \n\t" +#if defined(WORDS_BIGENDIAN) + " sll %[temp1], %[temp0], 16 \n\t" +#else + " srl %[temp1], %[temp0], 16 \n\t" +#endif + "addiu %[n1], %[n], -1 \n\t" + "movz %[temp0], %[n1], %[temp1] \n\t" + "movn %[temp0], %[n], %[temp1] \n\t" + "j 2f \n\t" + " addiu %[temp2], %[temp0], 0 \n\t" + "1: \n\t" + "addiu %[n], %[n], -2 \n\t" + "bgtz %[n], 0b \n\t" + " addiu %[p_coeffs], %[p_coeffs], -4 \n\t" + "2: \n\t" + ".set pop \n\t" + : [p_coeffs]"+&r"(p_coeffs), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [n]"=&r"(n), [n1]"=&r"(n1) + : + : "memory" + ); + res->last = temp2; + res->coeffs = coeffs; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspCostInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitMIPS32(void) { + VP8GetResidualCost = GetResidualCost_MIPS32; + VP8SetResidualCoeffs = SetResidualCoeffs_MIPS32; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8EncDspCostInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/media/libwebp/dsp/cost_mips_dsp_r2.c b/media/libwebp/dsp/cost_mips_dsp_r2.c new file mode 100644 index 0000000000..e9ee99f6ac --- /dev/null +++ b/media/libwebp/dsp/cost_mips_dsp_r2.c @@ -0,0 +1,107 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Author: Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../enc/cost_enc.h" + +static int GetResidualCost_MIPSdspR2(int ctx0, const VP8Residual* const res) { + int temp0, temp1; + int v_reg, ctx_reg; + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + const int16_t* res_coeffs = res->coeffs; + const int res_last = res->last; + const int const_max_level = MAX_VARIABLE_LEVEL; + const int const_2 = 2; + const uint16_t** p_costs = &costs[n][0]; + const size_t inc_p_costs = NUM_CTX * sizeof(*p_costs); + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "subu %[temp1], %[res_last], %[n] \n\t" + "blez %[temp1], 2f \n\t" + " nop \n\t" + "1: \n\t" + "sll %[temp0], %[n], 1 \n\t" + "lhx %[v_reg], %[temp0](%[res_coeffs]) \n\t" + "addiu %[n], %[n], 1 \n\t" + "absq_s.w %[v_reg], %[v_reg] \n\t" + "sltiu %[temp0], %[v_reg], 2 \n\t" + "move %[ctx_reg], %[v_reg] \n\t" + "movz %[ctx_reg], %[const_2], %[temp0] \n\t" + "sll %[temp1], %[v_reg], 1 \n\t" + "lhx %[temp1], %[temp1](%[VP8LevelFixedCosts]) \n\t" + "slt %[temp0], %[v_reg], %[const_max_level] \n\t" + "movz %[v_reg], %[const_max_level], %[temp0] \n\t" + "addu %[cost], %[cost], %[temp1] \n\t" + "sll %[v_reg], %[v_reg], 1 \n\t" + "sll %[ctx_reg], %[ctx_reg], 2 \n\t" + "lhx %[temp0], %[v_reg](%[t]) \n\t" + "addu %[p_costs], %[p_costs], %[inc_p_costs] \n\t" + "addu %[t], %[p_costs], %[ctx_reg] \n\t" + "addu %[cost], %[cost], %[temp0] \n\t" + "bne %[n], %[res_last], 1b \n\t" + " lw %[t], 0(%[t]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [cost]"+&r"(cost), [t]"+&r"(t), [n]"+&r"(n), [v_reg]"=&r"(v_reg), + [ctx_reg]"=&r"(ctx_reg), [p_costs]"+&r"(p_costs), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1) + : [const_2]"r"(const_2), [const_max_level]"r"(const_max_level), + [VP8LevelFixedCosts]"r"(VP8LevelFixedCosts), [res_last]"r"(res_last), + [res_coeffs]"r"(res_coeffs), [inc_p_costs]"r"(inc_p_costs) + : "memory" + ); + + // Last coefficient is always non-zero + { + const int v = abs(res->coeffs[n]); + assert(v != 0); + cost += VP8LevelCost(t, v); + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = (v == 1) ? 1 : 2; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspCostInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitMIPSdspR2(void) { + VP8GetResidualCost = GetResidualCost_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8EncDspCostInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/cost_neon.c b/media/libwebp/dsp/cost_neon.c new file mode 100644 index 0000000000..78f715ff27 --- /dev/null +++ b/media/libwebp/dsp/cost_neon.c @@ -0,0 +1,122 @@ +// Copyright 2018 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// ARM NEON version of cost functions + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_NEON) + +#include "../dsp/neon.h" +#include "../enc/cost_enc.h" + +static const uint8_t position[16] = { 1, 2, 3, 4, 5, 6, 7, 8, + 9, 10, 11, 12, 13, 14, 15, 16 }; + +static void SetResidualCoeffs_NEON(const int16_t* const coeffs, + VP8Residual* const res) { + const int16x8_t minus_one = vdupq_n_s16(-1); + const int16x8_t coeffs_0 = vld1q_s16(coeffs); + const int16x8_t coeffs_1 = vld1q_s16(coeffs + 8); + const uint16x8_t eob_0 = vtstq_s16(coeffs_0, minus_one); + const uint16x8_t eob_1 = vtstq_s16(coeffs_1, minus_one); + const uint8x16_t eob = vcombine_u8(vqmovn_u16(eob_0), vqmovn_u16(eob_1)); + const uint8x16_t masked = vandq_u8(eob, vld1q_u8(position)); + +#ifdef __aarch64__ + res->last = vmaxvq_u8(masked) - 1; +#else + const uint8x8_t eob_8x8 = vmax_u8(vget_low_u8(masked), vget_high_u8(masked)); + const uint16x8_t eob_16x8 = vmovl_u8(eob_8x8); + const uint16x4_t eob_16x4 = + vmax_u16(vget_low_u16(eob_16x8), vget_high_u16(eob_16x8)); + const uint32x4_t eob_32x4 = vmovl_u16(eob_16x4); + uint32x2_t eob_32x2 = + vmax_u32(vget_low_u32(eob_32x4), vget_high_u32(eob_32x4)); + eob_32x2 = vpmax_u32(eob_32x2, eob_32x2); + + vst1_lane_s32(&res->last, vreinterpret_s32_u32(eob_32x2), 0); + --res->last; +#endif // __aarch64__ + + res->coeffs = coeffs; +} + +static int GetResidualCost_NEON(int ctx0, const VP8Residual* const res) { + uint8_t levels[16], ctxs[16]; + uint16_t abs_levels[16]; + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + const int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + + { // precompute clamped levels and contexts, packed to 8b. + const uint8x16_t kCst2 = vdupq_n_u8(2); + const uint8x16_t kCst67 = vdupq_n_u8(MAX_VARIABLE_LEVEL); + const int16x8_t c0 = vld1q_s16(res->coeffs); + const int16x8_t c1 = vld1q_s16(res->coeffs + 8); + const uint16x8_t E0 = vreinterpretq_u16_s16(vabsq_s16(c0)); + const uint16x8_t E1 = vreinterpretq_u16_s16(vabsq_s16(c1)); + const uint8x16_t F = vcombine_u8(vqmovn_u16(E0), vqmovn_u16(E1)); + const uint8x16_t G = vminq_u8(F, kCst2); // context = 0,1,2 + const uint8x16_t H = vminq_u8(F, kCst67); // clamp_level in [0..67] + + vst1q_u8(ctxs, G); + vst1q_u8(levels, H); + + vst1q_u16(abs_levels, E0); + vst1q_u16(abs_levels + 8, E1); + } + for (; n < res->last; ++n) { + const int ctx = ctxs[n]; + const int level = levels[n]; + const int flevel = abs_levels[n]; // full level + cost += VP8LevelFixedCosts[flevel] + t[level]; // simplified VP8LevelCost() + t = costs[n + 1][ctx]; + } + // Last coefficient is always non-zero + { + const int level = levels[n]; + const int flevel = abs_levels[n]; + assert(flevel != 0); + cost += VP8LevelFixedCosts[flevel] + t[level]; + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = ctxs[n]; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspCostInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitNEON(void) { + VP8SetResidualCoeffs = SetResidualCoeffs_NEON; + VP8GetResidualCost = GetResidualCost_NEON; +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(VP8EncDspCostInitNEON) + +#endif // WEBP_USE_NEON diff --git a/media/libwebp/dsp/cost_sse2.c b/media/libwebp/dsp/cost_sse2.c new file mode 100644 index 0000000000..8cfe4e0091 --- /dev/null +++ b/media/libwebp/dsp/cost_sse2.c @@ -0,0 +1,119 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 version of cost functions +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <emmintrin.h> + +#include "../enc/cost_enc.h" +#include "../enc/vp8i_enc.h" +#include "../utils/utils.h" + +//------------------------------------------------------------------------------ + +static void SetResidualCoeffs_SSE2(const int16_t* const coeffs, + VP8Residual* const res) { + const __m128i c0 = _mm_loadu_si128((const __m128i*)(coeffs + 0)); + const __m128i c1 = _mm_loadu_si128((const __m128i*)(coeffs + 8)); + // Use SSE2 to compare 16 values with a single instruction. + const __m128i zero = _mm_setzero_si128(); + const __m128i m0 = _mm_packs_epi16(c0, c1); + const __m128i m1 = _mm_cmpeq_epi8(m0, zero); + // Get the comparison results as a bitmask into 16bits. Negate the mask to get + // the position of entries that are not equal to zero. We don't need to mask + // out least significant bits according to res->first, since coeffs[0] is 0 + // if res->first > 0. + const uint32_t mask = 0x0000ffffu ^ (uint32_t)_mm_movemask_epi8(m1); + // The position of the most significant non-zero bit indicates the position of + // the last non-zero value. + assert(res->first == 0 || coeffs[0] == 0); + res->last = mask ? BitsLog2Floor(mask) : -1; + res->coeffs = coeffs; +} + +static int GetResidualCost_SSE2(int ctx0, const VP8Residual* const res) { + uint8_t levels[16], ctxs[16]; + uint16_t abs_levels[16]; + int n = res->first; + // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 + const int p0 = res->prob[n][ctx0][0]; + CostArrayPtr const costs = res->costs; + const uint16_t* t = costs[n][ctx0]; + // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0 + // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll + // be missing during the loop. + int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0; + + if (res->last < 0) { + return VP8BitCost(0, p0); + } + + { // precompute clamped levels and contexts, packed to 8b. + const __m128i zero = _mm_setzero_si128(); + const __m128i kCst2 = _mm_set1_epi8(2); + const __m128i kCst67 = _mm_set1_epi8(MAX_VARIABLE_LEVEL); + const __m128i c0 = _mm_loadu_si128((const __m128i*)&res->coeffs[0]); + const __m128i c1 = _mm_loadu_si128((const __m128i*)&res->coeffs[8]); + const __m128i D0 = _mm_sub_epi16(zero, c0); + const __m128i D1 = _mm_sub_epi16(zero, c1); + const __m128i E0 = _mm_max_epi16(c0, D0); // abs(v), 16b + const __m128i E1 = _mm_max_epi16(c1, D1); + const __m128i F = _mm_packs_epi16(E0, E1); + const __m128i G = _mm_min_epu8(F, kCst2); // context = 0,1,2 + const __m128i H = _mm_min_epu8(F, kCst67); // clamp_level in [0..67] + + _mm_storeu_si128((__m128i*)&ctxs[0], G); + _mm_storeu_si128((__m128i*)&levels[0], H); + + _mm_storeu_si128((__m128i*)&abs_levels[0], E0); + _mm_storeu_si128((__m128i*)&abs_levels[8], E1); + } + for (; n < res->last; ++n) { + const int ctx = ctxs[n]; + const int level = levels[n]; + const int flevel = abs_levels[n]; // full level + cost += VP8LevelFixedCosts[flevel] + t[level]; // simplified VP8LevelCost() + t = costs[n + 1][ctx]; + } + // Last coefficient is always non-zero + { + const int level = levels[n]; + const int flevel = abs_levels[n]; + assert(flevel != 0); + cost += VP8LevelFixedCosts[flevel] + t[level]; + if (n < 15) { + const int b = VP8EncBands[n + 1]; + const int ctx = ctxs[n]; + const int last_p0 = res->prob[b][ctx][0]; + cost += VP8BitCost(0, last_p0); + } + } + return cost; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspCostInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitSSE2(void) { + VP8SetResidualCoeffs = SetResidualCoeffs_SSE2; + VP8GetResidualCost = GetResidualCost_SSE2; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8EncDspCostInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/media/libwebp/dsp/cpu.c b/media/libwebp/dsp/cpu.c new file mode 100644 index 0000000000..ff57d90224 --- /dev/null +++ b/media/libwebp/dsp/cpu.c @@ -0,0 +1,253 @@ +// Copyright 2011 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// CPU detection +// +// Author: Christian Duvivier (cduvivier@google.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_HAVE_NEON_RTCD) +#include <stdio.h> +#include <string.h> +#endif + +#if defined(WEBP_ANDROID_NEON) +#include <cpu-features.h> +#endif + +//------------------------------------------------------------------------------ +// SSE2 detection. +// + +// apple/darwin gcc-4.0.1 defines __PIC__, but not __pic__ with -fPIC. +#if (defined(__pic__) || defined(__PIC__)) && defined(__i386__) +static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) { + __asm__ volatile ( + "mov %%ebx, %%edi\n" + "cpuid\n" + "xchg %%edi, %%ebx\n" + : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3]) + : "a"(info_type), "c"(0)); +} +#elif defined(__x86_64__) && \ + (defined(__code_model_medium__) || defined(__code_model_large__)) && \ + defined(__PIC__) +static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) { + __asm__ volatile ( + "xchg{q}\t{%%rbx}, %q1\n" + "cpuid\n" + "xchg{q}\t{%%rbx}, %q1\n" + : "=a"(cpu_info[0]), "=&r"(cpu_info[1]), "=c"(cpu_info[2]), + "=d"(cpu_info[3]) + : "a"(info_type), "c"(0)); +} +#elif defined(__i386__) || defined(__x86_64__) +static WEBP_INLINE void GetCPUInfo(int cpu_info[4], int info_type) { + __asm__ volatile ( + "cpuid\n" + : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3]) + : "a"(info_type), "c"(0)); +} +#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) + +#if defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 150030729 // >= VS2008 SP1 +#include <intrin.h> +#define GetCPUInfo(info, type) __cpuidex(info, type, 0) // set ecx=0 +#define WEBP_HAVE_MSC_CPUID +#elif _MSC_VER > 1310 +#include <intrin.h> +#define GetCPUInfo __cpuid +#define WEBP_HAVE_MSC_CPUID +#endif + +#endif + +// NaCl has no support for xgetbv or the raw opcode. +#if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__)) +static WEBP_INLINE uint64_t xgetbv(void) { + const uint32_t ecx = 0; + uint32_t eax, edx; + // Use the raw opcode for xgetbv for compatibility with older toolchains. + __asm__ volatile ( + ".byte 0x0f, 0x01, 0xd0\n" + : "=a"(eax), "=d"(edx) : "c" (ecx)); + return ((uint64_t)edx << 32) | eax; +} +#elif (defined(_M_X64) || defined(_M_IX86)) && \ + defined(_MSC_FULL_VER) && _MSC_FULL_VER >= 160040219 // >= VS2010 SP1 +#include <immintrin.h> +#define xgetbv() _xgetbv(0) +#elif defined(_MSC_VER) && defined(_M_IX86) +static WEBP_INLINE uint64_t xgetbv(void) { + uint32_t eax_, edx_; + __asm { + xor ecx, ecx // ecx = 0 + // Use the raw opcode for xgetbv for compatibility with older toolchains. + __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0 + mov eax_, eax + mov edx_, edx + } + return ((uint64_t)edx_ << 32) | eax_; +} +#else +#define xgetbv() 0U // no AVX for older x64 or unrecognized toolchains. +#endif + +#if defined(__i386__) || defined(__x86_64__) || defined(WEBP_HAVE_MSC_CPUID) + +// helper function for run-time detection of slow SSSE3 platforms +static int CheckSlowModel(int info) { + // Table listing display models with longer latencies for the bsr instruction + // (ie 2 cycles vs 10/16 cycles) and some SSSE3 instructions like pshufb. + // Refer to Intel 64 and IA-32 Architectures Optimization Reference Manual. + static const uint8_t kSlowModels[] = { + 0x37, 0x4a, 0x4d, // Silvermont Microarchitecture + 0x1c, 0x26, 0x27 // Atom Microarchitecture + }; + const uint32_t model = ((info & 0xf0000) >> 12) | ((info >> 4) & 0xf); + const uint32_t family = (info >> 8) & 0xf; + if (family == 0x06) { + size_t i; + for (i = 0; i < sizeof(kSlowModels) / sizeof(kSlowModels[0]); ++i) { + if (model == kSlowModels[i]) return 1; + } + } + return 0; +} + +static int x86CPUInfo(CPUFeature feature) { + int max_cpuid_value; + int cpu_info[4]; + int is_intel = 0; + + // get the highest feature value cpuid supports + GetCPUInfo(cpu_info, 0); + max_cpuid_value = cpu_info[0]; + if (max_cpuid_value < 1) { + return 0; + } else { + const int VENDOR_ID_INTEL_EBX = 0x756e6547; // uneG + const int VENDOR_ID_INTEL_EDX = 0x49656e69; // Ieni + const int VENDOR_ID_INTEL_ECX = 0x6c65746e; // letn + is_intel = (cpu_info[1] == VENDOR_ID_INTEL_EBX && + cpu_info[2] == VENDOR_ID_INTEL_ECX && + cpu_info[3] == VENDOR_ID_INTEL_EDX); // genuine Intel? + } + + GetCPUInfo(cpu_info, 1); + if (feature == kSSE2) { + return !!(cpu_info[3] & (1 << 26)); + } + if (feature == kSSE3) { + return !!(cpu_info[2] & (1 << 0)); + } + if (feature == kSlowSSSE3) { + if (is_intel && (cpu_info[2] & (1 << 9))) { // SSSE3? + return CheckSlowModel(cpu_info[0]); + } + return 0; + } + + if (feature == kSSE4_1) { + return !!(cpu_info[2] & (1 << 19)); + } + if (feature == kAVX) { + // bits 27 (OSXSAVE) & 28 (256-bit AVX) + if ((cpu_info[2] & 0x18000000) == 0x18000000) { + // XMM state and YMM state enabled by the OS. + return (xgetbv() & 0x6) == 0x6; + } + } + if (feature == kAVX2) { + if (x86CPUInfo(kAVX) && max_cpuid_value >= 7) { + GetCPUInfo(cpu_info, 7); + return !!(cpu_info[1] & (1 << 5)); + } + } + return 0; +} +VP8CPUInfo VP8GetCPUInfo = x86CPUInfo; +#elif defined(WEBP_ANDROID_NEON) // NB: needs to be before generic NEON test. +static int AndroidCPUInfo(CPUFeature feature) { + const AndroidCpuFamily cpu_family = android_getCpuFamily(); + const uint64_t cpu_features = android_getCpuFeatures(); + if (feature == kNEON) { + return cpu_family == ANDROID_CPU_FAMILY_ARM && + (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON) != 0; + } + return 0; +} +VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo; +#elif defined(EMSCRIPTEN) // also needs to be before generic NEON test +// Use compile flags as an indicator of SIMD support instead of a runtime check. +static int wasmCPUInfo(CPUFeature feature) { + switch (feature) { +#ifdef WEBP_HAVE_SSE2 + case kSSE2: + return 1; +#endif +#ifdef WEBP_HAVE_SSE41 + case kSSE3: + case kSlowSSSE3: + case kSSE4_1: + return 1; +#endif +#ifdef WEBP_HAVE_NEON + case kNEON: + return 1; +#endif + default: + break; + } + return 0; +} +VP8CPUInfo VP8GetCPUInfo = wasmCPUInfo; +#elif defined(WEBP_HAVE_NEON) +// In most cases this function doesn't check for NEON support (it's assumed by +// the configuration), but enables turning off NEON at runtime, for testing +// purposes, by setting VP8DecGetCPUInfo = NULL. +static int armCPUInfo(CPUFeature feature) { + if (feature != kNEON) return 0; +#if defined(__linux__) && defined(WEBP_HAVE_NEON_RTCD) + { + int has_neon = 0; + char line[200]; + FILE* const cpuinfo = fopen("/proc/cpuinfo", "r"); + if (cpuinfo == NULL) return 0; + while (fgets(line, sizeof(line), cpuinfo)) { + if (!strncmp(line, "Features", 8)) { + if (strstr(line, " neon ") != NULL) { + has_neon = 1; + break; + } + } + } + fclose(cpuinfo); + return has_neon; + } +#else + return 1; +#endif +} +VP8CPUInfo VP8GetCPUInfo = armCPUInfo; +#elif defined(WEBP_USE_MIPS32) || defined(WEBP_USE_MIPS_DSP_R2) || \ + defined(WEBP_USE_MSA) +static int mipsCPUInfo(CPUFeature feature) { + if ((feature == kMIPS32) || (feature == kMIPSdspR2) || (feature == kMSA)) { + return 1; + } else { + return 0; + } + +} +VP8CPUInfo VP8GetCPUInfo = mipsCPUInfo; +#else +VP8CPUInfo VP8GetCPUInfo = NULL; +#endif diff --git a/media/libwebp/dsp/dec.c b/media/libwebp/dsp/dec.c index a599d26bc0..5c94b6d403 100644 --- a/media/libwebp/dsp/dec.c +++ b/media/libwebp/dsp/dec.c @@ -807,10 +807,10 @@ WEBP_DSP_INIT_FUNC(VP8DspInit) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8DspInitSSE2(); -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { VP8DspInitSSE41(); } @@ -834,7 +834,7 @@ WEBP_DSP_INIT_FUNC(VP8DspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8DspInitNEON(); diff --git a/media/libwebp/dsp/dec_mips32.c b/media/libwebp/dsp/dec_mips32.c new file mode 100644 index 0000000000..2d55214faa --- /dev/null +++ b/media/libwebp/dsp/dec_mips32.c @@ -0,0 +1,587 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of dsp functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "../dsp/mips_macro.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +static WEBP_INLINE int abs_mips32(int x) { + const int sign = x >> 31; + return (x ^ sign) - sign; +} + +// 4 pixels in, 2 pixels out +static WEBP_INLINE void do_filter2(uint8_t* p, int step) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + const int a = 3 * (q0 - p0) + VP8ksclip1[p1 - q1]; + const int a1 = VP8ksclip2[(a + 4) >> 3]; + const int a2 = VP8ksclip2[(a + 3) >> 3]; + p[-step] = VP8kclip1[p0 + a2]; + p[ 0] = VP8kclip1[q0 - a1]; +} + +// 4 pixels in, 4 pixels out +static WEBP_INLINE void do_filter4(uint8_t* p, int step) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + const int a = 3 * (q0 - p0); + const int a1 = VP8ksclip2[(a + 4) >> 3]; + const int a2 = VP8ksclip2[(a + 3) >> 3]; + const int a3 = (a1 + 1) >> 1; + p[-2 * step] = VP8kclip1[p1 + a3]; + p[- step] = VP8kclip1[p0 + a2]; + p[ 0] = VP8kclip1[q0 - a1]; + p[ step] = VP8kclip1[q1 - a3]; +} + +// 6 pixels in, 6 pixels out +static WEBP_INLINE void do_filter6(uint8_t* p, int step) { + const int p2 = p[-3 * step], p1 = p[-2 * step], p0 = p[-step]; + const int q0 = p[0], q1 = p[step], q2 = p[2 * step]; + const int a = VP8ksclip1[3 * (q0 - p0) + VP8ksclip1[p1 - q1]]; + // a is in [-128,127], a1 in [-27,27], a2 in [-18,18] and a3 in [-9,9] + const int a1 = (27 * a + 63) >> 7; // eq. to ((3 * a + 7) * 9) >> 7 + const int a2 = (18 * a + 63) >> 7; // eq. to ((2 * a + 7) * 9) >> 7 + const int a3 = (9 * a + 63) >> 7; // eq. to ((1 * a + 7) * 9) >> 7 + p[-3 * step] = VP8kclip1[p2 + a3]; + p[-2 * step] = VP8kclip1[p1 + a2]; + p[- step] = VP8kclip1[p0 + a1]; + p[ 0] = VP8kclip1[q0 - a1]; + p[ step] = VP8kclip1[q1 - a2]; + p[ 2 * step] = VP8kclip1[q2 - a3]; +} + +static WEBP_INLINE int hev(const uint8_t* p, int step, int thresh) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + return (abs_mips32(p1 - p0) > thresh) || (abs_mips32(q1 - q0) > thresh); +} + +static WEBP_INLINE int needs_filter(const uint8_t* p, int step, int t) { + const int p1 = p[-2 * step], p0 = p[-step], q0 = p[0], q1 = p[step]; + return ((4 * abs_mips32(p0 - q0) + abs_mips32(p1 - q1)) <= t); +} + +static WEBP_INLINE int needs_filter2(const uint8_t* p, + int step, int t, int it) { + const int p3 = p[-4 * step], p2 = p[-3 * step]; + const int p1 = p[-2 * step], p0 = p[-step]; + const int q0 = p[0], q1 = p[step], q2 = p[2 * step], q3 = p[3 * step]; + if ((4 * abs_mips32(p0 - q0) + abs_mips32(p1 - q1)) > t) { + return 0; + } + return abs_mips32(p3 - p2) <= it && abs_mips32(p2 - p1) <= it && + abs_mips32(p1 - p0) <= it && abs_mips32(q3 - q2) <= it && + abs_mips32(q2 - q1) <= it && abs_mips32(q1 - q0) <= it; +} + +static WEBP_INLINE void FilterLoop26(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; + while (size-- > 0) { + if (needs_filter2(p, hstride, thresh2, ithresh)) { + if (hev(p, hstride, hev_thresh)) { + do_filter2(p, hstride); + } else { + do_filter6(p, hstride); + } + } + p += vstride; + } +} + +static WEBP_INLINE void FilterLoop24(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; + while (size-- > 0) { + if (needs_filter2(p, hstride, thresh2, ithresh)) { + if (hev(p, hstride, hev_thresh)) { + do_filter2(p, hstride); + } else { + do_filter4(p, hstride); + } + } + p += vstride; + } +} + +// on macroblock edges +static void VFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, stride, 1, 16, thresh, ithresh, hev_thresh); +} + +static void HFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, 1, stride, 16, thresh, ithresh, hev_thresh); +} + +// 8-pixels wide variant, for chroma filtering +static void VFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +static void VFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +// on three inner edges +static void VFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh); + } +} + +static void HFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh); + } +} + +//------------------------------------------------------------------------------ +// Simple In-loop filtering (Paragraph 15.2) + +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + for (i = 0; i < 16; ++i) { + if (needs_filter(p + i, stride, thresh2)) { + do_filter2(p + i, stride); + } + } +} + +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + for (i = 0; i < 16; ++i) { + if (needs_filter(p + i * stride, 1, thresh2)) { + do_filter2(p + i * stride, 1); + } + } +} + +static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + SimpleVFilter16(p, stride, thresh); + } +} + +static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + SimpleHFilter16(p, stride, thresh); + } +} + +static void TransformOne(const int16_t* in, uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14; + int temp15, temp16, temp17, temp18; + int16_t* p_in = (int16_t*)in; + + // loops unrolled and merged to avoid usage of tmp buffer + // and to reduce number of stalls. MUL macro is written + // in assembler and inlined + __asm__ volatile( + "lh %[temp0], 0(%[in]) \n\t" + "lh %[temp8], 16(%[in]) \n\t" + "lh %[temp4], 8(%[in]) \n\t" + "lh %[temp12], 24(%[in]) \n\t" + "addu %[temp16], %[temp0], %[temp8] \n\t" + "subu %[temp0], %[temp0], %[temp8] \n\t" + "mul %[temp8], %[temp4], %[kC2] \n\t" + "mul %[temp17], %[temp12], %[kC1] \n\t" + "mul %[temp4], %[temp4], %[kC1] \n\t" + "mul %[temp12], %[temp12], %[kC2] \n\t" + "lh %[temp1], 2(%[in]) \n\t" + "lh %[temp5], 10(%[in]) \n\t" + "lh %[temp9], 18(%[in]) \n\t" + "lh %[temp13], 26(%[in]) \n\t" + "sra %[temp8], %[temp8], 16 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp4], %[temp4], 16 \n\t" + "sra %[temp12], %[temp12], 16 \n\t" + "lh %[temp2], 4(%[in]) \n\t" + "lh %[temp6], 12(%[in]) \n\t" + "lh %[temp10], 20(%[in]) \n\t" + "lh %[temp14], 28(%[in]) \n\t" + "subu %[temp17], %[temp8], %[temp17] \n\t" + "addu %[temp4], %[temp4], %[temp12] \n\t" + "addu %[temp8], %[temp16], %[temp4] \n\t" + "subu %[temp4], %[temp16], %[temp4] \n\t" + "addu %[temp16], %[temp1], %[temp9] \n\t" + "subu %[temp1], %[temp1], %[temp9] \n\t" + "lh %[temp3], 6(%[in]) \n\t" + "lh %[temp7], 14(%[in]) \n\t" + "lh %[temp11], 22(%[in]) \n\t" + "lh %[temp15], 30(%[in]) \n\t" + "addu %[temp12], %[temp0], %[temp17] \n\t" + "subu %[temp0], %[temp0], %[temp17] \n\t" + "mul %[temp9], %[temp5], %[kC2] \n\t" + "mul %[temp17], %[temp13], %[kC1] \n\t" + "mul %[temp5], %[temp5], %[kC1] \n\t" + "mul %[temp13], %[temp13], %[kC2] \n\t" + "sra %[temp9], %[temp9], 16 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "subu %[temp17], %[temp9], %[temp17] \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "sra %[temp13], %[temp13], 16 \n\t" + "addu %[temp5], %[temp5], %[temp13] \n\t" + "addu %[temp13], %[temp1], %[temp17] \n\t" + "subu %[temp1], %[temp1], %[temp17] \n\t" + "mul %[temp17], %[temp14], %[kC1] \n\t" + "mul %[temp14], %[temp14], %[kC2] \n\t" + "addu %[temp9], %[temp16], %[temp5] \n\t" + "subu %[temp5], %[temp16], %[temp5] \n\t" + "addu %[temp16], %[temp2], %[temp10] \n\t" + "subu %[temp2], %[temp2], %[temp10] \n\t" + "mul %[temp10], %[temp6], %[kC2] \n\t" + "mul %[temp6], %[temp6], %[kC1] \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp14], %[temp14], 16 \n\t" + "sra %[temp10], %[temp10], 16 \n\t" + "sra %[temp6], %[temp6], 16 \n\t" + "subu %[temp17], %[temp10], %[temp17] \n\t" + "addu %[temp6], %[temp6], %[temp14] \n\t" + "addu %[temp10], %[temp16], %[temp6] \n\t" + "subu %[temp6], %[temp16], %[temp6] \n\t" + "addu %[temp14], %[temp2], %[temp17] \n\t" + "subu %[temp2], %[temp2], %[temp17] \n\t" + "mul %[temp17], %[temp15], %[kC1] \n\t" + "mul %[temp15], %[temp15], %[kC2] \n\t" + "addu %[temp16], %[temp3], %[temp11] \n\t" + "subu %[temp3], %[temp3], %[temp11] \n\t" + "mul %[temp11], %[temp7], %[kC2] \n\t" + "mul %[temp7], %[temp7], %[kC1] \n\t" + "addiu %[temp8], %[temp8], 4 \n\t" + "addiu %[temp12], %[temp12], 4 \n\t" + "addiu %[temp0], %[temp0], 4 \n\t" + "addiu %[temp4], %[temp4], 4 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp15], %[temp15], 16 \n\t" + "sra %[temp11], %[temp11], 16 \n\t" + "sra %[temp7], %[temp7], 16 \n\t" + "subu %[temp17], %[temp11], %[temp17] \n\t" + "addu %[temp7], %[temp7], %[temp15] \n\t" + "addu %[temp15], %[temp3], %[temp17] \n\t" + "subu %[temp3], %[temp3], %[temp17] \n\t" + "addu %[temp11], %[temp16], %[temp7] \n\t" + "subu %[temp7], %[temp16], %[temp7] \n\t" + "addu %[temp16], %[temp8], %[temp10] \n\t" + "subu %[temp8], %[temp8], %[temp10] \n\t" + "mul %[temp10], %[temp9], %[kC2] \n\t" + "mul %[temp17], %[temp11], %[kC1] \n\t" + "mul %[temp9], %[temp9], %[kC1] \n\t" + "mul %[temp11], %[temp11], %[kC2] \n\t" + "sra %[temp10], %[temp10], 16 \n\t" + "sra %[temp17], %[temp17], 16 \n\t" + "sra %[temp9], %[temp9], 16 \n\t" + "sra %[temp11], %[temp11], 16 \n\t" + "subu %[temp17], %[temp10], %[temp17] \n\t" + "addu %[temp11], %[temp9], %[temp11] \n\t" + "addu %[temp10], %[temp12], %[temp14] \n\t" + "subu %[temp12], %[temp12], %[temp14] \n\t" + "mul %[temp14], %[temp13], %[kC2] \n\t" + "mul %[temp9], %[temp15], %[kC1] \n\t" + "mul %[temp13], %[temp13], %[kC1] \n\t" + "mul %[temp15], %[temp15], %[kC2] \n\t" + "sra %[temp14], %[temp14], 16 \n\t" + "sra %[temp9], %[temp9], 16 \n\t" + "sra %[temp13], %[temp13], 16 \n\t" + "sra %[temp15], %[temp15], 16 \n\t" + "subu %[temp9], %[temp14], %[temp9] \n\t" + "addu %[temp15], %[temp13], %[temp15] \n\t" + "addu %[temp14], %[temp0], %[temp2] \n\t" + "subu %[temp0], %[temp0], %[temp2] \n\t" + "mul %[temp2], %[temp1], %[kC2] \n\t" + "mul %[temp13], %[temp3], %[kC1] \n\t" + "mul %[temp1], %[temp1], %[kC1] \n\t" + "mul %[temp3], %[temp3], %[kC2] \n\t" + "sra %[temp2], %[temp2], 16 \n\t" + "sra %[temp13], %[temp13], 16 \n\t" + "sra %[temp1], %[temp1], 16 \n\t" + "sra %[temp3], %[temp3], 16 \n\t" + "subu %[temp13], %[temp2], %[temp13] \n\t" + "addu %[temp3], %[temp1], %[temp3] \n\t" + "addu %[temp2], %[temp4], %[temp6] \n\t" + "subu %[temp4], %[temp4], %[temp6] \n\t" + "mul %[temp6], %[temp5], %[kC2] \n\t" + "mul %[temp1], %[temp7], %[kC1] \n\t" + "mul %[temp5], %[temp5], %[kC1] \n\t" + "mul %[temp7], %[temp7], %[kC2] \n\t" + "sra %[temp6], %[temp6], 16 \n\t" + "sra %[temp1], %[temp1], 16 \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "sra %[temp7], %[temp7], 16 \n\t" + "subu %[temp1], %[temp6], %[temp1] \n\t" + "addu %[temp7], %[temp5], %[temp7] \n\t" + "addu %[temp5], %[temp16], %[temp11] \n\t" + "subu %[temp16], %[temp16], %[temp11] \n\t" + "addu %[temp11], %[temp8], %[temp17] \n\t" + "subu %[temp8], %[temp8], %[temp17] \n\t" + "sra %[temp5], %[temp5], 3 \n\t" + "sra %[temp16], %[temp16], 3 \n\t" + "sra %[temp11], %[temp11], 3 \n\t" + "sra %[temp8], %[temp8], 3 \n\t" + "addu %[temp17], %[temp10], %[temp15] \n\t" + "subu %[temp10], %[temp10], %[temp15] \n\t" + "addu %[temp15], %[temp12], %[temp9] \n\t" + "subu %[temp12], %[temp12], %[temp9] \n\t" + "sra %[temp17], %[temp17], 3 \n\t" + "sra %[temp10], %[temp10], 3 \n\t" + "sra %[temp15], %[temp15], 3 \n\t" + "sra %[temp12], %[temp12], 3 \n\t" + "addu %[temp9], %[temp14], %[temp3] \n\t" + "subu %[temp14], %[temp14], %[temp3] \n\t" + "addu %[temp3], %[temp0], %[temp13] \n\t" + "subu %[temp0], %[temp0], %[temp13] \n\t" + "sra %[temp9], %[temp9], 3 \n\t" + "sra %[temp14], %[temp14], 3 \n\t" + "sra %[temp3], %[temp3], 3 \n\t" + "sra %[temp0], %[temp0], 3 \n\t" + "addu %[temp13], %[temp2], %[temp7] \n\t" + "subu %[temp2], %[temp2], %[temp7] \n\t" + "addu %[temp7], %[temp4], %[temp1] \n\t" + "subu %[temp4], %[temp4], %[temp1] \n\t" + "sra %[temp13], %[temp13], 3 \n\t" + "sra %[temp2], %[temp2], 3 \n\t" + "sra %[temp7], %[temp7], 3 \n\t" + "sra %[temp4], %[temp4], 3 \n\t" + "addiu %[temp6], $zero, 255 \n\t" + "lbu %[temp1], 0+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp1], %[temp1], %[temp5] \n\t" + "sra %[temp5], %[temp1], 8 \n\t" + "sra %[temp18], %[temp1], 31 \n\t" + "beqz %[temp5], 1f \n\t" + "xor %[temp1], %[temp1], %[temp1] \n\t" + "movz %[temp1], %[temp6], %[temp18] \n\t" + "1: \n\t" + "lbu %[temp18], 1+0*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp1], 0+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp18], %[temp18], %[temp11] \n\t" + "sra %[temp11], %[temp18], 8 \n\t" + "sra %[temp1], %[temp18], 31 \n\t" + "beqz %[temp11], 2f \n\t" + "xor %[temp18], %[temp18], %[temp18] \n\t" + "movz %[temp18], %[temp6], %[temp1] \n\t" + "2: \n\t" + "lbu %[temp1], 2+0*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp18], 1+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp1], %[temp1], %[temp8] \n\t" + "sra %[temp8], %[temp1], 8 \n\t" + "sra %[temp18], %[temp1], 31 \n\t" + "beqz %[temp8], 3f \n\t" + "xor %[temp1], %[temp1], %[temp1] \n\t" + "movz %[temp1], %[temp6], %[temp18] \n\t" + "3: \n\t" + "lbu %[temp18], 3+0*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp1], 2+0*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp18], %[temp18], %[temp16] \n\t" + "sra %[temp16], %[temp18], 8 \n\t" + "sra %[temp1], %[temp18], 31 \n\t" + "beqz %[temp16], 4f \n\t" + "xor %[temp18], %[temp18], %[temp18] \n\t" + "movz %[temp18], %[temp6], %[temp1] \n\t" + "4: \n\t" + "sb %[temp18], 3+0*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp5], 0+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp8], 1+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp11], 2+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp16], 3+1*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp5], %[temp5], %[temp17] \n\t" + "addu %[temp8], %[temp8], %[temp15] \n\t" + "addu %[temp11], %[temp11], %[temp12] \n\t" + "addu %[temp16], %[temp16], %[temp10] \n\t" + "sra %[temp18], %[temp5], 8 \n\t" + "sra %[temp1], %[temp5], 31 \n\t" + "beqz %[temp18], 5f \n\t" + "xor %[temp5], %[temp5], %[temp5] \n\t" + "movz %[temp5], %[temp6], %[temp1] \n\t" + "5: \n\t" + "sra %[temp18], %[temp8], 8 \n\t" + "sra %[temp1], %[temp8], 31 \n\t" + "beqz %[temp18], 6f \n\t" + "xor %[temp8], %[temp8], %[temp8] \n\t" + "movz %[temp8], %[temp6], %[temp1] \n\t" + "6: \n\t" + "sra %[temp18], %[temp11], 8 \n\t" + "sra %[temp1], %[temp11], 31 \n\t" + "sra %[temp17], %[temp16], 8 \n\t" + "sra %[temp15], %[temp16], 31 \n\t" + "beqz %[temp18], 7f \n\t" + "xor %[temp11], %[temp11], %[temp11] \n\t" + "movz %[temp11], %[temp6], %[temp1] \n\t" + "7: \n\t" + "beqz %[temp17], 8f \n\t" + "xor %[temp16], %[temp16], %[temp16] \n\t" + "movz %[temp16], %[temp6], %[temp15] \n\t" + "8: \n\t" + "sb %[temp5], 0+1*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp8], 1+1*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp11], 2+1*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp16], 3+1*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp5], 0+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp8], 1+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp11], 2+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp16], 3+2*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp5], %[temp5], %[temp9] \n\t" + "addu %[temp8], %[temp8], %[temp3] \n\t" + "addu %[temp11], %[temp11], %[temp0] \n\t" + "addu %[temp16], %[temp16], %[temp14] \n\t" + "sra %[temp18], %[temp5], 8 \n\t" + "sra %[temp1], %[temp5], 31 \n\t" + "sra %[temp17], %[temp8], 8 \n\t" + "sra %[temp15], %[temp8], 31 \n\t" + "sra %[temp12], %[temp11], 8 \n\t" + "sra %[temp10], %[temp11], 31 \n\t" + "sra %[temp9], %[temp16], 8 \n\t" + "sra %[temp3], %[temp16], 31 \n\t" + "beqz %[temp18], 9f \n\t" + "xor %[temp5], %[temp5], %[temp5] \n\t" + "movz %[temp5], %[temp6], %[temp1] \n\t" + "9: \n\t" + "beqz %[temp17], 10f \n\t" + "xor %[temp8], %[temp8], %[temp8] \n\t" + "movz %[temp8], %[temp6], %[temp15] \n\t" + "10: \n\t" + "beqz %[temp12], 11f \n\t" + "xor %[temp11], %[temp11], %[temp11] \n\t" + "movz %[temp11], %[temp6], %[temp10] \n\t" + "11: \n\t" + "beqz %[temp9], 12f \n\t" + "xor %[temp16], %[temp16], %[temp16] \n\t" + "movz %[temp16], %[temp6], %[temp3] \n\t" + "12: \n\t" + "sb %[temp5], 0+2*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp8], 1+2*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp11], 2+2*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp16], 3+2*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp5], 0+3*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp8], 1+3*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp11], 2+3*" XSTR(BPS) "(%[dst]) \n\t" + "lbu %[temp16], 3+3*" XSTR(BPS) "(%[dst]) \n\t" + "addu %[temp5], %[temp5], %[temp13] \n\t" + "addu %[temp8], %[temp8], %[temp7] \n\t" + "addu %[temp11], %[temp11], %[temp4] \n\t" + "addu %[temp16], %[temp16], %[temp2] \n\t" + "sra %[temp18], %[temp5], 8 \n\t" + "sra %[temp1], %[temp5], 31 \n\t" + "sra %[temp17], %[temp8], 8 \n\t" + "sra %[temp15], %[temp8], 31 \n\t" + "sra %[temp12], %[temp11], 8 \n\t" + "sra %[temp10], %[temp11], 31 \n\t" + "sra %[temp9], %[temp16], 8 \n\t" + "sra %[temp3], %[temp16], 31 \n\t" + "beqz %[temp18], 13f \n\t" + "xor %[temp5], %[temp5], %[temp5] \n\t" + "movz %[temp5], %[temp6], %[temp1] \n\t" + "13: \n\t" + "beqz %[temp17], 14f \n\t" + "xor %[temp8], %[temp8], %[temp8] \n\t" + "movz %[temp8], %[temp6], %[temp15] \n\t" + "14: \n\t" + "beqz %[temp12], 15f \n\t" + "xor %[temp11], %[temp11], %[temp11] \n\t" + "movz %[temp11], %[temp6], %[temp10] \n\t" + "15: \n\t" + "beqz %[temp9], 16f \n\t" + "xor %[temp16], %[temp16], %[temp16] \n\t" + "movz %[temp16], %[temp6], %[temp3] \n\t" + "16: \n\t" + "sb %[temp5], 0+3*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp8], 1+3*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp11], 2+3*" XSTR(BPS) "(%[dst]) \n\t" + "sb %[temp16], 3+3*" XSTR(BPS) "(%[dst]) \n\t" + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11), + [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14), + [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17), + [temp18]"=&r"(temp18) + : [in]"r"(p_in), [kC1]"r"(kC1), [kC2]"r"(kC2), [dst]"r"(dst) + : "memory", "hi", "lo" + ); +} + +static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) { + TransformOne(in, dst); + if (do_two) { + TransformOne(in + 16, dst + 4); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitMIPS32(void) { + VP8InitClipTables(); + + VP8Transform = TransformTwo; + + VP8VFilter16 = VFilter16; + VP8HFilter16 = HFilter16; + VP8VFilter8 = VFilter8; + VP8HFilter8 = HFilter8; + VP8VFilter16i = VFilter16i; + VP8HFilter16i = HFilter16i; + VP8VFilter8i = VFilter8i; + VP8HFilter8i = HFilter8i; + + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8DspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/media/libwebp/dsp/dec_mips_dsp_r2.c b/media/libwebp/dsp/dec_mips_dsp_r2.c new file mode 100644 index 0000000000..dcc3041019 --- /dev/null +++ b/media/libwebp/dsp/dec_mips_dsp_r2.c @@ -0,0 +1,994 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of dsp functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../dsp/mips_macro.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +#define MUL(a, b) (((a) * (b)) >> 16) + +static void TransformDC(const int16_t* in, uint8_t* dst) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9, temp10; + + __asm__ volatile ( + LOAD_WITH_OFFSET_X4(temp1, temp2, temp3, temp4, dst, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + "lh %[temp5], 0(%[in]) \n\t" + "addiu %[temp5], %[temp5], 4 \n\t" + "ins %[temp5], %[temp5], 16, 16 \n\t" + "shra.ph %[temp5], %[temp5], 3 \n\t" + CONVERT_2_BYTES_TO_HALF(temp6, temp7, temp8, temp9, temp10, temp1, temp2, + temp3, temp1, temp2, temp3, temp4) + STORE_SAT_SUM_X2(temp6, temp7, temp8, temp9, temp10, temp1, temp2, temp3, + temp5, temp5, temp5, temp5, temp5, temp5, temp5, temp5, + dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_10() + : [in]"r"(in), [dst]"r"(dst) + : "memory" + ); +} + +static void TransformAC3(const int16_t* in, uint8_t* dst) { + const int a = in[0] + 4; + int c4 = MUL(in[4], kC2); + const int d4 = MUL(in[4], kC1); + const int c1 = MUL(in[1], kC2); + const int d1 = MUL(in[1], kC1); + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18; + + __asm__ volatile ( + "ins %[c4], %[d4], 16, 16 \n\t" + "replv.ph %[temp1], %[a] \n\t" + "replv.ph %[temp4], %[d1] \n\t" + ADD_SUB_HALVES(temp2, temp3, temp1, c4) + "replv.ph %[temp5], %[c1] \n\t" + SHIFT_R_SUM_X2(temp1, temp6, temp7, temp8, temp2, temp9, temp10, temp4, + temp2, temp2, temp3, temp3, temp4, temp5, temp4, temp5) + LOAD_WITH_OFFSET_X4(temp3, temp5, temp11, temp12, dst, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp13, temp14, temp3, temp15, temp5, temp16, + temp11, temp17, temp3, temp5, temp11, temp12) + PACK_2_HALVES_TO_WORD(temp12, temp18, temp7, temp6, temp1, temp8, temp2, + temp4, temp7, temp6, temp10, temp9) + STORE_SAT_SUM_X2(temp13, temp14, temp3, temp15, temp5, temp16, temp11, + temp17, temp12, temp18, temp1, temp8, temp2, temp4, + temp7, temp6, dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_18(), + [c4]"+&r"(c4) + : [dst]"r"(dst), [a]"r"(a), [d1]"r"(d1), [d4]"r"(d4), [c1]"r"(c1) + : "memory" + ); +} + +static void TransformOne(const int16_t* in, uint8_t* dst) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18; + + __asm__ volatile ( + "ulw %[temp1], 0(%[in]) \n\t" + "ulw %[temp2], 16(%[in]) \n\t" + LOAD_IN_X2(temp5, temp6, 24, 26) + ADD_SUB_HALVES(temp3, temp4, temp1, temp2) + LOAD_IN_X2(temp1, temp2, 8, 10) + MUL_SHIFT_SUM(temp7, temp8, temp9, temp10, temp11, temp12, temp13, temp14, + temp10, temp8, temp9, temp7, temp1, temp2, temp5, temp6, + temp13, temp11, temp14, temp12) + INSERT_HALF_X2(temp8, temp7, temp10, temp9) + "ulw %[temp17], 4(%[in]) \n\t" + "ulw %[temp18], 20(%[in]) \n\t" + ADD_SUB_HALVES(temp1, temp2, temp3, temp8) + ADD_SUB_HALVES(temp5, temp6, temp4, temp7) + ADD_SUB_HALVES(temp7, temp8, temp17, temp18) + LOAD_IN_X2(temp17, temp18, 12, 14) + LOAD_IN_X2(temp9, temp10, 28, 30) + MUL_SHIFT_SUM(temp11, temp12, temp13, temp14, temp15, temp16, temp4, temp17, + temp12, temp14, temp11, temp13, temp17, temp18, temp9, temp10, + temp15, temp4, temp16, temp17) + INSERT_HALF_X2(temp11, temp12, temp13, temp14) + ADD_SUB_HALVES(temp17, temp8, temp8, temp11) + ADD_SUB_HALVES(temp3, temp4, temp7, temp12) + + // horizontal + SRA_16(temp9, temp10, temp11, temp12, temp1, temp2, temp5, temp6) + INSERT_HALF_X2(temp1, temp6, temp5, temp2) + SRA_16(temp13, temp14, temp15, temp16, temp3, temp4, temp17, temp8) + "repl.ph %[temp2], 0x4 \n\t" + INSERT_HALF_X2(temp3, temp8, temp17, temp4) + "addq.ph %[temp1], %[temp1], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp2] \n\t" + ADD_SUB_HALVES(temp2, temp4, temp1, temp3) + ADD_SUB_HALVES(temp5, temp7, temp6, temp8) + MUL_SHIFT_SUM(temp1, temp3, temp6, temp8, temp9, temp13, temp17, temp18, + temp3, temp13, temp1, temp9, temp9, temp13, temp11, temp15, + temp6, temp17, temp8, temp18) + MUL_SHIFT_SUM(temp6, temp8, temp18, temp17, temp11, temp15, temp12, temp16, + temp8, temp15, temp6, temp11, temp12, temp16, temp10, temp14, + temp18, temp12, temp17, temp16) + INSERT_HALF_X2(temp1, temp3, temp9, temp13) + INSERT_HALF_X2(temp6, temp8, temp11, temp15) + SHIFT_R_SUM_X2(temp9, temp10, temp11, temp12, temp13, temp14, temp15, + temp16, temp2, temp4, temp5, temp7, temp3, temp1, temp8, + temp6) + PACK_2_HALVES_TO_WORD(temp1, temp2, temp3, temp4, temp9, temp12, temp13, + temp16, temp11, temp10, temp15, temp14) + LOAD_WITH_OFFSET_X4(temp10, temp11, temp14, temp15, dst, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5, temp6, temp7, temp8, temp17, temp18, temp10, + temp11, temp10, temp11, temp14, temp15) + STORE_SAT_SUM_X2(temp5, temp6, temp7, temp8, temp17, temp18, temp10, temp11, + temp9, temp12, temp1, temp2, temp13, temp16, temp3, temp4, + dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_18() + : [dst]"r"(dst), [in]"r"(in), [kC1]"r"(kC1), [kC2]"r"(kC2) + : "memory", "hi", "lo" + ); +} + +static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) { + TransformOne(in, dst); + if (do_two) { + TransformOne(in + 16, dst + 4); + } +} + +static WEBP_INLINE void FilterLoop26(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + const int thresh2 = 2 * thresh + 1; + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15; + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "1: \n\t" + "negu %[temp1], %[hstride] \n\t" + "addiu %[size], %[size], -1 \n\t" + "sll %[temp2], %[hstride], 1 \n\t" + "sll %[temp3], %[temp1], 1 \n\t" + "addu %[temp4], %[temp2], %[hstride] \n\t" + "addu %[temp5], %[temp3], %[temp1] \n\t" + "lbu %[temp7], 0(%[p]) \n\t" + "sll %[temp6], %[temp3], 1 \n\t" + "lbux %[temp8], %[temp5](%[p]) \n\t" + "lbux %[temp9], %[temp3](%[p]) \n\t" + "lbux %[temp10], %[temp1](%[p]) \n\t" + "lbux %[temp11], %[temp6](%[p]) \n\t" + "lbux %[temp12], %[hstride](%[p]) \n\t" + "lbux %[temp13], %[temp2](%[p]) \n\t" + "lbux %[temp14], %[temp4](%[p]) \n\t" + "subu %[temp1], %[temp10], %[temp7] \n\t" + "subu %[temp2], %[temp9], %[temp12] \n\t" + "absq_s.w %[temp3], %[temp1] \n\t" + "absq_s.w %[temp4], %[temp2] \n\t" + "negu %[temp1], %[temp1] \n\t" + "sll %[temp3], %[temp3], 2 \n\t" + "addu %[temp15], %[temp3], %[temp4] \n\t" + "subu %[temp3], %[temp15], %[thresh2] \n\t" + "sll %[temp6], %[temp1], 1 \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp4], %[temp11], %[temp8] \n\t" + "absq_s.w %[temp4], %[temp4] \n\t" + "shll_s.w %[temp2], %[temp2], 24 \n\t" + "subu %[temp4], %[temp4], %[ithresh] \n\t" + "bgtz %[temp4], 3f \n\t" + " subu %[temp3], %[temp8], %[temp9] \n\t" + "absq_s.w %[temp3], %[temp3] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp5], %[temp9], %[temp10] \n\t" + "absq_s.w %[temp3], %[temp5] \n\t" + "absq_s.w %[temp5], %[temp5] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp3], %[temp14], %[temp13] \n\t" + "absq_s.w %[temp3], %[temp3] \n\t" + "slt %[temp5], %[hev_thresh], %[temp5] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp3], %[temp13], %[temp12] \n\t" + "absq_s.w %[temp3], %[temp3] \n\t" + "sra %[temp4], %[temp2], 24 \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " subu %[temp15], %[temp12], %[temp7] \n\t" + "absq_s.w %[temp3], %[temp15] \n\t" + "absq_s.w %[temp15], %[temp15] \n\t" + "subu %[temp3], %[temp3], %[ithresh] \n\t" + "bgtz %[temp3], 3f \n\t" + " slt %[temp15], %[hev_thresh], %[temp15] \n\t" + "addu %[temp3], %[temp6], %[temp1] \n\t" + "or %[temp2], %[temp5], %[temp15] \n\t" + "addu %[temp5], %[temp4], %[temp3] \n\t" + "beqz %[temp2], 4f \n\t" + " shra_r.w %[temp1], %[temp5], 3 \n\t" + "addiu %[temp2], %[temp5], 3 \n\t" + "sra %[temp2], %[temp2], 3 \n\t" + "shll_s.w %[temp1], %[temp1], 27 \n\t" + "shll_s.w %[temp2], %[temp2], 27 \n\t" + "subu %[temp3], %[p], %[hstride] \n\t" + "sra %[temp1], %[temp1], 27 \n\t" + "sra %[temp2], %[temp2], 27 \n\t" + "subu %[temp1], %[temp7], %[temp1] \n\t" + "addu %[temp2], %[temp10], %[temp2] \n\t" + "lbux %[temp2], %[temp2](%[VP8kclip1]) \n\t" + "lbux %[temp1], %[temp1](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[temp3]) \n\t" + "j 3f \n\t" + " sb %[temp1], 0(%[p]) \n\t" + "4: \n\t" + "shll_s.w %[temp5], %[temp5], 24 \n\t" + "subu %[temp14], %[p], %[hstride] \n\t" + "subu %[temp11], %[temp14], %[hstride] \n\t" + "sra %[temp6], %[temp5], 24 \n\t" + "sll %[temp1], %[temp6], 3 \n\t" + "subu %[temp15], %[temp11], %[hstride] \n\t" + "addu %[temp2], %[temp6], %[temp1] \n\t" + "sll %[temp3], %[temp2], 1 \n\t" + "addu %[temp4], %[temp3], %[temp2] \n\t" + "addiu %[temp2], %[temp2], 63 \n\t" + "addiu %[temp3], %[temp3], 63 \n\t" + "addiu %[temp4], %[temp4], 63 \n\t" + "sra %[temp2], %[temp2], 7 \n\t" + "sra %[temp3], %[temp3], 7 \n\t" + "sra %[temp4], %[temp4], 7 \n\t" + "addu %[temp1], %[temp8], %[temp2] \n\t" + "addu %[temp5], %[temp9], %[temp3] \n\t" + "addu %[temp6], %[temp10], %[temp4] \n\t" + "subu %[temp8], %[temp7], %[temp4] \n\t" + "subu %[temp7], %[temp12], %[temp3] \n\t" + "addu %[temp10], %[p], %[hstride] \n\t" + "subu %[temp9], %[temp13], %[temp2] \n\t" + "addu %[temp12], %[temp10], %[hstride] \n\t" + "lbux %[temp2], %[temp1](%[VP8kclip1]) \n\t" + "lbux %[temp3], %[temp5](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[temp6](%[VP8kclip1]) \n\t" + "lbux %[temp5], %[temp8](%[VP8kclip1]) \n\t" + "lbux %[temp6], %[temp7](%[VP8kclip1]) \n\t" + "lbux %[temp8], %[temp9](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[temp15]) \n\t" + "sb %[temp3], 0(%[temp11]) \n\t" + "sb %[temp4], 0(%[temp14]) \n\t" + "sb %[temp5], 0(%[p]) \n\t" + "sb %[temp6], 0(%[temp10]) \n\t" + "sb %[temp8], 0(%[temp12]) \n\t" + "3: \n\t" + "bgtz %[size], 1b \n\t" + " addu %[p], %[p], %[vstride] \n\t" + ".set pop \n\t" + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),[temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), + [temp7]"=&r"(temp7),[temp8]"=&r"(temp8),[temp9]"=&r"(temp9), + [temp10]"=&r"(temp10),[temp11]"=&r"(temp11),[temp12]"=&r"(temp12), + [temp13]"=&r"(temp13),[temp14]"=&r"(temp14),[temp15]"=&r"(temp15), + [size]"+&r"(size), [p]"+&r"(p) + : [hstride]"r"(hstride), [thresh2]"r"(thresh2), + [ithresh]"r"(ithresh),[vstride]"r"(vstride), [hev_thresh]"r"(hev_thresh), + [VP8kclip1]"r"(VP8kclip1) + : "memory" + ); +} + +static WEBP_INLINE void FilterLoop24(uint8_t* p, + int hstride, int vstride, int size, + int thresh, int ithresh, int hev_thresh) { + int p0, q0, p1, q1, p2, q2, p3, q3; + int step1, step2, temp1, temp2, temp3, temp4; + uint8_t* pTemp0; + uint8_t* pTemp1; + const int thresh2 = 2 * thresh + 1; + + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "bltz %[size], 3f \n\t" + " nop \n\t" + "2: \n\t" + "negu %[step1], %[hstride] \n\t" + "lbu %[q0], 0(%[p]) \n\t" + "lbux %[p0], %[step1](%[p]) \n\t" + "subu %[step1], %[step1], %[hstride] \n\t" + "lbux %[q1], %[hstride](%[p]) \n\t" + "subu %[temp1], %[p0], %[q0] \n\t" + "lbux %[p1], %[step1](%[p]) \n\t" + "addu %[step2], %[hstride], %[hstride] \n\t" + "absq_s.w %[temp2], %[temp1] \n\t" + "subu %[temp3], %[p1], %[q1] \n\t" + "absq_s.w %[temp4], %[temp3] \n\t" + "sll %[temp2], %[temp2], 2 \n\t" + "addu %[temp2], %[temp2], %[temp4] \n\t" + "subu %[temp4], %[temp2], %[thresh2] \n\t" + "subu %[step1], %[step1], %[hstride] \n\t" + "bgtz %[temp4], 0f \n\t" + " lbux %[p2], %[step1](%[p]) \n\t" + "subu %[step1], %[step1], %[hstride] \n\t" + "lbux %[q2], %[step2](%[p]) \n\t" + "lbux %[p3], %[step1](%[p]) \n\t" + "subu %[temp4], %[p2], %[p1] \n\t" + "addu %[step2], %[step2], %[hstride] \n\t" + "subu %[temp2], %[p3], %[p2] \n\t" + "absq_s.w %[temp4], %[temp4] \n\t" + "absq_s.w %[temp2], %[temp2] \n\t" + "lbux %[q3], %[step2](%[p]) \n\t" + "subu %[temp4], %[temp4], %[ithresh] \n\t" + "negu %[temp1], %[temp1] \n\t" + "bgtz %[temp4], 0f \n\t" + " subu %[temp2], %[temp2], %[ithresh] \n\t" + "subu %[p3], %[p1], %[p0] \n\t" + "bgtz %[temp2], 0f \n\t" + " absq_s.w %[p3], %[p3] \n\t" + "subu %[temp4], %[q3], %[q2] \n\t" + "subu %[pTemp0], %[p], %[hstride] \n\t" + "absq_s.w %[temp4], %[temp4] \n\t" + "subu %[temp2], %[p3], %[ithresh] \n\t" + "sll %[step1], %[temp1], 1 \n\t" + "bgtz %[temp2], 0f \n\t" + " subu %[temp4], %[temp4], %[ithresh] \n\t" + "subu %[temp2], %[q2], %[q1] \n\t" + "bgtz %[temp4], 0f \n\t" + " absq_s.w %[temp2], %[temp2] \n\t" + "subu %[q3], %[q1], %[q0] \n\t" + "absq_s.w %[q3], %[q3] \n\t" + "subu %[temp2], %[temp2], %[ithresh] \n\t" + "addu %[temp1], %[temp1], %[step1] \n\t" + "bgtz %[temp2], 0f \n\t" + " subu %[temp4], %[q3], %[ithresh] \n\t" + "slt %[p3], %[hev_thresh], %[p3] \n\t" + "bgtz %[temp4], 0f \n\t" + " slt %[q3], %[hev_thresh], %[q3] \n\t" + "or %[q3], %[q3], %[p3] \n\t" + "bgtz %[q3], 1f \n\t" + " shra_r.w %[temp2], %[temp1], 3 \n\t" + "addiu %[temp1], %[temp1], 3 \n\t" + "sra %[temp1], %[temp1], 3 \n\t" + "shll_s.w %[temp2], %[temp2], 27 \n\t" + "shll_s.w %[temp1], %[temp1], 27 \n\t" + "addu %[pTemp1], %[p], %[hstride] \n\t" + "sra %[temp2], %[temp2], 27 \n\t" + "sra %[temp1], %[temp1], 27 \n\t" + "addiu %[step1], %[temp2], 1 \n\t" + "sra %[step1], %[step1], 1 \n\t" + "addu %[p0], %[p0], %[temp1] \n\t" + "addu %[p1], %[p1], %[step1] \n\t" + "subu %[q0], %[q0], %[temp2] \n\t" + "subu %[q1], %[q1], %[step1] \n\t" + "lbux %[temp2], %[p0](%[VP8kclip1]) \n\t" + "lbux %[temp3], %[q0](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[q1](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[pTemp0]) \n\t" + "lbux %[temp1], %[p1](%[VP8kclip1]) \n\t" + "subu %[pTemp0], %[pTemp0], %[hstride] \n\t" + "sb %[temp3], 0(%[p]) \n\t" + "sb %[temp4], 0(%[pTemp1]) \n\t" + "j 0f \n\t" + " sb %[temp1], 0(%[pTemp0]) \n\t" + "1: \n\t" + "shll_s.w %[temp3], %[temp3], 24 \n\t" + "sra %[temp3], %[temp3], 24 \n\t" + "addu %[temp1], %[temp1], %[temp3] \n\t" + "shra_r.w %[temp2], %[temp1], 3 \n\t" + "addiu %[temp1], %[temp1], 3 \n\t" + "shll_s.w %[temp2], %[temp2], 27 \n\t" + "sra %[temp1], %[temp1], 3 \n\t" + "shll_s.w %[temp1], %[temp1], 27 \n\t" + "sra %[temp2], %[temp2], 27 \n\t" + "sra %[temp1], %[temp1], 27 \n\t" + "addu %[p0], %[p0], %[temp1] \n\t" + "subu %[q0], %[q0], %[temp2] \n\t" + "lbux %[temp1], %[p0](%[VP8kclip1]) \n\t" + "lbux %[temp2], %[q0](%[VP8kclip1]) \n\t" + "sb %[temp2], 0(%[p]) \n\t" + "sb %[temp1], 0(%[pTemp0]) \n\t" + "0: \n\t" + "subu %[size], %[size], 1 \n\t" + "bgtz %[size], 2b \n\t" + " addu %[p], %[p], %[vstride] \n\t" + "3: \n\t" + ".set pop \n\t" + : [p0]"=&r"(p0), [q0]"=&r"(q0), [p1]"=&r"(p1), [q1]"=&r"(q1), + [p2]"=&r"(p2), [q2]"=&r"(q2), [p3]"=&r"(p3), [q3]"=&r"(q3), + [step2]"=&r"(step2), [step1]"=&r"(step1), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), + [pTemp0]"=&r"(pTemp0), [pTemp1]"=&r"(pTemp1), [p]"+&r"(p), + [size]"+&r"(size) + : [vstride]"r"(vstride), [ithresh]"r"(ithresh), + [hev_thresh]"r"(hev_thresh), [hstride]"r"(hstride), + [VP8kclip1]"r"(VP8kclip1), [thresh2]"r"(thresh2) + : "memory" + ); +} + +// on macroblock edges +static void VFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, stride, 1, 16, thresh, ithresh, hev_thresh); +} + +static void HFilter16(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(p, 1, stride, 16, thresh, ithresh, hev_thresh); +} + +// 8-pixels wide variant, for chroma filtering +static void VFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop26(u, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop26(v, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +// on three inner edges +static void VFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + FilterLoop24(p, stride, 1, 16, thresh, ithresh, hev_thresh); + } +} + +static void HFilter16i(uint8_t* p, int stride, + int thresh, int ithresh, int hev_thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + FilterLoop24(p, 1, stride, 16, thresh, ithresh, hev_thresh); + } +} + +static void VFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4 * stride, stride, 1, 8, thresh, ithresh, hev_thresh); +} + +static void HFilter8i(uint8_t* u, uint8_t* v, int stride, + int thresh, int ithresh, int hev_thresh) { + FilterLoop24(u + 4, 1, stride, 8, thresh, ithresh, hev_thresh); + FilterLoop24(v + 4, 1, stride, 8, thresh, ithresh, hev_thresh); +} + +#undef MUL + +//------------------------------------------------------------------------------ +// Simple In-loop filtering (Paragraph 15.2) + +static void SimpleVFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + uint8_t* p1 = p - stride; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "li %[i], 16 \n\t" + "0: \n\t" + "negu %[temp4], %[stride] \n\t" + "sll %[temp5], %[temp4], 1 \n\t" + "lbu %[temp2], 0(%[p]) \n\t" + "lbux %[temp3], %[stride](%[p]) \n\t" + "lbux %[temp1], %[temp4](%[p]) \n\t" + "lbux %[temp0], %[temp5](%[p]) \n\t" + "subu %[temp7], %[temp1], %[temp2] \n\t" + "subu %[temp6], %[temp0], %[temp3] \n\t" + "absq_s.w %[temp4], %[temp7] \n\t" + "absq_s.w %[temp5], %[temp6] \n\t" + "sll %[temp4], %[temp4], 2 \n\t" + "subu %[temp5], %[temp5], %[thresh2] \n\t" + "addu %[temp5], %[temp4], %[temp5] \n\t" + "negu %[temp8], %[temp7] \n\t" + "bgtz %[temp5], 1f \n\t" + " addiu %[i], %[i], -1 \n\t" + "sll %[temp4], %[temp8], 1 \n\t" + "shll_s.w %[temp5], %[temp6], 24 \n\t" + "addu %[temp3], %[temp4], %[temp8] \n\t" + "sra %[temp5], %[temp5], 24 \n\t" + "addu %[temp3], %[temp3], %[temp5] \n\t" + "addiu %[temp7], %[temp3], 3 \n\t" + "sra %[temp7], %[temp7], 3 \n\t" + "shra_r.w %[temp8], %[temp3], 3 \n\t" + "shll_s.w %[temp0], %[temp7], 27 \n\t" + "shll_s.w %[temp4], %[temp8], 27 \n\t" + "sra %[temp0], %[temp0], 27 \n\t" + "sra %[temp4], %[temp4], 27 \n\t" + "addu %[temp7], %[temp1], %[temp0] \n\t" + "subu %[temp2], %[temp2], %[temp4] \n\t" + "lbux %[temp3], %[temp7](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[temp2](%[VP8kclip1]) \n\t" + "sb %[temp3], 0(%[p1]) \n\t" + "sb %[temp4], 0(%[p]) \n\t" + "1: \n\t" + "addiu %[p1], %[p1], 1 \n\t" + "bgtz %[i], 0b \n\t" + " addiu %[p], %[p], 1 \n\t" + " .set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [p]"+&r"(p), [i]"=&r"(i), [p1]"+&r"(p1) + : [stride]"r"(stride), [VP8kclip1]"r"(VP8kclip1), [thresh2]"r"(thresh2) + : "memory" + ); +} + +// TEMP0 = SRC[A + A1 * BPS] +// TEMP1 = SRC[B + B1 * BPS] +// TEMP2 = SRC[C + C1 * BPS] +// TEMP3 = SRC[D + D1 * BPS] +#define LOAD_4_BYTES(TEMP0, TEMP1, TEMP2, TEMP3, \ + A, A1, B, B1, C, C1, D, D1, SRC) \ + "lbu %[" #TEMP0 "], " #A "+" #A1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "lbu %[" #TEMP1 "], " #B "+" #B1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "lbu %[" #TEMP2 "], " #C "+" #C1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "lbu %[" #TEMP3 "], " #D "+" #D1 "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + +static void SimpleHFilter16(uint8_t* p, int stride, int thresh) { + int i; + const int thresh2 = 2 * thresh + 1; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "li %[i], 16 \n\t" + "0: \n\t" + LOAD_4_BYTES(temp0, temp1, temp2, temp3, -2, 0, -1, 0, 0, 0, 1, 0, p) + "subu %[temp7], %[temp1], %[temp2] \n\t" + "subu %[temp6], %[temp0], %[temp3] \n\t" + "absq_s.w %[temp4], %[temp7] \n\t" + "absq_s.w %[temp5], %[temp6] \n\t" + "sll %[temp4], %[temp4], 2 \n\t" + "addu %[temp5], %[temp4], %[temp5] \n\t" + "subu %[temp5], %[temp5], %[thresh2] \n\t" + "negu %[temp8], %[temp7] \n\t" + "bgtz %[temp5], 1f \n\t" + " addiu %[i], %[i], -1 \n\t" + "sll %[temp4], %[temp8], 1 \n\t" + "shll_s.w %[temp5], %[temp6], 24 \n\t" + "addu %[temp3], %[temp4], %[temp8] \n\t" + "sra %[temp5], %[temp5], 24 \n\t" + "addu %[temp3], %[temp3], %[temp5] \n\t" + "addiu %[temp7], %[temp3], 3 \n\t" + "sra %[temp7], %[temp7], 3 \n\t" + "shra_r.w %[temp8], %[temp3], 3 \n\t" + "shll_s.w %[temp0], %[temp7], 27 \n\t" + "shll_s.w %[temp4], %[temp8], 27 \n\t" + "sra %[temp0], %[temp0], 27 \n\t" + "sra %[temp4], %[temp4], 27 \n\t" + "addu %[temp7], %[temp1], %[temp0] \n\t" + "subu %[temp2], %[temp2], %[temp4] \n\t" + "lbux %[temp3], %[temp7](%[VP8kclip1]) \n\t" + "lbux %[temp4], %[temp2](%[VP8kclip1]) \n\t" + "sb %[temp3], -1(%[p]) \n\t" + "sb %[temp4], 0(%[p]) \n\t" + "1: \n\t" + "bgtz %[i], 0b \n\t" + " addu %[p], %[p], %[stride] \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [p]"+&r"(p), [i]"=&r"(i) + : [stride]"r"(stride), [VP8kclip1]"r"(VP8kclip1), [thresh2]"r"(thresh2) + : "memory" + ); +} + +static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4 * stride; + SimpleVFilter16(p, stride, thresh); + } +} + +static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) { + int k; + for (k = 3; k > 0; --k) { + p += 4; + SimpleHFilter16(p, stride, thresh); + } +} + +// DST[A * BPS] = TEMP0 +// DST[B + C * BPS] = TEMP1 +#define STORE_8_BYTES(TEMP0, TEMP1, A, B, C, DST) \ + "usw %[" #TEMP0 "], " #A "*" XSTR(BPS) "(%[" #DST "]) \n\t" \ + "usw %[" #TEMP1 "], " #B "+" #C "*" XSTR(BPS) "(%[" #DST "]) \n\t" + +static void VE4(uint8_t* dst) { // vertical + const uint8_t* top = dst - BPS; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile ( + "ulw %[temp0], -1(%[top]) \n\t" + "ulh %[temp1], 3(%[top]) \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "packrl.ph %[temp5], %[temp3], %[temp2] \n\t" + "packrl.ph %[temp6], %[temp4], %[temp3] \n\t" + "shll.ph %[temp5], %[temp5], 1 \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp2], %[temp5], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp4] \n\t" + "addq.ph %[temp2], %[temp2], %[temp3] \n\t" + "addq.ph %[temp6], %[temp6], %[temp3] \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precr.qb.ph %[temp4], %[temp6], %[temp2] \n\t" + STORE_8_BYTES(temp4, temp4, 0, 0, 1, dst) + STORE_8_BYTES(temp4, temp4, 2, 0, 3, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void DC4(uint8_t* dst) { // DC + int temp0, temp1, temp2, temp3, temp4; + __asm__ volatile ( + "ulw %[temp0], -1*" XSTR(BPS) "(%[dst]) \n\t" + LOAD_4_BYTES(temp1, temp2, temp3, temp4, -1, 0, -1, 1, -1, 2, -1, 3, dst) + "ins %[temp1], %[temp2], 8, 8 \n\t" + "ins %[temp1], %[temp3], 16, 8 \n\t" + "ins %[temp1], %[temp4], 24, 8 \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "shra_r.w %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 0, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 0, 3, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4) + : [dst]"r"(dst) + : "memory" + ); +} + +static void RD4(uint8_t* dst) { // Down-right + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8; + __asm__ volatile ( + LOAD_4_BYTES(temp0, temp1, temp2, temp3, -1, 0, -1, 1, -1, 2, -1, 3, dst) + "ulw %[temp7], -1-" XSTR(BPS) "(%[dst]) \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "preceu.ph.qbr %[temp5], %[temp7] \n\t" + "ins %[temp2], %[temp1], 16, 16 \n\t" + "preceu.ph.qbl %[temp4], %[temp7] \n\t" + "ins %[temp3], %[temp2], 16, 16 \n\t" + "shll.ph %[temp2], %[temp2], 1 \n\t" + "addq.ph %[temp3], %[temp3], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp5], %[temp1] \n\t" + "addq.ph %[temp3], %[temp3], %[temp2] \n\t" + "addq.ph %[temp1], %[temp1], %[temp5] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp1], %[temp1], %[temp6] \n\t" + "packrl.ph %[temp0], %[temp4], %[temp5] \n\t" + "addq.ph %[temp8], %[temp5], %[temp4] \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shll.ph %[temp0], %[temp0], 1 \n\t" + "shra_r.ph %[temp1], %[temp1], 2 \n\t" + "addq.ph %[temp8], %[temp0], %[temp8] \n\t" + "lbu %[temp5], 3-" XSTR(BPS) "(%[dst]) \n\t" + "precrq.ph.w %[temp7], %[temp7], %[temp7] \n\t" + "shra_r.ph %[temp8], %[temp8], 2 \n\t" + "ins %[temp7], %[temp5], 0, 8 \n\t" + "precr.qb.ph %[temp2], %[temp1], %[temp3] \n\t" + "raddu.w.qb %[temp4], %[temp7] \n\t" + "precr.qb.ph %[temp6], %[temp8], %[temp1] \n\t" + "shra_r.w %[temp4], %[temp4], 2 \n\t" + STORE_8_BYTES(temp2, temp6, 3, 0, 1, dst) + "prepend %[temp2], %[temp8], 8 \n\t" + "prepend %[temp6], %[temp4], 8 \n\t" + STORE_8_BYTES(temp2, temp6, 2, 0, 0, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [dst]"r"(dst) + : "memory" + ); +} + +// TEMP0 = SRC[A * BPS] +// TEMP1 = SRC[B + C * BPS] +#define LOAD_8_BYTES(TEMP0, TEMP1, A, B, C, SRC) \ + "ulw %[" #TEMP0 "], " #A "*" XSTR(BPS) "(%[" #SRC "]) \n\t" \ + "ulw %[" #TEMP1 "], " #B "+" #C "*" XSTR(BPS) "(%[" #SRC "]) \n\t" + +static void LD4(uint8_t* dst) { // Down-Left + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + LOAD_8_BYTES(temp0, temp1, -1, 4, -1, dst) + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "preceu.ph.qbl %[temp5], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp3] \n\t" + "packrl.ph %[temp7], %[temp4], %[temp2] \n\t" + "packrl.ph %[temp8], %[temp5], %[temp4] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp9], %[temp2], %[temp6] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "addq.ph %[temp9], %[temp9], %[temp3] \n\t" + "shll.ph %[temp8], %[temp8], 1 \n\t" + "shra_r.ph %[temp9], %[temp9], 2 \n\t" + "addq.ph %[temp3], %[temp4], %[temp7] \n\t" + "addq.ph %[temp0], %[temp5], %[temp8] \n\t" + "addq.ph %[temp3], %[temp3], %[temp2] \n\t" + "addq.ph %[temp0], %[temp0], %[temp4] \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "srl %[temp1], %[temp1], 24 \n\t" + "sll %[temp1], %[temp1], 1 \n\t" + "raddu.w.qb %[temp5], %[temp5] \n\t" + "precr.qb.ph %[temp9], %[temp3], %[temp9] \n\t" + "precr.qb.ph %[temp3], %[temp0], %[temp3] \n\t" + "addu %[temp1], %[temp1], %[temp5] \n\t" + "shra_r.w %[temp1], %[temp1], 2 \n\t" + STORE_8_BYTES(temp9, temp3, 0, 0, 2, dst) + "prepend %[temp9], %[temp0], 8 \n\t" + "prepend %[temp3], %[temp1], 8 \n\t" + STORE_8_BYTES(temp9, temp3, 1, 0, 3, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [dst]"r"(dst) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Chroma + +static void DC8uv(uint8_t* dst) { // DC + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + LOAD_8_BYTES(temp0, temp1, -1, 4, -1, dst) + LOAD_4_BYTES(temp2, temp3, temp4, temp5, -1, 0, -1, 1, -1, 2, -1, 3, dst) + LOAD_4_BYTES(temp6, temp7, temp8, temp9, -1, 4, -1, 5, -1, 6, -1, 7, dst) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[temp4], %[temp4], %[temp5] \n\t" + "addu %[temp6], %[temp6], %[temp7] \n\t" + "addu %[temp8], %[temp8], %[temp9] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp4] \n\t" + "addu %[temp6], %[temp6], %[temp8] \n\t" + "addu %[temp0], %[temp0], %[temp2] \n\t" + "addu %[temp0], %[temp0], %[temp6] \n\t" + "shra_r.w %[temp0], %[temp0], 4 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 4, 0, dst) + STORE_8_BYTES(temp0, temp0, 1, 4, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 4, 2, dst) + STORE_8_BYTES(temp0, temp0, 3, 4, 3, dst) + STORE_8_BYTES(temp0, temp0, 4, 4, 4, dst) + STORE_8_BYTES(temp0, temp0, 5, 4, 5, dst) + STORE_8_BYTES(temp0, temp0, 6, 4, 6, dst) + STORE_8_BYTES(temp0, temp0, 7, 4, 7, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [dst]"r"(dst) + : "memory" + ); +} + +static void DC8uvNoLeft(uint8_t* dst) { // DC with no left samples + int temp0, temp1; + __asm__ volatile ( + LOAD_8_BYTES(temp0, temp1, -1, 4, -1, dst) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "shra_r.w %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 4, 0, dst) + STORE_8_BYTES(temp0, temp0, 1, 4, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 4, 2, dst) + STORE_8_BYTES(temp0, temp0, 3, 4, 3, dst) + STORE_8_BYTES(temp0, temp0, 4, 4, 4, dst) + STORE_8_BYTES(temp0, temp0, 5, 4, 5, dst) + STORE_8_BYTES(temp0, temp0, 6, 4, 6, dst) + STORE_8_BYTES(temp0, temp0, 7, 4, 7, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1) + : [dst]"r"(dst) + : "memory" + ); +} + +static void DC8uvNoTop(uint8_t* dst) { // DC with no top samples + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8; + __asm__ volatile ( + LOAD_4_BYTES(temp2, temp3, temp4, temp5, -1, 0, -1, 1, -1, 2, -1, 3, dst) + LOAD_4_BYTES(temp6, temp7, temp8, temp1, -1, 4, -1, 5, -1, 6, -1, 7, dst) + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[temp4], %[temp4], %[temp5] \n\t" + "addu %[temp6], %[temp6], %[temp7] \n\t" + "addu %[temp8], %[temp8], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp4] \n\t" + "addu %[temp6], %[temp6], %[temp8] \n\t" + "addu %[temp0], %[temp6], %[temp2] \n\t" + "shra_r.w %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + STORE_8_BYTES(temp0, temp0, 0, 4, 0, dst) + STORE_8_BYTES(temp0, temp0, 1, 4, 1, dst) + STORE_8_BYTES(temp0, temp0, 2, 4, 2, dst) + STORE_8_BYTES(temp0, temp0, 3, 4, 3, dst) + STORE_8_BYTES(temp0, temp0, 4, 4, 4, dst) + STORE_8_BYTES(temp0, temp0, 5, 4, 5, dst) + STORE_8_BYTES(temp0, temp0, 6, 4, 6, dst) + STORE_8_BYTES(temp0, temp0, 7, 4, 7, dst) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [dst]"r"(dst) + : "memory" + ); +} + +#undef LOAD_8_BYTES +#undef STORE_8_BYTES +#undef LOAD_4_BYTES + +#define CLIPPING(SIZE) \ + "preceu.ph.qbl %[temp2], %[temp0] \n\t" \ + "preceu.ph.qbr %[temp0], %[temp0] \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "preceu.ph.qbl %[temp3], %[temp1] \n\t" \ + "preceu.ph.qbr %[temp1], %[temp1] \n\t" \ +".endif \n\t" \ + "addu.ph %[temp2], %[temp2], %[dst_1] \n\t" \ + "addu.ph %[temp0], %[temp0], %[dst_1] \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "addu.ph %[temp3], %[temp3], %[dst_1] \n\t" \ + "addu.ph %[temp1], %[temp1], %[dst_1] \n\t" \ +".endif \n\t" \ + "shll_s.ph %[temp2], %[temp2], 7 \n\t" \ + "shll_s.ph %[temp0], %[temp0], 7 \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "shll_s.ph %[temp3], %[temp3], 7 \n\t" \ + "shll_s.ph %[temp1], %[temp1], 7 \n\t" \ +".endif \n\t" \ + "precrqu_s.qb.ph %[temp0], %[temp2], %[temp0] \n\t" \ +".if " #SIZE " == 8 \n\t" \ + "precrqu_s.qb.ph %[temp1], %[temp3], %[temp1] \n\t" \ +".endif \n\t" + + +#define CLIP_8B_TO_DST(DST, TOP, SIZE) do { \ + int dst_1 = ((int)(DST)[-1] << 16) + (DST)[-1]; \ + int temp0, temp1, temp2, temp3; \ + __asm__ volatile ( \ + ".if " #SIZE " < 8 \n\t" \ + "ulw %[temp0], 0(%[top]) \n\t" \ + "subu.ph %[dst_1], %[dst_1], %[top_1] \n\t" \ + CLIPPING(4) \ + "usw %[temp0], 0(%[dst]) \n\t" \ + ".else \n\t" \ + "ulw %[temp0], 0(%[top]) \n\t" \ + "ulw %[temp1], 4(%[top]) \n\t" \ + "subu.ph %[dst_1], %[dst_1], %[top_1] \n\t" \ + CLIPPING(8) \ + "usw %[temp0], 0(%[dst]) \n\t" \ + "usw %[temp1], 4(%[dst]) \n\t" \ + ".if " #SIZE " == 16 \n\t" \ + "ulw %[temp0], 8(%[top]) \n\t" \ + "ulw %[temp1], 12(%[top]) \n\t" \ + CLIPPING(8) \ + "usw %[temp0], 8(%[dst]) \n\t" \ + "usw %[temp1], 12(%[dst]) \n\t" \ + ".endif \n\t" \ + ".endif \n\t" \ + : [dst_1]"+&r"(dst_1), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), \ + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3) \ + : [top_1]"r"(top_1), [top]"r"((TOP)), [dst]"r"((DST)) \ + : "memory" \ + ); \ +} while (0) + +#define CLIP_TO_DST(DST, SIZE) do { \ + int y; \ + const uint8_t* top = (DST) - BPS; \ + const int top_1 = ((int)top[-1] << 16) + top[-1]; \ + for (y = 0; y < (SIZE); ++y) { \ + CLIP_8B_TO_DST((DST), top, (SIZE)); \ + (DST) += BPS; \ + } \ +} while (0) + +#define TRUE_MOTION(DST, SIZE) \ +static void TrueMotion##SIZE(uint8_t* (DST)) { \ + CLIP_TO_DST((DST), (SIZE)); \ +} + +TRUE_MOTION(dst, 4) +TRUE_MOTION(dst, 8) +TRUE_MOTION(dst, 16) + +#undef TRUE_MOTION +#undef CLIP_TO_DST +#undef CLIP_8B_TO_DST +#undef CLIPPING + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitMIPSdspR2(void) { + VP8TransformDC = TransformDC; + VP8TransformAC3 = TransformAC3; + VP8Transform = TransformTwo; + + VP8VFilter16 = VFilter16; + VP8HFilter16 = HFilter16; + VP8VFilter8 = VFilter8; + VP8HFilter8 = HFilter8; + VP8VFilter16i = VFilter16i; + VP8HFilter16i = HFilter16i; + VP8VFilter8i = VFilter8i; + VP8HFilter8i = HFilter8i; + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; + + VP8PredLuma4[0] = DC4; + VP8PredLuma4[1] = TrueMotion4; + VP8PredLuma4[2] = VE4; + VP8PredLuma4[4] = RD4; + VP8PredLuma4[6] = LD4; + + VP8PredChroma8[0] = DC8uv; + VP8PredChroma8[1] = TrueMotion8; + VP8PredChroma8[4] = DC8uvNoTop; + VP8PredChroma8[5] = DC8uvNoLeft; + + VP8PredLuma16[1] = TrueMotion16; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8DspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/dec_msa.c b/media/libwebp/dsp/dec_msa.c new file mode 100644 index 0000000000..5b0b14cc93 --- /dev/null +++ b/media/libwebp/dsp/dec_msa.c @@ -0,0 +1,1020 @@ +// Copyright 2016 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MSA version of dsp functions +// +// Author(s): Prashant Patil (prashant.patil@imgtec.com) + + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) + +#include "../dsp/msa_macro.h" + +//------------------------------------------------------------------------------ +// Transforms + +#define IDCT_1D_W(in0, in1, in2, in3, out0, out1, out2, out3) { \ + v4i32 a1_m, b1_m, c1_m, d1_m; \ + v4i32 c_tmp1_m, c_tmp2_m, d_tmp1_m, d_tmp2_m; \ + const v4i32 cospi8sqrt2minus1 = __msa_fill_w(20091); \ + const v4i32 sinpi8sqrt2 = __msa_fill_w(35468); \ + \ + a1_m = in0 + in2; \ + b1_m = in0 - in2; \ + c_tmp1_m = (in1 * sinpi8sqrt2) >> 16; \ + c_tmp2_m = in3 + ((in3 * cospi8sqrt2minus1) >> 16); \ + c1_m = c_tmp1_m - c_tmp2_m; \ + d_tmp1_m = in1 + ((in1 * cospi8sqrt2minus1) >> 16); \ + d_tmp2_m = (in3 * sinpi8sqrt2) >> 16; \ + d1_m = d_tmp1_m + d_tmp2_m; \ + BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3); \ +} +#define MULT1(a) ((((a) * 20091) >> 16) + (a)) +#define MULT2(a) (((a) * 35468) >> 16) + +static void TransformOne(const int16_t* in, uint8_t* dst) { + v8i16 input0, input1; + v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3; + v4i32 res0, res1, res2, res3; + const v16i8 zero = { 0 }; + v16i8 dest0, dest1, dest2, dest3; + + LD_SH2(in, 8, input0, input1); + UNPCK_SH_SW(input0, in0, in1); + UNPCK_SH_SW(input1, in2, in3); + IDCT_1D_W(in0, in1, in2, in3, hz0, hz1, hz2, hz3); + TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3); + IDCT_1D_W(hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3); + SRARI_W4_SW(vt0, vt1, vt2, vt3, 3); + TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3); + LD_SB4(dst, BPS, dest0, dest1, dest2, dest3); + ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3, + res0, res1, res2, res3); + ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3, + res0, res1, res2, res3); + ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3); + CLIP_SW4_0_255(res0, res1, res2, res3); + PCKEV_B2_SW(res0, res1, res2, res3, vt0, vt1); + res0 = (v4i32)__msa_pckev_b((v16i8)vt0, (v16i8)vt1); + ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS); +} + +static void TransformTwo(const int16_t* in, uint8_t* dst, int do_two) { + TransformOne(in, dst); + if (do_two) { + TransformOne(in + 16, dst + 4); + } +} + +static void TransformWHT(const int16_t* in, int16_t* out) { + v8i16 input0, input1; + const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 }; + const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 }; + const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 }; + const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 }; + v8i16 tmp0, tmp1, tmp2, tmp3; + v8i16 out0, out1; + + LD_SH2(in, 8, input0, input1); + input1 = SLDI_SH(input1, input1, 8); + tmp0 = input0 + input1; + tmp1 = input0 - input1; + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); + out0 = tmp2 + tmp3; + out1 = tmp2 - tmp3; + VSHF_H2_SH(out0, out1, out0, out1, mask2, mask3, input0, input1); + tmp0 = input0 + input1; + tmp1 = input0 - input1; + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); + tmp0 = tmp2 + tmp3; + tmp1 = tmp2 - tmp3; + ADDVI_H2_SH(tmp0, 3, tmp1, 3, out0, out1); + SRAI_H2_SH(out0, out1, 3); + out[0] = __msa_copy_s_h(out0, 0); + out[16] = __msa_copy_s_h(out0, 4); + out[32] = __msa_copy_s_h(out1, 0); + out[48] = __msa_copy_s_h(out1, 4); + out[64] = __msa_copy_s_h(out0, 1); + out[80] = __msa_copy_s_h(out0, 5); + out[96] = __msa_copy_s_h(out1, 1); + out[112] = __msa_copy_s_h(out1, 5); + out[128] = __msa_copy_s_h(out0, 2); + out[144] = __msa_copy_s_h(out0, 6); + out[160] = __msa_copy_s_h(out1, 2); + out[176] = __msa_copy_s_h(out1, 6); + out[192] = __msa_copy_s_h(out0, 3); + out[208] = __msa_copy_s_h(out0, 7); + out[224] = __msa_copy_s_h(out1, 3); + out[240] = __msa_copy_s_h(out1, 7); +} + +static void TransformDC(const int16_t* in, uint8_t* dst) { + const int DC = (in[0] + 4) >> 3; + const v8i16 tmp0 = __msa_fill_h(DC); + ADDBLK_ST4x4_UB(tmp0, tmp0, tmp0, tmp0, dst, BPS); +} + +static void TransformAC3(const int16_t* in, uint8_t* dst) { + const int a = in[0] + 4; + const int c4 = MULT2(in[4]); + const int d4 = MULT1(in[4]); + const int in2 = MULT2(in[1]); + const int in3 = MULT1(in[1]); + v4i32 tmp0 = { 0 }; + v4i32 out0 = __msa_fill_w(a + d4); + v4i32 out1 = __msa_fill_w(a + c4); + v4i32 out2 = __msa_fill_w(a - c4); + v4i32 out3 = __msa_fill_w(a - d4); + v4i32 res0, res1, res2, res3; + const v4i32 zero = { 0 }; + v16u8 dest0, dest1, dest2, dest3; + + INSERT_W4_SW(in3, in2, -in2, -in3, tmp0); + ADD4(out0, tmp0, out1, tmp0, out2, tmp0, out3, tmp0, + out0, out1, out2, out3); + SRAI_W4_SW(out0, out1, out2, out3, 3); + LD_UB4(dst, BPS, dest0, dest1, dest2, dest3); + ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3, + res0, res1, res2, res3); + ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3, + res0, res1, res2, res3); + ADD4(res0, out0, res1, out1, res2, out2, res3, out3, res0, res1, res2, res3); + CLIP_SW4_0_255(res0, res1, res2, res3); + PCKEV_B2_SW(res0, res1, res2, res3, out0, out1); + res0 = (v4i32)__msa_pckev_b((v16i8)out0, (v16i8)out1); + ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS); +} + +//------------------------------------------------------------------------------ +// Edge filtering functions + +#define FLIP_SIGN2(in0, in1, out0, out1) { \ + out0 = (v16i8)__msa_xori_b(in0, 0x80); \ + out1 = (v16i8)__msa_xori_b(in1, 0x80); \ +} + +#define FLIP_SIGN4(in0, in1, in2, in3, out0, out1, out2, out3) { \ + FLIP_SIGN2(in0, in1, out0, out1); \ + FLIP_SIGN2(in2, in3, out2, out3); \ +} + +#define FILT_VAL(q0_m, p0_m, mask, filt) do { \ + v16i8 q0_sub_p0; \ + q0_sub_p0 = __msa_subs_s_b(q0_m, p0_m); \ + filt = __msa_adds_s_b(filt, q0_sub_p0); \ + filt = __msa_adds_s_b(filt, q0_sub_p0); \ + filt = __msa_adds_s_b(filt, q0_sub_p0); \ + filt = filt & mask; \ +} while (0) + +#define FILT2(q_m, p_m, q, p) do { \ + u_r = SRAI_H(temp1, 7); \ + u_r = __msa_sat_s_h(u_r, 7); \ + u_l = SRAI_H(temp3, 7); \ + u_l = __msa_sat_s_h(u_l, 7); \ + u = __msa_pckev_b((v16i8)u_l, (v16i8)u_r); \ + q_m = __msa_subs_s_b(q_m, u); \ + p_m = __msa_adds_s_b(p_m, u); \ + q = __msa_xori_b((v16u8)q_m, 0x80); \ + p = __msa_xori_b((v16u8)p_m, 0x80); \ +} while (0) + +#define LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev) do { \ + v16i8 p1_m, p0_m, q0_m, q1_m; \ + v16i8 filt, t1, t2; \ + const v16i8 cnst4b = __msa_ldi_b(4); \ + const v16i8 cnst3b = __msa_ldi_b(3); \ + \ + FLIP_SIGN4(p1, p0, q0, q1, p1_m, p0_m, q0_m, q1_m); \ + filt = __msa_subs_s_b(p1_m, q1_m); \ + filt = filt & hev; \ + FILT_VAL(q0_m, p0_m, mask, filt); \ + t1 = __msa_adds_s_b(filt, cnst4b); \ + t1 = SRAI_B(t1, 3); \ + t2 = __msa_adds_s_b(filt, cnst3b); \ + t2 = SRAI_B(t2, 3); \ + q0_m = __msa_subs_s_b(q0_m, t1); \ + q0 = __msa_xori_b((v16u8)q0_m, 0x80); \ + p0_m = __msa_adds_s_b(p0_m, t2); \ + p0 = __msa_xori_b((v16u8)p0_m, 0x80); \ + filt = __msa_srari_b(t1, 1); \ + hev = __msa_xori_b(hev, 0xff); \ + filt = filt & hev; \ + q1_m = __msa_subs_s_b(q1_m, filt); \ + q1 = __msa_xori_b((v16u8)q1_m, 0x80); \ + p1_m = __msa_adds_s_b(p1_m, filt); \ + p1 = __msa_xori_b((v16u8)p1_m, 0x80); \ +} while (0) + +#define LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev) do { \ + v16i8 p2_m, p1_m, p0_m, q2_m, q1_m, q0_m; \ + v16i8 u, filt, t1, t2, filt_sign; \ + v8i16 filt_r, filt_l, u_r, u_l; \ + v8i16 temp0, temp1, temp2, temp3; \ + const v16i8 cnst4b = __msa_ldi_b(4); \ + const v16i8 cnst3b = __msa_ldi_b(3); \ + const v8i16 cnst9h = __msa_ldi_h(9); \ + const v8i16 cnst63h = __msa_ldi_h(63); \ + \ + FLIP_SIGN4(p1, p0, q0, q1, p1_m, p0_m, q0_m, q1_m); \ + filt = __msa_subs_s_b(p1_m, q1_m); \ + FILT_VAL(q0_m, p0_m, mask, filt); \ + FLIP_SIGN2(p2, q2, p2_m, q2_m); \ + t2 = filt & hev; \ + /* filt_val &= ~hev */ \ + hev = __msa_xori_b(hev, 0xff); \ + filt = filt & hev; \ + t1 = __msa_adds_s_b(t2, cnst4b); \ + t1 = SRAI_B(t1, 3); \ + t2 = __msa_adds_s_b(t2, cnst3b); \ + t2 = SRAI_B(t2, 3); \ + q0_m = __msa_subs_s_b(q0_m, t1); \ + p0_m = __msa_adds_s_b(p0_m, t2); \ + filt_sign = __msa_clti_s_b(filt, 0); \ + ILVRL_B2_SH(filt_sign, filt, filt_r, filt_l); \ + /* update q2/p2 */ \ + temp0 = filt_r * cnst9h; \ + temp1 = temp0 + cnst63h; \ + temp2 = filt_l * cnst9h; \ + temp3 = temp2 + cnst63h; \ + FILT2(q2_m, p2_m, q2, p2); \ + /* update q1/p1 */ \ + temp1 = temp1 + temp0; \ + temp3 = temp3 + temp2; \ + FILT2(q1_m, p1_m, q1, p1); \ + /* update q0/p0 */ \ + temp1 = temp1 + temp0; \ + temp3 = temp3 + temp2; \ + FILT2(q0_m, p0_m, q0, p0); \ +} while (0) + +#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in, \ + q0_in, q1_in, q2_in, q3_in, \ + limit_in, b_limit_in, thresh_in, \ + hev_out, mask_out) do { \ + v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m; \ + v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m; \ + v16u8 flat_out; \ + \ + /* absolute subtraction of pixel values */ \ + p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in); \ + p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in); \ + p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in); \ + q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in); \ + q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in); \ + q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in); \ + p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in); \ + p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in); \ + /* calculation of hev */ \ + flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m); \ + hev_out = (thresh_in < flat_out); \ + /* calculation of mask */ \ + p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m); \ + p1_asub_q1_m = SRAI_B(p1_asub_q1_m, 1); \ + p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m); \ + mask_out = (b_limit_in < p0_asub_q0_m); \ + mask_out = __msa_max_u_b(flat_out, mask_out); \ + p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m); \ + mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out); \ + q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m); \ + mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out); \ + mask_out = (limit_in < mask_out); \ + mask_out = __msa_xori_b(mask_out, 0xff); \ +} while (0) + +#define ST6x1_UB(in0, in0_idx, in1, in1_idx, pdst, stride) do { \ + const uint16_t tmp0_h = __msa_copy_s_h((v8i16)in1, in1_idx); \ + const uint32_t tmp0_w = __msa_copy_s_w((v4i32)in0, in0_idx); \ + SW(tmp0_w, pdst); \ + SH(tmp0_h, pdst + stride); \ +} while (0) + +#define ST6x4_UB(in0, start_in0_idx, in1, start_in1_idx, pdst, stride) do { \ + uint8_t* ptmp1 = (uint8_t*)pdst; \ + ST6x1_UB(in0, start_in0_idx, in1, start_in1_idx, ptmp1, 4); \ + ptmp1 += stride; \ + ST6x1_UB(in0, start_in0_idx + 1, in1, start_in1_idx + 1, ptmp1, 4); \ + ptmp1 += stride; \ + ST6x1_UB(in0, start_in0_idx + 2, in1, start_in1_idx + 2, ptmp1, 4); \ + ptmp1 += stride; \ + ST6x1_UB(in0, start_in0_idx + 3, in1, start_in1_idx + 3, ptmp1, 4); \ +} while (0) + +#define LPF_SIMPLE_FILT(p1_in, p0_in, q0_in, q1_in, mask) do { \ + v16i8 p1_m, p0_m, q0_m, q1_m, filt, filt1, filt2; \ + const v16i8 cnst4b = __msa_ldi_b(4); \ + const v16i8 cnst3b = __msa_ldi_b(3); \ + \ + FLIP_SIGN4(p1_in, p0_in, q0_in, q1_in, p1_m, p0_m, q0_m, q1_m); \ + filt = __msa_subs_s_b(p1_m, q1_m); \ + FILT_VAL(q0_m, p0_m, mask, filt); \ + filt1 = __msa_adds_s_b(filt, cnst4b); \ + filt1 = SRAI_B(filt1, 3); \ + filt2 = __msa_adds_s_b(filt, cnst3b); \ + filt2 = SRAI_B(filt2, 3); \ + q0_m = __msa_subs_s_b(q0_m, filt1); \ + p0_m = __msa_adds_s_b(p0_m, filt2); \ + q0_in = __msa_xori_b((v16u8)q0_m, 0x80); \ + p0_in = __msa_xori_b((v16u8)p0_m, 0x80); \ +} while (0) + +#define LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask) do { \ + v16u8 p1_a_sub_q1, p0_a_sub_q0; \ + \ + p0_a_sub_q0 = __msa_asub_u_b(p0, q0); \ + p1_a_sub_q1 = __msa_asub_u_b(p1, q1); \ + p1_a_sub_q1 = (v16u8)__msa_srli_b((v16i8)p1_a_sub_q1, 1); \ + p0_a_sub_q0 = __msa_adds_u_b(p0_a_sub_q0, p0_a_sub_q0); \ + mask = __msa_adds_u_b(p0_a_sub_q0, p1_a_sub_q1); \ + mask = (mask <= b_limit); \ +} while (0) + +static void VFilter16(uint8_t* src, int stride, + int b_limit_in, int limit_in, int thresh_in) { + uint8_t* ptemp = src - 4 * stride; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 mask, hev; + const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in); + const v16u8 limit = (v16u8)__msa_fill_b(limit_in); + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + + LD_UB8(ptemp, stride, p3, p2, p1, p0, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, + hev, mask); + LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev); + ptemp = src - 3 * stride; + ST_UB4(p2, p1, p0, q0, ptemp, stride); + ptemp += (4 * stride); + ST_UB2(q1, q2, ptemp, stride); +} + +static void HFilter16(uint8_t* src, int stride, + int b_limit_in, int limit_in, int thresh_in) { + uint8_t* ptmp = src - 4; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 mask, hev; + v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; + v16u8 row9, row10, row11, row12, row13, row14, row15; + v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + const v16u8 limit = (v16u8)__msa_fill_b(limit_in); + const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in); + + LD_UB8(ptmp, stride, row0, row1, row2, row3, row4, row5, row6, row7); + ptmp += (8 * stride); + LD_UB8(ptmp, stride, row8, row9, row10, row11, row12, row13, row14, row15); + TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, + row8, row9, row10, row11, row12, row13, row14, row15, + p3, p2, p1, p0, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, + hev, mask); + LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev); + ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4); + ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7); + ILVRL_B2_SH(q2, q1, tmp2, tmp5); + ptmp = src - 3; + ST6x1_UB(tmp3, 0, tmp2, 0, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp3, 1, tmp2, 1, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp3, 2, tmp2, 2, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp3, 3, tmp2, 3, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp4, 0, tmp2, 4, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp4, 1, tmp2, 5, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp4, 2, tmp2, 6, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp4, 3, tmp2, 7, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp6, 0, tmp5, 0, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp6, 1, tmp5, 1, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp6, 2, tmp5, 2, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp6, 3, tmp5, 3, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp7, 0, tmp5, 4, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp7, 1, tmp5, 5, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp7, 2, tmp5, 6, ptmp, 4); + ptmp += stride; + ST6x1_UB(tmp7, 3, tmp5, 7, ptmp, 4); +} + +// on three inner edges +static void VFilterHorEdge16i(uint8_t* src, int stride, + int b_limit, int limit, int thresh) { + v16u8 mask, hev; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + const v16u8 thresh0 = (v16u8)__msa_fill_b(thresh); + const v16u8 b_limit0 = (v16u8)__msa_fill_b(b_limit); + const v16u8 limit0 = (v16u8)__msa_fill_b(limit); + + LD_UB8((src - 4 * stride), stride, p3, p2, p1, p0, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0, + hev, mask); + LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev); + ST_UB4(p1, p0, q0, q1, (src - 2 * stride), stride); +} + +static void VFilter16i(uint8_t* src_y, int stride, + int b_limit, int limit, int thresh) { + VFilterHorEdge16i(src_y + 4 * stride, stride, b_limit, limit, thresh); + VFilterHorEdge16i(src_y + 8 * stride, stride, b_limit, limit, thresh); + VFilterHorEdge16i(src_y + 12 * stride, stride, b_limit, limit, thresh); +} + +static void HFilterVertEdge16i(uint8_t* src, int stride, + int b_limit, int limit, int thresh) { + v16u8 mask, hev; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0; + v16u8 row0, row1, row2, row3, row4, row5, row6, row7; + v16u8 row8, row9, row10, row11, row12, row13, row14, row15; + v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; + const v16u8 thresh0 = (v16u8)__msa_fill_b(thresh); + const v16u8 b_limit0 = (v16u8)__msa_fill_b(b_limit); + const v16u8 limit0 = (v16u8)__msa_fill_b(limit); + + LD_UB8(src - 4, stride, row0, row1, row2, row3, row4, row5, row6, row7); + LD_UB8(src - 4 + (8 * stride), stride, + row8, row9, row10, row11, row12, row13, row14, row15); + TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, + row8, row9, row10, row11, row12, row13, row14, row15, + p3, p2, p1, p0, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0, + hev, mask); + LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev); + ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3); + ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5); + src -= 2; + ST4x8_UB(tmp2, tmp3, src, stride); + src += (8 * stride); + ST4x8_UB(tmp4, tmp5, src, stride); +} + +static void HFilter16i(uint8_t* src_y, int stride, + int b_limit, int limit, int thresh) { + HFilterVertEdge16i(src_y + 4, stride, b_limit, limit, thresh); + HFilterVertEdge16i(src_y + 8, stride, b_limit, limit, thresh); + HFilterVertEdge16i(src_y + 12, stride, b_limit, limit, thresh); +} + +// 8-pixels wide variants, for chroma filtering +static void VFilter8(uint8_t* src_u, uint8_t* src_v, int stride, + int b_limit_in, int limit_in, int thresh_in) { + uint8_t* ptmp_src_u = src_u - 4 * stride; + uint8_t* ptmp_src_v = src_v - 4 * stride; + uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev; + v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u; + v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v; + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + const v16u8 limit = (v16u8)__msa_fill_b(limit_in); + const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in); + + LD_UB8(ptmp_src_u, stride, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u); + LD_UB8(ptmp_src_v, stride, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v); + ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0); + ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, + hev, mask); + LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev); + p2_d = __msa_copy_s_d((v2i64)p2, 0); + p1_d = __msa_copy_s_d((v2i64)p1, 0); + p0_d = __msa_copy_s_d((v2i64)p0, 0); + q0_d = __msa_copy_s_d((v2i64)q0, 0); + q1_d = __msa_copy_s_d((v2i64)q1, 0); + q2_d = __msa_copy_s_d((v2i64)q2, 0); + ptmp_src_u += stride; + SD4(p2_d, p1_d, p0_d, q0_d, ptmp_src_u, stride); + ptmp_src_u += (4 * stride); + SD(q1_d, ptmp_src_u); + ptmp_src_u += stride; + SD(q2_d, ptmp_src_u); + p2_d = __msa_copy_s_d((v2i64)p2, 1); + p1_d = __msa_copy_s_d((v2i64)p1, 1); + p0_d = __msa_copy_s_d((v2i64)p0, 1); + q0_d = __msa_copy_s_d((v2i64)q0, 1); + q1_d = __msa_copy_s_d((v2i64)q1, 1); + q2_d = __msa_copy_s_d((v2i64)q2, 1); + ptmp_src_v += stride; + SD4(p2_d, p1_d, p0_d, q0_d, ptmp_src_v, stride); + ptmp_src_v += (4 * stride); + SD(q1_d, ptmp_src_v); + ptmp_src_v += stride; + SD(q2_d, ptmp_src_v); +} + +static void HFilter8(uint8_t* src_u, uint8_t* src_v, int stride, + int b_limit_in, int limit_in, int thresh_in) { + uint8_t* ptmp_src_u = src_u - 4; + uint8_t* ptmp_src_v = src_v - 4; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev; + v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; + v16u8 row9, row10, row11, row12, row13, row14, row15; + v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + const v16u8 limit = (v16u8)__msa_fill_b(limit_in); + const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in); + + LD_UB8(ptmp_src_u, stride, row0, row1, row2, row3, row4, row5, row6, row7); + LD_UB8(ptmp_src_v, stride, + row8, row9, row10, row11, row12, row13, row14, row15); + TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, + row8, row9, row10, row11, row12, row13, row14, row15, + p3, p2, p1, p0, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, + hev, mask); + LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev); + ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4); + ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1); + ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7); + ILVRL_B2_SH(q2, q1, tmp2, tmp5); + ptmp_src_u += 1; + ST6x4_UB(tmp3, 0, tmp2, 0, ptmp_src_u, stride); + ptmp_src_u += 4 * stride; + ST6x4_UB(tmp4, 0, tmp2, 4, ptmp_src_u, stride); + ptmp_src_v += 1; + ST6x4_UB(tmp6, 0, tmp5, 0, ptmp_src_v, stride); + ptmp_src_v += 4 * stride; + ST6x4_UB(tmp7, 0, tmp5, 4, ptmp_src_v, stride); +} + +static void VFilter8i(uint8_t* src_u, uint8_t* src_v, int stride, + int b_limit_in, int limit_in, int thresh_in) { + uint64_t p1_d, p0_d, q0_d, q1_d; + v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev; + v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u; + v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v; + const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in); + const v16u8 limit = (v16u8)__msa_fill_b(limit_in); + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + + LD_UB8(src_u, stride, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u); + src_u += (5 * stride); + LD_UB8(src_v, stride, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v); + src_v += (5 * stride); + ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0); + ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, + hev, mask); + LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev); + p1_d = __msa_copy_s_d((v2i64)p1, 0); + p0_d = __msa_copy_s_d((v2i64)p0, 0); + q0_d = __msa_copy_s_d((v2i64)q0, 0); + q1_d = __msa_copy_s_d((v2i64)q1, 0); + SD4(q1_d, q0_d, p0_d, p1_d, src_u, -stride); + p1_d = __msa_copy_s_d((v2i64)p1, 1); + p0_d = __msa_copy_s_d((v2i64)p0, 1); + q0_d = __msa_copy_s_d((v2i64)q0, 1); + q1_d = __msa_copy_s_d((v2i64)q1, 1); + SD4(q1_d, q0_d, p0_d, p1_d, src_v, -stride); +} + +static void HFilter8i(uint8_t* src_u, uint8_t* src_v, int stride, + int b_limit_in, int limit_in, int thresh_in) { + v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev; + v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; + v16u8 row9, row10, row11, row12, row13, row14, row15; + v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; + const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in); + const v16u8 limit = (v16u8)__msa_fill_b(limit_in); + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + + LD_UB8(src_u, stride, row0, row1, row2, row3, row4, row5, row6, row7); + LD_UB8(src_v, stride, + row8, row9, row10, row11, row12, row13, row14, row15); + TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, + row8, row9, row10, row11, row12, row13, row14, row15, + p3, p2, p1, p0, q0, q1, q2, q3); + LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, + hev, mask); + LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev); + ILVR_B2_SW(p0, p1, q1, q0, tmp0, tmp1); + ILVRL_H2_SW(tmp1, tmp0, tmp2, tmp3); + ILVL_B2_SW(p0, p1, q1, q0, tmp0, tmp1); + ILVRL_H2_SW(tmp1, tmp0, tmp4, tmp5); + src_u += 2; + ST4x4_UB(tmp2, tmp2, 0, 1, 2, 3, src_u, stride); + src_u += 4 * stride; + ST4x4_UB(tmp3, tmp3, 0, 1, 2, 3, src_u, stride); + src_v += 2; + ST4x4_UB(tmp4, tmp4, 0, 1, 2, 3, src_v, stride); + src_v += 4 * stride; + ST4x4_UB(tmp5, tmp5, 0, 1, 2, 3, src_v, stride); +} + +static void SimpleVFilter16(uint8_t* src, int stride, int b_limit_in) { + v16u8 p1, p0, q1, q0, mask; + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + + LD_UB4(src - 2 * stride, stride, p1, p0, q0, q1); + LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask); + LPF_SIMPLE_FILT(p1, p0, q0, q1, mask); + ST_UB2(p0, q0, src - stride, stride); +} + +static void SimpleHFilter16(uint8_t* src, int stride, int b_limit_in) { + v16u8 p1, p0, q1, q0, mask, row0, row1, row2, row3, row4, row5, row6, row7; + v16u8 row8, row9, row10, row11, row12, row13, row14, row15; + v8i16 tmp0, tmp1; + const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in); + uint8_t* ptemp_src = src - 2; + + LD_UB8(ptemp_src, stride, row0, row1, row2, row3, row4, row5, row6, row7); + LD_UB8(ptemp_src + 8 * stride, stride, + row8, row9, row10, row11, row12, row13, row14, row15); + TRANSPOSE16x4_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, + row8, row9, row10, row11, row12, row13, row14, row15, + p1, p0, q0, q1); + LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask); + LPF_SIMPLE_FILT(p1, p0, q0, q1, mask); + ILVRL_B2_SH(q0, p0, tmp1, tmp0); + ptemp_src += 1; + ST2x4_UB(tmp1, 0, ptemp_src, stride); + ptemp_src += 4 * stride; + ST2x4_UB(tmp1, 4, ptemp_src, stride); + ptemp_src += 4 * stride; + ST2x4_UB(tmp0, 0, ptemp_src, stride); + ptemp_src += 4 * stride; + ST2x4_UB(tmp0, 4, ptemp_src, stride); + ptemp_src += 4 * stride; +} + +static void SimpleVFilter16i(uint8_t* src_y, int stride, int b_limit_in) { + SimpleVFilter16(src_y + 4 * stride, stride, b_limit_in); + SimpleVFilter16(src_y + 8 * stride, stride, b_limit_in); + SimpleVFilter16(src_y + 12 * stride, stride, b_limit_in); +} + +static void SimpleHFilter16i(uint8_t* src_y, int stride, int b_limit_in) { + SimpleHFilter16(src_y + 4, stride, b_limit_in); + SimpleHFilter16(src_y + 8, stride, b_limit_in); + SimpleHFilter16(src_y + 12, stride, b_limit_in); +} + +//------------------------------------------------------------------------------ +// Intra predictions +//------------------------------------------------------------------------------ + +// 4x4 + +static void DC4(uint8_t* dst) { // DC + uint32_t dc = 4; + int i; + for (i = 0; i < 4; ++i) dc += dst[i - BPS] + dst[-1 + i * BPS]; + dc >>= 3; + dc = dc | (dc << 8) | (dc << 16) | (dc << 24); + SW4(dc, dc, dc, dc, dst, BPS); +} + +static void TM4(uint8_t* dst) { + const uint8_t* const ptemp = dst - BPS - 1; + v8i16 T, d, r0, r1, r2, r3; + const v16i8 zero = { 0 }; + const v8i16 TL = (v8i16)__msa_fill_h(ptemp[0 * BPS]); + const v8i16 L0 = (v8i16)__msa_fill_h(ptemp[1 * BPS]); + const v8i16 L1 = (v8i16)__msa_fill_h(ptemp[2 * BPS]); + const v8i16 L2 = (v8i16)__msa_fill_h(ptemp[3 * BPS]); + const v8i16 L3 = (v8i16)__msa_fill_h(ptemp[4 * BPS]); + const v16u8 T1 = LD_UB(ptemp + 1); + + T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1); + d = T - TL; + ADD4(d, L0, d, L1, d, L2, d, L3, r0, r1, r2, r3); + CLIP_SH4_0_255(r0, r1, r2, r3); + PCKEV_ST4x4_UB(r0, r1, r2, r3, dst, BPS); +} + +static void VE4(uint8_t* dst) { // vertical + const uint8_t* const ptop = dst - BPS - 1; + const uint32_t val0 = LW(ptop + 0); + const uint32_t val1 = LW(ptop + 4); + uint32_t out; + v16u8 A = { 0 }, B, C, AC, B2, R; + + INSERT_W2_UB(val0, val1, A); + B = SLDI_UB(A, A, 1); + C = SLDI_UB(A, A, 2); + AC = __msa_ave_u_b(A, C); + B2 = __msa_ave_u_b(B, B); + R = __msa_aver_u_b(AC, B2); + out = __msa_copy_s_w((v4i32)R, 0); + SW4(out, out, out, out, dst, BPS); +} + +static void RD4(uint8_t* dst) { // Down-right + const uint8_t* const ptop = dst - 1 - BPS; + uint32_t val0 = LW(ptop + 0); + uint32_t val1 = LW(ptop + 4); + uint32_t val2, val3; + v16u8 A, B, C, AC, B2, R, A1 = { 0 }; + + INSERT_W2_UB(val0, val1, A1); + A = SLDI_UB(A1, A1, 12); + A = (v16u8)__msa_insert_b((v16i8)A, 3, ptop[1 * BPS]); + A = (v16u8)__msa_insert_b((v16i8)A, 2, ptop[2 * BPS]); + A = (v16u8)__msa_insert_b((v16i8)A, 1, ptop[3 * BPS]); + A = (v16u8)__msa_insert_b((v16i8)A, 0, ptop[4 * BPS]); + B = SLDI_UB(A, A, 1); + C = SLDI_UB(A, A, 2); + AC = __msa_ave_u_b(A, C); + B2 = __msa_ave_u_b(B, B); + R = __msa_aver_u_b(AC, B2); + val3 = __msa_copy_s_w((v4i32)R, 0); + R = SLDI_UB(R, R, 1); + val2 = __msa_copy_s_w((v4i32)R, 0); + R = SLDI_UB(R, R, 1); + val1 = __msa_copy_s_w((v4i32)R, 0); + R = SLDI_UB(R, R, 1); + val0 = __msa_copy_s_w((v4i32)R, 0); + SW4(val0, val1, val2, val3, dst, BPS); +} + +static void LD4(uint8_t* dst) { // Down-Left + const uint8_t* const ptop = dst - BPS; + uint32_t val0 = LW(ptop + 0); + uint32_t val1 = LW(ptop + 4); + uint32_t val2, val3; + v16u8 A = { 0 }, B, C, AC, B2, R; + + INSERT_W2_UB(val0, val1, A); + B = SLDI_UB(A, A, 1); + C = SLDI_UB(A, A, 2); + C = (v16u8)__msa_insert_b((v16i8)C, 6, ptop[7]); + AC = __msa_ave_u_b(A, C); + B2 = __msa_ave_u_b(B, B); + R = __msa_aver_u_b(AC, B2); + val0 = __msa_copy_s_w((v4i32)R, 0); + R = SLDI_UB(R, R, 1); + val1 = __msa_copy_s_w((v4i32)R, 0); + R = SLDI_UB(R, R, 1); + val2 = __msa_copy_s_w((v4i32)R, 0); + R = SLDI_UB(R, R, 1); + val3 = __msa_copy_s_w((v4i32)R, 0); + SW4(val0, val1, val2, val3, dst, BPS); +} + +// 16x16 + +static void DC16(uint8_t* dst) { // DC + uint32_t dc = 16; + int i; + const v16u8 rtop = LD_UB(dst - BPS); + const v8u16 dctop = __msa_hadd_u_h(rtop, rtop); + v16u8 out; + + for (i = 0; i < 16; ++i) { + dc += dst[-1 + i * BPS]; + } + dc += HADD_UH_U32(dctop); + out = (v16u8)__msa_fill_b(dc >> 5); + ST_UB8(out, out, out, out, out, out, out, out, dst, BPS); + ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS); +} + +static void TM16(uint8_t* dst) { + int j; + v8i16 d1, d2; + const v16i8 zero = { 0 }; + const v8i16 TL = (v8i16)__msa_fill_h(dst[-1 - BPS]); + const v16i8 T = LD_SB(dst - BPS); + + ILVRL_B2_SH(zero, T, d1, d2); + SUB2(d1, TL, d2, TL, d1, d2); + for (j = 0; j < 16; j += 4) { + v16i8 t0, t1, t2, t3; + v8i16 r0, r1, r2, r3, r4, r5, r6, r7; + const v8i16 L0 = (v8i16)__msa_fill_h(dst[-1 + 0 * BPS]); + const v8i16 L1 = (v8i16)__msa_fill_h(dst[-1 + 1 * BPS]); + const v8i16 L2 = (v8i16)__msa_fill_h(dst[-1 + 2 * BPS]); + const v8i16 L3 = (v8i16)__msa_fill_h(dst[-1 + 3 * BPS]); + ADD4(d1, L0, d1, L1, d1, L2, d1, L3, r0, r1, r2, r3); + ADD4(d2, L0, d2, L1, d2, L2, d2, L3, r4, r5, r6, r7); + CLIP_SH4_0_255(r0, r1, r2, r3); + CLIP_SH4_0_255(r4, r5, r6, r7); + PCKEV_B4_SB(r4, r0, r5, r1, r6, r2, r7, r3, t0, t1, t2, t3); + ST_SB4(t0, t1, t2, t3, dst, BPS); + dst += 4 * BPS; + } +} + +static void VE16(uint8_t* dst) { // vertical + const v16u8 rtop = LD_UB(dst - BPS); + ST_UB8(rtop, rtop, rtop, rtop, rtop, rtop, rtop, rtop, dst, BPS); + ST_UB8(rtop, rtop, rtop, rtop, rtop, rtop, rtop, rtop, dst + 8 * BPS, BPS); +} + +static void HE16(uint8_t* dst) { // horizontal + int j; + for (j = 16; j > 0; j -= 4) { + const v16u8 L0 = (v16u8)__msa_fill_b(dst[-1 + 0 * BPS]); + const v16u8 L1 = (v16u8)__msa_fill_b(dst[-1 + 1 * BPS]); + const v16u8 L2 = (v16u8)__msa_fill_b(dst[-1 + 2 * BPS]); + const v16u8 L3 = (v16u8)__msa_fill_b(dst[-1 + 3 * BPS]); + ST_UB4(L0, L1, L2, L3, dst, BPS); + dst += 4 * BPS; + } +} + +static void DC16NoTop(uint8_t* dst) { // DC with top samples not available + int j; + uint32_t dc = 8; + v16u8 out; + + for (j = 0; j < 16; ++j) { + dc += dst[-1 + j * BPS]; + } + out = (v16u8)__msa_fill_b(dc >> 4); + ST_UB8(out, out, out, out, out, out, out, out, dst, BPS); + ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS); +} + +static void DC16NoLeft(uint8_t* dst) { // DC with left samples not available + uint32_t dc = 8; + const v16u8 rtop = LD_UB(dst - BPS); + const v8u16 dctop = __msa_hadd_u_h(rtop, rtop); + v16u8 out; + + dc += HADD_UH_U32(dctop); + out = (v16u8)__msa_fill_b(dc >> 4); + ST_UB8(out, out, out, out, out, out, out, out, dst, BPS); + ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS); +} + +static void DC16NoTopLeft(uint8_t* dst) { // DC with nothing + const v16u8 out = (v16u8)__msa_fill_b(0x80); + ST_UB8(out, out, out, out, out, out, out, out, dst, BPS); + ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS); +} + +// Chroma + +#define STORE8x8(out, dst) do { \ + SD4(out, out, out, out, dst + 0 * BPS, BPS); \ + SD4(out, out, out, out, dst + 4 * BPS, BPS); \ +} while (0) + +static void DC8uv(uint8_t* dst) { // DC + uint32_t dc = 8; + int i; + uint64_t out; + const v16u8 rtop = LD_UB(dst - BPS); + const v8u16 temp0 = __msa_hadd_u_h(rtop, rtop); + const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0); + const v2u64 temp2 = __msa_hadd_u_d(temp1, temp1); + v16u8 dctemp; + + for (i = 0; i < 8; ++i) { + dc += dst[-1 + i * BPS]; + } + dc += __msa_copy_s_w((v4i32)temp2, 0); + dctemp = (v16u8)__msa_fill_b(dc >> 4); + out = __msa_copy_s_d((v2i64)dctemp, 0); + STORE8x8(out, dst); +} + +static void TM8uv(uint8_t* dst) { + int j; + const v16i8 T1 = LD_SB(dst - BPS); + const v16i8 zero = { 0 }; + const v8i16 T = (v8i16)__msa_ilvr_b(zero, T1); + const v8i16 TL = (v8i16)__msa_fill_h(dst[-1 - BPS]); + const v8i16 d = T - TL; + + for (j = 0; j < 8; j += 4) { + v16i8 t0, t1; + v8i16 r0 = (v8i16)__msa_fill_h(dst[-1 + 0 * BPS]); + v8i16 r1 = (v8i16)__msa_fill_h(dst[-1 + 1 * BPS]); + v8i16 r2 = (v8i16)__msa_fill_h(dst[-1 + 2 * BPS]); + v8i16 r3 = (v8i16)__msa_fill_h(dst[-1 + 3 * BPS]); + ADD4(d, r0, d, r1, d, r2, d, r3, r0, r1, r2, r3); + CLIP_SH4_0_255(r0, r1, r2, r3); + PCKEV_B2_SB(r1, r0, r3, r2, t0, t1); + ST4x4_UB(t0, t1, 0, 2, 0, 2, dst, BPS); + ST4x4_UB(t0, t1, 1, 3, 1, 3, dst + 4, BPS); + dst += 4 * BPS; + } +} + +static void VE8uv(uint8_t* dst) { // vertical + const v16u8 rtop = LD_UB(dst - BPS); + const uint64_t out = __msa_copy_s_d((v2i64)rtop, 0); + STORE8x8(out, dst); +} + +static void HE8uv(uint8_t* dst) { // horizontal + int j; + for (j = 0; j < 8; j += 4) { + const v16u8 L0 = (v16u8)__msa_fill_b(dst[-1 + 0 * BPS]); + const v16u8 L1 = (v16u8)__msa_fill_b(dst[-1 + 1 * BPS]); + const v16u8 L2 = (v16u8)__msa_fill_b(dst[-1 + 2 * BPS]); + const v16u8 L3 = (v16u8)__msa_fill_b(dst[-1 + 3 * BPS]); + const uint64_t out0 = __msa_copy_s_d((v2i64)L0, 0); + const uint64_t out1 = __msa_copy_s_d((v2i64)L1, 0); + const uint64_t out2 = __msa_copy_s_d((v2i64)L2, 0); + const uint64_t out3 = __msa_copy_s_d((v2i64)L3, 0); + SD4(out0, out1, out2, out3, dst, BPS); + dst += 4 * BPS; + } +} + +static void DC8uvNoLeft(uint8_t* dst) { // DC with no left samples + const uint32_t dc = 4; + const v16u8 rtop = LD_UB(dst - BPS); + const v8u16 temp0 = __msa_hadd_u_h(rtop, rtop); + const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0); + const v2u64 temp2 = __msa_hadd_u_d(temp1, temp1); + const uint32_t sum_m = __msa_copy_s_w((v4i32)temp2, 0); + const v16u8 dcval = (v16u8)__msa_fill_b((dc + sum_m) >> 3); + const uint64_t out = __msa_copy_s_d((v2i64)dcval, 0); + STORE8x8(out, dst); +} + +static void DC8uvNoTop(uint8_t* dst) { // DC with no top samples + uint32_t dc = 4; + int i; + uint64_t out; + v16u8 dctemp; + + for (i = 0; i < 8; ++i) { + dc += dst[-1 + i * BPS]; + } + dctemp = (v16u8)__msa_fill_b(dc >> 3); + out = __msa_copy_s_d((v2i64)dctemp, 0); + STORE8x8(out, dst); +} + +static void DC8uvNoTopLeft(uint8_t* dst) { // DC with nothing + const uint64_t out = 0x8080808080808080ULL; + STORE8x8(out, dst); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8DspInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitMSA(void) { + VP8TransformWHT = TransformWHT; + VP8Transform = TransformTwo; + VP8TransformDC = TransformDC; + VP8TransformAC3 = TransformAC3; + + VP8VFilter16 = VFilter16; + VP8HFilter16 = HFilter16; + VP8VFilter16i = VFilter16i; + VP8HFilter16i = HFilter16i; + VP8VFilter8 = VFilter8; + VP8HFilter8 = HFilter8; + VP8VFilter8i = VFilter8i; + VP8HFilter8i = HFilter8i; + VP8SimpleVFilter16 = SimpleVFilter16; + VP8SimpleHFilter16 = SimpleHFilter16; + VP8SimpleVFilter16i = SimpleVFilter16i; + VP8SimpleHFilter16i = SimpleHFilter16i; + + VP8PredLuma4[0] = DC4; + VP8PredLuma4[1] = TM4; + VP8PredLuma4[2] = VE4; + VP8PredLuma4[4] = RD4; + VP8PredLuma4[6] = LD4; + VP8PredLuma16[0] = DC16; + VP8PredLuma16[1] = TM16; + VP8PredLuma16[2] = VE16; + VP8PredLuma16[3] = HE16; + VP8PredLuma16[4] = DC16NoTop; + VP8PredLuma16[5] = DC16NoLeft; + VP8PredLuma16[6] = DC16NoTopLeft; + VP8PredChroma8[0] = DC8uv; + VP8PredChroma8[1] = TM8uv; + VP8PredChroma8[2] = VE8uv; + VP8PredChroma8[3] = HE8uv; + VP8PredChroma8[4] = DC8uvNoTop; + VP8PredChroma8[5] = DC8uvNoLeft; + VP8PredChroma8[6] = DC8uvNoTopLeft; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(VP8DspInitMSA) + +#endif // WEBP_USE_MSA diff --git a/media/libwebp/dsp/dec_neon.c b/media/libwebp/dsp/dec_neon.c index e8341327e4..ea690646d6 100644 --- a/media/libwebp/dsp/dec_neon.c +++ b/media/libwebp/dsp/dec_neon.c @@ -1283,12 +1283,12 @@ static void DC4_NEON(uint8_t* dst) { // DC const uint8x8_t A = vld1_u8(dst - BPS); // top row const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top const uint16x4_t p1 = vpadd_u16(p0, p0); - const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1)); - const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1)); - const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1)); - const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1)); - const uint16x8_t s0 = vaddq_u16(L0, L1); - const uint16x8_t s1 = vaddq_u16(L2, L3); + const uint8x8_t L0 = vld1_u8(dst + 0 * BPS - 1); + const uint8x8_t L1 = vld1_u8(dst + 1 * BPS - 1); + const uint8x8_t L2 = vld1_u8(dst + 2 * BPS - 1); + const uint8x8_t L3 = vld1_u8(dst + 3 * BPS - 1); + const uint16x8_t s0 = vaddl_u8(L0, L1); + const uint16x8_t s1 = vaddl_u8(L2, L3); const uint16x8_t s01 = vaddq_u16(s0, s1); const uint16x8_t sum = vaddq_u16(s01, vcombine_u16(p1, p1)); const uint8x8_t dc0 = vrshrn_n_u16(sum, 3); // (sum + 4) >> 3 @@ -1361,7 +1361,8 @@ static void RD4_NEON(uint8_t* dst) { // Down-right const uint32_t J = dst[-1 + 1 * BPS]; const uint32_t K = dst[-1 + 2 * BPS]; const uint32_t L = dst[-1 + 3 * BPS]; - const uint64x1_t LKJI____ = vcreate_u64(L | (K << 8) | (J << 16) | (I << 24)); + const uint64x1_t LKJI____ = + vcreate_u64((uint64_t)L | (K << 8) | (J << 16) | (I << 24)); const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC); const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8)); const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16)); @@ -1427,25 +1428,30 @@ static WEBP_INLINE void DC8_NEON(uint8_t* dst, int do_top, int do_left) { if (do_top) { const uint8x8_t A = vld1_u8(dst - BPS); // top row +#if defined(__aarch64__) + const uint16_t p2 = vaddlv_u8(A); + sum_top = vdupq_n_u16(p2); +#else const uint16x4_t p0 = vpaddl_u8(A); // cascading summation of the top const uint16x4_t p1 = vpadd_u16(p0, p0); const uint16x4_t p2 = vpadd_u16(p1, p1); sum_top = vcombine_u16(p2, p2); +#endif } if (do_left) { - const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + 0 * BPS - 1)); - const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + 1 * BPS - 1)); - const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + 2 * BPS - 1)); - const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + 3 * BPS - 1)); - const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + 4 * BPS - 1)); - const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + 5 * BPS - 1)); - const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + 6 * BPS - 1)); - const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + 7 * BPS - 1)); - const uint16x8_t s0 = vaddq_u16(L0, L1); - const uint16x8_t s1 = vaddq_u16(L2, L3); - const uint16x8_t s2 = vaddq_u16(L4, L5); - const uint16x8_t s3 = vaddq_u16(L6, L7); + const uint8x8_t L0 = vld1_u8(dst + 0 * BPS - 1); + const uint8x8_t L1 = vld1_u8(dst + 1 * BPS - 1); + const uint8x8_t L2 = vld1_u8(dst + 2 * BPS - 1); + const uint8x8_t L3 = vld1_u8(dst + 3 * BPS - 1); + const uint8x8_t L4 = vld1_u8(dst + 4 * BPS - 1); + const uint8x8_t L5 = vld1_u8(dst + 5 * BPS - 1); + const uint8x8_t L6 = vld1_u8(dst + 6 * BPS - 1); + const uint8x8_t L7 = vld1_u8(dst + 7 * BPS - 1); + const uint16x8_t s0 = vaddl_u8(L0, L1); + const uint16x8_t s1 = vaddl_u8(L2, L3); + const uint16x8_t s2 = vaddl_u8(L4, L5); + const uint16x8_t s3 = vaddl_u8(L6, L7); const uint16x8_t s01 = vaddq_u16(s0, s1); const uint16x8_t s23 = vaddq_u16(s2, s3); sum_left = vaddq_u16(s01, s23); @@ -1505,29 +1511,34 @@ static WEBP_INLINE void DC16_NEON(uint8_t* dst, int do_top, int do_left) { if (do_top) { const uint8x16_t A = vld1q_u8(dst - BPS); // top row +#if defined(__aarch64__) + const uint16_t p3 = vaddlvq_u8(A); + sum_top = vdupq_n_u16(p3); +#else const uint16x8_t p0 = vpaddlq_u8(A); // cascading summation of the top const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0)); const uint16x4_t p2 = vpadd_u16(p1, p1); const uint16x4_t p3 = vpadd_u16(p2, p2); sum_top = vcombine_u16(p3, p3); +#endif } if (do_left) { int i; sum_left = vdupq_n_u16(0); for (i = 0; i < 16; i += 8) { - const uint16x8_t L0 = vmovl_u8(vld1_u8(dst + (i + 0) * BPS - 1)); - const uint16x8_t L1 = vmovl_u8(vld1_u8(dst + (i + 1) * BPS - 1)); - const uint16x8_t L2 = vmovl_u8(vld1_u8(dst + (i + 2) * BPS - 1)); - const uint16x8_t L3 = vmovl_u8(vld1_u8(dst + (i + 3) * BPS - 1)); - const uint16x8_t L4 = vmovl_u8(vld1_u8(dst + (i + 4) * BPS - 1)); - const uint16x8_t L5 = vmovl_u8(vld1_u8(dst + (i + 5) * BPS - 1)); - const uint16x8_t L6 = vmovl_u8(vld1_u8(dst + (i + 6) * BPS - 1)); - const uint16x8_t L7 = vmovl_u8(vld1_u8(dst + (i + 7) * BPS - 1)); - const uint16x8_t s0 = vaddq_u16(L0, L1); - const uint16x8_t s1 = vaddq_u16(L2, L3); - const uint16x8_t s2 = vaddq_u16(L4, L5); - const uint16x8_t s3 = vaddq_u16(L6, L7); + const uint8x8_t L0 = vld1_u8(dst + (i + 0) * BPS - 1); + const uint8x8_t L1 = vld1_u8(dst + (i + 1) * BPS - 1); + const uint8x8_t L2 = vld1_u8(dst + (i + 2) * BPS - 1); + const uint8x8_t L3 = vld1_u8(dst + (i + 3) * BPS - 1); + const uint8x8_t L4 = vld1_u8(dst + (i + 4) * BPS - 1); + const uint8x8_t L5 = vld1_u8(dst + (i + 5) * BPS - 1); + const uint8x8_t L6 = vld1_u8(dst + (i + 6) * BPS - 1); + const uint8x8_t L7 = vld1_u8(dst + (i + 7) * BPS - 1); + const uint16x8_t s0 = vaddl_u8(L0, L1); + const uint16x8_t s1 = vaddl_u8(L2, L3); + const uint16x8_t s2 = vaddl_u8(L4, L5); + const uint16x8_t s3 = vaddl_u8(L6, L7); const uint16x8_t s01 = vaddq_u16(s0, s1); const uint16x8_t s23 = vaddq_u16(s2, s3); const uint16x8_t sum = vaddq_u16(s01, s23); diff --git a/media/libwebp/dsp/dec_sse2.c b/media/libwebp/dsp/dec_sse2.c index f187a5bb48..b90c082793 100644 --- a/media/libwebp/dsp/dec_sse2.c +++ b/media/libwebp/dsp/dec_sse2.c @@ -326,7 +326,7 @@ static WEBP_INLINE void Update2Pixels_SSE2(__m128i* const pi, __m128i* const qi, const __m128i a1_lo = _mm_srai_epi16(*a0_lo, 7); const __m128i a1_hi = _mm_srai_epi16(*a0_hi, 7); const __m128i delta = _mm_packs_epi16(a1_lo, a1_hi); - const __m128i sign_bit = _mm_set1_epi8(0x80); + const __m128i sign_bit = _mm_set1_epi8((char)0x80); *pi = _mm_adds_epi8(*pi, delta); *qi = _mm_subs_epi8(*qi, delta); FLIP_SIGN_BIT2(*pi, *qi); @@ -338,9 +338,9 @@ static WEBP_INLINE void NeedsFilter_SSE2(const __m128i* const p1, const __m128i* const q0, const __m128i* const q1, int thresh, __m128i* const mask) { - const __m128i m_thresh = _mm_set1_epi8(thresh); + const __m128i m_thresh = _mm_set1_epi8((char)thresh); const __m128i t1 = MM_ABS(*p1, *q1); // abs(p1 - q1) - const __m128i kFE = _mm_set1_epi8(0xFE); + const __m128i kFE = _mm_set1_epi8((char)0xFE); const __m128i t2 = _mm_and_si128(t1, kFE); // set lsb of each byte to zero const __m128i t3 = _mm_srli_epi16(t2, 1); // abs(p1 - q1) / 2 @@ -360,7 +360,7 @@ static WEBP_INLINE void DoFilter2_SSE2(__m128i* const p1, __m128i* const p0, __m128i* const q0, __m128i* const q1, int thresh) { __m128i a, mask; - const __m128i sign_bit = _mm_set1_epi8(0x80); + const __m128i sign_bit = _mm_set1_epi8((char)0x80); // convert p1/q1 to int8_t (for GetBaseDelta_SSE2) const __m128i p1s = _mm_xor_si128(*p1, sign_bit); const __m128i q1s = _mm_xor_si128(*q1, sign_bit); @@ -380,7 +380,7 @@ static WEBP_INLINE void DoFilter4_SSE2(__m128i* const p1, __m128i* const p0, const __m128i* const mask, int hev_thresh) { const __m128i zero = _mm_setzero_si128(); - const __m128i sign_bit = _mm_set1_epi8(0x80); + const __m128i sign_bit = _mm_set1_epi8((char)0x80); const __m128i k64 = _mm_set1_epi8(64); const __m128i k3 = _mm_set1_epi8(3); const __m128i k4 = _mm_set1_epi8(4); @@ -427,7 +427,7 @@ static WEBP_INLINE void DoFilter6_SSE2(__m128i* const p2, __m128i* const p1, const __m128i* const mask, int hev_thresh) { const __m128i zero = _mm_setzero_si128(); - const __m128i sign_bit = _mm_set1_epi8(0x80); + const __m128i sign_bit = _mm_set1_epi8((char)0x80); __m128i a, not_hev; // compute hev mask @@ -941,7 +941,7 @@ static void VR4_SSE2(uint8_t* dst) { // Vertical-Right const __m128i ABCD0 = _mm_srli_si128(XABCD, 1); const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0); const __m128i _XABCD = _mm_slli_si128(XABCD, 1); - const __m128i IXABCD = _mm_insert_epi16(_XABCD, I | (X << 8), 0); + const __m128i IXABCD = _mm_insert_epi16(_XABCD, (short)(I | (X << 8)), 0); const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0); const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one); const __m128i avg2 = _mm_subs_epu8(avg1, lsb); diff --git a/media/libwebp/dsp/dsp.h b/media/libwebp/dsp/dsp.h index 4e509bd2ca..ce1679ea53 100644 --- a/media/libwebp/dsp/dsp.h +++ b/media/libwebp/dsp/dsp.h @@ -27,6 +27,23 @@ extern "C" { #define BPS 32 // this is the common stride for enc/dec //------------------------------------------------------------------------------ +// WEBP_RESTRICT + +// Declares a pointer with the restrict type qualifier if available. +// This allows code to hint to the compiler that only this pointer references a +// particular object or memory region within the scope of the block in which it +// is declared. This may allow for improved optimizations due to the lack of +// pointer aliasing. See also: +// https://en.cppreference.com/w/c/language/restrict +#if defined(__GNUC__) +#define WEBP_RESTRICT __restrict__ +#elif defined(_MSC_VER) +#define WEBP_RESTRICT __restrict +#else +#define WEBP_RESTRICT +#endif + +//------------------------------------------------------------------------------ // CPU detection #if defined(__GNUC__) @@ -51,9 +68,7 @@ extern "C" { # define __has_builtin(x) 0 #endif -// for now, none of the optimizations below are available in emscripten -#if !defined(EMSCRIPTEN) - +#if !defined(HAVE_CONFIG_H) #if defined(_MSC_VER) && _MSC_VER > 1310 && \ (defined(_M_X64) || defined(_M_IX86)) #define WEBP_MSC_SSE2 // Visual C++ SSE2 targets @@ -63,23 +78,37 @@ extern "C" { (defined(_M_X64) || defined(_M_IX86)) #define WEBP_MSC_SSE41 // Visual C++ SSE4.1 targets #endif +#endif // WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp // files without intrinsics, allowing the corresponding Init() to be called. // Files containing intrinsics will need to be built targeting the instruction // set so should succeed on one of the earlier tests. -#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) || defined(WEBP_HAVE_SSE2) +#if (defined(__SSE2__) || defined(WEBP_MSC_SSE2)) && \ + (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE2)) #define WEBP_USE_SSE2 #endif -#if defined(__SSE4_1__) || defined(WEBP_MSC_SSE41) || defined(WEBP_HAVE_SSE41) +#if defined(WEBP_USE_SSE2) && !defined(WEBP_HAVE_SSE2) +#define WEBP_HAVE_SSE2 +#endif + +#if (defined(__SSE4_1__) || defined(WEBP_MSC_SSE41)) && \ + (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE41)) #define WEBP_USE_SSE41 #endif +#if defined(WEBP_USE_SSE41) && !defined(WEBP_HAVE_SSE41) +#define WEBP_HAVE_SSE41 +#endif + +#undef WEBP_MSC_SSE41 +#undef WEBP_MSC_SSE2 + // The intrinsics currently cause compiler errors with arm-nacl-gcc and the // inline assembly would need to be modified for use with Native Client. -#if (defined(__ARM_NEON__) || \ - defined(__aarch64__) || defined(WEBP_HAVE_NEON)) && \ +#if ((defined(__ARM_NEON__) || defined(__aarch64__)) && \ + (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_NEON))) && \ !defined(__native_client__) #define WEBP_USE_NEON #endif @@ -90,11 +119,20 @@ extern "C" { #define WEBP_USE_NEON #endif -#if defined(_MSC_VER) && _MSC_VER >= 1700 && defined(_M_ARM) +// Note: ARM64 is supported in Visual Studio 2017, but requires the direct +// inclusion of arm64_neon.h; Visual Studio 2019 includes this file in +// arm_neon.h. +#if defined(_MSC_VER) && \ + ((_MSC_VER >= 1700 && defined(_M_ARM)) || \ + (_MSC_VER >= 1920 && defined(_M_ARM64))) #define WEBP_USE_NEON #define WEBP_USE_INTRINSICS #endif +#if defined(WEBP_USE_NEON) && !defined(WEBP_HAVE_NEON) +#define WEBP_HAVE_NEON +#endif + #if defined(__mips__) && !defined(__mips64) && \ defined(__mips_isa_rev) && (__mips_isa_rev >= 1) && (__mips_isa_rev < 6) #define WEBP_USE_MIPS32 @@ -110,13 +148,11 @@ extern "C" { #define WEBP_USE_MSA #endif -#endif /* EMSCRIPTEN */ - #ifndef WEBP_DSP_OMIT_C_CODE #define WEBP_DSP_OMIT_C_CODE 1 #endif -#if (defined(__aarch64__) || defined(__ARM_NEON__)) && WEBP_DSP_OMIT_C_CODE +#if defined(WEBP_USE_NEON) && WEBP_DSP_OMIT_C_CODE #define WEBP_NEON_OMIT_C_CODE 1 #else #define WEBP_NEON_OMIT_C_CODE 0 @@ -193,6 +229,12 @@ extern "C" { #endif #endif +// If 'ptr' is NULL, returns NULL. Otherwise returns 'ptr + off'. +// Prevents undefined behavior sanitizer nullptr-with-nonzero-offset warning. +#if !defined(WEBP_OFFSET_PTR) +#define WEBP_OFFSET_PTR(ptr, off) (((ptr) == NULL) ? NULL : ((ptr) + (off))) +#endif + // Regularize the definition of WEBP_SWAP_16BIT_CSP (backward compatibility) #if !defined(WEBP_SWAP_16BIT_CSP) #define WEBP_SWAP_16BIT_CSP 0 @@ -246,9 +288,9 @@ extern VP8Fdct VP8FTransform2; // performs two transforms at a time extern VP8WHT VP8FTransformWHT; // Predictions // *dst is the destination block. *top and *left can be NULL. -typedef void (*VP8IntraPreds)(uint8_t *dst, const uint8_t* left, +typedef void (*VP8IntraPreds)(uint8_t* dst, const uint8_t* left, const uint8_t* top); -typedef void (*VP8Intra4Preds)(uint8_t *dst, const uint8_t* top); +typedef void (*VP8Intra4Preds)(uint8_t* dst, const uint8_t* top); extern VP8Intra4Preds VP8EncPredLuma4; extern VP8IntraPreds VP8EncPredLuma16; extern VP8IntraPreds VP8EncPredChroma8; @@ -572,26 +614,29 @@ extern void (*WebPApplyAlphaMultiply4444)( // Dispatch the values from alpha[] plane to the ARGB destination 'dst'. // Returns true if alpha[] plane has non-trivial values different from 0xff. -extern int (*WebPDispatchAlpha)(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint8_t* dst, int dst_stride); +extern int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint8_t* WEBP_RESTRICT dst, int dst_stride); // Transfer packed 8b alpha[] values to green channel in dst[], zero'ing the // A/R/B values. 'dst_stride' is the stride for dst[] in uint32_t units. -extern void (*WebPDispatchAlphaToGreen)(const uint8_t* alpha, int alpha_stride, - int width, int height, - uint32_t* dst, int dst_stride); +extern void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT alpha, + int alpha_stride, int width, int height, + uint32_t* WEBP_RESTRICT dst, + int dst_stride); // Extract the alpha values from 32b values in argb[] and pack them into alpha[] // (this is the opposite of WebPDispatchAlpha). // Returns true if there's only trivial 0xff alpha values. -extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride, - int width, int height, - uint8_t* alpha, int alpha_stride); +extern int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT argb, + int argb_stride, int width, int height, + uint8_t* WEBP_RESTRICT alpha, + int alpha_stride); // Extract the green values from 32b values in argb[] and pack them into alpha[] // (this is the opposite of WebPDispatchAlphaToGreen). -extern void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size); +extern void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb, + uint8_t* WEBP_RESTRICT alpha, int size); // Pre-Multiply operation transforms x into x * A / 255 (where x=Y,R,G or B). // Un-Multiply operation transforms x into x * 255 / A. @@ -604,34 +649,42 @@ void WebPMultARGBRows(uint8_t* ptr, int stride, int width, int num_rows, int inverse); // Same for a row of single values, with side alpha values. -extern void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha, +extern void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse); // Same a WebPMultRow(), but for several 'num_rows' rows. -void WebPMultRows(uint8_t* ptr, int stride, - const uint8_t* alpha, int alpha_stride, +void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride, + const uint8_t* WEBP_RESTRICT alpha, int alpha_stride, int width, int num_rows, int inverse); // Plain-C versions, used as fallback by some implementations. -void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha, +void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr, + const uint8_t* WEBP_RESTRICT const alpha, int width, int inverse); void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse); #ifdef WORDS_BIGENDIAN // ARGB packing function: a/r/g/b input is rgba or bgra order. -extern void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, - const uint8_t* g, const uint8_t* b, int len, - uint32_t* out); +extern void (*WebPPackARGB)(const uint8_t* WEBP_RESTRICT a, + const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, uint32_t* WEBP_RESTRICT out); #endif // RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order. -extern void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b, - int len, int step, uint32_t* out); +extern void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r, + const uint8_t* WEBP_RESTRICT g, + const uint8_t* WEBP_RESTRICT b, + int len, int step, uint32_t* WEBP_RESTRICT out); // This function returns true if src[i] contains a value different from 0xff. extern int (*WebPHasAlpha8b)(const uint8_t* src, int length); // This function returns true if src[4*i] contains a value different from 0xff. extern int (*WebPHasAlpha32b)(const uint8_t* src, int length); +// replaces transparent values in src[] by 'color'. +extern void (*WebPAlphaReplace)(uint32_t* src, int length, uint32_t color); // To be called first before using the above. void WebPInitAlphaProcessing(void); diff --git a/media/libwebp/dsp/enc.c b/media/libwebp/dsp/enc.c new file mode 100644 index 0000000000..69a2f5c577 --- /dev/null +++ b/media/libwebp/dsp/enc.c @@ -0,0 +1,830 @@ +// Copyright 2011 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Speed-critical encoding functions. +// +// Author: Skal (pascal.massimino@gmail.com) + +#include <assert.h> +#include <stdlib.h> // for abs() + +#include "../dsp/dsp.h" +#include "../enc/vp8i_enc.h" + +static WEBP_INLINE uint8_t clip_8b(int v) { + return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255; +} + +#if !WEBP_NEON_OMIT_C_CODE +static WEBP_INLINE int clip_max(int v, int max) { + return (v > max) ? max : v; +} +#endif // !WEBP_NEON_OMIT_C_CODE + +//------------------------------------------------------------------------------ +// Compute susceptibility based on DCT-coeff histograms: +// the higher, the "easier" the macroblock is to compress. + +const int VP8DspScan[16 + 4 + 4] = { + // Luma + 0 + 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS, + 0 + 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS, + 0 + 8 * BPS, 4 + 8 * BPS, 8 + 8 * BPS, 12 + 8 * BPS, + 0 + 12 * BPS, 4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS, + + 0 + 0 * BPS, 4 + 0 * BPS, 0 + 4 * BPS, 4 + 4 * BPS, // U + 8 + 0 * BPS, 12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS // V +}; + +// general-purpose util function +void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1], + VP8Histogram* const histo) { + int max_value = 0, last_non_zero = 1; + int k; + for (k = 0; k <= MAX_COEFF_THRESH; ++k) { + const int value = distribution[k]; + if (value > 0) { + if (value > max_value) max_value = value; + last_non_zero = k; + } + } + histo->max_value = max_value; + histo->last_non_zero = last_non_zero; +} + +#if !WEBP_NEON_OMIT_C_CODE +static void CollectHistogram_C(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int k; + int16_t out[16]; + + VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + + // Convert coefficients to bin. + for (k = 0; k < 16; ++k) { + const int v = abs(out[k]) >> 3; + const int clipped_value = clip_max(v, MAX_COEFF_THRESH); + ++distribution[clipped_value]; + } + } + VP8SetHistogramData(distribution, histo); +} +#endif // !WEBP_NEON_OMIT_C_CODE + +//------------------------------------------------------------------------------ +// run-time tables (~4k) + +static uint8_t clip1[255 + 510 + 1]; // clips [-255,510] to [0,255] + +// We declare this variable 'volatile' to prevent instruction reordering +// and make sure it's set to true _last_ (so as to be thread-safe) +static volatile int tables_ok = 0; + +static WEBP_TSAN_IGNORE_FUNCTION void InitTables(void) { + if (!tables_ok) { + int i; + for (i = -255; i <= 255 + 255; ++i) { + clip1[255 + i] = clip_8b(i); + } + tables_ok = 1; + } +} + + +//------------------------------------------------------------------------------ +// Transforms (Paragraph 14.4) + +#if !WEBP_NEON_OMIT_C_CODE + +#define STORE(x, y, v) \ + dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3)) + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; +#define MUL(a, b) (((a) * (b)) >> 16) + +static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in, + uint8_t* dst) { + int C[4 * 4], *tmp; + int i; + tmp = C; + for (i = 0; i < 4; ++i) { // vertical pass + const int a = in[0] + in[8]; + const int b = in[0] - in[8]; + const int c = MUL(in[4], kC2) - MUL(in[12], kC1); + const int d = MUL(in[4], kC1) + MUL(in[12], kC2); + tmp[0] = a + d; + tmp[1] = b + c; + tmp[2] = b - c; + tmp[3] = a - d; + tmp += 4; + in++; + } + + tmp = C; + for (i = 0; i < 4; ++i) { // horizontal pass + const int dc = tmp[0] + 4; + const int a = dc + tmp[8]; + const int b = dc - tmp[8]; + const int c = MUL(tmp[4], kC2) - MUL(tmp[12], kC1); + const int d = MUL(tmp[4], kC1) + MUL(tmp[12], kC2); + STORE(0, i, a + d); + STORE(1, i, b + c); + STORE(2, i, b - c); + STORE(3, i, a - d); + tmp++; + } +} + +static void ITransform_C(const uint8_t* ref, const int16_t* in, uint8_t* dst, + int do_two) { + ITransformOne(ref, in, dst); + if (do_two) { + ITransformOne(ref + 4, in + 16, dst + 4); + } +} + +static void FTransform_C(const uint8_t* src, const uint8_t* ref, int16_t* out) { + int i; + int tmp[16]; + for (i = 0; i < 4; ++i, src += BPS, ref += BPS) { + const int d0 = src[0] - ref[0]; // 9bit dynamic range ([-255,255]) + const int d1 = src[1] - ref[1]; + const int d2 = src[2] - ref[2]; + const int d3 = src[3] - ref[3]; + const int a0 = (d0 + d3); // 10b [-510,510] + const int a1 = (d1 + d2); + const int a2 = (d1 - d2); + const int a3 = (d0 - d3); + tmp[0 + i * 4] = (a0 + a1) * 8; // 14b [-8160,8160] + tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9; // [-7536,7542] + tmp[2 + i * 4] = (a0 - a1) * 8; + tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 + 937) >> 9; + } + for (i = 0; i < 4; ++i) { + const int a0 = (tmp[0 + i] + tmp[12 + i]); // 15b + const int a1 = (tmp[4 + i] + tmp[ 8 + i]); + const int a2 = (tmp[4 + i] - tmp[ 8 + i]); + const int a3 = (tmp[0 + i] - tmp[12 + i]); + out[0 + i] = (a0 + a1 + 7) >> 4; // 12b + out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0); + out[8 + i] = (a0 - a1 + 7) >> 4; + out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16); + } +} +#endif // !WEBP_NEON_OMIT_C_CODE + +static void FTransform2_C(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + VP8FTransform(src, ref, out); + VP8FTransform(src + 4, ref + 4, out + 16); +} + +#if !WEBP_NEON_OMIT_C_CODE +static void FTransformWHT_C(const int16_t* in, int16_t* out) { + // input is 12b signed + int32_t tmp[16]; + int i; + for (i = 0; i < 4; ++i, in += 64) { + const int a0 = (in[0 * 16] + in[2 * 16]); // 13b + const int a1 = (in[1 * 16] + in[3 * 16]); + const int a2 = (in[1 * 16] - in[3 * 16]); + const int a3 = (in[0 * 16] - in[2 * 16]); + tmp[0 + i * 4] = a0 + a1; // 14b + tmp[1 + i * 4] = a3 + a2; + tmp[2 + i * 4] = a3 - a2; + tmp[3 + i * 4] = a0 - a1; + } + for (i = 0; i < 4; ++i) { + const int a0 = (tmp[0 + i] + tmp[8 + i]); // 15b + const int a1 = (tmp[4 + i] + tmp[12+ i]); + const int a2 = (tmp[4 + i] - tmp[12+ i]); + const int a3 = (tmp[0 + i] - tmp[8 + i]); + const int b0 = a0 + a1; // 16b + const int b1 = a3 + a2; + const int b2 = a3 - a2; + const int b3 = a0 - a1; + out[ 0 + i] = b0 >> 1; // 15b + out[ 4 + i] = b1 >> 1; + out[ 8 + i] = b2 >> 1; + out[12 + i] = b3 >> 1; + } +} +#endif // !WEBP_NEON_OMIT_C_CODE + +#undef MUL +#undef STORE + +//------------------------------------------------------------------------------ +// Intra predictions + +static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) { + int j; + for (j = 0; j < size; ++j) { + memset(dst + j * BPS, value, size); + } +} + +static WEBP_INLINE void VerticalPred(uint8_t* dst, + const uint8_t* top, int size) { + int j; + if (top != NULL) { + for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size); + } else { + Fill(dst, 127, size); + } +} + +static WEBP_INLINE void HorizontalPred(uint8_t* dst, + const uint8_t* left, int size) { + if (left != NULL) { + int j; + for (j = 0; j < size; ++j) { + memset(dst + j * BPS, left[j], size); + } + } else { + Fill(dst, 129, size); + } +} + +static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left, + const uint8_t* top, int size) { + int y; + if (left != NULL) { + if (top != NULL) { + const uint8_t* const clip = clip1 + 255 - left[-1]; + for (y = 0; y < size; ++y) { + const uint8_t* const clip_table = clip + left[y]; + int x; + for (x = 0; x < size; ++x) { + dst[x] = clip_table[top[x]]; + } + dst += BPS; + } + } else { + HorizontalPred(dst, left, size); + } + } else { + // true motion without left samples (hence: with default 129 value) + // is equivalent to VE prediction where you just copy the top samples. + // Note that if top samples are not available, the default value is + // then 129, and not 127 as in the VerticalPred case. + if (top != NULL) { + VerticalPred(dst, top, size); + } else { + Fill(dst, 129, size); + } + } +} + +static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left, + const uint8_t* top, + int size, int round, int shift) { + int DC = 0; + int j; + if (top != NULL) { + for (j = 0; j < size; ++j) DC += top[j]; + if (left != NULL) { // top and left present + for (j = 0; j < size; ++j) DC += left[j]; + } else { // top, but no left + DC += DC; + } + DC = (DC + round) >> shift; + } else if (left != NULL) { // left but no top + for (j = 0; j < size; ++j) DC += left[j]; + DC += DC; + DC = (DC + round) >> shift; + } else { // no top, no left, nothing. + DC = 0x80; + } + Fill(dst, DC, size); +} + +//------------------------------------------------------------------------------ +// Chroma 8x8 prediction (paragraph 12.2) + +static void IntraChromaPreds_C(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + // U block + DCMode(C8DC8 + dst, left, top, 8, 8, 4); + VerticalPred(C8VE8 + dst, top, 8); + HorizontalPred(C8HE8 + dst, left, 8); + TrueMotion(C8TM8 + dst, left, top, 8); + // V block + dst += 8; + if (top != NULL) top += 8; + if (left != NULL) left += 16; + DCMode(C8DC8 + dst, left, top, 8, 8, 4); + VerticalPred(C8VE8 + dst, top, 8); + HorizontalPred(C8HE8 + dst, left, 8); + TrueMotion(C8TM8 + dst, left, top, 8); +} + +//------------------------------------------------------------------------------ +// luma 16x16 prediction (paragraph 12.3) + +static void Intra16Preds_C(uint8_t* dst, + const uint8_t* left, const uint8_t* top) { + DCMode(I16DC16 + dst, left, top, 16, 16, 5); + VerticalPred(I16VE16 + dst, top, 16); + HorizontalPred(I16HE16 + dst, left, 16); + TrueMotion(I16TM16 + dst, left, top, 16); +} + +//------------------------------------------------------------------------------ +// luma 4x4 prediction + +#define DST(x, y) dst[(x) + (y) * BPS] +#define AVG3(a, b, c) ((uint8_t)(((a) + 2 * (b) + (c) + 2) >> 2)) +#define AVG2(a, b) (((a) + (b) + 1) >> 1) + +static void VE4(uint8_t* dst, const uint8_t* top) { // vertical + const uint8_t vals[4] = { + AVG3(top[-1], top[0], top[1]), + AVG3(top[ 0], top[1], top[2]), + AVG3(top[ 1], top[2], top[3]), + AVG3(top[ 2], top[3], top[4]) + }; + int i; + for (i = 0; i < 4; ++i) { + memcpy(dst + i * BPS, vals, 4); + } +} + +static void HE4(uint8_t* dst, const uint8_t* top) { // horizontal + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J)); + WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K)); + WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L)); + WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L)); +} + +static void DC4(uint8_t* dst, const uint8_t* top) { + uint32_t dc = 4; + int i; + for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i]; + Fill(dst, dc >> 3, 4); +} + +static void RD4(uint8_t* dst, const uint8_t* top) { + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + const int D = top[3]; + DST(0, 3) = AVG3(J, K, L); + DST(0, 2) = DST(1, 3) = AVG3(I, J, K); + DST(0, 1) = DST(1, 2) = DST(2, 3) = AVG3(X, I, J); + DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I); + DST(1, 0) = DST(2, 1) = DST(3, 2) = AVG3(B, A, X); + DST(2, 0) = DST(3, 1) = AVG3(C, B, A); + DST(3, 0) = AVG3(D, C, B); +} + +static void LD4(uint8_t* dst, const uint8_t* top) { + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + const int D = top[3]; + const int E = top[4]; + const int F = top[5]; + const int G = top[6]; + const int H = top[7]; + DST(0, 0) = AVG3(A, B, C); + DST(1, 0) = DST(0, 1) = AVG3(B, C, D); + DST(2, 0) = DST(1, 1) = DST(0, 2) = AVG3(C, D, E); + DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F); + DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G); + DST(3, 2) = DST(2, 3) = AVG3(F, G, H); + DST(3, 3) = AVG3(G, H, H); +} + +static void VR4(uint8_t* dst, const uint8_t* top) { + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + const int D = top[3]; + DST(0, 0) = DST(1, 2) = AVG2(X, A); + DST(1, 0) = DST(2, 2) = AVG2(A, B); + DST(2, 0) = DST(3, 2) = AVG2(B, C); + DST(3, 0) = AVG2(C, D); + + DST(0, 3) = AVG3(K, J, I); + DST(0, 2) = AVG3(J, I, X); + DST(0, 1) = DST(1, 3) = AVG3(I, X, A); + DST(1, 1) = DST(2, 3) = AVG3(X, A, B); + DST(2, 1) = DST(3, 3) = AVG3(A, B, C); + DST(3, 1) = AVG3(B, C, D); +} + +static void VL4(uint8_t* dst, const uint8_t* top) { + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + const int D = top[3]; + const int E = top[4]; + const int F = top[5]; + const int G = top[6]; + const int H = top[7]; + DST(0, 0) = AVG2(A, B); + DST(1, 0) = DST(0, 2) = AVG2(B, C); + DST(2, 0) = DST(1, 2) = AVG2(C, D); + DST(3, 0) = DST(2, 2) = AVG2(D, E); + + DST(0, 1) = AVG3(A, B, C); + DST(1, 1) = DST(0, 3) = AVG3(B, C, D); + DST(2, 1) = DST(1, 3) = AVG3(C, D, E); + DST(3, 1) = DST(2, 3) = AVG3(D, E, F); + DST(3, 2) = AVG3(E, F, G); + DST(3, 3) = AVG3(F, G, H); +} + +static void HU4(uint8_t* dst, const uint8_t* top) { + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + DST(0, 0) = AVG2(I, J); + DST(2, 0) = DST(0, 1) = AVG2(J, K); + DST(2, 1) = DST(0, 2) = AVG2(K, L); + DST(1, 0) = AVG3(I, J, K); + DST(3, 0) = DST(1, 1) = AVG3(J, K, L); + DST(3, 1) = DST(1, 2) = AVG3(K, L, L); + DST(3, 2) = DST(2, 2) = + DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L; +} + +static void HD4(uint8_t* dst, const uint8_t* top) { + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + + DST(0, 0) = DST(2, 1) = AVG2(I, X); + DST(0, 1) = DST(2, 2) = AVG2(J, I); + DST(0, 2) = DST(2, 3) = AVG2(K, J); + DST(0, 3) = AVG2(L, K); + + DST(3, 0) = AVG3(A, B, C); + DST(2, 0) = AVG3(X, A, B); + DST(1, 0) = DST(3, 1) = AVG3(I, X, A); + DST(1, 1) = DST(3, 2) = AVG3(J, I, X); + DST(1, 2) = DST(3, 3) = AVG3(K, J, I); + DST(1, 3) = AVG3(L, K, J); +} + +static void TM4(uint8_t* dst, const uint8_t* top) { + int x, y; + const uint8_t* const clip = clip1 + 255 - top[-1]; + for (y = 0; y < 4; ++y) { + const uint8_t* const clip_table = clip + top[-2 - y]; + for (x = 0; x < 4; ++x) { + dst[x] = clip_table[top[x]]; + } + dst += BPS; + } +} + +#undef DST +#undef AVG3 +#undef AVG2 + +// Left samples are top[-5 .. -2], top_left is top[-1], top are +// located at top[0..3], and top right is top[4..7] +static void Intra4Preds_C(uint8_t* dst, const uint8_t* top) { + DC4(I4DC4 + dst, top); + TM4(I4TM4 + dst, top); + VE4(I4VE4 + dst, top); + HE4(I4HE4 + dst, top); + RD4(I4RD4 + dst, top); + VR4(I4VR4 + dst, top); + LD4(I4LD4 + dst, top); + VL4(I4VL4 + dst, top); + HD4(I4HD4 + dst, top); + HU4(I4HU4 + dst, top); +} + +//------------------------------------------------------------------------------ +// Metric + +#if !WEBP_NEON_OMIT_C_CODE +static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b, + int w, int h) { + int count = 0; + int y, x; + for (y = 0; y < h; ++y) { + for (x = 0; x < w; ++x) { + const int diff = (int)a[x] - b[x]; + count += diff * diff; + } + a += BPS; + b += BPS; + } + return count; +} + +static int SSE16x16_C(const uint8_t* a, const uint8_t* b) { + return GetSSE(a, b, 16, 16); +} +static int SSE16x8_C(const uint8_t* a, const uint8_t* b) { + return GetSSE(a, b, 16, 8); +} +static int SSE8x8_C(const uint8_t* a, const uint8_t* b) { + return GetSSE(a, b, 8, 8); +} +static int SSE4x4_C(const uint8_t* a, const uint8_t* b) { + return GetSSE(a, b, 4, 4); +} +#endif // !WEBP_NEON_OMIT_C_CODE + +static void Mean16x4_C(const uint8_t* ref, uint32_t dc[4]) { + int k, x, y; + for (k = 0; k < 4; ++k) { + uint32_t avg = 0; + for (y = 0; y < 4; ++y) { + for (x = 0; x < 4; ++x) { + avg += ref[x + y * BPS]; + } + } + dc[k] = avg; + ref += 4; // go to next 4x4 block. + } +} + +//------------------------------------------------------------------------------ +// Texture distortion +// +// We try to match the spectral content (weighted) between source and +// reconstructed samples. + +#if !WEBP_NEON_OMIT_C_CODE +// Hadamard transform +// Returns the weighted sum of the absolute value of transformed coefficients. +// w[] contains a row-major 4 by 4 symmetric matrix. +static int TTransform(const uint8_t* in, const uint16_t* w) { + int sum = 0; + int tmp[16]; + int i; + // horizontal pass + for (i = 0; i < 4; ++i, in += BPS) { + const int a0 = in[0] + in[2]; + const int a1 = in[1] + in[3]; + const int a2 = in[1] - in[3]; + const int a3 = in[0] - in[2]; + tmp[0 + i * 4] = a0 + a1; + tmp[1 + i * 4] = a3 + a2; + tmp[2 + i * 4] = a3 - a2; + tmp[3 + i * 4] = a0 - a1; + } + // vertical pass + for (i = 0; i < 4; ++i, ++w) { + const int a0 = tmp[0 + i] + tmp[8 + i]; + const int a1 = tmp[4 + i] + tmp[12+ i]; + const int a2 = tmp[4 + i] - tmp[12+ i]; + const int a3 = tmp[0 + i] - tmp[8 + i]; + const int b0 = a0 + a1; + const int b1 = a3 + a2; + const int b2 = a3 - a2; + const int b3 = a0 - a1; + + sum += w[ 0] * abs(b0); + sum += w[ 4] * abs(b1); + sum += w[ 8] * abs(b2); + sum += w[12] * abs(b3); + } + return sum; +} + +static int Disto4x4_C(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + const int sum1 = TTransform(a, w); + const int sum2 = TTransform(b, w); + return abs(sum2 - sum1) >> 5; +} + +static int Disto16x16_C(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4_C(a + x + y, b + x + y, w); + } + } + return D; +} +#endif // !WEBP_NEON_OMIT_C_CODE + +//------------------------------------------------------------------------------ +// Quantization +// + +static const uint8_t kZigzag[16] = { + 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15 +}; + +// Simple quantization +static int QuantizeBlock_C(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + int last = -1; + int n; + for (n = 0; n < 16; ++n) { + const int j = kZigzag[n]; + const int sign = (in[j] < 0); + const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j]; + if (coeff > mtx->zthresh_[j]) { + const uint32_t Q = mtx->q_[j]; + const uint32_t iQ = mtx->iq_[j]; + const uint32_t B = mtx->bias_[j]; + int level = QUANTDIV(coeff, iQ, B); + if (level > MAX_LEVEL) level = MAX_LEVEL; + if (sign) level = -level; + in[j] = level * (int)Q; + out[n] = level; + if (level) last = n; + } else { + out[n] = 0; + in[j] = 0; + } + } + return (last >= 0); +} + +#if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC +static int Quantize2Blocks_C(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} +#endif // !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC + +//------------------------------------------------------------------------------ +// Block copy + +static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int w, int h) { + int y; + for (y = 0; y < h; ++y) { + memcpy(dst, src, w); + src += BPS; + dst += BPS; + } +} + +static void Copy4x4_C(const uint8_t* src, uint8_t* dst) { + Copy(src, dst, 4, 4); +} + +static void Copy16x8_C(const uint8_t* src, uint8_t* dst) { + Copy(src, dst, 16, 8); +} + +//------------------------------------------------------------------------------ +// Initialization + +// Speed-critical function pointers. We have to initialize them to the default +// implementations within VP8EncDspInit(). +VP8CHisto VP8CollectHistogram; +VP8Idct VP8ITransform; +VP8Fdct VP8FTransform; +VP8Fdct VP8FTransform2; +VP8WHT VP8FTransformWHT; +VP8Intra4Preds VP8EncPredLuma4; +VP8IntraPreds VP8EncPredLuma16; +VP8IntraPreds VP8EncPredChroma8; +VP8Metric VP8SSE16x16; +VP8Metric VP8SSE8x8; +VP8Metric VP8SSE16x8; +VP8Metric VP8SSE4x4; +VP8WMetric VP8TDisto4x4; +VP8WMetric VP8TDisto16x16; +VP8MeanMetric VP8Mean16x4; +VP8QuantizeBlock VP8EncQuantizeBlock; +VP8Quantize2Blocks VP8EncQuantize2Blocks; +VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT; +VP8BlockCopy VP8Copy4x4; +VP8BlockCopy VP8Copy16x8; + +extern void VP8EncDspInitSSE2(void); +extern void VP8EncDspInitSSE41(void); +extern void VP8EncDspInitNEON(void); +extern void VP8EncDspInitMIPS32(void); +extern void VP8EncDspInitMIPSdspR2(void); +extern void VP8EncDspInitMSA(void); + +WEBP_DSP_INIT_FUNC(VP8EncDspInit) { + VP8DspInit(); // common inverse transforms + InitTables(); + + // default C implementations +#if !WEBP_NEON_OMIT_C_CODE + VP8ITransform = ITransform_C; + VP8FTransform = FTransform_C; + VP8FTransformWHT = FTransformWHT_C; + VP8TDisto4x4 = Disto4x4_C; + VP8TDisto16x16 = Disto16x16_C; + VP8CollectHistogram = CollectHistogram_C; + VP8SSE16x16 = SSE16x16_C; + VP8SSE16x8 = SSE16x8_C; + VP8SSE8x8 = SSE8x8_C; + VP8SSE4x4 = SSE4x4_C; +#endif + +#if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC + VP8EncQuantizeBlock = QuantizeBlock_C; + VP8EncQuantize2Blocks = Quantize2Blocks_C; +#endif + + VP8FTransform2 = FTransform2_C; + VP8EncPredLuma4 = Intra4Preds_C; + VP8EncPredLuma16 = Intra16Preds_C; + VP8EncPredChroma8 = IntraChromaPreds_C; + VP8Mean16x4 = Mean16x4_C; + VP8EncQuantizeBlockWHT = QuantizeBlock_C; + VP8Copy4x4 = Copy4x4_C; + VP8Copy16x8 = Copy16x8_C; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_HAVE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8EncDspInitSSE2(); +#if defined(WEBP_HAVE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + VP8EncDspInitSSE41(); + } +#endif + } +#endif +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + VP8EncDspInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8EncDspInitMIPSdspR2(); + } +#endif +#if defined(WEBP_USE_MSA) + if (VP8GetCPUInfo(kMSA)) { + VP8EncDspInitMSA(); + } +#endif + } + +#if defined(WEBP_HAVE_NEON) + if (WEBP_NEON_OMIT_C_CODE || + (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { + VP8EncDspInitNEON(); + } +#endif + + assert(VP8ITransform != NULL); + assert(VP8FTransform != NULL); + assert(VP8FTransformWHT != NULL); + assert(VP8TDisto4x4 != NULL); + assert(VP8TDisto16x16 != NULL); + assert(VP8CollectHistogram != NULL); + assert(VP8SSE16x16 != NULL); + assert(VP8SSE16x8 != NULL); + assert(VP8SSE8x8 != NULL); + assert(VP8SSE4x4 != NULL); + assert(VP8EncQuantizeBlock != NULL); + assert(VP8EncQuantize2Blocks != NULL); + assert(VP8FTransform2 != NULL); + assert(VP8EncPredLuma4 != NULL); + assert(VP8EncPredLuma16 != NULL); + assert(VP8EncPredChroma8 != NULL); + assert(VP8Mean16x4 != NULL); + assert(VP8EncQuantizeBlockWHT != NULL); + assert(VP8Copy4x4 != NULL); + assert(VP8Copy16x8 != NULL); +} diff --git a/media/libwebp/dsp/enc_mips32.c b/media/libwebp/dsp/enc_mips32.c new file mode 100644 index 0000000000..ee26dfb493 --- /dev/null +++ b/media/libwebp/dsp/enc_mips32.c @@ -0,0 +1,677 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of speed-critical encoding functions. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) +// Slobodan Prijic (slobodan.prijic@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "../dsp/mips_macro.h" +#include "../enc/vp8i_enc.h" +#include "../enc/cost_enc.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +// macro for one vertical pass in ITransformOne +// MUL macro inlined +// temp0..temp15 holds tmp[0]..tmp[15] +// A..D - offsets in bytes to load from in buffer +// TEMP0..TEMP3 - registers for corresponding tmp elements +// TEMP4..TEMP5 - temporary registers +#define VERTICAL_PASS(A, B, C, D, TEMP4, TEMP0, TEMP1, TEMP2, TEMP3) \ + "lh %[temp16], " #A "(%[temp20]) \n\t" \ + "lh %[temp18], " #B "(%[temp20]) \n\t" \ + "lh %[temp17], " #C "(%[temp20]) \n\t" \ + "lh %[temp19], " #D "(%[temp20]) \n\t" \ + "addu %[" #TEMP4 "], %[temp16], %[temp18] \n\t" \ + "subu %[temp16], %[temp16], %[temp18] \n\t" \ + "mul %[" #TEMP0 "], %[temp17], %[kC2] \n\t" \ + "mul %[temp18], %[temp19], %[kC1] \n\t" \ + "mul %[temp17], %[temp17], %[kC1] \n\t" \ + "mul %[temp19], %[temp19], %[kC2] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 16 \n\n" \ + "sra %[temp18], %[temp18], 16 \n\n" \ + "sra %[temp17], %[temp17], 16 \n\n" \ + "sra %[temp19], %[temp19], 16 \n\n" \ + "subu %[" #TEMP2 "], %[" #TEMP0 "], %[temp18] \n\t" \ + "addu %[" #TEMP3 "], %[temp17], %[temp19] \n\t" \ + "addu %[" #TEMP0 "], %[" #TEMP4 "], %[" #TEMP3 "] \n\t" \ + "addu %[" #TEMP1 "], %[temp16], %[" #TEMP2 "] \n\t" \ + "subu %[" #TEMP2 "], %[temp16], %[" #TEMP2 "] \n\t" \ + "subu %[" #TEMP3 "], %[" #TEMP4 "], %[" #TEMP3 "] \n\t" + +// macro for one horizontal pass in ITransformOne +// MUL and STORE macros inlined +// a = clip_8b(a) is replaced with: a = max(a, 0); a = min(a, 255) +// temp0..temp15 holds tmp[0]..tmp[15] +// A - offset in bytes to load from ref and store to dst buffer +// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements +#define HORIZONTAL_PASS(A, TEMP0, TEMP4, TEMP8, TEMP12) \ + "addiu %[" #TEMP0 "], %[" #TEMP0 "], 4 \n\t" \ + "addu %[temp16], %[" #TEMP0 "], %[" #TEMP8 "] \n\t" \ + "subu %[temp17], %[" #TEMP0 "], %[" #TEMP8 "] \n\t" \ + "mul %[" #TEMP0 "], %[" #TEMP4 "], %[kC2] \n\t" \ + "mul %[" #TEMP8 "], %[" #TEMP12 "], %[kC1] \n\t" \ + "mul %[" #TEMP4 "], %[" #TEMP4 "], %[kC1] \n\t" \ + "mul %[" #TEMP12 "], %[" #TEMP12 "], %[kC2] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 16 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 16 \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 16 \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 16 \n\t" \ + "subu %[temp18], %[" #TEMP0 "], %[" #TEMP8 "] \n\t" \ + "addu %[temp19], %[" #TEMP4 "], %[" #TEMP12 "] \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[temp19] \n\t" \ + "addu %[" #TEMP4 "], %[temp17], %[temp18] \n\t" \ + "subu %[" #TEMP8 "], %[temp17], %[temp18] \n\t" \ + "subu %[" #TEMP12 "], %[temp16], %[temp19] \n\t" \ + "lw %[temp20], 0(%[args]) \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 3 \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 3 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 3 \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 3 \n\t" \ + "lbu %[temp16], 0+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "lbu %[temp17], 1+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "lbu %[temp18], 2+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "lbu %[temp19], 3+" XSTR(BPS) "*" #A "(%[temp20]) \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[" #TEMP0 "] \n\t" \ + "addu %[" #TEMP4 "], %[temp17], %[" #TEMP4 "] \n\t" \ + "addu %[" #TEMP8 "], %[temp18], %[" #TEMP8 "] \n\t" \ + "addu %[" #TEMP12 "], %[temp19], %[" #TEMP12 "] \n\t" \ + "slt %[temp16], %[" #TEMP0 "], $zero \n\t" \ + "slt %[temp17], %[" #TEMP4 "], $zero \n\t" \ + "slt %[temp18], %[" #TEMP8 "], $zero \n\t" \ + "slt %[temp19], %[" #TEMP12 "], $zero \n\t" \ + "movn %[" #TEMP0 "], $zero, %[temp16] \n\t" \ + "movn %[" #TEMP4 "], $zero, %[temp17] \n\t" \ + "movn %[" #TEMP8 "], $zero, %[temp18] \n\t" \ + "movn %[" #TEMP12 "], $zero, %[temp19] \n\t" \ + "addiu %[temp20], $zero, 255 \n\t" \ + "slt %[temp16], %[" #TEMP0 "], %[temp20] \n\t" \ + "slt %[temp17], %[" #TEMP4 "], %[temp20] \n\t" \ + "slt %[temp18], %[" #TEMP8 "], %[temp20] \n\t" \ + "slt %[temp19], %[" #TEMP12 "], %[temp20] \n\t" \ + "movz %[" #TEMP0 "], %[temp20], %[temp16] \n\t" \ + "movz %[" #TEMP4 "], %[temp20], %[temp17] \n\t" \ + "lw %[temp16], 8(%[args]) \n\t" \ + "movz %[" #TEMP8 "], %[temp20], %[temp18] \n\t" \ + "movz %[" #TEMP12 "], %[temp20], %[temp19] \n\t" \ + "sb %[" #TEMP0 "], 0+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" \ + "sb %[" #TEMP4 "], 1+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" \ + "sb %[" #TEMP8 "], 2+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" \ + "sb %[" #TEMP12 "], 3+" XSTR(BPS) "*" #A "(%[temp16]) \n\t" + +// Does one or two inverse transforms. +static WEBP_INLINE void ITransformOne_MIPS32(const uint8_t* ref, + const int16_t* in, + uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + int temp7, temp8, temp9, temp10, temp11, temp12, temp13; + int temp14, temp15, temp16, temp17, temp18, temp19, temp20; + const int* args[3] = {(const int*)ref, (const int*)in, (const int*)dst}; + + __asm__ volatile( + "lw %[temp20], 4(%[args]) \n\t" + VERTICAL_PASS(0, 16, 8, 24, temp4, temp0, temp1, temp2, temp3) + VERTICAL_PASS(2, 18, 10, 26, temp8, temp4, temp5, temp6, temp7) + VERTICAL_PASS(4, 20, 12, 28, temp12, temp8, temp9, temp10, temp11) + VERTICAL_PASS(6, 22, 14, 30, temp20, temp12, temp13, temp14, temp15) + + HORIZONTAL_PASS(0, temp0, temp4, temp8, temp12) + HORIZONTAL_PASS(1, temp1, temp5, temp9, temp13) + HORIZONTAL_PASS(2, temp2, temp6, temp10, temp14) + HORIZONTAL_PASS(3, temp3, temp7, temp11, temp15) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11), + [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14), + [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17), + [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20) + : [args]"r"(args), [kC1]"r"(kC1), [kC2]"r"(kC2) + : "memory", "hi", "lo" + ); +} + +static void ITransform_MIPS32(const uint8_t* ref, const int16_t* in, + uint8_t* dst, int do_two) { + ITransformOne_MIPS32(ref, in, dst); + if (do_two) { + ITransformOne_MIPS32(ref + 4, in + 16, dst + 4); + } +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +// macro for one pass through for loop in QuantizeBlock +// QUANTDIV macro inlined +// J - offset in bytes (kZigzag[n] * 2) +// K - offset in bytes (kZigzag[n] * 4) +// N - offset in bytes (n * 2) +#define QUANTIZE_ONE(J, K, N) \ + "lh %[temp0], " #J "(%[ppin]) \n\t" \ + "lhu %[temp1], " #J "(%[ppsharpen]) \n\t" \ + "lw %[temp2], " #K "(%[ppzthresh]) \n\t" \ + "sra %[sign], %[temp0], 15 \n\t" \ + "xor %[coeff], %[temp0], %[sign] \n\t" \ + "subu %[coeff], %[coeff], %[sign] \n\t" \ + "addu %[coeff], %[coeff], %[temp1] \n\t" \ + "slt %[temp4], %[temp2], %[coeff] \n\t" \ + "addiu %[temp5], $zero, 0 \n\t" \ + "addiu %[level], $zero, 0 \n\t" \ + "beqz %[temp4], 2f \n\t" \ + "lhu %[temp1], " #J "(%[ppiq]) \n\t" \ + "lw %[temp2], " #K "(%[ppbias]) \n\t" \ + "lhu %[temp3], " #J "(%[ppq]) \n\t" \ + "mul %[level], %[coeff], %[temp1] \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "sra %[level], %[level], 17 \n\t" \ + "slt %[temp4], %[max_level], %[level] \n\t" \ + "movn %[level], %[max_level], %[temp4] \n\t" \ + "xor %[level], %[level], %[sign] \n\t" \ + "subu %[level], %[level], %[sign] \n\t" \ + "mul %[temp5], %[level], %[temp3] \n\t" \ +"2: \n\t" \ + "sh %[temp5], " #J "(%[ppin]) \n\t" \ + "sh %[level], " #N "(%[pout]) \n\t" + +static int QuantizeBlock_MIPS32(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + int temp0, temp1, temp2, temp3, temp4, temp5; + int sign, coeff, level, i; + int max_level = MAX_LEVEL; + + int16_t* ppin = &in[0]; + int16_t* pout = &out[0]; + const uint16_t* ppsharpen = &mtx->sharpen_[0]; + const uint32_t* ppzthresh = &mtx->zthresh_[0]; + const uint16_t* ppq = &mtx->q_[0]; + const uint16_t* ppiq = &mtx->iq_[0]; + const uint32_t* ppbias = &mtx->bias_[0]; + + __asm__ volatile( + QUANTIZE_ONE( 0, 0, 0) + QUANTIZE_ONE( 2, 4, 2) + QUANTIZE_ONE( 8, 16, 4) + QUANTIZE_ONE(16, 32, 6) + QUANTIZE_ONE(10, 20, 8) + QUANTIZE_ONE( 4, 8, 10) + QUANTIZE_ONE( 6, 12, 12) + QUANTIZE_ONE(12, 24, 14) + QUANTIZE_ONE(18, 36, 16) + QUANTIZE_ONE(24, 48, 18) + QUANTIZE_ONE(26, 52, 20) + QUANTIZE_ONE(20, 40, 22) + QUANTIZE_ONE(14, 28, 24) + QUANTIZE_ONE(22, 44, 26) + QUANTIZE_ONE(28, 56, 28) + QUANTIZE_ONE(30, 60, 30) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [sign]"=&r"(sign), [coeff]"=&r"(coeff), + [level]"=&r"(level) + : [pout]"r"(pout), [ppin]"r"(ppin), + [ppiq]"r"(ppiq), [max_level]"r"(max_level), + [ppbias]"r"(ppbias), [ppzthresh]"r"(ppzthresh), + [ppsharpen]"r"(ppsharpen), [ppq]"r"(ppq) + : "memory", "hi", "lo" + ); + + // moved out from macro to increase possibility for earlier breaking + for (i = 15; i >= 0; i--) { + if (out[i]) return 1; + } + return 0; +} + +static int Quantize2Blocks_MIPS32(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = QuantizeBlock_MIPS32(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= QuantizeBlock_MIPS32(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +#undef QUANTIZE_ONE + +// macro for one horizontal pass in Disto4x4 (TTransform) +// two calls of function TTransform are merged into single one +// A - offset in bytes to load from a and b buffers +// E..H - offsets in bytes to store first results to tmp buffer +// E1..H1 - offsets in bytes to store second results to tmp buffer +#define HORIZONTAL_PASS(A, E, F, G, H, E1, F1, G1, H1) \ + "lbu %[temp0], 0+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp1], 1+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp2], 2+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp3], 3+" XSTR(BPS) "*" #A "(%[a]) \n\t" \ + "lbu %[temp4], 0+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "lbu %[temp5], 1+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "lbu %[temp6], 2+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "lbu %[temp7], 3+" XSTR(BPS) "*" #A "(%[b]) \n\t" \ + "addu %[temp8], %[temp0], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp2] \n\t" \ + "addu %[temp2], %[temp1], %[temp3] \n\t" \ + "subu %[temp1], %[temp1], %[temp3] \n\t" \ + "addu %[temp3], %[temp4], %[temp6] \n\t" \ + "subu %[temp4], %[temp4], %[temp6] \n\t" \ + "addu %[temp6], %[temp5], %[temp7] \n\t" \ + "subu %[temp5], %[temp5], %[temp7] \n\t" \ + "addu %[temp7], %[temp8], %[temp2] \n\t" \ + "subu %[temp2], %[temp8], %[temp2] \n\t" \ + "addu %[temp8], %[temp0], %[temp1] \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "addu %[temp1], %[temp3], %[temp6] \n\t" \ + "subu %[temp3], %[temp3], %[temp6] \n\t" \ + "addu %[temp6], %[temp4], %[temp5] \n\t" \ + "subu %[temp4], %[temp4], %[temp5] \n\t" \ + "sw %[temp7], " #E "(%[tmp]) \n\t" \ + "sw %[temp2], " #H "(%[tmp]) \n\t" \ + "sw %[temp8], " #F "(%[tmp]) \n\t" \ + "sw %[temp0], " #G "(%[tmp]) \n\t" \ + "sw %[temp1], " #E1 "(%[tmp]) \n\t" \ + "sw %[temp3], " #H1 "(%[tmp]) \n\t" \ + "sw %[temp6], " #F1 "(%[tmp]) \n\t" \ + "sw %[temp4], " #G1 "(%[tmp]) \n\t" + +// macro for one vertical pass in Disto4x4 (TTransform) +// two calls of function TTransform are merged into single one +// since only one accu is available in mips32r1 instruction set +// first is done second call of function TTransform and after +// that first one. +// const int sum1 = TTransform(a, w); +// const int sum2 = TTransform(b, w); +// return abs(sum2 - sum1) >> 5; +// (sum2 - sum1) is calculated with madds (sub2) and msubs (sub1) +// A..D - offsets in bytes to load first results from tmp buffer +// A1..D1 - offsets in bytes to load second results from tmp buffer +// E..H - offsets in bytes to load from w buffer +#define VERTICAL_PASS(A, B, C, D, A1, B1, C1, D1, E, F, G, H) \ + "lw %[temp0], " #A1 "(%[tmp]) \n\t" \ + "lw %[temp1], " #C1 "(%[tmp]) \n\t" \ + "lw %[temp2], " #B1 "(%[tmp]) \n\t" \ + "lw %[temp3], " #D1 "(%[tmp]) \n\t" \ + "addu %[temp8], %[temp0], %[temp1] \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "addu %[temp1], %[temp2], %[temp3] \n\t" \ + "subu %[temp2], %[temp2], %[temp3] \n\t" \ + "addu %[temp3], %[temp8], %[temp1] \n\t" \ + "subu %[temp8], %[temp8], %[temp1] \n\t" \ + "addu %[temp1], %[temp0], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp2] \n\t" \ + "sra %[temp4], %[temp3], 31 \n\t" \ + "sra %[temp5], %[temp1], 31 \n\t" \ + "sra %[temp6], %[temp0], 31 \n\t" \ + "sra %[temp7], %[temp8], 31 \n\t" \ + "xor %[temp3], %[temp3], %[temp4] \n\t" \ + "xor %[temp1], %[temp1], %[temp5] \n\t" \ + "xor %[temp0], %[temp0], %[temp6] \n\t" \ + "xor %[temp8], %[temp8], %[temp7] \n\t" \ + "subu %[temp3], %[temp3], %[temp4] \n\t" \ + "subu %[temp1], %[temp1], %[temp5] \n\t" \ + "subu %[temp0], %[temp0], %[temp6] \n\t" \ + "subu %[temp8], %[temp8], %[temp7] \n\t" \ + "lhu %[temp4], " #E "(%[w]) \n\t" \ + "lhu %[temp5], " #F "(%[w]) \n\t" \ + "lhu %[temp6], " #G "(%[w]) \n\t" \ + "lhu %[temp7], " #H "(%[w]) \n\t" \ + "madd %[temp4], %[temp3] \n\t" \ + "madd %[temp5], %[temp1] \n\t" \ + "madd %[temp6], %[temp0] \n\t" \ + "madd %[temp7], %[temp8] \n\t" \ + "lw %[temp0], " #A "(%[tmp]) \n\t" \ + "lw %[temp1], " #C "(%[tmp]) \n\t" \ + "lw %[temp2], " #B "(%[tmp]) \n\t" \ + "lw %[temp3], " #D "(%[tmp]) \n\t" \ + "addu %[temp8], %[temp0], %[temp1] \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "addu %[temp1], %[temp2], %[temp3] \n\t" \ + "subu %[temp2], %[temp2], %[temp3] \n\t" \ + "addu %[temp3], %[temp8], %[temp1] \n\t" \ + "subu %[temp1], %[temp8], %[temp1] \n\t" \ + "addu %[temp8], %[temp0], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp2] \n\t" \ + "sra %[temp2], %[temp3], 31 \n\t" \ + "xor %[temp3], %[temp3], %[temp2] \n\t" \ + "subu %[temp3], %[temp3], %[temp2] \n\t" \ + "msub %[temp4], %[temp3] \n\t" \ + "sra %[temp2], %[temp8], 31 \n\t" \ + "sra %[temp3], %[temp0], 31 \n\t" \ + "sra %[temp4], %[temp1], 31 \n\t" \ + "xor %[temp8], %[temp8], %[temp2] \n\t" \ + "xor %[temp0], %[temp0], %[temp3] \n\t" \ + "xor %[temp1], %[temp1], %[temp4] \n\t" \ + "subu %[temp8], %[temp8], %[temp2] \n\t" \ + "subu %[temp0], %[temp0], %[temp3] \n\t" \ + "subu %[temp1], %[temp1], %[temp4] \n\t" \ + "msub %[temp5], %[temp8] \n\t" \ + "msub %[temp6], %[temp0] \n\t" \ + "msub %[temp7], %[temp1] \n\t" + +static int Disto4x4_MIPS32(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int tmp[32]; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + + __asm__ volatile( + HORIZONTAL_PASS(0, 0, 4, 8, 12, 64, 68, 72, 76) + HORIZONTAL_PASS(1, 16, 20, 24, 28, 80, 84, 88, 92) + HORIZONTAL_PASS(2, 32, 36, 40, 44, 96, 100, 104, 108) + HORIZONTAL_PASS(3, 48, 52, 56, 60, 112, 116, 120, 124) + "mthi $zero \n\t" + "mtlo $zero \n\t" + VERTICAL_PASS( 0, 16, 32, 48, 64, 80, 96, 112, 0, 8, 16, 24) + VERTICAL_PASS( 4, 20, 36, 52, 68, 84, 100, 116, 2, 10, 18, 26) + VERTICAL_PASS( 8, 24, 40, 56, 72, 88, 104, 120, 4, 12, 20, 28) + VERTICAL_PASS(12, 28, 44, 60, 76, 92, 108, 124, 6, 14, 22, 30) + "mflo %[temp0] \n\t" + "sra %[temp1], %[temp0], 31 \n\t" + "xor %[temp0], %[temp0], %[temp1] \n\t" + "subu %[temp0], %[temp0], %[temp1] \n\t" + "sra %[temp0], %[temp0], 5 \n\t" + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [a]"r"(a), [b]"r"(b), [w]"r"(w), [tmp]"r"(tmp) + : "memory", "hi", "lo" + ); + + return temp0; +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +static int Disto16x16_MIPS32(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4_MIPS32(a + x + y, b + x + y, w); + } + } + return D; +} + +// macro for one horizontal pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A - offset in bytes to load from src and ref buffers +// TEMP0..TEMP3 - registers for corresponding tmp elements +#define HORIZONTAL_PASS(A, TEMP0, TEMP1, TEMP2, TEMP3) \ + "lw %[" #TEMP1 "], 0(%[args]) \n\t" \ + "lw %[" #TEMP2 "], 4(%[args]) \n\t" \ + "lbu %[temp16], 0+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp17], 0+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "lbu %[temp18], 1+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp19], 1+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "subu %[temp20], %[temp16], %[temp17] \n\t" \ + "lbu %[temp16], 2+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp17], 2+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "subu %[" #TEMP0 "], %[temp18], %[temp19] \n\t" \ + "lbu %[temp18], 3+" XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "lbu %[temp19], 3+" XSTR(BPS) "*" #A "(%[" #TEMP2 "]) \n\t" \ + "subu %[" #TEMP1 "], %[temp16], %[temp17] \n\t" \ + "subu %[" #TEMP2 "], %[temp18], %[temp19] \n\t" \ + "addu %[" #TEMP3 "], %[temp20], %[" #TEMP2 "] \n\t" \ + "subu %[" #TEMP2 "], %[temp20], %[" #TEMP2 "] \n\t" \ + "addu %[temp20], %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "subu %[" #TEMP0 "], %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "mul %[temp16], %[" #TEMP2 "], %[c5352] \n\t" \ + "mul %[temp17], %[" #TEMP2 "], %[c2217] \n\t" \ + "mul %[temp18], %[" #TEMP0 "], %[c5352] \n\t" \ + "mul %[temp19], %[" #TEMP0 "], %[c2217] \n\t" \ + "addu %[" #TEMP1 "], %[" #TEMP3 "], %[temp20] \n\t" \ + "subu %[temp20], %[" #TEMP3 "], %[temp20] \n\t" \ + "sll %[" #TEMP0 "], %[" #TEMP1 "], 3 \n\t" \ + "sll %[" #TEMP2 "], %[temp20], 3 \n\t" \ + "addiu %[temp16], %[temp16], 1812 \n\t" \ + "addiu %[temp17], %[temp17], 937 \n\t" \ + "addu %[temp16], %[temp16], %[temp19] \n\t" \ + "subu %[temp17], %[temp17], %[temp18] \n\t" \ + "sra %[" #TEMP1 "], %[temp16], 9 \n\t" \ + "sra %[" #TEMP3 "], %[temp17], 9 \n\t" + +// macro for one vertical pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A..D - offsets in bytes to store to out buffer +// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements +#define VERTICAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12) \ + "addu %[temp16], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "subu %[temp19], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "addu %[temp17], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "subu %[temp18], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "mul %[" #TEMP8 "], %[temp19], %[c2217] \n\t" \ + "mul %[" #TEMP12 "], %[temp18], %[c2217] \n\t" \ + "mul %[" #TEMP4 "], %[temp19], %[c5352] \n\t" \ + "mul %[temp18], %[temp18], %[c5352] \n\t" \ + "addiu %[temp16], %[temp16], 7 \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 4 \n\t" \ + "addu %[" #TEMP12 "], %[" #TEMP12 "], %[" #TEMP4 "] \n\t" \ + "subu %[" #TEMP4 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 4 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 30000 \n\t" \ + "addiu %[" #TEMP12 "], %[" #TEMP12 "], 12000 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 21000 \n\t" \ + "subu %[" #TEMP8 "], %[" #TEMP8 "], %[temp18] \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 16 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 16 \n\t" \ + "addiu %[temp16], %[" #TEMP12 "], 1 \n\t" \ + "movn %[" #TEMP12 "], %[temp16], %[temp19] \n\t" \ + "sh %[" #TEMP0 "], " #A "(%[temp20]) \n\t" \ + "sh %[" #TEMP4 "], " #C "(%[temp20]) \n\t" \ + "sh %[" #TEMP8 "], " #D "(%[temp20]) \n\t" \ + "sh %[" #TEMP12 "], " #B "(%[temp20]) \n\t" + +static void FTransform_MIPS32(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + int temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16; + int temp17, temp18, temp19, temp20; + const int c2217 = 2217; + const int c5352 = 5352; + const int* const args[3] = + { (const int*)src, (const int*)ref, (const int*)out }; + + __asm__ volatile( + HORIZONTAL_PASS(0, temp0, temp1, temp2, temp3) + HORIZONTAL_PASS(1, temp4, temp5, temp6, temp7) + HORIZONTAL_PASS(2, temp8, temp9, temp10, temp11) + HORIZONTAL_PASS(3, temp12, temp13, temp14, temp15) + "lw %[temp20], 8(%[args]) \n\t" + VERTICAL_PASS(0, 8, 16, 24, temp0, temp4, temp8, temp12) + VERTICAL_PASS(2, 10, 18, 26, temp1, temp5, temp9, temp13) + VERTICAL_PASS(4, 12, 20, 28, temp2, temp6, temp10, temp14) + VERTICAL_PASS(6, 14, 22, 30, temp3, temp7, temp11, temp15) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11), + [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), [temp14]"=&r"(temp14), + [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), [temp17]"=&r"(temp17), + [temp18]"=&r"(temp18), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20) + : [args]"r"(args), [c2217]"r"(c2217), [c5352]"r"(c5352) + : "memory", "hi", "lo" + ); +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +#if !defined(WORK_AROUND_GCC) + +#define GET_SSE_INNER(A, B, C, D) \ + "lbu %[temp0], " #A "(%[a]) \n\t" \ + "lbu %[temp1], " #A "(%[b]) \n\t" \ + "lbu %[temp2], " #B "(%[a]) \n\t" \ + "lbu %[temp3], " #B "(%[b]) \n\t" \ + "lbu %[temp4], " #C "(%[a]) \n\t" \ + "lbu %[temp5], " #C "(%[b]) \n\t" \ + "lbu %[temp6], " #D "(%[a]) \n\t" \ + "lbu %[temp7], " #D "(%[b]) \n\t" \ + "subu %[temp0], %[temp0], %[temp1] \n\t" \ + "subu %[temp2], %[temp2], %[temp3] \n\t" \ + "subu %[temp4], %[temp4], %[temp5] \n\t" \ + "subu %[temp6], %[temp6], %[temp7] \n\t" \ + "madd %[temp0], %[temp0] \n\t" \ + "madd %[temp2], %[temp2] \n\t" \ + "madd %[temp4], %[temp4] \n\t" \ + "madd %[temp6], %[temp6] \n\t" + +#define GET_SSE(A, B, C, D) \ + GET_SSE_INNER(A, A + 1, A + 2, A + 3) \ + GET_SSE_INNER(B, B + 1, B + 2, B + 3) \ + GET_SSE_INNER(C, C + 1, C + 2, C + 3) \ + GET_SSE_INNER(D, D + 1, D + 2, D + 3) + +static int SSE16x16_MIPS32(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + GET_SSE( 8 * BPS, 4 + 8 * BPS, 8 + 8 * BPS, 12 + 8 * BPS) + GET_SSE( 9 * BPS, 4 + 9 * BPS, 8 + 9 * BPS, 12 + 9 * BPS) + GET_SSE(10 * BPS, 4 + 10 * BPS, 8 + 10 * BPS, 12 + 10 * BPS) + GET_SSE(11 * BPS, 4 + 11 * BPS, 8 + 11 * BPS, 12 + 11 * BPS) + GET_SSE(12 * BPS, 4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS) + GET_SSE(13 * BPS, 4 + 13 * BPS, 8 + 13 * BPS, 12 + 13 * BPS) + GET_SSE(14 * BPS, 4 + 14 * BPS, 8 + 14 * BPS, 12 + 14 * BPS) + GET_SSE(15 * BPS, 4 + 15 * BPS, 8 + 15 * BPS, 12 + 15 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE16x8_MIPS32(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE8x8_MIPS32(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE(0 * BPS, 4 + 0 * BPS, 1 * BPS, 4 + 1 * BPS) + GET_SSE(2 * BPS, 4 + 2 * BPS, 3 * BPS, 4 + 3 * BPS) + GET_SSE(4 * BPS, 4 + 4 * BPS, 5 * BPS, 4 + 5 * BPS) + GET_SSE(6 * BPS, 4 + 6 * BPS, 7 * BPS, 4 + 7 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE4x4_MIPS32(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + + __asm__ volatile( + "mult $zero, $zero \n\t" + + GET_SSE(0 * BPS, 1 * BPS, 2 * BPS, 3 * BPS) + + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +#undef GET_SSE +#undef GET_SSE_INNER + +#endif // !WORK_AROUND_GCC + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitMIPS32(void) { + VP8ITransform = ITransform_MIPS32; + VP8FTransform = FTransform_MIPS32; + + VP8EncQuantizeBlock = QuantizeBlock_MIPS32; + VP8EncQuantize2Blocks = Quantize2Blocks_MIPS32; + + VP8TDisto4x4 = Disto4x4_MIPS32; + VP8TDisto16x16 = Disto16x16_MIPS32; + +#if !defined(WORK_AROUND_GCC) + VP8SSE16x16 = SSE16x16_MIPS32; + VP8SSE8x8 = SSE8x8_MIPS32; + VP8SSE16x8 = SSE16x8_MIPS32; + VP8SSE4x4 = SSE4x4_MIPS32; +#endif +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8EncDspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/media/libwebp/dsp/enc_mips_dsp_r2.c b/media/libwebp/dsp/enc_mips_dsp_r2.c new file mode 100644 index 0000000000..cf4c85b59c --- /dev/null +++ b/media/libwebp/dsp/enc_mips_dsp_r2.c @@ -0,0 +1,1517 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of speed-critical encoding functions. +// +// Author(s): Darko Laus (darko.laus@imgtec.com) +// Mirko Raus (mirko.raus@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../dsp/mips_macro.h" +#include "../enc/cost_enc.h" +#include "../enc/vp8i_enc.h" + +static const int kC1 = 20091 + (1 << 16); +static const int kC2 = 35468; + +// O - output +// I - input (macro doesn't change it) +#define ADD_SUB_HALVES_X4(O0, O1, O2, O3, O4, O5, O6, O7, \ + I0, I1, I2, I3, I4, I5, I6, I7) \ + "addq.ph %[" #O0 "], %[" #I0 "], %[" #I1 "] \n\t" \ + "subq.ph %[" #O1 "], %[" #I0 "], %[" #I1 "] \n\t" \ + "addq.ph %[" #O2 "], %[" #I2 "], %[" #I3 "] \n\t" \ + "subq.ph %[" #O3 "], %[" #I2 "], %[" #I3 "] \n\t" \ + "addq.ph %[" #O4 "], %[" #I4 "], %[" #I5 "] \n\t" \ + "subq.ph %[" #O5 "], %[" #I4 "], %[" #I5 "] \n\t" \ + "addq.ph %[" #O6 "], %[" #I6 "], %[" #I7 "] \n\t" \ + "subq.ph %[" #O7 "], %[" #I6 "], %[" #I7 "] \n\t" + +// IO - input/output +#define ABS_X8(IO0, IO1, IO2, IO3, IO4, IO5, IO6, IO7) \ + "absq_s.ph %[" #IO0 "], %[" #IO0 "] \n\t" \ + "absq_s.ph %[" #IO1 "], %[" #IO1 "] \n\t" \ + "absq_s.ph %[" #IO2 "], %[" #IO2 "] \n\t" \ + "absq_s.ph %[" #IO3 "], %[" #IO3 "] \n\t" \ + "absq_s.ph %[" #IO4 "], %[" #IO4 "] \n\t" \ + "absq_s.ph %[" #IO5 "], %[" #IO5 "] \n\t" \ + "absq_s.ph %[" #IO6 "], %[" #IO6 "] \n\t" \ + "absq_s.ph %[" #IO7 "], %[" #IO7 "] \n\t" + +// dpa.w.ph $ac0 temp0 ,temp1 +// $ac += temp0[31..16] * temp1[31..16] + temp0[15..0] * temp1[15..0] +// dpax.w.ph $ac0 temp0 ,temp1 +// $ac += temp0[31..16] * temp1[15..0] + temp0[15..0] * temp1[31..16] +// O - output +// I - input (macro doesn't change it) +#define MUL_HALF(O0, I0, I1, I2, I3, I4, I5, I6, I7, \ + I8, I9, I10, I11, I12, I13, I14, I15) \ + "mult $ac0, $zero, $zero \n\t" \ + "dpa.w.ph $ac0, %[" #I2 "], %[" #I0 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I5 "], %[" #I6 "] \n\t" \ + "dpa.w.ph $ac0, %[" #I8 "], %[" #I9 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I11 "], %[" #I4 "] \n\t" \ + "dpa.w.ph $ac0, %[" #I12 "], %[" #I7 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I13 "], %[" #I1 "] \n\t" \ + "dpa.w.ph $ac0, %[" #I14 "], %[" #I3 "] \n\t" \ + "dpax.w.ph $ac0, %[" #I15 "], %[" #I10 "] \n\t" \ + "mflo %[" #O0 "], $ac0 \n\t" + +#define OUTPUT_EARLY_CLOBBER_REGS_17() \ + OUTPUT_EARLY_CLOBBER_REGS_10(), \ + [temp11]"=&r"(temp11), [temp12]"=&r"(temp12), [temp13]"=&r"(temp13), \ + [temp14]"=&r"(temp14), [temp15]"=&r"(temp15), [temp16]"=&r"(temp16), \ + [temp17]"=&r"(temp17) + +// macro for one horizontal pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A - offset in bytes to load from src and ref buffers +// TEMP0..TEMP3 - registers for corresponding tmp elements +#define HORIZONTAL_PASS(A, TEMP0, TEMP1, TEMP2, TEMP3) \ + "lw %[" #TEMP0 "], 0(%[args]) \n\t" \ + "lw %[" #TEMP1 "], 4(%[args]) \n\t" \ + "lw %[" #TEMP2 "], " XSTR(BPS) "*" #A "(%[" #TEMP0 "]) \n\t" \ + "lw %[" #TEMP3 "], " XSTR(BPS) "*" #A "(%[" #TEMP1 "]) \n\t" \ + "preceu.ph.qbl %[" #TEMP0 "], %[" #TEMP2 "] \n\t" \ + "preceu.ph.qbl %[" #TEMP1 "], %[" #TEMP3 "] \n\t" \ + "preceu.ph.qbr %[" #TEMP2 "], %[" #TEMP2 "] \n\t" \ + "preceu.ph.qbr %[" #TEMP3 "], %[" #TEMP3 "] \n\t" \ + "subq.ph %[" #TEMP0 "], %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "subq.ph %[" #TEMP2 "], %[" #TEMP2 "], %[" #TEMP3 "] \n\t" \ + "rotr %[" #TEMP0 "], %[" #TEMP0 "], 16 \n\t" \ + "addq.ph %[" #TEMP1 "], %[" #TEMP2 "], %[" #TEMP0 "] \n\t" \ + "subq.ph %[" #TEMP3 "], %[" #TEMP2 "], %[" #TEMP0 "] \n\t" \ + "seh %[" #TEMP0 "], %[" #TEMP1 "] \n\t" \ + "sra %[temp16], %[" #TEMP1 "], 16 \n\t" \ + "seh %[temp19], %[" #TEMP3 "] \n\t" \ + "sra %[" #TEMP3 "], %[" #TEMP3 "], 16 \n\t" \ + "subu %[" #TEMP2 "], %[" #TEMP0 "], %[temp16] \n\t" \ + "addu %[" #TEMP0 "], %[" #TEMP0 "], %[temp16] \n\t" \ + "mul %[temp17], %[temp19], %[c2217] \n\t" \ + "mul %[temp18], %[" #TEMP3 "], %[c5352] \n\t" \ + "mul %[" #TEMP1 "], %[temp19], %[c5352] \n\t" \ + "mul %[temp16], %[" #TEMP3 "], %[c2217] \n\t" \ + "sll %[" #TEMP2 "], %[" #TEMP2 "], 3 \n\t" \ + "sll %[" #TEMP0 "], %[" #TEMP0 "], 3 \n\t" \ + "subu %[" #TEMP3 "], %[temp17], %[temp18] \n\t" \ + "addu %[" #TEMP1 "], %[temp16], %[" #TEMP1 "] \n\t" \ + "addiu %[" #TEMP3 "], %[" #TEMP3 "], 937 \n\t" \ + "addiu %[" #TEMP1 "], %[" #TEMP1 "], 1812 \n\t" \ + "sra %[" #TEMP3 "], %[" #TEMP3 "], 9 \n\t" \ + "sra %[" #TEMP1 "], %[" #TEMP1 "], 9 \n\t" + +// macro for one vertical pass in FTransform +// temp0..temp15 holds tmp[0]..tmp[15] +// A..D - offsets in bytes to store to out buffer +// TEMP0, TEMP4, TEMP8 and TEMP12 - registers for corresponding tmp elements +#define VERTICAL_PASS(A, B, C, D, TEMP0, TEMP4, TEMP8, TEMP12) \ + "addu %[temp16], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "subu %[temp19], %[" #TEMP0 "], %[" #TEMP12 "] \n\t" \ + "addu %[temp17], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "subu %[temp18], %[" #TEMP4 "], %[" #TEMP8 "] \n\t" \ + "mul %[" #TEMP8 "], %[temp19], %[c2217] \n\t" \ + "mul %[" #TEMP12 "], %[temp18], %[c2217] \n\t" \ + "mul %[" #TEMP4 "], %[temp19], %[c5352] \n\t" \ + "mul %[temp18], %[temp18], %[c5352] \n\t" \ + "addiu %[temp16], %[temp16], 7 \n\t" \ + "addu %[" #TEMP0 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP0 "], %[" #TEMP0 "], 4 \n\t" \ + "addu %[" #TEMP12 "], %[" #TEMP12 "], %[" #TEMP4 "] \n\t" \ + "subu %[" #TEMP4 "], %[temp16], %[temp17] \n\t" \ + "sra %[" #TEMP4 "], %[" #TEMP4 "], 4 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 30000 \n\t" \ + "addiu %[" #TEMP12 "], %[" #TEMP12 "], 12000 \n\t" \ + "addiu %[" #TEMP8 "], %[" #TEMP8 "], 21000 \n\t" \ + "subu %[" #TEMP8 "], %[" #TEMP8 "], %[temp18] \n\t" \ + "sra %[" #TEMP12 "], %[" #TEMP12 "], 16 \n\t" \ + "sra %[" #TEMP8 "], %[" #TEMP8 "], 16 \n\t" \ + "addiu %[temp16], %[" #TEMP12 "], 1 \n\t" \ + "movn %[" #TEMP12 "], %[temp16], %[temp19] \n\t" \ + "sh %[" #TEMP0 "], " #A "(%[temp20]) \n\t" \ + "sh %[" #TEMP4 "], " #C "(%[temp20]) \n\t" \ + "sh %[" #TEMP8 "], " #D "(%[temp20]) \n\t" \ + "sh %[" #TEMP12 "], " #B "(%[temp20]) \n\t" + +static void FTransform_MIPSdspR2(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + const int c2217 = 2217; + const int c5352 = 5352; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + int temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16; + int temp17, temp18, temp19, temp20; + const int* const args[3] = + { (const int*)src, (const int*)ref, (const int*)out }; + + __asm__ volatile ( + HORIZONTAL_PASS(0, temp0, temp1, temp2, temp3) + HORIZONTAL_PASS(1, temp4, temp5, temp6, temp7) + HORIZONTAL_PASS(2, temp8, temp9, temp10, temp11) + HORIZONTAL_PASS(3, temp12, temp13, temp14, temp15) + "lw %[temp20], 8(%[args]) \n\t" + VERTICAL_PASS(0, 8, 16, 24, temp0, temp4, temp8, temp12) + VERTICAL_PASS(2, 10, 18, 26, temp1, temp5, temp9, temp13) + VERTICAL_PASS(4, 12, 20, 28, temp2, temp6, temp10, temp14) + VERTICAL_PASS(6, 14, 22, 30, temp3, temp7, temp11, temp15) + OUTPUT_EARLY_CLOBBER_REGS_18(), + [temp0]"=&r"(temp0), [temp19]"=&r"(temp19), [temp20]"=&r"(temp20) + : [args]"r"(args), [c2217]"r"(c2217), [c5352]"r"(c5352) + : "memory", "hi", "lo" + ); +} + +#undef VERTICAL_PASS +#undef HORIZONTAL_PASS + +static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in, + uint8_t* dst) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17, temp18; + + __asm__ volatile ( + "ulw %[temp1], 0(%[in]) \n\t" + "ulw %[temp2], 16(%[in]) \n\t" + LOAD_IN_X2(temp5, temp6, 24, 26) + ADD_SUB_HALVES(temp3, temp4, temp1, temp2) + LOAD_IN_X2(temp1, temp2, 8, 10) + MUL_SHIFT_SUM(temp7, temp8, temp9, temp10, temp11, temp12, temp13, temp14, + temp10, temp8, temp9, temp7, temp1, temp2, temp5, temp6, + temp13, temp11, temp14, temp12) + INSERT_HALF_X2(temp8, temp7, temp10, temp9) + "ulw %[temp17], 4(%[in]) \n\t" + "ulw %[temp18], 20(%[in]) \n\t" + ADD_SUB_HALVES(temp1, temp2, temp3, temp8) + ADD_SUB_HALVES(temp5, temp6, temp4, temp7) + ADD_SUB_HALVES(temp7, temp8, temp17, temp18) + LOAD_IN_X2(temp17, temp18, 12, 14) + LOAD_IN_X2(temp9, temp10, 28, 30) + MUL_SHIFT_SUM(temp11, temp12, temp13, temp14, temp15, temp16, temp4, temp17, + temp12, temp14, temp11, temp13, temp17, temp18, temp9, temp10, + temp15, temp4, temp16, temp17) + INSERT_HALF_X2(temp11, temp12, temp13, temp14) + ADD_SUB_HALVES(temp17, temp8, temp8, temp11) + ADD_SUB_HALVES(temp3, temp4, temp7, temp12) + + // horizontal + SRA_16(temp9, temp10, temp11, temp12, temp1, temp2, temp5, temp6) + INSERT_HALF_X2(temp1, temp6, temp5, temp2) + SRA_16(temp13, temp14, temp15, temp16, temp3, temp4, temp17, temp8) + "repl.ph %[temp2], 0x4 \n\t" + INSERT_HALF_X2(temp3, temp8, temp17, temp4) + "addq.ph %[temp1], %[temp1], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp2] \n\t" + ADD_SUB_HALVES(temp2, temp4, temp1, temp3) + ADD_SUB_HALVES(temp5, temp7, temp6, temp8) + MUL_SHIFT_SUM(temp1, temp3, temp6, temp8, temp9, temp13, temp17, temp18, + temp3, temp13, temp1, temp9, temp9, temp13, temp11, temp15, + temp6, temp17, temp8, temp18) + MUL_SHIFT_SUM(temp6, temp8, temp18, temp17, temp11, temp15, temp12, temp16, + temp8, temp15, temp6, temp11, temp12, temp16, temp10, temp14, + temp18, temp12, temp17, temp16) + INSERT_HALF_X2(temp1, temp3, temp9, temp13) + INSERT_HALF_X2(temp6, temp8, temp11, temp15) + SHIFT_R_SUM_X2(temp9, temp10, temp11, temp12, temp13, temp14, temp15, + temp16, temp2, temp4, temp5, temp7, temp3, temp1, temp8, + temp6) + PACK_2_HALVES_TO_WORD(temp1, temp2, temp3, temp4, temp9, temp12, temp13, + temp16, temp11, temp10, temp15, temp14) + LOAD_WITH_OFFSET_X4(temp10, temp11, temp14, temp15, ref, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5, temp6, temp7, temp8, temp17, temp18, temp10, + temp11, temp10, temp11, temp14, temp15) + STORE_SAT_SUM_X2(temp5, temp6, temp7, temp8, temp17, temp18, temp10, temp11, + temp9, temp12, temp1, temp2, temp13, temp16, temp3, temp4, + dst, 0, 1, 2, 3, BPS) + + OUTPUT_EARLY_CLOBBER_REGS_18() + : [dst]"r"(dst), [in]"r"(in), [kC1]"r"(kC1), [kC2]"r"(kC2), [ref]"r"(ref) + : "memory", "hi", "lo" + ); +} + +static void ITransform_MIPSdspR2(const uint8_t* ref, const int16_t* in, + uint8_t* dst, int do_two) { + ITransformOne(ref, in, dst); + if (do_two) { + ITransformOne(ref + 4, in + 16, dst + 4); + } +} + +static int Disto4x4_MIPSdspR2(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, temp9; + int temp10, temp11, temp12, temp13, temp14, temp15, temp16, temp17; + + __asm__ volatile ( + LOAD_WITH_OFFSET_X4(temp1, temp2, temp3, temp4, a, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5, temp6, temp7, temp8, temp9,temp10, temp11, + temp12, temp1, temp2, temp3, temp4) + ADD_SUB_HALVES_X4(temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp5, temp6, temp7, temp8, temp9, temp10, temp11, temp12) + PACK_2_HALVES_TO_WORD(temp9, temp10, temp11, temp12, temp1, temp3, temp5, + temp7, temp2, temp4, temp6, temp8) + ADD_SUB_HALVES_X4(temp2, temp4, temp6, temp8, temp9, temp1, temp3, temp10, + temp1, temp9, temp3, temp10, temp5, temp11, temp7, temp12) + ADD_SUB_HALVES_X4(temp5, temp11, temp7, temp2, temp9, temp3, temp6, temp12, + temp2, temp9, temp6, temp3, temp4, temp1, temp8, temp10) + ADD_SUB_HALVES_X4(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2, + temp5, temp7, temp11, temp2, temp9, temp6, temp3, temp12) + ABS_X8(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2) + LOAD_WITH_OFFSET_X4(temp3, temp6, temp9, temp12, w, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + LOAD_WITH_OFFSET_X4(temp13, temp14, temp15, temp16, w, + 0, 4, 8, 12, + 1, 1, 1, 1, + 16) + MUL_HALF(temp17, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16) + LOAD_WITH_OFFSET_X4(temp1, temp2, temp3, temp4, b, + 0, 0, 0, 0, + 0, 1, 2, 3, + BPS) + CONVERT_2_BYTES_TO_HALF(temp5,temp6, temp7, temp8, temp9,temp10, temp11, + temp12, temp1, temp2, temp3, temp4) + ADD_SUB_HALVES_X4(temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp5, temp6, temp7, temp8, temp9, temp10, temp11, temp12) + PACK_2_HALVES_TO_WORD(temp9, temp10, temp11, temp12, temp1, temp3, temp5, + temp7, temp2, temp4, temp6, temp8) + ADD_SUB_HALVES_X4(temp2, temp4, temp6, temp8, temp9, temp1, temp3, temp10, + temp1, temp9, temp3, temp10, temp5, temp11, temp7, temp12) + ADD_SUB_HALVES_X4(temp5, temp11, temp7, temp2, temp9, temp3, temp6, temp12, + temp2, temp9, temp6, temp3, temp4, temp1, temp8, temp10) + ADD_SUB_HALVES_X4(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2, + temp5, temp7, temp11, temp2, temp9, temp6, temp3, temp12) + ABS_X8(temp1, temp4, temp10, temp8, temp7, temp11, temp5, temp2) + LOAD_WITH_OFFSET_X4(temp3, temp6, temp9, temp12, w, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + LOAD_WITH_OFFSET_X4(temp13, temp14, temp15, temp16, w, + 0, 4, 8, 12, + 1, 1, 1, 1, + 16) + MUL_HALF(temp3, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8, + temp9, temp10, temp11, temp12, temp13, temp14, temp15, temp16) + OUTPUT_EARLY_CLOBBER_REGS_17() + : [a]"r"(a), [b]"r"(b), [w]"r"(w) + : "memory", "hi", "lo" + ); + return abs(temp3 - temp17) >> 5; +} + +static int Disto16x16_MIPSdspR2(const uint8_t* const a, + const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4_MIPSdspR2(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ +// Intra predictions + +#define FILL_PART(J, SIZE) \ + "usw %[value], 0+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + "usw %[value], 4+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + ".if " #SIZE " == 16 \n\t" \ + "usw %[value], 8+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + "usw %[value], 12+" #J "*" XSTR(BPS) "(%[dst]) \n\t" \ + ".endif \n\t" + +#define FILL_8_OR_16(DST, VALUE, SIZE) do { \ + int value = (VALUE); \ + __asm__ volatile ( \ + "replv.qb %[value], %[value] \n\t" \ + FILL_PART( 0, SIZE) \ + FILL_PART( 1, SIZE) \ + FILL_PART( 2, SIZE) \ + FILL_PART( 3, SIZE) \ + FILL_PART( 4, SIZE) \ + FILL_PART( 5, SIZE) \ + FILL_PART( 6, SIZE) \ + FILL_PART( 7, SIZE) \ + ".if " #SIZE " == 16 \n\t" \ + FILL_PART( 8, 16) \ + FILL_PART( 9, 16) \ + FILL_PART(10, 16) \ + FILL_PART(11, 16) \ + FILL_PART(12, 16) \ + FILL_PART(13, 16) \ + FILL_PART(14, 16) \ + FILL_PART(15, 16) \ + ".endif \n\t" \ + : [value]"+&r"(value) \ + : [dst]"r"((DST)) \ + : "memory" \ + ); \ +} while (0) + +#define VERTICAL_PRED(DST, TOP, SIZE) \ +static WEBP_INLINE void VerticalPred##SIZE(uint8_t* (DST), \ + const uint8_t* (TOP)) { \ + int j; \ + if ((TOP)) { \ + for (j = 0; j < (SIZE); ++j) memcpy((DST) + j * BPS, (TOP), (SIZE)); \ + } else { \ + FILL_8_OR_16((DST), 127, (SIZE)); \ + } \ +} + +VERTICAL_PRED(dst, top, 8) +VERTICAL_PRED(dst, top, 16) + +#undef VERTICAL_PRED + +#define HORIZONTAL_PRED(DST, LEFT, SIZE) \ +static WEBP_INLINE void HorizontalPred##SIZE(uint8_t* (DST), \ + const uint8_t* (LEFT)) { \ + if (LEFT) { \ + int j; \ + for (j = 0; j < (SIZE); ++j) { \ + memset((DST) + j * BPS, (LEFT)[j], (SIZE)); \ + } \ + } else { \ + FILL_8_OR_16((DST), 129, (SIZE)); \ + } \ +} + +HORIZONTAL_PRED(dst, left, 8) +HORIZONTAL_PRED(dst, left, 16) + +#undef HORIZONTAL_PRED + +#define CLIPPING() \ + "preceu.ph.qbl %[temp2], %[temp0] \n\t" \ + "preceu.ph.qbr %[temp0], %[temp0] \n\t" \ + "preceu.ph.qbl %[temp3], %[temp1] \n\t" \ + "preceu.ph.qbr %[temp1], %[temp1] \n\t" \ + "addu.ph %[temp2], %[temp2], %[leftY_1] \n\t" \ + "addu.ph %[temp0], %[temp0], %[leftY_1] \n\t" \ + "addu.ph %[temp3], %[temp3], %[leftY_1] \n\t" \ + "addu.ph %[temp1], %[temp1], %[leftY_1] \n\t" \ + "shll_s.ph %[temp2], %[temp2], 7 \n\t" \ + "shll_s.ph %[temp0], %[temp0], 7 \n\t" \ + "shll_s.ph %[temp3], %[temp3], 7 \n\t" \ + "shll_s.ph %[temp1], %[temp1], 7 \n\t" \ + "precrqu_s.qb.ph %[temp0], %[temp2], %[temp0] \n\t" \ + "precrqu_s.qb.ph %[temp1], %[temp3], %[temp1] \n\t" + +#define CLIP_8B_TO_DST(DST, LEFT, TOP, SIZE) do { \ + int leftY_1 = ((int)(LEFT)[y] << 16) + (LEFT)[y]; \ + int temp0, temp1, temp2, temp3; \ + __asm__ volatile ( \ + "replv.ph %[leftY_1], %[leftY_1] \n\t" \ + "ulw %[temp0], 0(%[top]) \n\t" \ + "ulw %[temp1], 4(%[top]) \n\t" \ + "subu.ph %[leftY_1], %[leftY_1], %[left_1] \n\t" \ + CLIPPING() \ + "usw %[temp0], 0(%[dst]) \n\t" \ + "usw %[temp1], 4(%[dst]) \n\t" \ + ".if " #SIZE " == 16 \n\t" \ + "ulw %[temp0], 8(%[top]) \n\t" \ + "ulw %[temp1], 12(%[top]) \n\t" \ + CLIPPING() \ + "usw %[temp0], 8(%[dst]) \n\t" \ + "usw %[temp1], 12(%[dst]) \n\t" \ + ".endif \n\t" \ + : [leftY_1]"+&r"(leftY_1), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), \ + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3) \ + : [left_1]"r"(left_1), [top]"r"((TOP)), [dst]"r"((DST)) \ + : "memory" \ + ); \ +} while (0) + +#define CLIP_TO_DST(DST, LEFT, TOP, SIZE) do { \ + int y; \ + const int left_1 = ((int)(LEFT)[-1] << 16) + (LEFT)[-1]; \ + for (y = 0; y < (SIZE); ++y) { \ + CLIP_8B_TO_DST((DST), (LEFT), (TOP), (SIZE)); \ + (DST) += BPS; \ + } \ +} while (0) + +#define TRUE_MOTION(DST, LEFT, TOP, SIZE) \ +static WEBP_INLINE void TrueMotion##SIZE(uint8_t* (DST), const uint8_t* (LEFT),\ + const uint8_t* (TOP)) { \ + if ((LEFT) != NULL) { \ + if ((TOP) != NULL) { \ + CLIP_TO_DST((DST), (LEFT), (TOP), (SIZE)); \ + } else { \ + HorizontalPred##SIZE((DST), (LEFT)); \ + } \ + } else { \ + /* true motion without left samples (hence: with default 129 value) */ \ + /* is equivalent to VE prediction where you just copy the top samples. */ \ + /* Note that if top samples are not available, the default value is */ \ + /* then 129, and not 127 as in the VerticalPred case. */ \ + if ((TOP) != NULL) { \ + VerticalPred##SIZE((DST), (TOP)); \ + } else { \ + FILL_8_OR_16((DST), 129, (SIZE)); \ + } \ + } \ +} + +TRUE_MOTION(dst, left, top, 8) +TRUE_MOTION(dst, left, top, 16) + +#undef TRUE_MOTION +#undef CLIP_TO_DST +#undef CLIP_8B_TO_DST +#undef CLIPPING + +static WEBP_INLINE void DCMode16(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + int DC, DC1; + int temp0, temp1, temp2, temp3; + + __asm__ volatile( + "beqz %[top], 2f \n\t" + LOAD_WITH_OFFSET_X4(temp0, temp1, temp2, temp3, top, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[DC], %[temp0], %[temp2] \n\t" + "move %[DC1], %[DC] \n\t" + "beqz %[left], 1f \n\t" + LOAD_WITH_OFFSET_X4(temp0, temp1, temp2, temp3, left, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[DC1], %[temp0], %[temp2] \n\t" + "1: \n\t" + "addu %[DC], %[DC], %[DC1] \n\t" + "j 3f \n\t" + "2: \n\t" + "beqz %[left], 4f \n\t" + LOAD_WITH_OFFSET_X4(temp0, temp1, temp2, temp3, left, + 0, 4, 8, 12, + 0, 0, 0, 0, + 0) + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addu %[DC], %[temp0], %[temp2] \n\t" + "addu %[DC], %[DC], %[DC] \n\t" + "3: \n\t" + "shra_r.w %[DC], %[DC], 5 \n\t" + "j 5f \n\t" + "4: \n\t" + "li %[DC], 0x80 \n\t" + "5: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [DC]"=&r"(DC), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [DC1]"=&r"(DC1) + : [left]"r"(left), [top]"r"(top) + : "memory" + ); + + FILL_8_OR_16(dst, DC, 16); +} + +static WEBP_INLINE void DCMode8(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + int DC, DC1; + int temp0, temp1, temp2, temp3; + + __asm__ volatile( + "beqz %[top], 2f \n\t" + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], 4(%[top]) \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[DC], %[temp0], %[temp1] \n\t" + "move %[DC1], %[DC] \n\t" + "beqz %[left], 1f \n\t" + "ulw %[temp2], 0(%[left]) \n\t" + "ulw %[temp3], 4(%[left]) \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[DC1], %[temp2], %[temp3] \n\t" + "1: \n\t" + "addu %[DC], %[DC], %[DC1] \n\t" + "j 3f \n\t" + "2: \n\t" + "beqz %[left], 4f \n\t" + "ulw %[temp2], 0(%[left]) \n\t" + "ulw %[temp3], 4(%[left]) \n\t" + "raddu.w.qb %[temp2], %[temp2] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "addu %[DC], %[temp2], %[temp3] \n\t" + "addu %[DC], %[DC], %[DC] \n\t" + "3: \n\t" + "shra_r.w %[DC], %[DC], 4 \n\t" + "j 5f \n\t" + "4: \n\t" + "li %[DC], 0x80 \n\t" + "5: \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [DC]"=&r"(DC), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [DC1]"=&r"(DC1) + : [left]"r"(left), [top]"r"(top) + : "memory" + ); + + FILL_8_OR_16(dst, DC, 8); +} + +static void DC4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1; + __asm__ volatile( + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], -5(%[top]) \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "raddu.w.qb %[temp1], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addiu %[temp0], %[temp0], 4 \n\t" + "srl %[temp0], %[temp0], 3 \n\t" + "replv.qb %[temp0], %[temp0] \n\t" + "usw %[temp0], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void TM4(uint8_t* dst, const uint8_t* top) { + int a10, a32, temp0, temp1, temp2, temp3, temp4, temp5; + const int c35 = 0xff00ff; + __asm__ volatile ( + "lbu %[temp1], 0(%[top]) \n\t" + "lbu %[a10], 1(%[top]) \n\t" + "lbu %[temp2], 2(%[top]) \n\t" + "lbu %[a32], 3(%[top]) \n\t" + "ulw %[temp0], -5(%[top]) \n\t" + "lbu %[temp4], -1(%[top]) \n\t" + "append %[a10], %[temp1], 16 \n\t" + "append %[a32], %[temp2], 16 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "shrl.ph %[temp1], %[temp0], 8 \n\t" + "and %[temp0], %[temp0], %[c35] \n\t" + "subu.ph %[temp1], %[temp1], %[temp4] \n\t" + "subu.ph %[temp0], %[temp0], %[temp4] \n\t" + "srl %[temp2], %[temp1], 16 \n\t" + "srl %[temp3], %[temp0], 16 \n\t" + "replv.ph %[temp2], %[temp2] \n\t" + "replv.ph %[temp3], %[temp3] \n\t" + "replv.ph %[temp4], %[temp1] \n\t" + "replv.ph %[temp5], %[temp0] \n\t" + "addu.ph %[temp0], %[temp3], %[a10] \n\t" + "addu.ph %[temp1], %[temp3], %[a32] \n\t" + "addu.ph %[temp3], %[temp2], %[a10] \n\t" + "addu.ph %[temp2], %[temp2], %[a32] \n\t" + "shll_s.ph %[temp0], %[temp0], 7 \n\t" + "shll_s.ph %[temp1], %[temp1], 7 \n\t" + "shll_s.ph %[temp3], %[temp3], 7 \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "precrqu_s.qb.ph %[temp0], %[temp1], %[temp0] \n\t" + "precrqu_s.qb.ph %[temp1], %[temp2], %[temp3] \n\t" + "addu.ph %[temp2], %[temp5], %[a10] \n\t" + "addu.ph %[temp3], %[temp5], %[a32] \n\t" + "addu.ph %[temp5], %[temp4], %[a10] \n\t" + "addu.ph %[temp4], %[temp4], %[a32] \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "shll_s.ph %[temp3], %[temp3], 7 \n\t" + "shll_s.ph %[temp4], %[temp4], 7 \n\t" + "shll_s.ph %[temp5], %[temp5], 7 \n\t" + "precrqu_s.qb.ph %[temp2], %[temp3], %[temp2] \n\t" + "precrqu_s.qb.ph %[temp3], %[temp4], %[temp5] \n\t" + "usw %[temp1], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp3], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp2], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [a10]"=&r"(a10), [a32]"=&r"(a32) + : [c35]"r"(c35), [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void VE4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile( + "ulw %[temp0], -1(%[top]) \n\t" + "ulh %[temp1], 3(%[top]) \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "packrl.ph %[temp5], %[temp3], %[temp2] \n\t" + "packrl.ph %[temp6], %[temp4], %[temp3] \n\t" + "shll.ph %[temp5], %[temp5], 1 \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp2], %[temp5], %[temp2] \n\t" + "addq.ph %[temp6], %[temp6], %[temp4] \n\t" + "addq.ph %[temp2], %[temp2], %[temp3] \n\t" + "addq.ph %[temp6], %[temp6], %[temp3] \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precr.qb.ph %[temp4], %[temp6], %[temp2] \n\t" + "usw %[temp4], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void HE4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile( + "ulw %[temp0], -4(%[top]) \n\t" + "lbu %[temp1], -5(%[top]) \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp0] \n\t" + "replv.ph %[temp4], %[temp1] \n\t" + "packrl.ph %[temp5], %[temp3], %[temp2] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp4] \n\t" + "shll.ph %[temp5], %[temp5], 1 \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp3], %[temp3], %[temp5] \n\t" + "addq.ph %[temp3], %[temp3], %[temp2] \n\t" + "addq.ph %[temp2], %[temp2], %[temp6] \n\t" + "addq.ph %[temp2], %[temp2], %[temp4] \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "replv.qb %[temp0], %[temp3] \n\t" + "replv.qb %[temp1], %[temp2] \n\t" + "srl %[temp3], %[temp3], 16 \n\t" + "srl %[temp2], %[temp2], 16 \n\t" + "replv.qb %[temp3], %[temp3] \n\t" + "replv.qb %[temp2], %[temp2] \n\t" + "usw %[temp3], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp0], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp2], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp1], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void RD4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5; + int temp6, temp7, temp8, temp9, temp10, temp11; + __asm__ volatile( + "ulw %[temp0], -5(%[top]) \n\t" + "ulw %[temp1], -1(%[top]) \n\t" + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "preceu.ph.qbl %[temp5], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp3] \n\t" + "packrl.ph %[temp7], %[temp4], %[temp2] \n\t" + "packrl.ph %[temp8], %[temp5], %[temp4] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp9], %[temp2], %[temp6] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "addq.ph %[temp9], %[temp9], %[temp3] \n\t" + "shll.ph %[temp8], %[temp8], 1 \n\t" + "shra_r.ph %[temp9], %[temp9], 2 \n\t" + "addq.ph %[temp10], %[temp4], %[temp7] \n\t" + "addq.ph %[temp11], %[temp5], %[temp8] \n\t" + "addq.ph %[temp10], %[temp10], %[temp2] \n\t" + "addq.ph %[temp11], %[temp11], %[temp4] \n\t" + "shra_r.ph %[temp10], %[temp10], 2 \n\t" + "shra_r.ph %[temp11], %[temp11], 2 \n\t" + "lbu %[temp0], 3(%[top]) \n\t" + "lbu %[temp1], 2(%[top]) \n\t" + "lbu %[temp2], 1(%[top]) \n\t" + "sll %[temp1], %[temp1], 1 \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp0], %[temp0], %[temp2] \n\t" + "precr.qb.ph %[temp9], %[temp10], %[temp9] \n\t" + "shra_r.w %[temp0], %[temp0], 2 \n\t" + "precr.qb.ph %[temp10], %[temp11], %[temp10] \n\t" + "usw %[temp9], 3*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 1*" XSTR(BPS) "(%[dst]) \n\t" + "prepend %[temp9], %[temp11], 8 \n\t" + "prepend %[temp10], %[temp0], 8 \n\t" + "usw %[temp9], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 0*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void VR4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + "ulw %[temp0], -4(%[top]) \n\t" + "ulw %[temp1], 0(%[top]) \n\t" + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp0], %[temp0] \n\t" + "preceu.ph.qbla %[temp3], %[temp1] \n\t" + "preceu.ph.qbra %[temp1], %[temp1] \n\t" + "packrl.ph %[temp7], %[temp3], %[temp2] \n\t" + "addqh_r.ph %[temp4], %[temp1], %[temp3] \n\t" + "move %[temp6], %[temp1] \n\t" + "append %[temp1], %[temp2], 16 \n\t" + "shll.ph %[temp9], %[temp6], 1 \n\t" + "addqh_r.ph %[temp5], %[temp7], %[temp6] \n\t" + "shll.ph %[temp8], %[temp7], 1 \n\t" + "addu.ph %[temp3], %[temp7], %[temp3] \n\t" + "addu.ph %[temp1], %[temp1], %[temp6] \n\t" + "packrl.ph %[temp7], %[temp2], %[temp0] \n\t" + "addu.ph %[temp6], %[temp0], %[temp2] \n\t" + "addu.ph %[temp3], %[temp3], %[temp9] \n\t" + "addu.ph %[temp1], %[temp1], %[temp8] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "shra_r.ph %[temp3], %[temp3], 2 \n\t" + "shra_r.ph %[temp1], %[temp1], 2 \n\t" + "addu.ph %[temp6], %[temp6], %[temp7] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precrq.ph.w %[temp8], %[temp4], %[temp5] \n\t" + "append %[temp4], %[temp5], 16 \n\t" + "precrq.ph.w %[temp2], %[temp3], %[temp1] \n\t" + "append %[temp3], %[temp1], 16 \n\t" + "precr.qb.ph %[temp8], %[temp8], %[temp4] \n\t" + "precr.qb.ph %[temp3], %[temp2], %[temp3] \n\t" + "usw %[temp8], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp3], 1*" XSTR(BPS) "(%[dst]) \n\t" + "append %[temp3], %[temp6], 8 \n\t" + "srl %[temp6], %[temp6], 16 \n\t" + "append %[temp8], %[temp6], 8 \n\t" + "usw %[temp3], 3*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp8], 2*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void LD4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5; + int temp6, temp7, temp8, temp9, temp10, temp11; + __asm__ volatile( + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], 4(%[top]) \n\t" + "preceu.ph.qbl %[temp2], %[temp0] \n\t" + "preceu.ph.qbr %[temp3], %[temp0] \n\t" + "preceu.ph.qbr %[temp4], %[temp1] \n\t" + "preceu.ph.qbl %[temp5], %[temp1] \n\t" + "packrl.ph %[temp6], %[temp2], %[temp3] \n\t" + "packrl.ph %[temp7], %[temp4], %[temp2] \n\t" + "packrl.ph %[temp8], %[temp5], %[temp4] \n\t" + "shll.ph %[temp6], %[temp6], 1 \n\t" + "addq.ph %[temp9], %[temp2], %[temp6] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "addq.ph %[temp9], %[temp9], %[temp3] \n\t" + "shll.ph %[temp8], %[temp8], 1 \n\t" + "shra_r.ph %[temp9], %[temp9], 2 \n\t" + "addq.ph %[temp10], %[temp4], %[temp7] \n\t" + "addq.ph %[temp11], %[temp5], %[temp8] \n\t" + "addq.ph %[temp10], %[temp10], %[temp2] \n\t" + "addq.ph %[temp11], %[temp11], %[temp4] \n\t" + "shra_r.ph %[temp10], %[temp10], 2 \n\t" + "shra_r.ph %[temp11], %[temp11], 2 \n\t" + "srl %[temp1], %[temp1], 24 \n\t" + "sll %[temp1], %[temp1], 1 \n\t" + "raddu.w.qb %[temp5], %[temp5] \n\t" + "precr.qb.ph %[temp9], %[temp10], %[temp9] \n\t" + "precr.qb.ph %[temp10], %[temp11], %[temp10] \n\t" + "addu %[temp1], %[temp1], %[temp5] \n\t" + "shra_r.w %[temp1], %[temp1], 2 \n\t" + "usw %[temp9], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 2*" XSTR(BPS) "(%[dst]) \n\t" + "prepend %[temp9], %[temp11], 8 \n\t" + "prepend %[temp10], %[temp1], 8 \n\t" + "usw %[temp9], 1*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp10], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), [temp11]"=&r"(temp11) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void VL4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + "ulw %[temp0], 0(%[top]) \n\t" + "ulw %[temp1], 4(%[top]) \n\t" + "preceu.ph.qbla %[temp2], %[temp0] \n\t" + "preceu.ph.qbra %[temp0], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp1] \n\t" + "preceu.ph.qbr %[temp1], %[temp1] \n\t" + "addqh_r.ph %[temp4], %[temp0], %[temp2] \n\t" + "packrl.ph %[temp7], %[temp1], %[temp0] \n\t" + "precrq.ph.w %[temp6], %[temp1], %[temp2] \n\t" + "shll.ph %[temp9], %[temp2], 1 \n\t" + "addqh_r.ph %[temp5], %[temp7], %[temp2] \n\t" + "shll.ph %[temp8], %[temp7], 1 \n\t" + "addu.ph %[temp2], %[temp2], %[temp6] \n\t" + "addu.ph %[temp0], %[temp0], %[temp7] \n\t" + "packrl.ph %[temp7], %[temp3], %[temp1] \n\t" + "addu.ph %[temp6], %[temp1], %[temp3] \n\t" + "addu.ph %[temp2], %[temp2], %[temp8] \n\t" + "addu.ph %[temp0], %[temp0], %[temp9] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "addu.ph %[temp6], %[temp6], %[temp7] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precrq.ph.w %[temp8], %[temp5], %[temp4] \n\t" + "append %[temp5], %[temp4], 16 \n\t" + "precrq.ph.w %[temp3], %[temp2], %[temp0] \n\t" + "append %[temp2], %[temp0], 16 \n\t" + "precr.qb.ph %[temp8], %[temp8], %[temp5] \n\t" + "precr.qb.ph %[temp3], %[temp3], %[temp2] \n\t" + "usw %[temp8], 0*" XSTR(BPS) "(%[dst]) \n\t" + "prepend %[temp8], %[temp6], 8 \n\t" + "usw %[temp3], 1*" XSTR(BPS) "(%[dst]) \n\t" + "srl %[temp6], %[temp6], 16 \n\t" + "prepend %[temp3], %[temp6], 8 \n\t" + "usw %[temp8], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp3], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void HD4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + __asm__ volatile ( + "ulw %[temp0], -5(%[top]) \n\t" + "ulw %[temp1], -1(%[top]) \n\t" + "preceu.ph.qbla %[temp2], %[temp0] \n\t" + "preceu.ph.qbra %[temp0], %[temp0] \n\t" + "preceu.ph.qbl %[temp3], %[temp1] \n\t" + "preceu.ph.qbr %[temp1], %[temp1] \n\t" + "addqh_r.ph %[temp4], %[temp0], %[temp2] \n\t" + "packrl.ph %[temp7], %[temp1], %[temp0] \n\t" + "precrq.ph.w %[temp6], %[temp1], %[temp2] \n\t" + "shll.ph %[temp9], %[temp2], 1 \n\t" + "addqh_r.ph %[temp5], %[temp7], %[temp2] \n\t" + "shll.ph %[temp8], %[temp7], 1 \n\t" + "addu.ph %[temp2], %[temp2], %[temp6] \n\t" + "addu.ph %[temp0], %[temp0], %[temp7] \n\t" + "packrl.ph %[temp7], %[temp3], %[temp1] \n\t" + "addu.ph %[temp6], %[temp1], %[temp3] \n\t" + "addu.ph %[temp2], %[temp2], %[temp8] \n\t" + "addu.ph %[temp0], %[temp0], %[temp9] \n\t" + "shll.ph %[temp7], %[temp7], 1 \n\t" + "shra_r.ph %[temp2], %[temp2], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "addu.ph %[temp6], %[temp6], %[temp7] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "precrq.ph.w %[temp1], %[temp2], %[temp5] \n\t" + "precrq.ph.w %[temp3], %[temp0], %[temp4] \n\t" + "precr.qb.ph %[temp7], %[temp6], %[temp1] \n\t" + "precr.qb.ph %[temp6], %[temp1], %[temp3] \n\t" + "usw %[temp7], 0*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp6], 1*" XSTR(BPS) "(%[dst]) \n\t" + "append %[temp2], %[temp5], 16 \n\t" + "append %[temp0], %[temp4], 16 \n\t" + "precr.qb.ph %[temp5], %[temp3], %[temp2] \n\t" + "precr.qb.ph %[temp4], %[temp2], %[temp0] \n\t" + "usw %[temp5], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp4], 3*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +static void HU4(uint8_t* dst, const uint8_t* top) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + __asm__ volatile ( + "ulw %[temp0], -5(%[top]) \n\t" + "preceu.ph.qbl %[temp1], %[temp0] \n\t" + "preceu.ph.qbr %[temp2], %[temp0] \n\t" + "packrl.ph %[temp3], %[temp1], %[temp2] \n\t" + "replv.qb %[temp7], %[temp2] \n\t" + "addqh_r.ph %[temp4], %[temp1], %[temp3] \n\t" + "addqh_r.ph %[temp5], %[temp3], %[temp2] \n\t" + "shll.ph %[temp6], %[temp3], 1 \n\t" + "addu.ph %[temp3], %[temp2], %[temp3] \n\t" + "addu.ph %[temp6], %[temp1], %[temp6] \n\t" + "shll.ph %[temp0], %[temp2], 1 \n\t" + "addu.ph %[temp6], %[temp6], %[temp2] \n\t" + "addu.ph %[temp0], %[temp3], %[temp0] \n\t" + "shra_r.ph %[temp6], %[temp6], 2 \n\t" + "shra_r.ph %[temp0], %[temp0], 2 \n\t" + "packrl.ph %[temp3], %[temp6], %[temp5] \n\t" + "precrq.ph.w %[temp2], %[temp6], %[temp4] \n\t" + "append %[temp0], %[temp5], 16 \n\t" + "precr.qb.ph %[temp3], %[temp3], %[temp2] \n\t" + "usw %[temp3], 0*" XSTR(BPS) "(%[dst]) \n\t" + "precr.qb.ph %[temp1], %[temp7], %[temp0] \n\t" + "usw %[temp7], 3*" XSTR(BPS) "(%[dst]) \n\t" + "packrl.ph %[temp2], %[temp1], %[temp3] \n\t" + "usw %[temp1], 2*" XSTR(BPS) "(%[dst]) \n\t" + "usw %[temp2], 1*" XSTR(BPS) "(%[dst]) \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7) + : [top]"r"(top), [dst]"r"(dst) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Chroma 8x8 prediction (paragraph 12.2) + +static void IntraChromaPreds_MIPSdspR2(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + // U block + DCMode8(C8DC8 + dst, left, top); + VerticalPred8(C8VE8 + dst, top); + HorizontalPred8(C8HE8 + dst, left); + TrueMotion8(C8TM8 + dst, left, top); + // V block + dst += 8; + if (top) top += 8; + if (left) left += 16; + DCMode8(C8DC8 + dst, left, top); + VerticalPred8(C8VE8 + dst, top); + HorizontalPred8(C8HE8 + dst, left); + TrueMotion8(C8TM8 + dst, left, top); +} + +//------------------------------------------------------------------------------ +// luma 16x16 prediction (paragraph 12.3) + +static void Intra16Preds_MIPSdspR2(uint8_t* dst, + const uint8_t* left, const uint8_t* top) { + DCMode16(I16DC16 + dst, left, top); + VerticalPred16(I16VE16 + dst, top); + HorizontalPred16(I16HE16 + dst, left); + TrueMotion16(I16TM16 + dst, left, top); +} + +// Left samples are top[-5 .. -2], top_left is top[-1], top are +// located at top[0..3], and top right is top[4..7] +static void Intra4Preds_MIPSdspR2(uint8_t* dst, const uint8_t* top) { + DC4(I4DC4 + dst, top); + TM4(I4TM4 + dst, top); + VE4(I4VE4 + dst, top); + HE4(I4HE4 + dst, top); + RD4(I4RD4 + dst, top); + VR4(I4VR4 + dst, top); + LD4(I4LD4 + dst, top); + VL4(I4VL4 + dst, top); + HD4(I4HD4 + dst, top); + HU4(I4HU4 + dst, top); +} + +//------------------------------------------------------------------------------ +// Metric + +#if !defined(WORK_AROUND_GCC) + +#define GET_SSE_INNER(A) \ + "lw %[temp0], " #A "(%[a]) \n\t" \ + "lw %[temp1], " #A "(%[b]) \n\t" \ + "preceu.ph.qbr %[temp2], %[temp0] \n\t" \ + "preceu.ph.qbl %[temp0], %[temp0] \n\t" \ + "preceu.ph.qbr %[temp3], %[temp1] \n\t" \ + "preceu.ph.qbl %[temp1], %[temp1] \n\t" \ + "subq.ph %[temp2], %[temp2], %[temp3] \n\t" \ + "subq.ph %[temp0], %[temp0], %[temp1] \n\t" \ + "dpa.w.ph $ac0, %[temp2], %[temp2] \n\t" \ + "dpa.w.ph $ac0, %[temp0], %[temp0] \n\t" + +#define GET_SSE(A, B, C, D) \ + GET_SSE_INNER(A) \ + GET_SSE_INNER(B) \ + GET_SSE_INNER(C) \ + GET_SSE_INNER(D) + +static int SSE16x16_MIPSdspR2(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + GET_SSE( 8 * BPS, 4 + 8 * BPS, 8 + 8 * BPS, 12 + 8 * BPS) + GET_SSE( 9 * BPS, 4 + 9 * BPS, 8 + 9 * BPS, 12 + 9 * BPS) + GET_SSE(10 * BPS, 4 + 10 * BPS, 8 + 10 * BPS, 12 + 10 * BPS) + GET_SSE(11 * BPS, 4 + 11 * BPS, 8 + 11 * BPS, 12 + 11 * BPS) + GET_SSE(12 * BPS, 4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS) + GET_SSE(13 * BPS, 4 + 13 * BPS, 8 + 13 * BPS, 12 + 13 * BPS) + GET_SSE(14 * BPS, 4 + 14 * BPS, 8 + 14 * BPS, 12 + 14 * BPS) + GET_SSE(15 * BPS, 4 + 15 * BPS, 8 + 15 * BPS, 12 + 15 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE16x8_MIPSdspR2(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE( 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS) + GET_SSE( 1 * BPS, 4 + 1 * BPS, 8 + 1 * BPS, 12 + 1 * BPS) + GET_SSE( 2 * BPS, 4 + 2 * BPS, 8 + 2 * BPS, 12 + 2 * BPS) + GET_SSE( 3 * BPS, 4 + 3 * BPS, 8 + 3 * BPS, 12 + 3 * BPS) + GET_SSE( 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS) + GET_SSE( 5 * BPS, 4 + 5 * BPS, 8 + 5 * BPS, 12 + 5 * BPS) + GET_SSE( 6 * BPS, 4 + 6 * BPS, 8 + 6 * BPS, 12 + 6 * BPS) + GET_SSE( 7 * BPS, 4 + 7 * BPS, 8 + 7 * BPS, 12 + 7 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE8x8_MIPSdspR2(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE(0 * BPS, 4 + 0 * BPS, 1 * BPS, 4 + 1 * BPS) + GET_SSE(2 * BPS, 4 + 2 * BPS, 3 * BPS, 4 + 3 * BPS) + GET_SSE(4 * BPS, 4 + 4 * BPS, 5 * BPS, 4 + 5 * BPS) + GET_SSE(6 * BPS, 4 + 6 * BPS, 7 * BPS, 4 + 7 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +static int SSE4x4_MIPSdspR2(const uint8_t* a, const uint8_t* b) { + int count; + int temp0, temp1, temp2, temp3; + __asm__ volatile ( + "mult $zero, $zero \n\t" + GET_SSE(0 * BPS, 1 * BPS, 2 * BPS, 3 * BPS) + "mflo %[count] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [count]"=&r"(count) + : [a]"r"(a), [b]"r"(b) + : "memory", "hi", "lo" + ); + return count; +} + +#undef GET_SSE +#undef GET_SSE_INNER + +#endif // !WORK_AROUND_GCC + +#undef FILL_8_OR_16 +#undef FILL_PART +#undef OUTPUT_EARLY_CLOBBER_REGS_17 +#undef MUL_HALF +#undef ABS_X8 +#undef ADD_SUB_HALVES_X4 + +//------------------------------------------------------------------------------ +// Quantization +// + +// macro for one pass through for loop in QuantizeBlock reading 2 values at time +// QUANTDIV macro inlined +// J - offset in bytes (kZigzag[n] * 2) +// K - offset in bytes (kZigzag[n] * 4) +// N - offset in bytes (n * 2) +// N1 - offset in bytes ((n + 1) * 2) +#define QUANTIZE_ONE(J, K, N, N1) \ + "ulw %[temp1], " #J "(%[ppin]) \n\t" \ + "ulw %[temp2], " #J "(%[ppsharpen]) \n\t" \ + "lhu %[temp3], " #K "(%[ppzthresh]) \n\t" \ + "lhu %[temp6], " #K "+4(%[ppzthresh]) \n\t" \ + "absq_s.ph %[temp4], %[temp1] \n\t" \ + "ins %[temp3], %[temp6], 16, 16 \n\t" \ + "addu.ph %[coeff], %[temp4], %[temp2] \n\t" \ + "shra.ph %[sign], %[temp1], 15 \n\t" \ + "li %[level], 0x10001 \n\t" \ + "cmp.lt.ph %[temp3], %[coeff] \n\t" \ + "lhu %[temp1], " #J "(%[ppiq]) \n\t" \ + "pick.ph %[temp5], %[level], $0 \n\t" \ + "lw %[temp2], " #K "(%[ppbias]) \n\t" \ + "beqz %[temp5], 0f \n\t" \ + "lhu %[temp3], " #J "(%[ppq]) \n\t" \ + "beq %[temp5], %[level], 1f \n\t" \ + "andi %[temp5], %[temp5], 0x1 \n\t" \ + "andi %[temp4], %[coeff], 0xffff \n\t" \ + "beqz %[temp5], 2f \n\t" \ + "mul %[level], %[temp4], %[temp1] \n\t" \ + "sh $0, " #J "+2(%[ppin]) \n\t" \ + "sh $0, " #N1 "(%[pout]) \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "sra %[level], %[level], 17 \n\t" \ + "slt %[temp4], %[max_level], %[level] \n\t" \ + "movn %[level], %[max_level], %[temp4] \n\t" \ + "andi %[temp6], %[sign], 0xffff \n\t" \ + "xor %[level], %[level], %[temp6] \n\t" \ + "subu %[level], %[level], %[temp6] \n\t" \ + "mul %[temp5], %[level], %[temp3] \n\t" \ + "or %[ret], %[ret], %[level] \n\t" \ + "sh %[level], " #N "(%[pout]) \n\t" \ + "sh %[temp5], " #J "(%[ppin]) \n\t" \ + "j 3f \n\t" \ +"2: \n\t" \ + "lhu %[temp1], " #J "+2(%[ppiq]) \n\t" \ + "srl %[temp5], %[coeff], 16 \n\t" \ + "mul %[level], %[temp5], %[temp1] \n\t" \ + "lw %[temp2], " #K "+4(%[ppbias]) \n\t" \ + "lhu %[temp3], " #J "+2(%[ppq]) \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "sra %[level], %[level], 17 \n\t" \ + "srl %[temp6], %[sign], 16 \n\t" \ + "slt %[temp4], %[max_level], %[level] \n\t" \ + "movn %[level], %[max_level], %[temp4] \n\t" \ + "xor %[level], %[level], %[temp6] \n\t" \ + "subu %[level], %[level], %[temp6] \n\t" \ + "mul %[temp5], %[level], %[temp3] \n\t" \ + "sh $0, " #J "(%[ppin]) \n\t" \ + "sh $0, " #N "(%[pout]) \n\t" \ + "or %[ret], %[ret], %[level] \n\t" \ + "sh %[temp5], " #J "+2(%[ppin]) \n\t" \ + "sh %[level], " #N1 "(%[pout]) \n\t" \ + "j 3f \n\t" \ +"1: \n\t" \ + "lhu %[temp1], " #J "(%[ppiq]) \n\t" \ + "lw %[temp2], " #K "(%[ppbias]) \n\t" \ + "ulw %[temp3], " #J "(%[ppq]) \n\t" \ + "andi %[temp5], %[coeff], 0xffff \n\t" \ + "srl %[temp0], %[coeff], 16 \n\t" \ + "lhu %[temp6], " #J "+2(%[ppiq]) \n\t" \ + "lw %[coeff], " #K "+4(%[ppbias]) \n\t" \ + "mul %[level], %[temp5], %[temp1] \n\t" \ + "mul %[temp4], %[temp0], %[temp6] \n\t" \ + "addu %[level], %[level], %[temp2] \n\t" \ + "addu %[temp4], %[temp4], %[coeff] \n\t" \ + "precrq.ph.w %[level], %[temp4], %[level] \n\t" \ + "shra.ph %[level], %[level], 1 \n\t" \ + "cmp.lt.ph %[max_level1],%[level] \n\t" \ + "pick.ph %[level], %[max_level], %[level] \n\t" \ + "xor %[level], %[level], %[sign] \n\t" \ + "subu.ph %[level], %[level], %[sign] \n\t" \ + "mul.ph %[temp3], %[level], %[temp3] \n\t" \ + "or %[ret], %[ret], %[level] \n\t" \ + "sh %[level], " #N "(%[pout]) \n\t" \ + "srl %[level], %[level], 16 \n\t" \ + "sh %[level], " #N1 "(%[pout]) \n\t" \ + "usw %[temp3], " #J "(%[ppin]) \n\t" \ + "j 3f \n\t" \ +"0: \n\t" \ + "sh $0, " #N "(%[pout]) \n\t" \ + "sh $0, " #N1 "(%[pout]) \n\t" \ + "usw $0, " #J "(%[ppin]) \n\t" \ +"3: \n\t" + +static int QuantizeBlock_MIPSdspR2(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + int temp0, temp1, temp2, temp3, temp4, temp5,temp6; + int sign, coeff, level; + int max_level = MAX_LEVEL; + int max_level1 = max_level << 16 | max_level; + int ret = 0; + + int16_t* ppin = &in[0]; + int16_t* pout = &out[0]; + const uint16_t* ppsharpen = &mtx->sharpen_[0]; + const uint32_t* ppzthresh = &mtx->zthresh_[0]; + const uint16_t* ppq = &mtx->q_[0]; + const uint16_t* ppiq = &mtx->iq_[0]; + const uint32_t* ppbias = &mtx->bias_[0]; + + __asm__ volatile ( + QUANTIZE_ONE( 0, 0, 0, 2) + QUANTIZE_ONE( 4, 8, 10, 12) + QUANTIZE_ONE( 8, 16, 4, 8) + QUANTIZE_ONE(12, 24, 14, 24) + QUANTIZE_ONE(16, 32, 6, 16) + QUANTIZE_ONE(20, 40, 22, 26) + QUANTIZE_ONE(24, 48, 18, 20) + QUANTIZE_ONE(28, 56, 28, 30) + + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [sign]"=&r"(sign), [coeff]"=&r"(coeff), + [level]"=&r"(level), [temp6]"=&r"(temp6), [ret]"+&r"(ret) + : [ppin]"r"(ppin), [pout]"r"(pout), [max_level1]"r"(max_level1), + [ppiq]"r"(ppiq), [max_level]"r"(max_level), + [ppbias]"r"(ppbias), [ppzthresh]"r"(ppzthresh), + [ppsharpen]"r"(ppsharpen), [ppq]"r"(ppq) + : "memory", "hi", "lo" + ); + + return (ret != 0); +} + +static int Quantize2Blocks_MIPSdspR2(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = QuantizeBlock_MIPSdspR2(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= QuantizeBlock_MIPSdspR2(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +#undef QUANTIZE_ONE + +// macro for one horizontal pass in FTransformWHT +// temp0..temp7 holds tmp[0]..tmp[15] +// A, B, C, D - offset in bytes to load from in buffer +// TEMP0, TEMP1 - registers for corresponding tmp elements +#define HORIZONTAL_PASS_WHT(A, B, C, D, TEMP0, TEMP1) \ + "lh %[" #TEMP0 "], " #A "(%[in]) \n\t" \ + "lh %[" #TEMP1 "], " #B "(%[in]) \n\t" \ + "lh %[temp8], " #C "(%[in]) \n\t" \ + "lh %[temp9], " #D "(%[in]) \n\t" \ + "ins %[" #TEMP1 "], %[" #TEMP0 "], 16, 16 \n\t" \ + "ins %[temp9], %[temp8], 16, 16 \n\t" \ + "subq.ph %[temp8], %[" #TEMP1 "], %[temp9] \n\t" \ + "addq.ph %[temp9], %[" #TEMP1 "], %[temp9] \n\t" \ + "precrq.ph.w %[" #TEMP0 "], %[temp8], %[temp9] \n\t" \ + "append %[temp8], %[temp9], 16 \n\t" \ + "subq.ph %[" #TEMP1 "], %[" #TEMP0 "], %[temp8] \n\t" \ + "addq.ph %[" #TEMP0 "], %[" #TEMP0 "], %[temp8] \n\t" \ + "rotr %[" #TEMP1 "], %[" #TEMP1 "], 16 \n\t" + +// macro for one vertical pass in FTransformWHT +// temp0..temp7 holds tmp[0]..tmp[15] +// A, B, C, D - offsets in bytes to store to out buffer +// TEMP0, TEMP2, TEMP4 and TEMP6 - registers for corresponding tmp elements +#define VERTICAL_PASS_WHT(A, B, C, D, TEMP0, TEMP2, TEMP4, TEMP6) \ + "addq.ph %[temp8], %[" #TEMP0 "], %[" #TEMP4 "] \n\t" \ + "addq.ph %[temp9], %[" #TEMP2 "], %[" #TEMP6 "] \n\t" \ + "subq.ph %[" #TEMP2 "], %[" #TEMP2 "], %[" #TEMP6 "] \n\t" \ + "subq.ph %[" #TEMP6 "], %[" #TEMP0 "], %[" #TEMP4 "] \n\t" \ + "addqh.ph %[" #TEMP0 "], %[temp8], %[temp9] \n\t" \ + "subqh.ph %[" #TEMP4 "], %[" #TEMP6 "], %[" #TEMP2 "] \n\t" \ + "addqh.ph %[" #TEMP2 "], %[" #TEMP2 "], %[" #TEMP6 "] \n\t" \ + "subqh.ph %[" #TEMP6 "], %[temp8], %[temp9] \n\t" \ + "usw %[" #TEMP0 "], " #A "(%[out]) \n\t" \ + "usw %[" #TEMP2 "], " #B "(%[out]) \n\t" \ + "usw %[" #TEMP4 "], " #C "(%[out]) \n\t" \ + "usw %[" #TEMP6 "], " #D "(%[out]) \n\t" + +static void FTransformWHT_MIPSdspR2(const int16_t* in, int16_t* out) { + int temp0, temp1, temp2, temp3, temp4; + int temp5, temp6, temp7, temp8, temp9; + + __asm__ volatile ( + HORIZONTAL_PASS_WHT( 0, 32, 64, 96, temp0, temp1) + HORIZONTAL_PASS_WHT(128, 160, 192, 224, temp2, temp3) + HORIZONTAL_PASS_WHT(256, 288, 320, 352, temp4, temp5) + HORIZONTAL_PASS_WHT(384, 416, 448, 480, temp6, temp7) + VERTICAL_PASS_WHT(0, 8, 16, 24, temp0, temp2, temp4, temp6) + VERTICAL_PASS_WHT(4, 12, 20, 28, temp1, temp3, temp5, temp7) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), + [temp9]"=&r"(temp9) + : [in]"r"(in), [out]"r"(out) + : "memory" + ); +} + +#undef VERTICAL_PASS_WHT +#undef HORIZONTAL_PASS_WHT + +// macro for converting coefficients to bin +// convert 8 coeffs at time +// A, B, C, D - offsets in bytes to load from out buffer +#define CONVERT_COEFFS_TO_BIN(A, B, C, D) \ + "ulw %[temp0], " #A "(%[out]) \n\t" \ + "ulw %[temp1], " #B "(%[out]) \n\t" \ + "ulw %[temp2], " #C "(%[out]) \n\t" \ + "ulw %[temp3], " #D "(%[out]) \n\t" \ + "absq_s.ph %[temp0], %[temp0] \n\t" \ + "absq_s.ph %[temp1], %[temp1] \n\t" \ + "absq_s.ph %[temp2], %[temp2] \n\t" \ + "absq_s.ph %[temp3], %[temp3] \n\t" \ + "shra.ph %[temp0], %[temp0], 3 \n\t" \ + "shra.ph %[temp1], %[temp1], 3 \n\t" \ + "shra.ph %[temp2], %[temp2], 3 \n\t" \ + "shra.ph %[temp3], %[temp3], 3 \n\t" \ + "shll_s.ph %[temp0], %[temp0], 10 \n\t" \ + "shll_s.ph %[temp1], %[temp1], 10 \n\t" \ + "shll_s.ph %[temp2], %[temp2], 10 \n\t" \ + "shll_s.ph %[temp3], %[temp3], 10 \n\t" \ + "shrl.ph %[temp0], %[temp0], 10 \n\t" \ + "shrl.ph %[temp1], %[temp1], 10 \n\t" \ + "shrl.ph %[temp2], %[temp2], 10 \n\t" \ + "shrl.ph %[temp3], %[temp3], 10 \n\t" \ + "shll.ph %[temp0], %[temp0], 2 \n\t" \ + "shll.ph %[temp1], %[temp1], 2 \n\t" \ + "shll.ph %[temp2], %[temp2], 2 \n\t" \ + "shll.ph %[temp3], %[temp3], 2 \n\t" \ + "ext %[temp4], %[temp0], 0, 16 \n\t" \ + "ext %[temp0], %[temp0], 16, 16 \n\t" \ + "addu %[temp4], %[temp4], %[dist] \n\t" \ + "addu %[temp0], %[temp0], %[dist] \n\t" \ + "ext %[temp5], %[temp1], 0, 16 \n\t" \ + "lw %[temp8], 0(%[temp4]) \n\t" \ + "ext %[temp1], %[temp1], 16, 16 \n\t" \ + "addu %[temp5], %[temp5], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp4]) \n\t" \ + "lw %[temp8], 0(%[temp0]) \n\t" \ + "addu %[temp1], %[temp1], %[dist] \n\t" \ + "ext %[temp6], %[temp2], 0, 16 \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp0]) \n\t" \ + "lw %[temp8], 0(%[temp5]) \n\t" \ + "ext %[temp2], %[temp2], 16, 16 \n\t" \ + "addu %[temp6], %[temp6], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp5]) \n\t" \ + "lw %[temp8], 0(%[temp1]) \n\t" \ + "addu %[temp2], %[temp2], %[dist] \n\t" \ + "ext %[temp7], %[temp3], 0, 16 \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp1]) \n\t" \ + "lw %[temp8], 0(%[temp6]) \n\t" \ + "ext %[temp3], %[temp3], 16, 16 \n\t" \ + "addu %[temp7], %[temp7], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp6]) \n\t" \ + "lw %[temp8], 0(%[temp2]) \n\t" \ + "addu %[temp3], %[temp3], %[dist] \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp2]) \n\t" \ + "lw %[temp8], 0(%[temp7]) \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp7]) \n\t" \ + "lw %[temp8], 0(%[temp3]) \n\t" \ + "addiu %[temp8], %[temp8], 1 \n\t" \ + "sw %[temp8], 0(%[temp3]) \n\t" + +static void CollectHistogram_MIPSdspR2(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + const int max_coeff = (MAX_COEFF_THRESH << 16) + MAX_COEFF_THRESH; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7, temp8; + + VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + + // Convert coefficients to bin. + __asm__ volatile ( + CONVERT_COEFFS_TO_BIN( 0, 4, 8, 12) + CONVERT_COEFFS_TO_BIN(16, 20, 24, 28) + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [temp8]"=&r"(temp8) + : [dist]"r"(distribution), [out]"r"(out), [max_coeff]"r"(max_coeff) + : "memory" + ); + } + VP8SetHistogramData(distribution, histo); +} + +#undef CONVERT_COEFFS_TO_BIN + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitMIPSdspR2(void) { + VP8FTransform = FTransform_MIPSdspR2; + VP8FTransformWHT = FTransformWHT_MIPSdspR2; + VP8ITransform = ITransform_MIPSdspR2; + + VP8TDisto4x4 = Disto4x4_MIPSdspR2; + VP8TDisto16x16 = Disto16x16_MIPSdspR2; + + VP8EncPredLuma16 = Intra16Preds_MIPSdspR2; + VP8EncPredChroma8 = IntraChromaPreds_MIPSdspR2; + VP8EncPredLuma4 = Intra4Preds_MIPSdspR2; + +#if !defined(WORK_AROUND_GCC) + VP8SSE16x16 = SSE16x16_MIPSdspR2; + VP8SSE8x8 = SSE8x8_MIPSdspR2; + VP8SSE16x8 = SSE16x8_MIPSdspR2; + VP8SSE4x4 = SSE4x4_MIPSdspR2; +#endif + + VP8EncQuantizeBlock = QuantizeBlock_MIPSdspR2; + VP8EncQuantize2Blocks = Quantize2Blocks_MIPSdspR2; + + VP8CollectHistogram = CollectHistogram_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8EncDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/enc_msa.c b/media/libwebp/dsp/enc_msa.c new file mode 100644 index 0000000000..229582e4a6 --- /dev/null +++ b/media/libwebp/dsp/enc_msa.c @@ -0,0 +1,896 @@ +// Copyright 2016 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MSA version of encoder dsp functions. +// +// Author: Prashant Patil (prashant.patil@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) + +#include <stdlib.h> +#include "../dsp/msa_macro.h" +#include "../enc/vp8i_enc.h" + +//------------------------------------------------------------------------------ +// Transforms + +#define IDCT_1D_W(in0, in1, in2, in3, out0, out1, out2, out3) do { \ + v4i32 a1_m, b1_m, c1_m, d1_m; \ + const v4i32 cospi8sqrt2minus1 = __msa_fill_w(20091); \ + const v4i32 sinpi8sqrt2 = __msa_fill_w(35468); \ + v4i32 c_tmp1_m = in1 * sinpi8sqrt2; \ + v4i32 c_tmp2_m = in3 * cospi8sqrt2minus1; \ + v4i32 d_tmp1_m = in1 * cospi8sqrt2minus1; \ + v4i32 d_tmp2_m = in3 * sinpi8sqrt2; \ + \ + ADDSUB2(in0, in2, a1_m, b1_m); \ + SRAI_W2_SW(c_tmp1_m, c_tmp2_m, 16); \ + c_tmp2_m = c_tmp2_m + in3; \ + c1_m = c_tmp1_m - c_tmp2_m; \ + SRAI_W2_SW(d_tmp1_m, d_tmp2_m, 16); \ + d_tmp1_m = d_tmp1_m + in1; \ + d1_m = d_tmp1_m + d_tmp2_m; \ + BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3); \ +} while (0) + +static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in, + uint8_t* dst) { + v8i16 input0, input1; + v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3; + v4i32 res0, res1, res2, res3; + v16i8 dest0, dest1, dest2, dest3; + const v16i8 zero = { 0 }; + + LD_SH2(in, 8, input0, input1); + UNPCK_SH_SW(input0, in0, in1); + UNPCK_SH_SW(input1, in2, in3); + IDCT_1D_W(in0, in1, in2, in3, hz0, hz1, hz2, hz3); + TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3); + IDCT_1D_W(hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3); + SRARI_W4_SW(vt0, vt1, vt2, vt3, 3); + TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3); + LD_SB4(ref, BPS, dest0, dest1, dest2, dest3); + ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3, + res0, res1, res2, res3); + ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3, + res0, res1, res2, res3); + ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3); + CLIP_SW4_0_255(res0, res1, res2, res3); + PCKEV_B2_SW(res0, res1, res2, res3, vt0, vt1); + res0 = (v4i32)__msa_pckev_b((v16i8)vt0, (v16i8)vt1); + ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS); +} + +static void ITransform_MSA(const uint8_t* ref, const int16_t* in, uint8_t* dst, + int do_two) { + ITransformOne(ref, in, dst); + if (do_two) { + ITransformOne(ref + 4, in + 16, dst + 4); + } +} + +static void FTransform_MSA(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + uint64_t out0, out1, out2, out3; + uint32_t in0, in1, in2, in3; + v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5; + v8i16 t0, t1, t2, t3; + v16u8 srcl0, srcl1, src0 = { 0 }, src1 = { 0 }; + const v8i16 mask0 = { 0, 4, 8, 12, 1, 5, 9, 13 }; + const v8i16 mask1 = { 3, 7, 11, 15, 2, 6, 10, 14 }; + const v8i16 mask2 = { 4, 0, 5, 1, 6, 2, 7, 3 }; + const v8i16 mask3 = { 0, 4, 1, 5, 2, 6, 3, 7 }; + const v8i16 cnst0 = { 2217, -5352, 2217, -5352, 2217, -5352, 2217, -5352 }; + const v8i16 cnst1 = { 5352, 2217, 5352, 2217, 5352, 2217, 5352, 2217 }; + + LW4(src, BPS, in0, in1, in2, in3); + INSERT_W4_UB(in0, in1, in2, in3, src0); + LW4(ref, BPS, in0, in1, in2, in3); + INSERT_W4_UB(in0, in1, in2, in3, src1); + ILVRL_B2_UB(src0, src1, srcl0, srcl1); + HSUB_UB2_SH(srcl0, srcl1, t0, t1); + VSHF_H2_SH(t0, t1, t0, t1, mask0, mask1, t2, t3); + ADDSUB2(t2, t3, t0, t1); + t0 = SRLI_H(t0, 3); + VSHF_H2_SH(t0, t0, t1, t1, mask2, mask3, t3, t2); + tmp0 = __msa_hadd_s_w(t3, t3); + tmp2 = __msa_hsub_s_w(t3, t3); + FILL_W2_SW(1812, 937, tmp1, tmp3); + DPADD_SH2_SW(t2, t2, cnst0, cnst1, tmp3, tmp1); + SRAI_W2_SW(tmp1, tmp3, 9); + PCKEV_H2_SH(tmp1, tmp0, tmp3, tmp2, t0, t1); + VSHF_H2_SH(t0, t1, t0, t1, mask0, mask1, t2, t3); + ADDSUB2(t2, t3, t0, t1); + VSHF_H2_SH(t0, t0, t1, t1, mask2, mask3, t3, t2); + tmp0 = __msa_hadd_s_w(t3, t3); + tmp2 = __msa_hsub_s_w(t3, t3); + ADDVI_W2_SW(tmp0, 7, tmp2, 7, tmp0, tmp2); + SRAI_W2_SW(tmp0, tmp2, 4); + FILL_W2_SW(12000, 51000, tmp1, tmp3); + DPADD_SH2_SW(t2, t2, cnst0, cnst1, tmp3, tmp1); + SRAI_W2_SW(tmp1, tmp3, 16); + UNPCK_R_SH_SW(t1, tmp4); + tmp5 = __msa_ceqi_w(tmp4, 0); + tmp4 = (v4i32)__msa_nor_v((v16u8)tmp5, (v16u8)tmp5); + tmp5 = __msa_fill_w(1); + tmp5 = (v4i32)__msa_and_v((v16u8)tmp5, (v16u8)tmp4); + tmp1 += tmp5; + PCKEV_H2_SH(tmp1, tmp0, tmp3, tmp2, t0, t1); + out0 = __msa_copy_s_d((v2i64)t0, 0); + out1 = __msa_copy_s_d((v2i64)t0, 1); + out2 = __msa_copy_s_d((v2i64)t1, 0); + out3 = __msa_copy_s_d((v2i64)t1, 1); + SD4(out0, out1, out2, out3, out, 8); +} + +static void FTransformWHT_MSA(const int16_t* in, int16_t* out) { + v8i16 in0 = { 0 }; + v8i16 in1 = { 0 }; + v8i16 tmp0, tmp1, tmp2, tmp3; + v8i16 out0, out1; + const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 }; + const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 }; + const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 }; + const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 }; + + in0 = __msa_insert_h(in0, 0, in[ 0]); + in0 = __msa_insert_h(in0, 1, in[ 64]); + in0 = __msa_insert_h(in0, 2, in[128]); + in0 = __msa_insert_h(in0, 3, in[192]); + in0 = __msa_insert_h(in0, 4, in[ 16]); + in0 = __msa_insert_h(in0, 5, in[ 80]); + in0 = __msa_insert_h(in0, 6, in[144]); + in0 = __msa_insert_h(in0, 7, in[208]); + in1 = __msa_insert_h(in1, 0, in[ 48]); + in1 = __msa_insert_h(in1, 1, in[112]); + in1 = __msa_insert_h(in1, 2, in[176]); + in1 = __msa_insert_h(in1, 3, in[240]); + in1 = __msa_insert_h(in1, 4, in[ 32]); + in1 = __msa_insert_h(in1, 5, in[ 96]); + in1 = __msa_insert_h(in1, 6, in[160]); + in1 = __msa_insert_h(in1, 7, in[224]); + ADDSUB2(in0, in1, tmp0, tmp1); + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); + ADDSUB2(tmp2, tmp3, tmp0, tmp1); + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1); + ADDSUB2(in0, in1, tmp0, tmp1); + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); + ADDSUB2(tmp2, tmp3, out0, out1); + SRAI_H2_SH(out0, out1, 1); + ST_SH2(out0, out1, out, 8); +} + +static int TTransform_MSA(const uint8_t* in, const uint16_t* w) { + int sum; + uint32_t in0_m, in1_m, in2_m, in3_m; + v16i8 src0 = { 0 }; + v8i16 in0, in1, tmp0, tmp1, tmp2, tmp3; + v4i32 dst0, dst1; + const v16i8 zero = { 0 }; + const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 }; + const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 }; + const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 }; + const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 }; + + LW4(in, BPS, in0_m, in1_m, in2_m, in3_m); + INSERT_W4_SB(in0_m, in1_m, in2_m, in3_m, src0); + ILVRL_B2_SH(zero, src0, tmp0, tmp1); + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1); + ADDSUB2(in0, in1, tmp0, tmp1); + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); + ADDSUB2(tmp2, tmp3, tmp0, tmp1); + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask2, mask3, in0, in1); + ADDSUB2(in0, in1, tmp0, tmp1); + VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3); + ADDSUB2(tmp2, tmp3, tmp0, tmp1); + tmp0 = __msa_add_a_h(tmp0, (v8i16)zero); + tmp1 = __msa_add_a_h(tmp1, (v8i16)zero); + LD_SH2(w, 8, tmp2, tmp3); + DOTP_SH2_SW(tmp0, tmp1, tmp2, tmp3, dst0, dst1); + dst0 = dst0 + dst1; + sum = HADD_SW_S32(dst0); + return sum; +} + +static int Disto4x4_MSA(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + const int sum1 = TTransform_MSA(a, w); + const int sum2 = TTransform_MSA(b, w); + return abs(sum2 - sum1) >> 5; +} + +static int Disto16x16_MSA(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4_MSA(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ +// Histogram + +static void CollectHistogram_MSA(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + { + int k; + v8i16 coeff0, coeff1; + const v8i16 zero = { 0 }; + const v8i16 max_coeff_thr = __msa_ldi_h(MAX_COEFF_THRESH); + LD_SH2(&out[0], 8, coeff0, coeff1); + coeff0 = __msa_add_a_h(coeff0, zero); + coeff1 = __msa_add_a_h(coeff1, zero); + SRAI_H2_SH(coeff0, coeff1, 3); + coeff0 = __msa_min_s_h(coeff0, max_coeff_thr); + coeff1 = __msa_min_s_h(coeff1, max_coeff_thr); + ST_SH2(coeff0, coeff1, &out[0], 8); + for (k = 0; k < 16; ++k) { + ++distribution[out[k]]; + } + } + } + VP8SetHistogramData(distribution, histo); +} + +//------------------------------------------------------------------------------ +// Intra predictions + +// luma 4x4 prediction + +#define DST(x, y) dst[(x) + (y) * BPS] +#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2) +#define AVG2(a, b) (((a) + (b) + 1) >> 1) + +static WEBP_INLINE void VE4(uint8_t* dst, const uint8_t* top) { // vertical + const v16u8 A1 = { 0 }; + const uint64_t val_m = LD(top - 1); + const v16u8 A = (v16u8)__msa_insert_d((v2i64)A1, 0, val_m); + const v16u8 B = SLDI_UB(A, A, 1); + const v16u8 C = SLDI_UB(A, A, 2); + const v16u8 AC = __msa_ave_u_b(A, C); + const v16u8 B2 = __msa_ave_u_b(B, B); + const v16u8 R = __msa_aver_u_b(AC, B2); + const uint32_t out = __msa_copy_s_w((v4i32)R, 0); + SW4(out, out, out, out, dst, BPS); +} + +static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) { // horizontal + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J)); + WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K)); + WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L)); + WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L)); +} + +static WEBP_INLINE void DC4(uint8_t* dst, const uint8_t* top) { + uint32_t dc = 4; + int i; + for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i]; + dc >>= 3; + dc = dc | (dc << 8) | (dc << 16) | (dc << 24); + SW4(dc, dc, dc, dc, dst, BPS); +} + +static WEBP_INLINE void RD4(uint8_t* dst, const uint8_t* top) { + const v16u8 A2 = { 0 }; + const uint64_t val_m = LD(top - 5); + const v16u8 A1 = (v16u8)__msa_insert_d((v2i64)A2, 0, val_m); + const v16u8 A = (v16u8)__msa_insert_b((v16i8)A1, 8, top[3]); + const v16u8 B = SLDI_UB(A, A, 1); + const v16u8 C = SLDI_UB(A, A, 2); + const v16u8 AC = __msa_ave_u_b(A, C); + const v16u8 B2 = __msa_ave_u_b(B, B); + const v16u8 R0 = __msa_aver_u_b(AC, B2); + const v16u8 R1 = SLDI_UB(R0, R0, 1); + const v16u8 R2 = SLDI_UB(R1, R1, 1); + const v16u8 R3 = SLDI_UB(R2, R2, 1); + const uint32_t val0 = __msa_copy_s_w((v4i32)R0, 0); + const uint32_t val1 = __msa_copy_s_w((v4i32)R1, 0); + const uint32_t val2 = __msa_copy_s_w((v4i32)R2, 0); + const uint32_t val3 = __msa_copy_s_w((v4i32)R3, 0); + SW4(val3, val2, val1, val0, dst, BPS); +} + +static WEBP_INLINE void LD4(uint8_t* dst, const uint8_t* top) { + const v16u8 A1 = { 0 }; + const uint64_t val_m = LD(top); + const v16u8 A = (v16u8)__msa_insert_d((v2i64)A1, 0, val_m); + const v16u8 B = SLDI_UB(A, A, 1); + const v16u8 C1 = SLDI_UB(A, A, 2); + const v16u8 C = (v16u8)__msa_insert_b((v16i8)C1, 6, top[7]); + const v16u8 AC = __msa_ave_u_b(A, C); + const v16u8 B2 = __msa_ave_u_b(B, B); + const v16u8 R0 = __msa_aver_u_b(AC, B2); + const v16u8 R1 = SLDI_UB(R0, R0, 1); + const v16u8 R2 = SLDI_UB(R1, R1, 1); + const v16u8 R3 = SLDI_UB(R2, R2, 1); + const uint32_t val0 = __msa_copy_s_w((v4i32)R0, 0); + const uint32_t val1 = __msa_copy_s_w((v4i32)R1, 0); + const uint32_t val2 = __msa_copy_s_w((v4i32)R2, 0); + const uint32_t val3 = __msa_copy_s_w((v4i32)R3, 0); + SW4(val0, val1, val2, val3, dst, BPS); +} + +static WEBP_INLINE void VR4(uint8_t* dst, const uint8_t* top) { + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + const int D = top[3]; + DST(0, 0) = DST(1, 2) = AVG2(X, A); + DST(1, 0) = DST(2, 2) = AVG2(A, B); + DST(2, 0) = DST(3, 2) = AVG2(B, C); + DST(3, 0) = AVG2(C, D); + DST(0, 3) = AVG3(K, J, I); + DST(0, 2) = AVG3(J, I, X); + DST(0, 1) = DST(1, 3) = AVG3(I, X, A); + DST(1, 1) = DST(2, 3) = AVG3(X, A, B); + DST(2, 1) = DST(3, 3) = AVG3(A, B, C); + DST(3, 1) = AVG3(B, C, D); +} + +static WEBP_INLINE void VL4(uint8_t* dst, const uint8_t* top) { + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + const int D = top[3]; + const int E = top[4]; + const int F = top[5]; + const int G = top[6]; + const int H = top[7]; + DST(0, 0) = AVG2(A, B); + DST(1, 0) = DST(0, 2) = AVG2(B, C); + DST(2, 0) = DST(1, 2) = AVG2(C, D); + DST(3, 0) = DST(2, 2) = AVG2(D, E); + DST(0, 1) = AVG3(A, B, C); + DST(1, 1) = DST(0, 3) = AVG3(B, C, D); + DST(2, 1) = DST(1, 3) = AVG3(C, D, E); + DST(3, 1) = DST(2, 3) = AVG3(D, E, F); + DST(3, 2) = AVG3(E, F, G); + DST(3, 3) = AVG3(F, G, H); +} + +static WEBP_INLINE void HU4(uint8_t* dst, const uint8_t* top) { + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + DST(0, 0) = AVG2(I, J); + DST(2, 0) = DST(0, 1) = AVG2(J, K); + DST(2, 1) = DST(0, 2) = AVG2(K, L); + DST(1, 0) = AVG3(I, J, K); + DST(3, 0) = DST(1, 1) = AVG3(J, K, L); + DST(3, 1) = DST(1, 2) = AVG3(K, L, L); + DST(3, 2) = DST(2, 2) = + DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L; +} + +static WEBP_INLINE void HD4(uint8_t* dst, const uint8_t* top) { + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + DST(0, 0) = DST(2, 1) = AVG2(I, X); + DST(0, 1) = DST(2, 2) = AVG2(J, I); + DST(0, 2) = DST(2, 3) = AVG2(K, J); + DST(0, 3) = AVG2(L, K); + DST(3, 0) = AVG3(A, B, C); + DST(2, 0) = AVG3(X, A, B); + DST(1, 0) = DST(3, 1) = AVG3(I, X, A); + DST(1, 1) = DST(3, 2) = AVG3(J, I, X); + DST(1, 2) = DST(3, 3) = AVG3(K, J, I); + DST(1, 3) = AVG3(L, K, J); +} + +static WEBP_INLINE void TM4(uint8_t* dst, const uint8_t* top) { + const v16i8 zero = { 0 }; + const v8i16 TL = (v8i16)__msa_fill_h(top[-1]); + const v8i16 L0 = (v8i16)__msa_fill_h(top[-2]); + const v8i16 L1 = (v8i16)__msa_fill_h(top[-3]); + const v8i16 L2 = (v8i16)__msa_fill_h(top[-4]); + const v8i16 L3 = (v8i16)__msa_fill_h(top[-5]); + const v16u8 T1 = LD_UB(top); + const v8i16 T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1); + const v8i16 d = T - TL; + v8i16 r0, r1, r2, r3; + ADD4(d, L0, d, L1, d, L2, d, L3, r0, r1, r2, r3); + CLIP_SH4_0_255(r0, r1, r2, r3); + PCKEV_ST4x4_UB(r0, r1, r2, r3, dst, BPS); +} + +#undef DST +#undef AVG3 +#undef AVG2 + +static void Intra4Preds_MSA(uint8_t* dst, const uint8_t* top) { + DC4(I4DC4 + dst, top); + TM4(I4TM4 + dst, top); + VE4(I4VE4 + dst, top); + HE4(I4HE4 + dst, top); + RD4(I4RD4 + dst, top); + VR4(I4VR4 + dst, top); + LD4(I4LD4 + dst, top); + VL4(I4VL4 + dst, top); + HD4(I4HD4 + dst, top); + HU4(I4HU4 + dst, top); +} + +// luma 16x16 prediction + +#define STORE16x16(out, dst) do { \ + ST_UB8(out, out, out, out, out, out, out, out, dst + 0 * BPS, BPS); \ + ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS); \ +} while (0) + +static WEBP_INLINE void VerticalPred16x16(uint8_t* dst, const uint8_t* top) { + if (top != NULL) { + const v16u8 out = LD_UB(top); + STORE16x16(out, dst); + } else { + const v16u8 out = (v16u8)__msa_fill_b(0x7f); + STORE16x16(out, dst); + } +} + +static WEBP_INLINE void HorizontalPred16x16(uint8_t* dst, + const uint8_t* left) { + if (left != NULL) { + int j; + for (j = 0; j < 16; j += 4) { + const v16u8 L0 = (v16u8)__msa_fill_b(left[0]); + const v16u8 L1 = (v16u8)__msa_fill_b(left[1]); + const v16u8 L2 = (v16u8)__msa_fill_b(left[2]); + const v16u8 L3 = (v16u8)__msa_fill_b(left[3]); + ST_UB4(L0, L1, L2, L3, dst, BPS); + dst += 4 * BPS; + left += 4; + } + } else { + const v16u8 out = (v16u8)__msa_fill_b(0x81); + STORE16x16(out, dst); + } +} + +static WEBP_INLINE void TrueMotion16x16(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + if (left != NULL) { + if (top != NULL) { + int j; + v8i16 d1, d2; + const v16i8 zero = { 0 }; + const v8i16 TL = (v8i16)__msa_fill_h(left[-1]); + const v16u8 T = LD_UB(top); + ILVRL_B2_SH(zero, T, d1, d2); + SUB2(d1, TL, d2, TL, d1, d2); + for (j = 0; j < 16; j += 4) { + v16i8 t0, t1, t2, t3; + v8i16 r0, r1, r2, r3, r4, r5, r6, r7; + const v8i16 L0 = (v8i16)__msa_fill_h(left[j + 0]); + const v8i16 L1 = (v8i16)__msa_fill_h(left[j + 1]); + const v8i16 L2 = (v8i16)__msa_fill_h(left[j + 2]); + const v8i16 L3 = (v8i16)__msa_fill_h(left[j + 3]); + ADD4(d1, L0, d1, L1, d1, L2, d1, L3, r0, r1, r2, r3); + ADD4(d2, L0, d2, L1, d2, L2, d2, L3, r4, r5, r6, r7); + CLIP_SH4_0_255(r0, r1, r2, r3); + CLIP_SH4_0_255(r4, r5, r6, r7); + PCKEV_B4_SB(r4, r0, r5, r1, r6, r2, r7, r3, t0, t1, t2, t3); + ST_SB4(t0, t1, t2, t3, dst, BPS); + dst += 4 * BPS; + } + } else { + HorizontalPred16x16(dst, left); + } + } else { + if (top != NULL) { + VerticalPred16x16(dst, top); + } else { + const v16u8 out = (v16u8)__msa_fill_b(0x81); + STORE16x16(out, dst); + } + } +} + +static WEBP_INLINE void DCMode16x16(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + int DC; + v16u8 out; + if (top != NULL && left != NULL) { + const v16u8 rtop = LD_UB(top); + const v8u16 dctop = __msa_hadd_u_h(rtop, rtop); + const v16u8 rleft = LD_UB(left); + const v8u16 dcleft = __msa_hadd_u_h(rleft, rleft); + const v8u16 dctemp = dctop + dcleft; + DC = HADD_UH_U32(dctemp); + DC = (DC + 16) >> 5; + } else if (left != NULL) { // left but no top + const v16u8 rleft = LD_UB(left); + const v8u16 dcleft = __msa_hadd_u_h(rleft, rleft); + DC = HADD_UH_U32(dcleft); + DC = (DC + DC + 16) >> 5; + } else if (top != NULL) { // top but no left + const v16u8 rtop = LD_UB(top); + const v8u16 dctop = __msa_hadd_u_h(rtop, rtop); + DC = HADD_UH_U32(dctop); + DC = (DC + DC + 16) >> 5; + } else { // no top, no left, nothing. + DC = 0x80; + } + out = (v16u8)__msa_fill_b(DC); + STORE16x16(out, dst); +} + +static void Intra16Preds_MSA(uint8_t* dst, + const uint8_t* left, const uint8_t* top) { + DCMode16x16(I16DC16 + dst, left, top); + VerticalPred16x16(I16VE16 + dst, top); + HorizontalPred16x16(I16HE16 + dst, left); + TrueMotion16x16(I16TM16 + dst, left, top); +} + +// Chroma 8x8 prediction + +#define CALC_DC8(in, out) do { \ + const v8u16 temp0 = __msa_hadd_u_h(in, in); \ + const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0); \ + const v2i64 temp2 = (v2i64)__msa_hadd_u_d(temp1, temp1); \ + const v2i64 temp3 = __msa_splati_d(temp2, 1); \ + const v2i64 temp4 = temp3 + temp2; \ + const v16i8 temp5 = (v16i8)__msa_srari_d(temp4, 4); \ + const v2i64 temp6 = (v2i64)__msa_splati_b(temp5, 0); \ + out = __msa_copy_s_d(temp6, 0); \ +} while (0) + +#define STORE8x8(out, dst) do { \ + SD4(out, out, out, out, dst + 0 * BPS, BPS); \ + SD4(out, out, out, out, dst + 4 * BPS, BPS); \ +} while (0) + +static WEBP_INLINE void VerticalPred8x8(uint8_t* dst, const uint8_t* top) { + if (top != NULL) { + const uint64_t out = LD(top); + STORE8x8(out, dst); + } else { + const uint64_t out = 0x7f7f7f7f7f7f7f7fULL; + STORE8x8(out, dst); + } +} + +static WEBP_INLINE void HorizontalPred8x8(uint8_t* dst, const uint8_t* left) { + if (left != NULL) { + int j; + for (j = 0; j < 8; j += 4) { + const v16u8 L0 = (v16u8)__msa_fill_b(left[0]); + const v16u8 L1 = (v16u8)__msa_fill_b(left[1]); + const v16u8 L2 = (v16u8)__msa_fill_b(left[2]); + const v16u8 L3 = (v16u8)__msa_fill_b(left[3]); + const uint64_t out0 = __msa_copy_s_d((v2i64)L0, 0); + const uint64_t out1 = __msa_copy_s_d((v2i64)L1, 0); + const uint64_t out2 = __msa_copy_s_d((v2i64)L2, 0); + const uint64_t out3 = __msa_copy_s_d((v2i64)L3, 0); + SD4(out0, out1, out2, out3, dst, BPS); + dst += 4 * BPS; + left += 4; + } + } else { + const uint64_t out = 0x8181818181818181ULL; + STORE8x8(out, dst); + } +} + +static WEBP_INLINE void TrueMotion8x8(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + if (left != NULL) { + if (top != NULL) { + int j; + const v8i16 TL = (v8i16)__msa_fill_h(left[-1]); + const v16u8 T1 = LD_UB(top); + const v16i8 zero = { 0 }; + const v8i16 T = (v8i16)__msa_ilvr_b(zero, (v16i8)T1); + const v8i16 d = T - TL; + for (j = 0; j < 8; j += 4) { + uint64_t out0, out1, out2, out3; + v16i8 t0, t1; + v8i16 r0 = (v8i16)__msa_fill_h(left[j + 0]); + v8i16 r1 = (v8i16)__msa_fill_h(left[j + 1]); + v8i16 r2 = (v8i16)__msa_fill_h(left[j + 2]); + v8i16 r3 = (v8i16)__msa_fill_h(left[j + 3]); + ADD4(d, r0, d, r1, d, r2, d, r3, r0, r1, r2, r3); + CLIP_SH4_0_255(r0, r1, r2, r3); + PCKEV_B2_SB(r1, r0, r3, r2, t0, t1); + out0 = __msa_copy_s_d((v2i64)t0, 0); + out1 = __msa_copy_s_d((v2i64)t0, 1); + out2 = __msa_copy_s_d((v2i64)t1, 0); + out3 = __msa_copy_s_d((v2i64)t1, 1); + SD4(out0, out1, out2, out3, dst, BPS); + dst += 4 * BPS; + } + } else { + HorizontalPred8x8(dst, left); + } + } else { + if (top != NULL) { + VerticalPred8x8(dst, top); + } else { + const uint64_t out = 0x8181818181818181ULL; + STORE8x8(out, dst); + } + } +} + +static WEBP_INLINE void DCMode8x8(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + uint64_t out; + v16u8 src = { 0 }; + if (top != NULL && left != NULL) { + const uint64_t left_m = LD(left); + const uint64_t top_m = LD(top); + INSERT_D2_UB(left_m, top_m, src); + CALC_DC8(src, out); + } else if (left != NULL) { // left but no top + const uint64_t left_m = LD(left); + INSERT_D2_UB(left_m, left_m, src); + CALC_DC8(src, out); + } else if (top != NULL) { // top but no left + const uint64_t top_m = LD(top); + INSERT_D2_UB(top_m, top_m, src); + CALC_DC8(src, out); + } else { // no top, no left, nothing. + src = (v16u8)__msa_fill_b(0x80); + out = __msa_copy_s_d((v2i64)src, 0); + } + STORE8x8(out, dst); +} + +static void IntraChromaPreds_MSA(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + // U block + DCMode8x8(C8DC8 + dst, left, top); + VerticalPred8x8(C8VE8 + dst, top); + HorizontalPred8x8(C8HE8 + dst, left); + TrueMotion8x8(C8TM8 + dst, left, top); + // V block + dst += 8; + if (top != NULL) top += 8; + if (left != NULL) left += 16; + DCMode8x8(C8DC8 + dst, left, top); + VerticalPred8x8(C8VE8 + dst, top); + HorizontalPred8x8(C8HE8 + dst, left); + TrueMotion8x8(C8TM8 + dst, left, top); +} + +//------------------------------------------------------------------------------ +// Metric + +#define PACK_DOTP_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do { \ + v16u8 tmp0, tmp1; \ + v8i16 tmp2, tmp3; \ + ILVRL_B2_UB(in0, in1, tmp0, tmp1); \ + HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ + DOTP_SH2_SW(tmp2, tmp3, tmp2, tmp3, out0, out1); \ + ILVRL_B2_UB(in2, in3, tmp0, tmp1); \ + HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ + DOTP_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3); \ +} while (0) + +#define PACK_DPADD_UB4_SW(in0, in1, in2, in3, out0, out1, out2, out3) do { \ + v16u8 tmp0, tmp1; \ + v8i16 tmp2, tmp3; \ + ILVRL_B2_UB(in0, in1, tmp0, tmp1); \ + HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ + DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out0, out1); \ + ILVRL_B2_UB(in2, in3, tmp0, tmp1); \ + HSUB_UB2_SH(tmp0, tmp1, tmp2, tmp3); \ + DPADD_SH2_SW(tmp2, tmp3, tmp2, tmp3, out2, out3); \ +} while (0) + +static int SSE16x16_MSA(const uint8_t* a, const uint8_t* b) { + uint32_t sum; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; + v4i32 out0, out1, out2, out3; + + LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); + LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); + PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3); + a += 8 * BPS; + b += 8 * BPS; + LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); + LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); + PACK_DPADD_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3); + out0 += out1; + out2 += out3; + out0 += out2; + sum = HADD_SW_S32(out0); + return sum; +} + +static int SSE16x8_MSA(const uint8_t* a, const uint8_t* b) { + uint32_t sum; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; + v4i32 out0, out1, out2, out3; + + LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); + LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); + PACK_DOTP_UB4_SW(src0, ref0, src1, ref1, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src2, ref2, src3, ref3, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src4, ref4, src5, ref5, out0, out1, out2, out3); + PACK_DPADD_UB4_SW(src6, ref6, src7, ref7, out0, out1, out2, out3); + out0 += out1; + out2 += out3; + out0 += out2; + sum = HADD_SW_S32(out0); + return sum; +} + +static int SSE8x8_MSA(const uint8_t* a, const uint8_t* b) { + uint32_t sum; + v16u8 src0, src1, src2, src3, src4, src5, src6, src7; + v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7; + v16u8 t0, t1, t2, t3; + v4i32 out0, out1, out2, out3; + + LD_UB8(a, BPS, src0, src1, src2, src3, src4, src5, src6, src7); + LD_UB8(b, BPS, ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7); + ILVR_B4_UB(src0, src1, src2, src3, ref0, ref1, ref2, ref3, t0, t1, t2, t3); + PACK_DOTP_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3); + ILVR_B4_UB(src4, src5, src6, src7, ref4, ref5, ref6, ref7, t0, t1, t2, t3); + PACK_DPADD_UB4_SW(t0, t2, t1, t3, out0, out1, out2, out3); + out0 += out1; + out2 += out3; + out0 += out2; + sum = HADD_SW_S32(out0); + return sum; +} + +static int SSE4x4_MSA(const uint8_t* a, const uint8_t* b) { + uint32_t sum = 0; + uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3; + v16u8 src = { 0 }, ref = { 0 }, tmp0, tmp1; + v8i16 diff0, diff1; + v4i32 out0, out1; + + LW4(a, BPS, src0, src1, src2, src3); + LW4(b, BPS, ref0, ref1, ref2, ref3); + INSERT_W4_UB(src0, src1, src2, src3, src); + INSERT_W4_UB(ref0, ref1, ref2, ref3, ref); + ILVRL_B2_UB(src, ref, tmp0, tmp1); + HSUB_UB2_SH(tmp0, tmp1, diff0, diff1); + DOTP_SH2_SW(diff0, diff1, diff0, diff1, out0, out1); + out0 += out1; + sum = HADD_SW_S32(out0); + return sum; +} + +//------------------------------------------------------------------------------ +// Quantization + +static int QuantizeBlock_MSA(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + int sum; + v8i16 in0, in1, sh0, sh1, out0, out1; + v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, sign0, sign1; + v4i32 s0, s1, s2, s3, b0, b1, b2, b3, t0, t1, t2, t3; + const v8i16 zero = { 0 }; + const v8i16 zigzag0 = { 0, 1, 4, 8, 5, 2, 3, 6 }; + const v8i16 zigzag1 = { 9, 12, 13, 10, 7, 11, 14, 15 }; + const v8i16 maxlevel = __msa_fill_h(MAX_LEVEL); + + LD_SH2(&in[0], 8, in0, in1); + LD_SH2(&mtx->sharpen_[0], 8, sh0, sh1); + tmp4 = __msa_add_a_h(in0, zero); + tmp5 = __msa_add_a_h(in1, zero); + ILVRL_H2_SH(sh0, tmp4, tmp0, tmp1); + ILVRL_H2_SH(sh1, tmp5, tmp2, tmp3); + HADD_SH4_SW(tmp0, tmp1, tmp2, tmp3, s0, s1, s2, s3); + sign0 = (in0 < zero); + sign1 = (in1 < zero); // sign + LD_SH2(&mtx->iq_[0], 8, tmp0, tmp1); // iq + ILVRL_H2_SW(zero, tmp0, t0, t1); + ILVRL_H2_SW(zero, tmp1, t2, t3); + LD_SW4(&mtx->bias_[0], 4, b0, b1, b2, b3); // bias + MUL4(t0, s0, t1, s1, t2, s2, t3, s3, t0, t1, t2, t3); + ADD4(b0, t0, b1, t1, b2, t2, b3, t3, b0, b1, b2, b3); + SRAI_W4_SW(b0, b1, b2, b3, 17); + PCKEV_H2_SH(b1, b0, b3, b2, tmp2, tmp3); + tmp0 = (tmp2 > maxlevel); + tmp1 = (tmp3 > maxlevel); + tmp2 = (v8i16)__msa_bmnz_v((v16u8)tmp2, (v16u8)maxlevel, (v16u8)tmp0); + tmp3 = (v8i16)__msa_bmnz_v((v16u8)tmp3, (v16u8)maxlevel, (v16u8)tmp1); + SUB2(zero, tmp2, zero, tmp3, tmp0, tmp1); + tmp2 = (v8i16)__msa_bmnz_v((v16u8)tmp2, (v16u8)tmp0, (v16u8)sign0); + tmp3 = (v8i16)__msa_bmnz_v((v16u8)tmp3, (v16u8)tmp1, (v16u8)sign1); + LD_SW4(&mtx->zthresh_[0], 4, t0, t1, t2, t3); // zthresh + t0 = (s0 > t0); + t1 = (s1 > t1); + t2 = (s2 > t2); + t3 = (s3 > t3); + PCKEV_H2_SH(t1, t0, t3, t2, tmp0, tmp1); + tmp4 = (v8i16)__msa_bmnz_v((v16u8)zero, (v16u8)tmp2, (v16u8)tmp0); + tmp5 = (v8i16)__msa_bmnz_v((v16u8)zero, (v16u8)tmp3, (v16u8)tmp1); + LD_SH2(&mtx->q_[0], 8, tmp0, tmp1); + MUL2(tmp4, tmp0, tmp5, tmp1, in0, in1); + VSHF_H2_SH(tmp4, tmp5, tmp4, tmp5, zigzag0, zigzag1, out0, out1); + ST_SH2(in0, in1, &in[0], 8); + ST_SH2(out0, out1, &out[0], 8); + out0 = __msa_add_a_h(out0, out1); + sum = HADD_SH_S32(out0); + return (sum > 0); +} + +static int Quantize2Blocks_MSA(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitMSA(void) { + VP8ITransform = ITransform_MSA; + VP8FTransform = FTransform_MSA; + VP8FTransformWHT = FTransformWHT_MSA; + + VP8TDisto4x4 = Disto4x4_MSA; + VP8TDisto16x16 = Disto16x16_MSA; + VP8CollectHistogram = CollectHistogram_MSA; + + VP8EncPredLuma4 = Intra4Preds_MSA; + VP8EncPredLuma16 = Intra16Preds_MSA; + VP8EncPredChroma8 = IntraChromaPreds_MSA; + + VP8SSE16x16 = SSE16x16_MSA; + VP8SSE16x8 = SSE16x8_MSA; + VP8SSE8x8 = SSE8x8_MSA; + VP8SSE4x4 = SSE4x4_MSA; + + VP8EncQuantizeBlock = QuantizeBlock_MSA; + VP8EncQuantize2Blocks = Quantize2Blocks_MSA; + VP8EncQuantizeBlockWHT = QuantizeBlock_MSA; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(VP8EncDspInitMSA) + +#endif // WEBP_USE_MSA diff --git a/media/libwebp/dsp/enc_neon.c b/media/libwebp/dsp/enc_neon.c new file mode 100644 index 0000000000..657be9b21b --- /dev/null +++ b/media/libwebp/dsp/enc_neon.c @@ -0,0 +1,938 @@ +// Copyright 2012 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// ARM NEON version of speed-critical encoding functions. +// +// adapted from libvpx (https://www.webmproject.org/code/) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_NEON) + +#include <assert.h> + +#include "../dsp/neon.h" +#include "../enc/vp8i_enc.h" + +//------------------------------------------------------------------------------ +// Transforms (Paragraph 14.4) + +// Inverse transform. +// This code is pretty much the same as TransformOne in the dec_neon.c, except +// for subtraction to *ref. See the comments there for algorithmic explanations. + +static const int16_t kC1 = 20091; +static const int16_t kC2 = 17734; // half of kC2, actually. See comment above. + +// This code works but is *slower* than the inlined-asm version below +// (with gcc-4.6). So we disable it for now. Later, it'll be conditional to +// WEBP_USE_INTRINSICS define. +// With gcc-4.8, it's a little faster speed than inlined-assembly. +#if defined(WEBP_USE_INTRINSICS) + +// Treats 'v' as an uint8x8_t and zero extends to an int16x8_t. +static WEBP_INLINE int16x8_t ConvertU8ToS16_NEON(uint32x2_t v) { + return vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(v))); +} + +// Performs unsigned 8b saturation on 'dst01' and 'dst23' storing the result +// to the corresponding rows of 'dst'. +static WEBP_INLINE void SaturateAndStore4x4_NEON(uint8_t* const dst, + const int16x8_t dst01, + const int16x8_t dst23) { + // Unsigned saturate to 8b. + const uint8x8_t dst01_u8 = vqmovun_s16(dst01); + const uint8x8_t dst23_u8 = vqmovun_s16(dst23); + + // Store the results. + vst1_lane_u32((uint32_t*)(dst + 0 * BPS), vreinterpret_u32_u8(dst01_u8), 0); + vst1_lane_u32((uint32_t*)(dst + 1 * BPS), vreinterpret_u32_u8(dst01_u8), 1); + vst1_lane_u32((uint32_t*)(dst + 2 * BPS), vreinterpret_u32_u8(dst23_u8), 0); + vst1_lane_u32((uint32_t*)(dst + 3 * BPS), vreinterpret_u32_u8(dst23_u8), 1); +} + +static WEBP_INLINE void Add4x4_NEON(const int16x8_t row01, + const int16x8_t row23, + const uint8_t* const ref, + uint8_t* const dst) { + uint32x2_t dst01 = vdup_n_u32(0); + uint32x2_t dst23 = vdup_n_u32(0); + + // Load the source pixels. + dst01 = vld1_lane_u32((uint32_t*)(ref + 0 * BPS), dst01, 0); + dst23 = vld1_lane_u32((uint32_t*)(ref + 2 * BPS), dst23, 0); + dst01 = vld1_lane_u32((uint32_t*)(ref + 1 * BPS), dst01, 1); + dst23 = vld1_lane_u32((uint32_t*)(ref + 3 * BPS), dst23, 1); + + { + // Convert to 16b. + const int16x8_t dst01_s16 = ConvertU8ToS16_NEON(dst01); + const int16x8_t dst23_s16 = ConvertU8ToS16_NEON(dst23); + + // Descale with rounding. + const int16x8_t out01 = vrsraq_n_s16(dst01_s16, row01, 3); + const int16x8_t out23 = vrsraq_n_s16(dst23_s16, row23, 3); + // Add the inverse transform. + SaturateAndStore4x4_NEON(dst, out01, out23); + } +} + +static WEBP_INLINE void Transpose8x2_NEON(const int16x8_t in0, + const int16x8_t in1, + int16x8x2_t* const out) { + // a0 a1 a2 a3 | b0 b1 b2 b3 => a0 b0 c0 d0 | a1 b1 c1 d1 + // c0 c1 c2 c3 | d0 d1 d2 d3 a2 b2 c2 d2 | a3 b3 c3 d3 + const int16x8x2_t tmp0 = vzipq_s16(in0, in1); // a0 c0 a1 c1 a2 c2 ... + // b0 d0 b1 d1 b2 d2 ... + *out = vzipq_s16(tmp0.val[0], tmp0.val[1]); +} + +static WEBP_INLINE void TransformPass_NEON(int16x8x2_t* const rows) { + // {rows} = in0 | in4 + // in8 | in12 + // B1 = in4 | in12 + const int16x8_t B1 = + vcombine_s16(vget_high_s16(rows->val[0]), vget_high_s16(rows->val[1])); + // C0 = kC1 * in4 | kC1 * in12 + // C1 = kC2 * in4 | kC2 * in12 + const int16x8_t C0 = vsraq_n_s16(B1, vqdmulhq_n_s16(B1, kC1), 1); + const int16x8_t C1 = vqdmulhq_n_s16(B1, kC2); + const int16x4_t a = vqadd_s16(vget_low_s16(rows->val[0]), + vget_low_s16(rows->val[1])); // in0 + in8 + const int16x4_t b = vqsub_s16(vget_low_s16(rows->val[0]), + vget_low_s16(rows->val[1])); // in0 - in8 + // c = kC2 * in4 - kC1 * in12 + // d = kC1 * in4 + kC2 * in12 + const int16x4_t c = vqsub_s16(vget_low_s16(C1), vget_high_s16(C0)); + const int16x4_t d = vqadd_s16(vget_low_s16(C0), vget_high_s16(C1)); + const int16x8_t D0 = vcombine_s16(a, b); // D0 = a | b + const int16x8_t D1 = vcombine_s16(d, c); // D1 = d | c + const int16x8_t E0 = vqaddq_s16(D0, D1); // a+d | b+c + const int16x8_t E_tmp = vqsubq_s16(D0, D1); // a-d | b-c + const int16x8_t E1 = vcombine_s16(vget_high_s16(E_tmp), vget_low_s16(E_tmp)); + Transpose8x2_NEON(E0, E1, rows); +} + +static void ITransformOne_NEON(const uint8_t* ref, + const int16_t* in, uint8_t* dst) { + int16x8x2_t rows; + INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8)); + TransformPass_NEON(&rows); + TransformPass_NEON(&rows); + Add4x4_NEON(rows.val[0], rows.val[1], ref, dst); +} + +#else + +static void ITransformOne_NEON(const uint8_t* ref, + const int16_t* in, uint8_t* dst) { + const int kBPS = BPS; + const int16_t kC1C2[] = { kC1, kC2, 0, 0 }; + + __asm__ volatile ( + "vld1.16 {q1, q2}, [%[in]] \n" + "vld1.16 {d0}, [%[kC1C2]] \n" + + // d2: in[0] + // d3: in[8] + // d4: in[4] + // d5: in[12] + "vswp d3, d4 \n" + + // q8 = {in[4], in[12]} * kC1 * 2 >> 16 + // q9 = {in[4], in[12]} * kC2 >> 16 + "vqdmulh.s16 q8, q2, d0[0] \n" + "vqdmulh.s16 q9, q2, d0[1] \n" + + // d22 = a = in[0] + in[8] + // d23 = b = in[0] - in[8] + "vqadd.s16 d22, d2, d3 \n" + "vqsub.s16 d23, d2, d3 \n" + + // q8 = in[4]/[12] * kC1 >> 16 + "vshr.s16 q8, q8, #1 \n" + + // Add {in[4], in[12]} back after the multiplication. + "vqadd.s16 q8, q2, q8 \n" + + // d20 = c = in[4]*kC2 - in[12]*kC1 + // d21 = d = in[4]*kC1 + in[12]*kC2 + "vqsub.s16 d20, d18, d17 \n" + "vqadd.s16 d21, d19, d16 \n" + + // d2 = tmp[0] = a + d + // d3 = tmp[1] = b + c + // d4 = tmp[2] = b - c + // d5 = tmp[3] = a - d + "vqadd.s16 d2, d22, d21 \n" + "vqadd.s16 d3, d23, d20 \n" + "vqsub.s16 d4, d23, d20 \n" + "vqsub.s16 d5, d22, d21 \n" + + "vzip.16 q1, q2 \n" + "vzip.16 q1, q2 \n" + + "vswp d3, d4 \n" + + // q8 = {tmp[4], tmp[12]} * kC1 * 2 >> 16 + // q9 = {tmp[4], tmp[12]} * kC2 >> 16 + "vqdmulh.s16 q8, q2, d0[0] \n" + "vqdmulh.s16 q9, q2, d0[1] \n" + + // d22 = a = tmp[0] + tmp[8] + // d23 = b = tmp[0] - tmp[8] + "vqadd.s16 d22, d2, d3 \n" + "vqsub.s16 d23, d2, d3 \n" + + "vshr.s16 q8, q8, #1 \n" + "vqadd.s16 q8, q2, q8 \n" + + // d20 = c = in[4]*kC2 - in[12]*kC1 + // d21 = d = in[4]*kC1 + in[12]*kC2 + "vqsub.s16 d20, d18, d17 \n" + "vqadd.s16 d21, d19, d16 \n" + + // d2 = tmp[0] = a + d + // d3 = tmp[1] = b + c + // d4 = tmp[2] = b - c + // d5 = tmp[3] = a - d + "vqadd.s16 d2, d22, d21 \n" + "vqadd.s16 d3, d23, d20 \n" + "vqsub.s16 d4, d23, d20 \n" + "vqsub.s16 d5, d22, d21 \n" + + "vld1.32 d6[0], [%[ref]], %[kBPS] \n" + "vld1.32 d6[1], [%[ref]], %[kBPS] \n" + "vld1.32 d7[0], [%[ref]], %[kBPS] \n" + "vld1.32 d7[1], [%[ref]], %[kBPS] \n" + + "sub %[ref], %[ref], %[kBPS], lsl #2 \n" + + // (val) + 4 >> 3 + "vrshr.s16 d2, d2, #3 \n" + "vrshr.s16 d3, d3, #3 \n" + "vrshr.s16 d4, d4, #3 \n" + "vrshr.s16 d5, d5, #3 \n" + + "vzip.16 q1, q2 \n" + "vzip.16 q1, q2 \n" + + // Must accumulate before saturating + "vmovl.u8 q8, d6 \n" + "vmovl.u8 q9, d7 \n" + + "vqadd.s16 q1, q1, q8 \n" + "vqadd.s16 q2, q2, q9 \n" + + "vqmovun.s16 d0, q1 \n" + "vqmovun.s16 d1, q2 \n" + + "vst1.32 d0[0], [%[dst]], %[kBPS] \n" + "vst1.32 d0[1], [%[dst]], %[kBPS] \n" + "vst1.32 d1[0], [%[dst]], %[kBPS] \n" + "vst1.32 d1[1], [%[dst]] \n" + + : [in] "+r"(in), [dst] "+r"(dst) // modified registers + : [kBPS] "r"(kBPS), [kC1C2] "r"(kC1C2), [ref] "r"(ref) // constants + : "memory", "q0", "q1", "q2", "q8", "q9", "q10", "q11" // clobbered + ); +} + +#endif // WEBP_USE_INTRINSICS + +static void ITransform_NEON(const uint8_t* ref, + const int16_t* in, uint8_t* dst, int do_two) { + ITransformOne_NEON(ref, in, dst); + if (do_two) { + ITransformOne_NEON(ref + 4, in + 16, dst + 4); + } +} + +// Load all 4x4 pixels into a single uint8x16_t variable. +static uint8x16_t Load4x4_NEON(const uint8_t* src) { + uint32x4_t out = vdupq_n_u32(0); + out = vld1q_lane_u32((const uint32_t*)(src + 0 * BPS), out, 0); + out = vld1q_lane_u32((const uint32_t*)(src + 1 * BPS), out, 1); + out = vld1q_lane_u32((const uint32_t*)(src + 2 * BPS), out, 2); + out = vld1q_lane_u32((const uint32_t*)(src + 3 * BPS), out, 3); + return vreinterpretq_u8_u32(out); +} + +// Forward transform. + +#if defined(WEBP_USE_INTRINSICS) + +static WEBP_INLINE void Transpose4x4_S16_NEON(const int16x4_t A, + const int16x4_t B, + const int16x4_t C, + const int16x4_t D, + int16x8_t* const out01, + int16x8_t* const out32) { + const int16x4x2_t AB = vtrn_s16(A, B); + const int16x4x2_t CD = vtrn_s16(C, D); + const int32x2x2_t tmp02 = vtrn_s32(vreinterpret_s32_s16(AB.val[0]), + vreinterpret_s32_s16(CD.val[0])); + const int32x2x2_t tmp13 = vtrn_s32(vreinterpret_s32_s16(AB.val[1]), + vreinterpret_s32_s16(CD.val[1])); + *out01 = vreinterpretq_s16_s64( + vcombine_s64(vreinterpret_s64_s32(tmp02.val[0]), + vreinterpret_s64_s32(tmp13.val[0]))); + *out32 = vreinterpretq_s16_s64( + vcombine_s64(vreinterpret_s64_s32(tmp13.val[1]), + vreinterpret_s64_s32(tmp02.val[1]))); +} + +static WEBP_INLINE int16x8_t DiffU8ToS16_NEON(const uint8x8_t a, + const uint8x8_t b) { + return vreinterpretq_s16_u16(vsubl_u8(a, b)); +} + +static void FTransform_NEON(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + int16x8_t d0d1, d3d2; // working 4x4 int16 variables + { + const uint8x16_t S0 = Load4x4_NEON(src); + const uint8x16_t R0 = Load4x4_NEON(ref); + const int16x8_t D0D1 = DiffU8ToS16_NEON(vget_low_u8(S0), vget_low_u8(R0)); + const int16x8_t D2D3 = DiffU8ToS16_NEON(vget_high_u8(S0), vget_high_u8(R0)); + const int16x4_t D0 = vget_low_s16(D0D1); + const int16x4_t D1 = vget_high_s16(D0D1); + const int16x4_t D2 = vget_low_s16(D2D3); + const int16x4_t D3 = vget_high_s16(D2D3); + Transpose4x4_S16_NEON(D0, D1, D2, D3, &d0d1, &d3d2); + } + { // 1rst pass + const int32x4_t kCst937 = vdupq_n_s32(937); + const int32x4_t kCst1812 = vdupq_n_s32(1812); + const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2); // d0+d3 | d1+d2 (=a0|a1) + const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2); // d0-d3 | d1-d2 (=a3|a2) + const int16x8_t a0a1_2 = vshlq_n_s16(a0a1, 3); + const int16x4_t tmp0 = vadd_s16(vget_low_s16(a0a1_2), + vget_high_s16(a0a1_2)); + const int16x4_t tmp2 = vsub_s16(vget_low_s16(a0a1_2), + vget_high_s16(a0a1_2)); + const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217); + const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217); + const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352); + const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352); + const int16x4_t tmp1 = vshrn_n_s32(vaddq_s32(a2_p_a3, kCst1812), 9); + const int16x4_t tmp3 = vshrn_n_s32(vaddq_s32(a3_m_a2, kCst937), 9); + Transpose4x4_S16_NEON(tmp0, tmp1, tmp2, tmp3, &d0d1, &d3d2); + } + { // 2nd pass + // the (1<<16) addition is for the replacement: a3!=0 <-> 1-(a3==0) + const int32x4_t kCst12000 = vdupq_n_s32(12000 + (1 << 16)); + const int32x4_t kCst51000 = vdupq_n_s32(51000); + const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2); // d0+d3 | d1+d2 (=a0|a1) + const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2); // d0-d3 | d1-d2 (=a3|a2) + const int16x4_t a0_k7 = vadd_s16(vget_low_s16(a0a1), vdup_n_s16(7)); + const int16x4_t out0 = vshr_n_s16(vadd_s16(a0_k7, vget_high_s16(a0a1)), 4); + const int16x4_t out2 = vshr_n_s16(vsub_s16(a0_k7, vget_high_s16(a0a1)), 4); + const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217); + const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217); + const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352); + const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352); + const int16x4_t tmp1 = vaddhn_s32(a2_p_a3, kCst12000); + const int16x4_t out3 = vaddhn_s32(a3_m_a2, kCst51000); + const int16x4_t a3_eq_0 = + vreinterpret_s16_u16(vceq_s16(vget_low_s16(a3a2), vdup_n_s16(0))); + const int16x4_t out1 = vadd_s16(tmp1, a3_eq_0); + vst1_s16(out + 0, out0); + vst1_s16(out + 4, out1); + vst1_s16(out + 8, out2); + vst1_s16(out + 12, out3); + } +} + +#else + +// adapted from vp8/encoder/arm/neon/shortfdct_neon.asm +static const int16_t kCoeff16[] = { + 5352, 5352, 5352, 5352, 2217, 2217, 2217, 2217 +}; +static const int32_t kCoeff32[] = { + 1812, 1812, 1812, 1812, + 937, 937, 937, 937, + 12000, 12000, 12000, 12000, + 51000, 51000, 51000, 51000 +}; + +static void FTransform_NEON(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + const int kBPS = BPS; + const uint8_t* src_ptr = src; + const uint8_t* ref_ptr = ref; + const int16_t* coeff16 = kCoeff16; + const int32_t* coeff32 = kCoeff32; + + __asm__ volatile ( + // load src into q4, q5 in high half + "vld1.8 {d8}, [%[src_ptr]], %[kBPS] \n" + "vld1.8 {d10}, [%[src_ptr]], %[kBPS] \n" + "vld1.8 {d9}, [%[src_ptr]], %[kBPS] \n" + "vld1.8 {d11}, [%[src_ptr]] \n" + + // load ref into q6, q7 in high half + "vld1.8 {d12}, [%[ref_ptr]], %[kBPS] \n" + "vld1.8 {d14}, [%[ref_ptr]], %[kBPS] \n" + "vld1.8 {d13}, [%[ref_ptr]], %[kBPS] \n" + "vld1.8 {d15}, [%[ref_ptr]] \n" + + // Pack the high values in to q4 and q6 + "vtrn.32 q4, q5 \n" + "vtrn.32 q6, q7 \n" + + // d[0-3] = src - ref + "vsubl.u8 q0, d8, d12 \n" + "vsubl.u8 q1, d9, d13 \n" + + // load coeff16 into q8(d16=5352, d17=2217) + "vld1.16 {q8}, [%[coeff16]] \n" + + // load coeff32 high half into q9 = 1812, q10 = 937 + "vld1.32 {q9, q10}, [%[coeff32]]! \n" + + // load coeff32 low half into q11=12000, q12=51000 + "vld1.32 {q11,q12}, [%[coeff32]] \n" + + // part 1 + // Transpose. Register dN is the same as dN in C + "vtrn.32 d0, d2 \n" + "vtrn.32 d1, d3 \n" + "vtrn.16 d0, d1 \n" + "vtrn.16 d2, d3 \n" + + "vadd.s16 d4, d0, d3 \n" // a0 = d0 + d3 + "vadd.s16 d5, d1, d2 \n" // a1 = d1 + d2 + "vsub.s16 d6, d1, d2 \n" // a2 = d1 - d2 + "vsub.s16 d7, d0, d3 \n" // a3 = d0 - d3 + + "vadd.s16 d0, d4, d5 \n" // a0 + a1 + "vshl.s16 d0, d0, #3 \n" // temp[0+i*4] = (a0+a1) << 3 + "vsub.s16 d2, d4, d5 \n" // a0 - a1 + "vshl.s16 d2, d2, #3 \n" // (temp[2+i*4] = (a0-a1) << 3 + + "vmlal.s16 q9, d7, d16 \n" // a3*5352 + 1812 + "vmlal.s16 q10, d7, d17 \n" // a3*2217 + 937 + "vmlal.s16 q9, d6, d17 \n" // a2*2217 + a3*5352 + 1812 + "vmlsl.s16 q10, d6, d16 \n" // a3*2217 + 937 - a2*5352 + + // temp[1+i*4] = (d2*2217 + d3*5352 + 1812) >> 9 + // temp[3+i*4] = (d3*2217 + 937 - d2*5352) >> 9 + "vshrn.s32 d1, q9, #9 \n" + "vshrn.s32 d3, q10, #9 \n" + + // part 2 + // transpose d0=ip[0], d1=ip[4], d2=ip[8], d3=ip[12] + "vtrn.32 d0, d2 \n" + "vtrn.32 d1, d3 \n" + "vtrn.16 d0, d1 \n" + "vtrn.16 d2, d3 \n" + + "vmov.s16 d26, #7 \n" + + "vadd.s16 d4, d0, d3 \n" // a1 = ip[0] + ip[12] + "vadd.s16 d5, d1, d2 \n" // b1 = ip[4] + ip[8] + "vsub.s16 d6, d1, d2 \n" // c1 = ip[4] - ip[8] + "vadd.s16 d4, d4, d26 \n" // a1 + 7 + "vsub.s16 d7, d0, d3 \n" // d1 = ip[0] - ip[12] + + "vadd.s16 d0, d4, d5 \n" // op[0] = a1 + b1 + 7 + "vsub.s16 d2, d4, d5 \n" // op[8] = a1 - b1 + 7 + + "vmlal.s16 q11, d7, d16 \n" // d1*5352 + 12000 + "vmlal.s16 q12, d7, d17 \n" // d1*2217 + 51000 + + "vceq.s16 d4, d7, #0 \n" + + "vshr.s16 d0, d0, #4 \n" + "vshr.s16 d2, d2, #4 \n" + + "vmlal.s16 q11, d6, d17 \n" // c1*2217 + d1*5352 + 12000 + "vmlsl.s16 q12, d6, d16 \n" // d1*2217 - c1*5352 + 51000 + + "vmvn d4, d4 \n" // !(d1 == 0) + // op[4] = (c1*2217 + d1*5352 + 12000)>>16 + "vshrn.s32 d1, q11, #16 \n" + // op[4] += (d1!=0) + "vsub.s16 d1, d1, d4 \n" + // op[12]= (d1*2217 - c1*5352 + 51000)>>16 + "vshrn.s32 d3, q12, #16 \n" + + // set result to out array + "vst1.16 {q0, q1}, [%[out]] \n" + : [src_ptr] "+r"(src_ptr), [ref_ptr] "+r"(ref_ptr), + [coeff32] "+r"(coeff32) // modified registers + : [kBPS] "r"(kBPS), [coeff16] "r"(coeff16), + [out] "r"(out) // constants + : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", + "q10", "q11", "q12", "q13" // clobbered + ); +} + +#endif + +#define LOAD_LANE_16b(VALUE, LANE) do { \ + (VALUE) = vld1_lane_s16(src, (VALUE), (LANE)); \ + src += stride; \ +} while (0) + +static void FTransformWHT_NEON(const int16_t* src, int16_t* out) { + const int stride = 16; + const int16x4_t zero = vdup_n_s16(0); + int32x4x4_t tmp0; + int16x4x4_t in; + INIT_VECTOR4(in, zero, zero, zero, zero); + LOAD_LANE_16b(in.val[0], 0); + LOAD_LANE_16b(in.val[1], 0); + LOAD_LANE_16b(in.val[2], 0); + LOAD_LANE_16b(in.val[3], 0); + LOAD_LANE_16b(in.val[0], 1); + LOAD_LANE_16b(in.val[1], 1); + LOAD_LANE_16b(in.val[2], 1); + LOAD_LANE_16b(in.val[3], 1); + LOAD_LANE_16b(in.val[0], 2); + LOAD_LANE_16b(in.val[1], 2); + LOAD_LANE_16b(in.val[2], 2); + LOAD_LANE_16b(in.val[3], 2); + LOAD_LANE_16b(in.val[0], 3); + LOAD_LANE_16b(in.val[1], 3); + LOAD_LANE_16b(in.val[2], 3); + LOAD_LANE_16b(in.val[3], 3); + + { + // a0 = in[0 * 16] + in[2 * 16] + // a1 = in[1 * 16] + in[3 * 16] + // a2 = in[1 * 16] - in[3 * 16] + // a3 = in[0 * 16] - in[2 * 16] + const int32x4_t a0 = vaddl_s16(in.val[0], in.val[2]); + const int32x4_t a1 = vaddl_s16(in.val[1], in.val[3]); + const int32x4_t a2 = vsubl_s16(in.val[1], in.val[3]); + const int32x4_t a3 = vsubl_s16(in.val[0], in.val[2]); + tmp0.val[0] = vaddq_s32(a0, a1); + tmp0.val[1] = vaddq_s32(a3, a2); + tmp0.val[2] = vsubq_s32(a3, a2); + tmp0.val[3] = vsubq_s32(a0, a1); + } + { + const int32x4x4_t tmp1 = Transpose4x4_NEON(tmp0); + // a0 = tmp[0 + i] + tmp[ 8 + i] + // a1 = tmp[4 + i] + tmp[12 + i] + // a2 = tmp[4 + i] - tmp[12 + i] + // a3 = tmp[0 + i] - tmp[ 8 + i] + const int32x4_t a0 = vaddq_s32(tmp1.val[0], tmp1.val[2]); + const int32x4_t a1 = vaddq_s32(tmp1.val[1], tmp1.val[3]); + const int32x4_t a2 = vsubq_s32(tmp1.val[1], tmp1.val[3]); + const int32x4_t a3 = vsubq_s32(tmp1.val[0], tmp1.val[2]); + const int32x4_t b0 = vhaddq_s32(a0, a1); // (a0 + a1) >> 1 + const int32x4_t b1 = vhaddq_s32(a3, a2); // (a3 + a2) >> 1 + const int32x4_t b2 = vhsubq_s32(a3, a2); // (a3 - a2) >> 1 + const int32x4_t b3 = vhsubq_s32(a0, a1); // (a0 - a1) >> 1 + const int16x4_t out0 = vmovn_s32(b0); + const int16x4_t out1 = vmovn_s32(b1); + const int16x4_t out2 = vmovn_s32(b2); + const int16x4_t out3 = vmovn_s32(b3); + + vst1_s16(out + 0, out0); + vst1_s16(out + 4, out1); + vst1_s16(out + 8, out2); + vst1_s16(out + 12, out3); + } +} +#undef LOAD_LANE_16b + +//------------------------------------------------------------------------------ +// Texture distortion +// +// We try to match the spectral content (weighted) between source and +// reconstructed samples. + +// a 0123, b 0123 +// a 4567, b 4567 +// a 89ab, b 89ab +// a cdef, b cdef +// +// transpose +// +// a 048c, b 048c +// a 159d, b 159d +// a 26ae, b 26ae +// a 37bf, b 37bf +// +static WEBP_INLINE int16x8x4_t DistoTranspose4x4S16_NEON(int16x8x4_t q4_in) { + const int16x8x2_t q2_tmp0 = vtrnq_s16(q4_in.val[0], q4_in.val[1]); + const int16x8x2_t q2_tmp1 = vtrnq_s16(q4_in.val[2], q4_in.val[3]); + const int32x4x2_t q2_tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q2_tmp0.val[0]), + vreinterpretq_s32_s16(q2_tmp1.val[0])); + const int32x4x2_t q2_tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q2_tmp0.val[1]), + vreinterpretq_s32_s16(q2_tmp1.val[1])); + q4_in.val[0] = vreinterpretq_s16_s32(q2_tmp2.val[0]); + q4_in.val[2] = vreinterpretq_s16_s32(q2_tmp2.val[1]); + q4_in.val[1] = vreinterpretq_s16_s32(q2_tmp3.val[0]); + q4_in.val[3] = vreinterpretq_s16_s32(q2_tmp3.val[1]); + return q4_in; +} + +static WEBP_INLINE int16x8x4_t DistoHorizontalPass_NEON( + const int16x8x4_t q4_in) { + // {a0, a1} = {in[0] + in[2], in[1] + in[3]} + // {a3, a2} = {in[0] - in[2], in[1] - in[3]} + const int16x8_t q_a0 = vaddq_s16(q4_in.val[0], q4_in.val[2]); + const int16x8_t q_a1 = vaddq_s16(q4_in.val[1], q4_in.val[3]); + const int16x8_t q_a3 = vsubq_s16(q4_in.val[0], q4_in.val[2]); + const int16x8_t q_a2 = vsubq_s16(q4_in.val[1], q4_in.val[3]); + int16x8x4_t q4_out; + // tmp[0] = a0 + a1 + // tmp[1] = a3 + a2 + // tmp[2] = a3 - a2 + // tmp[3] = a0 - a1 + INIT_VECTOR4(q4_out, + vabsq_s16(vaddq_s16(q_a0, q_a1)), + vabsq_s16(vaddq_s16(q_a3, q_a2)), + vabdq_s16(q_a3, q_a2), vabdq_s16(q_a0, q_a1)); + return q4_out; +} + +static WEBP_INLINE int16x8x4_t DistoVerticalPass_NEON(const uint8x8x4_t q4_in) { + const int16x8_t q_a0 = vreinterpretq_s16_u16(vaddl_u8(q4_in.val[0], + q4_in.val[2])); + const int16x8_t q_a1 = vreinterpretq_s16_u16(vaddl_u8(q4_in.val[1], + q4_in.val[3])); + const int16x8_t q_a2 = vreinterpretq_s16_u16(vsubl_u8(q4_in.val[1], + q4_in.val[3])); + const int16x8_t q_a3 = vreinterpretq_s16_u16(vsubl_u8(q4_in.val[0], + q4_in.val[2])); + int16x8x4_t q4_out; + + INIT_VECTOR4(q4_out, + vaddq_s16(q_a0, q_a1), vaddq_s16(q_a3, q_a2), + vsubq_s16(q_a3, q_a2), vsubq_s16(q_a0, q_a1)); + return q4_out; +} + +static WEBP_INLINE int16x4x4_t DistoLoadW_NEON(const uint16_t* w) { + const uint16x8_t q_w07 = vld1q_u16(&w[0]); + const uint16x8_t q_w8f = vld1q_u16(&w[8]); + int16x4x4_t d4_w; + INIT_VECTOR4(d4_w, + vget_low_s16(vreinterpretq_s16_u16(q_w07)), + vget_high_s16(vreinterpretq_s16_u16(q_w07)), + vget_low_s16(vreinterpretq_s16_u16(q_w8f)), + vget_high_s16(vreinterpretq_s16_u16(q_w8f))); + return d4_w; +} + +static WEBP_INLINE int32x2_t DistoSum_NEON(const int16x8x4_t q4_in, + const int16x4x4_t d4_w) { + int32x2_t d_sum; + // sum += w[ 0] * abs(b0); + // sum += w[ 4] * abs(b1); + // sum += w[ 8] * abs(b2); + // sum += w[12] * abs(b3); + int32x4_t q_sum0 = vmull_s16(d4_w.val[0], vget_low_s16(q4_in.val[0])); + int32x4_t q_sum1 = vmull_s16(d4_w.val[1], vget_low_s16(q4_in.val[1])); + int32x4_t q_sum2 = vmull_s16(d4_w.val[2], vget_low_s16(q4_in.val[2])); + int32x4_t q_sum3 = vmull_s16(d4_w.val[3], vget_low_s16(q4_in.val[3])); + q_sum0 = vmlsl_s16(q_sum0, d4_w.val[0], vget_high_s16(q4_in.val[0])); + q_sum1 = vmlsl_s16(q_sum1, d4_w.val[1], vget_high_s16(q4_in.val[1])); + q_sum2 = vmlsl_s16(q_sum2, d4_w.val[2], vget_high_s16(q4_in.val[2])); + q_sum3 = vmlsl_s16(q_sum3, d4_w.val[3], vget_high_s16(q4_in.val[3])); + + q_sum0 = vaddq_s32(q_sum0, q_sum1); + q_sum2 = vaddq_s32(q_sum2, q_sum3); + q_sum2 = vaddq_s32(q_sum0, q_sum2); + d_sum = vpadd_s32(vget_low_s32(q_sum2), vget_high_s32(q_sum2)); + d_sum = vpadd_s32(d_sum, d_sum); + return d_sum; +} + +#define LOAD_LANE_32b(src, VALUE, LANE) \ + (VALUE) = vld1_lane_u32((const uint32_t*)(src), (VALUE), (LANE)) + +// Hadamard transform +// Returns the weighted sum of the absolute value of transformed coefficients. +// w[] contains a row-major 4 by 4 symmetric matrix. +static int Disto4x4_NEON(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + uint32x2_t d_in_ab_0123 = vdup_n_u32(0); + uint32x2_t d_in_ab_4567 = vdup_n_u32(0); + uint32x2_t d_in_ab_89ab = vdup_n_u32(0); + uint32x2_t d_in_ab_cdef = vdup_n_u32(0); + uint8x8x4_t d4_in; + + // load data a, b + LOAD_LANE_32b(a + 0 * BPS, d_in_ab_0123, 0); + LOAD_LANE_32b(a + 1 * BPS, d_in_ab_4567, 0); + LOAD_LANE_32b(a + 2 * BPS, d_in_ab_89ab, 0); + LOAD_LANE_32b(a + 3 * BPS, d_in_ab_cdef, 0); + LOAD_LANE_32b(b + 0 * BPS, d_in_ab_0123, 1); + LOAD_LANE_32b(b + 1 * BPS, d_in_ab_4567, 1); + LOAD_LANE_32b(b + 2 * BPS, d_in_ab_89ab, 1); + LOAD_LANE_32b(b + 3 * BPS, d_in_ab_cdef, 1); + INIT_VECTOR4(d4_in, + vreinterpret_u8_u32(d_in_ab_0123), + vreinterpret_u8_u32(d_in_ab_4567), + vreinterpret_u8_u32(d_in_ab_89ab), + vreinterpret_u8_u32(d_in_ab_cdef)); + + { + // Vertical pass first to avoid a transpose (vertical and horizontal passes + // are commutative because w/kWeightY is symmetric) and subsequent + // transpose. + const int16x8x4_t q4_v = DistoVerticalPass_NEON(d4_in); + const int16x4x4_t d4_w = DistoLoadW_NEON(w); + // horizontal pass + const int16x8x4_t q4_t = DistoTranspose4x4S16_NEON(q4_v); + const int16x8x4_t q4_h = DistoHorizontalPass_NEON(q4_t); + int32x2_t d_sum = DistoSum_NEON(q4_h, d4_w); + + // abs(sum2 - sum1) >> 5 + d_sum = vabs_s32(d_sum); + d_sum = vshr_n_s32(d_sum, 5); + return vget_lane_s32(d_sum, 0); + } +} +#undef LOAD_LANE_32b + +static int Disto16x16_NEON(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4_NEON(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ + +static void CollectHistogram_NEON(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + const uint16x8_t max_coeff_thresh = vdupq_n_u16(MAX_COEFF_THRESH); + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + FTransform_NEON(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + { + int k; + const int16x8_t a0 = vld1q_s16(out + 0); + const int16x8_t b0 = vld1q_s16(out + 8); + const uint16x8_t a1 = vreinterpretq_u16_s16(vabsq_s16(a0)); + const uint16x8_t b1 = vreinterpretq_u16_s16(vabsq_s16(b0)); + const uint16x8_t a2 = vshrq_n_u16(a1, 3); + const uint16x8_t b2 = vshrq_n_u16(b1, 3); + const uint16x8_t a3 = vminq_u16(a2, max_coeff_thresh); + const uint16x8_t b3 = vminq_u16(b2, max_coeff_thresh); + vst1q_s16(out + 0, vreinterpretq_s16_u16(a3)); + vst1q_s16(out + 8, vreinterpretq_s16_u16(b3)); + // Convert coefficients to bin. + for (k = 0; k < 16; ++k) { + ++distribution[out[k]]; + } + } + } + VP8SetHistogramData(distribution, histo); +} + +//------------------------------------------------------------------------------ + +static WEBP_INLINE void AccumulateSSE16_NEON(const uint8_t* const a, + const uint8_t* const b, + uint32x4_t* const sum) { + const uint8x16_t a0 = vld1q_u8(a); + const uint8x16_t b0 = vld1q_u8(b); + const uint8x16_t abs_diff = vabdq_u8(a0, b0); + const uint16x8_t prod1 = vmull_u8(vget_low_u8(abs_diff), + vget_low_u8(abs_diff)); + const uint16x8_t prod2 = vmull_u8(vget_high_u8(abs_diff), + vget_high_u8(abs_diff)); + /* pair-wise adds and widen */ + const uint32x4_t sum1 = vpaddlq_u16(prod1); + const uint32x4_t sum2 = vpaddlq_u16(prod2); + *sum = vaddq_u32(*sum, vaddq_u32(sum1, sum2)); +} + +// Horizontal sum of all four uint32_t values in 'sum'. +static int SumToInt_NEON(uint32x4_t sum) { + const uint64x2_t sum2 = vpaddlq_u32(sum); + const uint64_t sum3 = vgetq_lane_u64(sum2, 0) + vgetq_lane_u64(sum2, 1); + return (int)sum3; +} + +static int SSE16x16_NEON(const uint8_t* a, const uint8_t* b) { + uint32x4_t sum = vdupq_n_u32(0); + int y; + for (y = 0; y < 16; ++y) { + AccumulateSSE16_NEON(a + y * BPS, b + y * BPS, &sum); + } + return SumToInt_NEON(sum); +} + +static int SSE16x8_NEON(const uint8_t* a, const uint8_t* b) { + uint32x4_t sum = vdupq_n_u32(0); + int y; + for (y = 0; y < 8; ++y) { + AccumulateSSE16_NEON(a + y * BPS, b + y * BPS, &sum); + } + return SumToInt_NEON(sum); +} + +static int SSE8x8_NEON(const uint8_t* a, const uint8_t* b) { + uint32x4_t sum = vdupq_n_u32(0); + int y; + for (y = 0; y < 8; ++y) { + const uint8x8_t a0 = vld1_u8(a + y * BPS); + const uint8x8_t b0 = vld1_u8(b + y * BPS); + const uint8x8_t abs_diff = vabd_u8(a0, b0); + const uint16x8_t prod = vmull_u8(abs_diff, abs_diff); + sum = vpadalq_u16(sum, prod); + } + return SumToInt_NEON(sum); +} + +static int SSE4x4_NEON(const uint8_t* a, const uint8_t* b) { + const uint8x16_t a0 = Load4x4_NEON(a); + const uint8x16_t b0 = Load4x4_NEON(b); + const uint8x16_t abs_diff = vabdq_u8(a0, b0); + const uint16x8_t prod1 = vmull_u8(vget_low_u8(abs_diff), + vget_low_u8(abs_diff)); + const uint16x8_t prod2 = vmull_u8(vget_high_u8(abs_diff), + vget_high_u8(abs_diff)); + /* pair-wise adds and widen */ + const uint32x4_t sum1 = vpaddlq_u16(prod1); + const uint32x4_t sum2 = vpaddlq_u16(prod2); + return SumToInt_NEON(vaddq_u32(sum1, sum2)); +} + +//------------------------------------------------------------------------------ + +// Compilation with gcc-4.6.x is problematic for now. +#if !defined(WORK_AROUND_GCC) + +static int16x8_t Quantize_NEON(int16_t* const in, + const VP8Matrix* const mtx, int offset) { + const uint16x8_t sharp = vld1q_u16(&mtx->sharpen_[offset]); + const uint16x8_t q = vld1q_u16(&mtx->q_[offset]); + const uint16x8_t iq = vld1q_u16(&mtx->iq_[offset]); + const uint32x4_t bias0 = vld1q_u32(&mtx->bias_[offset + 0]); + const uint32x4_t bias1 = vld1q_u32(&mtx->bias_[offset + 4]); + + const int16x8_t a = vld1q_s16(in + offset); // in + const uint16x8_t b = vreinterpretq_u16_s16(vabsq_s16(a)); // coeff = abs(in) + const int16x8_t sign = vshrq_n_s16(a, 15); // sign + const uint16x8_t c = vaddq_u16(b, sharp); // + sharpen + const uint32x4_t m0 = vmull_u16(vget_low_u16(c), vget_low_u16(iq)); + const uint32x4_t m1 = vmull_u16(vget_high_u16(c), vget_high_u16(iq)); + const uint32x4_t m2 = vhaddq_u32(m0, bias0); + const uint32x4_t m3 = vhaddq_u32(m1, bias1); // (coeff * iQ + bias) >> 1 + const uint16x8_t c0 = vcombine_u16(vshrn_n_u32(m2, 16), + vshrn_n_u32(m3, 16)); // QFIX=17 = 16+1 + const uint16x8_t c1 = vminq_u16(c0, vdupq_n_u16(MAX_LEVEL)); + const int16x8_t c2 = veorq_s16(vreinterpretq_s16_u16(c1), sign); + const int16x8_t c3 = vsubq_s16(c2, sign); // restore sign + const int16x8_t c4 = vmulq_s16(c3, vreinterpretq_s16_u16(q)); + vst1q_s16(in + offset, c4); + assert(QFIX == 17); // this function can't work as is if QFIX != 16+1 + return c3; +} + +static const uint8_t kShuffles[4][8] = { + { 0, 1, 2, 3, 8, 9, 16, 17 }, + { 10, 11, 4, 5, 6, 7, 12, 13 }, + { 18, 19, 24, 25, 26, 27, 20, 21 }, + { 14, 15, 22, 23, 28, 29, 30, 31 } +}; + +static int QuantizeBlock_NEON(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + const int16x8_t out0 = Quantize_NEON(in, mtx, 0); + const int16x8_t out1 = Quantize_NEON(in, mtx, 8); + uint8x8x4_t shuffles; + // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use + // non-standard versions there. +#if defined(__APPLE__) && defined(__aarch64__) && \ + defined(__apple_build_version__) && (__apple_build_version__< 6020037) + uint8x16x2_t all_out; + INIT_VECTOR2(all_out, vreinterpretq_u8_s16(out0), vreinterpretq_u8_s16(out1)); + INIT_VECTOR4(shuffles, + vtbl2q_u8(all_out, vld1_u8(kShuffles[0])), + vtbl2q_u8(all_out, vld1_u8(kShuffles[1])), + vtbl2q_u8(all_out, vld1_u8(kShuffles[2])), + vtbl2q_u8(all_out, vld1_u8(kShuffles[3]))); +#else + uint8x8x4_t all_out; + INIT_VECTOR4(all_out, + vreinterpret_u8_s16(vget_low_s16(out0)), + vreinterpret_u8_s16(vget_high_s16(out0)), + vreinterpret_u8_s16(vget_low_s16(out1)), + vreinterpret_u8_s16(vget_high_s16(out1))); + INIT_VECTOR4(shuffles, + vtbl4_u8(all_out, vld1_u8(kShuffles[0])), + vtbl4_u8(all_out, vld1_u8(kShuffles[1])), + vtbl4_u8(all_out, vld1_u8(kShuffles[2])), + vtbl4_u8(all_out, vld1_u8(kShuffles[3]))); +#endif + // Zigzag reordering + vst1_u8((uint8_t*)(out + 0), shuffles.val[0]); + vst1_u8((uint8_t*)(out + 4), shuffles.val[1]); + vst1_u8((uint8_t*)(out + 8), shuffles.val[2]); + vst1_u8((uint8_t*)(out + 12), shuffles.val[3]); + // test zeros + if (*(uint64_t*)(out + 0) != 0) return 1; + if (*(uint64_t*)(out + 4) != 0) return 1; + if (*(uint64_t*)(out + 8) != 0) return 1; + if (*(uint64_t*)(out + 12) != 0) return 1; + return 0; +} + +static int Quantize2Blocks_NEON(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + nz = QuantizeBlock_NEON(in + 0 * 16, out + 0 * 16, mtx) << 0; + nz |= QuantizeBlock_NEON(in + 1 * 16, out + 1 * 16, mtx) << 1; + return nz; +} + +#endif // !WORK_AROUND_GCC + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitNEON(void) { + VP8ITransform = ITransform_NEON; + VP8FTransform = FTransform_NEON; + + VP8FTransformWHT = FTransformWHT_NEON; + + VP8TDisto4x4 = Disto4x4_NEON; + VP8TDisto16x16 = Disto16x16_NEON; + VP8CollectHistogram = CollectHistogram_NEON; + + VP8SSE16x16 = SSE16x16_NEON; + VP8SSE16x8 = SSE16x8_NEON; + VP8SSE8x8 = SSE8x8_NEON; + VP8SSE4x4 = SSE4x4_NEON; + +#if !defined(WORK_AROUND_GCC) + VP8EncQuantizeBlock = QuantizeBlock_NEON; + VP8EncQuantize2Blocks = Quantize2Blocks_NEON; +#endif +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(VP8EncDspInitNEON) + +#endif // WEBP_USE_NEON diff --git a/media/libwebp/dsp/enc_sse2.c b/media/libwebp/dsp/enc_sse2.c new file mode 100644 index 0000000000..ff78755111 --- /dev/null +++ b/media/libwebp/dsp/enc_sse2.c @@ -0,0 +1,1381 @@ +// Copyright 2011 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 version of speed-critical encoding functions. +// +// Author: Christian Duvivier (cduvivier@google.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <assert.h> +#include <stdlib.h> // for abs() +#include <emmintrin.h> + +#include "../dsp/common_sse2.h" +#include "../enc/cost_enc.h" +#include "../enc/vp8i_enc.h" + +//------------------------------------------------------------------------------ +// Transforms (Paragraph 14.4) + +// Does one or two inverse transforms. +static void ITransform_SSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst, + int do_two) { + // This implementation makes use of 16-bit fixed point versions of two + // multiply constants: + // K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16 + // K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16 + // + // To be able to use signed 16-bit integers, we use the following trick to + // have constants within range: + // - Associated constants are obtained by subtracting the 16-bit fixed point + // version of one: + // k = K - (1 << 16) => K = k + (1 << 16) + // K1 = 85267 => k1 = 20091 + // K2 = 35468 => k2 = -30068 + // - The multiplication of a variable by a constant become the sum of the + // variable and the multiplication of that variable by the associated + // constant: + // (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x + const __m128i k1 = _mm_set1_epi16(20091); + const __m128i k2 = _mm_set1_epi16(-30068); + __m128i T0, T1, T2, T3; + + // Load and concatenate the transform coefficients (we'll do two inverse + // transforms in parallel). In the case of only one inverse transform, the + // second half of the vectors will just contain random value we'll never + // use nor store. + __m128i in0, in1, in2, in3; + { + in0 = _mm_loadl_epi64((const __m128i*)&in[0]); + in1 = _mm_loadl_epi64((const __m128i*)&in[4]); + in2 = _mm_loadl_epi64((const __m128i*)&in[8]); + in3 = _mm_loadl_epi64((const __m128i*)&in[12]); + // a00 a10 a20 a30 x x x x + // a01 a11 a21 a31 x x x x + // a02 a12 a22 a32 x x x x + // a03 a13 a23 a33 x x x x + if (do_two) { + const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]); + const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]); + const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]); + const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]); + in0 = _mm_unpacklo_epi64(in0, inB0); + in1 = _mm_unpacklo_epi64(in1, inB1); + in2 = _mm_unpacklo_epi64(in2, inB2); + in3 = _mm_unpacklo_epi64(in3, inB3); + // a00 a10 a20 a30 b00 b10 b20 b30 + // a01 a11 a21 a31 b01 b11 b21 b31 + // a02 a12 a22 a32 b02 b12 b22 b32 + // a03 a13 a23 a33 b03 b13 b23 b33 + } + } + + // Vertical pass and subsequent transpose. + { + // First pass, c and d calculations are longer because of the "trick" + // multiplications. + const __m128i a = _mm_add_epi16(in0, in2); + const __m128i b = _mm_sub_epi16(in0, in2); + // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3 + const __m128i c1 = _mm_mulhi_epi16(in1, k2); + const __m128i c2 = _mm_mulhi_epi16(in3, k1); + const __m128i c3 = _mm_sub_epi16(in1, in3); + const __m128i c4 = _mm_sub_epi16(c1, c2); + const __m128i c = _mm_add_epi16(c3, c4); + // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3 + const __m128i d1 = _mm_mulhi_epi16(in1, k1); + const __m128i d2 = _mm_mulhi_epi16(in3, k2); + const __m128i d3 = _mm_add_epi16(in1, in3); + const __m128i d4 = _mm_add_epi16(d1, d2); + const __m128i d = _mm_add_epi16(d3, d4); + + // Second pass. + const __m128i tmp0 = _mm_add_epi16(a, d); + const __m128i tmp1 = _mm_add_epi16(b, c); + const __m128i tmp2 = _mm_sub_epi16(b, c); + const __m128i tmp3 = _mm_sub_epi16(a, d); + + // Transpose the two 4x4. + VP8Transpose_2_4x4_16b(&tmp0, &tmp1, &tmp2, &tmp3, &T0, &T1, &T2, &T3); + } + + // Horizontal pass and subsequent transpose. + { + // First pass, c and d calculations are longer because of the "trick" + // multiplications. + const __m128i four = _mm_set1_epi16(4); + const __m128i dc = _mm_add_epi16(T0, four); + const __m128i a = _mm_add_epi16(dc, T2); + const __m128i b = _mm_sub_epi16(dc, T2); + // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3 + const __m128i c1 = _mm_mulhi_epi16(T1, k2); + const __m128i c2 = _mm_mulhi_epi16(T3, k1); + const __m128i c3 = _mm_sub_epi16(T1, T3); + const __m128i c4 = _mm_sub_epi16(c1, c2); + const __m128i c = _mm_add_epi16(c3, c4); + // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3 + const __m128i d1 = _mm_mulhi_epi16(T1, k1); + const __m128i d2 = _mm_mulhi_epi16(T3, k2); + const __m128i d3 = _mm_add_epi16(T1, T3); + const __m128i d4 = _mm_add_epi16(d1, d2); + const __m128i d = _mm_add_epi16(d3, d4); + + // Second pass. + const __m128i tmp0 = _mm_add_epi16(a, d); + const __m128i tmp1 = _mm_add_epi16(b, c); + const __m128i tmp2 = _mm_sub_epi16(b, c); + const __m128i tmp3 = _mm_sub_epi16(a, d); + const __m128i shifted0 = _mm_srai_epi16(tmp0, 3); + const __m128i shifted1 = _mm_srai_epi16(tmp1, 3); + const __m128i shifted2 = _mm_srai_epi16(tmp2, 3); + const __m128i shifted3 = _mm_srai_epi16(tmp3, 3); + + // Transpose the two 4x4. + VP8Transpose_2_4x4_16b(&shifted0, &shifted1, &shifted2, &shifted3, &T0, &T1, + &T2, &T3); + } + + // Add inverse transform to 'ref' and store. + { + const __m128i zero = _mm_setzero_si128(); + // Load the reference(s). + __m128i ref0, ref1, ref2, ref3; + if (do_two) { + // Load eight bytes/pixels per line. + ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]); + ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]); + ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]); + ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]); + } else { + // Load four bytes/pixels per line. + ref0 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[0 * BPS])); + ref1 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[1 * BPS])); + ref2 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[2 * BPS])); + ref3 = _mm_cvtsi32_si128(WebPMemToUint32(&ref[3 * BPS])); + } + // Convert to 16b. + ref0 = _mm_unpacklo_epi8(ref0, zero); + ref1 = _mm_unpacklo_epi8(ref1, zero); + ref2 = _mm_unpacklo_epi8(ref2, zero); + ref3 = _mm_unpacklo_epi8(ref3, zero); + // Add the inverse transform(s). + ref0 = _mm_add_epi16(ref0, T0); + ref1 = _mm_add_epi16(ref1, T1); + ref2 = _mm_add_epi16(ref2, T2); + ref3 = _mm_add_epi16(ref3, T3); + // Unsigned saturate to 8b. + ref0 = _mm_packus_epi16(ref0, ref0); + ref1 = _mm_packus_epi16(ref1, ref1); + ref2 = _mm_packus_epi16(ref2, ref2); + ref3 = _mm_packus_epi16(ref3, ref3); + // Store the results. + if (do_two) { + // Store eight bytes/pixels per line. + _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0); + _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1); + _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2); + _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3); + } else { + // Store four bytes/pixels per line. + WebPUint32ToMem(&dst[0 * BPS], _mm_cvtsi128_si32(ref0)); + WebPUint32ToMem(&dst[1 * BPS], _mm_cvtsi128_si32(ref1)); + WebPUint32ToMem(&dst[2 * BPS], _mm_cvtsi128_si32(ref2)); + WebPUint32ToMem(&dst[3 * BPS], _mm_cvtsi128_si32(ref3)); + } + } +} + +static void FTransformPass1_SSE2(const __m128i* const in01, + const __m128i* const in23, + __m128i* const out01, + __m128i* const out32) { + const __m128i k937 = _mm_set1_epi32(937); + const __m128i k1812 = _mm_set1_epi32(1812); + + const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8); + const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8); + const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352, + 2217, 5352, 2217, 5352); + const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217, + -5352, 2217, -5352, 2217); + + // *in01 = 00 01 10 11 02 03 12 13 + // *in23 = 20 21 30 31 22 23 32 33 + const __m128i shuf01_p = _mm_shufflehi_epi16(*in01, _MM_SHUFFLE(2, 3, 0, 1)); + const __m128i shuf23_p = _mm_shufflehi_epi16(*in23, _MM_SHUFFLE(2, 3, 0, 1)); + // 00 01 10 11 03 02 13 12 + // 20 21 30 31 23 22 33 32 + const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p); + const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p); + // 00 01 10 11 20 21 30 31 + // 03 02 13 12 23 22 33 32 + const __m128i a01 = _mm_add_epi16(s01, s32); + const __m128i a32 = _mm_sub_epi16(s01, s32); + // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ] + // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ] + + const __m128i tmp0 = _mm_madd_epi16(a01, k88p); // [ (a0 + a1) << 3, ... ] + const __m128i tmp2 = _mm_madd_epi16(a01, k88m); // [ (a0 - a1) << 3, ... ] + const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p); + const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m); + const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812); + const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937); + const __m128i tmp1 = _mm_srai_epi32(tmp1_2, 9); + const __m128i tmp3 = _mm_srai_epi32(tmp3_2, 9); + const __m128i s03 = _mm_packs_epi32(tmp0, tmp2); + const __m128i s12 = _mm_packs_epi32(tmp1, tmp3); + const __m128i s_lo = _mm_unpacklo_epi16(s03, s12); // 0 1 0 1 0 1... + const __m128i s_hi = _mm_unpackhi_epi16(s03, s12); // 2 3 2 3 2 3 + const __m128i v23 = _mm_unpackhi_epi32(s_lo, s_hi); + *out01 = _mm_unpacklo_epi32(s_lo, s_hi); + *out32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2)); // 3 2 3 2 3 2.. +} + +static void FTransformPass2_SSE2(const __m128i* const v01, + const __m128i* const v32, + int16_t* out) { + const __m128i zero = _mm_setzero_si128(); + const __m128i seven = _mm_set1_epi16(7); + const __m128i k5352_2217 = _mm_set_epi16(5352, 2217, 5352, 2217, + 5352, 2217, 5352, 2217); + const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352, + 2217, -5352, 2217, -5352); + const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16)); + const __m128i k51000 = _mm_set1_epi32(51000); + + // Same operations are done on the (0,3) and (1,2) pairs. + // a3 = v0 - v3 + // a2 = v1 - v2 + const __m128i a32 = _mm_sub_epi16(*v01, *v32); + const __m128i a22 = _mm_unpackhi_epi64(a32, a32); + + const __m128i b23 = _mm_unpacklo_epi16(a22, a32); + const __m128i c1 = _mm_madd_epi16(b23, k5352_2217); + const __m128i c3 = _mm_madd_epi16(b23, k2217_5352); + const __m128i d1 = _mm_add_epi32(c1, k12000_plus_one); + const __m128i d3 = _mm_add_epi32(c3, k51000); + const __m128i e1 = _mm_srai_epi32(d1, 16); + const __m128i e3 = _mm_srai_epi32(d3, 16); + // f1 = ((b3 * 5352 + b2 * 2217 + 12000) >> 16) + // f3 = ((b3 * 2217 - b2 * 5352 + 51000) >> 16) + const __m128i f1 = _mm_packs_epi32(e1, e1); + const __m128i f3 = _mm_packs_epi32(e3, e3); + // g1 = f1 + (a3 != 0); + // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the + // desired (0, 1), we add one earlier through k12000_plus_one. + // -> g1 = f1 + 1 - (a3 == 0) + const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero)); + + // a0 = v0 + v3 + // a1 = v1 + v2 + const __m128i a01 = _mm_add_epi16(*v01, *v32); + const __m128i a01_plus_7 = _mm_add_epi16(a01, seven); + const __m128i a11 = _mm_unpackhi_epi64(a01, a01); + const __m128i c0 = _mm_add_epi16(a01_plus_7, a11); + const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11); + // d0 = (a0 + a1 + 7) >> 4; + // d2 = (a0 - a1 + 7) >> 4; + const __m128i d0 = _mm_srai_epi16(c0, 4); + const __m128i d2 = _mm_srai_epi16(c2, 4); + + const __m128i d0_g1 = _mm_unpacklo_epi64(d0, g1); + const __m128i d2_f3 = _mm_unpacklo_epi64(d2, f3); + _mm_storeu_si128((__m128i*)&out[0], d0_g1); + _mm_storeu_si128((__m128i*)&out[8], d2_f3); +} + +static void FTransform_SSE2(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + const __m128i zero = _mm_setzero_si128(); + // Load src. + const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]); + const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]); + const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]); + const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]); + // 00 01 02 03 * + // 10 11 12 13 * + // 20 21 22 23 * + // 30 31 32 33 * + // Shuffle. + const __m128i src_0 = _mm_unpacklo_epi16(src0, src1); + const __m128i src_1 = _mm_unpacklo_epi16(src2, src3); + // 00 01 10 11 02 03 12 13 * * ... + // 20 21 30 31 22 22 32 33 * * ... + + // Load ref. + const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]); + const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]); + const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]); + const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]); + const __m128i ref_0 = _mm_unpacklo_epi16(ref0, ref1); + const __m128i ref_1 = _mm_unpacklo_epi16(ref2, ref3); + + // Convert both to 16 bit. + const __m128i src_0_16b = _mm_unpacklo_epi8(src_0, zero); + const __m128i src_1_16b = _mm_unpacklo_epi8(src_1, zero); + const __m128i ref_0_16b = _mm_unpacklo_epi8(ref_0, zero); + const __m128i ref_1_16b = _mm_unpacklo_epi8(ref_1, zero); + + // Compute the difference. + const __m128i row01 = _mm_sub_epi16(src_0_16b, ref_0_16b); + const __m128i row23 = _mm_sub_epi16(src_1_16b, ref_1_16b); + __m128i v01, v32; + + // First pass + FTransformPass1_SSE2(&row01, &row23, &v01, &v32); + + // Second pass + FTransformPass2_SSE2(&v01, &v32, out); +} + +static void FTransform2_SSE2(const uint8_t* src, const uint8_t* ref, + int16_t* out) { + const __m128i zero = _mm_setzero_si128(); + + // Load src and convert to 16b. + const __m128i src0 = _mm_loadl_epi64((const __m128i*)&src[0 * BPS]); + const __m128i src1 = _mm_loadl_epi64((const __m128i*)&src[1 * BPS]); + const __m128i src2 = _mm_loadl_epi64((const __m128i*)&src[2 * BPS]); + const __m128i src3 = _mm_loadl_epi64((const __m128i*)&src[3 * BPS]); + const __m128i src_0 = _mm_unpacklo_epi8(src0, zero); + const __m128i src_1 = _mm_unpacklo_epi8(src1, zero); + const __m128i src_2 = _mm_unpacklo_epi8(src2, zero); + const __m128i src_3 = _mm_unpacklo_epi8(src3, zero); + // Load ref and convert to 16b. + const __m128i ref0 = _mm_loadl_epi64((const __m128i*)&ref[0 * BPS]); + const __m128i ref1 = _mm_loadl_epi64((const __m128i*)&ref[1 * BPS]); + const __m128i ref2 = _mm_loadl_epi64((const __m128i*)&ref[2 * BPS]); + const __m128i ref3 = _mm_loadl_epi64((const __m128i*)&ref[3 * BPS]); + const __m128i ref_0 = _mm_unpacklo_epi8(ref0, zero); + const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero); + const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero); + const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero); + // Compute difference. -> 00 01 02 03 00' 01' 02' 03' + const __m128i diff0 = _mm_sub_epi16(src_0, ref_0); + const __m128i diff1 = _mm_sub_epi16(src_1, ref_1); + const __m128i diff2 = _mm_sub_epi16(src_2, ref_2); + const __m128i diff3 = _mm_sub_epi16(src_3, ref_3); + + // Unpack and shuffle + // 00 01 02 03 0 0 0 0 + // 10 11 12 13 0 0 0 0 + // 20 21 22 23 0 0 0 0 + // 30 31 32 33 0 0 0 0 + const __m128i shuf01l = _mm_unpacklo_epi32(diff0, diff1); + const __m128i shuf23l = _mm_unpacklo_epi32(diff2, diff3); + const __m128i shuf01h = _mm_unpackhi_epi32(diff0, diff1); + const __m128i shuf23h = _mm_unpackhi_epi32(diff2, diff3); + __m128i v01l, v32l; + __m128i v01h, v32h; + + // First pass + FTransformPass1_SSE2(&shuf01l, &shuf23l, &v01l, &v32l); + FTransformPass1_SSE2(&shuf01h, &shuf23h, &v01h, &v32h); + + // Second pass + FTransformPass2_SSE2(&v01l, &v32l, out + 0); + FTransformPass2_SSE2(&v01h, &v32h, out + 16); +} + +static void FTransformWHTRow_SSE2(const int16_t* const in, __m128i* const out) { + const __m128i kMult = _mm_set_epi16(-1, 1, -1, 1, 1, 1, 1, 1); + const __m128i src0 = _mm_loadl_epi64((__m128i*)&in[0 * 16]); + const __m128i src1 = _mm_loadl_epi64((__m128i*)&in[1 * 16]); + const __m128i src2 = _mm_loadl_epi64((__m128i*)&in[2 * 16]); + const __m128i src3 = _mm_loadl_epi64((__m128i*)&in[3 * 16]); + const __m128i A01 = _mm_unpacklo_epi16(src0, src1); // A0 A1 | ... + const __m128i A23 = _mm_unpacklo_epi16(src2, src3); // A2 A3 | ... + const __m128i B0 = _mm_adds_epi16(A01, A23); // a0 | a1 | ... + const __m128i B1 = _mm_subs_epi16(A01, A23); // a3 | a2 | ... + const __m128i C0 = _mm_unpacklo_epi32(B0, B1); // a0 | a1 | a3 | a2 | ... + const __m128i C1 = _mm_unpacklo_epi32(B1, B0); // a3 | a2 | a0 | a1 | ... + const __m128i D = _mm_unpacklo_epi64(C0, C1); // a0 a1 a3 a2 a3 a2 a0 a1 + *out = _mm_madd_epi16(D, kMult); +} + +static void FTransformWHT_SSE2(const int16_t* in, int16_t* out) { + // Input is 12b signed. + __m128i row0, row1, row2, row3; + // Rows are 14b signed. + FTransformWHTRow_SSE2(in + 0 * 64, &row0); + FTransformWHTRow_SSE2(in + 1 * 64, &row1); + FTransformWHTRow_SSE2(in + 2 * 64, &row2); + FTransformWHTRow_SSE2(in + 3 * 64, &row3); + + { + // The a* are 15b signed. + const __m128i a0 = _mm_add_epi32(row0, row2); + const __m128i a1 = _mm_add_epi32(row1, row3); + const __m128i a2 = _mm_sub_epi32(row1, row3); + const __m128i a3 = _mm_sub_epi32(row0, row2); + const __m128i a0a3 = _mm_packs_epi32(a0, a3); + const __m128i a1a2 = _mm_packs_epi32(a1, a2); + + // The b* are 16b signed. + const __m128i b0b1 = _mm_add_epi16(a0a3, a1a2); + const __m128i b3b2 = _mm_sub_epi16(a0a3, a1a2); + const __m128i tmp_b2b3 = _mm_unpackhi_epi64(b3b2, b3b2); + const __m128i b2b3 = _mm_unpacklo_epi64(tmp_b2b3, b3b2); + + _mm_storeu_si128((__m128i*)&out[0], _mm_srai_epi16(b0b1, 1)); + _mm_storeu_si128((__m128i*)&out[8], _mm_srai_epi16(b2b3, 1)); + } +} + +//------------------------------------------------------------------------------ +// Compute susceptibility based on DCT-coeff histograms: +// the higher, the "easier" the macroblock is to compress. + +static void CollectHistogram_SSE2(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + const __m128i zero = _mm_setzero_si128(); + const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + int k; + + FTransform_SSE2(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + + // Convert coefficients to bin (within out[]). + { + // Load. + const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); + const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); + const __m128i d0 = _mm_sub_epi16(zero, out0); + const __m128i d1 = _mm_sub_epi16(zero, out1); + const __m128i abs0 = _mm_max_epi16(out0, d0); // abs(v), 16b + const __m128i abs1 = _mm_max_epi16(out1, d1); + // v = abs(out) >> 3 + const __m128i v0 = _mm_srai_epi16(abs0, 3); + const __m128i v1 = _mm_srai_epi16(abs1, 3); + // bin = min(v, MAX_COEFF_THRESH) + const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); + const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); + // Store. + _mm_storeu_si128((__m128i*)&out[0], bin0); + _mm_storeu_si128((__m128i*)&out[8], bin1); + } + + // Convert coefficients to bin. + for (k = 0; k < 16; ++k) { + ++distribution[out[k]]; + } + } + VP8SetHistogramData(distribution, histo); +} + +//------------------------------------------------------------------------------ +// Intra predictions + +// helper for chroma-DC predictions +static WEBP_INLINE void Put8x8uv_SSE2(uint8_t v, uint8_t* dst) { + int j; + const __m128i values = _mm_set1_epi8(v); + for (j = 0; j < 8; ++j) { + _mm_storel_epi64((__m128i*)(dst + j * BPS), values); + } +} + +static WEBP_INLINE void Put16_SSE2(uint8_t v, uint8_t* dst) { + int j; + const __m128i values = _mm_set1_epi8(v); + for (j = 0; j < 16; ++j) { + _mm_store_si128((__m128i*)(dst + j * BPS), values); + } +} + +static WEBP_INLINE void Fill_SSE2(uint8_t* dst, int value, int size) { + if (size == 4) { + int j; + for (j = 0; j < 4; ++j) { + memset(dst + j * BPS, value, 4); + } + } else if (size == 8) { + Put8x8uv_SSE2(value, dst); + } else { + Put16_SSE2(value, dst); + } +} + +static WEBP_INLINE void VE8uv_SSE2(uint8_t* dst, const uint8_t* top) { + int j; + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + for (j = 0; j < 8; ++j) { + _mm_storel_epi64((__m128i*)(dst + j * BPS), top_values); + } +} + +static WEBP_INLINE void VE16_SSE2(uint8_t* dst, const uint8_t* top) { + const __m128i top_values = _mm_load_si128((const __m128i*)top); + int j; + for (j = 0; j < 16; ++j) { + _mm_store_si128((__m128i*)(dst + j * BPS), top_values); + } +} + +static WEBP_INLINE void VerticalPred_SSE2(uint8_t* dst, + const uint8_t* top, int size) { + if (top != NULL) { + if (size == 8) { + VE8uv_SSE2(dst, top); + } else { + VE16_SSE2(dst, top); + } + } else { + Fill_SSE2(dst, 127, size); + } +} + +static WEBP_INLINE void HE8uv_SSE2(uint8_t* dst, const uint8_t* left) { + int j; + for (j = 0; j < 8; ++j) { + const __m128i values = _mm_set1_epi8(left[j]); + _mm_storel_epi64((__m128i*)dst, values); + dst += BPS; + } +} + +static WEBP_INLINE void HE16_SSE2(uint8_t* dst, const uint8_t* left) { + int j; + for (j = 0; j < 16; ++j) { + const __m128i values = _mm_set1_epi8(left[j]); + _mm_store_si128((__m128i*)dst, values); + dst += BPS; + } +} + +static WEBP_INLINE void HorizontalPred_SSE2(uint8_t* dst, + const uint8_t* left, int size) { + if (left != NULL) { + if (size == 8) { + HE8uv_SSE2(dst, left); + } else { + HE16_SSE2(dst, left); + } + } else { + Fill_SSE2(dst, 129, size); + } +} + +static WEBP_INLINE void TM_SSE2(uint8_t* dst, const uint8_t* left, + const uint8_t* top, int size) { + const __m128i zero = _mm_setzero_si128(); + int y; + if (size == 8) { + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + const __m128i top_base = _mm_unpacklo_epi8(top_values, zero); + for (y = 0; y < 8; ++y, dst += BPS) { + const int val = left[y] - left[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero); + _mm_storel_epi64((__m128i*)dst, out); + } + } else { + const __m128i top_values = _mm_load_si128((const __m128i*)top); + const __m128i top_base_0 = _mm_unpacklo_epi8(top_values, zero); + const __m128i top_base_1 = _mm_unpackhi_epi8(top_values, zero); + for (y = 0; y < 16; ++y, dst += BPS) { + const int val = left[y] - left[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out_0 = _mm_add_epi16(base, top_base_0); + const __m128i out_1 = _mm_add_epi16(base, top_base_1); + const __m128i out = _mm_packus_epi16(out_0, out_1); + _mm_store_si128((__m128i*)dst, out); + } + } +} + +static WEBP_INLINE void TrueMotion_SSE2(uint8_t* dst, const uint8_t* left, + const uint8_t* top, int size) { + if (left != NULL) { + if (top != NULL) { + TM_SSE2(dst, left, top, size); + } else { + HorizontalPred_SSE2(dst, left, size); + } + } else { + // true motion without left samples (hence: with default 129 value) + // is equivalent to VE prediction where you just copy the top samples. + // Note that if top samples are not available, the default value is + // then 129, and not 127 as in the VerticalPred case. + if (top != NULL) { + VerticalPred_SSE2(dst, top, size); + } else { + Fill_SSE2(dst, 129, size); + } + } +} + +static WEBP_INLINE void DC8uv_SSE2(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + const __m128i left_values = _mm_loadl_epi64((const __m128i*)left); + const __m128i combined = _mm_unpacklo_epi64(top_values, left_values); + const int DC = VP8HorizontalAdd8b(&combined) + 8; + Put8x8uv_SSE2(DC >> 4, dst); +} + +static WEBP_INLINE void DC8uvNoLeft_SSE2(uint8_t* dst, const uint8_t* top) { + const __m128i zero = _mm_setzero_si128(); + const __m128i top_values = _mm_loadl_epi64((const __m128i*)top); + const __m128i sum = _mm_sad_epu8(top_values, zero); + const int DC = _mm_cvtsi128_si32(sum) + 4; + Put8x8uv_SSE2(DC >> 3, dst); +} + +static WEBP_INLINE void DC8uvNoTop_SSE2(uint8_t* dst, const uint8_t* left) { + // 'left' is contiguous so we can reuse the top summation. + DC8uvNoLeft_SSE2(dst, left); +} + +static WEBP_INLINE void DC8uvNoTopLeft_SSE2(uint8_t* dst) { + Put8x8uv_SSE2(0x80, dst); +} + +static WEBP_INLINE void DC8uvMode_SSE2(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + if (top != NULL) { + if (left != NULL) { // top and left present + DC8uv_SSE2(dst, left, top); + } else { // top, but no left + DC8uvNoLeft_SSE2(dst, top); + } + } else if (left != NULL) { // left but no top + DC8uvNoTop_SSE2(dst, left); + } else { // no top, no left, nothing. + DC8uvNoTopLeft_SSE2(dst); + } +} + +static WEBP_INLINE void DC16_SSE2(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + const __m128i top_row = _mm_load_si128((const __m128i*)top); + const __m128i left_row = _mm_load_si128((const __m128i*)left); + const int DC = + VP8HorizontalAdd8b(&top_row) + VP8HorizontalAdd8b(&left_row) + 16; + Put16_SSE2(DC >> 5, dst); +} + +static WEBP_INLINE void DC16NoLeft_SSE2(uint8_t* dst, const uint8_t* top) { + const __m128i top_row = _mm_load_si128((const __m128i*)top); + const int DC = VP8HorizontalAdd8b(&top_row) + 8; + Put16_SSE2(DC >> 4, dst); +} + +static WEBP_INLINE void DC16NoTop_SSE2(uint8_t* dst, const uint8_t* left) { + // 'left' is contiguous so we can reuse the top summation. + DC16NoLeft_SSE2(dst, left); +} + +static WEBP_INLINE void DC16NoTopLeft_SSE2(uint8_t* dst) { + Put16_SSE2(0x80, dst); +} + +static WEBP_INLINE void DC16Mode_SSE2(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + if (top != NULL) { + if (left != NULL) { // top and left present + DC16_SSE2(dst, left, top); + } else { // top, but no left + DC16NoLeft_SSE2(dst, top); + } + } else if (left != NULL) { // left but no top + DC16NoTop_SSE2(dst, left); + } else { // no top, no left, nothing. + DC16NoTopLeft_SSE2(dst); + } +} + +//------------------------------------------------------------------------------ +// 4x4 predictions + +#define DST(x, y) dst[(x) + (y) * BPS] +#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2) +#define AVG2(a, b) (((a) + (b) + 1) >> 1) + +// We use the following 8b-arithmetic tricks: +// (a + 2 * b + c + 2) >> 2 = (AC + b + 1) >> 1 +// where: AC = (a + c) >> 1 = [(a + c + 1) >> 1] - [(a^c) & 1] +// and: +// (a + 2 * b + c + 2) >> 2 = (AB + BC + 1) >> 1 - (ab|bc)&lsb +// where: AC = (a + b + 1) >> 1, BC = (b + c + 1) >> 1 +// and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1 + +static WEBP_INLINE void VE4_SSE2(uint8_t* dst, + const uint8_t* top) { // vertical + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(top - 1)); + const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2); + const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one); + const __m128i b = _mm_subs_epu8(a, lsb); + const __m128i avg = _mm_avg_epu8(b, BCDEFGH0); + const uint32_t vals = _mm_cvtsi128_si32(avg); + int i; + for (i = 0; i < 4; ++i) { + WebPUint32ToMem(dst + i * BPS, vals); + } +} + +static WEBP_INLINE void HE4_SSE2(uint8_t* dst, + const uint8_t* top) { // horizontal + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J)); + WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K)); + WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L)); + WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L)); +} + +static WEBP_INLINE void DC4_SSE2(uint8_t* dst, const uint8_t* top) { + uint32_t dc = 4; + int i; + for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i]; + Fill_SSE2(dst, dc >> 3, 4); +} + +static WEBP_INLINE void LD4_SSE2(uint8_t* dst, + const uint8_t* top) { // Down-Left + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top); + const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2); + const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, top[7], 3); + const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0); + WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( abcdefg )); + WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1))); + WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2))); + WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3))); +} + +static WEBP_INLINE void VR4_SSE2(uint8_t* dst, + const uint8_t* top) { // Vertical-Right + const __m128i one = _mm_set1_epi8(1); + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int X = top[-1]; + const __m128i XABCD = _mm_loadl_epi64((const __m128i*)(top - 1)); + const __m128i ABCD0 = _mm_srli_si128(XABCD, 1); + const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0); + const __m128i _XABCD = _mm_slli_si128(XABCD, 1); + const __m128i IXABCD = _mm_insert_epi16(_XABCD, (short)(I | (X << 8)), 0); + const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i efgh = _mm_avg_epu8(avg2, XABCD); + WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( abcd )); + WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32( efgh )); + WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1))); + WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1))); + + // these two are hard to implement in SSE2, so we keep the C-version: + DST(0, 2) = AVG3(J, I, X); + DST(0, 3) = AVG3(K, J, I); +} + +static WEBP_INLINE void VL4_SSE2(uint8_t* dst, + const uint8_t* top) { // Vertical-Left + const __m128i one = _mm_set1_epi8(1); + const __m128i ABCDEFGH = _mm_loadl_epi64((const __m128i*)top); + const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1); + const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2); + const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_); + const __m128i avg2 = _mm_avg_epu8(CDEFGH__, BCDEFGH_); + const __m128i avg3 = _mm_avg_epu8(avg1, avg2); + const __m128i lsb1 = _mm_and_si128(_mm_xor_si128(avg1, avg2), one); + const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_); + const __m128i bc = _mm_xor_si128(CDEFGH__, BCDEFGH_); + const __m128i abbc = _mm_or_si128(ab, bc); + const __m128i lsb2 = _mm_and_si128(abbc, lsb1); + const __m128i avg4 = _mm_subs_epu8(avg3, lsb2); + const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4)); + WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( avg1 )); + WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32( avg4 )); + WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1))); + WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1))); + + // these two are hard to get and irregular + DST(3, 2) = (extra_out >> 0) & 0xff; + DST(3, 3) = (extra_out >> 8) & 0xff; +} + +static WEBP_INLINE void RD4_SSE2(uint8_t* dst, + const uint8_t* top) { // Down-right + const __m128i one = _mm_set1_epi8(1); + const __m128i LKJIXABC = _mm_loadl_epi64((const __m128i*)(top - 5)); + const __m128i LKJIXABCD = _mm_insert_epi16(LKJIXABC, top[3], 4); + const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1); + const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2); + const __m128i avg1 = _mm_avg_epu8(JIXABCD__, LKJIXABCD); + const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one); + const __m128i avg2 = _mm_subs_epu8(avg1, lsb); + const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_); + WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32( abcdefg )); + WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1))); + WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2))); + WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3))); +} + +static WEBP_INLINE void HU4_SSE2(uint8_t* dst, const uint8_t* top) { + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + DST(0, 0) = AVG2(I, J); + DST(2, 0) = DST(0, 1) = AVG2(J, K); + DST(2, 1) = DST(0, 2) = AVG2(K, L); + DST(1, 0) = AVG3(I, J, K); + DST(3, 0) = DST(1, 1) = AVG3(J, K, L); + DST(3, 1) = DST(1, 2) = AVG3(K, L, L); + DST(3, 2) = DST(2, 2) = + DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L; +} + +static WEBP_INLINE void HD4_SSE2(uint8_t* dst, const uint8_t* top) { + const int X = top[-1]; + const int I = top[-2]; + const int J = top[-3]; + const int K = top[-4]; + const int L = top[-5]; + const int A = top[0]; + const int B = top[1]; + const int C = top[2]; + + DST(0, 0) = DST(2, 1) = AVG2(I, X); + DST(0, 1) = DST(2, 2) = AVG2(J, I); + DST(0, 2) = DST(2, 3) = AVG2(K, J); + DST(0, 3) = AVG2(L, K); + + DST(3, 0) = AVG3(A, B, C); + DST(2, 0) = AVG3(X, A, B); + DST(1, 0) = DST(3, 1) = AVG3(I, X, A); + DST(1, 1) = DST(3, 2) = AVG3(J, I, X); + DST(1, 2) = DST(3, 3) = AVG3(K, J, I); + DST(1, 3) = AVG3(L, K, J); +} + +static WEBP_INLINE void TM4_SSE2(uint8_t* dst, const uint8_t* top) { + const __m128i zero = _mm_setzero_si128(); + const __m128i top_values = _mm_cvtsi32_si128(WebPMemToUint32(top)); + const __m128i top_base = _mm_unpacklo_epi8(top_values, zero); + int y; + for (y = 0; y < 4; ++y, dst += BPS) { + const int val = top[-2 - y] - top[-1]; + const __m128i base = _mm_set1_epi16(val); + const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero); + WebPUint32ToMem(dst, _mm_cvtsi128_si32(out)); + } +} + +#undef DST +#undef AVG3 +#undef AVG2 + +//------------------------------------------------------------------------------ +// luma 4x4 prediction + +// Left samples are top[-5 .. -2], top_left is top[-1], top are +// located at top[0..3], and top right is top[4..7] +static void Intra4Preds_SSE2(uint8_t* dst, const uint8_t* top) { + DC4_SSE2(I4DC4 + dst, top); + TM4_SSE2(I4TM4 + dst, top); + VE4_SSE2(I4VE4 + dst, top); + HE4_SSE2(I4HE4 + dst, top); + RD4_SSE2(I4RD4 + dst, top); + VR4_SSE2(I4VR4 + dst, top); + LD4_SSE2(I4LD4 + dst, top); + VL4_SSE2(I4VL4 + dst, top); + HD4_SSE2(I4HD4 + dst, top); + HU4_SSE2(I4HU4 + dst, top); +} + +//------------------------------------------------------------------------------ +// Chroma 8x8 prediction (paragraph 12.2) + +static void IntraChromaPreds_SSE2(uint8_t* dst, const uint8_t* left, + const uint8_t* top) { + // U block + DC8uvMode_SSE2(C8DC8 + dst, left, top); + VerticalPred_SSE2(C8VE8 + dst, top, 8); + HorizontalPred_SSE2(C8HE8 + dst, left, 8); + TrueMotion_SSE2(C8TM8 + dst, left, top, 8); + // V block + dst += 8; + if (top != NULL) top += 8; + if (left != NULL) left += 16; + DC8uvMode_SSE2(C8DC8 + dst, left, top); + VerticalPred_SSE2(C8VE8 + dst, top, 8); + HorizontalPred_SSE2(C8HE8 + dst, left, 8); + TrueMotion_SSE2(C8TM8 + dst, left, top, 8); +} + +//------------------------------------------------------------------------------ +// luma 16x16 prediction (paragraph 12.3) + +static void Intra16Preds_SSE2(uint8_t* dst, + const uint8_t* left, const uint8_t* top) { + DC16Mode_SSE2(I16DC16 + dst, left, top); + VerticalPred_SSE2(I16VE16 + dst, top, 16); + HorizontalPred_SSE2(I16HE16 + dst, left, 16); + TrueMotion_SSE2(I16TM16 + dst, left, top, 16); +} + +//------------------------------------------------------------------------------ +// Metric + +static WEBP_INLINE void SubtractAndAccumulate_SSE2(const __m128i a, + const __m128i b, + __m128i* const sum) { + // take abs(a-b) in 8b + const __m128i a_b = _mm_subs_epu8(a, b); + const __m128i b_a = _mm_subs_epu8(b, a); + const __m128i abs_a_b = _mm_or_si128(a_b, b_a); + // zero-extend to 16b + const __m128i zero = _mm_setzero_si128(); + const __m128i C0 = _mm_unpacklo_epi8(abs_a_b, zero); + const __m128i C1 = _mm_unpackhi_epi8(abs_a_b, zero); + // multiply with self + const __m128i sum1 = _mm_madd_epi16(C0, C0); + const __m128i sum2 = _mm_madd_epi16(C1, C1); + *sum = _mm_add_epi32(sum1, sum2); +} + +static WEBP_INLINE int SSE_16xN_SSE2(const uint8_t* a, const uint8_t* b, + int num_pairs) { + __m128i sum = _mm_setzero_si128(); + int32_t tmp[4]; + int i; + + for (i = 0; i < num_pairs; ++i) { + const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[BPS * 0]); + const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[BPS * 0]); + const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[BPS * 1]); + const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[BPS * 1]); + __m128i sum1, sum2; + SubtractAndAccumulate_SSE2(a0, b0, &sum1); + SubtractAndAccumulate_SSE2(a1, b1, &sum2); + sum = _mm_add_epi32(sum, _mm_add_epi32(sum1, sum2)); + a += 2 * BPS; + b += 2 * BPS; + } + _mm_storeu_si128((__m128i*)tmp, sum); + return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); +} + +static int SSE16x16_SSE2(const uint8_t* a, const uint8_t* b) { + return SSE_16xN_SSE2(a, b, 8); +} + +static int SSE16x8_SSE2(const uint8_t* a, const uint8_t* b) { + return SSE_16xN_SSE2(a, b, 4); +} + +#define LOAD_8x16b(ptr) \ + _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i*)(ptr)), zero) + +static int SSE8x8_SSE2(const uint8_t* a, const uint8_t* b) { + const __m128i zero = _mm_setzero_si128(); + int num_pairs = 4; + __m128i sum = zero; + int32_t tmp[4]; + while (num_pairs-- > 0) { + const __m128i a0 = LOAD_8x16b(&a[BPS * 0]); + const __m128i a1 = LOAD_8x16b(&a[BPS * 1]); + const __m128i b0 = LOAD_8x16b(&b[BPS * 0]); + const __m128i b1 = LOAD_8x16b(&b[BPS * 1]); + // subtract + const __m128i c0 = _mm_subs_epi16(a0, b0); + const __m128i c1 = _mm_subs_epi16(a1, b1); + // multiply/accumulate with self + const __m128i d0 = _mm_madd_epi16(c0, c0); + const __m128i d1 = _mm_madd_epi16(c1, c1); + // collect + const __m128i sum01 = _mm_add_epi32(d0, d1); + sum = _mm_add_epi32(sum, sum01); + a += 2 * BPS; + b += 2 * BPS; + } + _mm_storeu_si128((__m128i*)tmp, sum); + return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); +} +#undef LOAD_8x16b + +static int SSE4x4_SSE2(const uint8_t* a, const uint8_t* b) { + const __m128i zero = _mm_setzero_si128(); + + // Load values. Note that we read 8 pixels instead of 4, + // but the a/b buffers are over-allocated to that effect. + const __m128i a0 = _mm_loadl_epi64((const __m128i*)&a[BPS * 0]); + const __m128i a1 = _mm_loadl_epi64((const __m128i*)&a[BPS * 1]); + const __m128i a2 = _mm_loadl_epi64((const __m128i*)&a[BPS * 2]); + const __m128i a3 = _mm_loadl_epi64((const __m128i*)&a[BPS * 3]); + const __m128i b0 = _mm_loadl_epi64((const __m128i*)&b[BPS * 0]); + const __m128i b1 = _mm_loadl_epi64((const __m128i*)&b[BPS * 1]); + const __m128i b2 = _mm_loadl_epi64((const __m128i*)&b[BPS * 2]); + const __m128i b3 = _mm_loadl_epi64((const __m128i*)&b[BPS * 3]); + // Combine pair of lines. + const __m128i a01 = _mm_unpacklo_epi32(a0, a1); + const __m128i a23 = _mm_unpacklo_epi32(a2, a3); + const __m128i b01 = _mm_unpacklo_epi32(b0, b1); + const __m128i b23 = _mm_unpacklo_epi32(b2, b3); + // Convert to 16b. + const __m128i a01s = _mm_unpacklo_epi8(a01, zero); + const __m128i a23s = _mm_unpacklo_epi8(a23, zero); + const __m128i b01s = _mm_unpacklo_epi8(b01, zero); + const __m128i b23s = _mm_unpacklo_epi8(b23, zero); + // subtract, square and accumulate + const __m128i d0 = _mm_subs_epi16(a01s, b01s); + const __m128i d1 = _mm_subs_epi16(a23s, b23s); + const __m128i e0 = _mm_madd_epi16(d0, d0); + const __m128i e1 = _mm_madd_epi16(d1, d1); + const __m128i sum = _mm_add_epi32(e0, e1); + + int32_t tmp[4]; + _mm_storeu_si128((__m128i*)tmp, sum); + return (tmp[3] + tmp[2] + tmp[1] + tmp[0]); +} + +//------------------------------------------------------------------------------ + +static void Mean16x4_SSE2(const uint8_t* ref, uint32_t dc[4]) { + const __m128i mask = _mm_set1_epi16(0x00ff); + const __m128i a0 = _mm_loadu_si128((const __m128i*)&ref[BPS * 0]); + const __m128i a1 = _mm_loadu_si128((const __m128i*)&ref[BPS * 1]); + const __m128i a2 = _mm_loadu_si128((const __m128i*)&ref[BPS * 2]); + const __m128i a3 = _mm_loadu_si128((const __m128i*)&ref[BPS * 3]); + const __m128i b0 = _mm_srli_epi16(a0, 8); // hi byte + const __m128i b1 = _mm_srli_epi16(a1, 8); + const __m128i b2 = _mm_srli_epi16(a2, 8); + const __m128i b3 = _mm_srli_epi16(a3, 8); + const __m128i c0 = _mm_and_si128(a0, mask); // lo byte + const __m128i c1 = _mm_and_si128(a1, mask); + const __m128i c2 = _mm_and_si128(a2, mask); + const __m128i c3 = _mm_and_si128(a3, mask); + const __m128i d0 = _mm_add_epi32(b0, c0); + const __m128i d1 = _mm_add_epi32(b1, c1); + const __m128i d2 = _mm_add_epi32(b2, c2); + const __m128i d3 = _mm_add_epi32(b3, c3); + const __m128i e0 = _mm_add_epi32(d0, d1); + const __m128i e1 = _mm_add_epi32(d2, d3); + const __m128i f0 = _mm_add_epi32(e0, e1); + uint16_t tmp[8]; + _mm_storeu_si128((__m128i*)tmp, f0); + dc[0] = tmp[0] + tmp[1]; + dc[1] = tmp[2] + tmp[3]; + dc[2] = tmp[4] + tmp[5]; + dc[3] = tmp[6] + tmp[7]; +} + +//------------------------------------------------------------------------------ +// Texture distortion +// +// We try to match the spectral content (weighted) between source and +// reconstructed samples. + +// Hadamard transform +// Returns the weighted sum of the absolute value of transformed coefficients. +// w[] contains a row-major 4 by 4 symmetric matrix. +static int TTransform_SSE2(const uint8_t* inA, const uint8_t* inB, + const uint16_t* const w) { + int32_t sum[4]; + __m128i tmp_0, tmp_1, tmp_2, tmp_3; + const __m128i zero = _mm_setzero_si128(); + + // Load and combine inputs. + { + const __m128i inA_0 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 0]); + const __m128i inA_1 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 1]); + const __m128i inA_2 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 2]); + const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]); + const __m128i inB_0 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 0]); + const __m128i inB_1 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 1]); + const __m128i inB_2 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 2]); + const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]); + + // Combine inA and inB (we'll do two transforms in parallel). + const __m128i inAB_0 = _mm_unpacklo_epi32(inA_0, inB_0); + const __m128i inAB_1 = _mm_unpacklo_epi32(inA_1, inB_1); + const __m128i inAB_2 = _mm_unpacklo_epi32(inA_2, inB_2); + const __m128i inAB_3 = _mm_unpacklo_epi32(inA_3, inB_3); + tmp_0 = _mm_unpacklo_epi8(inAB_0, zero); + tmp_1 = _mm_unpacklo_epi8(inAB_1, zero); + tmp_2 = _mm_unpacklo_epi8(inAB_2, zero); + tmp_3 = _mm_unpacklo_epi8(inAB_3, zero); + // a00 a01 a02 a03 b00 b01 b02 b03 + // a10 a11 a12 a13 b10 b11 b12 b13 + // a20 a21 a22 a23 b20 b21 b22 b23 + // a30 a31 a32 a33 b30 b31 b32 b33 + } + + // Vertical pass first to avoid a transpose (vertical and horizontal passes + // are commutative because w/kWeightY is symmetric) and subsequent transpose. + { + // Calculate a and b (two 4x4 at once). + const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); + const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); + const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); + const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); + const __m128i b0 = _mm_add_epi16(a0, a1); + const __m128i b1 = _mm_add_epi16(a3, a2); + const __m128i b2 = _mm_sub_epi16(a3, a2); + const __m128i b3 = _mm_sub_epi16(a0, a1); + // a00 a01 a02 a03 b00 b01 b02 b03 + // a10 a11 a12 a13 b10 b11 b12 b13 + // a20 a21 a22 a23 b20 b21 b22 b23 + // a30 a31 a32 a33 b30 b31 b32 b33 + + // Transpose the two 4x4. + VP8Transpose_2_4x4_16b(&b0, &b1, &b2, &b3, &tmp_0, &tmp_1, &tmp_2, &tmp_3); + } + + // Horizontal pass and difference of weighted sums. + { + // Load all inputs. + const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]); + const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]); + + // Calculate a and b (two 4x4 at once). + const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); + const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); + const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); + const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); + const __m128i b0 = _mm_add_epi16(a0, a1); + const __m128i b1 = _mm_add_epi16(a3, a2); + const __m128i b2 = _mm_sub_epi16(a3, a2); + const __m128i b3 = _mm_sub_epi16(a0, a1); + + // Separate the transforms of inA and inB. + __m128i A_b0 = _mm_unpacklo_epi64(b0, b1); + __m128i A_b2 = _mm_unpacklo_epi64(b2, b3); + __m128i B_b0 = _mm_unpackhi_epi64(b0, b1); + __m128i B_b2 = _mm_unpackhi_epi64(b2, b3); + + { + const __m128i d0 = _mm_sub_epi16(zero, A_b0); + const __m128i d1 = _mm_sub_epi16(zero, A_b2); + const __m128i d2 = _mm_sub_epi16(zero, B_b0); + const __m128i d3 = _mm_sub_epi16(zero, B_b2); + A_b0 = _mm_max_epi16(A_b0, d0); // abs(v), 16b + A_b2 = _mm_max_epi16(A_b2, d1); + B_b0 = _mm_max_epi16(B_b0, d2); + B_b2 = _mm_max_epi16(B_b2, d3); + } + + // weighted sums + A_b0 = _mm_madd_epi16(A_b0, w_0); + A_b2 = _mm_madd_epi16(A_b2, w_8); + B_b0 = _mm_madd_epi16(B_b0, w_0); + B_b2 = _mm_madd_epi16(B_b2, w_8); + A_b0 = _mm_add_epi32(A_b0, A_b2); + B_b0 = _mm_add_epi32(B_b0, B_b2); + + // difference of weighted sums + A_b0 = _mm_sub_epi32(A_b0, B_b0); + _mm_storeu_si128((__m128i*)&sum[0], A_b0); + } + return sum[0] + sum[1] + sum[2] + sum[3]; +} + +static int Disto4x4_SSE2(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + const int diff_sum = TTransform_SSE2(a, b, w); + return abs(diff_sum) >> 5; +} + +static int Disto16x16_SSE2(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4_SSE2(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ +// Quantization +// + +static WEBP_INLINE int DoQuantizeBlock_SSE2(int16_t in[16], int16_t out[16], + const uint16_t* const sharpen, + const VP8Matrix* const mtx) { + const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL); + const __m128i zero = _mm_setzero_si128(); + __m128i coeff0, coeff8; + __m128i out0, out8; + __m128i packed_out; + + // Load all inputs. + __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]); + __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]); + const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]); + const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]); + const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]); + const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]); + + // extract sign(in) (0x0000 if positive, 0xffff if negative) + const __m128i sign0 = _mm_cmpgt_epi16(zero, in0); + const __m128i sign8 = _mm_cmpgt_epi16(zero, in8); + + // coeff = abs(in) = (in ^ sign) - sign + coeff0 = _mm_xor_si128(in0, sign0); + coeff8 = _mm_xor_si128(in8, sign8); + coeff0 = _mm_sub_epi16(coeff0, sign0); + coeff8 = _mm_sub_epi16(coeff8, sign8); + + // coeff = abs(in) + sharpen + if (sharpen != NULL) { + const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]); + const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]); + coeff0 = _mm_add_epi16(coeff0, sharpen0); + coeff8 = _mm_add_epi16(coeff8, sharpen8); + } + + // out = (coeff * iQ + B) >> QFIX + { + // doing calculations with 32b precision (QFIX=17) + // out = (coeff * iQ) + const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0); + const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0); + const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8); + const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8); + __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H); + __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H); + __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H); + __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H); + // out = (coeff * iQ + B) + const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]); + const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]); + const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]); + const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]); + out_00 = _mm_add_epi32(out_00, bias_00); + out_04 = _mm_add_epi32(out_04, bias_04); + out_08 = _mm_add_epi32(out_08, bias_08); + out_12 = _mm_add_epi32(out_12, bias_12); + // out = QUANTDIV(coeff, iQ, B, QFIX) + out_00 = _mm_srai_epi32(out_00, QFIX); + out_04 = _mm_srai_epi32(out_04, QFIX); + out_08 = _mm_srai_epi32(out_08, QFIX); + out_12 = _mm_srai_epi32(out_12, QFIX); + + // pack result as 16b + out0 = _mm_packs_epi32(out_00, out_04); + out8 = _mm_packs_epi32(out_08, out_12); + + // if (coeff > 2047) coeff = 2047 + out0 = _mm_min_epi16(out0, max_coeff_2047); + out8 = _mm_min_epi16(out8, max_coeff_2047); + } + + // get sign back (if (sign[j]) out_n = -out_n) + out0 = _mm_xor_si128(out0, sign0); + out8 = _mm_xor_si128(out8, sign8); + out0 = _mm_sub_epi16(out0, sign0); + out8 = _mm_sub_epi16(out8, sign8); + + // in = out * Q + in0 = _mm_mullo_epi16(out0, q0); + in8 = _mm_mullo_epi16(out8, q8); + + _mm_storeu_si128((__m128i*)&in[0], in0); + _mm_storeu_si128((__m128i*)&in[8], in8); + + // zigzag the output before storing it. + // + // The zigzag pattern can almost be reproduced with a small sequence of + // shuffles. After it, we only need to swap the 7th (ending up in third + // position instead of twelfth) and 8th values. + { + __m128i outZ0, outZ8; + outZ0 = _mm_shufflehi_epi16(out0, _MM_SHUFFLE(2, 1, 3, 0)); + outZ0 = _mm_shuffle_epi32 (outZ0, _MM_SHUFFLE(3, 1, 2, 0)); + outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2)); + outZ8 = _mm_shufflelo_epi16(out8, _MM_SHUFFLE(3, 0, 2, 1)); + outZ8 = _mm_shuffle_epi32 (outZ8, _MM_SHUFFLE(3, 1, 2, 0)); + outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0)); + _mm_storeu_si128((__m128i*)&out[0], outZ0); + _mm_storeu_si128((__m128i*)&out[8], outZ8); + packed_out = _mm_packs_epi16(outZ0, outZ8); + } + { + const int16_t outZ_12 = out[12]; + const int16_t outZ_3 = out[3]; + out[3] = outZ_12; + out[12] = outZ_3; + } + + // detect if all 'out' values are zeroes or not + return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff); +} + +static int QuantizeBlock_SSE2(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock_SSE2(in, out, &mtx->sharpen_[0], mtx); +} + +static int QuantizeBlockWHT_SSE2(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock_SSE2(in, out, NULL, mtx); +} + +static int Quantize2Blocks_SSE2(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + const uint16_t* const sharpen = &mtx->sharpen_[0]; + nz = DoQuantizeBlock_SSE2(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0; + nz |= DoQuantizeBlock_SSE2(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1; + return nz; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE2(void) { + VP8CollectHistogram = CollectHistogram_SSE2; + VP8EncPredLuma16 = Intra16Preds_SSE2; + VP8EncPredChroma8 = IntraChromaPreds_SSE2; + VP8EncPredLuma4 = Intra4Preds_SSE2; + VP8EncQuantizeBlock = QuantizeBlock_SSE2; + VP8EncQuantize2Blocks = Quantize2Blocks_SSE2; + VP8EncQuantizeBlockWHT = QuantizeBlockWHT_SSE2; + VP8ITransform = ITransform_SSE2; + VP8FTransform = FTransform_SSE2; + VP8FTransform2 = FTransform2_SSE2; + VP8FTransformWHT = FTransformWHT_SSE2; + VP8SSE16x16 = SSE16x16_SSE2; + VP8SSE16x8 = SSE16x8_SSE2; + VP8SSE8x8 = SSE8x8_SSE2; + VP8SSE4x4 = SSE4x4_SSE2; + VP8TDisto4x4 = Disto4x4_SSE2; + VP8TDisto16x16 = Disto16x16_SSE2; + VP8Mean16x4 = Mean16x4_SSE2; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8EncDspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/media/libwebp/dsp/enc_sse41.c b/media/libwebp/dsp/enc_sse41.c new file mode 100644 index 0000000000..09ea29361d --- /dev/null +++ b/media/libwebp/dsp/enc_sse41.c @@ -0,0 +1,339 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE4 version of some encoding functions. +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_SSE41) +#include <smmintrin.h> +#include <stdlib.h> // for abs() + +#include "../dsp/common_sse2.h" +#include "../enc/vp8i_enc.h" + +//------------------------------------------------------------------------------ +// Compute susceptibility based on DCT-coeff histograms. + +static void CollectHistogram_SSE41(const uint8_t* ref, const uint8_t* pred, + int start_block, int end_block, + VP8Histogram* const histo) { + const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH); + int j; + int distribution[MAX_COEFF_THRESH + 1] = { 0 }; + for (j = start_block; j < end_block; ++j) { + int16_t out[16]; + int k; + + VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); + + // Convert coefficients to bin (within out[]). + { + // Load. + const __m128i out0 = _mm_loadu_si128((__m128i*)&out[0]); + const __m128i out1 = _mm_loadu_si128((__m128i*)&out[8]); + // v = abs(out) >> 3 + const __m128i abs0 = _mm_abs_epi16(out0); + const __m128i abs1 = _mm_abs_epi16(out1); + const __m128i v0 = _mm_srai_epi16(abs0, 3); + const __m128i v1 = _mm_srai_epi16(abs1, 3); + // bin = min(v, MAX_COEFF_THRESH) + const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh); + const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh); + // Store. + _mm_storeu_si128((__m128i*)&out[0], bin0); + _mm_storeu_si128((__m128i*)&out[8], bin1); + } + + // Convert coefficients to bin. + for (k = 0; k < 16; ++k) { + ++distribution[out[k]]; + } + } + VP8SetHistogramData(distribution, histo); +} + +//------------------------------------------------------------------------------ +// Texture distortion +// +// We try to match the spectral content (weighted) between source and +// reconstructed samples. + +// Hadamard transform +// Returns the weighted sum of the absolute value of transformed coefficients. +// w[] contains a row-major 4 by 4 symmetric matrix. +static int TTransform_SSE41(const uint8_t* inA, const uint8_t* inB, + const uint16_t* const w) { + int32_t sum[4]; + __m128i tmp_0, tmp_1, tmp_2, tmp_3; + + // Load and combine inputs. + { + const __m128i inA_0 = _mm_loadu_si128((const __m128i*)&inA[BPS * 0]); + const __m128i inA_1 = _mm_loadu_si128((const __m128i*)&inA[BPS * 1]); + const __m128i inA_2 = _mm_loadu_si128((const __m128i*)&inA[BPS * 2]); + // In SSE4.1, with gcc 4.8 at least (maybe other versions), + // _mm_loadu_si128 is faster than _mm_loadl_epi64. But for the last lump + // of inA and inB, _mm_loadl_epi64 is still used not to have an out of + // bound read. + const __m128i inA_3 = _mm_loadl_epi64((const __m128i*)&inA[BPS * 3]); + const __m128i inB_0 = _mm_loadu_si128((const __m128i*)&inB[BPS * 0]); + const __m128i inB_1 = _mm_loadu_si128((const __m128i*)&inB[BPS * 1]); + const __m128i inB_2 = _mm_loadu_si128((const __m128i*)&inB[BPS * 2]); + const __m128i inB_3 = _mm_loadl_epi64((const __m128i*)&inB[BPS * 3]); + + // Combine inA and inB (we'll do two transforms in parallel). + const __m128i inAB_0 = _mm_unpacklo_epi32(inA_0, inB_0); + const __m128i inAB_1 = _mm_unpacklo_epi32(inA_1, inB_1); + const __m128i inAB_2 = _mm_unpacklo_epi32(inA_2, inB_2); + const __m128i inAB_3 = _mm_unpacklo_epi32(inA_3, inB_3); + tmp_0 = _mm_cvtepu8_epi16(inAB_0); + tmp_1 = _mm_cvtepu8_epi16(inAB_1); + tmp_2 = _mm_cvtepu8_epi16(inAB_2); + tmp_3 = _mm_cvtepu8_epi16(inAB_3); + // a00 a01 a02 a03 b00 b01 b02 b03 + // a10 a11 a12 a13 b10 b11 b12 b13 + // a20 a21 a22 a23 b20 b21 b22 b23 + // a30 a31 a32 a33 b30 b31 b32 b33 + } + + // Vertical pass first to avoid a transpose (vertical and horizontal passes + // are commutative because w/kWeightY is symmetric) and subsequent transpose. + { + // Calculate a and b (two 4x4 at once). + const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); + const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); + const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); + const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); + const __m128i b0 = _mm_add_epi16(a0, a1); + const __m128i b1 = _mm_add_epi16(a3, a2); + const __m128i b2 = _mm_sub_epi16(a3, a2); + const __m128i b3 = _mm_sub_epi16(a0, a1); + // a00 a01 a02 a03 b00 b01 b02 b03 + // a10 a11 a12 a13 b10 b11 b12 b13 + // a20 a21 a22 a23 b20 b21 b22 b23 + // a30 a31 a32 a33 b30 b31 b32 b33 + + // Transpose the two 4x4. + VP8Transpose_2_4x4_16b(&b0, &b1, &b2, &b3, &tmp_0, &tmp_1, &tmp_2, &tmp_3); + } + + // Horizontal pass and difference of weighted sums. + { + // Load all inputs. + const __m128i w_0 = _mm_loadu_si128((const __m128i*)&w[0]); + const __m128i w_8 = _mm_loadu_si128((const __m128i*)&w[8]); + + // Calculate a and b (two 4x4 at once). + const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2); + const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3); + const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3); + const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2); + const __m128i b0 = _mm_add_epi16(a0, a1); + const __m128i b1 = _mm_add_epi16(a3, a2); + const __m128i b2 = _mm_sub_epi16(a3, a2); + const __m128i b3 = _mm_sub_epi16(a0, a1); + + // Separate the transforms of inA and inB. + __m128i A_b0 = _mm_unpacklo_epi64(b0, b1); + __m128i A_b2 = _mm_unpacklo_epi64(b2, b3); + __m128i B_b0 = _mm_unpackhi_epi64(b0, b1); + __m128i B_b2 = _mm_unpackhi_epi64(b2, b3); + + A_b0 = _mm_abs_epi16(A_b0); + A_b2 = _mm_abs_epi16(A_b2); + B_b0 = _mm_abs_epi16(B_b0); + B_b2 = _mm_abs_epi16(B_b2); + + // weighted sums + A_b0 = _mm_madd_epi16(A_b0, w_0); + A_b2 = _mm_madd_epi16(A_b2, w_8); + B_b0 = _mm_madd_epi16(B_b0, w_0); + B_b2 = _mm_madd_epi16(B_b2, w_8); + A_b0 = _mm_add_epi32(A_b0, A_b2); + B_b0 = _mm_add_epi32(B_b0, B_b2); + + // difference of weighted sums + A_b2 = _mm_sub_epi32(A_b0, B_b0); + _mm_storeu_si128((__m128i*)&sum[0], A_b2); + } + return sum[0] + sum[1] + sum[2] + sum[3]; +} + +static int Disto4x4_SSE41(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + const int diff_sum = TTransform_SSE41(a, b, w); + return abs(diff_sum) >> 5; +} + +static int Disto16x16_SSE41(const uint8_t* const a, const uint8_t* const b, + const uint16_t* const w) { + int D = 0; + int x, y; + for (y = 0; y < 16 * BPS; y += 4 * BPS) { + for (x = 0; x < 16; x += 4) { + D += Disto4x4_SSE41(a + x + y, b + x + y, w); + } + } + return D; +} + +//------------------------------------------------------------------------------ +// Quantization +// + +// Generates a pshufb constant for shuffling 16b words. +#define PSHUFB_CST(A,B,C,D,E,F,G,H) \ + _mm_set_epi8(2 * (H) + 1, 2 * (H) + 0, 2 * (G) + 1, 2 * (G) + 0, \ + 2 * (F) + 1, 2 * (F) + 0, 2 * (E) + 1, 2 * (E) + 0, \ + 2 * (D) + 1, 2 * (D) + 0, 2 * (C) + 1, 2 * (C) + 0, \ + 2 * (B) + 1, 2 * (B) + 0, 2 * (A) + 1, 2 * (A) + 0) + +static WEBP_INLINE int DoQuantizeBlock_SSE41(int16_t in[16], int16_t out[16], + const uint16_t* const sharpen, + const VP8Matrix* const mtx) { + const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL); + const __m128i zero = _mm_setzero_si128(); + __m128i out0, out8; + __m128i packed_out; + + // Load all inputs. + __m128i in0 = _mm_loadu_si128((__m128i*)&in[0]); + __m128i in8 = _mm_loadu_si128((__m128i*)&in[8]); + const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]); + const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]); + const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]); + const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]); + + // coeff = abs(in) + __m128i coeff0 = _mm_abs_epi16(in0); + __m128i coeff8 = _mm_abs_epi16(in8); + + // coeff = abs(in) + sharpen + if (sharpen != NULL) { + const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]); + const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]); + coeff0 = _mm_add_epi16(coeff0, sharpen0); + coeff8 = _mm_add_epi16(coeff8, sharpen8); + } + + // out = (coeff * iQ + B) >> QFIX + { + // doing calculations with 32b precision (QFIX=17) + // out = (coeff * iQ) + const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0); + const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0); + const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8); + const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8); + __m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H); + __m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H); + __m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H); + __m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H); + // out = (coeff * iQ + B) + const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]); + const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]); + const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]); + const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]); + out_00 = _mm_add_epi32(out_00, bias_00); + out_04 = _mm_add_epi32(out_04, bias_04); + out_08 = _mm_add_epi32(out_08, bias_08); + out_12 = _mm_add_epi32(out_12, bias_12); + // out = QUANTDIV(coeff, iQ, B, QFIX) + out_00 = _mm_srai_epi32(out_00, QFIX); + out_04 = _mm_srai_epi32(out_04, QFIX); + out_08 = _mm_srai_epi32(out_08, QFIX); + out_12 = _mm_srai_epi32(out_12, QFIX); + + // pack result as 16b + out0 = _mm_packs_epi32(out_00, out_04); + out8 = _mm_packs_epi32(out_08, out_12); + + // if (coeff > 2047) coeff = 2047 + out0 = _mm_min_epi16(out0, max_coeff_2047); + out8 = _mm_min_epi16(out8, max_coeff_2047); + } + + // put sign back + out0 = _mm_sign_epi16(out0, in0); + out8 = _mm_sign_epi16(out8, in8); + + // in = out * Q + in0 = _mm_mullo_epi16(out0, q0); + in8 = _mm_mullo_epi16(out8, q8); + + _mm_storeu_si128((__m128i*)&in[0], in0); + _mm_storeu_si128((__m128i*)&in[8], in8); + + // zigzag the output before storing it. The re-ordering is: + // 0 1 2 3 4 5 6 7 | 8 9 10 11 12 13 14 15 + // -> 0 1 4[8]5 2 3 6 | 9 12 13 10 [7]11 14 15 + // There's only two misplaced entries ([8] and [7]) that are crossing the + // reg's boundaries. + // We use pshufb instead of pshuflo/pshufhi. + { + const __m128i kCst_lo = PSHUFB_CST(0, 1, 4, -1, 5, 2, 3, 6); + const __m128i kCst_7 = PSHUFB_CST(-1, -1, -1, -1, 7, -1, -1, -1); + const __m128i tmp_lo = _mm_shuffle_epi8(out0, kCst_lo); + const __m128i tmp_7 = _mm_shuffle_epi8(out0, kCst_7); // extract #7 + const __m128i kCst_hi = PSHUFB_CST(1, 4, 5, 2, -1, 3, 6, 7); + const __m128i kCst_8 = PSHUFB_CST(-1, -1, -1, 0, -1, -1, -1, -1); + const __m128i tmp_hi = _mm_shuffle_epi8(out8, kCst_hi); + const __m128i tmp_8 = _mm_shuffle_epi8(out8, kCst_8); // extract #8 + const __m128i out_z0 = _mm_or_si128(tmp_lo, tmp_8); + const __m128i out_z8 = _mm_or_si128(tmp_hi, tmp_7); + _mm_storeu_si128((__m128i*)&out[0], out_z0); + _mm_storeu_si128((__m128i*)&out[8], out_z8); + packed_out = _mm_packs_epi16(out_z0, out_z8); + } + + // detect if all 'out' values are zeroes or not + return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff); +} + +#undef PSHUFB_CST + +static int QuantizeBlock_SSE41(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock_SSE41(in, out, &mtx->sharpen_[0], mtx); +} + +static int QuantizeBlockWHT_SSE41(int16_t in[16], int16_t out[16], + const VP8Matrix* const mtx) { + return DoQuantizeBlock_SSE41(in, out, NULL, mtx); +} + +static int Quantize2Blocks_SSE41(int16_t in[32], int16_t out[32], + const VP8Matrix* const mtx) { + int nz; + const uint16_t* const sharpen = &mtx->sharpen_[0]; + nz = DoQuantizeBlock_SSE41(in + 0 * 16, out + 0 * 16, sharpen, mtx) << 0; + nz |= DoQuantizeBlock_SSE41(in + 1 * 16, out + 1 * 16, sharpen, mtx) << 1; + return nz; +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8EncDspInitSSE41(void); +WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInitSSE41(void) { + VP8CollectHistogram = CollectHistogram_SSE41; + VP8EncQuantizeBlock = QuantizeBlock_SSE41; + VP8EncQuantize2Blocks = Quantize2Blocks_SSE41; + VP8EncQuantizeBlockWHT = QuantizeBlockWHT_SSE41; + VP8TDisto4x4 = Disto4x4_SSE41; + VP8TDisto16x16 = Disto16x16_SSE41; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(VP8EncDspInitSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/media/libwebp/dsp/filters.c b/media/libwebp/dsp/filters.c index dea3eb4101..b0c659478f 100644 --- a/media/libwebp/dsp/filters.c +++ b/media/libwebp/dsp/filters.c @@ -33,9 +33,9 @@ static WEBP_INLINE void PredictLine_C(const uint8_t* src, const uint8_t* pred, uint8_t* dst, int length, int inverse) { int i; if (inverse) { - for (i = 0; i < length; ++i) dst[i] = src[i] + pred[i]; + for (i = 0; i < length; ++i) dst[i] = (uint8_t)(src[i] + pred[i]); } else { - for (i = 0; i < length; ++i) dst[i] = src[i] - pred[i]; + for (i = 0; i < length; ++i) dst[i] = (uint8_t)(src[i] - pred[i]); } } @@ -155,7 +155,7 @@ static WEBP_INLINE void DoGradientFilter_C(const uint8_t* in, const int pred = GradientPredictor_C(preds[w - 1], preds[w - stride], preds[w - stride - 1]); - out[w] = in[w] + (inverse ? pred : -pred); + out[w] = (uint8_t)(in[w] + (inverse ? pred : -pred)); } ++row; preds += stride; @@ -194,7 +194,7 @@ static void HorizontalUnfilter_C(const uint8_t* prev, const uint8_t* in, uint8_t pred = (prev == NULL) ? 0 : prev[0]; int i; for (i = 0; i < width; ++i) { - out[i] = pred + in[i]; + out[i] = (uint8_t)(pred + in[i]); pred = out[i]; } } @@ -206,7 +206,7 @@ static void VerticalUnfilter_C(const uint8_t* prev, const uint8_t* in, HorizontalUnfilter_C(NULL, in, out, width); } else { int i; - for (i = 0; i < width; ++i) out[i] = prev[i] + in[i]; + for (i = 0; i < width; ++i) out[i] = (uint8_t)(prev[i] + in[i]); } } #endif // !WEBP_NEON_OMIT_C_CODE @@ -220,7 +220,7 @@ static void GradientUnfilter_C(const uint8_t* prev, const uint8_t* in, int i; for (i = 0; i < width; ++i) { top = prev[i]; // need to read this first, in case prev==out - left = in[i] + GradientPredictor_C(left, top, top_left); + left = (uint8_t)(in[i] + GradientPredictor_C(left, top, top_left)); top_left = top; out[i] = left; } @@ -254,7 +254,7 @@ WEBP_DSP_INIT_FUNC(VP8FiltersInit) { #endif if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8FiltersInitSSE2(); } @@ -271,7 +271,7 @@ WEBP_DSP_INIT_FUNC(VP8FiltersInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8FiltersInitNEON(); diff --git a/media/libwebp/dsp/filters_mips_dsp_r2.c b/media/libwebp/dsp/filters_mips_dsp_r2.c new file mode 100644 index 0000000000..edb1eaac26 --- /dev/null +++ b/media/libwebp/dsp/filters_mips_dsp_r2.c @@ -0,0 +1,402 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Spatial prediction using various filters +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../dsp/dsp.h" +#include <assert.h> +#include <stdlib.h> +#include <string.h> + +//------------------------------------------------------------------------------ +// Helpful macro. + +# define SANITY_CHECK(in, out) \ + assert(in != NULL); \ + assert(out != NULL); \ + assert(width > 0); \ + assert(height > 0); \ + assert(stride >= width); \ + assert(row >= 0 && num_rows > 0 && row + num_rows <= height); \ + (void)height; // Silence unused warning. + +#define DO_PREDICT_LINE(SRC, DST, LENGTH, INVERSE) do { \ + const uint8_t* psrc = (uint8_t*)(SRC); \ + uint8_t* pdst = (uint8_t*)(DST); \ + const int ilength = (int)(LENGTH); \ + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; \ + __asm__ volatile ( \ + ".set push \n\t" \ + ".set noreorder \n\t" \ + "srl %[temp0], %[length], 2 \n\t" \ + "beqz %[temp0], 4f \n\t" \ + " andi %[temp6], %[length], 3 \n\t" \ + ".if " #INVERSE " \n\t" \ + "1: \n\t" \ + "lbu %[temp1], -1(%[dst]) \n\t" \ + "lbu %[temp2], 0(%[src]) \n\t" \ + "lbu %[temp3], 1(%[src]) \n\t" \ + "lbu %[temp4], 2(%[src]) \n\t" \ + "lbu %[temp5], 3(%[src]) \n\t" \ + "addu %[temp1], %[temp1], %[temp2] \n\t" \ + "addu %[temp2], %[temp1], %[temp3] \n\t" \ + "addu %[temp3], %[temp2], %[temp4] \n\t" \ + "addu %[temp4], %[temp3], %[temp5] \n\t" \ + "sb %[temp1], 0(%[dst]) \n\t" \ + "sb %[temp2], 1(%[dst]) \n\t" \ + "sb %[temp3], 2(%[dst]) \n\t" \ + "sb %[temp4], 3(%[dst]) \n\t" \ + "addiu %[src], %[src], 4 \n\t" \ + "addiu %[temp0], %[temp0], -1 \n\t" \ + "bnez %[temp0], 1b \n\t" \ + " addiu %[dst], %[dst], 4 \n\t" \ + ".else \n\t" \ + "1: \n\t" \ + "ulw %[temp1], -1(%[src]) \n\t" \ + "ulw %[temp2], 0(%[src]) \n\t" \ + "addiu %[src], %[src], 4 \n\t" \ + "addiu %[temp0], %[temp0], -1 \n\t" \ + "subu.qb %[temp3], %[temp2], %[temp1] \n\t" \ + "usw %[temp3], 0(%[dst]) \n\t" \ + "bnez %[temp0], 1b \n\t" \ + " addiu %[dst], %[dst], 4 \n\t" \ + ".endif \n\t" \ + "4: \n\t" \ + "beqz %[temp6], 3f \n\t" \ + " nop \n\t" \ + "2: \n\t" \ + "lbu %[temp2], 0(%[src]) \n\t" \ + ".if " #INVERSE " \n\t" \ + "lbu %[temp1], -1(%[dst]) \n\t" \ + "addu %[temp3], %[temp1], %[temp2] \n\t" \ + ".else \n\t" \ + "lbu %[temp1], -1(%[src]) \n\t" \ + "subu %[temp3], %[temp1], %[temp2] \n\t" \ + ".endif \n\t" \ + "addiu %[src], %[src], 1 \n\t" \ + "sb %[temp3], 0(%[dst]) \n\t" \ + "addiu %[temp6], %[temp6], -1 \n\t" \ + "bnez %[temp6], 2b \n\t" \ + " addiu %[dst], %[dst], 1 \n\t" \ + "3: \n\t" \ + ".set pop \n\t" \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [dst]"+&r"(pdst), [src]"+&r"(psrc) \ + : [length]"r"(ilength) \ + : "memory" \ + ); \ + } while (0) + +static WEBP_INLINE void PredictLine_MIPSdspR2(const uint8_t* src, uint8_t* dst, + int length) { + DO_PREDICT_LINE(src, dst, length, 0); +} + +#define DO_PREDICT_LINE_VERTICAL(SRC, PRED, DST, LENGTH, INVERSE) do { \ + const uint8_t* psrc = (uint8_t*)(SRC); \ + const uint8_t* ppred = (uint8_t*)(PRED); \ + uint8_t* pdst = (uint8_t*)(DST); \ + const int ilength = (int)(LENGTH); \ + int temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; \ + __asm__ volatile ( \ + ".set push \n\t" \ + ".set noreorder \n\t" \ + "srl %[temp0], %[length], 0x3 \n\t" \ + "beqz %[temp0], 4f \n\t" \ + " andi %[temp7], %[length], 0x7 \n\t" \ + "1: \n\t" \ + "ulw %[temp1], 0(%[src]) \n\t" \ + "ulw %[temp2], 0(%[pred]) \n\t" \ + "ulw %[temp3], 4(%[src]) \n\t" \ + "ulw %[temp4], 4(%[pred]) \n\t" \ + "addiu %[src], %[src], 8 \n\t" \ + ".if " #INVERSE " \n\t" \ + "addu.qb %[temp5], %[temp1], %[temp2] \n\t" \ + "addu.qb %[temp6], %[temp3], %[temp4] \n\t" \ + ".else \n\t" \ + "subu.qb %[temp5], %[temp1], %[temp2] \n\t" \ + "subu.qb %[temp6], %[temp3], %[temp4] \n\t" \ + ".endif \n\t" \ + "addiu %[pred], %[pred], 8 \n\t" \ + "usw %[temp5], 0(%[dst]) \n\t" \ + "usw %[temp6], 4(%[dst]) \n\t" \ + "addiu %[temp0], %[temp0], -1 \n\t" \ + "bnez %[temp0], 1b \n\t" \ + " addiu %[dst], %[dst], 8 \n\t" \ + "4: \n\t" \ + "beqz %[temp7], 3f \n\t" \ + " nop \n\t" \ + "2: \n\t" \ + "lbu %[temp1], 0(%[src]) \n\t" \ + "lbu %[temp2], 0(%[pred]) \n\t" \ + "addiu %[src], %[src], 1 \n\t" \ + "addiu %[pred], %[pred], 1 \n\t" \ + ".if " #INVERSE " \n\t" \ + "addu %[temp3], %[temp1], %[temp2] \n\t" \ + ".else \n\t" \ + "subu %[temp3], %[temp1], %[temp2] \n\t" \ + ".endif \n\t" \ + "sb %[temp3], 0(%[dst]) \n\t" \ + "addiu %[temp7], %[temp7], -1 \n\t" \ + "bnez %[temp7], 2b \n\t" \ + " addiu %[dst], %[dst], 1 \n\t" \ + "3: \n\t" \ + ".set pop \n\t" \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), [pred]"+&r"(ppred), \ + [dst]"+&r"(pdst), [src]"+&r"(psrc) \ + : [length]"r"(ilength) \ + : "memory" \ + ); \ + } while (0) + +#define PREDICT_LINE_ONE_PASS(SRC, PRED, DST) do { \ + int temp1, temp2, temp3; \ + __asm__ volatile ( \ + "lbu %[temp1], 0(%[src]) \n\t" \ + "lbu %[temp2], 0(%[pred]) \n\t" \ + "subu %[temp3], %[temp1], %[temp2] \n\t" \ + "sb %[temp3], 0(%[dst]) \n\t" \ + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3) \ + : [pred]"r"((PRED)), [dst]"r"((DST)), [src]"r"((SRC)) \ + : "memory" \ + ); \ + } while (0) + +//------------------------------------------------------------------------------ +// Horizontal filter. + +#define FILTER_LINE_BY_LINE do { \ + while (row < last_row) { \ + PREDICT_LINE_ONE_PASS(in, preds - stride, out); \ + DO_PREDICT_LINE(in + 1, out + 1, width - 1, 0); \ + ++row; \ + preds += stride; \ + in += stride; \ + out += stride; \ + } \ + } while (0) + +static WEBP_INLINE void DoHorizontalFilter_MIPSdspR2(const uint8_t* in, + int width, int height, + int stride, + int row, int num_rows, + uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = in; + + if (row == 0) { + // Leftmost pixel is the same as input for topmost scanline. + out[0] = in[0]; + PredictLine_MIPSdspR2(in + 1, out + 1, width - 1); + row = 1; + preds += stride; + in += stride; + out += stride; + } + + // Filter line-by-line. + FILTER_LINE_BY_LINE; +} +#undef FILTER_LINE_BY_LINE + +static void HorizontalFilter_MIPSdspR2(const uint8_t* data, + int width, int height, + int stride, uint8_t* filtered_data) { + DoHorizontalFilter_MIPSdspR2(data, width, height, stride, 0, height, + filtered_data); +} + +//------------------------------------------------------------------------------ +// Vertical filter. + +#define FILTER_LINE_BY_LINE do { \ + while (row < last_row) { \ + DO_PREDICT_LINE_VERTICAL(in, preds, out, width, 0); \ + ++row; \ + preds += stride; \ + in += stride; \ + out += stride; \ + } \ + } while (0) + +static WEBP_INLINE void DoVerticalFilter_MIPSdspR2(const uint8_t* in, + int width, int height, + int stride, + int row, int num_rows, + uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = in; + + if (row == 0) { + // Very first top-left pixel is copied. + out[0] = in[0]; + // Rest of top scan-line is left-predicted. + PredictLine_MIPSdspR2(in + 1, out + 1, width - 1); + row = 1; + in += stride; + out += stride; + } else { + // We are starting from in-between. Make sure 'preds' points to prev row. + preds -= stride; + } + + // Filter line-by-line. + FILTER_LINE_BY_LINE; +} +#undef FILTER_LINE_BY_LINE + +static void VerticalFilter_MIPSdspR2(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoVerticalFilter_MIPSdspR2(data, width, height, stride, 0, height, + filtered_data); +} + +//------------------------------------------------------------------------------ +// Gradient filter. + +static int GradientPredictor_MIPSdspR2(uint8_t a, uint8_t b, uint8_t c) { + int temp0; + __asm__ volatile ( + "addu %[temp0], %[a], %[b] \n\t" + "subu %[temp0], %[temp0], %[c] \n\t" + "shll_s.w %[temp0], %[temp0], 23 \n\t" + "precrqu_s.qb.ph %[temp0], %[temp0], $zero \n\t" + "srl %[temp0], %[temp0], 24 \n\t" + : [temp0]"=&r"(temp0) + : [a]"r"(a),[b]"r"(b),[c]"r"(c) + ); + return temp0; +} + +#define FILTER_LINE_BY_LINE(PREDS, OPERATION) do { \ + while (row < last_row) { \ + int w; \ + PREDICT_LINE_ONE_PASS(in, PREDS - stride, out); \ + for (w = 1; w < width; ++w) { \ + const int pred = GradientPredictor_MIPSdspR2(PREDS[w - 1], \ + PREDS[w - stride], \ + PREDS[w - stride - 1]); \ + out[w] = in[w] OPERATION pred; \ + } \ + ++row; \ + in += stride; \ + out += stride; \ + } \ + } while (0) + +static void DoGradientFilter_MIPSdspR2(const uint8_t* in, + int width, int height, int stride, + int row, int num_rows, uint8_t* out) { + const uint8_t* preds; + const size_t start_offset = row * stride; + const int last_row = row + num_rows; + SANITY_CHECK(in, out); + in += start_offset; + out += start_offset; + preds = in; + + // left prediction for top scan-line + if (row == 0) { + out[0] = in[0]; + PredictLine_MIPSdspR2(in + 1, out + 1, width - 1); + row = 1; + preds += stride; + in += stride; + out += stride; + } + + // Filter line-by-line. + FILTER_LINE_BY_LINE(in, -); +} +#undef FILTER_LINE_BY_LINE + +static void GradientFilter_MIPSdspR2(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + DoGradientFilter_MIPSdspR2(data, width, height, stride, 0, height, + filtered_data); +} + +//------------------------------------------------------------------------------ + +static void HorizontalUnfilter_MIPSdspR2(const uint8_t* prev, const uint8_t* in, + uint8_t* out, int width) { + out[0] = in[0] + (prev == NULL ? 0 : prev[0]); + DO_PREDICT_LINE(in + 1, out + 1, width - 1, 1); +} + +static void VerticalUnfilter_MIPSdspR2(const uint8_t* prev, const uint8_t* in, + uint8_t* out, int width) { + if (prev == NULL) { + HorizontalUnfilter_MIPSdspR2(NULL, in, out, width); + } else { + DO_PREDICT_LINE_VERTICAL(in, prev, out, width, 1); + } +} + +static void GradientUnfilter_MIPSdspR2(const uint8_t* prev, const uint8_t* in, + uint8_t* out, int width) { + if (prev == NULL) { + HorizontalUnfilter_MIPSdspR2(NULL, in, out, width); + } else { + uint8_t top = prev[0], top_left = top, left = top; + int i; + for (i = 0; i < width; ++i) { + top = prev[i]; // need to read this first, in case prev==dst + left = in[i] + GradientPredictor_MIPSdspR2(left, top, top_left); + top_left = top; + out[i] = left; + } + } +} + +#undef DO_PREDICT_LINE_VERTICAL +#undef PREDICT_LINE_ONE_PASS +#undef DO_PREDICT_LINE +#undef SANITY_CHECK + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8FiltersInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitMIPSdspR2(void) { + WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_MIPSdspR2; + WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_MIPSdspR2; + WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_MIPSdspR2; + + WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_MIPSdspR2; + WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter_MIPSdspR2; + WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8FiltersInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/filters_msa.c b/media/libwebp/dsp/filters_msa.c new file mode 100644 index 0000000000..cd32cdabaf --- /dev/null +++ b/media/libwebp/dsp/filters_msa.c @@ -0,0 +1,202 @@ +// Copyright 2016 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MSA variant of alpha filters +// +// Author: Prashant Patil (prashant.patil@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) + +#include "../dsp/msa_macro.h" + +#include <assert.h> + +static WEBP_INLINE void PredictLineInverse0(const uint8_t* src, + const uint8_t* pred, + uint8_t* dst, int length) { + v16u8 src0, pred0, dst0; + assert(length >= 0); + while (length >= 32) { + v16u8 src1, pred1, dst1; + LD_UB2(src, 16, src0, src1); + LD_UB2(pred, 16, pred0, pred1); + SUB2(src0, pred0, src1, pred1, dst0, dst1); + ST_UB2(dst0, dst1, dst, 16); + src += 32; + pred += 32; + dst += 32; + length -= 32; + } + if (length > 0) { + int i; + if (length >= 16) { + src0 = LD_UB(src); + pred0 = LD_UB(pred); + dst0 = src0 - pred0; + ST_UB(dst0, dst); + src += 16; + pred += 16; + dst += 16; + length -= 16; + } + for (i = 0; i < length; i++) { + dst[i] = src[i] - pred[i]; + } + } +} + +//------------------------------------------------------------------------------ +// Helpful macro. + +#define SANITY_CHECK(in, out) \ + assert(in != NULL); \ + assert(out != NULL); \ + assert(width > 0); \ + assert(height > 0); \ + assert(stride >= width); + +//------------------------------------------------------------------------------ +// Horrizontal filter + +static void HorizontalFilter_MSA(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + const uint8_t* preds = data; + const uint8_t* in = data; + uint8_t* out = filtered_data; + int row = 1; + SANITY_CHECK(in, out); + + // Leftmost pixel is the same as input for topmost scanline. + out[0] = in[0]; + PredictLineInverse0(in + 1, preds, out + 1, width - 1); + preds += stride; + in += stride; + out += stride; + // Filter line-by-line. + while (row < height) { + // Leftmost pixel is predicted from above. + PredictLineInverse0(in, preds - stride, out, 1); + PredictLineInverse0(in + 1, preds, out + 1, width - 1); + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +//------------------------------------------------------------------------------ +// Gradient filter + +static WEBP_INLINE void PredictLineGradient(const uint8_t* pinput, + const uint8_t* ppred, + uint8_t* poutput, int stride, + int size) { + int w; + const v16i8 zero = { 0 }; + while (size >= 16) { + v16u8 pred0, dst0; + v8i16 a0, a1, b0, b1, c0, c1; + const v16u8 tmp0 = LD_UB(ppred - 1); + const v16u8 tmp1 = LD_UB(ppred - stride); + const v16u8 tmp2 = LD_UB(ppred - stride - 1); + const v16u8 src0 = LD_UB(pinput); + ILVRL_B2_SH(zero, tmp0, a0, a1); + ILVRL_B2_SH(zero, tmp1, b0, b1); + ILVRL_B2_SH(zero, tmp2, c0, c1); + ADD2(a0, b0, a1, b1, a0, a1); + SUB2(a0, c0, a1, c1, a0, a1); + CLIP_SH2_0_255(a0, a1); + pred0 = (v16u8)__msa_pckev_b((v16i8)a1, (v16i8)a0); + dst0 = src0 - pred0; + ST_UB(dst0, poutput); + ppred += 16; + pinput += 16; + poutput += 16; + size -= 16; + } + for (w = 0; w < size; ++w) { + const int pred = ppred[w - 1] + ppred[w - stride] - ppred[w - stride - 1]; + poutput[w] = pinput[w] - (pred < 0 ? 0 : pred > 255 ? 255 : pred); + } +} + + +static void GradientFilter_MSA(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + const uint8_t* in = data; + const uint8_t* preds = data; + uint8_t* out = filtered_data; + int row = 1; + SANITY_CHECK(in, out); + + // left prediction for top scan-line + out[0] = in[0]; + PredictLineInverse0(in + 1, preds, out + 1, width - 1); + preds += stride; + in += stride; + out += stride; + // Filter line-by-line. + while (row < height) { + out[0] = in[0] - preds[- stride]; + PredictLineGradient(preds + 1, in + 1, out + 1, stride, width - 1); + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +//------------------------------------------------------------------------------ +// Vertical filter + +static void VerticalFilter_MSA(const uint8_t* data, int width, int height, + int stride, uint8_t* filtered_data) { + const uint8_t* in = data; + const uint8_t* preds = data; + uint8_t* out = filtered_data; + int row = 1; + SANITY_CHECK(in, out); + + // Very first top-left pixel is copied. + out[0] = in[0]; + // Rest of top scan-line is left-predicted. + PredictLineInverse0(in + 1, preds, out + 1, width - 1); + in += stride; + out += stride; + + // Filter line-by-line. + while (row < height) { + PredictLineInverse0(in, preds, out, width); + ++row; + preds += stride; + in += stride; + out += stride; + } +} + +#undef SANITY_CHECK + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8FiltersInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitMSA(void) { + WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_MSA; + WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter_MSA; + WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter_MSA; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(VP8FiltersInitMSA) + +#endif // WEBP_USE_MSA diff --git a/media/libwebp/dsp/filters_sse2.c b/media/libwebp/dsp/filters_sse2.c index 2cc9bb9766..9b91ab680f 100644 --- a/media/libwebp/dsp/filters_sse2.c +++ b/media/libwebp/dsp/filters_sse2.c @@ -163,7 +163,8 @@ static void GradientPredictDirect_SSE2(const uint8_t* const row, _mm_storel_epi64((__m128i*)(out + i), H); } for (; i < length; ++i) { - out[i] = row[i] - GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]); + const int delta = GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]); + out[i] = (uint8_t)(row[i] - delta); } } @@ -188,7 +189,7 @@ static WEBP_INLINE void DoGradientFilter_SSE2(const uint8_t* in, // Filter line-by-line. while (row < last_row) { - out[0] = in[0] - in[-stride]; + out[0] = (uint8_t)(in[0] - in[-stride]); GradientPredictDirect_SSE2(in + 1, in + 1 - stride, out + 1, width - 1); ++row; in += stride; @@ -223,7 +224,7 @@ static void HorizontalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in, uint8_t* out, int width) { int i; __m128i last; - out[0] = in[0] + (prev == NULL ? 0 : prev[0]); + out[0] = (uint8_t)(in[0] + (prev == NULL ? 0 : prev[0])); if (width <= 1) return; last = _mm_set_epi32(0, 0, 0, out[0]); for (i = 1; i + 8 <= width; i += 8) { @@ -238,7 +239,7 @@ static void HorizontalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in, _mm_storel_epi64((__m128i*)(out + i), A7); last = _mm_srli_epi64(A7, 56); } - for (; i < width; ++i) out[i] = in[i] + out[i - 1]; + for (; i < width; ++i) out[i] = (uint8_t)(in[i] + out[i - 1]); } static void VerticalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in, @@ -259,7 +260,7 @@ static void VerticalUnfilter_SSE2(const uint8_t* prev, const uint8_t* in, _mm_storeu_si128((__m128i*)&out[i + 0], C0); _mm_storeu_si128((__m128i*)&out[i + 16], C1); } - for (; i < width; ++i) out[i] = in[i] + prev[i]; + for (; i < width; ++i) out[i] = (uint8_t)(in[i] + prev[i]); } } @@ -296,7 +297,8 @@ static void GradientPredictInverse_SSE2(const uint8_t* const in, _mm_storel_epi64((__m128i*)&row[i], out); } for (; i < length; ++i) { - row[i] = in[i] + GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]); + const int delta = GradientPredictor_SSE2(row[i - 1], top[i], top[i - 1]); + row[i] = (uint8_t)(in[i] + delta); } } } @@ -306,7 +308,7 @@ static void GradientUnfilter_SSE2(const uint8_t* prev, const uint8_t* in, if (prev == NULL) { HorizontalUnfilter_SSE2(NULL, in, out, width); } else { - out[0] = in[0] + prev[0]; // predict from above + out[0] = (uint8_t)(in[0] + prev[0]); // predict from above GradientPredictInverse_SSE2(in + 1, prev + 1, out + 1, width - 1); } } @@ -318,7 +320,12 @@ extern void VP8FiltersInitSSE2(void); WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitSSE2(void) { WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_SSE2; +#if defined(CHROMIUM) + // TODO(crbug.com/654974) + (void)VerticalUnfilter_SSE2; +#else WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_SSE2; +#endif WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_SSE2; WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_SSE2; diff --git a/media/libwebp/dsp/lossless.c b/media/libwebp/dsp/lossless.c index 1a1523d221..763c425ff8 100644 --- a/media/libwebp/dsp/lossless.c +++ b/media/libwebp/dsp/lossless.c @@ -81,7 +81,7 @@ static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, // gcc <= 4.9 on ARM generates incorrect code in Select() when Sub3() is // inlined. -#if defined(__arm__) && LOCAL_GCC_VERSION <= 0x409 +#if defined(__arm__) && defined(__GNUC__) && LOCAL_GCC_VERSION <= 0x409 # define LOCAL_INLINE __attribute__ ((noinline)) #else # define LOCAL_INLINE WEBP_INLINE @@ -107,88 +107,107 @@ static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { //------------------------------------------------------------------------------ // Predictors -static uint32_t Predictor0_C(uint32_t left, const uint32_t* const top) { +uint32_t VP8LPredictor0_C(const uint32_t* const left, + const uint32_t* const top) { (void)top; (void)left; return ARGB_BLACK; } -static uint32_t Predictor1_C(uint32_t left, const uint32_t* const top) { +uint32_t VP8LPredictor1_C(const uint32_t* const left, + const uint32_t* const top) { (void)top; - return left; + return *left; } -static uint32_t Predictor2_C(uint32_t left, const uint32_t* const top) { +uint32_t VP8LPredictor2_C(const uint32_t* const left, + const uint32_t* const top) { (void)left; return top[0]; } -static uint32_t Predictor3_C(uint32_t left, const uint32_t* const top) { +uint32_t VP8LPredictor3_C(const uint32_t* const left, + const uint32_t* const top) { (void)left; return top[1]; } -static uint32_t Predictor4_C(uint32_t left, const uint32_t* const top) { +uint32_t VP8LPredictor4_C(const uint32_t* const left, + const uint32_t* const top) { (void)left; return top[-1]; } -static uint32_t Predictor5_C(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average3(left, top[0], top[1]); +uint32_t VP8LPredictor5_C(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average3(*left, top[0], top[1]); return pred; } -static uint32_t Predictor6_C(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average2(left, top[-1]); +uint32_t VP8LPredictor6_C(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average2(*left, top[-1]); return pred; } -static uint32_t Predictor7_C(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average2(left, top[0]); +uint32_t VP8LPredictor7_C(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average2(*left, top[0]); return pred; } -static uint32_t Predictor8_C(uint32_t left, const uint32_t* const top) { +uint32_t VP8LPredictor8_C(const uint32_t* const left, + const uint32_t* const top) { const uint32_t pred = Average2(top[-1], top[0]); (void)left; return pred; } -static uint32_t Predictor9_C(uint32_t left, const uint32_t* const top) { +uint32_t VP8LPredictor9_C(const uint32_t* const left, + const uint32_t* const top) { const uint32_t pred = Average2(top[0], top[1]); (void)left; return pred; } -static uint32_t Predictor10_C(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average4(left, top[-1], top[0], top[1]); +uint32_t VP8LPredictor10_C(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average4(*left, top[-1], top[0], top[1]); return pred; } -static uint32_t Predictor11_C(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Select(top[0], left, top[-1]); +uint32_t VP8LPredictor11_C(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Select(top[0], *left, top[-1]); return pred; } -static uint32_t Predictor12_C(uint32_t left, const uint32_t* const top) { - const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); +uint32_t VP8LPredictor12_C(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractFull(*left, top[0], top[-1]); return pred; } -static uint32_t Predictor13_C(uint32_t left, const uint32_t* const top) { - const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); +uint32_t VP8LPredictor13_C(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractHalf(*left, top[0], top[-1]); return pred; } -GENERATE_PREDICTOR_ADD(Predictor0_C, PredictorAdd0_C) +static void PredictorAdd0_C(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int x; + (void)upper; + for (x = 0; x < num_pixels; ++x) out[x] = VP8LAddPixels(in[x], ARGB_BLACK); +} static void PredictorAdd1_C(const uint32_t* in, const uint32_t* upper, int num_pixels, uint32_t* out) { int i; uint32_t left = out[-1]; + (void)upper; for (i = 0; i < num_pixels; ++i) { out[i] = left = VP8LAddPixels(in[i], left); } - (void)upper; } -GENERATE_PREDICTOR_ADD(Predictor2_C, PredictorAdd2_C) -GENERATE_PREDICTOR_ADD(Predictor3_C, PredictorAdd3_C) -GENERATE_PREDICTOR_ADD(Predictor4_C, PredictorAdd4_C) -GENERATE_PREDICTOR_ADD(Predictor5_C, PredictorAdd5_C) -GENERATE_PREDICTOR_ADD(Predictor6_C, PredictorAdd6_C) -GENERATE_PREDICTOR_ADD(Predictor7_C, PredictorAdd7_C) -GENERATE_PREDICTOR_ADD(Predictor8_C, PredictorAdd8_C) -GENERATE_PREDICTOR_ADD(Predictor9_C, PredictorAdd9_C) -GENERATE_PREDICTOR_ADD(Predictor10_C, PredictorAdd10_C) -GENERATE_PREDICTOR_ADD(Predictor11_C, PredictorAdd11_C) -GENERATE_PREDICTOR_ADD(Predictor12_C, PredictorAdd12_C) -GENERATE_PREDICTOR_ADD(Predictor13_C, PredictorAdd13_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor2_C, PredictorAdd2_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor3_C, PredictorAdd3_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor4_C, PredictorAdd4_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor5_C, PredictorAdd5_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor6_C, PredictorAdd6_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor7_C, PredictorAdd7_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor8_C, PredictorAdd8_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor9_C, PredictorAdd9_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor10_C, PredictorAdd10_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor11_C, PredictorAdd11_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor12_C, PredictorAdd12_C) +GENERATE_PREDICTOR_ADD(VP8LPredictor13_C, PredictorAdd13_C) //------------------------------------------------------------------------------ @@ -270,14 +289,14 @@ void VP8LTransformColorInverse_C(const VP8LMultipliers* const m, int i; for (i = 0; i < num_pixels; ++i) { const uint32_t argb = src[i]; - const uint32_t green = argb >> 8; + const int8_t green = (int8_t)(argb >> 8); const uint32_t red = argb >> 16; int new_red = red & 0xff; int new_blue = argb & 0xff; new_red += ColorTransformDelta(m->green_to_red_, green); new_red &= 0xff; new_blue += ColorTransformDelta(m->green_to_blue_, green); - new_blue += ColorTransformDelta(m->red_to_blue_, new_red); + new_blue += ColorTransformDelta(m->red_to_blue_, (int8_t)new_red); new_blue &= 0xff; dst[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue); } @@ -557,7 +576,6 @@ VP8LPredictorFunc VP8LPredictors[16]; // exposed plain-C implementations VP8LPredictorAddSubFunc VP8LPredictorsAdd_C[16]; -VP8LPredictorFunc VP8LPredictors_C[16]; VP8LTransformColorInverseFunc VP8LTransformColorInverse; @@ -571,6 +589,7 @@ VP8LMapARGBFunc VP8LMapColor32b; VP8LMapAlphaFunc VP8LMapColor8b; extern void VP8LDspInitSSE2(void); +extern void VP8LDspInitSSE41(void); extern void VP8LDspInitNEON(void); extern void VP8LDspInitMIPSdspR2(void); extern void VP8LDspInitMSA(void); @@ -595,8 +614,7 @@ extern void VP8LDspInitMSA(void); } while (0); WEBP_DSP_INIT_FUNC(VP8LDspInit) { - COPY_PREDICTOR_ARRAY(Predictor, VP8LPredictors) - COPY_PREDICTOR_ARRAY(Predictor, VP8LPredictors_C) + COPY_PREDICTOR_ARRAY(VP8LPredictor, VP8LPredictors) COPY_PREDICTOR_ARRAY(PredictorAdd, VP8LPredictorsAdd) COPY_PREDICTOR_ARRAY(PredictorAdd, VP8LPredictorsAdd_C) @@ -618,9 +636,14 @@ WEBP_DSP_INIT_FUNC(VP8LDspInit) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { VP8LDspInitSSE2(); +#if defined(WEBP_HAVE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + VP8LDspInitSSE41(); + } +#endif } #endif #if defined(WEBP_USE_MIPS_DSP_R2) @@ -635,7 +658,7 @@ WEBP_DSP_INIT_FUNC(VP8LDspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { VP8LDspInitNEON(); diff --git a/media/libwebp/dsp/lossless.h b/media/libwebp/dsp/lossless.h index 6db5fafc13..0c129d2860 100644 --- a/media/libwebp/dsp/lossless.h +++ b/media/libwebp/dsp/lossless.h @@ -28,9 +28,39 @@ extern "C" { //------------------------------------------------------------------------------ // Decoding -typedef uint32_t (*VP8LPredictorFunc)(uint32_t left, const uint32_t* const top); +typedef uint32_t (*VP8LPredictorFunc)(const uint32_t* const left, + const uint32_t* const top); extern VP8LPredictorFunc VP8LPredictors[16]; -extern VP8LPredictorFunc VP8LPredictors_C[16]; + +uint32_t VP8LPredictor0_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor1_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor2_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor3_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor4_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor5_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor6_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor7_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor8_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor9_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor10_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor11_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor12_C(const uint32_t* const left, + const uint32_t* const top); +uint32_t VP8LPredictor13_C(const uint32_t* const left, + const uint32_t* const top); + // These Add/Sub function expects upper[-1] and out[-1] to be readable. typedef void (*VP8LPredictorAddSubFunc)(const uint32_t* in, const uint32_t* upper, int num_pixels, diff --git a/media/libwebp/dsp/lossless_common.h b/media/libwebp/dsp/lossless_common.h index dd2e4f247e..2b20637a28 100644 --- a/media/libwebp/dsp/lossless_common.h +++ b/media/libwebp/dsp/lossless_common.h @@ -177,24 +177,13 @@ uint32_t VP8LSubPixels(uint32_t a, uint32_t b) { static void PREDICTOR_ADD(const uint32_t* in, const uint32_t* upper, \ int num_pixels, uint32_t* out) { \ int x; \ + assert(upper != NULL); \ for (x = 0; x < num_pixels; ++x) { \ - const uint32_t pred = (PREDICTOR)(out[x - 1], upper + x); \ + const uint32_t pred = (PREDICTOR)(&out[x - 1], upper + x); \ out[x] = VP8LAddPixels(in[x], pred); \ } \ } -// It subtracts the prediction from the input pixel and stores the residual -// in the output pixel. -#define GENERATE_PREDICTOR_SUB(PREDICTOR, PREDICTOR_SUB) \ -static void PREDICTOR_SUB(const uint32_t* in, const uint32_t* upper, \ - int num_pixels, uint32_t* out) { \ - int x; \ - for (x = 0; x < num_pixels; ++x) { \ - const uint32_t pred = (PREDICTOR)(in[x - 1], upper + x); \ - out[x] = VP8LSubPixels(in[x], pred); \ - } \ -} - #ifdef __cplusplus } // extern "C" #endif diff --git a/media/libwebp/dsp/lossless_enc.c b/media/libwebp/dsp/lossless_enc.c new file mode 100644 index 0000000000..dca5e26be8 --- /dev/null +++ b/media/libwebp/dsp/lossless_enc.c @@ -0,0 +1,948 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Image transform methods for lossless encoder. +// +// Authors: Vikas Arora (vikaas.arora@gmail.com) +// Jyrki Alakuijala (jyrki@google.com) +// Urvang Joshi (urvang@google.com) + +#include "../dsp/dsp.h" + +#include <assert.h> +#include <math.h> +#include <stdlib.h> +#include "../dec/vp8li_dec.h" +#include "../utils/endian_inl_utils.h" +#include "../dsp/lossless.h" +#include "../dsp/lossless_common.h" +#include "../dsp/yuv.h" + +// lookup table for small values of log2(int) +const float kLog2Table[LOG_LOOKUP_IDX_MAX] = { + 0.0000000000000000f, 0.0000000000000000f, + 1.0000000000000000f, 1.5849625007211560f, + 2.0000000000000000f, 2.3219280948873621f, + 2.5849625007211560f, 2.8073549220576041f, + 3.0000000000000000f, 3.1699250014423121f, + 3.3219280948873621f, 3.4594316186372973f, + 3.5849625007211560f, 3.7004397181410921f, + 3.8073549220576041f, 3.9068905956085187f, + 4.0000000000000000f, 4.0874628412503390f, + 4.1699250014423121f, 4.2479275134435852f, + 4.3219280948873626f, 4.3923174227787606f, + 4.4594316186372973f, 4.5235619560570130f, + 4.5849625007211560f, 4.6438561897747243f, + 4.7004397181410917f, 4.7548875021634682f, + 4.8073549220576037f, 4.8579809951275718f, + 4.9068905956085187f, 4.9541963103868749f, + 5.0000000000000000f, 5.0443941193584533f, + 5.0874628412503390f, 5.1292830169449663f, + 5.1699250014423121f, 5.2094533656289501f, + 5.2479275134435852f, 5.2854022188622487f, + 5.3219280948873626f, 5.3575520046180837f, + 5.3923174227787606f, 5.4262647547020979f, + 5.4594316186372973f, 5.4918530963296747f, + 5.5235619560570130f, 5.5545888516776376f, + 5.5849625007211560f, 5.6147098441152083f, + 5.6438561897747243f, 5.6724253419714951f, + 5.7004397181410917f, 5.7279204545631987f, + 5.7548875021634682f, 5.7813597135246599f, + 5.8073549220576037f, 5.8328900141647412f, + 5.8579809951275718f, 5.8826430493618415f, + 5.9068905956085187f, 5.9307373375628866f, + 5.9541963103868749f, 5.9772799234999167f, + 6.0000000000000000f, 6.0223678130284543f, + 6.0443941193584533f, 6.0660891904577720f, + 6.0874628412503390f, 6.1085244567781691f, + 6.1292830169449663f, 6.1497471195046822f, + 6.1699250014423121f, 6.1898245588800175f, + 6.2094533656289501f, 6.2288186904958804f, + 6.2479275134435852f, 6.2667865406949010f, + 6.2854022188622487f, 6.3037807481771030f, + 6.3219280948873626f, 6.3398500028846243f, + 6.3575520046180837f, 6.3750394313469245f, + 6.3923174227787606f, 6.4093909361377017f, + 6.4262647547020979f, 6.4429434958487279f, + 6.4594316186372973f, 6.4757334309663976f, + 6.4918530963296747f, 6.5077946401986963f, + 6.5235619560570130f, 6.5391588111080309f, + 6.5545888516776376f, 6.5698556083309478f, + 6.5849625007211560f, 6.5999128421871278f, + 6.6147098441152083f, 6.6293566200796094f, + 6.6438561897747243f, 6.6582114827517946f, + 6.6724253419714951f, 6.6865005271832185f, + 6.7004397181410917f, 6.7142455176661224f, + 6.7279204545631987f, 6.7414669864011464f, + 6.7548875021634682f, 6.7681843247769259f, + 6.7813597135246599f, 6.7944158663501061f, + 6.8073549220576037f, 6.8201789624151878f, + 6.8328900141647412f, 6.8454900509443747f, + 6.8579809951275718f, 6.8703647195834047f, + 6.8826430493618415f, 6.8948177633079437f, + 6.9068905956085187f, 6.9188632372745946f, + 6.9307373375628866f, 6.9425145053392398f, + 6.9541963103868749f, 6.9657842846620869f, + 6.9772799234999167f, 6.9886846867721654f, + 7.0000000000000000f, 7.0112272554232539f, + 7.0223678130284543f, 7.0334230015374501f, + 7.0443941193584533f, 7.0552824355011898f, + 7.0660891904577720f, 7.0768155970508308f, + 7.0874628412503390f, 7.0980320829605263f, + 7.1085244567781691f, 7.1189410727235076f, + 7.1292830169449663f, 7.1395513523987936f, + 7.1497471195046822f, 7.1598713367783890f, + 7.1699250014423121f, 7.1799090900149344f, + 7.1898245588800175f, 7.1996723448363644f, + 7.2094533656289501f, 7.2191685204621611f, + 7.2288186904958804f, 7.2384047393250785f, + 7.2479275134435852f, 7.2573878426926521f, + 7.2667865406949010f, 7.2761244052742375f, + 7.2854022188622487f, 7.2946207488916270f, + 7.3037807481771030f, 7.3128829552843557f, + 7.3219280948873626f, 7.3309168781146167f, + 7.3398500028846243f, 7.3487281542310771f, + 7.3575520046180837f, 7.3663222142458160f, + 7.3750394313469245f, 7.3837042924740519f, + 7.3923174227787606f, 7.4008794362821843f, + 7.4093909361377017f, 7.4178525148858982f, + 7.4262647547020979f, 7.4346282276367245f, + 7.4429434958487279f, 7.4512111118323289f, + 7.4594316186372973f, 7.4676055500829976f, + 7.4757334309663976f, 7.4838157772642563f, + 7.4918530963296747f, 7.4998458870832056f, + 7.5077946401986963f, 7.5156998382840427f, + 7.5235619560570130f, 7.5313814605163118f, + 7.5391588111080309f, 7.5468944598876364f, + 7.5545888516776376f, 7.5622424242210728f, + 7.5698556083309478f, 7.5774288280357486f, + 7.5849625007211560f, 7.5924570372680806f, + 7.5999128421871278f, 7.6073303137496104f, + 7.6147098441152083f, 7.6220518194563764f, + 7.6293566200796094f, 7.6366246205436487f, + 7.6438561897747243f, 7.6510516911789281f, + 7.6582114827517946f, 7.6653359171851764f, + 7.6724253419714951f, 7.6794800995054464f, + 7.6865005271832185f, 7.6934869574993252f, + 7.7004397181410917f, 7.7073591320808825f, + 7.7142455176661224f, 7.7210991887071855f, + 7.7279204545631987f, 7.7347096202258383f, + 7.7414669864011464f, 7.7481928495894605f, + 7.7548875021634682f, 7.7615512324444795f, + 7.7681843247769259f, 7.7747870596011736f, + 7.7813597135246599f, 7.7879025593914317f, + 7.7944158663501061f, 7.8008998999203047f, + 7.8073549220576037f, 7.8137811912170374f, + 7.8201789624151878f, 7.8265484872909150f, + 7.8328900141647412f, 7.8392037880969436f, + 7.8454900509443747f, 7.8517490414160571f, + 7.8579809951275718f, 7.8641861446542797f, + 7.8703647195834047f, 7.8765169465649993f, + 7.8826430493618415f, 7.8887432488982591f, + 7.8948177633079437f, 7.9008668079807486f, + 7.9068905956085187f, 7.9128893362299619f, + 7.9188632372745946f, 7.9248125036057812f, + 7.9307373375628866f, 7.9366379390025709f, + 7.9425145053392398f, 7.9483672315846778f, + 7.9541963103868749f, 7.9600019320680805f, + 7.9657842846620869f, 7.9715435539507719f, + 7.9772799234999167f, 7.9829935746943103f, + 7.9886846867721654f, 7.9943534368588577f +}; + +const float kSLog2Table[LOG_LOOKUP_IDX_MAX] = { + 0.00000000f, 0.00000000f, 2.00000000f, 4.75488750f, + 8.00000000f, 11.60964047f, 15.50977500f, 19.65148445f, + 24.00000000f, 28.52932501f, 33.21928095f, 38.05374781f, + 43.01955001f, 48.10571634f, 53.30296891f, 58.60335893f, + 64.00000000f, 69.48686830f, 75.05865003f, 80.71062276f, + 86.43856190f, 92.23866588f, 98.10749561f, 104.04192499f, + 110.03910002f, 116.09640474f, 122.21143267f, 128.38196256f, + 134.60593782f, 140.88144886f, 147.20671787f, 153.58008562f, + 160.00000000f, 166.46500594f, 172.97373660f, 179.52490559f, + 186.11730005f, 192.74977453f, 199.42124551f, 206.13068654f, + 212.87712380f, 219.65963219f, 226.47733176f, 233.32938445f, + 240.21499122f, 247.13338933f, 254.08384998f, 261.06567603f, + 268.07820003f, 275.12078236f, 282.19280949f, 289.29369244f, + 296.42286534f, 303.57978409f, 310.76392512f, 317.97478424f, + 325.21187564f, 332.47473081f, 339.76289772f, 347.07593991f, + 354.41343574f, 361.77497759f, 369.16017124f, 376.56863518f, + 384.00000000f, 391.45390785f, 398.93001188f, 406.42797576f, + 413.94747321f, 421.48818752f, 429.04981119f, 436.63204548f, + 444.23460010f, 451.85719280f, 459.49954906f, 467.16140179f, + 474.84249102f, 482.54256363f, 490.26137307f, 497.99867911f, + 505.75424759f, 513.52785023f, 521.31926438f, 529.12827280f, + 536.95466351f, 544.79822957f, 552.65876890f, 560.53608414f, + 568.42998244f, 576.34027536f, 584.26677867f, 592.20931226f, + 600.16769996f, 608.14176943f, 616.13135206f, 624.13628279f, + 632.15640007f, 640.19154569f, 648.24156472f, 656.30630539f, + 664.38561898f, 672.47935976f, 680.58738488f, 688.70955430f, + 696.84573069f, 704.99577935f, 713.15956818f, 721.33696754f, + 729.52785023f, 737.73209140f, 745.94956849f, 754.18016116f, + 762.42375127f, 770.68022275f, 778.94946161f, 787.23135586f, + 795.52579543f, 803.83267219f, 812.15187982f, 820.48331383f, + 828.82687147f, 837.18245171f, 845.54995518f, 853.92928416f, + 862.32034249f, 870.72303558f, 879.13727036f, 887.56295522f, + 896.00000000f, 904.44831595f, 912.90781569f, 921.37841320f, + 929.86002376f, 938.35256392f, 946.85595152f, 955.37010560f, + 963.89494641f, 972.43039537f, 980.97637504f, 989.53280911f, + 998.09962237f, 1006.67674069f, 1015.26409097f, 1023.86160116f, + 1032.46920021f, 1041.08681805f, 1049.71438560f, 1058.35183469f, + 1066.99909811f, 1075.65610955f, 1084.32280357f, 1092.99911564f, + 1101.68498204f, 1110.38033993f, 1119.08512727f, 1127.79928282f, + 1136.52274614f, 1145.25545758f, 1153.99735821f, 1162.74838989f, + 1171.50849518f, 1180.27761738f, 1189.05570047f, 1197.84268914f, + 1206.63852876f, 1215.44316535f, 1224.25654560f, 1233.07861684f, + 1241.90932703f, 1250.74862473f, 1259.59645914f, 1268.45278005f, + 1277.31753781f, 1286.19068338f, 1295.07216828f, 1303.96194457f, + 1312.85996488f, 1321.76618236f, 1330.68055071f, 1339.60302413f, + 1348.53355734f, 1357.47210556f, 1366.41862452f, 1375.37307041f, + 1384.33539991f, 1393.30557020f, 1402.28353887f, 1411.26926400f, + 1420.26270412f, 1429.26381818f, 1438.27256558f, 1447.28890615f, + 1456.31280014f, 1465.34420819f, 1474.38309138f, 1483.42941118f, + 1492.48312945f, 1501.54420843f, 1510.61261078f, 1519.68829949f, + 1528.77123795f, 1537.86138993f, 1546.95871952f, 1556.06319119f, + 1565.17476976f, 1574.29342040f, 1583.41910860f, 1592.55180020f, + 1601.69146137f, 1610.83805860f, 1619.99155871f, 1629.15192882f, + 1638.31913637f, 1647.49314911f, 1656.67393509f, 1665.86146266f, + 1675.05570047f, 1684.25661744f, 1693.46418280f, 1702.67836605f, + 1711.89913698f, 1721.12646563f, 1730.36032233f, 1739.60067768f, + 1748.84750254f, 1758.10076802f, 1767.36044551f, 1776.62650662f, + 1785.89892323f, 1795.17766747f, 1804.46271172f, 1813.75402857f, + 1823.05159087f, 1832.35537170f, 1841.66534438f, 1850.98148244f, + 1860.30375965f, 1869.63214999f, 1878.96662767f, 1888.30716711f, + 1897.65374295f, 1907.00633003f, 1916.36490342f, 1925.72943838f, + 1935.09991037f, 1944.47629506f, 1953.85856831f, 1963.24670620f, + 1972.64068498f, 1982.04048108f, 1991.44607117f, 2000.85743204f, + 2010.27454072f, 2019.69737440f, 2029.12591044f, 2038.56012640f +}; + +const VP8LPrefixCode kPrefixEncodeCode[PREFIX_LOOKUP_IDX_MAX] = { + { 0, 0}, { 0, 0}, { 1, 0}, { 2, 0}, { 3, 0}, { 4, 1}, { 4, 1}, { 5, 1}, + { 5, 1}, { 6, 2}, { 6, 2}, { 6, 2}, { 6, 2}, { 7, 2}, { 7, 2}, { 7, 2}, + { 7, 2}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, { 8, 3}, + { 8, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, { 9, 3}, + { 9, 3}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, + {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, {10, 4}, + {10, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, + {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, {11, 4}, + {11, 4}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, {12, 5}, + {12, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, {13, 5}, + {13, 5}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, {14, 6}, + {14, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, {15, 6}, + {15, 6}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, {16, 7}, + {16, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, + {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, {17, 7}, +}; + +const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX] = { + 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 1, 2, 3, 0, 1, 2, 3, + 0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, + 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, + 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, + 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, + 127, + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, + 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, + 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, + 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, + 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, + 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, + 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126 +}; + +static float FastSLog2Slow_C(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { +#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ) + // use clz if available + const int log_cnt = BitsLog2Floor(v) - 7; + const uint32_t y = 1 << log_cnt; + int correction = 0; + const float v_f = (float)v; + const uint32_t orig_v = v; + v >>= log_cnt; +#else + int log_cnt = 0; + uint32_t y = 1; + int correction = 0; + const float v_f = (float)v; + const uint32_t orig_v = v; + do { + ++log_cnt; + v = v >> 1; + y = y << 1; + } while (v >= LOG_LOOKUP_IDX_MAX); +#endif + // vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256 + // Xf = floor(Xf) * (1 + (v % y) / v) + // log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v) + // The correction factor: log(1 + d) ~ d; for very small d values, so + // log2(1 + (v % y) / v) ~ LOG_2_RECIPROCAL * (v % y)/v + // LOG_2_RECIPROCAL ~ 23/16 + correction = (23 * (orig_v & (y - 1))) >> 4; + return v_f * (kLog2Table[v] + log_cnt) + correction; + } else { + return (float)(LOG_2_RECIPROCAL * v * log((double)v)); + } +} + +static float FastLog2Slow_C(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { +#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ) + // use clz if available + const int log_cnt = BitsLog2Floor(v) - 7; + const uint32_t y = 1 << log_cnt; + const uint32_t orig_v = v; + double log_2; + v >>= log_cnt; +#else + int log_cnt = 0; + uint32_t y = 1; + const uint32_t orig_v = v; + double log_2; + do { + ++log_cnt; + v = v >> 1; + y = y << 1; + } while (v >= LOG_LOOKUP_IDX_MAX); +#endif + log_2 = kLog2Table[v] + log_cnt; + if (orig_v >= APPROX_LOG_MAX) { + // Since the division is still expensive, add this correction factor only + // for large values of 'v'. + const int correction = (23 * (orig_v & (y - 1))) >> 4; + log_2 += (double)correction / orig_v; + } + return (float)log_2; + } else { + return (float)(LOG_2_RECIPROCAL * log((double)v)); + } +} + +//------------------------------------------------------------------------------ +// Methods to calculate Entropy (Shannon). + +// Compute the combined Shanon's entropy for distribution {X} and {X+Y} +static float CombinedShannonEntropy_C(const int X[256], const int Y[256]) { + int i; + double retval = 0.; + int sumX = 0, sumXY = 0; + for (i = 0; i < 256; ++i) { + const int x = X[i]; + if (x != 0) { + const int xy = x + Y[i]; + sumX += x; + retval -= VP8LFastSLog2(x); + sumXY += xy; + retval -= VP8LFastSLog2(xy); + } else if (Y[i] != 0) { + sumXY += Y[i]; + retval -= VP8LFastSLog2(Y[i]); + } + } + retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY); + return (float)retval; +} + +void VP8LBitEntropyInit(VP8LBitEntropy* const entropy) { + entropy->entropy = 0.; + entropy->sum = 0; + entropy->nonzeros = 0; + entropy->max_val = 0; + entropy->nonzero_code = VP8L_NON_TRIVIAL_SYM; +} + +void VP8LBitsEntropyUnrefined(const uint32_t* const array, int n, + VP8LBitEntropy* const entropy) { + int i; + + VP8LBitEntropyInit(entropy); + + for (i = 0; i < n; ++i) { + if (array[i] != 0) { + entropy->sum += array[i]; + entropy->nonzero_code = i; + ++entropy->nonzeros; + entropy->entropy -= VP8LFastSLog2(array[i]); + if (entropy->max_val < array[i]) { + entropy->max_val = array[i]; + } + } + } + entropy->entropy += VP8LFastSLog2(entropy->sum); +} + +static WEBP_INLINE void GetEntropyUnrefinedHelper( + uint32_t val, int i, uint32_t* const val_prev, int* const i_prev, + VP8LBitEntropy* const bit_entropy, VP8LStreaks* const stats) { + const int streak = i - *i_prev; + + // Gather info for the bit entropy. + if (*val_prev != 0) { + bit_entropy->sum += (*val_prev) * streak; + bit_entropy->nonzeros += streak; + bit_entropy->nonzero_code = *i_prev; + bit_entropy->entropy -= VP8LFastSLog2(*val_prev) * streak; + if (bit_entropy->max_val < *val_prev) { + bit_entropy->max_val = *val_prev; + } + } + + // Gather info for the Huffman cost. + stats->counts[*val_prev != 0] += (streak > 3); + stats->streaks[*val_prev != 0][(streak > 3)] += streak; + + *val_prev = val; + *i_prev = i; +} + +static void GetEntropyUnrefined_C(const uint32_t X[], int length, + VP8LBitEntropy* const bit_entropy, + VP8LStreaks* const stats) { + int i; + int i_prev = 0; + uint32_t x_prev = X[0]; + + memset(stats, 0, sizeof(*stats)); + VP8LBitEntropyInit(bit_entropy); + + for (i = 1; i < length; ++i) { + const uint32_t x = X[i]; + if (x != x_prev) { + GetEntropyUnrefinedHelper(x, i, &x_prev, &i_prev, bit_entropy, stats); + } + } + GetEntropyUnrefinedHelper(0, i, &x_prev, &i_prev, bit_entropy, stats); + + bit_entropy->entropy += VP8LFastSLog2(bit_entropy->sum); +} + +static void GetCombinedEntropyUnrefined_C(const uint32_t X[], + const uint32_t Y[], + int length, + VP8LBitEntropy* const bit_entropy, + VP8LStreaks* const stats) { + int i = 1; + int i_prev = 0; + uint32_t xy_prev = X[0] + Y[0]; + + memset(stats, 0, sizeof(*stats)); + VP8LBitEntropyInit(bit_entropy); + + for (i = 1; i < length; ++i) { + const uint32_t xy = X[i] + Y[i]; + if (xy != xy_prev) { + GetEntropyUnrefinedHelper(xy, i, &xy_prev, &i_prev, bit_entropy, stats); + } + } + GetEntropyUnrefinedHelper(0, i, &xy_prev, &i_prev, bit_entropy, stats); + + bit_entropy->entropy += VP8LFastSLog2(bit_entropy->sum); +} + +//------------------------------------------------------------------------------ + +void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels) { + int i; + for (i = 0; i < num_pixels; ++i) { + const int argb = argb_data[i]; + const int green = (argb >> 8) & 0xff; + const uint32_t new_r = (((argb >> 16) & 0xff) - green) & 0xff; + const uint32_t new_b = (((argb >> 0) & 0xff) - green) & 0xff; + argb_data[i] = (argb & 0xff00ff00u) | (new_r << 16) | new_b; + } +} + +static WEBP_INLINE int ColorTransformDelta(int8_t color_pred, int8_t color) { + return ((int)color_pred * color) >> 5; +} + +static WEBP_INLINE int8_t U32ToS8(uint32_t v) { + return (int8_t)(v & 0xff); +} + +void VP8LTransformColor_C(const VP8LMultipliers* const m, uint32_t* data, + int num_pixels) { + int i; + for (i = 0; i < num_pixels; ++i) { + const uint32_t argb = data[i]; + const int8_t green = U32ToS8(argb >> 8); + const int8_t red = U32ToS8(argb >> 16); + int new_red = red & 0xff; + int new_blue = argb & 0xff; + new_red -= ColorTransformDelta(m->green_to_red_, green); + new_red &= 0xff; + new_blue -= ColorTransformDelta(m->green_to_blue_, green); + new_blue -= ColorTransformDelta(m->red_to_blue_, red); + new_blue &= 0xff; + data[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue); + } +} + +static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red, + uint32_t argb) { + const int8_t green = U32ToS8(argb >> 8); + int new_red = argb >> 16; + new_red -= ColorTransformDelta(green_to_red, green); + return (new_red & 0xff); +} + +static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue, + uint8_t red_to_blue, + uint32_t argb) { + const int8_t green = U32ToS8(argb >> 8); + const int8_t red = U32ToS8(argb >> 16); + uint8_t new_blue = argb & 0xff; + new_blue -= ColorTransformDelta(green_to_blue, green); + new_blue -= ColorTransformDelta(red_to_blue, red); + return (new_blue & 0xff); +} + +void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]) { + while (tile_height-- > 0) { + int x; + for (x = 0; x < tile_width; ++x) { + ++histo[TransformColorRed((uint8_t)green_to_red, argb[x])]; + } + argb += stride; + } +} + +void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, + int histo[]) { + while (tile_height-- > 0) { + int x; + for (x = 0; x < tile_width; ++x) { + ++histo[TransformColorBlue((uint8_t)green_to_blue, (uint8_t)red_to_blue, + argb[x])]; + } + argb += stride; + } +} + +//------------------------------------------------------------------------------ + +static int VectorMismatch_C(const uint32_t* const array1, + const uint32_t* const array2, int length) { + int match_len = 0; + + while (match_len < length && array1[match_len] == array2[match_len]) { + ++match_len; + } + return match_len; +} + +// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel. +void VP8LBundleColorMap_C(const uint8_t* const row, int width, int xbits, + uint32_t* dst) { + int x; + if (xbits > 0) { + const int bit_depth = 1 << (3 - xbits); + const int mask = (1 << xbits) - 1; + uint32_t code = 0xff000000; + for (x = 0; x < width; ++x) { + const int xsub = x & mask; + if (xsub == 0) { + code = 0xff000000; + } + code |= row[x] << (8 + bit_depth * xsub); + dst[x >> xbits] = code; + } + } else { + for (x = 0; x < width; ++x) dst[x] = 0xff000000 | (row[x] << 8); + } +} + +//------------------------------------------------------------------------------ + +static double ExtraCost_C(const uint32_t* population, int length) { + int i; + double cost = 0.; + for (i = 2; i < length - 2; ++i) cost += (i >> 1) * population[i + 2]; + return cost; +} + +static double ExtraCostCombined_C(const uint32_t* X, const uint32_t* Y, + int length) { + int i; + double cost = 0.; + for (i = 2; i < length - 2; ++i) { + const int xy = X[i + 2] + Y[i + 2]; + cost += (i >> 1) * xy; + } + return cost; +} + +//------------------------------------------------------------------------------ + +static void AddVector_C(const uint32_t* a, const uint32_t* b, uint32_t* out, + int size) { + int i; + for (i = 0; i < size; ++i) out[i] = a[i] + b[i]; +} + +static void AddVectorEq_C(const uint32_t* a, uint32_t* out, int size) { + int i; + for (i = 0; i < size; ++i) out[i] += a[i]; +} + +#define ADD(X, ARG, LEN) do { \ + if (a->is_used_[X]) { \ + if (b->is_used_[X]) { \ + VP8LAddVector(a->ARG, b->ARG, out->ARG, (LEN)); \ + } else { \ + memcpy(&out->ARG[0], &a->ARG[0], (LEN) * sizeof(out->ARG[0])); \ + } \ + } else if (b->is_used_[X]) { \ + memcpy(&out->ARG[0], &b->ARG[0], (LEN) * sizeof(out->ARG[0])); \ + } else { \ + memset(&out->ARG[0], 0, (LEN) * sizeof(out->ARG[0])); \ + } \ +} while (0) + +#define ADD_EQ(X, ARG, LEN) do { \ + if (a->is_used_[X]) { \ + if (out->is_used_[X]) { \ + VP8LAddVectorEq(a->ARG, out->ARG, (LEN)); \ + } else { \ + memcpy(&out->ARG[0], &a->ARG[0], (LEN) * sizeof(out->ARG[0])); \ + } \ + } \ +} while (0) + +void VP8LHistogramAdd(const VP8LHistogram* const a, + const VP8LHistogram* const b, VP8LHistogram* const out) { + int i; + const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_); + assert(a->palette_code_bits_ == b->palette_code_bits_); + + if (b != out) { + ADD(0, literal_, literal_size); + ADD(1, red_, NUM_LITERAL_CODES); + ADD(2, blue_, NUM_LITERAL_CODES); + ADD(3, alpha_, NUM_LITERAL_CODES); + ADD(4, distance_, NUM_DISTANCE_CODES); + for (i = 0; i < 5; ++i) { + out->is_used_[i] = (a->is_used_[i] | b->is_used_[i]); + } + } else { + ADD_EQ(0, literal_, literal_size); + ADD_EQ(1, red_, NUM_LITERAL_CODES); + ADD_EQ(2, blue_, NUM_LITERAL_CODES); + ADD_EQ(3, alpha_, NUM_LITERAL_CODES); + ADD_EQ(4, distance_, NUM_DISTANCE_CODES); + for (i = 0; i < 5; ++i) out->is_used_[i] |= a->is_used_[i]; + } +} +#undef ADD +#undef ADD_EQ + +//------------------------------------------------------------------------------ +// Image transforms. + +static void PredictorSub0_C(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i < num_pixels; ++i) out[i] = VP8LSubPixels(in[i], ARGB_BLACK); + (void)upper; +} + +static void PredictorSub1_C(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i < num_pixels; ++i) out[i] = VP8LSubPixels(in[i], in[i - 1]); + (void)upper; +} + +// It subtracts the prediction from the input pixel and stores the residual +// in the output pixel. +#define GENERATE_PREDICTOR_SUB(PREDICTOR_I) \ +static void PredictorSub##PREDICTOR_I##_C(const uint32_t* in, \ + const uint32_t* upper, \ + int num_pixels, uint32_t* out) { \ + int x; \ + assert(upper != NULL); \ + for (x = 0; x < num_pixels; ++x) { \ + const uint32_t pred = \ + VP8LPredictor##PREDICTOR_I##_C(&in[x - 1], upper + x); \ + out[x] = VP8LSubPixels(in[x], pred); \ + } \ +} + +GENERATE_PREDICTOR_SUB(2) +GENERATE_PREDICTOR_SUB(3) +GENERATE_PREDICTOR_SUB(4) +GENERATE_PREDICTOR_SUB(5) +GENERATE_PREDICTOR_SUB(6) +GENERATE_PREDICTOR_SUB(7) +GENERATE_PREDICTOR_SUB(8) +GENERATE_PREDICTOR_SUB(9) +GENERATE_PREDICTOR_SUB(10) +GENERATE_PREDICTOR_SUB(11) +GENERATE_PREDICTOR_SUB(12) +GENERATE_PREDICTOR_SUB(13) + +//------------------------------------------------------------------------------ + +VP8LProcessEncBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed; + +VP8LTransformColorFunc VP8LTransformColor; + +VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms; +VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms; + +VP8LFastLog2SlowFunc VP8LFastLog2Slow; +VP8LFastLog2SlowFunc VP8LFastSLog2Slow; + +VP8LCostFunc VP8LExtraCost; +VP8LCostCombinedFunc VP8LExtraCostCombined; +VP8LCombinedShannonEntropyFunc VP8LCombinedShannonEntropy; + +VP8LGetEntropyUnrefinedFunc VP8LGetEntropyUnrefined; +VP8LGetCombinedEntropyUnrefinedFunc VP8LGetCombinedEntropyUnrefined; + +VP8LAddVectorFunc VP8LAddVector; +VP8LAddVectorEqFunc VP8LAddVectorEq; + +VP8LVectorMismatchFunc VP8LVectorMismatch; +VP8LBundleColorMapFunc VP8LBundleColorMap; + +VP8LPredictorAddSubFunc VP8LPredictorsSub[16]; +VP8LPredictorAddSubFunc VP8LPredictorsSub_C[16]; + +extern void VP8LEncDspInitSSE2(void); +extern void VP8LEncDspInitSSE41(void); +extern void VP8LEncDspInitNEON(void); +extern void VP8LEncDspInitMIPS32(void); +extern void VP8LEncDspInitMIPSdspR2(void); +extern void VP8LEncDspInitMSA(void); + +WEBP_DSP_INIT_FUNC(VP8LEncDspInit) { + VP8LDspInit(); + +#if !WEBP_NEON_OMIT_C_CODE + VP8LSubtractGreenFromBlueAndRed = VP8LSubtractGreenFromBlueAndRed_C; + + VP8LTransformColor = VP8LTransformColor_C; +#endif + + VP8LCollectColorBlueTransforms = VP8LCollectColorBlueTransforms_C; + VP8LCollectColorRedTransforms = VP8LCollectColorRedTransforms_C; + + VP8LFastLog2Slow = FastLog2Slow_C; + VP8LFastSLog2Slow = FastSLog2Slow_C; + + VP8LExtraCost = ExtraCost_C; + VP8LExtraCostCombined = ExtraCostCombined_C; + VP8LCombinedShannonEntropy = CombinedShannonEntropy_C; + + VP8LGetEntropyUnrefined = GetEntropyUnrefined_C; + VP8LGetCombinedEntropyUnrefined = GetCombinedEntropyUnrefined_C; + + VP8LAddVector = AddVector_C; + VP8LAddVectorEq = AddVectorEq_C; + + VP8LVectorMismatch = VectorMismatch_C; + VP8LBundleColorMap = VP8LBundleColorMap_C; + + VP8LPredictorsSub[0] = PredictorSub0_C; + VP8LPredictorsSub[1] = PredictorSub1_C; + VP8LPredictorsSub[2] = PredictorSub2_C; + VP8LPredictorsSub[3] = PredictorSub3_C; + VP8LPredictorsSub[4] = PredictorSub4_C; + VP8LPredictorsSub[5] = PredictorSub5_C; + VP8LPredictorsSub[6] = PredictorSub6_C; + VP8LPredictorsSub[7] = PredictorSub7_C; + VP8LPredictorsSub[8] = PredictorSub8_C; + VP8LPredictorsSub[9] = PredictorSub9_C; + VP8LPredictorsSub[10] = PredictorSub10_C; + VP8LPredictorsSub[11] = PredictorSub11_C; + VP8LPredictorsSub[12] = PredictorSub12_C; + VP8LPredictorsSub[13] = PredictorSub13_C; + VP8LPredictorsSub[14] = PredictorSub0_C; // <- padding security sentinels + VP8LPredictorsSub[15] = PredictorSub0_C; + + VP8LPredictorsSub_C[0] = PredictorSub0_C; + VP8LPredictorsSub_C[1] = PredictorSub1_C; + VP8LPredictorsSub_C[2] = PredictorSub2_C; + VP8LPredictorsSub_C[3] = PredictorSub3_C; + VP8LPredictorsSub_C[4] = PredictorSub4_C; + VP8LPredictorsSub_C[5] = PredictorSub5_C; + VP8LPredictorsSub_C[6] = PredictorSub6_C; + VP8LPredictorsSub_C[7] = PredictorSub7_C; + VP8LPredictorsSub_C[8] = PredictorSub8_C; + VP8LPredictorsSub_C[9] = PredictorSub9_C; + VP8LPredictorsSub_C[10] = PredictorSub10_C; + VP8LPredictorsSub_C[11] = PredictorSub11_C; + VP8LPredictorsSub_C[12] = PredictorSub12_C; + VP8LPredictorsSub_C[13] = PredictorSub13_C; + VP8LPredictorsSub_C[14] = PredictorSub0_C; // <- padding security sentinels + VP8LPredictorsSub_C[15] = PredictorSub0_C; + + // If defined, use CPUInfo() to overwrite some pointers with faster versions. + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_HAVE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8LEncDspInitSSE2(); +#if defined(WEBP_HAVE_SSE41) + if (VP8GetCPUInfo(kSSE4_1)) { + VP8LEncDspInitSSE41(); + } +#endif + } +#endif +#if defined(WEBP_USE_MIPS32) + if (VP8GetCPUInfo(kMIPS32)) { + VP8LEncDspInitMIPS32(); + } +#endif +#if defined(WEBP_USE_MIPS_DSP_R2) + if (VP8GetCPUInfo(kMIPSdspR2)) { + VP8LEncDspInitMIPSdspR2(); + } +#endif +#if defined(WEBP_USE_MSA) + if (VP8GetCPUInfo(kMSA)) { + VP8LEncDspInitMSA(); + } +#endif + } + +#if defined(WEBP_HAVE_NEON) + if (WEBP_NEON_OMIT_C_CODE || + (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { + VP8LEncDspInitNEON(); + } +#endif + + assert(VP8LSubtractGreenFromBlueAndRed != NULL); + assert(VP8LTransformColor != NULL); + assert(VP8LCollectColorBlueTransforms != NULL); + assert(VP8LCollectColorRedTransforms != NULL); + assert(VP8LFastLog2Slow != NULL); + assert(VP8LFastSLog2Slow != NULL); + assert(VP8LExtraCost != NULL); + assert(VP8LExtraCostCombined != NULL); + assert(VP8LCombinedShannonEntropy != NULL); + assert(VP8LGetEntropyUnrefined != NULL); + assert(VP8LGetCombinedEntropyUnrefined != NULL); + assert(VP8LAddVector != NULL); + assert(VP8LAddVectorEq != NULL); + assert(VP8LVectorMismatch != NULL); + assert(VP8LBundleColorMap != NULL); + assert(VP8LPredictorsSub[0] != NULL); + assert(VP8LPredictorsSub[1] != NULL); + assert(VP8LPredictorsSub[2] != NULL); + assert(VP8LPredictorsSub[3] != NULL); + assert(VP8LPredictorsSub[4] != NULL); + assert(VP8LPredictorsSub[5] != NULL); + assert(VP8LPredictorsSub[6] != NULL); + assert(VP8LPredictorsSub[7] != NULL); + assert(VP8LPredictorsSub[8] != NULL); + assert(VP8LPredictorsSub[9] != NULL); + assert(VP8LPredictorsSub[10] != NULL); + assert(VP8LPredictorsSub[11] != NULL); + assert(VP8LPredictorsSub[12] != NULL); + assert(VP8LPredictorsSub[13] != NULL); + assert(VP8LPredictorsSub[14] != NULL); + assert(VP8LPredictorsSub[15] != NULL); + assert(VP8LPredictorsSub_C[0] != NULL); + assert(VP8LPredictorsSub_C[1] != NULL); + assert(VP8LPredictorsSub_C[2] != NULL); + assert(VP8LPredictorsSub_C[3] != NULL); + assert(VP8LPredictorsSub_C[4] != NULL); + assert(VP8LPredictorsSub_C[5] != NULL); + assert(VP8LPredictorsSub_C[6] != NULL); + assert(VP8LPredictorsSub_C[7] != NULL); + assert(VP8LPredictorsSub_C[8] != NULL); + assert(VP8LPredictorsSub_C[9] != NULL); + assert(VP8LPredictorsSub_C[10] != NULL); + assert(VP8LPredictorsSub_C[11] != NULL); + assert(VP8LPredictorsSub_C[12] != NULL); + assert(VP8LPredictorsSub_C[13] != NULL); + assert(VP8LPredictorsSub_C[14] != NULL); + assert(VP8LPredictorsSub_C[15] != NULL); +} + +//------------------------------------------------------------------------------ diff --git a/media/libwebp/dsp/lossless_enc_mips32.c b/media/libwebp/dsp/lossless_enc_mips32.c new file mode 100644 index 0000000000..088e608b44 --- /dev/null +++ b/media/libwebp/dsp/lossless_enc_mips32.c @@ -0,0 +1,397 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of lossless functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "../dsp/dsp.h" +#include "../dsp/lossless.h" +#include "../dsp/lossless_common.h" + +#if defined(WEBP_USE_MIPS32) + +#include <assert.h> +#include <math.h> +#include <stdlib.h> +#include <string.h> + +static float FastSLog2Slow_MIPS32(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { + uint32_t log_cnt, y, correction; + const int c24 = 24; + const float v_f = (float)v; + uint32_t temp; + + // Xf = 256 = 2^8 + // log_cnt is index of leading one in upper 24 bits + __asm__ volatile( + "clz %[log_cnt], %[v] \n\t" + "addiu %[y], $zero, 1 \n\t" + "subu %[log_cnt], %[c24], %[log_cnt] \n\t" + "sllv %[y], %[y], %[log_cnt] \n\t" + "srlv %[temp], %[v], %[log_cnt] \n\t" + : [log_cnt]"=&r"(log_cnt), [y]"=&r"(y), + [temp]"=r"(temp) + : [c24]"r"(c24), [v]"r"(v) + ); + + // vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256 + // Xf = floor(Xf) * (1 + (v % y) / v) + // log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v) + // The correction factor: log(1 + d) ~ d; for very small d values, so + // log2(1 + (v % y) / v) ~ LOG_2_RECIPROCAL * (v % y)/v + // LOG_2_RECIPROCAL ~ 23/16 + + // (v % y) = (v % 2^log_cnt) = v & (2^log_cnt - 1) + correction = (23 * (v & (y - 1))) >> 4; + return v_f * (kLog2Table[temp] + log_cnt) + correction; + } else { + return (float)(LOG_2_RECIPROCAL * v * log((double)v)); + } +} + +static float FastLog2Slow_MIPS32(uint32_t v) { + assert(v >= LOG_LOOKUP_IDX_MAX); + if (v < APPROX_LOG_WITH_CORRECTION_MAX) { + uint32_t log_cnt, y; + const int c24 = 24; + double log_2; + uint32_t temp; + + __asm__ volatile( + "clz %[log_cnt], %[v] \n\t" + "addiu %[y], $zero, 1 \n\t" + "subu %[log_cnt], %[c24], %[log_cnt] \n\t" + "sllv %[y], %[y], %[log_cnt] \n\t" + "srlv %[temp], %[v], %[log_cnt] \n\t" + : [log_cnt]"=&r"(log_cnt), [y]"=&r"(y), + [temp]"=r"(temp) + : [c24]"r"(c24), [v]"r"(v) + ); + + log_2 = kLog2Table[temp] + log_cnt; + if (v >= APPROX_LOG_MAX) { + // Since the division is still expensive, add this correction factor only + // for large values of 'v'. + + const uint32_t correction = (23 * (v & (y - 1))) >> 4; + log_2 += (double)correction / v; + } + return (float)log_2; + } else { + return (float)(LOG_2_RECIPROCAL * log((double)v)); + } +} + +// C version of this function: +// int i = 0; +// int64_t cost = 0; +// const uint32_t* pop = &population[4]; +// const uint32_t* LoopEnd = &population[length]; +// while (pop != LoopEnd) { +// ++i; +// cost += i * *pop; +// cost += i * *(pop + 1); +// pop += 2; +// } +// return (double)cost; +static double ExtraCost_MIPS32(const uint32_t* const population, int length) { + int i, temp0, temp1; + const uint32_t* pop = &population[4]; + const uint32_t* const LoopEnd = &population[length]; + + __asm__ volatile( + "mult $zero, $zero \n\t" + "xor %[i], %[i], %[i] \n\t" + "beq %[pop], %[LoopEnd], 2f \n\t" + "1: \n\t" + "lw %[temp0], 0(%[pop]) \n\t" + "lw %[temp1], 4(%[pop]) \n\t" + "addiu %[i], %[i], 1 \n\t" + "addiu %[pop], %[pop], 8 \n\t" + "madd %[i], %[temp0] \n\t" + "madd %[i], %[temp1] \n\t" + "bne %[pop], %[LoopEnd], 1b \n\t" + "2: \n\t" + "mfhi %[temp0] \n\t" + "mflo %[temp1] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [i]"=&r"(i), [pop]"+r"(pop) + : [LoopEnd]"r"(LoopEnd) + : "memory", "hi", "lo" + ); + + return (double)((int64_t)temp0 << 32 | temp1); +} + +// C version of this function: +// int i = 0; +// int64_t cost = 0; +// const uint32_t* pX = &X[4]; +// const uint32_t* pY = &Y[4]; +// const uint32_t* LoopEnd = &X[length]; +// while (pX != LoopEnd) { +// const uint32_t xy0 = *pX + *pY; +// const uint32_t xy1 = *(pX + 1) + *(pY + 1); +// ++i; +// cost += i * xy0; +// cost += i * xy1; +// pX += 2; +// pY += 2; +// } +// return (double)cost; +static double ExtraCostCombined_MIPS32(const uint32_t* const X, + const uint32_t* const Y, int length) { + int i, temp0, temp1, temp2, temp3; + const uint32_t* pX = &X[4]; + const uint32_t* pY = &Y[4]; + const uint32_t* const LoopEnd = &X[length]; + + __asm__ volatile( + "mult $zero, $zero \n\t" + "xor %[i], %[i], %[i] \n\t" + "beq %[pX], %[LoopEnd], 2f \n\t" + "1: \n\t" + "lw %[temp0], 0(%[pX]) \n\t" + "lw %[temp1], 0(%[pY]) \n\t" + "lw %[temp2], 4(%[pX]) \n\t" + "lw %[temp3], 4(%[pY]) \n\t" + "addiu %[i], %[i], 1 \n\t" + "addu %[temp0], %[temp0], %[temp1] \n\t" + "addu %[temp2], %[temp2], %[temp3] \n\t" + "addiu %[pX], %[pX], 8 \n\t" + "addiu %[pY], %[pY], 8 \n\t" + "madd %[i], %[temp0] \n\t" + "madd %[i], %[temp2] \n\t" + "bne %[pX], %[LoopEnd], 1b \n\t" + "2: \n\t" + "mfhi %[temp0] \n\t" + "mflo %[temp1] \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [i]"=&r"(i), [pX]"+r"(pX), [pY]"+r"(pY) + : [LoopEnd]"r"(LoopEnd) + : "memory", "hi", "lo" + ); + + return (double)((int64_t)temp0 << 32 | temp1); +} + +#define HUFFMAN_COST_PASS \ + __asm__ volatile( \ + "sll %[temp1], %[temp0], 3 \n\t" \ + "addiu %[temp3], %[streak], -3 \n\t" \ + "addu %[temp2], %[pstreaks], %[temp1] \n\t" \ + "blez %[temp3], 1f \n\t" \ + "srl %[temp1], %[temp1], 1 \n\t" \ + "addu %[temp3], %[pcnts], %[temp1] \n\t" \ + "lw %[temp0], 4(%[temp2]) \n\t" \ + "lw %[temp1], 0(%[temp3]) \n\t" \ + "addu %[temp0], %[temp0], %[streak] \n\t" \ + "addiu %[temp1], %[temp1], 1 \n\t" \ + "sw %[temp0], 4(%[temp2]) \n\t" \ + "sw %[temp1], 0(%[temp3]) \n\t" \ + "b 2f \n\t" \ + "1: \n\t" \ + "lw %[temp0], 0(%[temp2]) \n\t" \ + "addu %[temp0], %[temp0], %[streak] \n\t" \ + "sw %[temp0], 0(%[temp2]) \n\t" \ + "2: \n\t" \ + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp0]"+r"(temp0) \ + : [pstreaks]"r"(pstreaks), [pcnts]"r"(pcnts), \ + [streak]"r"(streak) \ + : "memory" \ + ); + +// Returns the various RLE counts +static WEBP_INLINE void GetEntropyUnrefinedHelper( + uint32_t val, int i, uint32_t* const val_prev, int* const i_prev, + VP8LBitEntropy* const bit_entropy, VP8LStreaks* const stats) { + int* const pstreaks = &stats->streaks[0][0]; + int* const pcnts = &stats->counts[0]; + int temp0, temp1, temp2, temp3; + const int streak = i - *i_prev; + + // Gather info for the bit entropy. + if (*val_prev != 0) { + bit_entropy->sum += (*val_prev) * streak; + bit_entropy->nonzeros += streak; + bit_entropy->nonzero_code = *i_prev; + bit_entropy->entropy -= VP8LFastSLog2(*val_prev) * streak; + if (bit_entropy->max_val < *val_prev) { + bit_entropy->max_val = *val_prev; + } + } + + // Gather info for the Huffman cost. + temp0 = (*val_prev != 0); + HUFFMAN_COST_PASS + + *val_prev = val; + *i_prev = i; +} + +static void GetEntropyUnrefined_MIPS32(const uint32_t X[], int length, + VP8LBitEntropy* const bit_entropy, + VP8LStreaks* const stats) { + int i; + int i_prev = 0; + uint32_t x_prev = X[0]; + + memset(stats, 0, sizeof(*stats)); + VP8LBitEntropyInit(bit_entropy); + + for (i = 1; i < length; ++i) { + const uint32_t x = X[i]; + if (x != x_prev) { + GetEntropyUnrefinedHelper(x, i, &x_prev, &i_prev, bit_entropy, stats); + } + } + GetEntropyUnrefinedHelper(0, i, &x_prev, &i_prev, bit_entropy, stats); + + bit_entropy->entropy += VP8LFastSLog2(bit_entropy->sum); +} + +static void GetCombinedEntropyUnrefined_MIPS32(const uint32_t X[], + const uint32_t Y[], + int length, + VP8LBitEntropy* const entropy, + VP8LStreaks* const stats) { + int i = 1; + int i_prev = 0; + uint32_t xy_prev = X[0] + Y[0]; + + memset(stats, 0, sizeof(*stats)); + VP8LBitEntropyInit(entropy); + + for (i = 1; i < length; ++i) { + const uint32_t xy = X[i] + Y[i]; + if (xy != xy_prev) { + GetEntropyUnrefinedHelper(xy, i, &xy_prev, &i_prev, entropy, stats); + } + } + GetEntropyUnrefinedHelper(0, i, &xy_prev, &i_prev, entropy, stats); + + entropy->entropy += VP8LFastSLog2(entropy->sum); +} + +#define ASM_START \ + __asm__ volatile( \ + ".set push \n\t" \ + ".set at \n\t" \ + ".set macro \n\t" \ + "1: \n\t" + +// P2 = P0 + P1 +// A..D - offsets +// E - temp variable to tell macro +// if pointer should be incremented +// literal_ and successive histograms could be unaligned +// so we must use ulw and usw +#define ADD_TO_OUT(A, B, C, D, E, P0, P1, P2) \ + "ulw %[temp0], " #A "(%[" #P0 "]) \n\t" \ + "ulw %[temp1], " #B "(%[" #P0 "]) \n\t" \ + "ulw %[temp2], " #C "(%[" #P0 "]) \n\t" \ + "ulw %[temp3], " #D "(%[" #P0 "]) \n\t" \ + "ulw %[temp4], " #A "(%[" #P1 "]) \n\t" \ + "ulw %[temp5], " #B "(%[" #P1 "]) \n\t" \ + "ulw %[temp6], " #C "(%[" #P1 "]) \n\t" \ + "ulw %[temp7], " #D "(%[" #P1 "]) \n\t" \ + "addu %[temp4], %[temp4], %[temp0] \n\t" \ + "addu %[temp5], %[temp5], %[temp1] \n\t" \ + "addu %[temp6], %[temp6], %[temp2] \n\t" \ + "addu %[temp7], %[temp7], %[temp3] \n\t" \ + "addiu %[" #P0 "], %[" #P0 "], 16 \n\t" \ + ".if " #E " == 1 \n\t" \ + "addiu %[" #P1 "], %[" #P1 "], 16 \n\t" \ + ".endif \n\t" \ + "usw %[temp4], " #A "(%[" #P2 "]) \n\t" \ + "usw %[temp5], " #B "(%[" #P2 "]) \n\t" \ + "usw %[temp6], " #C "(%[" #P2 "]) \n\t" \ + "usw %[temp7], " #D "(%[" #P2 "]) \n\t" \ + "addiu %[" #P2 "], %[" #P2 "], 16 \n\t" \ + "bne %[" #P0 "], %[LoopEnd], 1b \n\t" \ + ".set pop \n\t" \ + +#define ASM_END_COMMON_0 \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), \ + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), \ + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7), \ + [pa]"+r"(pa), [pout]"+r"(pout) + +#define ASM_END_COMMON_1 \ + : [LoopEnd]"r"(LoopEnd) \ + : "memory", "at" \ + ); + +#define ASM_END_0 \ + ASM_END_COMMON_0 \ + , [pb]"+r"(pb) \ + ASM_END_COMMON_1 + +#define ASM_END_1 \ + ASM_END_COMMON_0 \ + ASM_END_COMMON_1 + +static void AddVector_MIPS32(const uint32_t* pa, const uint32_t* pb, + uint32_t* pout, int size) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + const uint32_t end = ((size) / 4) * 4; + const uint32_t* const LoopEnd = pa + end; + int i; + ASM_START + ADD_TO_OUT(0, 4, 8, 12, 1, pa, pb, pout) + ASM_END_0 + for (i = end; i < size; ++i) pout[i] = pa[i] + pb[i]; +} + +static void AddVectorEq_MIPS32(const uint32_t* pa, uint32_t* pout, int size) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + const uint32_t end = ((size) / 4) * 4; + const uint32_t* const LoopEnd = pa + end; + int i; + ASM_START + ADD_TO_OUT(0, 4, 8, 12, 0, pa, pout, pout) + ASM_END_1 + for (i = end; i < size; ++i) pout[i] += pa[i]; +} + +#undef ASM_END_1 +#undef ASM_END_0 +#undef ASM_END_COMMON_1 +#undef ASM_END_COMMON_0 +#undef ADD_TO_OUT +#undef ASM_START + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPS32(void) { + VP8LFastSLog2Slow = FastSLog2Slow_MIPS32; + VP8LFastLog2Slow = FastLog2Slow_MIPS32; + VP8LExtraCost = ExtraCost_MIPS32; + VP8LExtraCostCombined = ExtraCostCombined_MIPS32; + VP8LGetEntropyUnrefined = GetEntropyUnrefined_MIPS32; + VP8LGetCombinedEntropyUnrefined = GetCombinedEntropyUnrefined_MIPS32; + VP8LAddVector = AddVector_MIPS32; + VP8LAddVectorEq = AddVectorEq_MIPS32; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/media/libwebp/dsp/lossless_enc_mips_dsp_r2.c b/media/libwebp/dsp/lossless_enc_mips_dsp_r2.c new file mode 100644 index 0000000000..157dfc2e01 --- /dev/null +++ b/media/libwebp/dsp/lossless_enc_mips_dsp_r2.c @@ -0,0 +1,281 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Image transform methods for lossless encoder. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../dsp/lossless.h" + +static void SubtractGreenFromBlueAndRed_MIPSdspR2(uint32_t* argb_data, + int num_pixels) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + uint32_t* const p_loop1_end = argb_data + (num_pixels & ~3); + uint32_t* const p_loop2_end = p_loop1_end + (num_pixels & 3); + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[argb_data], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[argb_data]) \n\t" + "lw %[temp1], 4(%[argb_data]) \n\t" + "lw %[temp2], 8(%[argb_data]) \n\t" + "lw %[temp3], 12(%[argb_data]) \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "ext %[temp5], %[temp1], 8, 8 \n\t" + "ext %[temp6], %[temp2], 8, 8 \n\t" + "ext %[temp7], %[temp3], 8, 8 \n\t" + "addiu %[argb_data], %[argb_data], 16 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "replv.ph %[temp5], %[temp5] \n\t" + "replv.ph %[temp6], %[temp6] \n\t" + "replv.ph %[temp7], %[temp7] \n\t" + "subu.qb %[temp0], %[temp0], %[temp4] \n\t" + "subu.qb %[temp1], %[temp1], %[temp5] \n\t" + "subu.qb %[temp2], %[temp2], %[temp6] \n\t" + "subu.qb %[temp3], %[temp3], %[temp7] \n\t" + "sw %[temp0], -16(%[argb_data]) \n\t" + "sw %[temp1], -12(%[argb_data]) \n\t" + "sw %[temp2], -8(%[argb_data]) \n\t" + "bne %[argb_data], %[p_loop1_end], 0b \n\t" + " sw %[temp3], -4(%[argb_data]) \n\t" + "3: \n\t" + "beq %[argb_data], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[argb_data]) \n\t" + "addiu %[argb_data], %[argb_data], 4 \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "subu.qb %[temp0], %[temp0], %[temp4] \n\t" + "bne %[argb_data], %[p_loop2_end], 1b \n\t" + " sw %[temp0], -4(%[argb_data]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [argb_data]"+&r"(argb_data), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), + [temp7]"=&r"(temp7) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static WEBP_INLINE uint32_t ColorTransformDelta(int8_t color_pred, + int8_t color) { + return (uint32_t)((int)(color_pred) * color) >> 5; +} + +static void TransformColor_MIPSdspR2(const VP8LMultipliers* const m, + uint32_t* data, int num_pixels) { + int temp0, temp1, temp2, temp3, temp4, temp5; + uint32_t argb, argb1, new_red, new_red1; + const uint32_t G_to_R = m->green_to_red_; + const uint32_t G_to_B = m->green_to_blue_; + const uint32_t R_to_B = m->red_to_blue_; + uint32_t* const p_loop_end = data + (num_pixels & ~1); + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[data], %[p_loop_end], 1f \n\t" + " nop \n\t" + "replv.ph %[temp0], %[G_to_R] \n\t" + "replv.ph %[temp1], %[G_to_B] \n\t" + "replv.ph %[temp2], %[R_to_B] \n\t" + "shll.ph %[temp0], %[temp0], 8 \n\t" + "shll.ph %[temp1], %[temp1], 8 \n\t" + "shll.ph %[temp2], %[temp2], 8 \n\t" + "shra.ph %[temp0], %[temp0], 8 \n\t" + "shra.ph %[temp1], %[temp1], 8 \n\t" + "shra.ph %[temp2], %[temp2], 8 \n\t" + "0: \n\t" + "lw %[argb], 0(%[data]) \n\t" + "lw %[argb1], 4(%[data]) \n\t" + "lhu %[new_red], 2(%[data]) \n\t" + "lhu %[new_red1], 6(%[data]) \n\t" + "precrq.qb.ph %[temp3], %[argb], %[argb1] \n\t" + "precr.qb.ph %[temp4], %[argb], %[argb1] \n\t" + "preceu.ph.qbra %[temp3], %[temp3] \n\t" + "preceu.ph.qbla %[temp4], %[temp4] \n\t" + "shll.ph %[temp3], %[temp3], 8 \n\t" + "shll.ph %[temp4], %[temp4], 8 \n\t" + "shra.ph %[temp3], %[temp3], 8 \n\t" + "shra.ph %[temp4], %[temp4], 8 \n\t" + "mul.ph %[temp5], %[temp3], %[temp0] \n\t" + "mul.ph %[temp3], %[temp3], %[temp1] \n\t" + "mul.ph %[temp4], %[temp4], %[temp2] \n\t" + "addiu %[data], %[data], 8 \n\t" + "ins %[new_red1], %[new_red], 16, 16 \n\t" + "ins %[argb1], %[argb], 16, 16 \n\t" + "shra.ph %[temp5], %[temp5], 5 \n\t" + "shra.ph %[temp3], %[temp3], 5 \n\t" + "shra.ph %[temp4], %[temp4], 5 \n\t" + "subu.ph %[new_red1], %[new_red1], %[temp5] \n\t" + "subu.ph %[argb1], %[argb1], %[temp3] \n\t" + "preceu.ph.qbra %[temp5], %[new_red1] \n\t" + "subu.ph %[argb1], %[argb1], %[temp4] \n\t" + "preceu.ph.qbra %[temp3], %[argb1] \n\t" + "sb %[temp5], -2(%[data]) \n\t" + "sb %[temp3], -4(%[data]) \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "sra %[temp3], %[temp3], 16 \n\t" + "sb %[temp5], -6(%[data]) \n\t" + "bne %[data], %[p_loop_end], 0b \n\t" + " sb %[temp3], -8(%[data]) \n\t" + "1: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [new_red1]"=&r"(new_red1), [new_red]"=&r"(new_red), + [argb]"=&r"(argb), [argb1]"=&r"(argb1), [data]"+&r"(data) + : [G_to_R]"r"(G_to_R), [R_to_B]"r"(R_to_B), + [G_to_B]"r"(G_to_B), [p_loop_end]"r"(p_loop_end) + : "memory", "hi", "lo" + ); + + if (num_pixels & 1) { + const uint32_t argb_ = data[0]; + const uint32_t green = argb_ >> 8; + const uint32_t red = argb_ >> 16; + uint32_t new_blue = argb_; + new_red = red; + new_red -= ColorTransformDelta(m->green_to_red_, green); + new_red &= 0xff; + new_blue -= ColorTransformDelta(m->green_to_blue_, green); + new_blue -= ColorTransformDelta(m->red_to_blue_, red); + new_blue &= 0xff; + data[0] = (argb_ & 0xff00ff00u) | (new_red << 16) | (new_blue); + } +} + +static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue, + uint8_t red_to_blue, + uint32_t argb) { + const uint32_t green = argb >> 8; + const uint32_t red = argb >> 16; + uint8_t new_blue = argb; + new_blue -= ColorTransformDelta(green_to_blue, green); + new_blue -= ColorTransformDelta(red_to_blue, red); + return (new_blue & 0xff); +} + +static void CollectColorBlueTransforms_MIPSdspR2(const uint32_t* argb, + int stride, + int tile_width, + int tile_height, + int green_to_blue, + int red_to_blue, + int histo[]) { + const int rtb = (red_to_blue << 16) | (red_to_blue & 0xffff); + const int gtb = (green_to_blue << 16) | (green_to_blue & 0xffff); + const uint32_t mask = 0xff00ffu; + while (tile_height-- > 0) { + int x; + const uint32_t* p_argb = argb; + argb += stride; + for (x = 0; x < (tile_width >> 1); ++x) { + int temp0, temp1, temp2, temp3, temp4, temp5, temp6; + __asm__ volatile ( + "lw %[temp0], 0(%[p_argb]) \n\t" + "lw %[temp1], 4(%[p_argb]) \n\t" + "precr.qb.ph %[temp2], %[temp0], %[temp1] \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "shra.ph %[temp2], %[temp2], 8 \n\t" + "shra.ph %[temp3], %[temp1], 8 \n\t" + "mul.ph %[temp5], %[temp2], %[rtb] \n\t" + "mul.ph %[temp6], %[temp3], %[gtb] \n\t" + "and %[temp4], %[temp1], %[mask] \n\t" + "addiu %[p_argb], %[p_argb], 8 \n\t" + "shra.ph %[temp5], %[temp5], 5 \n\t" + "shra.ph %[temp6], %[temp6], 5 \n\t" + "subu.qb %[temp2], %[temp4], %[temp5] \n\t" + "subu.qb %[temp2], %[temp2], %[temp6] \n\t" + : [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), + [temp5]"=&r"(temp5), [temp6]"=&r"(temp6) + : [rtb]"r"(rtb), [gtb]"r"(gtb), [mask]"r"(mask) + : "memory", "hi", "lo" + ); + ++histo[(uint8_t)(temp2 >> 16)]; + ++histo[(uint8_t)temp2]; + } + if (tile_width & 1) { + ++histo[TransformColorBlue(green_to_blue, red_to_blue, *p_argb)]; + } + } +} + +static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red, + uint32_t argb) { + const uint32_t green = argb >> 8; + uint32_t new_red = argb >> 16; + new_red -= ColorTransformDelta(green_to_red, green); + return (new_red & 0xff); +} + +static void CollectColorRedTransforms_MIPSdspR2(const uint32_t* argb, + int stride, + int tile_width, + int tile_height, + int green_to_red, + int histo[]) { + const int gtr = (green_to_red << 16) | (green_to_red & 0xffff); + while (tile_height-- > 0) { + int x; + const uint32_t* p_argb = argb; + argb += stride; + for (x = 0; x < (tile_width >> 1); ++x) { + int temp0, temp1, temp2, temp3, temp4; + __asm__ volatile ( + "lw %[temp0], 0(%[p_argb]) \n\t" + "lw %[temp1], 4(%[p_argb]) \n\t" + "precrq.ph.w %[temp4], %[temp0], %[temp1] \n\t" + "ins %[temp1], %[temp0], 16, 16 \n\t" + "shra.ph %[temp3], %[temp1], 8 \n\t" + "mul.ph %[temp2], %[temp3], %[gtr] \n\t" + "addiu %[p_argb], %[p_argb], 8 \n\t" + "shra.ph %[temp2], %[temp2], 5 \n\t" + "subu.qb %[temp2], %[temp4], %[temp2] \n\t" + : [p_argb]"+&r"(p_argb), [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4) + : [gtr]"r"(gtr) + : "memory", "hi", "lo" + ); + ++histo[(uint8_t)(temp2 >> 16)]; + ++histo[(uint8_t)temp2]; + } + if (tile_width & 1) { + ++histo[TransformColorRed(green_to_red, *p_argb)]; + } + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMIPSdspR2(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_MIPSdspR2; + VP8LTransformColor = TransformColor_MIPSdspR2; + VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_MIPSdspR2; + VP8LCollectColorRedTransforms = CollectColorRedTransforms_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/lossless_enc_msa.c b/media/libwebp/dsp/lossless_enc_msa.c new file mode 100644 index 0000000000..f8a5f8c56f --- /dev/null +++ b/media/libwebp/dsp/lossless_enc_msa.c @@ -0,0 +1,148 @@ +// Copyright 2016 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MSA variant of Image transform methods for lossless encoder. +// +// Authors: Prashant Patil (Prashant.Patil@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) + +#include "../dsp/lossless.h" +#include "../dsp/msa_macro.h" + +#define TRANSFORM_COLOR_8(src0, src1, dst0, dst1, c0, c1, mask0, mask1) do { \ + v8i16 g0, g1, t0, t1, t2, t3; \ + v4i32 t4, t5; \ + VSHF_B2_SH(src0, src0, src1, src1, mask0, mask0, g0, g1); \ + DOTP_SB2_SH(g0, g1, c0, c0, t0, t1); \ + SRAI_H2_SH(t0, t1, 5); \ + t0 = __msa_subv_h((v8i16)src0, t0); \ + t1 = __msa_subv_h((v8i16)src1, t1); \ + t4 = __msa_srli_w((v4i32)src0, 16); \ + t5 = __msa_srli_w((v4i32)src1, 16); \ + DOTP_SB2_SH(t4, t5, c1, c1, t2, t3); \ + SRAI_H2_SH(t2, t3, 5); \ + SUB2(t0, t2, t1, t3, t0, t1); \ + VSHF_B2_UB(src0, t0, src1, t1, mask1, mask1, dst0, dst1); \ +} while (0) + +#define TRANSFORM_COLOR_4(src, dst, c0, c1, mask0, mask1) do { \ + const v16i8 g0 = VSHF_SB(src, src, mask0); \ + v8i16 t0 = __msa_dotp_s_h(c0, g0); \ + v8i16 t1; \ + v4i32 t2; \ + t0 = SRAI_H(t0, 5); \ + t0 = __msa_subv_h((v8i16)src, t0); \ + t2 = __msa_srli_w((v4i32)src, 16); \ + t1 = __msa_dotp_s_h(c1, (v16i8)t2); \ + t1 = SRAI_H(t1, 5); \ + t0 = t0 - t1; \ + dst = VSHF_UB(src, t0, mask1); \ +} while (0) + +static void TransformColor_MSA(const VP8LMultipliers* const m, uint32_t* data, + int num_pixels) { + v16u8 src0, dst0; + const v16i8 g2br = (v16i8)__msa_fill_w(m->green_to_blue_ | + (m->green_to_red_ << 16)); + const v16i8 r2b = (v16i8)__msa_fill_w(m->red_to_blue_); + const v16u8 mask0 = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, + 13, 255, 13, 255 }; + const v16u8 mask1 = { 16, 1, 18, 3, 20, 5, 22, 7, 24, 9, 26, 11, + 28, 13, 30, 15 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1; + LD_UB2(data, 4, src0, src1); + TRANSFORM_COLOR_8(src0, src1, dst0, dst1, g2br, r2b, mask0, mask1); + ST_UB2(dst0, dst1, data, 4); + data += 8; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(data); + TRANSFORM_COLOR_4(src0, dst0, g2br, r2b, mask0, mask1); + ST_UB(dst0, data); + data += 4; + num_pixels -= 4; + } + if (num_pixels > 0) { + src0 = LD_UB(data); + TRANSFORM_COLOR_4(src0, dst0, g2br, r2b, mask0, mask1); + if (num_pixels == 3) { + const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); + const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2); + SD(pix_d, data + 0); + SW(pix_w, data + 2); + } else if (num_pixels == 2) { + const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); + SD(pix_d, data); + } else { + const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 0); + SW(pix_w, data); + } + } + } +} + +static void SubtractGreenFromBlueAndRed_MSA(uint32_t* argb_data, + int num_pixels) { + int i; + uint8_t* ptemp_data = (uint8_t*)argb_data; + v16u8 src0, dst0, tmp0; + const v16u8 mask = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, + 13, 255, 13, 255 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1, tmp1; + LD_UB2(ptemp_data, 16, src0, src1); + VSHF_B2_UB(src0, src1, src1, src0, mask, mask, tmp0, tmp1); + SUB2(src0, tmp0, src1, tmp1, dst0, dst1); + ST_UB2(dst0, dst1, ptemp_data, 16); + ptemp_data += 8 * 4; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(ptemp_data); + tmp0 = VSHF_UB(src0, src0, mask); + dst0 = src0 - tmp0; + ST_UB(dst0, ptemp_data); + ptemp_data += 4 * 4; + num_pixels -= 4; + } + for (i = 0; i < num_pixels; i++) { + const uint8_t b = ptemp_data[0]; + const uint8_t g = ptemp_data[1]; + const uint8_t r = ptemp_data[2]; + ptemp_data[0] = (b - g) & 0xff; + ptemp_data[2] = (r - g) & 0xff; + ptemp_data += 4; + } + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitMSA(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_MSA; + VP8LTransformColor = TransformColor_MSA; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(VP8LEncDspInitMSA) + +#endif // WEBP_USE_MSA diff --git a/media/libwebp/dsp/lossless_enc_neon.c b/media/libwebp/dsp/lossless_enc_neon.c new file mode 100644 index 0000000000..89d5439e49 --- /dev/null +++ b/media/libwebp/dsp/lossless_enc_neon.c @@ -0,0 +1,144 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// NEON variant of methods for lossless encoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_NEON) + +#include <arm_neon.h> + +#include "../dsp/lossless.h" +#include "../dsp/neon.h" + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use +// non-standard versions there. +#if defined(__APPLE__) && defined(__aarch64__) && \ + defined(__apple_build_version__) && (__apple_build_version__< 6020037) +#define USE_VTBLQ +#endif + +#ifdef USE_VTBLQ +// 255 = byte will be zeroed +static const uint8_t kGreenShuffle[16] = { + 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255 +}; + +static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb, + const uint8x16_t shuffle) { + return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)), + vtbl1q_u8(argb, vget_high_u8(shuffle))); +} +#else // !USE_VTBLQ +// 255 = byte will be zeroed +static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 }; + +static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb, + const uint8x8_t shuffle) { + return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle), + vtbl1_u8(vget_high_u8(argb), shuffle)); +} +#endif // USE_VTBLQ + +static void SubtractGreenFromBlueAndRed_NEON(uint32_t* argb_data, + int num_pixels) { + const uint32_t* const end = argb_data + (num_pixels & ~3); +#ifdef USE_VTBLQ + const uint8x16_t shuffle = vld1q_u8(kGreenShuffle); +#else + const uint8x8_t shuffle = vld1_u8(kGreenShuffle); +#endif + for (; argb_data < end; argb_data += 4) { + const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data); + const uint8x16_t greens = DoGreenShuffle_NEON(argb, shuffle); + vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens)); + } + // fallthrough and finish off with plain-C + VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3); +} + +//------------------------------------------------------------------------------ +// Color Transform + +static void TransformColor_NEON(const VP8LMultipliers* const m, + uint32_t* argb_data, int num_pixels) { + // sign-extended multiplying constants, pre-shifted by 6. +#define CST(X) (((int16_t)(m->X << 8)) >> 6) + const int16_t rb[8] = { + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_), + CST(green_to_blue_), CST(green_to_red_) + }; + const int16x8_t mults_rb = vld1q_s16(rb); + const int16_t b2[8] = { + 0, CST(red_to_blue_), 0, CST(red_to_blue_), + 0, CST(red_to_blue_), 0, CST(red_to_blue_), + }; + const int16x8_t mults_b2 = vld1q_s16(b2); +#undef CST +#ifdef USE_VTBLQ + static const uint8_t kg0g0[16] = { + 255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13 + }; + const uint8x16_t shuffle = vld1q_u8(kg0g0); +#else + static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 }; + const uint8x8_t shuffle = vld1_u8(k0g0g); +#endif + const uint32x4_t mask_rb = vdupq_n_u32(0x00ff00ffu); // red-blue masks + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i)); + // 0 g 0 g + const uint8x16_t greens = DoGreenShuffle_NEON(in, shuffle); + // x dr x db1 + const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb); + // r 0 b 0 + const int16x8_t B = vshlq_n_s16(vreinterpretq_s16_u8(in), 8); + // x db2 0 0 + const int16x8_t C = vqdmulhq_s16(B, mults_b2); + // 0 0 x db2 + const uint32x4_t D = vshrq_n_u32(vreinterpretq_u32_s16(C), 16); + // x dr x db + const int8x16_t E = vaddq_s8(vreinterpretq_s8_u32(D), + vreinterpretq_s8_s16(A)); + // 0 dr 0 db + const uint32x4_t F = vandq_u32(vreinterpretq_u32_s8(E), mask_rb); + const int8x16_t out = vsubq_s8(vreinterpretq_s8_u8(in), + vreinterpretq_s8_u32(F)); + vst1q_s8((int8_t*)(argb_data + i), out); + } + // fallthrough and finish off with plain-C + VP8LTransformColor_C(m, argb_data + i, num_pixels - i); +} + +#undef USE_VTBLQ + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitNEON(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_NEON; + VP8LTransformColor = TransformColor_NEON; +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(VP8LEncDspInitNEON) + +#endif // WEBP_USE_NEON diff --git a/media/libwebp/dsp/lossless_enc_sse2.c b/media/libwebp/dsp/lossless_enc_sse2.c new file mode 100644 index 0000000000..665ceb669c --- /dev/null +++ b/media/libwebp/dsp/lossless_enc_sse2.c @@ -0,0 +1,669 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 variant of methods for lossless encoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_SSE2) +#include <assert.h> +#include <emmintrin.h> +#include "../dsp/lossless.h" +#include "../dsp/common_sse2.h" +#include "../dsp/lossless_common.h" + +// For sign-extended multiplying constants, pre-shifted by 5: +#define CST_5b(X) (((int16_t)((uint16_t)(X) << 8)) >> 5) + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +static void SubtractGreenFromBlueAndRed_SSE2(uint32_t* argb_data, + int num_pixels) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb + const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g + const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); + const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g + const __m128i out = _mm_sub_epi8(in, C); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + if (i != num_pixels) { + VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i); + } +} + +//------------------------------------------------------------------------------ +// Color Transform + +#define MK_CST_16(HI, LO) \ + _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff))) + +static void TransformColor_SSE2(const VP8LMultipliers* const m, + uint32_t* argb_data, int num_pixels) { + const __m128i mults_rb = MK_CST_16(CST_5b(m->green_to_red_), + CST_5b(m->green_to_blue_)); + const __m128i mults_b2 = MK_CST_16(CST_5b(m->red_to_blue_), 0); + const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks + const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff); // red-blue masks + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb + const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0 + const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); + const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0 + const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1 + const __m128i E = _mm_slli_epi16(in, 8); // r 0 b 0 + const __m128i F = _mm_mulhi_epi16(E, mults_b2); // x db2 0 0 + const __m128i G = _mm_srli_epi32(F, 16); // 0 0 x db2 + const __m128i H = _mm_add_epi8(G, D); // x dr x db + const __m128i I = _mm_and_si128(H, mask_rb); // 0 dr 0 db + const __m128i out = _mm_sub_epi8(in, I); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + if (i != num_pixels) { + VP8LTransformColor_C(m, argb_data + i, num_pixels - i); + } +} + +//------------------------------------------------------------------------------ +#define SPAN 8 +static void CollectColorBlueTransforms_SSE2(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, + int histo[]) { + const __m128i mults_r = MK_CST_16(CST_5b(red_to_blue), 0); + const __m128i mults_g = MK_CST_16(0, CST_5b(green_to_blue)); + const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask + const __m128i mask_b = _mm_set1_epi32(0x0000ff); // blue mask + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + int i, x; + for (x = 0; x + SPAN <= tile_width; x += SPAN) { + uint16_t values[SPAN]; + const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); + const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); + const __m128i A0 = _mm_slli_epi16(in0, 8); // r 0 | b 0 + const __m128i A1 = _mm_slli_epi16(in1, 8); + const __m128i B0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0 + const __m128i B1 = _mm_and_si128(in1, mask_g); + const __m128i C0 = _mm_mulhi_epi16(A0, mults_r); // x db | 0 0 + const __m128i C1 = _mm_mulhi_epi16(A1, mults_r); + const __m128i D0 = _mm_mulhi_epi16(B0, mults_g); // 0 0 | x db + const __m128i D1 = _mm_mulhi_epi16(B1, mults_g); + const __m128i E0 = _mm_sub_epi8(in0, D0); // x x | x b' + const __m128i E1 = _mm_sub_epi8(in1, D1); + const __m128i F0 = _mm_srli_epi32(C0, 16); // 0 0 | x db + const __m128i F1 = _mm_srli_epi32(C1, 16); + const __m128i G0 = _mm_sub_epi8(E0, F0); // 0 0 | x b' + const __m128i G1 = _mm_sub_epi8(E1, F1); + const __m128i H0 = _mm_and_si128(G0, mask_b); // 0 0 | 0 b + const __m128i H1 = _mm_and_si128(G1, mask_b); + const __m128i I = _mm_packs_epi32(H0, H1); // 0 b' | 0 b' + _mm_storeu_si128((__m128i*)values, I); + for (i = 0; i < SPAN; ++i) ++histo[values[i]]; + } + } + { + const int left_over = tile_width & (SPAN - 1); + if (left_over > 0) { + VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride, + left_over, tile_height, + green_to_blue, red_to_blue, histo); + } + } +} + +static void CollectColorRedTransforms_SSE2(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]) { + const __m128i mults_g = MK_CST_16(0, CST_5b(green_to_red)); + const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask + const __m128i mask = _mm_set1_epi32(0xff); + + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + int i, x; + for (x = 0; x + SPAN <= tile_width; x += SPAN) { + uint16_t values[SPAN]; + const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]); + const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]); + const __m128i A0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0 + const __m128i A1 = _mm_and_si128(in1, mask_g); + const __m128i B0 = _mm_srli_epi32(in0, 16); // 0 0 | x r + const __m128i B1 = _mm_srli_epi32(in1, 16); + const __m128i C0 = _mm_mulhi_epi16(A0, mults_g); // 0 0 | x dr + const __m128i C1 = _mm_mulhi_epi16(A1, mults_g); + const __m128i E0 = _mm_sub_epi8(B0, C0); // x x | x r' + const __m128i E1 = _mm_sub_epi8(B1, C1); + const __m128i F0 = _mm_and_si128(E0, mask); // 0 0 | 0 r' + const __m128i F1 = _mm_and_si128(E1, mask); + const __m128i I = _mm_packs_epi32(F0, F1); + _mm_storeu_si128((__m128i*)values, I); + for (i = 0; i < SPAN; ++i) ++histo[values[i]]; + } + } + { + const int left_over = tile_width & (SPAN - 1); + if (left_over > 0) { + VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride, + left_over, tile_height, + green_to_red, histo); + } + } +} +#undef SPAN +#undef MK_CST_16 + +//------------------------------------------------------------------------------ + +// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But +// that's ok since the histogram values are less than 1<<28 (max picture size). +#define LINE_SIZE 16 // 8 or 16 +static void AddVector_SSE2(const uint32_t* a, const uint32_t* b, uint32_t* out, + int size) { + int i; + for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) { + const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]); + const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]); +#if (LINE_SIZE == 16) + const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]); + const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]); +#endif + const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]); + const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]); +#if (LINE_SIZE == 16) + const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i + 8]); + const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]); +#endif + _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0)); + _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1)); +#if (LINE_SIZE == 16) + _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2)); + _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3)); +#endif + } + for (; i < size; ++i) { + out[i] = a[i] + b[i]; + } +} + +static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) { + int i; + for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) { + const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]); + const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]); +#if (LINE_SIZE == 16) + const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]); + const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]); +#endif + const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i + 0]); + const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i + 4]); +#if (LINE_SIZE == 16) + const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i + 8]); + const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]); +#endif + _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0)); + _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1)); +#if (LINE_SIZE == 16) + _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2)); + _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3)); +#endif + } + for (; i < size; ++i) { + out[i] += a[i]; + } +} +#undef LINE_SIZE + +//------------------------------------------------------------------------------ +// Entropy + +// TODO(https://crbug.com/webp/499): this function produces different results +// from the C code due to use of double/float resulting in output differences +// when compared to -noasm. +#if !(defined(WEBP_HAVE_SLOW_CLZ_CTZ) || defined(__i386__) || defined(_M_IX86)) + +static float CombinedShannonEntropy_SSE2(const int X[256], const int Y[256]) { + int i; + double retval = 0.; + int sumX = 0, sumXY = 0; + const __m128i zero = _mm_setzero_si128(); + + for (i = 0; i < 256; i += 16) { + const __m128i x0 = _mm_loadu_si128((const __m128i*)(X + i + 0)); + const __m128i y0 = _mm_loadu_si128((const __m128i*)(Y + i + 0)); + const __m128i x1 = _mm_loadu_si128((const __m128i*)(X + i + 4)); + const __m128i y1 = _mm_loadu_si128((const __m128i*)(Y + i + 4)); + const __m128i x2 = _mm_loadu_si128((const __m128i*)(X + i + 8)); + const __m128i y2 = _mm_loadu_si128((const __m128i*)(Y + i + 8)); + const __m128i x3 = _mm_loadu_si128((const __m128i*)(X + i + 12)); + const __m128i y3 = _mm_loadu_si128((const __m128i*)(Y + i + 12)); + const __m128i x4 = _mm_packs_epi16(_mm_packs_epi32(x0, x1), + _mm_packs_epi32(x2, x3)); + const __m128i y4 = _mm_packs_epi16(_mm_packs_epi32(y0, y1), + _mm_packs_epi32(y2, y3)); + const int32_t mx = _mm_movemask_epi8(_mm_cmpgt_epi8(x4, zero)); + int32_t my = _mm_movemask_epi8(_mm_cmpgt_epi8(y4, zero)) | mx; + while (my) { + const int32_t j = BitsCtz(my); + int xy; + if ((mx >> j) & 1) { + const int x = X[i + j]; + sumXY += x; + retval -= VP8LFastSLog2(x); + } + xy = X[i + j] + Y[i + j]; + sumX += xy; + retval -= VP8LFastSLog2(xy); + my &= my - 1; + } + } + retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY); + return (float)retval; +} + +#else + +#define DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC // won't be faster + +#endif + +//------------------------------------------------------------------------------ + +static int VectorMismatch_SSE2(const uint32_t* const array1, + const uint32_t* const array2, int length) { + int match_len; + + if (length >= 12) { + __m128i A0 = _mm_loadu_si128((const __m128i*)&array1[0]); + __m128i A1 = _mm_loadu_si128((const __m128i*)&array2[0]); + match_len = 0; + do { + // Loop unrolling and early load both provide a speedup of 10% for the + // current function. Also, max_limit can be MAX_LENGTH=4096 at most. + const __m128i cmpA = _mm_cmpeq_epi32(A0, A1); + const __m128i B0 = + _mm_loadu_si128((const __m128i*)&array1[match_len + 4]); + const __m128i B1 = + _mm_loadu_si128((const __m128i*)&array2[match_len + 4]); + if (_mm_movemask_epi8(cmpA) != 0xffff) break; + match_len += 4; + + { + const __m128i cmpB = _mm_cmpeq_epi32(B0, B1); + A0 = _mm_loadu_si128((const __m128i*)&array1[match_len + 4]); + A1 = _mm_loadu_si128((const __m128i*)&array2[match_len + 4]); + if (_mm_movemask_epi8(cmpB) != 0xffff) break; + match_len += 4; + } + } while (match_len + 12 < length); + } else { + match_len = 0; + // Unroll the potential first two loops. + if (length >= 4 && + _mm_movemask_epi8(_mm_cmpeq_epi32( + _mm_loadu_si128((const __m128i*)&array1[0]), + _mm_loadu_si128((const __m128i*)&array2[0]))) == 0xffff) { + match_len = 4; + if (length >= 8 && + _mm_movemask_epi8(_mm_cmpeq_epi32( + _mm_loadu_si128((const __m128i*)&array1[4]), + _mm_loadu_si128((const __m128i*)&array2[4]))) == 0xffff) { + match_len = 8; + } + } + } + + while (match_len < length && array1[match_len] == array2[match_len]) { + ++match_len; + } + return match_len; +} + +// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel. +static void BundleColorMap_SSE2(const uint8_t* const row, int width, int xbits, + uint32_t* dst) { + int x; + assert(xbits >= 0); + assert(xbits <= 3); + switch (xbits) { + case 0: { + const __m128i ff = _mm_set1_epi16((short)0xff00); + const __m128i zero = _mm_setzero_si128(); + // Store 0xff000000 | (row[x] << 8). + for (x = 0; x + 16 <= width; x += 16, dst += 16) { + const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]); + const __m128i in_lo = _mm_unpacklo_epi8(zero, in); + const __m128i dst0 = _mm_unpacklo_epi16(in_lo, ff); + const __m128i dst1 = _mm_unpackhi_epi16(in_lo, ff); + const __m128i in_hi = _mm_unpackhi_epi8(zero, in); + const __m128i dst2 = _mm_unpacklo_epi16(in_hi, ff); + const __m128i dst3 = _mm_unpackhi_epi16(in_hi, ff); + _mm_storeu_si128((__m128i*)&dst[0], dst0); + _mm_storeu_si128((__m128i*)&dst[4], dst1); + _mm_storeu_si128((__m128i*)&dst[8], dst2); + _mm_storeu_si128((__m128i*)&dst[12], dst3); + } + break; + } + case 1: { + const __m128i ff = _mm_set1_epi16((short)0xff00); + const __m128i mul = _mm_set1_epi16(0x110); + for (x = 0; x + 16 <= width; x += 16, dst += 8) { + // 0a0b | (where a/b are 4 bits). + const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]); + const __m128i tmp = _mm_mullo_epi16(in, mul); // aba0 + const __m128i pack = _mm_and_si128(tmp, ff); // ab00 + const __m128i dst0 = _mm_unpacklo_epi16(pack, ff); + const __m128i dst1 = _mm_unpackhi_epi16(pack, ff); + _mm_storeu_si128((__m128i*)&dst[0], dst0); + _mm_storeu_si128((__m128i*)&dst[4], dst1); + } + break; + } + case 2: { + const __m128i mask_or = _mm_set1_epi32(0xff000000); + const __m128i mul_cst = _mm_set1_epi16(0x0104); + const __m128i mask_mul = _mm_set1_epi16(0x0f00); + for (x = 0; x + 16 <= width; x += 16, dst += 4) { + // 000a000b000c000d | (where a/b/c/d are 2 bits). + const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]); + const __m128i mul = _mm_mullo_epi16(in, mul_cst); // 00ab00b000cd00d0 + const __m128i tmp = _mm_and_si128(mul, mask_mul); // 00ab000000cd0000 + const __m128i shift = _mm_srli_epi32(tmp, 12); // 00000000ab000000 + const __m128i pack = _mm_or_si128(shift, tmp); // 00000000abcd0000 + // Convert to 0xff00**00. + const __m128i res = _mm_or_si128(pack, mask_or); + _mm_storeu_si128((__m128i*)dst, res); + } + break; + } + default: { + assert(xbits == 3); + for (x = 0; x + 16 <= width; x += 16, dst += 2) { + // 0000000a00000000b... | (where a/b are 1 bit). + const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]); + const __m128i shift = _mm_slli_epi64(in, 7); + const uint32_t move = _mm_movemask_epi8(shift); + dst[0] = 0xff000000 | ((move & 0xff) << 8); + dst[1] = 0xff000000 | (move & 0xff00); + } + break; + } + } + if (x != width) { + VP8LBundleColorMap_C(row + x, width - x, xbits, dst); + } +} + +//------------------------------------------------------------------------------ +// Batch version of Predictor Transform subtraction + +static WEBP_INLINE void Average2_m128i(const __m128i* const a0, + const __m128i* const a1, + __m128i* const avg) { + // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1) + const __m128i ones = _mm_set1_epi8(1); + const __m128i avg1 = _mm_avg_epu8(*a0, *a1); + const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones); + *avg = _mm_sub_epi8(avg1, one); +} + +// Predictor0: ARGB_BLACK. +static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + const __m128i black = _mm_set1_epi32(ARGB_BLACK); + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + const __m128i res = _mm_sub_epi8(src, black); + _mm_storeu_si128((__m128i*)&out[i], res); + } + if (i != num_pixels) { + VP8LPredictorsSub_C[0](in + i, NULL, num_pixels - i, out + i); + } + (void)upper; +} + +#define GENERATE_PREDICTOR_1(X, IN) \ + static void PredictorSub##X##_SSE2(const uint32_t* const in, \ + const uint32_t* const upper, \ + int num_pixels, uint32_t* const out) { \ + int i; \ + for (i = 0; i + 4 <= num_pixels; i += 4) { \ + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ + const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN)); \ + const __m128i res = _mm_sub_epi8(src, pred); \ + _mm_storeu_si128((__m128i*)&out[i], res); \ + } \ + if (i != num_pixels) { \ + VP8LPredictorsSub_C[(X)](in + i, WEBP_OFFSET_PTR(upper, i), \ + num_pixels - i, out + i); \ + } \ + } + +GENERATE_PREDICTOR_1(1, in[i - 1]) // Predictor1: L +GENERATE_PREDICTOR_1(2, upper[i]) // Predictor2: T +GENERATE_PREDICTOR_1(3, upper[i + 1]) // Predictor3: TR +GENERATE_PREDICTOR_1(4, upper[i - 1]) // Predictor4: TL +#undef GENERATE_PREDICTOR_1 + +// Predictor5: avg2(avg2(L, TR), T) +static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]); + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + __m128i avg, pred, res; + Average2_m128i(&L, &TR, &avg); + Average2_m128i(&avg, &T, &pred); + res = _mm_sub_epi8(src, pred); + _mm_storeu_si128((__m128i*)&out[i], res); + } + if (i != num_pixels) { + VP8LPredictorsSub_C[5](in + i, upper + i, num_pixels - i, out + i); + } +} + +#define GENERATE_PREDICTOR_2(X, A, B) \ +static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \ + int num_pixels, uint32_t* out) { \ + int i; \ + for (i = 0; i + 4 <= num_pixels; i += 4) { \ + const __m128i tA = _mm_loadu_si128((const __m128i*)&(A)); \ + const __m128i tB = _mm_loadu_si128((const __m128i*)&(B)); \ + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ + __m128i pred, res; \ + Average2_m128i(&tA, &tB, &pred); \ + res = _mm_sub_epi8(src, pred); \ + _mm_storeu_si128((__m128i*)&out[i], res); \ + } \ + if (i != num_pixels) { \ + VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \ + } \ +} + +GENERATE_PREDICTOR_2(6, in[i - 1], upper[i - 1]) // Predictor6: avg(L, TL) +GENERATE_PREDICTOR_2(7, in[i - 1], upper[i]) // Predictor7: avg(L, T) +GENERATE_PREDICTOR_2(8, upper[i - 1], upper[i]) // Predictor8: avg(TL, T) +GENERATE_PREDICTOR_2(9, upper[i], upper[i + 1]) // Predictor9: average(T, TR) +#undef GENERATE_PREDICTOR_2 + +// Predictor10: avg(avg(L,TL), avg(T, TR)). +static void PredictorSub10_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]); + __m128i avgTTR, avgLTL, avg, res; + Average2_m128i(&T, &TR, &avgTTR); + Average2_m128i(&L, &TL, &avgLTL); + Average2_m128i(&avgTTR, &avgLTL, &avg); + res = _mm_sub_epi8(src, avg); + _mm_storeu_si128((__m128i*)&out[i], res); + } + if (i != num_pixels) { + VP8LPredictorsSub_C[10](in + i, upper + i, num_pixels - i, out + i); + } +} + +// Predictor11: select. +static void GetSumAbsDiff32_SSE2(const __m128i* const A, const __m128i* const B, + __m128i* const out) { + // We can unpack with any value on the upper 32 bits, provided it's the same + // on both operands (to that their sum of abs diff is zero). Here we use *A. + const __m128i A_lo = _mm_unpacklo_epi32(*A, *A); + const __m128i B_lo = _mm_unpacklo_epi32(*B, *A); + const __m128i A_hi = _mm_unpackhi_epi32(*A, *A); + const __m128i B_hi = _mm_unpackhi_epi32(*B, *A); + const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo); + const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi); + *out = _mm_packs_epi32(s_lo, s_hi); +} + +static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + __m128i pa, pb; + GetSumAbsDiff32_SSE2(&T, &TL, &pa); // pa = sum |T-TL| + GetSumAbsDiff32_SSE2(&L, &TL, &pb); // pb = sum |L-TL| + { + const __m128i mask = _mm_cmpgt_epi32(pb, pa); + const __m128i A = _mm_and_si128(mask, L); + const __m128i B = _mm_andnot_si128(mask, T); + const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T + const __m128i res = _mm_sub_epi8(src, pred); + _mm_storeu_si128((__m128i*)&out[i], res); + } + } + if (i != num_pixels) { + VP8LPredictorsSub_C[11](in + i, upper + i, num_pixels - i, out + i); + } +} + +// Predictor12: ClampedSubSubtractFull. +static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + const __m128i zero = _mm_setzero_si128(); + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); + const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]); + const __m128i L_lo = _mm_unpacklo_epi8(L, zero); + const __m128i L_hi = _mm_unpackhi_epi8(L, zero); + const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); + const __m128i T_lo = _mm_unpacklo_epi8(T, zero); + const __m128i T_hi = _mm_unpackhi_epi8(T, zero); + const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); + const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero); + const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero); + const __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo); + const __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi); + const __m128i pred_lo = _mm_add_epi16(L_lo, diff_lo); + const __m128i pred_hi = _mm_add_epi16(L_hi, diff_hi); + const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi); + const __m128i res = _mm_sub_epi8(src, pred); + _mm_storeu_si128((__m128i*)&out[i], res); + } + if (i != num_pixels) { + VP8LPredictorsSub_C[12](in + i, upper + i, num_pixels - i, out + i); + } +} + +// Predictors13: ClampedAddSubtractHalf +static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper, + int num_pixels, uint32_t* out) { + int i; + const __m128i zero = _mm_setzero_si128(); + for (i = 0; i + 2 <= num_pixels; i += 2) { + // we can only process two pixels at a time + const __m128i L = _mm_loadl_epi64((const __m128i*)&in[i - 1]); + const __m128i src = _mm_loadl_epi64((const __m128i*)&in[i]); + const __m128i T = _mm_loadl_epi64((const __m128i*)&upper[i]); + const __m128i TL = _mm_loadl_epi64((const __m128i*)&upper[i - 1]); + const __m128i L_lo = _mm_unpacklo_epi8(L, zero); + const __m128i T_lo = _mm_unpacklo_epi8(T, zero); + const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero); + const __m128i sum = _mm_add_epi16(T_lo, L_lo); + const __m128i avg = _mm_srli_epi16(sum, 1); + const __m128i A1 = _mm_sub_epi16(avg, TL_lo); + const __m128i bit_fix = _mm_cmpgt_epi16(TL_lo, avg); + const __m128i A2 = _mm_sub_epi16(A1, bit_fix); + const __m128i A3 = _mm_srai_epi16(A2, 1); + const __m128i A4 = _mm_add_epi16(avg, A3); + const __m128i pred = _mm_packus_epi16(A4, A4); + const __m128i res = _mm_sub_epi8(src, pred); + _mm_storel_epi64((__m128i*)&out[i], res); + } + if (i != num_pixels) { + VP8LPredictorsSub_C[13](in + i, upper + i, num_pixels - i, out + i); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE2; + VP8LTransformColor = TransformColor_SSE2; + VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE2; + VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE2; + VP8LAddVector = AddVector_SSE2; + VP8LAddVectorEq = AddVectorEq_SSE2; +#if !defined(DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC) + VP8LCombinedShannonEntropy = CombinedShannonEntropy_SSE2; +#endif + VP8LVectorMismatch = VectorMismatch_SSE2; + VP8LBundleColorMap = BundleColorMap_SSE2; + + VP8LPredictorsSub[0] = PredictorSub0_SSE2; + VP8LPredictorsSub[1] = PredictorSub1_SSE2; + VP8LPredictorsSub[2] = PredictorSub2_SSE2; + VP8LPredictorsSub[3] = PredictorSub3_SSE2; + VP8LPredictorsSub[4] = PredictorSub4_SSE2; + VP8LPredictorsSub[5] = PredictorSub5_SSE2; + VP8LPredictorsSub[6] = PredictorSub6_SSE2; + VP8LPredictorsSub[7] = PredictorSub7_SSE2; + VP8LPredictorsSub[8] = PredictorSub8_SSE2; + VP8LPredictorsSub[9] = PredictorSub9_SSE2; + VP8LPredictorsSub[10] = PredictorSub10_SSE2; + VP8LPredictorsSub[11] = PredictorSub11_SSE2; + VP8LPredictorsSub[12] = PredictorSub12_SSE2; + VP8LPredictorsSub[13] = PredictorSub13_SSE2; + VP8LPredictorsSub[14] = PredictorSub0_SSE2; // <- padding security sentinels + VP8LPredictorsSub[15] = PredictorSub0_SSE2; +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/media/libwebp/dsp/lossless_enc_sse41.c b/media/libwebp/dsp/lossless_enc_sse41.c new file mode 100644 index 0000000000..2c6bc5bb00 --- /dev/null +++ b/media/libwebp/dsp/lossless_enc_sse41.c @@ -0,0 +1,155 @@ +// Copyright 2015 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE4.1 variant of methods for lossless encoder +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_SSE41) +#include <assert.h> +#include <smmintrin.h> +#include "../dsp/lossless.h" + +// For sign-extended multiplying constants, pre-shifted by 5: +#define CST_5b(X) (((int16_t)((uint16_t)(X) << 8)) >> 5) + +//------------------------------------------------------------------------------ +// Subtract-Green Transform + +static void SubtractGreenFromBlueAndRed_SSE41(uint32_t* argb_data, + int num_pixels) { + int i; + const __m128i kCstShuffle = _mm_set_epi8(-1, 13, -1, 13, -1, 9, -1, 9, + -1, 5, -1, 5, -1, 1, -1, 1); + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); + const __m128i in_0g0g = _mm_shuffle_epi8(in, kCstShuffle); + const __m128i out = _mm_sub_epi8(in, in_0g0g); + _mm_storeu_si128((__m128i*)&argb_data[i], out); + } + // fallthrough and finish off with plain-C + if (i != num_pixels) { + VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i); + } +} + +//------------------------------------------------------------------------------ +// Color Transform + +#define MK_CST_16(HI, LO) \ + _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff))) + +static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_blue, int red_to_blue, + int histo[]) { + const __m128i mult = + MK_CST_16(CST_5b(red_to_blue) + 256,CST_5b(green_to_blue)); + const __m128i perm = + _mm_setr_epi8(-1, 1, -1, 2, -1, 5, -1, 6, -1, 9, -1, 10, -1, 13, -1, 14); + if (tile_width >= 4) { + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + const __m128i A1 = _mm_loadu_si128((const __m128i*)src); + const __m128i B1 = _mm_shuffle_epi8(A1, perm); + const __m128i C1 = _mm_mulhi_epi16(B1, mult); + const __m128i D1 = _mm_sub_epi16(A1, C1); + __m128i E = _mm_add_epi16(_mm_srli_epi32(D1, 16), D1); + int x; + for (x = 4; x + 4 <= tile_width; x += 4) { + const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x)); + __m128i B2, C2, D2; + ++histo[_mm_extract_epi8(E, 0)]; + B2 = _mm_shuffle_epi8(A2, perm); + ++histo[_mm_extract_epi8(E, 4)]; + C2 = _mm_mulhi_epi16(B2, mult); + ++histo[_mm_extract_epi8(E, 8)]; + D2 = _mm_sub_epi16(A2, C2); + ++histo[_mm_extract_epi8(E, 12)]; + E = _mm_add_epi16(_mm_srli_epi32(D2, 16), D2); + } + ++histo[_mm_extract_epi8(E, 0)]; + ++histo[_mm_extract_epi8(E, 4)]; + ++histo[_mm_extract_epi8(E, 8)]; + ++histo[_mm_extract_epi8(E, 12)]; + } + } + { + const int left_over = tile_width & 3; + if (left_over > 0) { + VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride, + left_over, tile_height, + green_to_blue, red_to_blue, histo); + } + } +} + +static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride, + int tile_width, int tile_height, + int green_to_red, int histo[]) { + + const __m128i mult = MK_CST_16(0, CST_5b(green_to_red)); + const __m128i mask_g = _mm_set1_epi32(0x0000ff00); + if (tile_width >= 4) { + int y; + for (y = 0; y < tile_height; ++y) { + const uint32_t* const src = argb + y * stride; + const __m128i A1 = _mm_loadu_si128((const __m128i*)src); + const __m128i B1 = _mm_and_si128(A1, mask_g); + const __m128i C1 = _mm_madd_epi16(B1, mult); + __m128i D = _mm_sub_epi16(A1, C1); + int x; + for (x = 4; x + 4 <= tile_width; x += 4) { + const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x)); + __m128i B2, C2; + ++histo[_mm_extract_epi8(D, 2)]; + B2 = _mm_and_si128(A2, mask_g); + ++histo[_mm_extract_epi8(D, 6)]; + C2 = _mm_madd_epi16(B2, mult); + ++histo[_mm_extract_epi8(D, 10)]; + ++histo[_mm_extract_epi8(D, 14)]; + D = _mm_sub_epi16(A2, C2); + } + ++histo[_mm_extract_epi8(D, 2)]; + ++histo[_mm_extract_epi8(D, 6)]; + ++histo[_mm_extract_epi8(D, 10)]; + ++histo[_mm_extract_epi8(D, 14)]; + } + } + { + const int left_over = tile_width & 3; + if (left_over > 0) { + VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride, + left_over, tile_height, green_to_red, + histo); + } + } +} + +#undef MK_CST_16 + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LEncDspInitSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE41(void) { + VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE41; + VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE41; + VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE41; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/media/libwebp/dsp/lossless_mips_dsp_r2.c b/media/libwebp/dsp/lossless_mips_dsp_r2.c new file mode 100644 index 0000000000..ec98834f84 --- /dev/null +++ b/media/libwebp/dsp/lossless_mips_dsp_r2.c @@ -0,0 +1,701 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// Image transforms and color space conversion methods for lossless decoder. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../dsp/lossless.h" +#include "../dsp/lossless_common.h" + +#define MAP_COLOR_FUNCS(FUNC_NAME, TYPE, GET_INDEX, GET_VALUE) \ +static void FUNC_NAME(const TYPE* src, \ + const uint32_t* const color_map, \ + TYPE* dst, int y_start, int y_end, \ + int width) { \ + int y; \ + for (y = y_start; y < y_end; ++y) { \ + int x; \ + for (x = 0; x < (width >> 2); ++x) { \ + int tmp1, tmp2, tmp3, tmp4; \ + __asm__ volatile ( \ + ".ifc " #TYPE ", uint8_t \n\t" \ + "lbu %[tmp1], 0(%[src]) \n\t" \ + "lbu %[tmp2], 1(%[src]) \n\t" \ + "lbu %[tmp3], 2(%[src]) \n\t" \ + "lbu %[tmp4], 3(%[src]) \n\t" \ + "addiu %[src], %[src], 4 \n\t" \ + ".endif \n\t" \ + ".ifc " #TYPE ", uint32_t \n\t" \ + "lw %[tmp1], 0(%[src]) \n\t" \ + "lw %[tmp2], 4(%[src]) \n\t" \ + "lw %[tmp3], 8(%[src]) \n\t" \ + "lw %[tmp4], 12(%[src]) \n\t" \ + "ext %[tmp1], %[tmp1], 8, 8 \n\t" \ + "ext %[tmp2], %[tmp2], 8, 8 \n\t" \ + "ext %[tmp3], %[tmp3], 8, 8 \n\t" \ + "ext %[tmp4], %[tmp4], 8, 8 \n\t" \ + "addiu %[src], %[src], 16 \n\t" \ + ".endif \n\t" \ + "sll %[tmp1], %[tmp1], 2 \n\t" \ + "sll %[tmp2], %[tmp2], 2 \n\t" \ + "sll %[tmp3], %[tmp3], 2 \n\t" \ + "sll %[tmp4], %[tmp4], 2 \n\t" \ + "lwx %[tmp1], %[tmp1](%[color_map]) \n\t" \ + "lwx %[tmp2], %[tmp2](%[color_map]) \n\t" \ + "lwx %[tmp3], %[tmp3](%[color_map]) \n\t" \ + "lwx %[tmp4], %[tmp4](%[color_map]) \n\t" \ + ".ifc " #TYPE ", uint8_t \n\t" \ + "ext %[tmp1], %[tmp1], 8, 8 \n\t" \ + "ext %[tmp2], %[tmp2], 8, 8 \n\t" \ + "ext %[tmp3], %[tmp3], 8, 8 \n\t" \ + "ext %[tmp4], %[tmp4], 8, 8 \n\t" \ + "sb %[tmp1], 0(%[dst]) \n\t" \ + "sb %[tmp2], 1(%[dst]) \n\t" \ + "sb %[tmp3], 2(%[dst]) \n\t" \ + "sb %[tmp4], 3(%[dst]) \n\t" \ + "addiu %[dst], %[dst], 4 \n\t" \ + ".endif \n\t" \ + ".ifc " #TYPE ", uint32_t \n\t" \ + "sw %[tmp1], 0(%[dst]) \n\t" \ + "sw %[tmp2], 4(%[dst]) \n\t" \ + "sw %[tmp3], 8(%[dst]) \n\t" \ + "sw %[tmp4], 12(%[dst]) \n\t" \ + "addiu %[dst], %[dst], 16 \n\t" \ + ".endif \n\t" \ + : [tmp1]"=&r"(tmp1), [tmp2]"=&r"(tmp2), [tmp3]"=&r"(tmp3), \ + [tmp4]"=&r"(tmp4), [src]"+&r"(src), [dst]"+r"(dst) \ + : [color_map]"r"(color_map) \ + : "memory" \ + ); \ + } \ + for (x = 0; x < (width & 3); ++x) { \ + *dst++ = GET_VALUE(color_map[GET_INDEX(*src++)]); \ + } \ + } \ +} + +MAP_COLOR_FUNCS(MapARGB_MIPSdspR2, uint32_t, VP8GetARGBIndex, VP8GetARGBValue) +MAP_COLOR_FUNCS(MapAlpha_MIPSdspR2, uint8_t, VP8GetAlphaIndex, VP8GetAlphaValue) + +#undef MAP_COLOR_FUNCS + +static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, + uint32_t c2) { + int temp0, temp1, temp2, temp3, temp4, temp5; + __asm__ volatile ( + "preceu.ph.qbr %[temp1], %[c0] \n\t" + "preceu.ph.qbl %[temp2], %[c0] \n\t" + "preceu.ph.qbr %[temp3], %[c1] \n\t" + "preceu.ph.qbl %[temp4], %[c1] \n\t" + "preceu.ph.qbr %[temp5], %[c2] \n\t" + "preceu.ph.qbl %[temp0], %[c2] \n\t" + "subq.ph %[temp3], %[temp3], %[temp5] \n\t" + "subq.ph %[temp4], %[temp4], %[temp0] \n\t" + "addq.ph %[temp1], %[temp1], %[temp3] \n\t" + "addq.ph %[temp2], %[temp2], %[temp4] \n\t" + "shll_s.ph %[temp1], %[temp1], 7 \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "precrqu_s.qb.ph %[temp2], %[temp2], %[temp1] \n\t" + : [temp0]"=r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5) + : [c0]"r"(c0), [c1]"r"(c1), [c2]"r"(c2) + : "memory" + ); + return temp2; +} + +static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, + uint32_t c2) { + int temp0, temp1, temp2, temp3, temp4, temp5; + __asm__ volatile ( + "adduh.qb %[temp5], %[c0], %[c1] \n\t" + "preceu.ph.qbr %[temp3], %[c2] \n\t" + "preceu.ph.qbr %[temp1], %[temp5] \n\t" + "preceu.ph.qbl %[temp2], %[temp5] \n\t" + "preceu.ph.qbl %[temp4], %[c2] \n\t" + "subq.ph %[temp3], %[temp1], %[temp3] \n\t" + "subq.ph %[temp4], %[temp2], %[temp4] \n\t" + "shrl.ph %[temp5], %[temp3], 15 \n\t" + "shrl.ph %[temp0], %[temp4], 15 \n\t" + "addq.ph %[temp3], %[temp3], %[temp5] \n\t" + "addq.ph %[temp4], %[temp0], %[temp4] \n\t" + "shra.ph %[temp3], %[temp3], 1 \n\t" + "shra.ph %[temp4], %[temp4], 1 \n\t" + "addq.ph %[temp1], %[temp1], %[temp3] \n\t" + "addq.ph %[temp2], %[temp2], %[temp4] \n\t" + "shll_s.ph %[temp1], %[temp1], 7 \n\t" + "shll_s.ph %[temp2], %[temp2], 7 \n\t" + "precrqu_s.qb.ph %[temp1], %[temp2], %[temp1] \n\t" + : [temp0]"=r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=r"(temp4), [temp5]"=&r"(temp5) + : [c0]"r"(c0), [c1]"r"(c1), [c2]"r"(c2) + : "memory" + ); + return temp1; +} + +static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { + int temp0, temp1, temp2, temp3, temp4, temp5; + __asm__ volatile ( + "cmpgdu.lt.qb %[temp1], %[c], %[b] \n\t" + "pick.qb %[temp1], %[b], %[c] \n\t" + "pick.qb %[temp2], %[c], %[b] \n\t" + "cmpgdu.lt.qb %[temp4], %[c], %[a] \n\t" + "pick.qb %[temp4], %[a], %[c] \n\t" + "pick.qb %[temp5], %[c], %[a] \n\t" + "subu.qb %[temp3], %[temp1], %[temp2] \n\t" + "subu.qb %[temp0], %[temp4], %[temp5] \n\t" + "raddu.w.qb %[temp3], %[temp3] \n\t" + "raddu.w.qb %[temp0], %[temp0] \n\t" + "subu %[temp3], %[temp3], %[temp0] \n\t" + "slti %[temp0], %[temp3], 0x1 \n\t" + "movz %[a], %[b], %[temp0] \n\t" + : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp0]"=&r"(temp0), + [a]"+&r"(a) + : [b]"r"(b), [c]"r"(c) + ); + return a; +} + +static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { + __asm__ volatile ( + "adduh.qb %[a0], %[a0], %[a1] \n\t" + : [a0]"+r"(a0) + : [a1]"r"(a1) + ); + return a0; +} + +static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { + return Average2(Average2(a0, a2), a1); +} + +static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1, + uint32_t a2, uint32_t a3) { + return Average2(Average2(a0, a1), Average2(a2, a3)); +} + +static uint32_t Predictor5_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + return Average3(*left, top[0], top[1]); +} + +static uint32_t Predictor6_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + return Average2(*left, top[-1]); +} + +static uint32_t Predictor7_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + return Average2(*left, top[0]); +} + +static uint32_t Predictor8_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + (void)left; + return Average2(top[-1], top[0]); +} + +static uint32_t Predictor9_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + (void)left; + return Average2(top[0], top[1]); +} + +static uint32_t Predictor10_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + return Average4(*left, top[-1], top[0], top[1]); +} + +static uint32_t Predictor11_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + return Select(top[0], *left, top[-1]); +} + +static uint32_t Predictor12_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + return ClampedAddSubtractFull(*left, top[0], top[-1]); +} + +static uint32_t Predictor13_MIPSdspR2(const uint32_t* const left, + const uint32_t* const top) { + return ClampedAddSubtractHalf(*left, top[0], top[-1]); +} + +// Add green to blue and red channels (i.e. perform the inverse transform of +// 'subtract green'). +static void AddGreenToBlueAndRed_MIPSdspR2(const uint32_t* src, int num_pixels, + uint32_t* dst) { + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "ext %[temp5], %[temp1], 8, 8 \n\t" + "ext %[temp6], %[temp2], 8, 8 \n\t" + "ext %[temp7], %[temp3], 8, 8 \n\t" + "addiu %[src], %[src], 16 \n\t" + "addiu %[dst], %[dst], 16 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "replv.ph %[temp5], %[temp5] \n\t" + "replv.ph %[temp6], %[temp6] \n\t" + "replv.ph %[temp7], %[temp7] \n\t" + "addu.qb %[temp0], %[temp0], %[temp4] \n\t" + "addu.qb %[temp1], %[temp1], %[temp5] \n\t" + "addu.qb %[temp2], %[temp2], %[temp6] \n\t" + "addu.qb %[temp3], %[temp3], %[temp7] \n\t" + "sw %[temp0], -16(%[dst]) \n\t" + "sw %[temp1], -12(%[dst]) \n\t" + "sw %[temp2], -8(%[dst]) \n\t" + "bne %[src], %[p_loop1_end], 0b \n\t" + " sw %[temp3], -4(%[dst]) \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "addiu %[src], %[src], 4 \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "ext %[temp4], %[temp0], 8, 8 \n\t" + "replv.ph %[temp4], %[temp4] \n\t" + "addu.qb %[temp0], %[temp0], %[temp4] \n\t" + "bne %[src], %[p_loop2_end], 1b \n\t" + " sw %[temp0], -4(%[dst]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [dst]"+&r"(dst), [src]"+&r"(src), [temp0]"=&r"(temp0), + [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6), + [temp7]"=&r"(temp7) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void TransformColorInverse_MIPSdspR2(const VP8LMultipliers* const m, + const uint32_t* src, int num_pixels, + uint32_t* dst) { + int temp0, temp1, temp2, temp3, temp4, temp5; + uint32_t argb, argb1, new_red; + const uint32_t G_to_R = m->green_to_red_; + const uint32_t G_to_B = m->green_to_blue_; + const uint32_t R_to_B = m->red_to_blue_; + const uint32_t* const p_loop_end = src + (num_pixels & ~1); + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop_end], 1f \n\t" + " nop \n\t" + "replv.ph %[temp0], %[G_to_R] \n\t" + "replv.ph %[temp1], %[G_to_B] \n\t" + "replv.ph %[temp2], %[R_to_B] \n\t" + "shll.ph %[temp0], %[temp0], 8 \n\t" + "shll.ph %[temp1], %[temp1], 8 \n\t" + "shll.ph %[temp2], %[temp2], 8 \n\t" + "shra.ph %[temp0], %[temp0], 8 \n\t" + "shra.ph %[temp1], %[temp1], 8 \n\t" + "shra.ph %[temp2], %[temp2], 8 \n\t" + "0: \n\t" + "lw %[argb], 0(%[src]) \n\t" + "lw %[argb1], 4(%[src]) \n\t" + "sw %[argb], 0(%[dst]) \n\t" + "sw %[argb1], 4(%[dst]) \n\t" + "addiu %[src], %[src], 8 \n\t" + "addiu %[dst], %[dst], 8 \n\t" + "precrq.qb.ph %[temp3], %[argb], %[argb1] \n\t" + "preceu.ph.qbra %[temp3], %[temp3] \n\t" + "shll.ph %[temp3], %[temp3], 8 \n\t" + "shra.ph %[temp3], %[temp3], 8 \n\t" + "mul.ph %[temp5], %[temp3], %[temp0] \n\t" + "mul.ph %[temp3], %[temp3], %[temp1] \n\t" + "precrq.ph.w %[new_red], %[argb], %[argb1] \n\t" + "ins %[argb1], %[argb], 16, 16 \n\t" + "shra.ph %[temp5], %[temp5], 5 \n\t" + "shra.ph %[temp3], %[temp3], 5 \n\t" + "addu.ph %[new_red], %[new_red], %[temp5] \n\t" + "addu.ph %[argb1], %[argb1], %[temp3] \n\t" + "preceu.ph.qbra %[temp5], %[new_red] \n\t" + "shll.ph %[temp4], %[temp5], 8 \n\t" + "shra.ph %[temp4], %[temp4], 8 \n\t" + "mul.ph %[temp4], %[temp4], %[temp2] \n\t" + "sb %[temp5], -2(%[dst]) \n\t" + "sra %[temp5], %[temp5], 16 \n\t" + "shra.ph %[temp4], %[temp4], 5 \n\t" + "addu.ph %[argb1], %[argb1], %[temp4] \n\t" + "preceu.ph.qbra %[temp3], %[argb1] \n\t" + "sb %[temp5], -6(%[dst]) \n\t" + "sb %[temp3], -4(%[dst]) \n\t" + "sra %[temp3], %[temp3], 16 \n\t" + "bne %[src], %[p_loop_end], 0b \n\t" + " sb %[temp3], -8(%[dst]) \n\t" + "1: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [new_red]"=&r"(new_red), [argb]"=&r"(argb), + [argb1]"=&r"(argb1), [dst]"+&r"(dst), [src]"+&r"(src) + : [G_to_R]"r"(G_to_R), [R_to_B]"r"(R_to_B), + [G_to_B]"r"(G_to_B), [p_loop_end]"r"(p_loop_end) + : "memory", "hi", "lo" + ); + + // Fall-back to C-version for left-overs. + if (num_pixels & 1) VP8LTransformColorInverse_C(m, src, 1, dst); +} + +static void ConvertBGRAToRGB_MIPSdspR2(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "ins %[temp3], %[temp2], 24, 8 \n\t" + "sll %[temp2], %[temp2], 8 \n\t" + "rotr %[temp3], %[temp3], 16 \n\t" + "ins %[temp2], %[temp1], 0, 16 \n\t" + "sll %[temp1], %[temp1], 8 \n\t" + "wsbh %[temp3], %[temp3] \n\t" + "balign %[temp0], %[temp1], 1 \n\t" + "wsbh %[temp2], %[temp2] \n\t" + "wsbh %[temp0], %[temp0] \n\t" + "usw %[temp3], 8(%[dst]) \n\t" + "rotr %[temp0], %[temp0], 16 \n\t" + "usw %[temp2], 4(%[dst]) \n\t" + "addiu %[src], %[src], 16 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 12 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "addiu %[src], %[src], 4 \n\t" + "wsbh %[temp1], %[temp0] \n\t" + "addiu %[dst], %[dst], 3 \n\t" + "ush %[temp1], -2(%[dst]) \n\t" + "sra %[temp0], %[temp0], 16 \n\t" + "bne %[src], %[p_loop2_end], 1b \n\t" + " sb %[temp0], -3(%[dst]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToRGBA_MIPSdspR2(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "wsbh %[temp0], %[temp0] \n\t" + "wsbh %[temp1], %[temp1] \n\t" + "wsbh %[temp2], %[temp2] \n\t" + "wsbh %[temp3], %[temp3] \n\t" + "addiu %[src], %[src], 16 \n\t" + "balign %[temp0], %[temp0], 1 \n\t" + "balign %[temp1], %[temp1], 1 \n\t" + "balign %[temp2], %[temp2], 1 \n\t" + "balign %[temp3], %[temp3], 1 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp1], 4(%[dst]) \n\t" + "usw %[temp2], 8(%[dst]) \n\t" + "usw %[temp3], 12(%[dst]) \n\t" + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 16 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "wsbh %[temp0], %[temp0] \n\t" + "addiu %[src], %[src], 4 \n\t" + "balign %[temp0], %[temp0], 1 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "bne %[src], %[p_loop2_end], 1b \n\t" + " addiu %[dst], %[dst], 4 \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToRGBA4444_MIPSdspR2(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4, temp5; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "ext %[temp4], %[temp0], 28, 4 \n\t" + "ext %[temp5], %[temp0], 12, 4 \n\t" + "ins %[temp0], %[temp4], 0, 4 \n\t" + "ext %[temp4], %[temp1], 28, 4 \n\t" + "ins %[temp0], %[temp5], 16, 4 \n\t" + "ext %[temp5], %[temp1], 12, 4 \n\t" + "ins %[temp1], %[temp4], 0, 4 \n\t" + "ext %[temp4], %[temp2], 28, 4 \n\t" + "ins %[temp1], %[temp5], 16, 4 \n\t" + "ext %[temp5], %[temp2], 12, 4 \n\t" + "ins %[temp2], %[temp4], 0, 4 \n\t" + "ext %[temp4], %[temp3], 28, 4 \n\t" + "ins %[temp2], %[temp5], 16, 4 \n\t" + "ext %[temp5], %[temp3], 12, 4 \n\t" + "ins %[temp3], %[temp4], 0, 4 \n\t" + "precr.qb.ph %[temp1], %[temp1], %[temp0] \n\t" + "ins %[temp3], %[temp5], 16, 4 \n\t" + "addiu %[src], %[src], 16 \n\t" + "precr.qb.ph %[temp3], %[temp3], %[temp2] \n\t" +#if (WEBP_SWAP_16BIT_CSP == 1) + "usw %[temp1], 0(%[dst]) \n\t" + "usw %[temp3], 4(%[dst]) \n\t" +#else + "wsbh %[temp1], %[temp1] \n\t" + "wsbh %[temp3], %[temp3] \n\t" + "usw %[temp1], 0(%[dst]) \n\t" + "usw %[temp3], 4(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 8 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "ext %[temp4], %[temp0], 28, 4 \n\t" + "ext %[temp5], %[temp0], 12, 4 \n\t" + "ins %[temp0], %[temp4], 0, 4 \n\t" + "ins %[temp0], %[temp5], 16, 4 \n\t" + "addiu %[src], %[src], 4 \n\t" + "precr.qb.ph %[temp0], %[temp0], %[temp0] \n\t" +#if (WEBP_SWAP_16BIT_CSP == 1) + "ush %[temp0], 0(%[dst]) \n\t" +#else + "wsbh %[temp0], %[temp0] \n\t" + "ush %[temp0], 0(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop2_end], 1b \n\t" + " addiu %[dst], %[dst], 2 \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToRGB565_MIPSdspR2(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3, temp4, temp5; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "ext %[temp4], %[temp0], 8, 16 \n\t" + "ext %[temp5], %[temp0], 5, 11 \n\t" + "ext %[temp0], %[temp0], 3, 5 \n\t" + "ins %[temp4], %[temp5], 0, 11 \n\t" + "ext %[temp5], %[temp1], 5, 11 \n\t" + "ins %[temp4], %[temp0], 0, 5 \n\t" + "ext %[temp0], %[temp1], 8, 16 \n\t" + "ext %[temp1], %[temp1], 3, 5 \n\t" + "ins %[temp0], %[temp5], 0, 11 \n\t" + "ext %[temp5], %[temp2], 5, 11 \n\t" + "ins %[temp0], %[temp1], 0, 5 \n\t" + "ext %[temp1], %[temp2], 8, 16 \n\t" + "ext %[temp2], %[temp2], 3, 5 \n\t" + "ins %[temp1], %[temp5], 0, 11 \n\t" + "ext %[temp5], %[temp3], 5, 11 \n\t" + "ins %[temp1], %[temp2], 0, 5 \n\t" + "ext %[temp2], %[temp3], 8, 16 \n\t" + "ext %[temp3], %[temp3], 3, 5 \n\t" + "ins %[temp2], %[temp5], 0, 11 \n\t" + "append %[temp0], %[temp4], 16 \n\t" + "ins %[temp2], %[temp3], 0, 5 \n\t" + "addiu %[src], %[src], 16 \n\t" + "append %[temp2], %[temp1], 16 \n\t" +#if (WEBP_SWAP_16BIT_CSP == 1) + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp2], 4(%[dst]) \n\t" +#else + "wsbh %[temp0], %[temp0] \n\t" + "wsbh %[temp2], %[temp2] \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp2], 4(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 8 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "ext %[temp4], %[temp0], 8, 16 \n\t" + "ext %[temp5], %[temp0], 5, 11 \n\t" + "ext %[temp0], %[temp0], 3, 5 \n\t" + "ins %[temp4], %[temp5], 0, 11 \n\t" + "addiu %[src], %[src], 4 \n\t" + "ins %[temp4], %[temp0], 0, 5 \n\t" +#if (WEBP_SWAP_16BIT_CSP == 1) + "ush %[temp4], 0(%[dst]) \n\t" +#else + "wsbh %[temp4], %[temp4] \n\t" + "ush %[temp4], 0(%[dst]) \n\t" +#endif + "bne %[src], %[p_loop2_end], 1b \n\t" + " addiu %[dst], %[dst], 2 \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), + [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +static void ConvertBGRAToBGR_MIPSdspR2(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int temp0, temp1, temp2, temp3; + const uint32_t* const p_loop1_end = src + (num_pixels & ~3); + const uint32_t* const p_loop2_end = src + num_pixels; + __asm__ volatile ( + ".set push \n\t" + ".set noreorder \n\t" + "beq %[src], %[p_loop1_end], 3f \n\t" + " nop \n\t" + "0: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "lw %[temp1], 4(%[src]) \n\t" + "lw %[temp2], 8(%[src]) \n\t" + "lw %[temp3], 12(%[src]) \n\t" + "ins %[temp0], %[temp1], 24, 8 \n\t" + "sra %[temp1], %[temp1], 8 \n\t" + "ins %[temp1], %[temp2], 16, 16 \n\t" + "sll %[temp2], %[temp2], 8 \n\t" + "balign %[temp3], %[temp2], 1 \n\t" + "addiu %[src], %[src], 16 \n\t" + "usw %[temp0], 0(%[dst]) \n\t" + "usw %[temp1], 4(%[dst]) \n\t" + "usw %[temp3], 8(%[dst]) \n\t" + "bne %[src], %[p_loop1_end], 0b \n\t" + " addiu %[dst], %[dst], 12 \n\t" + "3: \n\t" + "beq %[src], %[p_loop2_end], 2f \n\t" + " nop \n\t" + "1: \n\t" + "lw %[temp0], 0(%[src]) \n\t" + "addiu %[src], %[src], 4 \n\t" + "addiu %[dst], %[dst], 3 \n\t" + "ush %[temp0], -3(%[dst]) \n\t" + "sra %[temp0], %[temp0], 16 \n\t" + "bne %[src], %[p_loop2_end], 1b \n\t" + " sb %[temp0], -1(%[dst]) \n\t" + "2: \n\t" + ".set pop \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), + [temp3]"=&r"(temp3), [dst]"+&r"(dst), [src]"+&r"(src) + : [p_loop1_end]"r"(p_loop1_end), [p_loop2_end]"r"(p_loop2_end) + : "memory" + ); +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMIPSdspR2(void) { + VP8LMapColor32b = MapARGB_MIPSdspR2; + VP8LMapColor8b = MapAlpha_MIPSdspR2; + + VP8LPredictors[5] = Predictor5_MIPSdspR2; + VP8LPredictors[6] = Predictor6_MIPSdspR2; + VP8LPredictors[7] = Predictor7_MIPSdspR2; + VP8LPredictors[8] = Predictor8_MIPSdspR2; + VP8LPredictors[9] = Predictor9_MIPSdspR2; + VP8LPredictors[10] = Predictor10_MIPSdspR2; + VP8LPredictors[11] = Predictor11_MIPSdspR2; + VP8LPredictors[12] = Predictor12_MIPSdspR2; + VP8LPredictors[13] = Predictor13_MIPSdspR2; + + VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_MIPSdspR2; + VP8LTransformColorInverse = TransformColorInverse_MIPSdspR2; + + VP8LConvertBGRAToRGB = ConvertBGRAToRGB_MIPSdspR2; + VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_MIPSdspR2; + VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444_MIPSdspR2; + VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565_MIPSdspR2; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(VP8LDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/lossless_msa.c b/media/libwebp/dsp/lossless_msa.c new file mode 100644 index 0000000000..16256ab57f --- /dev/null +++ b/media/libwebp/dsp/lossless_msa.c @@ -0,0 +1,356 @@ +// Copyright 2016 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MSA variant of methods for lossless decoder +// +// Author: Prashant Patil (prashant.patil@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) + +#include "../dsp/lossless.h" +#include "../dsp/msa_macro.h" + +//------------------------------------------------------------------------------ +// Colorspace conversion functions + +#define CONVERT16_BGRA_XXX(psrc, pdst, m0, m1, m2) do { \ + v16u8 src0, src1, src2, src3, dst0, dst1, dst2; \ + LD_UB4(psrc, 16, src0, src1, src2, src3); \ + VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \ + dst2 = VSHF_UB(src2, src3, m2); \ + ST_UB2(dst0, dst1, pdst, 16); \ + ST_UB(dst2, pdst + 32); \ +} while (0) + +#define CONVERT12_BGRA_XXX(psrc, pdst, m0, m1, m2) do { \ + uint32_t pix_w; \ + v16u8 src0, src1, src2, dst0, dst1, dst2; \ + LD_UB3(psrc, 16, src0, src1, src2); \ + VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \ + dst2 = VSHF_UB(src2, src2, m2); \ + ST_UB2(dst0, dst1, pdst, 16); \ + pix_w = __msa_copy_s_w((v4i32)dst2, 0); \ + SW(pix_w, pdst + 32); \ +} while (0) + +#define CONVERT8_BGRA_XXX(psrc, pdst, m0, m1) do { \ + uint64_t pix_d; \ + v16u8 src0, src1, src2 = { 0 }, dst0, dst1; \ + LD_UB2(psrc, 16, src0, src1); \ + VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \ + ST_UB(dst0, pdst); \ + pix_d = __msa_copy_s_d((v2i64)dst1, 0); \ + SD(pix_d, pdst + 16); \ +} while (0) + +#define CONVERT4_BGRA_XXX(psrc, pdst, m) do { \ + const v16u8 src0 = LD_UB(psrc); \ + const v16u8 dst0 = VSHF_UB(src0, src0, m); \ + uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); \ + uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2); \ + SD(pix_d, pdst + 0); \ + SW(pix_w, pdst + 8); \ +} while (0) + +#define CONVERT1_BGRA_BGR(psrc, pdst) do { \ + const int32_t b = (psrc)[0]; \ + const int32_t g = (psrc)[1]; \ + const int32_t r = (psrc)[2]; \ + (pdst)[0] = b; \ + (pdst)[1] = g; \ + (pdst)[2] = r; \ +} while (0) + +#define CONVERT1_BGRA_RGB(psrc, pdst) do { \ + const int32_t b = (psrc)[0]; \ + const int32_t g = (psrc)[1]; \ + const int32_t r = (psrc)[2]; \ + (pdst)[0] = r; \ + (pdst)[1] = g; \ + (pdst)[2] = b; \ +} while (0) + +#define TRANSFORM_COLOR_INVERSE_8(src0, src1, dst0, dst1, \ + c0, c1, mask0, mask1) do { \ + v8i16 g0, g1, t0, t1, t2, t3; \ + v4i32 t4, t5; \ + VSHF_B2_SH(src0, src0, src1, src1, mask0, mask0, g0, g1); \ + DOTP_SB2_SH(g0, g1, c0, c0, t0, t1); \ + SRAI_H2_SH(t0, t1, 5); \ + t0 = __msa_addv_h(t0, (v8i16)src0); \ + t1 = __msa_addv_h(t1, (v8i16)src1); \ + t4 = __msa_srli_w((v4i32)t0, 16); \ + t5 = __msa_srli_w((v4i32)t1, 16); \ + DOTP_SB2_SH(t4, t5, c1, c1, t2, t3); \ + SRAI_H2_SH(t2, t3, 5); \ + ADD2(t0, t2, t1, t3, t0, t1); \ + VSHF_B2_UB(src0, t0, src1, t1, mask1, mask1, dst0, dst1); \ +} while (0) + +#define TRANSFORM_COLOR_INVERSE_4(src, dst, c0, c1, mask0, mask1) do { \ + const v16i8 g0 = VSHF_SB(src, src, mask0); \ + v8i16 t0 = __msa_dotp_s_h(c0, g0); \ + v8i16 t1; \ + v4i32 t2; \ + t0 = SRAI_H(t0, 5); \ + t0 = __msa_addv_h(t0, (v8i16)src); \ + t2 = __msa_srli_w((v4i32)t0, 16); \ + t1 = __msa_dotp_s_h(c1, (v16i8)t2); \ + t1 = SRAI_H(t1, 5); \ + t0 = t0 + t1; \ + dst = VSHF_UB(src, t0, mask1); \ +} while (0) + +static void ConvertBGRAToRGBA_MSA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int i; + const uint8_t* ptemp_src = (const uint8_t*)src; + uint8_t* ptemp_dst = (uint8_t*)dst; + v16u8 src0, dst0; + const v16u8 mask = { 2, 1, 0, 3, 6, 5, 4, 7, 10, 9, 8, 11, 14, 13, 12, 15 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1; + LD_UB2(ptemp_src, 16, src0, src1); + VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1); + ST_UB2(dst0, dst1, ptemp_dst, 16); + ptemp_src += 32; + ptemp_dst += 32; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(ptemp_src); + dst0 = VSHF_UB(src0, src0, mask); + ST_UB(dst0, ptemp_dst); + ptemp_src += 16; + ptemp_dst += 16; + num_pixels -= 4; + } + for (i = 0; i < num_pixels; i++) { + const uint8_t b = ptemp_src[2]; + const uint8_t g = ptemp_src[1]; + const uint8_t r = ptemp_src[0]; + const uint8_t a = ptemp_src[3]; + ptemp_dst[0] = b; + ptemp_dst[1] = g; + ptemp_dst[2] = r; + ptemp_dst[3] = a; + ptemp_src += 4; + ptemp_dst += 4; + } + } +} + +static void ConvertBGRAToBGR_MSA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint8_t* ptemp_src = (const uint8_t*)src; + uint8_t* ptemp_dst = (uint8_t*)dst; + const v16u8 mask0 = { 0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, + 16, 17, 18, 20 }; + const v16u8 mask1 = { 5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20, + 21, 22, 24, 25 }; + const v16u8 mask2 = { 10, 12, 13, 14, 16, 17, 18, 20, 21, 22, 24, 25, + 26, 28, 29, 30 }; + + while (num_pixels >= 16) { + CONVERT16_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 64; + ptemp_dst += 48; + num_pixels -= 16; + } + if (num_pixels > 0) { + if (num_pixels >= 12) { + CONVERT12_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 48; + ptemp_dst += 36; + num_pixels -= 12; + } else if (num_pixels >= 8) { + CONVERT8_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1); + ptemp_src += 32; + ptemp_dst += 24; + num_pixels -= 8; + } else if (num_pixels >= 4) { + CONVERT4_BGRA_XXX(ptemp_src, ptemp_dst, mask0); + ptemp_src += 16; + ptemp_dst += 12; + num_pixels -= 4; + } + if (num_pixels == 3) { + CONVERT1_BGRA_BGR(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_BGR(ptemp_src + 4, ptemp_dst + 3); + CONVERT1_BGRA_BGR(ptemp_src + 8, ptemp_dst + 6); + } else if (num_pixels == 2) { + CONVERT1_BGRA_BGR(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_BGR(ptemp_src + 4, ptemp_dst + 3); + } else if (num_pixels == 1) { + CONVERT1_BGRA_BGR(ptemp_src, ptemp_dst); + } + } +} + +static void ConvertBGRAToRGB_MSA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint8_t* ptemp_src = (const uint8_t*)src; + uint8_t* ptemp_dst = (uint8_t*)dst; + const v16u8 mask0 = { 2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, + 18, 17, 16, 22 }; + const v16u8 mask1 = { 5, 4, 10, 9, 8, 14, 13, 12, 18, 17, 16, 22, + 21, 20, 26, 25 }; + const v16u8 mask2 = { 8, 14, 13, 12, 18, 17, 16, 22, 21, 20, 26, 25, + 24, 30, 29, 28 }; + + while (num_pixels >= 16) { + CONVERT16_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 64; + ptemp_dst += 48; + num_pixels -= 16; + } + if (num_pixels) { + if (num_pixels >= 12) { + CONVERT12_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 48; + ptemp_dst += 36; + num_pixels -= 12; + } else if (num_pixels >= 8) { + CONVERT8_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1); + ptemp_src += 32; + ptemp_dst += 24; + num_pixels -= 8; + } else if (num_pixels >= 4) { + CONVERT4_BGRA_XXX(ptemp_src, ptemp_dst, mask0); + ptemp_src += 16; + ptemp_dst += 12; + num_pixels -= 4; + } + if (num_pixels == 3) { + CONVERT1_BGRA_RGB(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_RGB(ptemp_src + 4, ptemp_dst + 3); + CONVERT1_BGRA_RGB(ptemp_src + 8, ptemp_dst + 6); + } else if (num_pixels == 2) { + CONVERT1_BGRA_RGB(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_RGB(ptemp_src + 4, ptemp_dst + 3); + } else if (num_pixels == 1) { + CONVERT1_BGRA_RGB(ptemp_src, ptemp_dst); + } + } +} + +static void AddGreenToBlueAndRed_MSA(const uint32_t* const src, int num_pixels, + uint32_t* dst) { + int i; + const uint8_t* in = (const uint8_t*)src; + uint8_t* out = (uint8_t*)dst; + v16u8 src0, dst0, tmp0; + const v16u8 mask = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, + 13, 255, 13, 255 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1, tmp1; + LD_UB2(in, 16, src0, src1); + VSHF_B2_UB(src0, src1, src1, src0, mask, mask, tmp0, tmp1); + ADD2(src0, tmp0, src1, tmp1, dst0, dst1); + ST_UB2(dst0, dst1, out, 16); + in += 32; + out += 32; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(in); + tmp0 = VSHF_UB(src0, src0, mask); + dst0 = src0 + tmp0; + ST_UB(dst0, out); + in += 16; + out += 16; + num_pixels -= 4; + } + for (i = 0; i < num_pixels; i++) { + const uint8_t b = in[0]; + const uint8_t g = in[1]; + const uint8_t r = in[2]; + out[0] = (b + g) & 0xff; + out[1] = g; + out[2] = (r + g) & 0xff; + out[4] = in[4]; + out += 4; + } + } +} + +static void TransformColorInverse_MSA(const VP8LMultipliers* const m, + const uint32_t* src, int num_pixels, + uint32_t* dst) { + v16u8 src0, dst0; + const v16i8 g2br = (v16i8)__msa_fill_w(m->green_to_blue_ | + (m->green_to_red_ << 16)); + const v16i8 r2b = (v16i8)__msa_fill_w(m->red_to_blue_); + const v16u8 mask0 = { 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, + 13, 255, 13, 255 }; + const v16u8 mask1 = { 16, 1, 18, 3, 20, 5, 22, 7, 24, 9, 26, 11, + 28, 13, 30, 15 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1; + LD_UB2(src, 4, src0, src1); + TRANSFORM_COLOR_INVERSE_8(src0, src1, dst0, dst1, g2br, r2b, mask0, mask1); + ST_UB2(dst0, dst1, dst, 4); + src += 8; + dst += 8; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(src); + TRANSFORM_COLOR_INVERSE_4(src0, dst0, g2br, r2b, mask0, mask1); + ST_UB(dst0, dst); + src += 4; + dst += 4; + num_pixels -= 4; + } + if (num_pixels > 0) { + src0 = LD_UB(src); + TRANSFORM_COLOR_INVERSE_4(src0, dst0, g2br, r2b, mask0, mask1); + if (num_pixels == 3) { + const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); + const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2); + SD(pix_d, dst + 0); + SW(pix_w, dst + 2); + } else if (num_pixels == 2) { + const uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); + SD(pix_d, dst); + } else { + const uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 0); + SW(pix_w, dst); + } + } + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMSA(void) { + VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_MSA; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR_MSA; + VP8LConvertBGRAToRGB = ConvertBGRAToRGB_MSA; + + VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_MSA; + VP8LTransformColorInverse = TransformColorInverse_MSA; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(VP8LDspInitMSA) + +#endif // WEBP_USE_MSA diff --git a/media/libwebp/dsp/lossless_neon.c b/media/libwebp/dsp/lossless_neon.c index a7bf47f3c4..2122e46f7a 100644 --- a/media/libwebp/dsp/lossless_neon.c +++ b/media/libwebp/dsp/lossless_neon.c @@ -188,17 +188,21 @@ static WEBP_INLINE uint32_t Average3_NEON(uint32_t a0, uint32_t a1, return avg; } -static uint32_t Predictor5_NEON(uint32_t left, const uint32_t* const top) { - return Average3_NEON(left, top[0], top[1]); +static uint32_t Predictor5_NEON(const uint32_t* const left, + const uint32_t* const top) { + return Average3_NEON(*left, top[0], top[1]); } -static uint32_t Predictor6_NEON(uint32_t left, const uint32_t* const top) { - return Average2_NEON(left, top[-1]); +static uint32_t Predictor6_NEON(const uint32_t* const left, + const uint32_t* const top) { + return Average2_NEON(*left, top[-1]); } -static uint32_t Predictor7_NEON(uint32_t left, const uint32_t* const top) { - return Average2_NEON(left, top[0]); +static uint32_t Predictor7_NEON(const uint32_t* const left, + const uint32_t* const top) { + return Average2_NEON(*left, top[0]); } -static uint32_t Predictor13_NEON(uint32_t left, const uint32_t* const top) { - return ClampedAddSubtractHalf_NEON(left, top[0], top[-1]); +static uint32_t Predictor13_NEON(const uint32_t* const left, + const uint32_t* const top) { + return ClampedAddSubtractHalf_NEON(*left, top[0], top[-1]); } // Batch versions of those functions. diff --git a/media/libwebp/dsp/lossless_sse2.c b/media/libwebp/dsp/lossless_sse2.c index c40fcfb769..03796493de 100644 --- a/media/libwebp/dsp/lossless_sse2.c +++ b/media/libwebp/dsp/lossless_sse2.c @@ -18,7 +18,6 @@ #include "../dsp/common_sse2.h" #include "../dsp/lossless.h" #include "../dsp/lossless_common.h" -#include <assert.h> #include <emmintrin.h> //------------------------------------------------------------------------------ @@ -139,42 +138,51 @@ static WEBP_INLINE uint32_t Average4_SSE2(uint32_t a0, uint32_t a1, return output; } -static uint32_t Predictor5_SSE2(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average3_SSE2(left, top[0], top[1]); +static uint32_t Predictor5_SSE2(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average3_SSE2(*left, top[0], top[1]); return pred; } -static uint32_t Predictor6_SSE2(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average2_SSE2(left, top[-1]); +static uint32_t Predictor6_SSE2(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average2_SSE2(*left, top[-1]); return pred; } -static uint32_t Predictor7_SSE2(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average2_SSE2(left, top[0]); +static uint32_t Predictor7_SSE2(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average2_SSE2(*left, top[0]); return pred; } -static uint32_t Predictor8_SSE2(uint32_t left, const uint32_t* const top) { +static uint32_t Predictor8_SSE2(const uint32_t* const left, + const uint32_t* const top) { const uint32_t pred = Average2_SSE2(top[-1], top[0]); (void)left; return pred; } -static uint32_t Predictor9_SSE2(uint32_t left, const uint32_t* const top) { +static uint32_t Predictor9_SSE2(const uint32_t* const left, + const uint32_t* const top) { const uint32_t pred = Average2_SSE2(top[0], top[1]); (void)left; return pred; } -static uint32_t Predictor10_SSE2(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Average4_SSE2(left, top[-1], top[0], top[1]); +static uint32_t Predictor10_SSE2(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Average4_SSE2(*left, top[-1], top[0], top[1]); return pred; } -static uint32_t Predictor11_SSE2(uint32_t left, const uint32_t* const top) { - const uint32_t pred = Select_SSE2(top[0], left, top[-1]); +static uint32_t Predictor11_SSE2(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = Select_SSE2(top[0], *left, top[-1]); return pred; } -static uint32_t Predictor12_SSE2(uint32_t left, const uint32_t* const top) { - const uint32_t pred = ClampedAddSubtractFull_SSE2(left, top[0], top[-1]); +static uint32_t Predictor12_SSE2(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractFull_SSE2(*left, top[0], top[-1]); return pred; } -static uint32_t Predictor13_SSE2(uint32_t left, const uint32_t* const top) { - const uint32_t pred = ClampedAddSubtractHalf_SSE2(left, top[0], top[-1]); +static uint32_t Predictor13_SSE2(const uint32_t* const left, + const uint32_t* const top) { + const uint32_t pred = ClampedAddSubtractHalf_SSE2(*left, top[0], top[-1]); return pred; } @@ -191,8 +199,9 @@ static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper, _mm_storeu_si128((__m128i*)&out[i], res); } if (i != num_pixels) { - VP8LPredictorsAdd_C[0](in + i, upper + i, num_pixels - i, out + i); + VP8LPredictorsAdd_C[0](in + i, NULL, num_pixels - i, out + i); } + (void)upper; } // Predictor1: left. diff --git a/media/libwebp/dsp/lossless_sse41.c b/media/libwebp/dsp/lossless_sse41.c new file mode 100644 index 0000000000..3308ac31ee --- /dev/null +++ b/media/libwebp/dsp/lossless_sse41.c @@ -0,0 +1,132 @@ +// Copyright 2021 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE41 variant of methods for lossless decoder + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_SSE41) + +#include "../dsp/common_sse41.h" +#include "../dsp/lossless.h" +#include "../dsp/lossless_common.h" + +//------------------------------------------------------------------------------ +// Color-space conversion functions + +static void TransformColorInverse_SSE41(const VP8LMultipliers* const m, + const uint32_t* const src, + int num_pixels, uint32_t* dst) { +// sign-extended multiplying constants, pre-shifted by 5. +#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend + const __m128i mults_rb = _mm_set1_epi32((uint32_t)CST(green_to_red_) << 16 | + (CST(green_to_blue_) & 0xffff)); + const __m128i mults_b2 = _mm_set1_epi32(CST(red_to_blue_)); +#undef CST + const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); + const __m128i perm1 = _mm_setr_epi8(-1, 1, -1, 1, -1, 5, -1, 5, + -1, 9, -1, 9, -1, 13, -1, 13); + const __m128i perm2 = _mm_setr_epi8(-1, 2, -1, -1, -1, 6, -1, -1, + -1, 10, -1, -1, -1, 14, -1, -1); + int i; + for (i = 0; i + 4 <= num_pixels; i += 4) { + const __m128i A = _mm_loadu_si128((const __m128i*)(src + i)); + const __m128i B = _mm_shuffle_epi8(A, perm1); // argb -> g0g0 + const __m128i C = _mm_mulhi_epi16(B, mults_rb); + const __m128i D = _mm_add_epi8(A, C); + const __m128i E = _mm_shuffle_epi8(D, perm2); + const __m128i F = _mm_mulhi_epi16(E, mults_b2); + const __m128i G = _mm_add_epi8(D, F); + const __m128i out = _mm_blendv_epi8(G, A, mask_ag); + _mm_storeu_si128((__m128i*)&dst[i], out); + } + // Fall-back to C-version for left-overs. + if (i != num_pixels) { + VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i); + } +} + +//------------------------------------------------------------------------------ + +#define ARGB_TO_RGB_SSE41 do { \ + while (num_pixels >= 16) { \ + const __m128i in0 = _mm_loadu_si128(in + 0); \ + const __m128i in1 = _mm_loadu_si128(in + 1); \ + const __m128i in2 = _mm_loadu_si128(in + 2); \ + const __m128i in3 = _mm_loadu_si128(in + 3); \ + const __m128i a0 = _mm_shuffle_epi8(in0, perm0); \ + const __m128i a1 = _mm_shuffle_epi8(in1, perm1); \ + const __m128i a2 = _mm_shuffle_epi8(in2, perm2); \ + const __m128i a3 = _mm_shuffle_epi8(in3, perm3); \ + const __m128i b0 = _mm_blend_epi16(a0, a1, 0xc0); \ + const __m128i b1 = _mm_blend_epi16(a1, a2, 0xf0); \ + const __m128i b2 = _mm_blend_epi16(a2, a3, 0xfc); \ + _mm_storeu_si128(out + 0, b0); \ + _mm_storeu_si128(out + 1, b1); \ + _mm_storeu_si128(out + 2, b2); \ + in += 4; \ + out += 3; \ + num_pixels -= 16; \ + } \ +} while (0) + +static void ConvertBGRAToRGB_SSE41(const uint32_t* src, int num_pixels, + uint8_t* dst) { + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + const __m128i perm0 = _mm_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, + 8, 14, 13, 12, -1, -1, -1, -1); + const __m128i perm1 = _mm_shuffle_epi32(perm0, 0x39); + const __m128i perm2 = _mm_shuffle_epi32(perm0, 0x4e); + const __m128i perm3 = _mm_shuffle_epi32(perm0, 0x93); + + ARGB_TO_RGB_SSE41; + + // left-overs + if (num_pixels > 0) { + VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out); + } +} + +static void ConvertBGRAToBGR_SSE41(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const __m128i* in = (const __m128i*)src; + __m128i* out = (__m128i*)dst; + const __m128i perm0 = _mm_setr_epi8(0, 1, 2, 4, 5, 6, 8, 9, 10, + 12, 13, 14, -1, -1, -1, -1); + const __m128i perm1 = _mm_shuffle_epi32(perm0, 0x39); + const __m128i perm2 = _mm_shuffle_epi32(perm0, 0x4e); + const __m128i perm3 = _mm_shuffle_epi32(perm0, 0x93); + + ARGB_TO_RGB_SSE41; + + // left-overs + if (num_pixels > 0) { + VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, (uint8_t*)out); + } +} + +#undef ARGB_TO_RGB_SSE41 + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitSSE41(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE41(void) { + VP8LTransformColorInverse = TransformColorInverse_SSE41; + VP8LConvertBGRAToRGB = ConvertBGRAToRGB_SSE41; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR_SSE41; +} + +#else // !WEBP_USE_SSE41 + +WEBP_DSP_INIT_STUB(VP8LDspInitSSE41) + +#endif // WEBP_USE_SSE41 diff --git a/media/libwebp/dsp/msa_macro.h b/media/libwebp/dsp/msa_macro.h index de026a1d9e..717e3b7b9f 100644 --- a/media/libwebp/dsp/msa_macro.h +++ b/media/libwebp/dsp/msa_macro.h @@ -14,6 +14,10 @@ #ifndef WEBP_DSP_MSA_MACRO_H_ #define WEBP_DSP_MSA_MACRO_H_ +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) + #include <stdint.h> #include <msa.h> @@ -1389,4 +1393,5 @@ static WEBP_INLINE uint32_t func_hadd_uh_u32(v8u16 in) { } while (0) #define AVER_UB2_UB(...) AVER_UB2(v16u8, __VA_ARGS__) +#endif // WEBP_USE_MSA #endif // WEBP_DSP_MSA_MACRO_H_ diff --git a/media/libwebp/dsp/neon.h b/media/libwebp/dsp/neon.h index 63c27a2901..9a0f630de0 100644 --- a/media/libwebp/dsp/neon.h +++ b/media/libwebp/dsp/neon.h @@ -12,10 +12,12 @@ #ifndef WEBP_DSP_NEON_H_ #define WEBP_DSP_NEON_H_ -#include <arm_neon.h> - #include "../dsp/dsp.h" +#if defined(WEBP_USE_NEON) + +#include <arm_neon.h> + // Right now, some intrinsics functions seem slower, so we disable them // everywhere except newer clang/gcc or aarch64 where the inline assembly is // incompatible. @@ -98,4 +100,5 @@ static WEBP_INLINE int32x4x4_t Transpose4x4_NEON(const int32x4x4_t rows) { } while (0) #endif +#endif // WEBP_USE_NEON #endif // WEBP_DSP_NEON_H_ diff --git a/media/libwebp/dsp/quant.h b/media/libwebp/dsp/quant.h index b82e728a53..14d8613431 100644 --- a/media/libwebp/dsp/quant.h +++ b/media/libwebp/dsp/quant.h @@ -10,6 +10,8 @@ #ifndef WEBP_DSP_QUANT_H_ #define WEBP_DSP_QUANT_H_ +#include <string.h> + #include "../dsp/dsp.h" #include "../webp/types.h" @@ -67,4 +69,17 @@ static WEBP_INLINE int IsFlat(const int16_t* levels, int num_blocks, #endif // defined(WEBP_USE_NEON) && !defined(WEBP_ANDROID_NEON) && // !defined(WEBP_HAVE_NEON_RTCD) +static WEBP_INLINE int IsFlatSource16(const uint8_t* src) { + const uint32_t v = src[0] * 0x01010101u; + int i; + for (i = 0; i < 16; ++i) { + if (memcmp(src + 0, &v, 4) || memcmp(src + 4, &v, 4) || + memcmp(src + 8, &v, 4) || memcmp(src + 12, &v, 4)) { + return 0; + } + src += BPS; + } + return 1; +} + #endif // WEBP_DSP_QUANT_H_ diff --git a/media/libwebp/dsp/rescaler.c b/media/libwebp/dsp/rescaler.c index 6bf387f8e0..4bbd281b1c 100644 --- a/media/libwebp/dsp/rescaler.c +++ b/media/libwebp/dsp/rescaler.c @@ -38,8 +38,9 @@ void WebPRescalerImportRowExpand_C(WebPRescaler* const wrk, int x_out = channel; // simple bilinear interpolation int accum = wrk->x_add; - int left = src[x_in]; - int right = (wrk->src_width > 1) ? src[x_in + x_stride] : left; + rescaler_t left = (rescaler_t)src[x_in]; + rescaler_t right = + (wrk->src_width > 1) ? (rescaler_t)src[x_in + x_stride] : left; x_in += x_stride; while (1) { wrk->frow[x_out] = right * wrk->x_add + (left - right) * accum; @@ -50,7 +51,7 @@ void WebPRescalerImportRowExpand_C(WebPRescaler* const wrk, left = right; x_in += x_stride; assert(x_in < wrk->src_width * x_stride); - right = src[x_in]; + right = (rescaler_t)src[x_in]; accum += wrk->x_add; } } @@ -109,8 +110,7 @@ void WebPRescalerExportRowExpand_C(WebPRescaler* const wrk) { for (x_out = 0; x_out < x_out_max; ++x_out) { const uint32_t J = frow[x_out]; const int v = (int)MULT_FIX(J, wrk->fy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; } } else { const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); @@ -120,8 +120,7 @@ void WebPRescalerExportRowExpand_C(WebPRescaler* const wrk) { + (uint64_t)B * irow[x_out]; const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); const int v = (int)MULT_FIX(J, wrk->fy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; } } } @@ -138,17 +137,15 @@ void WebPRescalerExportRowShrink_C(WebPRescaler* const wrk) { assert(!wrk->y_expand); if (yscale) { for (x_out = 0; x_out < x_out_max; ++x_out) { - const uint32_t frac = (uint32_t)MULT_FIX(frow[x_out], yscale); - const int v = (int)MULT_FIX_FLOOR(irow[x_out] - frac, wrk->fxy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + const uint32_t frac = (uint32_t)MULT_FIX_FLOOR(frow[x_out], yscale); + const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale); + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; irow[x_out] = frac; // new fractional start } } else { for (x_out = 0; x_out < x_out_max; ++x_out) { const int v = (int)MULT_FIX(irow[x_out], wrk->fxy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; irow[x_out] = 0; } } @@ -217,7 +214,7 @@ WEBP_DSP_INIT_FUNC(WebPRescalerDspInit) { WebPRescalerImportRowShrink = WebPRescalerImportRowShrink_C; if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPRescalerDspInitSSE2(); } @@ -239,7 +236,7 @@ WEBP_DSP_INIT_FUNC(WebPRescalerDspInit) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPRescalerDspInitNEON(); diff --git a/media/libwebp/dsp/rescaler_mips32.c b/media/libwebp/dsp/rescaler_mips32.c new file mode 100644 index 0000000000..44fad3fbe7 --- /dev/null +++ b/media/libwebp/dsp/rescaler_mips32.c @@ -0,0 +1,295 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of rescaling functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS32) && !defined(WEBP_REDUCE_SIZE) + +#include <assert.h> +#include "../utils/rescaler_utils.h" + +//------------------------------------------------------------------------------ +// Row import + +static void ImportRowShrink_MIPS32(WebPRescaler* const wrk, + const uint8_t* src) { + const int x_stride = wrk->num_channels; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const int fx_scale = wrk->fx_scale; + const int x_add = wrk->x_add; + const int x_sub = wrk->x_sub; + const int x_stride1 = x_stride << 2; + int channel; + assert(!wrk->x_expand); + assert(!WebPRescalerInputDone(wrk)); + + for (channel = 0; channel < x_stride; ++channel) { + const uint8_t* src1 = src + channel; + rescaler_t* frow = wrk->frow + channel; + int temp1, temp2, temp3; + int base, frac, sum; + int accum, accum1; + int loop_c = x_out_max - channel; + + __asm__ volatile ( + "li %[temp1], 0x8000 \n\t" + "li %[temp2], 0x10000 \n\t" + "li %[sum], 0 \n\t" + "li %[accum], 0 \n\t" + "1: \n\t" + "addu %[accum], %[accum], %[x_add] \n\t" + "li %[base], 0 \n\t" + "blez %[accum], 3f \n\t" + "2: \n\t" + "lbu %[base], 0(%[src1]) \n\t" + "subu %[accum], %[accum], %[x_sub] \n\t" + "addu %[src1], %[src1], %[x_stride] \n\t" + "addu %[sum], %[sum], %[base] \n\t" + "bgtz %[accum], 2b \n\t" + "3: \n\t" + "negu %[accum1], %[accum] \n\t" + "mul %[frac], %[base], %[accum1] \n\t" + "mul %[temp3], %[sum], %[x_sub] \n\t" + "subu %[loop_c], %[loop_c], %[x_stride] \n\t" + "mult %[temp1], %[temp2] \n\t" + "maddu %[frac], %[fx_scale] \n\t" + "mfhi %[sum] \n\t" + "subu %[temp3], %[temp3], %[frac] \n\t" + "sw %[temp3], 0(%[frow]) \n\t" + "addu %[frow], %[frow], %[x_stride1] \n\t" + "bgtz %[loop_c], 1b \n\t" + : [accum]"=&r"(accum), [src1]"+r"(src1), [temp3]"=&r"(temp3), + [sum]"=&r"(sum), [base]"=&r"(base), [frac]"=&r"(frac), + [frow]"+r"(frow), [accum1]"=&r"(accum1), + [temp2]"=&r"(temp2), [temp1]"=&r"(temp1) + : [x_stride]"r"(x_stride), [fx_scale]"r"(fx_scale), + [x_sub]"r"(x_sub), [x_add]"r"(x_add), + [loop_c]"r"(loop_c), [x_stride1]"r"(x_stride1) + : "memory", "hi", "lo" + ); + assert(accum == 0); + } +} + +static void ImportRowExpand_MIPS32(WebPRescaler* const wrk, + const uint8_t* src) { + const int x_stride = wrk->num_channels; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const int x_add = wrk->x_add; + const int x_sub = wrk->x_sub; + const int src_width = wrk->src_width; + const int x_stride1 = x_stride << 2; + int channel; + assert(wrk->x_expand); + assert(!WebPRescalerInputDone(wrk)); + + for (channel = 0; channel < x_stride; ++channel) { + const uint8_t* src1 = src + channel; + rescaler_t* frow = wrk->frow + channel; + int temp1, temp2, temp3, temp4; + int frac; + int accum; + int x_out = channel; + + __asm__ volatile ( + "addiu %[temp3], %[src_width], -1 \n\t" + "lbu %[temp2], 0(%[src1]) \n\t" + "addu %[src1], %[src1], %[x_stride] \n\t" + "bgtz %[temp3], 0f \n\t" + "addiu %[temp1], %[temp2], 0 \n\t" + "b 3f \n\t" + "0: \n\t" + "lbu %[temp1], 0(%[src1]) \n\t" + "3: \n\t" + "addiu %[accum], %[x_add], 0 \n\t" + "1: \n\t" + "subu %[temp3], %[temp2], %[temp1] \n\t" + "mul %[temp3], %[temp3], %[accum] \n\t" + "mul %[temp4], %[temp1], %[x_add] \n\t" + "addu %[temp3], %[temp4], %[temp3] \n\t" + "sw %[temp3], 0(%[frow]) \n\t" + "addu %[frow], %[frow], %[x_stride1] \n\t" + "addu %[x_out], %[x_out], %[x_stride] \n\t" + "subu %[temp3], %[x_out], %[x_out_max] \n\t" + "bgez %[temp3], 2f \n\t" + "subu %[accum], %[accum], %[x_sub] \n\t" + "bgez %[accum], 4f \n\t" + "addiu %[temp2], %[temp1], 0 \n\t" + "addu %[src1], %[src1], %[x_stride] \n\t" + "lbu %[temp1], 0(%[src1]) \n\t" + "addu %[accum], %[accum], %[x_add] \n\t" + "4: \n\t" + "b 1b \n\t" + "2: \n\t" + : [src1]"+r"(src1), [accum]"=&r"(accum), [temp1]"=&r"(temp1), + [temp2]"=&r"(temp2), [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), + [x_out]"+r"(x_out), [frac]"=&r"(frac), [frow]"+r"(frow) + : [x_stride]"r"(x_stride), [x_add]"r"(x_add), [x_sub]"r"(x_sub), + [x_stride1]"r"(x_stride1), [src_width]"r"(src_width), + [x_out_max]"r"(x_out_max) + : "memory", "hi", "lo" + ); + assert(wrk->x_sub == 0 /* <- special case for src_width=1 */ || accum == 0); + } +} + +//------------------------------------------------------------------------------ +// Row export + +static void ExportRowExpand_MIPS32(WebPRescaler* const wrk) { + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + int temp0, temp1, temp3, temp4, temp5, loop_end; + const int temp2 = (int)wrk->fy_scale; + const int temp6 = x_out_max << 2; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "addiu %[frow], %[frow], 4 \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp0], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [temp6]"r"(temp6) + : "memory", "hi", "lo" + ); + } else { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 0(%[irow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[A], %[temp0] \n\t" + "maddu %[B], %[temp1] \n\t" + "addiu %[frow], %[frow], 4 \n\t" + "addiu %[irow], %[irow], 4 \n\t" + "mfhi %[temp5] \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp5], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [temp6]"r"(temp6), [A]"r"(A), [B]"r"(B) + : "memory", "hi", "lo" + ); + } +} + +#if 0 // disabled for now. TODO(skal): make match the C-code +static void ExportRowShrink_MIPS32(WebPRescaler* const wrk) { + const int x_out_max = wrk->dst_width * wrk->num_channels; + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const rescaler_t* frow = wrk->frow; + const int yscale = wrk->fy_scale * (-wrk->y_accum); + int temp0, temp1, temp3, temp4, temp5, loop_end; + const int temp2 = (int)wrk->fxy_scale; + const int temp6 = x_out_max << 2; + + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + assert(wrk->fxy_scale != 0); + if (yscale) { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "mult %[temp3], %[temp4] \n\t" + "addiu %[frow], %[frow], 4 \n\t" + "maddu %[temp0], %[yscale] \n\t" + "mfhi %[temp1] \n\t" + "lw %[temp0], 0(%[irow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "addiu %[irow], %[irow], 4 \n\t" + "subu %[temp0], %[temp0], %[temp1] \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp0], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sw %[temp1], -4(%[irow]) \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [yscale]"r"(yscale), [temp6]"r"(temp6) + : "memory", "hi", "lo" + ); + } else { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[irow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[irow]) \n\t" + "addiu %[dst], %[dst], 1 \n\t" + "addiu %[irow], %[irow], 4 \n\t" + "mult %[temp3], %[temp4] \n\t" + "maddu %[temp0], %[temp2] \n\t" + "mfhi %[temp5] \n\t" + "sw $zero, -4(%[irow]) \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[irow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [irow]"+r"(irow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end) + : [temp2]"r"(temp2), [temp6]"r"(temp6) + : "memory", "hi", "lo" + ); + } +} +#endif // 0 + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPRescalerDspInitMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitMIPS32(void) { + WebPRescalerImportRowExpand = ImportRowExpand_MIPS32; + WebPRescalerImportRowShrink = ImportRowShrink_MIPS32; + WebPRescalerExportRowExpand = ExportRowExpand_MIPS32; +// WebPRescalerExportRowShrink = ExportRowShrink_MIPS32; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/media/libwebp/dsp/rescaler_mips_dsp_r2.c b/media/libwebp/dsp/rescaler_mips_dsp_r2.c new file mode 100644 index 0000000000..d6f2996578 --- /dev/null +++ b/media/libwebp/dsp/rescaler_mips_dsp_r2.c @@ -0,0 +1,314 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of rescaling functions +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) && !defined(WEBP_REDUCE_SIZE) + +#include <assert.h> +#include "../utils/rescaler_utils.h" + +#define ROUNDER (WEBP_RESCALER_ONE >> 1) +#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX) +#define MULT_FIX_FLOOR(x, y) (((uint64_t)(x) * (y)) >> WEBP_RESCALER_RFIX) + +//------------------------------------------------------------------------------ +// Row export + +#if 0 // disabled for now. TODO(skal): make match the C-code +static void ExportRowShrink_MIPSdspR2(WebPRescaler* const wrk) { + int i; + const int x_out_max = wrk->dst_width * wrk->num_channels; + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const rescaler_t* frow = wrk->frow; + const int yscale = wrk->fy_scale * (-wrk->y_accum); + int temp0, temp1, temp2, temp3, temp4, temp5, loop_end; + const int temp7 = (int)wrk->fxy_scale; + const int temp6 = (x_out_max & ~0x3) << 2; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + assert(wrk->fxy_scale != 0); + if (yscale) { + if (x_out_max >= 4) { + int temp8, temp9, temp10, temp11; + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 4(%[frow]) \n\t" + "lw %[temp2], 8(%[frow]) \n\t" + "lw %[temp5], 12(%[frow]) \n\t" + "mult $ac0, %[temp3], %[temp4] \n\t" + "maddu $ac0, %[temp0], %[yscale] \n\t" + "mult $ac1, %[temp3], %[temp4] \n\t" + "maddu $ac1, %[temp1], %[yscale] \n\t" + "mult $ac2, %[temp3], %[temp4] \n\t" + "maddu $ac2, %[temp2], %[yscale] \n\t" + "mult $ac3, %[temp3], %[temp4] \n\t" + "maddu $ac3, %[temp5], %[yscale] \n\t" + "addiu %[frow], %[frow], 16 \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp5], $ac3 \n\t" + "lw %[temp8], 0(%[irow]) \n\t" + "lw %[temp9], 4(%[irow]) \n\t" + "lw %[temp10], 8(%[irow]) \n\t" + "lw %[temp11], 12(%[irow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "addiu %[irow], %[irow], 16 \n\t" + "subu %[temp8], %[temp8], %[temp0] \n\t" + "subu %[temp9], %[temp9], %[temp1] \n\t" + "subu %[temp10], %[temp10], %[temp2] \n\t" + "subu %[temp11], %[temp11], %[temp5] \n\t" + "mult $ac0, %[temp3], %[temp4] \n\t" + "maddu $ac0, %[temp8], %[temp7] \n\t" + "mult $ac1, %[temp3], %[temp4] \n\t" + "maddu $ac1, %[temp9], %[temp7] \n\t" + "mult $ac2, %[temp3], %[temp4] \n\t" + "maddu $ac2, %[temp10], %[temp7] \n\t" + "mult $ac3, %[temp3], %[temp4] \n\t" + "maddu $ac3, %[temp11], %[temp7] \n\t" + "mfhi %[temp8], $ac0 \n\t" + "mfhi %[temp9], $ac1 \n\t" + "mfhi %[temp10], $ac2 \n\t" + "mfhi %[temp11], $ac3 \n\t" + "sw %[temp0], -16(%[irow]) \n\t" + "sw %[temp1], -12(%[irow]) \n\t" + "sw %[temp2], -8(%[irow]) \n\t" + "sw %[temp5], -4(%[irow]) \n\t" + "sb %[temp8], -4(%[dst]) \n\t" + "sb %[temp9], -3(%[dst]) \n\t" + "sb %[temp10], -2(%[dst]) \n\t" + "sb %[temp11], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end), + [temp8]"=&r"(temp8), [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), + [temp11]"=&r"(temp11), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [yscale]"r"(yscale), [temp6]"r"(temp6) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const uint32_t frac = (uint32_t)MULT_FIX_FLOOR(*frow++, yscale); + const int v = (int)MULT_FIX(*irow - frac, wrk->fxy_scale); + *dst++ = (v > 255) ? 255u : (uint8_t)v; + *irow++ = frac; // new fractional start + } + } else { + if (x_out_max >= 4) { + __asm__ volatile ( + "li %[temp3], 0x10000 \n\t" + "li %[temp4], 0x8000 \n\t" + "addu %[loop_end], %[irow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[irow]) \n\t" + "lw %[temp1], 4(%[irow]) \n\t" + "lw %[temp2], 8(%[irow]) \n\t" + "lw %[temp5], 12(%[irow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "addiu %[irow], %[irow], 16 \n\t" + "mult $ac0, %[temp3], %[temp4] \n\t" + "maddu $ac0, %[temp0], %[temp7] \n\t" + "mult $ac1, %[temp3], %[temp4] \n\t" + "maddu $ac1, %[temp1], %[temp7] \n\t" + "mult $ac2, %[temp3], %[temp4] \n\t" + "maddu $ac2, %[temp2], %[temp7] \n\t" + "mult $ac3, %[temp3], %[temp4] \n\t" + "maddu $ac3, %[temp5], %[temp7] \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp5], $ac3 \n\t" + "sw $zero, -16(%[irow]) \n\t" + "sw $zero, -12(%[irow]) \n\t" + "sw $zero, -8(%[irow]) \n\t" + "sw $zero, -4(%[irow]) \n\t" + "sb %[temp0], -4(%[dst]) \n\t" + "sb %[temp1], -3(%[dst]) \n\t" + "sb %[temp2], -2(%[dst]) \n\t" + "sb %[temp5], -1(%[dst]) \n\t" + "bne %[irow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [irow]"+r"(irow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [temp6]"r"(temp6) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const int v = (int)MULT_FIX_FLOOR(*irow, wrk->fxy_scale); + *dst++ = (v > 255) ? 255u : (uint8_t)v; + *irow++ = 0; + } + } +} +#endif // 0 + +static void ExportRowExpand_MIPSdspR2(WebPRescaler* const wrk) { + int i; + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + int temp0, temp1, temp2, temp3, temp4, temp5, loop_end; + const int temp6 = (x_out_max & ~0x3) << 2; + const int temp7 = (int)wrk->fy_scale; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + if (x_out_max >= 4) { + __asm__ volatile ( + "li %[temp4], 0x10000 \n\t" + "li %[temp5], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 4(%[frow]) \n\t" + "lw %[temp2], 8(%[frow]) \n\t" + "lw %[temp3], 12(%[frow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "addiu %[frow], %[frow], 16 \n\t" + "mult $ac0, %[temp4], %[temp5] \n\t" + "maddu $ac0, %[temp0], %[temp7] \n\t" + "mult $ac1, %[temp4], %[temp5] \n\t" + "maddu $ac1, %[temp1], %[temp7] \n\t" + "mult $ac2, %[temp4], %[temp5] \n\t" + "maddu $ac2, %[temp2], %[temp7] \n\t" + "mult $ac3, %[temp4], %[temp5] \n\t" + "maddu $ac3, %[temp3], %[temp7] \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp3], $ac3 \n\t" + "sb %[temp0], -4(%[dst]) \n\t" + "sb %[temp1], -3(%[dst]) \n\t" + "sb %[temp2], -2(%[dst]) \n\t" + "sb %[temp3], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [dst]"+r"(dst), [loop_end]"=&r"(loop_end), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [temp6]"r"(temp6) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const uint32_t J = *frow++; + const int v = (int)MULT_FIX(J, wrk->fy_scale); + *dst++ = (v > 255) ? 255u : (uint8_t)v; + } + } else { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + if (x_out_max >= 4) { + int temp8, temp9, temp10, temp11; + __asm__ volatile ( + "li %[temp8], 0x10000 \n\t" + "li %[temp9], 0x8000 \n\t" + "addu %[loop_end], %[frow], %[temp6] \n\t" + "1: \n\t" + "lw %[temp0], 0(%[frow]) \n\t" + "lw %[temp1], 4(%[frow]) \n\t" + "lw %[temp2], 8(%[frow]) \n\t" + "lw %[temp3], 12(%[frow]) \n\t" + "lw %[temp4], 0(%[irow]) \n\t" + "lw %[temp5], 4(%[irow]) \n\t" + "lw %[temp10], 8(%[irow]) \n\t" + "lw %[temp11], 12(%[irow]) \n\t" + "addiu %[dst], %[dst], 4 \n\t" + "mult $ac0, %[temp8], %[temp9] \n\t" + "maddu $ac0, %[A], %[temp0] \n\t" + "maddu $ac0, %[B], %[temp4] \n\t" + "mult $ac1, %[temp8], %[temp9] \n\t" + "maddu $ac1, %[A], %[temp1] \n\t" + "maddu $ac1, %[B], %[temp5] \n\t" + "mult $ac2, %[temp8], %[temp9] \n\t" + "maddu $ac2, %[A], %[temp2] \n\t" + "maddu $ac2, %[B], %[temp10] \n\t" + "mult $ac3, %[temp8], %[temp9] \n\t" + "maddu $ac3, %[A], %[temp3] \n\t" + "maddu $ac3, %[B], %[temp11] \n\t" + "addiu %[frow], %[frow], 16 \n\t" + "addiu %[irow], %[irow], 16 \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp3], $ac3 \n\t" + "mult $ac0, %[temp8], %[temp9] \n\t" + "maddu $ac0, %[temp0], %[temp7] \n\t" + "mult $ac1, %[temp8], %[temp9] \n\t" + "maddu $ac1, %[temp1], %[temp7] \n\t" + "mult $ac2, %[temp8], %[temp9] \n\t" + "maddu $ac2, %[temp2], %[temp7] \n\t" + "mult $ac3, %[temp8], %[temp9] \n\t" + "maddu $ac3, %[temp3], %[temp7] \n\t" + "mfhi %[temp0], $ac0 \n\t" + "mfhi %[temp1], $ac1 \n\t" + "mfhi %[temp2], $ac2 \n\t" + "mfhi %[temp3], $ac3 \n\t" + "sb %[temp0], -4(%[dst]) \n\t" + "sb %[temp1], -3(%[dst]) \n\t" + "sb %[temp2], -2(%[dst]) \n\t" + "sb %[temp3], -1(%[dst]) \n\t" + "bne %[frow], %[loop_end], 1b \n\t" + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp3]"=&r"(temp3), + [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [frow]"+r"(frow), + [irow]"+r"(irow), [dst]"+r"(dst), [loop_end]"=&r"(loop_end), + [temp8]"=&r"(temp8), [temp9]"=&r"(temp9), [temp10]"=&r"(temp10), + [temp11]"=&r"(temp11), [temp2]"=&r"(temp2) + : [temp7]"r"(temp7), [temp6]"r"(temp6), [A]"r"(A), [B]"r"(B) + : "memory", "hi", "lo", "$ac1hi", "$ac1lo", + "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" + ); + } + for (i = 0; i < (x_out_max & 0x3); ++i) { + const uint64_t I = (uint64_t)A * *frow++ + + (uint64_t)B * *irow++; + const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); + const int v = (int)MULT_FIX(J, wrk->fy_scale); + *dst++ = (v > 255) ? 255u : (uint8_t)v; + } + } +} + +#undef MULT_FIX_FLOOR +#undef MULT_FIX +#undef ROUNDER + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPRescalerDspInitMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitMIPSdspR2(void) { + WebPRescalerExportRowExpand = ExportRowExpand_MIPSdspR2; +// WebPRescalerExportRowShrink = ExportRowShrink_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 diff --git a/media/libwebp/dsp/rescaler_msa.c b/media/libwebp/dsp/rescaler_msa.c new file mode 100644 index 0000000000..3366b6d637 --- /dev/null +++ b/media/libwebp/dsp/rescaler_msa.c @@ -0,0 +1,443 @@ +// Copyright 2016 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MSA version of rescaling functions +// +// Author: Prashant Patil (prashant.patil@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) && !defined(WEBP_REDUCE_SIZE) + +#include <assert.h> + +#include "../utils/rescaler_utils.h" +#include "../dsp/msa_macro.h" + +#define ROUNDER (WEBP_RESCALER_ONE >> 1) +#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX) +#define MULT_FIX_FLOOR(x, y) (((uint64_t)(x) * (y)) >> WEBP_RESCALER_RFIX) + +#define CALC_MULT_FIX_16(in0, in1, in2, in3, scale, shift, dst) do { \ + v4u32 tmp0, tmp1, tmp2, tmp3; \ + v16u8 t0, t1, t2, t3, t4, t5; \ + v2u64 out0, out1, out2, out3; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in1, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, scale, scale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_B2_UB(out1, out0, out3, out2, t0, t1); \ + ILVRL_W2_UW(zero, in2, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in3, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, scale, scale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_B2_UB(out1, out0, out3, out2, t2, t3); \ + PCKEV_B2_UB(t1, t0, t3, t2, t4, t5); \ + dst = (v16u8)__msa_pckev_b((v16i8)t5, (v16i8)t4); \ +} while (0) + +#define CALC_MULT_FIX_4(in0, scale, shift, dst) do { \ + v4u32 tmp0, tmp1; \ + v16i8 t0, t1; \ + v2u64 out0, out1; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + t0 = __msa_pckev_b((v16i8)out1, (v16i8)out0); \ + t1 = __msa_pckev_b(t0, t0); \ + t0 = __msa_pckev_b(t1, t1); \ + dst = __msa_copy_s_w((v4i32)t0, 0); \ +} while (0) + +#define CALC_MULT_FIX1_16(in0, in1, in2, in3, fyscale, shift, \ + dst0, dst1, dst2, dst3) do { \ + v4u32 tmp0, tmp1, tmp2, tmp3; \ + v2u64 out0, out1, out2, out3; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in1, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, fyscale, fyscale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, fyscale, fyscale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_W2_UW(out1, out0, out3, out2, dst0, dst1); \ + ILVRL_W2_UW(zero, in2, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in3, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, fyscale, fyscale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, fyscale, fyscale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_W2_UW(out1, out0, out3, out2, dst2, dst3); \ +} while (0) + +#define CALC_MULT_FIX1_4(in0, scale, shift, dst) do { \ + v4u32 tmp0, tmp1; \ + v2u64 out0, out1; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + dst = (v4u32)__msa_pckev_w((v4i32)out1, (v4i32)out0); \ +} while (0) + +#define CALC_MULT_FIX2_16(in0, in1, in2, in3, mult, scale, shift, \ + dst0, dst1) do { \ + v4u32 tmp0, tmp1, tmp2, tmp3; \ + v2u64 out0, out1, out2, out3; \ + ILVRL_W2_UW(in0, in2, tmp0, tmp1); \ + ILVRL_W2_UW(in1, in3, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, mult, mult, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, mult, mult, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + DOTP_UW2_UD(out0, out1, scale, scale, out0, out1); \ + DOTP_UW2_UD(out2, out3, scale, scale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_B2_UB(out1, out0, out3, out2, dst0, dst1); \ +} while (0) + +#define CALC_MULT_FIX2_4(in0, in1, mult, scale, shift, dst) do { \ + v4u32 tmp0, tmp1; \ + v2u64 out0, out1; \ + v16i8 t0, t1; \ + ILVRL_W2_UW(in0, in1, tmp0, tmp1); \ + DOTP_UW2_UD(tmp0, tmp1, mult, mult, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + DOTP_UW2_UD(out0, out1, scale, scale, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + t0 = __msa_pckev_b((v16i8)out1, (v16i8)out0); \ + t1 = __msa_pckev_b(t0, t0); \ + t0 = __msa_pckev_b(t1, t1); \ + dst = __msa_copy_s_w((v4i32)t0, 0); \ +} while (0) + +static WEBP_INLINE void ExportRowExpand_0(const uint32_t* frow, uint8_t* dst, + int length, + WebPRescaler* const wrk) { + const v4u32 scale = (v4u32)__msa_fill_w(wrk->fy_scale); + const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + const v4i32 zero = { 0 }; + + while (length >= 16) { + v4u32 src0, src1, src2, src3; + v16u8 out; + LD_UW4(frow, 4, src0, src1, src2, src3); + CALC_MULT_FIX_16(src0, src1, src2, src3, scale, shift, out); + ST_UB(out, dst); + length -= 16; + frow += 16; + dst += 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 src0, src1, src2; + LD_UW3(frow, 4, src0, src1, src2); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + CALC_MULT_FIX_4(src2, scale, shift, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + length -= 12; + frow += 12; + dst += 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 src0, src1; + LD_UW2(frow, 4, src0, src1); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + SW2(val0_m, val1_m, dst, 4); + length -= 8; + frow += 8; + dst += 8; + } else if (length >= 4) { + uint32_t val0_m; + const v4u32 src0 = LD_UW(frow); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + SW(val0_m, dst); + length -= 4; + frow += 4; + dst += 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const uint32_t J = frow[x_out]; + const int v = (int)MULT_FIX(J, wrk->fy_scale); + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; + } + } +} + +static WEBP_INLINE void ExportRowExpand_1(const uint32_t* frow, uint32_t* irow, + uint8_t* dst, int length, + WebPRescaler* const wrk) { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + const v4i32 B1 = __msa_fill_w(B); + const v4i32 A1 = __msa_fill_w(A); + const v4i32 AB = __msa_ilvr_w(A1, B1); + const v4u32 scale = (v4u32)__msa_fill_w(wrk->fy_scale); + const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + + while (length >= 16) { + v4u32 frow0, frow1, frow2, frow3, irow0, irow1, irow2, irow3; + v16u8 t0, t1, t2, t3, t4, t5; + LD_UW4(frow, 4, frow0, frow1, frow2, frow3); + LD_UW4(irow, 4, irow0, irow1, irow2, irow3); + CALC_MULT_FIX2_16(frow0, frow1, irow0, irow1, AB, scale, shift, t0, t1); + CALC_MULT_FIX2_16(frow2, frow3, irow2, irow3, AB, scale, shift, t2, t3); + PCKEV_B2_UB(t1, t0, t3, t2, t4, t5); + t0 = (v16u8)__msa_pckev_b((v16i8)t5, (v16i8)t4); + ST_UB(t0, dst); + frow += 16; + irow += 16; + dst += 16; + length -= 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 frow0, frow1, frow2, irow0, irow1, irow2; + LD_UW3(frow, 4, frow0, frow1, frow2); + LD_UW3(irow, 4, irow0, irow1, irow2); + CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m); + CALC_MULT_FIX2_4(frow1, irow1, AB, scale, shift, val1_m); + CALC_MULT_FIX2_4(frow2, irow2, AB, scale, shift, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + frow += 12; + irow += 12; + dst += 12; + length -= 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 frow0, frow1, irow0, irow1; + LD_UW2(frow, 4, frow0, frow1); + LD_UW2(irow, 4, irow0, irow1); + CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m); + CALC_MULT_FIX2_4(frow1, irow1, AB, scale, shift, val1_m); + SW2(val0_m, val1_m, dst, 4); + frow += 4; + irow += 4; + dst += 4; + length -= 4; + } else if (length >= 4) { + uint32_t val0_m; + const v4u32 frow0 = LD_UW(frow + 0); + const v4u32 irow0 = LD_UW(irow + 0); + CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m); + SW(val0_m, dst); + frow += 4; + irow += 4; + dst += 4; + length -= 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const uint64_t I = (uint64_t)A * frow[x_out] + + (uint64_t)B * irow[x_out]; + const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); + const int v = (int)MULT_FIX(J, wrk->fy_scale); + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; + } + } +} + +static void RescalerExportRowExpand_MIPSdspR2(WebPRescaler* const wrk) { + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + ExportRowExpand_0(frow, dst, x_out_max, wrk); + } else { + ExportRowExpand_1(frow, irow, dst, x_out_max, wrk); + } +} + +#if 0 // disabled for now. TODO(skal): make match the C-code +static WEBP_INLINE void ExportRowShrink_0(const uint32_t* frow, uint32_t* irow, + uint8_t* dst, int length, + const uint32_t yscale, + WebPRescaler* const wrk) { + const v4u32 y_scale = (v4u32)__msa_fill_w(yscale); + const v4u32 fxyscale = (v4u32)__msa_fill_w(wrk->fxy_scale); + const v4u32 shiftval = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + const v4i32 zero = { 0 }; + + while (length >= 16) { + v4u32 src0, src1, src2, src3, frac0, frac1, frac2, frac3; + v16u8 out; + LD_UW4(frow, 4, src0, src1, src2, src3); + CALC_MULT_FIX1_16(src0, src1, src2, src3, y_scale, shiftval, + frac0, frac1, frac2, frac3); + LD_UW4(irow, 4, src0, src1, src2, src3); + SUB4(src0, frac0, src1, frac1, src2, frac2, src3, frac3, + src0, src1, src2, src3); + CALC_MULT_FIX_16(src0, src1, src2, src3, fxyscale, shiftval, out); + ST_UB(out, dst); + ST_UW4(frac0, frac1, frac2, frac3, irow, 4); + frow += 16; + irow += 16; + dst += 16; + length -= 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 src0, src1, src2, frac0, frac1, frac2; + LD_UW3(frow, 4, src0, src1, src2); + CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0); + CALC_MULT_FIX1_4(src1, y_scale, shiftval, frac1); + CALC_MULT_FIX1_4(src2, y_scale, shiftval, frac2); + LD_UW3(irow, 4, src0, src1, src2); + SUB3(src0, frac0, src1, frac1, src2, frac2, src0, src1, src2); + CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m); + CALC_MULT_FIX_4(src1, fxyscale, shiftval, val1_m); + CALC_MULT_FIX_4(src2, fxyscale, shiftval, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + ST_UW3(frac0, frac1, frac2, irow, 4); + frow += 12; + irow += 12; + dst += 12; + length -= 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 src0, src1, frac0, frac1; + LD_UW2(frow, 4, src0, src1); + CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0); + CALC_MULT_FIX1_4(src1, y_scale, shiftval, frac1); + LD_UW2(irow, 4, src0, src1); + SUB2(src0, frac0, src1, frac1, src0, src1); + CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m); + CALC_MULT_FIX_4(src1, fxyscale, shiftval, val1_m); + SW2(val0_m, val1_m, dst, 4); + ST_UW2(frac0, frac1, irow, 4); + frow += 8; + irow += 8; + dst += 8; + length -= 8; + } else if (length >= 4) { + uint32_t val0_m; + v4u32 frac0; + v4u32 src0 = LD_UW(frow); + CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0); + src0 = LD_UW(irow); + src0 = src0 - frac0; + CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m); + SW(val0_m, dst); + ST_UW(frac0, irow); + frow += 4; + irow += 4; + dst += 4; + length -= 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const uint32_t frac = (uint32_t)MULT_FIX_FLOOR(frow[x_out], yscale); + const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale); + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; + irow[x_out] = frac; + } + } +} + +static WEBP_INLINE void ExportRowShrink_1(uint32_t* irow, uint8_t* dst, + int length, + WebPRescaler* const wrk) { + const v4u32 scale = (v4u32)__msa_fill_w(wrk->fxy_scale); + const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + const v4i32 zero = { 0 }; + + while (length >= 16) { + v4u32 src0, src1, src2, src3; + v16u8 dst0; + LD_UW4(irow, 4, src0, src1, src2, src3); + CALC_MULT_FIX_16(src0, src1, src2, src3, scale, shift, dst0); + ST_UB(dst0, dst); + ST_SW4(zero, zero, zero, zero, irow, 4); + length -= 16; + irow += 16; + dst += 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 src0, src1, src2; + LD_UW3(irow, 4, src0, src1, src2); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + CALC_MULT_FIX_4(src2, scale, shift, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + ST_SW3(zero, zero, zero, irow, 4); + length -= 12; + irow += 12; + dst += 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 src0, src1; + LD_UW2(irow, 4, src0, src1); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + SW2(val0_m, val1_m, dst, 4); + ST_SW2(zero, zero, irow, 4); + length -= 8; + irow += 8; + dst += 8; + } else if (length >= 4) { + uint32_t val0_m; + const v4u32 src0 = LD_UW(irow + 0); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + SW(val0_m, dst); + ST_SW(zero, irow); + length -= 4; + irow += 4; + dst += 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const int v = (int)MULT_FIX(irow[x_out], wrk->fxy_scale); + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; + irow[x_out] = 0; + } + } +} + +static void RescalerExportRowShrink_MIPSdspR2(WebPRescaler* const wrk) { + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + const uint32_t yscale = wrk->fy_scale * (-wrk->y_accum); + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + if (yscale) { + ExportRowShrink_0(frow, irow, dst, x_out_max, yscale, wrk); + } else { + ExportRowShrink_1(irow, dst, x_out_max, wrk); + } +} +#endif // 0 + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPRescalerDspInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitMSA(void) { + WebPRescalerExportRowExpand = RescalerExportRowExpand_MIPSdspR2; +// WebPRescalerExportRowShrink = RescalerExportRowShrink_MIPSdspR2; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitMSA) + +#endif // WEBP_USE_MSA diff --git a/media/libwebp/dsp/rescaler_neon.c b/media/libwebp/dsp/rescaler_neon.c index b560d0cdcc..62bef72113 100644 --- a/media/libwebp/dsp/rescaler_neon.c +++ b/media/libwebp/dsp/rescaler_neon.c @@ -81,14 +81,13 @@ static void RescalerExportRowExpand_NEON(WebPRescaler* const wrk) { const uint32x4_t B1 = MULT_FIX(A1, fy_scale_half); const uint16x4_t C0 = vmovn_u32(B0); const uint16x4_t C1 = vmovn_u32(B1); - const uint8x8_t D = vmovn_u16(vcombine_u16(C0, C1)); + const uint8x8_t D = vqmovn_u16(vcombine_u16(C0, C1)); vst1_u8(dst + x_out, D); } for (; x_out < x_out_max; ++x_out) { const uint32_t J = frow[x_out]; const int v = (int)MULT_FIX_C(J, fy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; } } else { const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); @@ -102,7 +101,7 @@ static void RescalerExportRowExpand_NEON(WebPRescaler* const wrk) { const uint32x4_t D1 = MULT_FIX(C1, fy_scale_half); const uint16x4_t E0 = vmovn_u32(D0); const uint16x4_t E1 = vmovn_u32(D1); - const uint8x8_t F = vmovn_u16(vcombine_u16(E0, E1)); + const uint8x8_t F = vqmovn_u16(vcombine_u16(E0, E1)); vst1_u8(dst + x_out, F); } for (; x_out < x_out_max; ++x_out) { @@ -110,8 +109,7 @@ static void RescalerExportRowExpand_NEON(WebPRescaler* const wrk) { + (uint64_t)B * irow[x_out]; const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); const int v = (int)MULT_FIX_C(J, fy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; } } } @@ -135,23 +133,22 @@ static void RescalerExportRowShrink_NEON(WebPRescaler* const wrk) { for (x_out = 0; x_out < max_span; x_out += 8) { LOAD_32x8(frow + x_out, in0, in1); LOAD_32x8(irow + x_out, in2, in3); - const uint32x4_t A0 = MULT_FIX(in0, yscale_half); - const uint32x4_t A1 = MULT_FIX(in1, yscale_half); + const uint32x4_t A0 = MULT_FIX_FLOOR(in0, yscale_half); + const uint32x4_t A1 = MULT_FIX_FLOOR(in1, yscale_half); const uint32x4_t B0 = vqsubq_u32(in2, A0); const uint32x4_t B1 = vqsubq_u32(in3, A1); - const uint32x4_t C0 = MULT_FIX_FLOOR(B0, fxy_scale_half); - const uint32x4_t C1 = MULT_FIX_FLOOR(B1, fxy_scale_half); + const uint32x4_t C0 = MULT_FIX(B0, fxy_scale_half); + const uint32x4_t C1 = MULT_FIX(B1, fxy_scale_half); const uint16x4_t D0 = vmovn_u32(C0); const uint16x4_t D1 = vmovn_u32(C1); - const uint8x8_t E = vmovn_u16(vcombine_u16(D0, D1)); + const uint8x8_t E = vqmovn_u16(vcombine_u16(D0, D1)); vst1_u8(dst + x_out, E); STORE_32x8(A0, A1, irow + x_out); } for (; x_out < x_out_max; ++x_out) { - const uint32_t frac = (uint32_t)MULT_FIX_C(frow[x_out], yscale); - const int v = (int)MULT_FIX_FLOOR_C(irow[x_out] - frac, fxy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + const uint32_t frac = (uint32_t)MULT_FIX_FLOOR_C(frow[x_out], yscale); + const int v = (int)MULT_FIX_C(irow[x_out] - frac, fxy_scale); + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; irow[x_out] = frac; // new fractional start } } else { @@ -161,14 +158,13 @@ static void RescalerExportRowShrink_NEON(WebPRescaler* const wrk) { const uint32x4_t A1 = MULT_FIX(in1, fxy_scale_half); const uint16x4_t B0 = vmovn_u32(A0); const uint16x4_t B1 = vmovn_u32(A1); - const uint8x8_t C = vmovn_u16(vcombine_u16(B0, B1)); + const uint8x8_t C = vqmovn_u16(vcombine_u16(B0, B1)); vst1_u8(dst + x_out, C); STORE_32x8(zero, zero, irow + x_out); } for (; x_out < x_out_max; ++x_out) { const int v = (int)MULT_FIX_C(irow[x_out], fxy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; irow[x_out] = 0; } } diff --git a/media/libwebp/dsp/rescaler_sse2.c b/media/libwebp/dsp/rescaler_sse2.c index 2d35f76ab0..237997c808 100644 --- a/media/libwebp/dsp/rescaler_sse2.c +++ b/media/libwebp/dsp/rescaler_sse2.c @@ -225,35 +225,6 @@ static WEBP_INLINE void ProcessRow_SSE2(const __m128i* const A0, _mm_storel_epi64((__m128i*)dst, G); } -static WEBP_INLINE void ProcessRow_Floor_SSE2(const __m128i* const A0, - const __m128i* const A1, - const __m128i* const A2, - const __m128i* const A3, - const __m128i* const mult, - uint8_t* const dst) { - const __m128i mask = _mm_set_epi32(0xffffffffu, 0, 0xffffffffu, 0); - const __m128i B0 = _mm_mul_epu32(*A0, *mult); - const __m128i B1 = _mm_mul_epu32(*A1, *mult); - const __m128i B2 = _mm_mul_epu32(*A2, *mult); - const __m128i B3 = _mm_mul_epu32(*A3, *mult); - const __m128i D0 = _mm_srli_epi64(B0, WEBP_RESCALER_RFIX); - const __m128i D1 = _mm_srli_epi64(B1, WEBP_RESCALER_RFIX); -#if (WEBP_RESCALER_RFIX < 32) - const __m128i D2 = - _mm_and_si128(_mm_slli_epi64(B2, 32 - WEBP_RESCALER_RFIX), mask); - const __m128i D3 = - _mm_and_si128(_mm_slli_epi64(B3, 32 - WEBP_RESCALER_RFIX), mask); -#else - const __m128i D2 = _mm_and_si128(B2, mask); - const __m128i D3 = _mm_and_si128(B3, mask); -#endif - const __m128i E0 = _mm_or_si128(D0, D2); - const __m128i E1 = _mm_or_si128(D1, D3); - const __m128i F = _mm_packs_epi32(E0, E1); - const __m128i G = _mm_packus_epi16(F, F); - _mm_storel_epi64((__m128i*)dst, G); -} - static void RescalerExportRowExpand_SSE2(WebPRescaler* const wrk) { int x_out; uint8_t* const dst = wrk->dst; @@ -274,8 +245,7 @@ static void RescalerExportRowExpand_SSE2(WebPRescaler* const wrk) { for (; x_out < x_out_max; ++x_out) { const uint32_t J = frow[x_out]; const int v = (int)MULT_FIX(J, wrk->fy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; } } else { const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); @@ -308,8 +278,7 @@ static void RescalerExportRowExpand_SSE2(WebPRescaler* const wrk) { + (uint64_t)B * irow[x_out]; const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); const int v = (int)MULT_FIX(J, wrk->fy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; } } } @@ -328,20 +297,15 @@ static void RescalerExportRowShrink_SSE2(WebPRescaler* const wrk) { const int scale_xy = wrk->fxy_scale; const __m128i mult_xy = _mm_set_epi32(0, scale_xy, 0, scale_xy); const __m128i mult_y = _mm_set_epi32(0, yscale, 0, yscale); - const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER); for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) { __m128i A0, A1, A2, A3, B0, B1, B2, B3; LoadDispatchAndMult_SSE2(irow + x_out, NULL, &A0, &A1, &A2, &A3); LoadDispatchAndMult_SSE2(frow + x_out, &mult_y, &B0, &B1, &B2, &B3); { - const __m128i C0 = _mm_add_epi64(B0, rounder); - const __m128i C1 = _mm_add_epi64(B1, rounder); - const __m128i C2 = _mm_add_epi64(B2, rounder); - const __m128i C3 = _mm_add_epi64(B3, rounder); - const __m128i D0 = _mm_srli_epi64(C0, WEBP_RESCALER_RFIX); // = frac - const __m128i D1 = _mm_srli_epi64(C1, WEBP_RESCALER_RFIX); - const __m128i D2 = _mm_srli_epi64(C2, WEBP_RESCALER_RFIX); - const __m128i D3 = _mm_srli_epi64(C3, WEBP_RESCALER_RFIX); + const __m128i D0 = _mm_srli_epi64(B0, WEBP_RESCALER_RFIX); // = frac + const __m128i D1 = _mm_srli_epi64(B1, WEBP_RESCALER_RFIX); + const __m128i D2 = _mm_srli_epi64(B2, WEBP_RESCALER_RFIX); + const __m128i D3 = _mm_srli_epi64(B3, WEBP_RESCALER_RFIX); const __m128i E0 = _mm_sub_epi64(A0, D0); // irow[x] - frac const __m128i E1 = _mm_sub_epi64(A1, D1); const __m128i E2 = _mm_sub_epi64(A2, D2); @@ -352,14 +316,13 @@ static void RescalerExportRowShrink_SSE2(WebPRescaler* const wrk) { const __m128i G1 = _mm_or_si128(D1, F3); _mm_storeu_si128((__m128i*)(irow + x_out + 0), G0); _mm_storeu_si128((__m128i*)(irow + x_out + 4), G1); - ProcessRow_Floor_SSE2(&E0, &E1, &E2, &E3, &mult_xy, dst + x_out); + ProcessRow_SSE2(&E0, &E1, &E2, &E3, &mult_xy, dst + x_out); } } for (; x_out < x_out_max; ++x_out) { - const uint32_t frac = (int)MULT_FIX(frow[x_out], yscale); - const int v = (int)MULT_FIX_FLOOR(irow[x_out] - frac, wrk->fxy_scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + const uint32_t frac = (int)MULT_FIX_FLOOR(frow[x_out], yscale); + const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale); + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; irow[x_out] = frac; // new fractional start } } else { @@ -375,8 +338,7 @@ static void RescalerExportRowShrink_SSE2(WebPRescaler* const wrk) { } for (; x_out < x_out_max; ++x_out) { const int v = (int)MULT_FIX(irow[x_out], scale); - assert(v >= 0 && v <= 255); - dst[x_out] = v; + dst[x_out] = (v > 255) ? 255u : (uint8_t)v; irow[x_out] = 0; } } diff --git a/media/libwebp/dsp/ssim.c b/media/libwebp/dsp/ssim.c new file mode 100644 index 0000000000..4141e32fd8 --- /dev/null +++ b/media/libwebp/dsp/ssim.c @@ -0,0 +1,159 @@ +// Copyright 2017 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// distortion calculation +// +// Author: Skal (pascal.massimino@gmail.com) + +#include <assert.h> +#include <stdlib.h> // for abs() + +#include "../dsp/dsp.h" + +#if !defined(WEBP_REDUCE_SIZE) + +//------------------------------------------------------------------------------ +// SSIM / PSNR + +// hat-shaped filter. Sum of coefficients is equal to 16. +static const uint32_t kWeight[2 * VP8_SSIM_KERNEL + 1] = { + 1, 2, 3, 4, 3, 2, 1 +}; +static const uint32_t kWeightSum = 16 * 16; // sum{kWeight}^2 + +static WEBP_INLINE double SSIMCalculation( + const VP8DistoStats* const stats, uint32_t N /*num samples*/) { + const uint32_t w2 = N * N; + const uint32_t C1 = 20 * w2; + const uint32_t C2 = 60 * w2; + const uint32_t C3 = 8 * 8 * w2; // 'dark' limit ~= 6 + const uint64_t xmxm = (uint64_t)stats->xm * stats->xm; + const uint64_t ymym = (uint64_t)stats->ym * stats->ym; + if (xmxm + ymym >= C3) { + const int64_t xmym = (int64_t)stats->xm * stats->ym; + const int64_t sxy = (int64_t)stats->xym * N - xmym; // can be negative + const uint64_t sxx = (uint64_t)stats->xxm * N - xmxm; + const uint64_t syy = (uint64_t)stats->yym * N - ymym; + // we descale by 8 to prevent overflow during the fnum/fden multiply. + const uint64_t num_S = (2 * (uint64_t)(sxy < 0 ? 0 : sxy) + C2) >> 8; + const uint64_t den_S = (sxx + syy + C2) >> 8; + const uint64_t fnum = (2 * xmym + C1) * num_S; + const uint64_t fden = (xmxm + ymym + C1) * den_S; + const double r = (double)fnum / fden; + assert(r >= 0. && r <= 1.0); + return r; + } + return 1.; // area is too dark to contribute meaningfully +} + +double VP8SSIMFromStats(const VP8DistoStats* const stats) { + return SSIMCalculation(stats, kWeightSum); +} + +double VP8SSIMFromStatsClipped(const VP8DistoStats* const stats) { + return SSIMCalculation(stats, stats->w); +} + +static double SSIMGetClipped_C(const uint8_t* src1, int stride1, + const uint8_t* src2, int stride2, + int xo, int yo, int W, int H) { + VP8DistoStats stats = { 0, 0, 0, 0, 0, 0 }; + const int ymin = (yo - VP8_SSIM_KERNEL < 0) ? 0 : yo - VP8_SSIM_KERNEL; + const int ymax = (yo + VP8_SSIM_KERNEL > H - 1) ? H - 1 + : yo + VP8_SSIM_KERNEL; + const int xmin = (xo - VP8_SSIM_KERNEL < 0) ? 0 : xo - VP8_SSIM_KERNEL; + const int xmax = (xo + VP8_SSIM_KERNEL > W - 1) ? W - 1 + : xo + VP8_SSIM_KERNEL; + int x, y; + src1 += ymin * stride1; + src2 += ymin * stride2; + for (y = ymin; y <= ymax; ++y, src1 += stride1, src2 += stride2) { + for (x = xmin; x <= xmax; ++x) { + const uint32_t w = kWeight[VP8_SSIM_KERNEL + x - xo] + * kWeight[VP8_SSIM_KERNEL + y - yo]; + const uint32_t s1 = src1[x]; + const uint32_t s2 = src2[x]; + stats.w += w; + stats.xm += w * s1; + stats.ym += w * s2; + stats.xxm += w * s1 * s1; + stats.xym += w * s1 * s2; + stats.yym += w * s2 * s2; + } + } + return VP8SSIMFromStatsClipped(&stats); +} + +static double SSIMGet_C(const uint8_t* src1, int stride1, + const uint8_t* src2, int stride2) { + VP8DistoStats stats = { 0, 0, 0, 0, 0, 0 }; + int x, y; + for (y = 0; y <= 2 * VP8_SSIM_KERNEL; ++y, src1 += stride1, src2 += stride2) { + for (x = 0; x <= 2 * VP8_SSIM_KERNEL; ++x) { + const uint32_t w = kWeight[x] * kWeight[y]; + const uint32_t s1 = src1[x]; + const uint32_t s2 = src2[x]; + stats.xm += w * s1; + stats.ym += w * s2; + stats.xxm += w * s1 * s1; + stats.xym += w * s1 * s2; + stats.yym += w * s2 * s2; + } + } + return VP8SSIMFromStats(&stats); +} + +#endif // !defined(WEBP_REDUCE_SIZE) + +//------------------------------------------------------------------------------ + +#if !defined(WEBP_DISABLE_STATS) +static uint32_t AccumulateSSE_C(const uint8_t* src1, + const uint8_t* src2, int len) { + int i; + uint32_t sse2 = 0; + assert(len <= 65535); // to ensure that accumulation fits within uint32_t + for (i = 0; i < len; ++i) { + const int32_t diff = src1[i] - src2[i]; + sse2 += diff * diff; + } + return sse2; +} +#endif + +//------------------------------------------------------------------------------ + +#if !defined(WEBP_REDUCE_SIZE) +VP8SSIMGetFunc VP8SSIMGet; +VP8SSIMGetClippedFunc VP8SSIMGetClipped; +#endif +#if !defined(WEBP_DISABLE_STATS) +VP8AccumulateSSEFunc VP8AccumulateSSE; +#endif + +extern void VP8SSIMDspInitSSE2(void); + +WEBP_DSP_INIT_FUNC(VP8SSIMDspInit) { +#if !defined(WEBP_REDUCE_SIZE) + VP8SSIMGetClipped = SSIMGetClipped_C; + VP8SSIMGet = SSIMGet_C; +#endif + +#if !defined(WEBP_DISABLE_STATS) + VP8AccumulateSSE = AccumulateSSE_C; +#endif + + if (VP8GetCPUInfo != NULL) { +#if defined(WEBP_HAVE_SSE2) + if (VP8GetCPUInfo(kSSE2)) { + VP8SSIMDspInitSSE2(); + } +#endif + } +} diff --git a/media/libwebp/dsp/ssim_sse2.c b/media/libwebp/dsp/ssim_sse2.c new file mode 100644 index 0000000000..89810a77c8 --- /dev/null +++ b/media/libwebp/dsp/ssim_sse2.c @@ -0,0 +1,165 @@ +// Copyright 2017 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// SSE2 version of distortion calculation +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_SSE2) + +#include <assert.h> +#include <emmintrin.h> + +#include "../dsp/common_sse2.h" + +#if !defined(WEBP_DISABLE_STATS) + +// Helper function +static WEBP_INLINE void SubtractAndSquare_SSE2(const __m128i a, const __m128i b, + __m128i* const sum) { + // take abs(a-b) in 8b + const __m128i a_b = _mm_subs_epu8(a, b); + const __m128i b_a = _mm_subs_epu8(b, a); + const __m128i abs_a_b = _mm_or_si128(a_b, b_a); + // zero-extend to 16b + const __m128i zero = _mm_setzero_si128(); + const __m128i C0 = _mm_unpacklo_epi8(abs_a_b, zero); + const __m128i C1 = _mm_unpackhi_epi8(abs_a_b, zero); + // multiply with self + const __m128i sum1 = _mm_madd_epi16(C0, C0); + const __m128i sum2 = _mm_madd_epi16(C1, C1); + *sum = _mm_add_epi32(sum1, sum2); +} + +//------------------------------------------------------------------------------ +// SSIM / PSNR entry point + +static uint32_t AccumulateSSE_SSE2(const uint8_t* src1, + const uint8_t* src2, int len) { + int i = 0; + uint32_t sse2 = 0; + if (len >= 16) { + const int limit = len - 32; + int32_t tmp[4]; + __m128i sum1; + __m128i sum = _mm_setzero_si128(); + __m128i a0 = _mm_loadu_si128((const __m128i*)&src1[i]); + __m128i b0 = _mm_loadu_si128((const __m128i*)&src2[i]); + i += 16; + while (i <= limit) { + const __m128i a1 = _mm_loadu_si128((const __m128i*)&src1[i]); + const __m128i b1 = _mm_loadu_si128((const __m128i*)&src2[i]); + __m128i sum2; + i += 16; + SubtractAndSquare_SSE2(a0, b0, &sum1); + sum = _mm_add_epi32(sum, sum1); + a0 = _mm_loadu_si128((const __m128i*)&src1[i]); + b0 = _mm_loadu_si128((const __m128i*)&src2[i]); + i += 16; + SubtractAndSquare_SSE2(a1, b1, &sum2); + sum = _mm_add_epi32(sum, sum2); + } + SubtractAndSquare_SSE2(a0, b0, &sum1); + sum = _mm_add_epi32(sum, sum1); + _mm_storeu_si128((__m128i*)tmp, sum); + sse2 += (tmp[3] + tmp[2] + tmp[1] + tmp[0]); + } + + for (; i < len; ++i) { + const int32_t diff = src1[i] - src2[i]; + sse2 += diff * diff; + } + return sse2; +} +#endif // !defined(WEBP_DISABLE_STATS) + +#if !defined(WEBP_REDUCE_SIZE) + +static uint32_t HorizontalAdd16b_SSE2(const __m128i* const m) { + uint16_t tmp[8]; + const __m128i a = _mm_srli_si128(*m, 8); + const __m128i b = _mm_add_epi16(*m, a); + _mm_storeu_si128((__m128i*)tmp, b); + return (uint32_t)tmp[3] + tmp[2] + tmp[1] + tmp[0]; +} + +static uint32_t HorizontalAdd32b_SSE2(const __m128i* const m) { + const __m128i a = _mm_srli_si128(*m, 8); + const __m128i b = _mm_add_epi32(*m, a); + const __m128i c = _mm_add_epi32(b, _mm_srli_si128(b, 4)); + return (uint32_t)_mm_cvtsi128_si32(c); +} + +static const uint16_t kWeight[] = { 1, 2, 3, 4, 3, 2, 1, 0 }; + +#define ACCUMULATE_ROW(WEIGHT) do { \ + /* compute row weight (Wx * Wy) */ \ + const __m128i Wy = _mm_set1_epi16((WEIGHT)); \ + const __m128i W = _mm_mullo_epi16(Wx, Wy); \ + /* process 8 bytes at a time (7 bytes, actually) */ \ + const __m128i a0 = _mm_loadl_epi64((const __m128i*)src1); \ + const __m128i b0 = _mm_loadl_epi64((const __m128i*)src2); \ + /* convert to 16b and multiply by weight */ \ + const __m128i a1 = _mm_unpacklo_epi8(a0, zero); \ + const __m128i b1 = _mm_unpacklo_epi8(b0, zero); \ + const __m128i wa1 = _mm_mullo_epi16(a1, W); \ + const __m128i wb1 = _mm_mullo_epi16(b1, W); \ + /* accumulate */ \ + xm = _mm_add_epi16(xm, wa1); \ + ym = _mm_add_epi16(ym, wb1); \ + xxm = _mm_add_epi32(xxm, _mm_madd_epi16(a1, wa1)); \ + xym = _mm_add_epi32(xym, _mm_madd_epi16(a1, wb1)); \ + yym = _mm_add_epi32(yym, _mm_madd_epi16(b1, wb1)); \ + src1 += stride1; \ + src2 += stride2; \ +} while (0) + +static double SSIMGet_SSE2(const uint8_t* src1, int stride1, + const uint8_t* src2, int stride2) { + VP8DistoStats stats; + const __m128i zero = _mm_setzero_si128(); + __m128i xm = zero, ym = zero; // 16b accums + __m128i xxm = zero, yym = zero, xym = zero; // 32b accum + const __m128i Wx = _mm_loadu_si128((const __m128i*)kWeight); + assert(2 * VP8_SSIM_KERNEL + 1 == 7); + ACCUMULATE_ROW(1); + ACCUMULATE_ROW(2); + ACCUMULATE_ROW(3); + ACCUMULATE_ROW(4); + ACCUMULATE_ROW(3); + ACCUMULATE_ROW(2); + ACCUMULATE_ROW(1); + stats.xm = HorizontalAdd16b_SSE2(&xm); + stats.ym = HorizontalAdd16b_SSE2(&ym); + stats.xxm = HorizontalAdd32b_SSE2(&xxm); + stats.xym = HorizontalAdd32b_SSE2(&xym); + stats.yym = HorizontalAdd32b_SSE2(&yym); + return VP8SSIMFromStats(&stats); +} + +#endif // !defined(WEBP_REDUCE_SIZE) + +extern void VP8SSIMDspInitSSE2(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8SSIMDspInitSSE2(void) { +#if !defined(WEBP_DISABLE_STATS) + VP8AccumulateSSE = AccumulateSSE_SSE2; +#endif +#if !defined(WEBP_REDUCE_SIZE) + VP8SSIMGet = SSIMGet_SSE2; +#endif +} + +#else // !WEBP_USE_SSE2 + +WEBP_DSP_INIT_STUB(VP8SSIMDspInitSSE2) + +#endif // WEBP_USE_SSE2 diff --git a/media/libwebp/dsp/upsampling.c b/media/libwebp/dsp/upsampling.c index b76483a3a6..aa1c807e89 100644 --- a/media/libwebp/dsp/upsampling.c +++ b/media/libwebp/dsp/upsampling.c @@ -233,12 +233,12 @@ WEBP_DSP_INIT_FUNC(WebPInitYUV444Converters) { WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444_C; if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitYUV444ConvertersSSE2(); } #endif -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitYUV444ConvertersSSE41(); } @@ -278,12 +278,12 @@ WEBP_DSP_INIT_FUNC(WebPInitUpsamplers) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitUpsamplersSSE2(); } #endif -#if defined(WEBP_USE_SSE41) +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitUpsamplersSSE41(); } @@ -300,7 +300,7 @@ WEBP_DSP_INIT_FUNC(WebPInitUpsamplers) { #endif } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPInitUpsamplersNEON(); diff --git a/media/libwebp/dsp/upsampling_mips_dsp_r2.c b/media/libwebp/dsp/upsampling_mips_dsp_r2.c new file mode 100644 index 0000000000..2789c29e02 --- /dev/null +++ b/media/libwebp/dsp/upsampling_mips_dsp_r2.c @@ -0,0 +1,291 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// YUV to RGB upsampling functions. +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include <assert.h> +#include "../dsp/yuv.h" + +#define YUV_TO_RGB(Y, U, V, R, G, B) do { \ + const int t1 = MultHi(Y, 19077); \ + const int t2 = MultHi(V, 13320); \ + R = MultHi(V, 26149); \ + G = MultHi(U, 6419); \ + B = MultHi(U, 33050); \ + R = t1 + R; \ + G = t1 - G; \ + B = t1 + B; \ + R = R - 14234; \ + G = G - t2 + 8708; \ + B = B - 17685; \ + __asm__ volatile ( \ + "shll_s.w %[" #R "], %[" #R "], 17 \n\t" \ + "shll_s.w %[" #G "], %[" #G "], 17 \n\t" \ + "shll_s.w %[" #B "], %[" #B "], 17 \n\t" \ + "precrqu_s.qb.ph %[" #R "], %[" #R "], $zero \n\t" \ + "precrqu_s.qb.ph %[" #G "], %[" #G "], $zero \n\t" \ + "precrqu_s.qb.ph %[" #B "], %[" #B "], $zero \n\t" \ + "srl %[" #R "], %[" #R "], 24 \n\t" \ + "srl %[" #G "], %[" #G "], 24 \n\t" \ + "srl %[" #B "], %[" #B "], 24 \n\t" \ + : [R]"+r"(R), [G]"+r"(G), [B]"+r"(B) \ + : \ + ); \ + } while (0) + +#if !defined(WEBP_REDUCE_CSP) +static WEBP_INLINE void YuvToRgb(int y, int u, int v, uint8_t* const rgb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + rgb[0] = r; + rgb[1] = g; + rgb[2] = b; +} +static WEBP_INLINE void YuvToBgr(int y, int u, int v, uint8_t* const bgr) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + bgr[0] = b; + bgr[1] = g; + bgr[2] = r; +} +static WEBP_INLINE void YuvToRgb565(int y, int u, int v, uint8_t* const rgb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + { + const int rg = (r & 0xf8) | (g >> 5); + const int gb = ((g << 3) & 0xe0) | (b >> 3); +#if (WEBP_SWAP_16BIT_CSP == 1) + rgb[0] = gb; + rgb[1] = rg; +#else + rgb[0] = rg; + rgb[1] = gb; +#endif + } +} +static WEBP_INLINE void YuvToRgba4444(int y, int u, int v, + uint8_t* const argb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + { + const int rg = (r & 0xf0) | (g >> 4); + const int ba = (b & 0xf0) | 0x0f; // overwrite the lower 4 bits +#if (WEBP_SWAP_16BIT_CSP == 1) + argb[0] = ba; + argb[1] = rg; +#else + argb[0] = rg; + argb[1] = ba; +#endif + } +} +#endif // WEBP_REDUCE_CSP + +//----------------------------------------------------------------------------- +// Alpha handling variants + +#if !defined(WEBP_REDUCE_CSP) +static WEBP_INLINE void YuvToArgb(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const argb) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + argb[0] = 0xff; + argb[1] = r; + argb[2] = g; + argb[3] = b; +} +#endif // WEBP_REDUCE_CSP +static WEBP_INLINE void YuvToBgra(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const bgra) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + bgra[0] = b; + bgra[1] = g; + bgra[2] = r; + bgra[3] = 0xff; +} +static WEBP_INLINE void YuvToRgba(uint8_t y, uint8_t u, uint8_t v, + uint8_t* const rgba) { + int r, g, b; + YUV_TO_RGB(y, u, v, r, g, b); + rgba[0] = r; + rgba[1] = g; + rgba[2] = b; + rgba[3] = 0xff; +} + +//------------------------------------------------------------------------------ +// Fancy upsampler + +#ifdef FANCY_UPSAMPLING + +// Given samples laid out in a square as: +// [a b] +// [c d] +// we interpolate u/v as: +// ([9*a + 3*b + 3*c + d 3*a + 9*b + 3*c + d] + [8 8]) / 16 +// ([3*a + b + 9*c + 3*d a + 3*b + 3*c + 9*d] [8 8]) / 16 + +// We process u and v together stashed into 32bit (16bit each). +#define LOAD_UV(u, v) ((u) | ((v) << 16)) + +#define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ +static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ + const uint8_t* top_u, const uint8_t* top_v, \ + const uint8_t* cur_u, const uint8_t* cur_v, \ + uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ + int x; \ + const int last_pixel_pair = (len - 1) >> 1; \ + uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]); /* top-left sample */ \ + uint32_t l_uv = LOAD_UV(cur_u[0], cur_v[0]); /* left-sample */ \ + assert(top_y != NULL); \ + { \ + const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ + FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \ + } \ + if (bottom_y != NULL) { \ + const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ + FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst); \ + } \ + for (x = 1; x <= last_pixel_pair; ++x) { \ + const uint32_t t_uv = LOAD_UV(top_u[x], top_v[x]); /* top sample */ \ + const uint32_t uv = LOAD_UV(cur_u[x], cur_v[x]); /* sample */ \ + /* precompute invariant values associated with first and second diagonals*/\ + const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u; \ + const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3; \ + const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3; \ + { \ + const uint32_t uv0 = (diag_12 + tl_uv) >> 1; \ + const uint32_t uv1 = (diag_03 + t_uv) >> 1; \ + FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ + top_dst + (2 * x - 1) * XSTEP); \ + FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16), \ + top_dst + (2 * x - 0) * XSTEP); \ + } \ + if (bottom_y != NULL) { \ + const uint32_t uv0 = (diag_03 + l_uv) >> 1; \ + const uint32_t uv1 = (diag_12 + uv) >> 1; \ + FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ + bottom_dst + (2 * x - 1) * XSTEP); \ + FUNC(bottom_y[2 * x + 0], uv1 & 0xff, (uv1 >> 16), \ + bottom_dst + (2 * x + 0) * XSTEP); \ + } \ + tl_uv = t_uv; \ + l_uv = uv; \ + } \ + if (!(len & 1)) { \ + { \ + const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ + FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ + top_dst + (len - 1) * XSTEP); \ + } \ + if (bottom_y != NULL) { \ + const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ + FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ + bottom_dst + (len - 1) * XSTEP); \ + } \ + } \ +} + +// All variants implemented. +UPSAMPLE_FUNC(UpsampleRgbaLinePair, YuvToRgba, 4) +UPSAMPLE_FUNC(UpsampleBgraLinePair, YuvToBgra, 4) +#if !defined(WEBP_REDUCE_CSP) +UPSAMPLE_FUNC(UpsampleRgbLinePair, YuvToRgb, 3) +UPSAMPLE_FUNC(UpsampleBgrLinePair, YuvToBgr, 3) +UPSAMPLE_FUNC(UpsampleArgbLinePair, YuvToArgb, 4) +UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2) +UPSAMPLE_FUNC(UpsampleRgb565LinePair, YuvToRgb565, 2) +#endif // WEBP_REDUCE_CSP + +#undef LOAD_UV +#undef UPSAMPLE_FUNC + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitUpsamplersMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMIPSdspR2(void) { + WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair; + WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair; +#if !defined(WEBP_REDUCE_CSP) + WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair; + WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair; + WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair; + WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair; + WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair; + WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair; + WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair; +#endif // WEBP_REDUCE_CSP +} + +#endif // FANCY_UPSAMPLING + +//------------------------------------------------------------------------------ +// YUV444 converter + +#define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP) \ +static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + int i; \ + for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]); \ +} + +YUV444_FUNC(Yuv444ToRgba, YuvToRgba, 4) +YUV444_FUNC(Yuv444ToBgra, YuvToBgra, 4) +#if !defined(WEBP_REDUCE_CSP) +YUV444_FUNC(Yuv444ToRgb, YuvToRgb, 3) +YUV444_FUNC(Yuv444ToBgr, YuvToBgr, 3) +YUV444_FUNC(Yuv444ToArgb, YuvToArgb, 4) +YUV444_FUNC(Yuv444ToRgba4444, YuvToRgba4444, 2) +YUV444_FUNC(Yuv444ToRgb565, YuvToRgb565, 2) +#endif // WEBP_REDUCE_CSP + +#undef YUV444_FUNC + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitYUV444ConvertersMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersMIPSdspR2(void) { + WebPYUV444Converters[MODE_RGBA] = Yuv444ToRgba; + WebPYUV444Converters[MODE_BGRA] = Yuv444ToBgra; + WebPYUV444Converters[MODE_rgbA] = Yuv444ToRgba; + WebPYUV444Converters[MODE_bgrA] = Yuv444ToBgra; +#if !defined(WEBP_REDUCE_CSP) + WebPYUV444Converters[MODE_RGB] = Yuv444ToRgb; + WebPYUV444Converters[MODE_BGR] = Yuv444ToBgr; + WebPYUV444Converters[MODE_ARGB] = Yuv444ToArgb; + WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444; + WebPYUV444Converters[MODE_RGB_565] = Yuv444ToRgb565; + WebPYUV444Converters[MODE_Argb] = Yuv444ToArgb; + WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444; +#endif // WEBP_REDUCE_CSP +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPInitYUV444ConvertersMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 + +#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_MIPS_DSP_R2)) +WEBP_DSP_INIT_STUB(WebPInitUpsamplersMIPSdspR2) +#endif diff --git a/media/libwebp/dsp/upsampling_msa.c b/media/libwebp/dsp/upsampling_msa.c new file mode 100644 index 0000000000..d8ef6feb58 --- /dev/null +++ b/media/libwebp/dsp/upsampling_msa.c @@ -0,0 +1,688 @@ +// Copyright 2016 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MSA version of YUV to RGB upsampling functions. +// +// Author: Prashant Patil (prashant.patil@imgtec.com) + +#include <string.h> +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MSA) + +#include "../dsp/msa_macro.h" +#include "../dsp/yuv.h" + +#ifdef FANCY_UPSAMPLING + +#define ILVR_UW2(in, out0, out1) do { \ + const v8i16 t0 = (v8i16)__msa_ilvr_b((v16i8)zero, (v16i8)in); \ + out0 = (v4u32)__msa_ilvr_h((v8i16)zero, t0); \ + out1 = (v4u32)__msa_ilvl_h((v8i16)zero, t0); \ +} while (0) + +#define ILVRL_UW4(in, out0, out1, out2, out3) do { \ + v16u8 t0, t1; \ + ILVRL_B2_UB(zero, in, t0, t1); \ + ILVRL_H2_UW(zero, t0, out0, out1); \ + ILVRL_H2_UW(zero, t1, out2, out3); \ +} while (0) + +#define MULTHI_16(in0, in1, in2, in3, cnst, out0, out1) do { \ + const v4i32 const0 = (v4i32)__msa_fill_w(cnst * 256); \ + v4u32 temp0, temp1, temp2, temp3; \ + MUL4(in0, const0, in1, const0, in2, const0, in3, const0, \ + temp0, temp1, temp2, temp3); \ + PCKOD_H2_UH(temp1, temp0, temp3, temp2, out0, out1); \ +} while (0) + +#define MULTHI_8(in0, in1, cnst, out0) do { \ + const v4i32 const0 = (v4i32)__msa_fill_w(cnst * 256); \ + v4u32 temp0, temp1; \ + MUL2(in0, const0, in1, const0, temp0, temp1); \ + out0 = (v8u16)__msa_pckod_h((v8i16)temp1, (v8i16)temp0); \ +} while (0) + +#define CALC_R16(y0, y1, v0, v1, dst) do { \ + const v8i16 const_a = (v8i16)__msa_fill_h(14234); \ + const v8i16 a0 = __msa_adds_s_h((v8i16)y0, (v8i16)v0); \ + const v8i16 a1 = __msa_adds_s_h((v8i16)y1, (v8i16)v1); \ + v8i16 b0 = __msa_subs_s_h(a0, const_a); \ + v8i16 b1 = __msa_subs_s_h(a1, const_a); \ + SRAI_H2_SH(b0, b1, 6); \ + CLIP_SH2_0_255(b0, b1); \ + dst = (v16u8)__msa_pckev_b((v16i8)b1, (v16i8)b0); \ +} while (0) + +#define CALC_R8(y0, v0, dst) do { \ + const v8i16 const_a = (v8i16)__msa_fill_h(14234); \ + const v8i16 a0 = __msa_adds_s_h((v8i16)y0, (v8i16)v0); \ + v8i16 b0 = __msa_subs_s_h(a0, const_a); \ + b0 = SRAI_H(b0, 6); \ + CLIP_SH_0_255(b0); \ + dst = (v16u8)__msa_pckev_b((v16i8)b0, (v16i8)b0); \ +} while (0) + +#define CALC_G16(y0, y1, u0, u1, v0, v1, dst) do { \ + const v8i16 const_a = (v8i16)__msa_fill_h(8708); \ + v8i16 a0 = __msa_subs_s_h((v8i16)y0, (v8i16)u0); \ + v8i16 a1 = __msa_subs_s_h((v8i16)y1, (v8i16)u1); \ + const v8i16 b0 = __msa_subs_s_h(a0, (v8i16)v0); \ + const v8i16 b1 = __msa_subs_s_h(a1, (v8i16)v1); \ + a0 = __msa_adds_s_h(b0, const_a); \ + a1 = __msa_adds_s_h(b1, const_a); \ + SRAI_H2_SH(a0, a1, 6); \ + CLIP_SH2_0_255(a0, a1); \ + dst = (v16u8)__msa_pckev_b((v16i8)a1, (v16i8)a0); \ +} while (0) + +#define CALC_G8(y0, u0, v0, dst) do { \ + const v8i16 const_a = (v8i16)__msa_fill_h(8708); \ + v8i16 a0 = __msa_subs_s_h((v8i16)y0, (v8i16)u0); \ + const v8i16 b0 = __msa_subs_s_h(a0, (v8i16)v0); \ + a0 = __msa_adds_s_h(b0, const_a); \ + a0 = SRAI_H(a0, 6); \ + CLIP_SH_0_255(a0); \ + dst = (v16u8)__msa_pckev_b((v16i8)a0, (v16i8)a0); \ +} while (0) + +#define CALC_B16(y0, y1, u0, u1, dst) do { \ + const v8u16 const_a = (v8u16)__msa_fill_h(17685); \ + const v8u16 a0 = __msa_adds_u_h((v8u16)y0, u0); \ + const v8u16 a1 = __msa_adds_u_h((v8u16)y1, u1); \ + v8u16 b0 = __msa_subs_u_h(a0, const_a); \ + v8u16 b1 = __msa_subs_u_h(a1, const_a); \ + SRAI_H2_UH(b0, b1, 6); \ + CLIP_UH2_0_255(b0, b1); \ + dst = (v16u8)__msa_pckev_b((v16i8)b1, (v16i8)b0); \ +} while (0) + +#define CALC_B8(y0, u0, dst) do { \ + const v8u16 const_a = (v8u16)__msa_fill_h(17685); \ + const v8u16 a0 = __msa_adds_u_h((v8u16)y0, u0); \ + v8u16 b0 = __msa_subs_u_h(a0, const_a); \ + b0 = SRAI_H(b0, 6); \ + CLIP_UH_0_255(b0); \ + dst = (v16u8)__msa_pckev_b((v16i8)b0, (v16i8)b0); \ +} while (0) + +#define CALC_RGB16(y, u, v, R, G, B) do { \ + const v16u8 zero = { 0 }; \ + v8u16 y0, y1, u0, u1, v0, v1; \ + v4u32 p0, p1, p2, p3; \ + const v16u8 in_y = LD_UB(y); \ + const v16u8 in_u = LD_UB(u); \ + const v16u8 in_v = LD_UB(v); \ + ILVRL_UW4(in_y, p0, p1, p2, p3); \ + MULTHI_16(p0, p1, p2, p3, 19077, y0, y1); \ + ILVRL_UW4(in_v, p0, p1, p2, p3); \ + MULTHI_16(p0, p1, p2, p3, 26149, v0, v1); \ + CALC_R16(y0, y1, v0, v1, R); \ + MULTHI_16(p0, p1, p2, p3, 13320, v0, v1); \ + ILVRL_UW4(in_u, p0, p1, p2, p3); \ + MULTHI_16(p0, p1, p2, p3, 6419, u0, u1); \ + CALC_G16(y0, y1, u0, u1, v0, v1, G); \ + MULTHI_16(p0, p1, p2, p3, 33050, u0, u1); \ + CALC_B16(y0, y1, u0, u1, B); \ +} while (0) + +#define CALC_RGB8(y, u, v, R, G, B) do { \ + const v16u8 zero = { 0 }; \ + v8u16 y0, u0, v0; \ + v4u32 p0, p1; \ + const v16u8 in_y = LD_UB(y); \ + const v16u8 in_u = LD_UB(u); \ + const v16u8 in_v = LD_UB(v); \ + ILVR_UW2(in_y, p0, p1); \ + MULTHI_8(p0, p1, 19077, y0); \ + ILVR_UW2(in_v, p0, p1); \ + MULTHI_8(p0, p1, 26149, v0); \ + CALC_R8(y0, v0, R); \ + MULTHI_8(p0, p1, 13320, v0); \ + ILVR_UW2(in_u, p0, p1); \ + MULTHI_8(p0, p1, 6419, u0); \ + CALC_G8(y0, u0, v0, G); \ + MULTHI_8(p0, p1, 33050, u0); \ + CALC_B8(y0, u0, B); \ +} while (0) + +#define STORE16_3(a0, a1, a2, dst) do { \ + const v16u8 mask0 = { 0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, \ + 8, 9, 20, 10 }; \ + const v16u8 mask1 = { 0, 21, 1, 2, 22, 3, 4, 23, 5, 6, 24, 7, \ + 8, 25, 9, 10 }; \ + const v16u8 mask2 = { 26, 0, 1, 27, 2, 3, 28, 4, 5, 29, 6, 7, \ + 30, 8, 9, 31 }; \ + v16u8 out0, out1, out2, tmp0, tmp1, tmp2; \ + ILVRL_B2_UB(a1, a0, tmp0, tmp1); \ + out0 = VSHF_UB(tmp0, a2, mask0); \ + tmp2 = SLDI_UB(tmp1, tmp0, 11); \ + out1 = VSHF_UB(tmp2, a2, mask1); \ + tmp2 = SLDI_UB(tmp1, tmp1, 6); \ + out2 = VSHF_UB(tmp2, a2, mask2); \ + ST_UB(out0, dst + 0); \ + ST_UB(out1, dst + 16); \ + ST_UB(out2, dst + 32); \ +} while (0) + +#define STORE8_3(a0, a1, a2, dst) do { \ + int64_t out_m; \ + const v16u8 mask0 = { 0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, \ + 8, 9, 20, 10 }; \ + const v16u8 mask1 = { 11, 21, 12, 13, 22, 14, 15, 23, \ + 255, 255, 255, 255, 255, 255, 255, 255 }; \ + const v16u8 tmp0 = (v16u8)__msa_ilvr_b((v16i8)a1, (v16i8)a0); \ + v16u8 out0, out1; \ + VSHF_B2_UB(tmp0, a2, tmp0, a2, mask0, mask1, out0, out1); \ + ST_UB(out0, dst); \ + out_m = __msa_copy_s_d((v2i64)out1, 0); \ + SD(out_m, dst + 16); \ +} while (0) + +#define STORE16_4(a0, a1, a2, a3, dst) do { \ + v16u8 tmp0, tmp1, tmp2, tmp3; \ + v16u8 out0, out1, out2, out3; \ + ILVRL_B2_UB(a1, a0, tmp0, tmp1); \ + ILVRL_B2_UB(a3, a2, tmp2, tmp3); \ + ILVRL_H2_UB(tmp2, tmp0, out0, out1); \ + ILVRL_H2_UB(tmp3, tmp1, out2, out3); \ + ST_UB(out0, dst + 0); \ + ST_UB(out1, dst + 16); \ + ST_UB(out2, dst + 32); \ + ST_UB(out3, dst + 48); \ +} while (0) + +#define STORE8_4(a0, a1, a2, a3, dst) do { \ + v16u8 tmp0, tmp1, tmp2, tmp3; \ + ILVR_B2_UB(a1, a0, a3, a2, tmp0, tmp1); \ + ILVRL_H2_UB(tmp1, tmp0, tmp2, tmp3); \ + ST_UB(tmp2, dst + 0); \ + ST_UB(tmp3, dst + 16); \ +} while (0) + +#define STORE2_16(a0, a1, dst) do { \ + v16u8 out0, out1; \ + ILVRL_B2_UB(a1, a0, out0, out1); \ + ST_UB(out0, dst + 0); \ + ST_UB(out1, dst + 16); \ +} while (0) + +#define STORE2_8(a0, a1, dst) do { \ + const v16u8 out0 = (v16u8)__msa_ilvr_b((v16i8)a1, (v16i8)a0); \ + ST_UB(out0, dst); \ +} while (0) + +#define CALC_RGBA4444(y, u, v, out0, out1, N, dst) do { \ + CALC_RGB##N(y, u, v, R, G, B); \ + tmp0 = ANDI_B(R, 0xf0); \ + tmp1 = SRAI_B(G, 4); \ + RG = tmp0 | tmp1; \ + tmp0 = ANDI_B(B, 0xf0); \ + BA = ORI_B(tmp0, 0x0f); \ + STORE2_##N(out0, out1, dst); \ +} while (0) + +#define CALC_RGB565(y, u, v, out0, out1, N, dst) do { \ + CALC_RGB##N(y, u, v, R, G, B); \ + tmp0 = ANDI_B(R, 0xf8); \ + tmp1 = SRAI_B(G, 5); \ + RG = tmp0 | tmp1; \ + tmp0 = SLLI_B(G, 3); \ + tmp1 = ANDI_B(tmp0, 0xe0); \ + tmp0 = SRAI_B(B, 3); \ + GB = tmp0 | tmp1; \ + STORE2_##N(out0, out1, dst); \ +} while (0) + +static WEBP_INLINE int Clip8(int v) { + return v < 0 ? 0 : v > 255 ? 255 : v; +} + +static void YuvToRgb(int y, int u, int v, uint8_t* const rgb) { + const int y1 = MultHi(y, 19077); + const int r1 = y1 + MultHi(v, 26149) - 14234; + const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708; + const int b1 = y1 + MultHi(u, 33050) - 17685; + rgb[0] = Clip8(r1 >> 6); + rgb[1] = Clip8(g1 >> 6); + rgb[2] = Clip8(b1 >> 6); +} + +static void YuvToBgr(int y, int u, int v, uint8_t* const bgr) { + const int y1 = MultHi(y, 19077); + const int r1 = y1 + MultHi(v, 26149) - 14234; + const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708; + const int b1 = y1 + MultHi(u, 33050) - 17685; + bgr[0] = Clip8(b1 >> 6); + bgr[1] = Clip8(g1 >> 6); + bgr[2] = Clip8(r1 >> 6); +} + +#if !defined(WEBP_REDUCE_CSP) +static void YuvToRgb565(int y, int u, int v, uint8_t* const rgb) { + const int y1 = MultHi(y, 19077); + const int r1 = y1 + MultHi(v, 26149) - 14234; + const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708; + const int b1 = y1 + MultHi(u, 33050) - 17685; + const int r = Clip8(r1 >> 6); + const int g = Clip8(g1 >> 6); + const int b = Clip8(b1 >> 6); + const int rg = (r & 0xf8) | (g >> 5); + const int gb = ((g << 3) & 0xe0) | (b >> 3); +#if (WEBP_SWAP_16BIT_CSP == 1) + rgb[0] = gb; + rgb[1] = rg; +#else + rgb[0] = rg; + rgb[1] = gb; +#endif +} + +static void YuvToRgba4444(int y, int u, int v, uint8_t* const argb) { + const int y1 = MultHi(y, 19077); + const int r1 = y1 + MultHi(v, 26149) - 14234; + const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708; + const int b1 = y1 + MultHi(u, 33050) - 17685; + const int r = Clip8(r1 >> 6); + const int g = Clip8(g1 >> 6); + const int b = Clip8(b1 >> 6); + const int rg = (r & 0xf0) | (g >> 4); + const int ba = (b & 0xf0) | 0x0f; // overwrite the lower 4 bits +#if (WEBP_SWAP_16BIT_CSP == 1) + argb[0] = ba; + argb[1] = rg; +#else + argb[0] = rg; + argb[1] = ba; +#endif +} + +static void YuvToArgb(uint8_t y, uint8_t u, uint8_t v, uint8_t* const argb) { + argb[0] = 0xff; + YuvToRgb(y, u, v, argb + 1); +} +#endif // WEBP_REDUCE_CSP + +static void YuvToBgra(uint8_t y, uint8_t u, uint8_t v, uint8_t* const bgra) { + YuvToBgr(y, u, v, bgra); + bgra[3] = 0xff; +} + +static void YuvToRgba(uint8_t y, uint8_t u, uint8_t v, uint8_t* const rgba) { + YuvToRgb(y, u, v, rgba); + rgba[3] = 0xff; +} + +#if !defined(WEBP_REDUCE_CSP) +static void YuvToRgbLine(const uint8_t* y, const uint8_t* u, + const uint8_t* v, uint8_t* dst, int length) { + v16u8 R, G, B; + while (length >= 16) { + CALC_RGB16(y, u, v, R, G, B); + STORE16_3(R, G, B, dst); + y += 16; + u += 16; + v += 16; + dst += 16 * 3; + length -= 16; + } + if (length > 8) { + uint8_t temp[3 * 16] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB16(temp, u, v, R, G, B); + STORE16_3(R, G, B, temp); + memcpy(dst, temp, length * 3 * sizeof(*dst)); + } else if (length > 0) { + uint8_t temp[3 * 8] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB8(temp, u, v, R, G, B); + STORE8_3(R, G, B, temp); + memcpy(dst, temp, length * 3 * sizeof(*dst)); + } +} + +static void YuvToBgrLine(const uint8_t* y, const uint8_t* u, + const uint8_t* v, uint8_t* dst, int length) { + v16u8 R, G, B; + while (length >= 16) { + CALC_RGB16(y, u, v, R, G, B); + STORE16_3(B, G, R, dst); + y += 16; + u += 16; + v += 16; + dst += 16 * 3; + length -= 16; + } + if (length > 8) { + uint8_t temp[3 * 16] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB16(temp, u, v, R, G, B); + STORE16_3(B, G, R, temp); + memcpy(dst, temp, length * 3 * sizeof(*dst)); + } else if (length > 0) { + uint8_t temp[3 * 8] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB8(temp, u, v, R, G, B); + STORE8_3(B, G, R, temp); + memcpy(dst, temp, length * 3 * sizeof(*dst)); + } +} +#endif // WEBP_REDUCE_CSP + +static void YuvToRgbaLine(const uint8_t* y, const uint8_t* u, + const uint8_t* v, uint8_t* dst, int length) { + v16u8 R, G, B; + const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL); + while (length >= 16) { + CALC_RGB16(y, u, v, R, G, B); + STORE16_4(R, G, B, A, dst); + y += 16; + u += 16; + v += 16; + dst += 16 * 4; + length -= 16; + } + if (length > 8) { + uint8_t temp[4 * 16] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB16(&temp[0], u, v, R, G, B); + STORE16_4(R, G, B, A, temp); + memcpy(dst, temp, length * 4 * sizeof(*dst)); + } else if (length > 0) { + uint8_t temp[4 * 8] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB8(temp, u, v, R, G, B); + STORE8_4(R, G, B, A, temp); + memcpy(dst, temp, length * 4 * sizeof(*dst)); + } +} + +static void YuvToBgraLine(const uint8_t* y, const uint8_t* u, + const uint8_t* v, uint8_t* dst, int length) { + v16u8 R, G, B; + const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL); + while (length >= 16) { + CALC_RGB16(y, u, v, R, G, B); + STORE16_4(B, G, R, A, dst); + y += 16; + u += 16; + v += 16; + dst += 16 * 4; + length -= 16; + } + if (length > 8) { + uint8_t temp[4 * 16] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB16(temp, u, v, R, G, B); + STORE16_4(B, G, R, A, temp); + memcpy(dst, temp, length * 4 * sizeof(*dst)); + } else if (length > 0) { + uint8_t temp[4 * 8] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB8(temp, u, v, R, G, B); + STORE8_4(B, G, R, A, temp); + memcpy(dst, temp, length * 4 * sizeof(*dst)); + } +} + +#if !defined(WEBP_REDUCE_CSP) +static void YuvToArgbLine(const uint8_t* y, const uint8_t* u, + const uint8_t* v, uint8_t* dst, int length) { + v16u8 R, G, B; + const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL); + while (length >= 16) { + CALC_RGB16(y, u, v, R, G, B); + STORE16_4(A, R, G, B, dst); + y += 16; + u += 16; + v += 16; + dst += 16 * 4; + length -= 16; + } + if (length > 8) { + uint8_t temp[4 * 16] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB16(temp, u, v, R, G, B); + STORE16_4(A, R, G, B, temp); + memcpy(dst, temp, length * 4 * sizeof(*dst)); + } else if (length > 0) { + uint8_t temp[4 * 8] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); + CALC_RGB8(temp, u, v, R, G, B); + STORE8_4(A, R, G, B, temp); + memcpy(dst, temp, length * 4 * sizeof(*dst)); + } +} + +static void YuvToRgba4444Line(const uint8_t* y, const uint8_t* u, + const uint8_t* v, uint8_t* dst, int length) { + v16u8 R, G, B, RG, BA, tmp0, tmp1; + while (length >= 16) { +#if (WEBP_SWAP_16BIT_CSP == 1) + CALC_RGBA4444(y, u, v, BA, RG, 16, dst); +#else + CALC_RGBA4444(y, u, v, RG, BA, 16, dst); +#endif + y += 16; + u += 16; + v += 16; + dst += 16 * 2; + length -= 16; + } + if (length > 8) { + uint8_t temp[2 * 16] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); +#if (WEBP_SWAP_16BIT_CSP == 1) + CALC_RGBA4444(temp, u, v, BA, RG, 16, temp); +#else + CALC_RGBA4444(temp, u, v, RG, BA, 16, temp); +#endif + memcpy(dst, temp, length * 2 * sizeof(*dst)); + } else if (length > 0) { + uint8_t temp[2 * 8] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); +#if (WEBP_SWAP_16BIT_CSP == 1) + CALC_RGBA4444(temp, u, v, BA, RG, 8, temp); +#else + CALC_RGBA4444(temp, u, v, RG, BA, 8, temp); +#endif + memcpy(dst, temp, length * 2 * sizeof(*dst)); + } +} + +static void YuvToRgb565Line(const uint8_t* y, const uint8_t* u, + const uint8_t* v, uint8_t* dst, int length) { + v16u8 R, G, B, RG, GB, tmp0, tmp1; + while (length >= 16) { +#if (WEBP_SWAP_16BIT_CSP == 1) + CALC_RGB565(y, u, v, GB, RG, 16, dst); +#else + CALC_RGB565(y, u, v, RG, GB, 16, dst); +#endif + y += 16; + u += 16; + v += 16; + dst += 16 * 2; + length -= 16; + } + if (length > 8) { + uint8_t temp[2 * 16] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); +#if (WEBP_SWAP_16BIT_CSP == 1) + CALC_RGB565(temp, u, v, GB, RG, 16, temp); +#else + CALC_RGB565(temp, u, v, RG, GB, 16, temp); +#endif + memcpy(dst, temp, length * 2 * sizeof(*dst)); + } else if (length > 0) { + uint8_t temp[2 * 8] = { 0 }; + memcpy(temp, y, length * sizeof(*temp)); +#if (WEBP_SWAP_16BIT_CSP == 1) + CALC_RGB565(temp, u, v, GB, RG, 8, temp); +#else + CALC_RGB565(temp, u, v, RG, GB, 8, temp); +#endif + memcpy(dst, temp, length * 2 * sizeof(*dst)); + } +} +#endif // WEBP_REDUCE_CSP + +#define UPSAMPLE_32PIXELS(a, b, c, d) do { \ + v16u8 s = __msa_aver_u_b(a, d); \ + v16u8 t = __msa_aver_u_b(b, c); \ + const v16u8 st = s ^ t; \ + v16u8 ad = a ^ d; \ + v16u8 bc = b ^ c; \ + v16u8 t0 = ad | bc; \ + v16u8 t1 = t0 | st; \ + v16u8 t2 = ANDI_B(t1, 1); \ + v16u8 t3 = __msa_aver_u_b(s, t); \ + const v16u8 k = t3 - t2; \ + v16u8 diag1, diag2; \ + AVER_UB2_UB(t, k, s, k, t0, t1); \ + bc = bc & st; \ + ad = ad & st; \ + t = t ^ k; \ + s = s ^ k; \ + t2 = bc | t; \ + t3 = ad | s; \ + t2 = ANDI_B(t2, 1); \ + t3 = ANDI_B(t3, 1); \ + SUB2(t0, t2, t1, t3, diag1, diag2); \ + AVER_UB2_UB(a, diag1, b, diag2, t0, t1); \ + ILVRL_B2_UB(t1, t0, a, b); \ + if (pbot_y != NULL) { \ + AVER_UB2_UB(c, diag2, d, diag1, t0, t1); \ + ILVRL_B2_UB(t1, t0, c, d); \ + } \ +} while (0) + +#define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ +static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y, \ + const uint8_t* top_u, const uint8_t* top_v, \ + const uint8_t* cur_u, const uint8_t* cur_v, \ + uint8_t* top_dst, uint8_t* bot_dst, int len) \ +{ \ + int size = (len - 1) >> 1; \ + uint8_t temp_u[64]; \ + uint8_t temp_v[64]; \ + const uint32_t tl_uv = ((top_u[0]) | ((top_v[0]) << 16)); \ + const uint32_t l_uv = ((cur_u[0]) | ((cur_v[0]) << 16)); \ + const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ + const uint8_t* ptop_y = &top_y[1]; \ + uint8_t* ptop_dst = top_dst + XSTEP; \ + const uint8_t* pbot_y = &bot_y[1]; \ + uint8_t* pbot_dst = bot_dst + XSTEP; \ + \ + FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \ + if (bot_y != NULL) { \ + const uint32_t uv1 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ + FUNC(bot_y[0], uv1 & 0xff, (uv1 >> 16), bot_dst); \ + } \ + while (size >= 16) { \ + v16u8 tu0, tu1, tv0, tv1, cu0, cu1, cv0, cv1; \ + LD_UB2(top_u, 1, tu0, tu1); \ + LD_UB2(cur_u, 1, cu0, cu1); \ + LD_UB2(top_v, 1, tv0, tv1); \ + LD_UB2(cur_v, 1, cv0, cv1); \ + UPSAMPLE_32PIXELS(tu0, tu1, cu0, cu1); \ + UPSAMPLE_32PIXELS(tv0, tv1, cv0, cv1); \ + ST_UB4(tu0, tu1, cu0, cu1, &temp_u[0], 16); \ + ST_UB4(tv0, tv1, cv0, cv1, &temp_v[0], 16); \ + FUNC##Line(ptop_y, &temp_u[ 0], &temp_v[0], ptop_dst, 32); \ + if (bot_y != NULL) { \ + FUNC##Line(pbot_y, &temp_u[32], &temp_v[32], pbot_dst, 32); \ + } \ + ptop_y += 32; \ + pbot_y += 32; \ + ptop_dst += XSTEP * 32; \ + pbot_dst += XSTEP * 32; \ + top_u += 16; \ + top_v += 16; \ + cur_u += 16; \ + cur_v += 16; \ + size -= 16; \ + } \ + if (size > 0) { \ + v16u8 tu0, tu1, tv0, tv1, cu0, cu1, cv0, cv1; \ + memcpy(&temp_u[ 0], top_u, 17 * sizeof(uint8_t)); \ + memcpy(&temp_u[32], cur_u, 17 * sizeof(uint8_t)); \ + memcpy(&temp_v[ 0], top_v, 17 * sizeof(uint8_t)); \ + memcpy(&temp_v[32], cur_v, 17 * sizeof(uint8_t)); \ + LD_UB2(&temp_u[ 0], 1, tu0, tu1); \ + LD_UB2(&temp_u[32], 1, cu0, cu1); \ + LD_UB2(&temp_v[ 0], 1, tv0, tv1); \ + LD_UB2(&temp_v[32], 1, cv0, cv1); \ + UPSAMPLE_32PIXELS(tu0, tu1, cu0, cu1); \ + UPSAMPLE_32PIXELS(tv0, tv1, cv0, cv1); \ + ST_UB4(tu0, tu1, cu0, cu1, &temp_u[0], 16); \ + ST_UB4(tv0, tv1, cv0, cv1, &temp_v[0], 16); \ + FUNC##Line(ptop_y, &temp_u[ 0], &temp_v[0], ptop_dst, size * 2); \ + if (bot_y != NULL) { \ + FUNC##Line(pbot_y, &temp_u[32], &temp_v[32], pbot_dst, size * 2); \ + } \ + top_u += size; \ + top_v += size; \ + cur_u += size; \ + cur_v += size; \ + } \ + if (!(len & 1)) { \ + const uint32_t t0 = ((top_u[0]) | ((top_v[0]) << 16)); \ + const uint32_t c0 = ((cur_u[0]) | ((cur_v[0]) << 16)); \ + const uint32_t tmp0 = (3 * t0 + c0 + 0x00020002u) >> 2; \ + FUNC(top_y[len - 1], tmp0 & 0xff, (tmp0 >> 16), \ + top_dst + (len - 1) * XSTEP); \ + if (bot_y != NULL) { \ + const uint32_t tmp1 = (3 * c0 + t0 + 0x00020002u) >> 2; \ + FUNC(bot_y[len - 1], tmp1 & 0xff, (tmp1 >> 16), \ + bot_dst + (len - 1) * XSTEP); \ + } \ + } \ +} + +UPSAMPLE_FUNC(UpsampleRgbaLinePair, YuvToRgba, 4) +UPSAMPLE_FUNC(UpsampleBgraLinePair, YuvToBgra, 4) +#if !defined(WEBP_REDUCE_CSP) +UPSAMPLE_FUNC(UpsampleRgbLinePair, YuvToRgb, 3) +UPSAMPLE_FUNC(UpsampleBgrLinePair, YuvToBgr, 3) +UPSAMPLE_FUNC(UpsampleArgbLinePair, YuvToArgb, 4) +UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2) +UPSAMPLE_FUNC(UpsampleRgb565LinePair, YuvToRgb565, 2) +#endif // WEBP_REDUCE_CSP + +//------------------------------------------------------------------------------ +// Entry point + +extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; + +extern void WebPInitUpsamplersMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMSA(void) { + WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair; + WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair; + WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair; +#if !defined(WEBP_REDUCE_CSP) + WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair; + WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair; + WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair; + WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair; + WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair; + WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair; + WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair; +#endif // WEBP_REDUCE_CSP +} + +#endif // FANCY_UPSAMPLING + +#endif // WEBP_USE_MSA + +#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_MSA)) +WEBP_DSP_INIT_STUB(WebPInitUpsamplersMSA) +#endif diff --git a/media/libwebp/dsp/upsampling_neon.c b/media/libwebp/dsp/upsampling_neon.c index c847d70d46..41cb44b03f 100644 --- a/media/libwebp/dsp/upsampling_neon.c +++ b/media/libwebp/dsp/upsampling_neon.c @@ -58,8 +58,8 @@ } while (0) // Turn the macro into a function for reducing code-size when non-critical -static void Upsample16Pixels_NEON(const uint8_t *r1, const uint8_t *r2, - uint8_t *out) { +static void Upsample16Pixels_NEON(const uint8_t* r1, const uint8_t* r2, + uint8_t* out) { UPSAMPLE_16PIXELS(r1, r2, out); } @@ -190,14 +190,14 @@ static const int16_t kCoeffs1[4] = { 19077, 26149, 6419, 13320 }; } #define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP) \ -static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y, \ - const uint8_t *top_u, const uint8_t *top_v, \ - const uint8_t *cur_u, const uint8_t *cur_v, \ - uint8_t *top_dst, uint8_t *bottom_dst, int len) { \ +static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ + const uint8_t* top_u, const uint8_t* top_v, \ + const uint8_t* cur_u, const uint8_t* cur_v, \ + uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ int block; \ /* 16 byte aligned array to cache reconstructed u and v */ \ uint8_t uv_buf[2 * 32 + 15]; \ - uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ + uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ const int uv_len = (len + 1) >> 1; \ /* 9 pixels must be read-able for each block */ \ const int num_blocks = (uv_len - 1) >> 3; \ diff --git a/media/libwebp/dsp/yuv.c b/media/libwebp/dsp/yuv.c index 12c04ca426..bd9db04149 100644 --- a/media/libwebp/dsp/yuv.c +++ b/media/libwebp/dsp/yuv.c @@ -90,16 +90,16 @@ WEBP_DSP_INIT_FUNC(WebPInitSamplers) { // If defined, use CPUInfo() to overwrite some pointers with faster versions. if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitSamplersSSE2(); } -#endif // WEBP_USE_SSE2 -#if defined(WEBP_USE_SSE41) +#endif // WEBP_HAVE_SSE2 +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitSamplersSSE41(); } -#endif // WEBP_USE_SSE41 +#endif // WEBP_HAVE_SSE41 #if defined(WEBP_USE_MIPS32) if (VP8GetCPUInfo(kMIPS32)) { WebPInitSamplersMIPS32(); @@ -276,26 +276,26 @@ WEBP_DSP_INIT_FUNC(WebPInitConvertARGBToYUV) { #endif if (VP8GetCPUInfo != NULL) { -#if defined(WEBP_USE_SSE2) +#if defined(WEBP_HAVE_SSE2) if (VP8GetCPUInfo(kSSE2)) { WebPInitConvertARGBToYUVSSE2(); WebPInitSharpYUVSSE2(); } -#endif // WEBP_USE_SSE2 -#if defined(WEBP_USE_SSE41) +#endif // WEBP_HAVE_SSE2 +#if defined(WEBP_HAVE_SSE41) if (VP8GetCPUInfo(kSSE4_1)) { WebPInitConvertARGBToYUVSSE41(); } -#endif // WEBP_USE_SSE41 +#endif // WEBP_HAVE_SSE41 } -#if defined(WEBP_USE_NEON) +#if defined(WEBP_HAVE_NEON) if (WEBP_NEON_OMIT_C_CODE || (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { WebPInitConvertARGBToYUVNEON(); WebPInitSharpYUVNEON(); } -#endif // WEBP_USE_NEON +#endif // WEBP_HAVE_NEON assert(WebPConvertARGBToY != NULL); assert(WebPConvertARGBToUV != NULL); diff --git a/media/libwebp/dsp/yuv.h b/media/libwebp/dsp/yuv.h index 947b89e13c..28524ec422 100644 --- a/media/libwebp/dsp/yuv.h +++ b/media/libwebp/dsp/yuv.h @@ -10,7 +10,7 @@ // inline YUV<->RGB conversion function // // The exact naming is Y'CbCr, following the ITU-R BT.601 standard. -// More information at: http://en.wikipedia.org/wiki/YCbCr +// More information at: https://en.wikipedia.org/wiki/YCbCr // Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16 // U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128 // V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128 diff --git a/media/libwebp/dsp/yuv_mips32.c b/media/libwebp/dsp/yuv_mips32.c new file mode 100644 index 0000000000..fc7c2cda0e --- /dev/null +++ b/media/libwebp/dsp/yuv_mips32.c @@ -0,0 +1,103 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS version of YUV to RGB upsampling functions. +// +// Author(s): Djordje Pesut (djordje.pesut@imgtec.com) +// Jovan Zelincevic (jovan.zelincevic@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS32) + +#include "../dsp/yuv.h" + +//------------------------------------------------------------------------------ +// simple point-sampling + +#define ROW_FUNC(FUNC_NAME, XSTEP, R, G, B, A) \ +static void FUNC_NAME(const uint8_t* y, \ + const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + int i, r, g, b; \ + int temp0, temp1, temp2, temp3, temp4; \ + for (i = 0; i < (len >> 1); i++) { \ + temp1 = MultHi(v[0], 26149); \ + temp3 = MultHi(v[0], 13320); \ + temp2 = MultHi(u[0], 6419); \ + temp4 = MultHi(u[0], 33050); \ + temp0 = MultHi(y[0], 19077); \ + temp1 -= 14234; \ + temp3 -= 8708; \ + temp2 += temp3; \ + temp4 -= 17685; \ + r = VP8Clip8(temp0 + temp1); \ + g = VP8Clip8(temp0 - temp2); \ + b = VP8Clip8(temp0 + temp4); \ + temp0 = MultHi(y[1], 19077); \ + dst[R] = r; \ + dst[G] = g; \ + dst[B] = b; \ + if (A) dst[A] = 0xff; \ + r = VP8Clip8(temp0 + temp1); \ + g = VP8Clip8(temp0 - temp2); \ + b = VP8Clip8(temp0 + temp4); \ + dst[R + XSTEP] = r; \ + dst[G + XSTEP] = g; \ + dst[B + XSTEP] = b; \ + if (A) dst[A + XSTEP] = 0xff; \ + y += 2; \ + ++u; \ + ++v; \ + dst += 2 * XSTEP; \ + } \ + if (len & 1) { \ + temp1 = MultHi(v[0], 26149); \ + temp3 = MultHi(v[0], 13320); \ + temp2 = MultHi(u[0], 6419); \ + temp4 = MultHi(u[0], 33050); \ + temp0 = MultHi(y[0], 19077); \ + temp1 -= 14234; \ + temp3 -= 8708; \ + temp2 += temp3; \ + temp4 -= 17685; \ + r = VP8Clip8(temp0 + temp1); \ + g = VP8Clip8(temp0 - temp2); \ + b = VP8Clip8(temp0 + temp4); \ + dst[R] = r; \ + dst[G] = g; \ + dst[B] = b; \ + if (A) dst[A] = 0xff; \ + } \ +} + +ROW_FUNC(YuvToRgbRow_MIPS32, 3, 0, 1, 2, 0) +ROW_FUNC(YuvToRgbaRow_MIPS32, 4, 0, 1, 2, 3) +ROW_FUNC(YuvToBgrRow_MIPS32, 3, 2, 1, 0, 0) +ROW_FUNC(YuvToBgraRow_MIPS32, 4, 2, 1, 0, 3) + +#undef ROW_FUNC + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitSamplersMIPS32(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersMIPS32(void) { + WebPSamplers[MODE_RGB] = YuvToRgbRow_MIPS32; + WebPSamplers[MODE_RGBA] = YuvToRgbaRow_MIPS32; + WebPSamplers[MODE_BGR] = YuvToBgrRow_MIPS32; + WebPSamplers[MODE_BGRA] = YuvToBgraRow_MIPS32; +} + +#else // !WEBP_USE_MIPS32 + +WEBP_DSP_INIT_STUB(WebPInitSamplersMIPS32) + +#endif // WEBP_USE_MIPS32 diff --git a/media/libwebp/dsp/yuv_mips_dsp_r2.c b/media/libwebp/dsp/yuv_mips_dsp_r2.c new file mode 100644 index 0000000000..1418a9fba1 --- /dev/null +++ b/media/libwebp/dsp/yuv_mips_dsp_r2.c @@ -0,0 +1,134 @@ +// Copyright 2014 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// MIPS DSPr2 version of YUV to RGB upsampling functions. +// +// Author(s): Branimir Vasic (branimir.vasic@imgtec.com) +// Djordje Pesut (djordje.pesut@imgtec.com) + +#include "../dsp/dsp.h" + +#if defined(WEBP_USE_MIPS_DSP_R2) + +#include "../dsp/yuv.h" + +//------------------------------------------------------------------------------ +// simple point-sampling + +#define ROW_FUNC_PART_1() \ + "lbu %[temp3], 0(%[v]) \n\t" \ + "lbu %[temp4], 0(%[u]) \n\t" \ + "lbu %[temp0], 0(%[y]) \n\t" \ + "mul %[temp1], %[t_con_1], %[temp3] \n\t" \ + "mul %[temp3], %[t_con_2], %[temp3] \n\t" \ + "mul %[temp2], %[t_con_3], %[temp4] \n\t" \ + "mul %[temp4], %[t_con_4], %[temp4] \n\t" \ + "mul %[temp0], %[t_con_5], %[temp0] \n\t" \ + "subu %[temp1], %[temp1], %[t_con_6] \n\t" \ + "subu %[temp3], %[temp3], %[t_con_7] \n\t" \ + "addu %[temp2], %[temp2], %[temp3] \n\t" \ + "subu %[temp4], %[temp4], %[t_con_8] \n\t" \ + +#define ROW_FUNC_PART_2(R, G, B, K) \ + "addu %[temp5], %[temp0], %[temp1] \n\t" \ + "subu %[temp6], %[temp0], %[temp2] \n\t" \ + "addu %[temp7], %[temp0], %[temp4] \n\t" \ +".if " #K " \n\t" \ + "lbu %[temp0], 1(%[y]) \n\t" \ +".endif \n\t" \ + "shll_s.w %[temp5], %[temp5], 17 \n\t" \ + "shll_s.w %[temp6], %[temp6], 17 \n\t" \ +".if " #K " \n\t" \ + "mul %[temp0], %[t_con_5], %[temp0] \n\t" \ +".endif \n\t" \ + "shll_s.w %[temp7], %[temp7], 17 \n\t" \ + "precrqu_s.qb.ph %[temp5], %[temp5], $zero \n\t" \ + "precrqu_s.qb.ph %[temp6], %[temp6], $zero \n\t" \ + "precrqu_s.qb.ph %[temp7], %[temp7], $zero \n\t" \ + "srl %[temp5], %[temp5], 24 \n\t" \ + "srl %[temp6], %[temp6], 24 \n\t" \ + "srl %[temp7], %[temp7], 24 \n\t" \ + "sb %[temp5], " #R "(%[dst]) \n\t" \ + "sb %[temp6], " #G "(%[dst]) \n\t" \ + "sb %[temp7], " #B "(%[dst]) \n\t" \ + +#define ASM_CLOBBER_LIST() \ + : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), \ + [temp3]"=&r"(temp3), [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), \ + [temp6]"=&r"(temp6), [temp7]"=&r"(temp7) \ + : [t_con_1]"r"(t_con_1), [t_con_2]"r"(t_con_2), [t_con_3]"r"(t_con_3), \ + [t_con_4]"r"(t_con_4), [t_con_5]"r"(t_con_5), [t_con_6]"r"(t_con_6), \ + [u]"r"(u), [v]"r"(v), [y]"r"(y), [dst]"r"(dst), \ + [t_con_7]"r"(t_con_7), [t_con_8]"r"(t_con_8) \ + : "memory", "hi", "lo" \ + +#define ROW_FUNC(FUNC_NAME, XSTEP, R, G, B, A) \ +static void FUNC_NAME(const uint8_t* y, \ + const uint8_t* u, const uint8_t* v, \ + uint8_t* dst, int len) { \ + int i; \ + uint32_t temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7; \ + const int t_con_1 = 26149; \ + const int t_con_2 = 13320; \ + const int t_con_3 = 6419; \ + const int t_con_4 = 33050; \ + const int t_con_5 = 19077; \ + const int t_con_6 = 14234; \ + const int t_con_7 = 8708; \ + const int t_con_8 = 17685; \ + for (i = 0; i < (len >> 1); i++) { \ + __asm__ volatile ( \ + ROW_FUNC_PART_1() \ + ROW_FUNC_PART_2(R, G, B, 1) \ + ROW_FUNC_PART_2(R + XSTEP, G + XSTEP, B + XSTEP, 0) \ + ASM_CLOBBER_LIST() \ + ); \ + if (A) dst[A] = dst[A + XSTEP] = 0xff; \ + y += 2; \ + ++u; \ + ++v; \ + dst += 2 * XSTEP; \ + } \ + if (len & 1) { \ + __asm__ volatile ( \ + ROW_FUNC_PART_1() \ + ROW_FUNC_PART_2(R, G, B, 0) \ + ASM_CLOBBER_LIST() \ + ); \ + if (A) dst[A] = 0xff; \ + } \ +} + +ROW_FUNC(YuvToRgbRow_MIPSdspR2, 3, 0, 1, 2, 0) +ROW_FUNC(YuvToRgbaRow_MIPSdspR2, 4, 0, 1, 2, 3) +ROW_FUNC(YuvToBgrRow_MIPSdspR2, 3, 2, 1, 0, 0) +ROW_FUNC(YuvToBgraRow_MIPSdspR2, 4, 2, 1, 0, 3) + +#undef ROW_FUNC +#undef ASM_CLOBBER_LIST +#undef ROW_FUNC_PART_2 +#undef ROW_FUNC_PART_1 + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPInitSamplersMIPSdspR2(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersMIPSdspR2(void) { + WebPSamplers[MODE_RGB] = YuvToRgbRow_MIPSdspR2; + WebPSamplers[MODE_RGBA] = YuvToRgbaRow_MIPSdspR2; + WebPSamplers[MODE_BGR] = YuvToBgrRow_MIPSdspR2; + WebPSamplers[MODE_BGRA] = YuvToBgraRow_MIPSdspR2; +} + +#else // !WEBP_USE_MIPS_DSP_R2 + +WEBP_DSP_INIT_STUB(WebPInitSamplersMIPSdspR2) + +#endif // WEBP_USE_MIPS_DSP_R2 |