diff options
Diffstat (limited to 'libs/libopus/celt')
57 files changed, 1933 insertions, 723 deletions
diff --git a/libs/libopus/celt/_kiss_fft_guts.h b/libs/libopus/celt/_kiss_fft_guts.h index 5e3d58fd6..17392b3e9 100644 --- a/libs/libopus/celt/_kiss_fft_guts.h +++ b/libs/libopus/celt/_kiss_fft_guts.h @@ -58,12 +58,12 @@ # define S_MUL(a,b) MULT16_32_Q15(b, a) # define C_MUL(m,a,b) \ - do{ (m).r = SUB32(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ - (m).i = ADD32(S_MUL((a).r,(b).i) , S_MUL((a).i,(b).r)); }while(0) + do{ (m).r = SUB32_ovflw(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ + (m).i = ADD32_ovflw(S_MUL((a).r,(b).i) , S_MUL((a).i,(b).r)); }while(0) # define C_MULC(m,a,b) \ - do{ (m).r = ADD32(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ - (m).i = SUB32(S_MUL((a).i,(b).r) , S_MUL((a).r,(b).i)); }while(0) + do{ (m).r = ADD32_ovflw(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \ + (m).i = SUB32_ovflw(S_MUL((a).i,(b).r) , S_MUL((a).r,(b).i)); }while(0) # define C_MULBYSCALAR( c, s ) \ do{ (c).r = S_MUL( (c).r , s ) ;\ @@ -77,17 +77,17 @@ DIVSCALAR( (c).i , div); }while (0) #define C_ADD( res, a,b)\ - do {(res).r=ADD32((a).r,(b).r); (res).i=ADD32((a).i,(b).i); \ + do {(res).r=ADD32_ovflw((a).r,(b).r); (res).i=ADD32_ovflw((a).i,(b).i); \ }while(0) #define C_SUB( res, a,b)\ - do {(res).r=SUB32((a).r,(b).r); (res).i=SUB32((a).i,(b).i); \ + do {(res).r=SUB32_ovflw((a).r,(b).r); (res).i=SUB32_ovflw((a).i,(b).i); \ }while(0) #define C_ADDTO( res , a)\ - do {(res).r = ADD32((res).r, (a).r); (res).i = ADD32((res).i,(a).i);\ + do {(res).r = ADD32_ovflw((res).r, (a).r); (res).i = ADD32_ovflw((res).i,(a).i);\ }while(0) #define C_SUBFROM( res , a)\ - do {(res).r = ADD32((res).r,(a).r); (res).i = SUB32((res).i,(a).i); \ + do {(res).r = ADD32_ovflw((res).r,(a).r); (res).i = SUB32_ovflw((res).i,(a).i); \ }while(0) #if defined(OPUS_ARM_INLINE_ASM) diff --git a/libs/libopus/celt/arch.h b/libs/libopus/celt/arch.h index 8ceab5fe1..3845c3a08 100644 --- a/libs/libopus/celt/arch.h +++ b/libs/libopus/celt/arch.h @@ -46,25 +46,53 @@ # endif # endif +#if OPUS_GNUC_PREREQ(3, 0) +#define opus_likely(x) (__builtin_expect(!!(x), 1)) +#define opus_unlikely(x) (__builtin_expect(!!(x), 0)) +#else +#define opus_likely(x) (!!(x)) +#define opus_unlikely(x) (!!(x)) +#endif + #define CELT_SIG_SCALE 32768.f -#define celt_fatal(str) _celt_fatal(str, __FILE__, __LINE__); -#ifdef ENABLE_ASSERTIONS +#define CELT_FATAL(str) celt_fatal(str, __FILE__, __LINE__); + +#if defined(ENABLE_ASSERTIONS) || defined(ENABLE_HARDENING) +#ifdef __GNUC__ +__attribute__((noreturn)) +#endif +void celt_fatal(const char *str, const char *file, int line); + +#if defined(CELT_C) && !defined(OVERRIDE_celt_fatal) #include <stdio.h> #include <stdlib.h> #ifdef __GNUC__ __attribute__((noreturn)) #endif -static OPUS_INLINE void _celt_fatal(const char *str, const char *file, int line) +void celt_fatal(const char *str, const char *file, int line) { fprintf (stderr, "Fatal (internal) error in %s, line %d: %s\n", file, line, str); +#if defined(_MSC_VER) + _set_abort_behavior( 0, _WRITE_ABORT_MSG); +#endif abort(); } -#define celt_assert(cond) {if (!(cond)) {celt_fatal("assertion failed: " #cond);}} -#define celt_assert2(cond, message) {if (!(cond)) {celt_fatal("assertion failed: " #cond "\n" message);}} +#endif + +#define celt_assert(cond) {if (!(cond)) {CELT_FATAL("assertion failed: " #cond);}} +#define celt_assert2(cond, message) {if (!(cond)) {CELT_FATAL("assertion failed: " #cond "\n" message);}} +#define MUST_SUCCEED(call) celt_assert((call) == OPUS_OK) #else #define celt_assert(cond) #define celt_assert2(cond, message) +#define MUST_SUCCEED(call) do {if((call) != OPUS_OK) {RESTORE_STACK; return OPUS_INTERNAL_ERROR;} } while (0) +#endif + +#if defined(ENABLE_ASSERTIONS) +#define celt_sig_assert(cond) {if (!(cond)) {CELT_FATAL("signal assertion failed: " #cond);}} +#else +#define celt_sig_assert(cond) #endif #define IMUL32(a,b) ((a)*(b)) @@ -93,14 +121,20 @@ static OPUS_INLINE void _celt_fatal(const char *str, const char *file, int line) typedef opus_int16 opus_val16; typedef opus_int32 opus_val32; +typedef opus_int64 opus_val64; typedef opus_val32 celt_sig; typedef opus_val16 celt_norm; typedef opus_val32 celt_ener; +#define celt_isnan(x) 0 + #define Q15ONE 32767 #define SIG_SHIFT 12 +/* Safe saturation value for 32-bit signals. Should be less than + 2^31*(1-0.85) to avoid blowing up on DC at deemphasis.*/ +#define SIG_SAT (300000000) #define NORM_SCALING 16384 @@ -129,7 +163,7 @@ static OPUS_INLINE opus_int16 SAT16(opus_int32 x) { #ifdef OPUS_ARM_PRESUME_AARCH64_NEON_INTR #include "arm/fixed_arm64.h" -#elif OPUS_ARM_INLINE_EDSP +#elif defined (OPUS_ARM_INLINE_EDSP) #include "arm/fixed_armv5e.h" #elif defined (OPUS_ARM_INLINE_ASM) #include "arm/fixed_armv4.h" @@ -147,6 +181,7 @@ static OPUS_INLINE opus_int16 SAT16(opus_int32 x) { typedef float opus_val16; typedef float opus_val32; +typedef float opus_val64; typedef float celt_sig; typedef float celt_norm; @@ -186,6 +221,7 @@ static OPUS_INLINE int celt_isnan(float x) #define NEG16(x) (-(x)) #define NEG32(x) (-(x)) +#define NEG32_ovflw(x) (-(x)) #define EXTRACT16(x) (x) #define EXTEND32(x) (x) #define SHR16(a,shift) (a) @@ -202,6 +238,7 @@ static OPUS_INLINE int celt_isnan(float x) #define SATURATE16(x) (x) #define ROUND16(a,shift) (a) +#define SROUND16(a,shift) (a) #define HALF16(x) (.5f*(x)) #define HALF32(x) (.5f*(x)) @@ -209,6 +246,8 @@ static OPUS_INLINE int celt_isnan(float x) #define SUB16(a,b) ((a)-(b)) #define ADD32(a,b) ((a)+(b)) #define SUB32(a,b) ((a)-(b)) +#define ADD32_ovflw(a,b) ((a)+(b)) +#define SUB32_ovflw(a,b) ((a)-(b)) #define MULT16_16_16(a,b) ((a)*(b)) #define MULT16_16(a,b) ((opus_val32)(a)*(opus_val32)(b)) #define MAC16_16(c,a,b) ((c)+(opus_val32)(a)*(opus_val32)(b)) @@ -243,9 +282,9 @@ static OPUS_INLINE int celt_isnan(float x) #ifndef GLOBAL_STACK_SIZE #ifdef FIXED_POINT -#define GLOBAL_STACK_SIZE 100000 +#define GLOBAL_STACK_SIZE 120000 #else -#define GLOBAL_STACK_SIZE 100000 +#define GLOBAL_STACK_SIZE 120000 #endif #endif diff --git a/libs/libopus/celt/arm/arm2gnu.pl b/libs/libopus/celt/arm/arm2gnu.pl index 6c922ac81..a2895f744 100755 --- a/libs/libopus/celt/arm/arm2gnu.pl +++ b/libs/libopus/celt/arm/arm2gnu.pl @@ -164,11 +164,11 @@ while (<>) { $prefix = ""; if ($proc) { - $prefix = $prefix.sprintf("\t.type\t%s, %%function; ",$proc) unless ($apple); + $prefix = $prefix.sprintf("\t.type\t%s, %%function", $proc) unless ($apple); # Make sure we $prefix isn't empty here (for the $apple case). # We handle mangling the label here, make sure it doesn't match # the label handling below (if $prefix would be empty). - $prefix = "; "; + $prefix = $prefix."; "; push(@proc_stack, $proc); s/^[A-Za-z_\.]\w+/$symprefix$&:/; } diff --git a/libs/libopus/celt/arm/arm_celt_map.c b/libs/libopus/celt/arm/arm_celt_map.c index 4d4d069a8..ca988b66f 100644 --- a/libs/libopus/celt/arm/arm_celt_map.c +++ b/libs/libopus/celt/arm/arm_celt_map.c @@ -35,12 +35,29 @@ #if defined(OPUS_HAVE_RTCD) +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR) +opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y, int N) = { + celt_inner_prod_c, /* ARMv4 */ + celt_inner_prod_c, /* EDSP */ + celt_inner_prod_c, /* Media */ + celt_inner_prod_neon /* NEON */ +}; + +void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) = { + dual_inner_prod_c, /* ARMv4 */ + dual_inner_prod_c, /* EDSP */ + dual_inner_prod_c, /* Media */ + dual_inner_prod_neon /* NEON */ +}; +# endif + # if defined(FIXED_POINT) # if ((defined(OPUS_ARM_MAY_HAVE_NEON) && !defined(OPUS_ARM_PRESUME_NEON)) || \ (defined(OPUS_ARM_MAY_HAVE_MEDIA) && !defined(OPUS_ARM_PRESUME_MEDIA)) || \ (defined(OPUS_ARM_MAY_HAVE_EDSP) && !defined(OPUS_ARM_PRESUME_EDSP))) opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int , int) = { + const opus_val16 *, opus_val32 *, int, int, int) = { celt_pitch_xcorr_c, /* ARMv4 */ MAY_HAVE_EDSP(celt_pitch_xcorr), /* EDSP */ MAY_HAVE_MEDIA(celt_pitch_xcorr), /* Media */ @@ -51,7 +68,7 @@ opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, # else /* !FIXED_POINT */ # if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR) void (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int, int) = { + const opus_val16 *, opus_val32 *, int, int, int) = { celt_pitch_xcorr_c, /* ARMv4 */ celt_pitch_xcorr_c, /* EDSP */ celt_pitch_xcorr_c, /* Media */ diff --git a/libs/libopus/celt/arm/armcpu.c b/libs/libopus/celt/arm/armcpu.c index 694a63b78..cce3ae3a9 100644 --- a/libs/libopus/celt/arm/armcpu.c +++ b/libs/libopus/celt/arm/armcpu.c @@ -93,6 +93,8 @@ static OPUS_INLINE opus_uint32 opus_cpu_capabilities(void){ #elif defined(__linux__) /* Linux based */ +#include <stdio.h> + opus_uint32 opus_cpu_capabilities(void) { opus_uint32 flags = 0; diff --git a/libs/libopus/celt/arm/celt_ne10_fft.c b/libs/libopus/celt/arm/celt_fft_ne10.c index 42d96a711..ea5fd7808 100644 --- a/libs/libopus/celt/arm/celt_ne10_fft.c +++ b/libs/libopus/celt/arm/celt_fft_ne10.c @@ -1,7 +1,7 @@ /* Copyright (c) 2015 Xiph.Org Foundation Written by Viswanath Puttagunta */ /** - @file celt_ne10_fft.c + @file celt_fft_ne10.c @brief ARM Neon optimizations for fft using NE10 library */ @@ -36,7 +36,6 @@ #endif #endif -#include <NE10_init.h> #include <NE10_dsp.h> #include "os_support.h" #include "kiss_fft.h" diff --git a/libs/libopus/celt/arm/celt_ne10_mdct.c b/libs/libopus/celt/arm/celt_mdct_ne10.c index 293c3efd7..3531d02d1 100644 --- a/libs/libopus/celt/arm/celt_ne10_mdct.c +++ b/libs/libopus/celt/arm/celt_mdct_ne10.c @@ -1,7 +1,7 @@ /* Copyright (c) 2015 Xiph.Org Foundation Written by Viswanath Puttagunta */ /** - @file celt_ne10_mdct.c + @file celt_mdct_ne10.c @brief ARM Neon optimizations for mdct using NE10 library */ diff --git a/libs/libopus/celt/arm/celt_neon_intr.c b/libs/libopus/celt/arm/celt_neon_intr.c index 47bbe3dc2..effda769d 100644 --- a/libs/libopus/celt/arm/celt_neon_intr.c +++ b/libs/libopus/celt/arm/celt_neon_intr.c @@ -191,121 +191,21 @@ static void xcorr_kernel_neon_float(const float32_t *x, const float32_t *y, vst1q_f32(sum, SUMM); } -/* - * Function: xcorr_kernel_neon_float_process1 - * --------------------------------- - * Computes single correlation values and stores in *sum - */ -static void xcorr_kernel_neon_float_process1(const float32_t *x, - const float32_t *y, float32_t *sum, int len) { - float32x4_t XX[4]; - float32x4_t YY[4]; - float32x2_t XX_2; - float32x2_t YY_2; - float32x4_t SUMM; - float32x2_t SUMM_2[2]; - const float32_t *xi = x; - const float32_t *yi = y; - - SUMM = vdupq_n_f32(0); - - /* Work on 16 values per iteration */ - while (len >= 16) { - XX[0] = vld1q_f32(xi); - xi += 4; - XX[1] = vld1q_f32(xi); - xi += 4; - XX[2] = vld1q_f32(xi); - xi += 4; - XX[3] = vld1q_f32(xi); - xi += 4; - - YY[0] = vld1q_f32(yi); - yi += 4; - YY[1] = vld1q_f32(yi); - yi += 4; - YY[2] = vld1q_f32(yi); - yi += 4; - YY[3] = vld1q_f32(yi); - yi += 4; - - SUMM = vmlaq_f32(SUMM, YY[0], XX[0]); - SUMM = vmlaq_f32(SUMM, YY[1], XX[1]); - SUMM = vmlaq_f32(SUMM, YY[2], XX[2]); - SUMM = vmlaq_f32(SUMM, YY[3], XX[3]); - len -= 16; - } - - /* Work on 8 values */ - if (len >= 8) { - XX[0] = vld1q_f32(xi); - xi += 4; - XX[1] = vld1q_f32(xi); - xi += 4; - - YY[0] = vld1q_f32(yi); - yi += 4; - YY[1] = vld1q_f32(yi); - yi += 4; - - SUMM = vmlaq_f32(SUMM, YY[0], XX[0]); - SUMM = vmlaq_f32(SUMM, YY[1], XX[1]); - len -= 8; - } - - /* Work on 4 values */ - if (len >= 4) { - XX[0] = vld1q_f32(xi); - xi += 4; - YY[0] = vld1q_f32(yi); - yi += 4; - SUMM = vmlaq_f32(SUMM, YY[0], XX[0]); - len -= 4; - } - - /* Start accumulating results */ - SUMM_2[0] = vget_low_f32(SUMM); - if (len >= 2) { - /* While at it, consume 2 more values if available */ - XX_2 = vld1_f32(xi); - xi += 2; - YY_2 = vld1_f32(yi); - yi += 2; - SUMM_2[0] = vmla_f32(SUMM_2[0], YY_2, XX_2); - len -= 2; - } - SUMM_2[1] = vget_high_f32(SUMM); - SUMM_2[0] = vadd_f32(SUMM_2[0], SUMM_2[1]); - SUMM_2[0] = vpadd_f32(SUMM_2[0], SUMM_2[0]); - /* Ok, now we have result accumulated in SUMM_2[0].0 */ - - if (len > 0) { - /* Case when you have one value left */ - XX_2 = vld1_dup_f32(xi); - YY_2 = vld1_dup_f32(yi); - SUMM_2[0] = vmla_f32(SUMM_2[0], XX_2, YY_2); - } - - vst1_lane_f32(sum, SUMM_2[0], 0); -} - void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch) { + opus_val32 *xcorr, int len, int max_pitch, int arch) { int i; + (void)arch; celt_assert(max_pitch > 0); - celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0); + celt_sig_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0); for (i = 0; i < (max_pitch-3); i += 4) { xcorr_kernel_neon_float((const float32_t *)_x, (const float32_t *)_y+i, (float32_t *)xcorr+i, len); } - /* In case max_pitch isn't multiple of 4 - * compute single correlation value per iteration - */ + /* In case max_pitch isn't a multiple of 4, do non-unrolled version. */ for (; i < max_pitch; i++) { - xcorr_kernel_neon_float_process1((const float32_t *)_x, - (const float32_t *)_y+i, (float32_t *)xcorr+i, len); + xcorr[i] = celt_inner_prod_neon(_x, _y+i, len); } } #endif diff --git a/libs/libopus/celt/arm/celt_pitch_xcorr_arm.s b/libs/libopus/celt/arm/celt_pitch_xcorr_arm.s index f96e0a88b..6e873afc3 100644 --- a/libs/libopus/celt/arm/celt_pitch_xcorr_arm.s +++ b/libs/libopus/celt/arm/celt_pitch_xcorr_arm.s @@ -153,7 +153,7 @@ xcorr_kernel_neon_process1 ENDP ; opus_val32 celt_pitch_xcorr_neon(opus_val16 *_x, opus_val16 *_y, -; opus_val32 *xcorr, int len, int max_pitch) +; opus_val32 *xcorr, int len, int max_pitch, int arch) celt_pitch_xcorr_neon PROC ; input: ; r0 = opus_val16 *_x @@ -168,6 +168,8 @@ celt_pitch_xcorr_neon PROC ; r6 = int max_pitch ; r12 = int j ; q15 = int maxcorr[4] (q15 is not used by xcorr_kernel_neon()) + ; ignored: + ; int arch STMFD sp!, {r4-r6, lr} LDR r6, [sp, #16] VMOV.S32 q15, #1 @@ -358,6 +360,8 @@ celt_pitch_xcorr_edsp PROC ; r9 = opus_val32 sum3 ; r1 = int max_pitch ; r12 = int j + ; ignored: + ; int arch STMFD sp!, {r4-r11, lr} MOV r5, r1 LDR r1, [sp, #36] diff --git a/libs/libopus/celt/arm/fft_arm.h b/libs/libopus/celt/arm/fft_arm.h index 0cb55d8e2..0b78175f3 100644 --- a/libs/libopus/celt/arm/fft_arm.h +++ b/libs/libopus/celt/arm/fft_arm.h @@ -34,7 +34,6 @@ #if !defined(FFT_ARM_H) #define FFT_ARM_H -#include "config.h" #include "kiss_fft.h" #if defined(HAVE_ARM_NE10) diff --git a/libs/libopus/celt/arm/fixed_armv4.h b/libs/libopus/celt/arm/fixed_armv4.h index efb3b1896..d84888a77 100644 --- a/libs/libopus/celt/arm/fixed_armv4.h +++ b/libs/libopus/celt/arm/fixed_armv4.h @@ -37,7 +37,7 @@ static OPUS_INLINE opus_val32 MULT16_32_Q16_armv4(opus_val16 a, opus_val32 b) "#MULT16_32_Q16\n\t" "smull %0, %1, %2, %3\n\t" : "=&r"(rd_lo), "=&r"(rd_hi) - : "%r"(b),"r"(a<<16) + : "%r"(b),"r"(SHL32(a,16)) ); return rd_hi; } @@ -54,10 +54,10 @@ static OPUS_INLINE opus_val32 MULT16_32_Q15_armv4(opus_val16 a, opus_val32 b) "#MULT16_32_Q15\n\t" "smull %0, %1, %2, %3\n\t" : "=&r"(rd_lo), "=&r"(rd_hi) - : "%r"(b), "r"(a<<16) + : "%r"(b), "r"(SHL32(a,16)) ); /*We intentionally don't OR in the high bit of rd_lo for speed.*/ - return rd_hi<<1; + return SHL32(rd_hi,1); } #define MULT16_32_Q15(a, b) (MULT16_32_Q15_armv4(a, b)) diff --git a/libs/libopus/celt/arm/fixed_armv5e.h b/libs/libopus/celt/arm/fixed_armv5e.h index 36a632110..6bf73cbac 100644 --- a/libs/libopus/celt/arm/fixed_armv5e.h +++ b/libs/libopus/celt/arm/fixed_armv5e.h @@ -59,7 +59,7 @@ static OPUS_INLINE opus_val32 MULT16_32_Q15_armv5e(opus_val16 a, opus_val32 b) : "=r"(res) : "r"(b), "r"(a) ); - return res<<1; + return SHL32(res,1); } #define MULT16_32_Q15(a, b) (MULT16_32_Q15_armv5e(a, b)) @@ -76,7 +76,7 @@ static OPUS_INLINE opus_val32 MAC16_32_Q15_armv5e(opus_val32 c, opus_val16 a, "#MAC16_32_Q15\n\t" "smlawb %0, %1, %2, %3;\n" : "=r"(res) - : "r"(b<<1), "r"(a), "r"(c) + : "r"(SHL32(b,1)), "r"(a), "r"(c) ); return res; } diff --git a/libs/libopus/celt/arm/mdct_arm.h b/libs/libopus/celt/arm/mdct_arm.h index 49cbb4457..14200bac4 100644 --- a/libs/libopus/celt/arm/mdct_arm.h +++ b/libs/libopus/celt/arm/mdct_arm.h @@ -33,7 +33,6 @@ #if !defined(MDCT_ARM_H) #define MDCT_ARM_H -#include "config.h" #include "mdct.h" #if defined(HAVE_ARM_NE10) diff --git a/libs/libopus/celt/arm/pitch_arm.h b/libs/libopus/celt/arm/pitch_arm.h index 14331169e..bed8b04ea 100644 --- a/libs/libopus/celt/arm/pitch_arm.h +++ b/libs/libopus/celt/arm/pitch_arm.h @@ -30,11 +30,47 @@ # include "armcpu.h" +# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) +opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N); +void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, + const opus_val16 *y02, int N, opus_val32 *xy1, opus_val32 *xy2); + +# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON) +# define OVERRIDE_CELT_INNER_PROD (1) +# define OVERRIDE_DUAL_INNER_PROD (1) +# define celt_inner_prod(x, y, N, arch) ((void)(arch), PRESUME_NEON(celt_inner_prod)(x, y, N)) +# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((void)(arch), PRESUME_NEON(dual_inner_prod)(x, y01, y02, N, xy1, xy2)) +# endif +# endif + +# if !defined(OVERRIDE_CELT_INNER_PROD) +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, const opus_val16 *y, int N); +# define OVERRIDE_CELT_INNER_PROD (1) +# define celt_inner_prod(x, y, N, arch) ((*CELT_INNER_PROD_IMPL[(arch)&OPUS_ARCHMASK])(x, y, N)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_CELT_INNER_PROD (1) +# define celt_inner_prod(x, y, N, arch) ((void)(arch), celt_inner_prod_neon(x, y, N)) +# endif +# endif + +# if !defined(OVERRIDE_DUAL_INNER_PROD) +# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) +extern void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *x, + const opus_val16 *y01, const opus_val16 *y02, int N, opus_val32 *xy1, opus_val32 *xy2); +# define OVERRIDE_DUAL_INNER_PROD (1) +# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((*DUAL_INNER_PROD_IMPL[(arch)&OPUS_ARCHMASK])(x, y01, y02, N, xy1, xy2)) +# elif defined(OPUS_ARM_PRESUME_NEON_INTR) +# define OVERRIDE_DUAL_INNER_PROD (1) +# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) ((void)(arch), dual_inner_prod_neon(x, y01, y02, N, xy1, xy2)) +# endif +# endif + # if defined(FIXED_POINT) # if defined(OPUS_ARM_MAY_HAVE_NEON) opus_val32 celt_pitch_xcorr_neon(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); + opus_val32 *xcorr, int len, int max_pitch, int arch); # endif # if defined(OPUS_ARM_MAY_HAVE_MEDIA) @@ -43,7 +79,7 @@ opus_val32 celt_pitch_xcorr_neon(const opus_val16 *_x, const opus_val16 *_y, # if defined(OPUS_ARM_MAY_HAVE_EDSP) opus_val32 celt_pitch_xcorr_edsp(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); + opus_val32 *xcorr, int len, int max_pitch, int arch); # endif # if defined(OPUS_HAVE_RTCD) && \ @@ -52,18 +88,17 @@ opus_val32 celt_pitch_xcorr_edsp(const opus_val16 *_x, const opus_val16 *_y, (defined(OPUS_ARM_MAY_HAVE_EDSP) && !defined(OPUS_ARM_PRESUME_EDSP))) extern opus_val32 (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int, int); + const opus_val16 *, opus_val32 *, int, int, int); # define OVERRIDE_PITCH_XCORR (1) # define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ ((*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])(_x, _y, \ - xcorr, len, max_pitch)) + xcorr, len, max_pitch, arch)) # elif defined(OPUS_ARM_PRESUME_EDSP) || \ defined(OPUS_ARM_PRESUME_MEDIA) || \ defined(OPUS_ARM_PRESUME_NEON) # define OVERRIDE_PITCH_XCORR (1) -# define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ - ((void)(arch),PRESUME_NEON(celt_pitch_xcorr)(_x, _y, xcorr, len, max_pitch)) +# define celt_pitch_xcorr (PRESUME_NEON(celt_pitch_xcorr)) # endif @@ -99,25 +134,24 @@ extern void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])( /* Float case */ #if defined(OPUS_ARM_MAY_HAVE_NEON_INTR) void celt_pitch_xcorr_float_neon(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); + opus_val32 *xcorr, int len, int max_pitch, int arch); #endif # if defined(OPUS_HAVE_RTCD) && \ (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)) extern void (*const CELT_PITCH_XCORR_IMPL[OPUS_ARCHMASK+1])(const opus_val16 *, - const opus_val16 *, opus_val32 *, int, int); + const opus_val16 *, opus_val32 *, int, int, int); # define OVERRIDE_PITCH_XCORR (1) # define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ ((*CELT_PITCH_XCORR_IMPL[(arch)&OPUS_ARCHMASK])(_x, _y, \ - xcorr, len, max_pitch)) + xcorr, len, max_pitch, arch)) # elif defined(OPUS_ARM_PRESUME_NEON_INTR) # define OVERRIDE_PITCH_XCORR (1) -# define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \ - ((void)(arch),celt_pitch_xcorr_float_neon(_x, _y, xcorr, len, max_pitch)) +# define celt_pitch_xcorr celt_pitch_xcorr_float_neon # endif diff --git a/libs/libopus/celt/arm/pitch_neon_intr.c b/libs/libopus/celt/arm/pitch_neon_intr.c new file mode 100644 index 000000000..1ac38c433 --- /dev/null +++ b/libs/libopus/celt/arm/pitch_neon_intr.c @@ -0,0 +1,290 @@ +/*********************************************************************** +Copyright (c) 2017 Google Inc. +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions +are met: +- Redistributions of source code must retain the above copyright notice, +this list of conditions and the following disclaimer. +- Redistributions in binary form must reproduce the above copyright +notice, this list of conditions and the following disclaimer in the +documentation and/or other materials provided with the distribution. +- Neither the name of Internet Society, IETF or IETF Trust, nor the +names of specific contributors, may be used to endorse or promote +products derived from this software without specific prior written +permission. +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE +POSSIBILITY OF SUCH DAMAGE. +***********************************************************************/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include <arm_neon.h> +#include "pitch.h" + +#ifdef FIXED_POINT + +opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N) +{ + int i; + opus_val32 xy; + int16x8_t x_s16x8, y_s16x8; + int32x4_t xy_s32x4 = vdupq_n_s32(0); + int64x2_t xy_s64x2; + int64x1_t xy_s64x1; + + for (i = 0; i < N - 7; i += 8) { + x_s16x8 = vld1q_s16(&x[i]); + y_s16x8 = vld1q_s16(&y[i]); + xy_s32x4 = vmlal_s16(xy_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y_s16x8)); + xy_s32x4 = vmlal_s16(xy_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y_s16x8)); + } + + if (N - i >= 4) { + const int16x4_t x_s16x4 = vld1_s16(&x[i]); + const int16x4_t y_s16x4 = vld1_s16(&y[i]); + xy_s32x4 = vmlal_s16(xy_s32x4, x_s16x4, y_s16x4); + i += 4; + } + + xy_s64x2 = vpaddlq_s32(xy_s32x4); + xy_s64x1 = vadd_s64(vget_low_s64(xy_s64x2), vget_high_s64(xy_s64x2)); + xy = vget_lane_s32(vreinterpret_s32_s64(xy_s64x1), 0); + + for (; i < N; i++) { + xy = MAC16_16(xy, x[i], y[i]); + } + +#ifdef OPUS_CHECK_ASM + celt_assert(celt_inner_prod_c(x, y, N) == xy); +#endif + + return xy; +} + +void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) +{ + int i; + opus_val32 xy01, xy02; + int16x8_t x_s16x8, y01_s16x8, y02_s16x8; + int32x4_t xy01_s32x4 = vdupq_n_s32(0); + int32x4_t xy02_s32x4 = vdupq_n_s32(0); + int64x2_t xy01_s64x2, xy02_s64x2; + int64x1_t xy01_s64x1, xy02_s64x1; + + for (i = 0; i < N - 7; i += 8) { + x_s16x8 = vld1q_s16(&x[i]); + y01_s16x8 = vld1q_s16(&y01[i]); + y02_s16x8 = vld1q_s16(&y02[i]); + xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y01_s16x8)); + xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_low_s16 (x_s16x8), vget_low_s16 (y02_s16x8)); + xy01_s32x4 = vmlal_s16(xy01_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y01_s16x8)); + xy02_s32x4 = vmlal_s16(xy02_s32x4, vget_high_s16(x_s16x8), vget_high_s16(y02_s16x8)); + } + + if (N - i >= 4) { + const int16x4_t x_s16x4 = vld1_s16(&x[i]); + const int16x4_t y01_s16x4 = vld1_s16(&y01[i]); + const int16x4_t y02_s16x4 = vld1_s16(&y02[i]); + xy01_s32x4 = vmlal_s16(xy01_s32x4, x_s16x4, y01_s16x4); + xy02_s32x4 = vmlal_s16(xy02_s32x4, x_s16x4, y02_s16x4); + i += 4; + } + + xy01_s64x2 = vpaddlq_s32(xy01_s32x4); + xy02_s64x2 = vpaddlq_s32(xy02_s32x4); + xy01_s64x1 = vadd_s64(vget_low_s64(xy01_s64x2), vget_high_s64(xy01_s64x2)); + xy02_s64x1 = vadd_s64(vget_low_s64(xy02_s64x2), vget_high_s64(xy02_s64x2)); + xy01 = vget_lane_s32(vreinterpret_s32_s64(xy01_s64x1), 0); + xy02 = vget_lane_s32(vreinterpret_s32_s64(xy02_s64x1), 0); + + for (; i < N; i++) { + xy01 = MAC16_16(xy01, x[i], y01[i]); + xy02 = MAC16_16(xy02, x[i], y02[i]); + } + *xy1 = xy01; + *xy2 = xy02; + +#ifdef OPUS_CHECK_ASM + { + opus_val32 xy1_c, xy2_c; + dual_inner_prod_c(x, y01, y02, N, &xy1_c, &xy2_c); + celt_assert(xy1_c == *xy1); + celt_assert(xy2_c == *xy2); + } +#endif +} + +#else /* !FIXED_POINT */ + +/* ========================================================================== */ + +#ifdef OPUS_CHECK_ASM + +/* This part of code simulates floating-point NEON operations. */ + +/* celt_inner_prod_neon_float_c_simulation() simulates the floating-point */ +/* operations of celt_inner_prod_neon(), and both functions should have bit */ +/* exact output. */ +static opus_val32 celt_inner_prod_neon_float_c_simulation(const opus_val16 *x, const opus_val16 *y, int N) +{ + int i; + opus_val32 xy, xy0 = 0, xy1 = 0, xy2 = 0, xy3 = 0; + for (i = 0; i < N - 3; i += 4) { + xy0 = MAC16_16(xy0, x[i + 0], y[i + 0]); + xy1 = MAC16_16(xy1, x[i + 1], y[i + 1]); + xy2 = MAC16_16(xy2, x[i + 2], y[i + 2]); + xy3 = MAC16_16(xy3, x[i + 3], y[i + 3]); + } + xy0 += xy2; + xy1 += xy3; + xy = xy0 + xy1; + for (; i < N; i++) { + xy = MAC16_16(xy, x[i], y[i]); + } + return xy; +} + +/* dual_inner_prod_neon_float_c_simulation() simulates the floating-point */ +/* operations of dual_inner_prod_neon(), and both functions should have bit */ +/* exact output. */ +static void dual_inner_prod_neon_float_c_simulation(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) +{ + int i; + opus_val32 xy01, xy02, xy01_0 = 0, xy01_1 = 0, xy01_2 = 0, xy01_3 = 0, xy02_0 = 0, xy02_1 = 0, xy02_2 = 0, xy02_3 = 0; + for (i = 0; i < N - 3; i += 4) { + xy01_0 = MAC16_16(xy01_0, x[i + 0], y01[i + 0]); + xy01_1 = MAC16_16(xy01_1, x[i + 1], y01[i + 1]); + xy01_2 = MAC16_16(xy01_2, x[i + 2], y01[i + 2]); + xy01_3 = MAC16_16(xy01_3, x[i + 3], y01[i + 3]); + xy02_0 = MAC16_16(xy02_0, x[i + 0], y02[i + 0]); + xy02_1 = MAC16_16(xy02_1, x[i + 1], y02[i + 1]); + xy02_2 = MAC16_16(xy02_2, x[i + 2], y02[i + 2]); + xy02_3 = MAC16_16(xy02_3, x[i + 3], y02[i + 3]); + } + xy01_0 += xy01_2; + xy02_0 += xy02_2; + xy01_1 += xy01_3; + xy02_1 += xy02_3; + xy01 = xy01_0 + xy01_1; + xy02 = xy02_0 + xy02_1; + for (; i < N; i++) { + xy01 = MAC16_16(xy01, x[i], y01[i]); + xy02 = MAC16_16(xy02, x[i], y02[i]); + } + *xy1 = xy01; + *xy2 = xy02; +} + +#endif /* OPUS_CHECK_ASM */ + +/* ========================================================================== */ + +opus_val32 celt_inner_prod_neon(const opus_val16 *x, const opus_val16 *y, int N) +{ + int i; + opus_val32 xy; + float32x4_t xy_f32x4 = vdupq_n_f32(0); + float32x2_t xy_f32x2; + + for (i = 0; i < N - 7; i += 8) { + float32x4_t x_f32x4, y_f32x4; + x_f32x4 = vld1q_f32(&x[i]); + y_f32x4 = vld1q_f32(&y[i]); + xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4); + x_f32x4 = vld1q_f32(&x[i + 4]); + y_f32x4 = vld1q_f32(&y[i + 4]); + xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4); + } + + if (N - i >= 4) { + const float32x4_t x_f32x4 = vld1q_f32(&x[i]); + const float32x4_t y_f32x4 = vld1q_f32(&y[i]); + xy_f32x4 = vmlaq_f32(xy_f32x4, x_f32x4, y_f32x4); + i += 4; + } + + xy_f32x2 = vadd_f32(vget_low_f32(xy_f32x4), vget_high_f32(xy_f32x4)); + xy_f32x2 = vpadd_f32(xy_f32x2, xy_f32x2); + xy = vget_lane_f32(xy_f32x2, 0); + + for (; i < N; i++) { + xy = MAC16_16(xy, x[i], y[i]); + } + +#ifdef OPUS_CHECK_ASM + celt_assert(ABS32(celt_inner_prod_neon_float_c_simulation(x, y, N) - xy) <= VERY_SMALL); +#endif + + return xy; +} + +void dual_inner_prod_neon(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02, + int N, opus_val32 *xy1, opus_val32 *xy2) +{ + int i; + opus_val32 xy01, xy02; + float32x4_t xy01_f32x4 = vdupq_n_f32(0); + float32x4_t xy02_f32x4 = vdupq_n_f32(0); + float32x2_t xy01_f32x2, xy02_f32x2; + + for (i = 0; i < N - 7; i += 8) { + float32x4_t x_f32x4, y01_f32x4, y02_f32x4; + x_f32x4 = vld1q_f32(&x[i]); + y01_f32x4 = vld1q_f32(&y01[i]); + y02_f32x4 = vld1q_f32(&y02[i]); + xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4); + xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4); + x_f32x4 = vld1q_f32(&x[i + 4]); + y01_f32x4 = vld1q_f32(&y01[i + 4]); + y02_f32x4 = vld1q_f32(&y02[i + 4]); + xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4); + xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4); + } + + if (N - i >= 4) { + const float32x4_t x_f32x4 = vld1q_f32(&x[i]); + const float32x4_t y01_f32x4 = vld1q_f32(&y01[i]); + const float32x4_t y02_f32x4 = vld1q_f32(&y02[i]); + xy01_f32x4 = vmlaq_f32(xy01_f32x4, x_f32x4, y01_f32x4); + xy02_f32x4 = vmlaq_f32(xy02_f32x4, x_f32x4, y02_f32x4); + i += 4; + } + + xy01_f32x2 = vadd_f32(vget_low_f32(xy01_f32x4), vget_high_f32(xy01_f32x4)); + xy02_f32x2 = vadd_f32(vget_low_f32(xy02_f32x4), vget_high_f32(xy02_f32x4)); + xy01_f32x2 = vpadd_f32(xy01_f32x2, xy01_f32x2); + xy02_f32x2 = vpadd_f32(xy02_f32x2, xy02_f32x2); + xy01 = vget_lane_f32(xy01_f32x2, 0); + xy02 = vget_lane_f32(xy02_f32x2, 0); + + for (; i < N; i++) { + xy01 = MAC16_16(xy01, x[i], y01[i]); + xy02 = MAC16_16(xy02, x[i], y02[i]); + } + *xy1 = xy01; + *xy2 = xy02; + +#ifdef OPUS_CHECK_ASM + { + opus_val32 xy1_c, xy2_c; + dual_inner_prod_neon_float_c_simulation(x, y01, y02, N, &xy1_c, &xy2_c); + celt_assert(ABS32(xy1_c - *xy1) <= VERY_SMALL); + celt_assert(ABS32(xy2_c - *xy2) <= VERY_SMALL); + } +#endif +} + +#endif /* FIXED_POINT */ diff --git a/libs/libopus/celt/bands.c b/libs/libopus/celt/bands.c index 87eaa6c03..bd54036af 100644 --- a/libs/libopus/celt/bands.c +++ b/libs/libopus/celt/bands.c @@ -65,19 +65,19 @@ opus_uint32 celt_lcg_rand(opus_uint32 seed) /* This is a cos() approximation designed to be bit-exact on any platform. Bit exactness with this approximation is important because it has an impact on the bit allocation */ -static opus_int16 bitexact_cos(opus_int16 x) +opus_int16 bitexact_cos(opus_int16 x) { opus_int32 tmp; opus_int16 x2; tmp = (4096+((opus_int32)(x)*(x)))>>13; - celt_assert(tmp<=32767); + celt_sig_assert(tmp<=32767); x2 = tmp; x2 = (32767-x2) + FRAC_MUL16(x2, (-7651 + FRAC_MUL16(x2, (8277 + FRAC_MUL16(-626, x2))))); - celt_assert(x2<=32766); + celt_sig_assert(x2<=32766); return 1+x2; } -static int bitexact_log2tan(int isin,int icos) +int bitexact_log2tan(int isin,int icos) { int lc; int ls; @@ -92,10 +92,11 @@ static int bitexact_log2tan(int isin,int icos) #ifdef FIXED_POINT /* Compute the amplitude (sqrt energy) in each of the bands */ -void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM) +void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch) { int i, c, N; const opus_int16 *eBands = m->eBands; + (void)arch; N = m->shortMdctSize<<LM; c=0; do { for (i=0;i<end;i++) @@ -155,7 +156,7 @@ void normalise_bands(const CELTMode *m, const celt_sig * OPUS_RESTRICT freq, cel #else /* FIXED_POINT */ /* Compute the amplitude (sqrt energy) in each of the bands */ -void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM) +void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch) { int i, c, N; const opus_int16 *eBands = m->eBands; @@ -164,7 +165,7 @@ void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *band for (i=0;i<end;i++) { opus_val32 sum; - sum = 1e-27f + celt_inner_prod_c(&X[c*N+(eBands[i]<<LM)], &X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM); + sum = 1e-27f + celt_inner_prod(&X[c*N+(eBands[i]<<LM)], &X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM, arch); bandE[i+c*m->nbEBands] = celt_sqrt(sum); /*printf ("%f ", bandE[i+c*m->nbEBands]);*/ } @@ -224,9 +225,9 @@ void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X, #endif j=M*eBands[i]; band_end = M*eBands[i+1]; - lg = ADD16(bandLogE[i], SHL16((opus_val16)eMeans[i],6)); + lg = SATURATE16(ADD32(bandLogE[i], SHL32((opus_val32)eMeans[i],6))); #ifndef FIXED_POINT - g = celt_exp2(lg); + g = celt_exp2(MIN32(32.f, lg)); #else /* Handle the integer part of the log energy */ shift = 16-(lg>>DB_SHIFT); @@ -241,12 +242,12 @@ void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X, /* Handle extreme gains with negative shift. */ if (shift<0) { - /* For shift < -2 we'd be likely to overflow, so we're capping - the gain here. This shouldn't happen unless the bitstream is - already corrupted. */ - if (shift < -2) + /* For shift <= -2 and g > 16384 we'd be likely to overflow, so we're + capping the gain here, which is equivalent to a cap of 18 on lg. + This shouldn't trigger unless the bitstream is already corrupted. */ + if (shift <= -2) { - g = 32767; + g = 16384; shift = -2; } do { @@ -281,7 +282,7 @@ void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_mas N0 = m->eBands[i+1]-m->eBands[i]; /* depth in 1/8 bits */ - celt_assert(pulses[i]>=0); + celt_sig_assert(pulses[i]>=0); depth = celt_udiv(1+pulses[i], (m->eBands[i+1]-m->eBands[i]))>>LM; #ifdef FIXED_POINT @@ -360,6 +361,30 @@ void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_mas } } +/* Compute the weights to use for optimizing normalized distortion across + channels. We use the amplitude to weight square distortion, which means + that we use the square root of the value we would have been using if we + wanted to minimize the MSE in the non-normalized domain. This roughly + corresponds to some quick-and-dirty perceptual experiments I ran to + measure inter-aural masking (there doesn't seem to be any published data + on the topic). */ +static void compute_channel_weights(celt_ener Ex, celt_ener Ey, opus_val16 w[2]) +{ + celt_ener minE; +#ifdef FIXED_POINT + int shift; +#endif + minE = MIN32(Ex, Ey); + /* Adjustment to make the weights a bit more conservative. */ + Ex = ADD32(Ex, minE/3); + Ey = ADD32(Ey, minE/3); +#ifdef FIXED_POINT + shift = celt_ilog2(EPSILON+MAX32(Ex, Ey))-14; +#endif + w[0] = VSHR32(Ex, shift); + w[1] = VSHR32(Ey, shift); +} + static void intensity_stereo(const CELTMode *m, celt_norm * OPUS_RESTRICT X, const celt_norm * OPUS_RESTRICT Y, const celt_ener *bandE, int bandID, int N) { int i = bandID; @@ -453,7 +478,7 @@ static void stereo_merge(celt_norm * OPUS_RESTRICT X, celt_norm * OPUS_RESTRICT /* Decide whether we should spread the pulses in the current frame */ int spreading_decision(const CELTMode *m, const celt_norm *X, int *average, int last_decision, int *hf_average, int *tapset_decision, int update_hf, - int end, int C, int M) + int end, int C, int M, const int *spread_weight) { int i, c, N0; int sum = 0, nbBands=0; @@ -494,8 +519,8 @@ int spreading_decision(const CELTMode *m, const celt_norm *X, int *average, if (i>m->nbEBands-4) hf_sum += celt_udiv(32*(tcount[1]+tcount[0]), N); tmp = (2*tcount[2] >= N) + (2*tcount[1] >= N) + (2*tcount[0] >= N); - sum += tmp*256; - nbBands++; + sum += tmp*spread_weight[i]; + nbBands+=spread_weight[i]; } } while (++c<C); @@ -519,7 +544,7 @@ int spreading_decision(const CELTMode *m, const celt_norm *X, int *average, /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/ celt_assert(nbBands>0); /* end has to be non-zero */ celt_assert(sum>=0); - sum = celt_udiv(sum, nbBands); + sum = celt_udiv((opus_int32)sum<<8, nbBands); /* Recursive averaging */ sum = (sum+*average)>>1; *average = sum; @@ -647,6 +672,7 @@ static int compute_qn(int N, int b, int offset, int pulse_cap, int stereo) struct band_ctx { int encode; + int resynth; const CELTMode *m; int i; int intensity; @@ -657,6 +683,9 @@ struct band_ctx { const celt_ener *bandE; opus_uint32 seed; int arch; + int theta_round; + int disable_inv; + int avoid_split_noise; }; struct split_ctx { @@ -714,8 +743,35 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, if (qn!=1) { if (encode) - itheta = (itheta*(opus_int32)qn+8192)>>14; - + { + if (!stereo || ctx->theta_round == 0) + { + itheta = (itheta*(opus_int32)qn+8192)>>14; + if (!stereo && ctx->avoid_split_noise && itheta > 0 && itheta < qn) + { + /* Check if the selected value of theta will cause the bit allocation + to inject noise on one side. If so, make sure the energy of that side + is zero. */ + int unquantized = celt_udiv((opus_int32)itheta*16384, qn); + imid = bitexact_cos((opus_int16)unquantized); + iside = bitexact_cos((opus_int16)(16384-unquantized)); + delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid)); + if (delta > *b) + itheta = qn; + else if (delta < -*b) + itheta = 0; + } + } else { + int down; + /* Bias quantization towards itheta=0 and itheta=16384. */ + int bias = itheta > 8192 ? 32767/qn : -32767/qn; + down = IMIN(qn-1, IMAX(0, (itheta*(opus_int32)qn + bias)>>14)); + if (ctx->theta_round < 0) + itheta = down; + else + itheta = down+1; + } + } /* Entropy coding of the angle. We use a uniform pdf for the time split, a step for stereo, and a triangular one for the rest. */ if (stereo && N>2) @@ -793,7 +849,7 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, } else if (stereo) { if (encode) { - inv = itheta > 8192; + inv = itheta > 8192 && !ctx->disable_inv; if (inv) { int j; @@ -810,6 +866,9 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, inv = ec_dec_bit_logp(ec, 2); } else inv = 0; + /* inv flag override to avoid problems with downmixing. */ + if (ctx->disable_inv) + inv = 0; itheta = 0; } qalloc = ec_tell_frac(ec) - tell; @@ -842,14 +901,9 @@ static void compute_theta(struct band_ctx *ctx, struct split_ctx *sctx, sctx->itheta = itheta; sctx->qalloc = qalloc; } -static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, int b, +static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, celt_norm *lowband_out) { -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif int c; int stereo; celt_norm *x = X; @@ -872,9 +926,8 @@ static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, sign = ec_dec_bits(ec, 1); } ctx->remaining_bits -= 1<<BITRES; - b-=1<<BITRES; } - if (resynth) + if (ctx->resynth) x[0] = sign ? -NORM_SCALING : NORM_SCALING; x = Y; } while (++c<1+stereo); @@ -899,11 +952,6 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X, int B0=B; opus_val16 mid=0, side=0; unsigned cm=0; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif celt_norm *Y=NULL; int encode; const CELTMode *m; @@ -935,8 +983,7 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X, fill = (fill&1)|(fill<<1); B = (B+1)>>1; - compute_theta(ctx, &sctx, X, Y, N, &b, B, B0, - LM, 0, &fill); + compute_theta(ctx, &sctx, X, Y, N, &b, B, B0, LM, 0, &fill); imid = sctx.imid; iside = sctx.iside; delta = sctx.delta; @@ -970,24 +1017,20 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X, rebalance = ctx->remaining_bits; if (mbits >= sbits) { - cm = quant_partition(ctx, X, N, mbits, B, - lowband, LM, + cm = quant_partition(ctx, X, N, mbits, B, lowband, LM, MULT16_16_P15(gain,mid), fill); rebalance = mbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<<BITRES && itheta!=0) sbits += rebalance - (3<<BITRES); - cm |= quant_partition(ctx, Y, N, sbits, B, - next_lowband2, LM, + cm |= quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM, MULT16_16_P15(gain,side), fill>>B)<<(B0>>1); } else { - cm = quant_partition(ctx, Y, N, sbits, B, - next_lowband2, LM, + cm = quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM, MULT16_16_P15(gain,side), fill>>B)<<(B0>>1); rebalance = sbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<<BITRES && itheta!=16384) mbits += rebalance - (3<<BITRES); - cm |= quant_partition(ctx, X, N, mbits, B, - lowband, LM, + cm |= quant_partition(ctx, X, N, mbits, B, lowband, LM, MULT16_16_P15(gain,mid), fill); } } else { @@ -1012,18 +1055,14 @@ static unsigned quant_partition(struct band_ctx *ctx, celt_norm *X, /* Finally do the actual quantization */ if (encode) { - cm = alg_quant(X, N, K, spread, B, ec -#ifdef RESYNTH - , gain -#endif - ); + cm = alg_quant(X, N, K, spread, B, ec, gain, ctx->resynth, ctx->arch); } else { cm = alg_unquant(X, N, K, spread, B, ec, gain); } } else { /* If there's no pulse, fill the band anyway */ int j; - if (resynth) + if (ctx->resynth) { unsigned cm_mask; /* B can be as large as 16, so this shift might overflow an int on a @@ -1080,11 +1119,6 @@ static unsigned quant_band(struct band_ctx *ctx, celt_norm *X, int recombine=0; int longBlocks; unsigned cm=0; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif int k; int encode; int tf_change; @@ -1099,7 +1133,7 @@ static unsigned quant_band(struct band_ctx *ctx, celt_norm *X, /* Special case for one sample */ if (N==1) { - return quant_band_n1(ctx, X, NULL, b, lowband_out); + return quant_band_n1(ctx, X, NULL, lowband_out); } if (tf_change>0) @@ -1151,11 +1185,10 @@ static unsigned quant_band(struct band_ctx *ctx, celt_norm *X, deinterleave_hadamard(lowband, N_B>>recombine, B0<<recombine, longBlocks); } - cm = quant_partition(ctx, X, N, b, B, lowband, - LM, gain, fill); + cm = quant_partition(ctx, X, N, b, B, lowband, LM, gain, fill); /* This code is used by the decoder and by the resynthesis-enabled encoder */ - if (resynth) + if (ctx->resynth) { /* Undo the sample reorganization going from time order to frequency order */ if (B0>1) @@ -1208,11 +1241,6 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm int inv = 0; opus_val16 mid=0, side=0; unsigned cm=0; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !ctx->encode; -#endif int mbits, sbits, delta; int itheta; int qalloc; @@ -1227,13 +1255,12 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm /* Special case for one sample */ if (N==1) { - return quant_band_n1(ctx, X, Y, b, lowband_out); + return quant_band_n1(ctx, X, Y, lowband_out); } orig_fill = fill; - compute_theta(ctx, &sctx, X, Y, N, &b, B, B, - LM, 1, &fill); + compute_theta(ctx, &sctx, X, Y, N, &b, B, B, LM, 1, &fill); inv = sctx.inv; imid = sctx.imid; iside = sctx.iside; @@ -1281,13 +1308,13 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm sign = 1-2*sign; /* We use orig_fill here because we want to fold the side, but if itheta==16384, we'll have cleared the low bits of fill. */ - cm = quant_band(ctx, x2, N, mbits, B, lowband, - LM, lowband_out, Q15ONE, lowband_scratch, orig_fill); + cm = quant_band(ctx, x2, N, mbits, B, lowband, LM, lowband_out, Q15ONE, + lowband_scratch, orig_fill); /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse), and there's no need to worry about mixing with the other channel. */ y2[0] = -sign*x2[1]; y2[1] = sign*x2[0]; - if (resynth) + if (ctx->resynth) { celt_norm tmp; X[0] = MULT16_16_Q15(mid, X[0]); @@ -1314,38 +1341,32 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm { /* In stereo mode, we do not apply a scaling to the mid because we need the normalized mid for folding later. */ - cm = quant_band(ctx, X, N, mbits, B, - lowband, LM, lowband_out, - Q15ONE, lowband_scratch, fill); + cm = quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE, + lowband_scratch, fill); rebalance = mbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<<BITRES && itheta!=0) sbits += rebalance - (3<<BITRES); /* For a stereo split, the high bits of fill are always zero, so no folding will be done to the side. */ - cm |= quant_band(ctx, Y, N, sbits, B, - NULL, LM, NULL, - side, NULL, fill>>B); + cm |= quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B); } else { /* For a stereo split, the high bits of fill are always zero, so no folding will be done to the side. */ - cm = quant_band(ctx, Y, N, sbits, B, - NULL, LM, NULL, - side, NULL, fill>>B); + cm = quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B); rebalance = sbits - (rebalance-ctx->remaining_bits); if (rebalance > 3<<BITRES && itheta!=16384) mbits += rebalance - (3<<BITRES); /* In stereo mode, we do not apply a scaling to the mid because we need the normalized mid for folding later. */ - cm |= quant_band(ctx, X, N, mbits, B, - lowband, LM, lowband_out, - Q15ONE, lowband_scratch, fill); + cm |= quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE, + lowband_scratch, fill); } } /* This code is used by the decoder and by the resynthesis-enabled encoder */ - if (resynth) + if (ctx->resynth) { if (N!=2) stereo_merge(X, Y, mid, N, ctx->arch); @@ -1359,19 +1380,38 @@ static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm return cm; } +static void special_hybrid_folding(const CELTMode *m, celt_norm *norm, celt_norm *norm2, int start, int M, int dual_stereo) +{ + int n1, n2; + const opus_int16 * OPUS_RESTRICT eBands = m->eBands; + n1 = M*(eBands[start+1]-eBands[start]); + n2 = M*(eBands[start+2]-eBands[start+1]); + /* Duplicate enough of the first band folding data to be able to fold the second band. + Copies no data for CELT-only mode. */ + OPUS_COPY(&norm[n1], &norm[2*n1 - n2], n2-n1); + if (dual_stereo) + OPUS_COPY(&norm2[n1], &norm2[2*n1 - n2], n2-n1); +} void quant_all_bands(int encode, const CELTMode *m, int start, int end, celt_norm *X_, celt_norm *Y_, unsigned char *collapse_masks, const celt_ener *bandE, int *pulses, int shortBlocks, int spread, int dual_stereo, int intensity, int *tf_res, opus_int32 total_bits, opus_int32 balance, ec_ctx *ec, int LM, int codedBands, - opus_uint32 *seed, int arch) + opus_uint32 *seed, int complexity, int arch, int disable_inv) { int i; opus_int32 remaining_bits; const opus_int16 * OPUS_RESTRICT eBands = m->eBands; celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2; VARDECL(celt_norm, _norm); + VARDECL(celt_norm, _lowband_scratch); + VARDECL(celt_norm, X_save); + VARDECL(celt_norm, Y_save); + VARDECL(celt_norm, X_save2); + VARDECL(celt_norm, Y_save2); + VARDECL(celt_norm, norm_save2); + int resynth_alloc; celt_norm *lowband_scratch; int B; int M; @@ -1379,10 +1419,11 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, int update_lowband = 1; int C = Y_ != NULL ? 2 : 1; int norm_offset; + int theta_rdo = encode && Y_!=NULL && !dual_stereo && complexity>=8; #ifdef RESYNTH int resynth = 1; #else - int resynth = !encode; + int resynth = !encode || theta_rdo; #endif struct band_ctx ctx; SAVE_STACK; @@ -1395,9 +1436,24 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, ALLOC(_norm, C*(M*eBands[m->nbEBands-1]-norm_offset), celt_norm); norm = _norm; norm2 = norm + M*eBands[m->nbEBands-1]-norm_offset; - /* We can use the last band as scratch space because we don't need that - scratch space for the last band. */ - lowband_scratch = X_+M*eBands[m->nbEBands-1]; + + /* For decoding, we can use the last band as scratch space because we don't need that + scratch space for the last band and we don't care about the data there until we're + decoding the last band. */ + if (encode && resynth) + resynth_alloc = M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]); + else + resynth_alloc = ALLOC_NONE; + ALLOC(_lowband_scratch, resynth_alloc, celt_norm); + if (encode && resynth) + lowband_scratch = _lowband_scratch; + else + lowband_scratch = X_+M*eBands[m->nbEBands-1]; + ALLOC(X_save, resynth_alloc, celt_norm); + ALLOC(Y_save, resynth_alloc, celt_norm); + ALLOC(X_save2, resynth_alloc, celt_norm); + ALLOC(Y_save2, resynth_alloc, celt_norm); + ALLOC(norm_save2, resynth_alloc, celt_norm); lowband_offset = 0; ctx.bandE = bandE; @@ -1408,6 +1464,11 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, ctx.seed = *seed; ctx.spread = spread; ctx.arch = arch; + ctx.disable_inv = disable_inv; + ctx.resynth = resynth; + ctx.theta_round = 0; + /* Avoid injecting noise in the first band on transients. */ + ctx.avoid_split_noise = B > 1; for (i=start;i<end;i++) { opus_int32 tell; @@ -1430,6 +1491,7 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, else Y = NULL; N = M*eBands[i+1]-M*eBands[i]; + celt_assert(N > 0); tell = ec_tell_frac(ec); /* Compute how many bits we want to allocate to this band */ @@ -1445,8 +1507,15 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, b = 0; } +#ifndef DISABLE_UPDATE_DRAFT + if (resynth && (M*eBands[i]-N >= M*eBands[start] || i==start+1) && (update_lowband || lowband_offset==0)) + lowband_offset = i; + if (i == start+1) + special_hybrid_folding(m, norm, norm2, start, M, dual_stereo); +#else if (resynth && M*eBands[i]-N >= M*eBands[start] && (update_lowband || lowband_offset==0)) lowband_offset = i; +#endif tf_change = tf_res[i]; ctx.tf_change = tf_change; @@ -1457,7 +1526,7 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, Y = norm; lowband_scratch = NULL; } - if (i==end-1) + if (last && !theta_rdo) lowband_scratch = NULL; /* Get a conservative estimate of the collapse_mask's for the bands we're @@ -1472,7 +1541,11 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, fold_start = lowband_offset; while(M*eBands[--fold_start] > effective_lowband+norm_offset); fold_end = lowband_offset-1; +#ifndef DISABLE_UPDATE_DRAFT + while(++fold_end < i && M*eBands[fold_end] < effective_lowband+norm_offset+N); +#else while(M*eBands[++fold_end] < effective_lowband+norm_offset+N); +#endif x_cm = y_cm = 0; fold_i = fold_start; do { x_cm |= collapse_masks[fold_i*C+0]; @@ -1505,13 +1578,79 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, } else { if (Y!=NULL) { - x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, - effective_lowband != -1 ? norm+effective_lowband : NULL, LM, - last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm); + if (theta_rdo && i < intensity) + { + ec_ctx ec_save, ec_save2; + struct band_ctx ctx_save, ctx_save2; + opus_val32 dist0, dist1; + unsigned cm, cm2; + int nstart_bytes, nend_bytes, save_bytes; + unsigned char *bytes_buf; + unsigned char bytes_save[1275]; + opus_val16 w[2]; + compute_channel_weights(bandE[i], bandE[i+m->nbEBands], w); + /* Make a copy. */ + cm = x_cm|y_cm; + ec_save = *ec; + ctx_save = ctx; + OPUS_COPY(X_save, X, N); + OPUS_COPY(Y_save, Y, N); + /* Encode and round down. */ + ctx.theta_round = -1; + x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, + effective_lowband != -1 ? norm+effective_lowband : NULL, LM, + last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm); + dist0 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch)); + + /* Save first result. */ + cm2 = x_cm; + ec_save2 = *ec; + ctx_save2 = ctx; + OPUS_COPY(X_save2, X, N); + OPUS_COPY(Y_save2, Y, N); + if (!last) + OPUS_COPY(norm_save2, norm+M*eBands[i]-norm_offset, N); + nstart_bytes = ec_save.offs; + nend_bytes = ec_save.storage; + bytes_buf = ec_save.buf+nstart_bytes; + save_bytes = nend_bytes-nstart_bytes; + OPUS_COPY(bytes_save, bytes_buf, save_bytes); + + /* Restore */ + *ec = ec_save; + ctx = ctx_save; + OPUS_COPY(X, X_save, N); + OPUS_COPY(Y, Y_save, N); +#ifndef DISABLE_UPDATE_DRAFT + if (i == start+1) + special_hybrid_folding(m, norm, norm2, start, M, dual_stereo); +#endif + /* Encode and round up. */ + ctx.theta_round = 1; + x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, + effective_lowband != -1 ? norm+effective_lowband : NULL, LM, + last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm); + dist1 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch)); + if (dist0 >= dist1) { + x_cm = cm2; + *ec = ec_save2; + ctx = ctx_save2; + OPUS_COPY(X, X_save2, N); + OPUS_COPY(Y, Y_save2, N); + if (!last) + OPUS_COPY(norm+M*eBands[i]-norm_offset, norm_save2, N); + OPUS_COPY(bytes_buf, bytes_save, save_bytes); + } + } else { + ctx.theta_round = 0; + x_cm = quant_band_stereo(&ctx, X, Y, N, b, B, + effective_lowband != -1 ? norm+effective_lowband : NULL, LM, + last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm); + } } else { x_cm = quant_band(&ctx, X, N, b, B, effective_lowband != -1 ? norm+effective_lowband : NULL, LM, - last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm); + last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm); } y_cm = x_cm; } @@ -1521,6 +1660,9 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, /* Update the folding position only as long as we have 1 bit/sample depth. */ update_lowband = b>(N<<BITRES); + /* We only need to avoid noise on a split for the first band. After that, we + have folding. */ + ctx.avoid_split_noise = 0; } *seed = ctx.seed; diff --git a/libs/libopus/celt/bands.h b/libs/libopus/celt/bands.h index e8bef4bad..422b32cf7 100644 --- a/libs/libopus/celt/bands.h +++ b/libs/libopus/celt/bands.h @@ -36,12 +36,15 @@ #include "entdec.h" #include "rate.h" +opus_int16 bitexact_cos(opus_int16 x); +int bitexact_log2tan(int isin,int icos); + /** Compute the amplitude (sqrt energy) in each of the bands * @param m Mode data * @param X Spectrum * @param bandE Square root of the energy for each band (returned) */ -void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM); +void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch); /*void compute_noise_energies(const CELTMode *m, const celt_sig *X, const opus_val16 *tonality, celt_ener *bandE);*/ @@ -69,7 +72,7 @@ void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X, int spreading_decision(const CELTMode *m, const celt_norm *X, int *average, int last_decision, int *hf_average, int *tapset_decision, int update_hf, - int end, int C, int M); + int end, int C, int M, const int *spread_weight); #ifdef MEASURE_NORM_MSE void measure_norm_mse(const CELTMode *m, float *X, float *X0, float *bandE, float *bandE0, int M, int N, int C); @@ -105,7 +108,7 @@ void quant_all_bands(int encode, const CELTMode *m, int start, int end, const celt_ener *bandE, int *pulses, int shortBlocks, int spread, int dual_stereo, int intensity, int *tf_res, opus_int32 total_bits, opus_int32 balance, ec_ctx *ec, int M, int codedBands, opus_uint32 *seed, - int arch); + int complexity, int arch, int disable_inv); void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_masks, int LM, int C, int size, int start, diff --git a/libs/libopus/celt/celt.c b/libs/libopus/celt/celt.c index b121c51a1..9ce234695 100644 --- a/libs/libopus/celt/celt.c +++ b/libs/libopus/celt/celt.c @@ -111,26 +111,31 @@ void comb_filter_const_c(opus_val32 *y, opus_val32 *x, int T, int N, t = MAC16_32_Q16(x[i], g10, x2); t = MAC16_32_Q16(t, g11, ADD32(x1,x3)); t = MAC16_32_Q16(t, g12, ADD32(x0,x4)); + t = SATURATE(t, SIG_SAT); y[i] = t; x4=SHL32(x[i-T+3],1); t = MAC16_32_Q16(x[i+1], g10, x1); t = MAC16_32_Q16(t, g11, ADD32(x0,x2)); t = MAC16_32_Q16(t, g12, ADD32(x4,x3)); + t = SATURATE(t, SIG_SAT); y[i+1] = t; x3=SHL32(x[i-T+4],1); t = MAC16_32_Q16(x[i+2], g10, x0); t = MAC16_32_Q16(t, g11, ADD32(x4,x1)); t = MAC16_32_Q16(t, g12, ADD32(x3,x2)); + t = SATURATE(t, SIG_SAT); y[i+2] = t; x2=SHL32(x[i-T+5],1); t = MAC16_32_Q16(x[i+3], g10, x4); t = MAC16_32_Q16(t, g11, ADD32(x3,x0)); t = MAC16_32_Q16(t, g12, ADD32(x2,x1)); + t = SATURATE(t, SIG_SAT); y[i+3] = t; x1=SHL32(x[i-T+6],1); t = MAC16_32_Q16(x[i+4], g10, x3); t = MAC16_32_Q16(t, g11, ADD32(x2,x4)); t = MAC16_32_Q16(t, g12, ADD32(x1,x0)); + t = SATURATE(t, SIG_SAT); y[i+4] = t; } #ifdef CUSTOM_MODES @@ -141,6 +146,7 @@ void comb_filter_const_c(opus_val32 *y, opus_val32 *x, int T, int N, t = MAC16_32_Q16(x[i], g10, x2); t = MAC16_32_Q16(t, g11, ADD32(x1,x3)); t = MAC16_32_Q16(t, g12, ADD32(x0,x4)); + t = SATURATE(t, SIG_SAT); y[i] = t; x4=x3; x3=x2; @@ -169,6 +175,7 @@ void comb_filter_const_c(opus_val32 *y, opus_val32 *x, int T, int N, + MULT16_32_Q15(g10,x2) + MULT16_32_Q15(g11,ADD32(x1,x3)) + MULT16_32_Q15(g12,ADD32(x0,x4)); + y[i] = SATURATE(y[i], SIG_SAT); x4=x3; x3=x2; x2=x1; @@ -200,6 +207,10 @@ void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N, OPUS_MOVE(y, x, N); return; } + /* When the gain is zero, T0 and/or T1 is set to zero. We need + to have then be at least 2 to avoid processing garbage data. */ + T0 = IMAX(T0, COMBFILTER_MINPERIOD); + T1 = IMAX(T1, COMBFILTER_MINPERIOD); g00 = MULT16_16_P15(g0, gains[tapset0][0]); g01 = MULT16_16_P15(g0, gains[tapset0][1]); g02 = MULT16_16_P15(g0, gains[tapset0][2]); @@ -225,6 +236,7 @@ void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N, + MULT16_32_Q15(MULT16_16_Q15(f,g10),x2) + MULT16_32_Q15(MULT16_16_Q15(f,g11),ADD32(x1,x3)) + MULT16_32_Q15(MULT16_16_Q15(f,g12),ADD32(x0,x4)); + y[i] = SATURATE(y[i], SIG_SAT); x4=x3; x3=x2; x2=x1; @@ -244,11 +256,16 @@ void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N, } #endif /* OVERRIDE_comb_filter */ +/* TF change table. Positive values mean better frequency resolution (longer + effective window), whereas negative values mean better time resolution + (shorter effective window). The second index is computed as: + 4*isTransient + 2*tf_select + per_band_flag */ const signed char tf_select_table[4][8] = { - {0, -1, 0, -1, 0,-1, 0,-1}, - {0, -1, 0, -2, 1, 0, 1,-1}, - {0, -2, 0, -3, 2, 0, 1,-1}, - {0, -2, 0, -3, 3, 0, 1,-1}, + /*isTransient=0 isTransient=1 */ + {0, -1, 0, -1, 0,-1, 0,-1}, /* 2.5 ms */ + {0, -1, 0, -2, 1, 0, 1,-1}, /* 5 ms */ + {0, -2, 0, -3, 2, 0, 1,-1}, /* 10 ms */ + {0, -2, 0, -3, 3, 0, 1,-1}, /* 20 ms */ }; diff --git a/libs/libopus/celt/celt.h b/libs/libopus/celt/celt.h index d1f7eb690..24b6b2b52 100644 --- a/libs/libopus/celt/celt.h +++ b/libs/libopus/celt/celt.h @@ -50,6 +50,8 @@ extern "C" { #define CELTDecoder OpusCustomDecoder #define CELTMode OpusCustomMode +#define LEAK_BANDS 19 + typedef struct { int valid; float tonality; @@ -57,17 +59,27 @@ typedef struct { float noisiness; float activity; float music_prob; - int bandwidth; -}AnalysisInfo; + float music_prob_min; + float music_prob_max; + int bandwidth; + float activity_probability; + float max_pitch_ratio; + /* Store as Q6 char to save space. */ + unsigned char leak_boost[LEAK_BANDS]; +} AnalysisInfo; + +typedef struct { + int signalType; + int offset; +} SILKInfo; #define __celt_check_mode_ptr_ptr(ptr) ((ptr) + ((ptr) - (const CELTMode**)(ptr))) #define __celt_check_analysis_ptr(ptr) ((ptr) + ((ptr) - (const AnalysisInfo*)(ptr))) -/* Encoder/decoder Requests */ +#define __celt_check_silkinfo_ptr(ptr) ((ptr) + ((ptr) - (const SILKInfo*)(ptr))) -/* Expose this option again when variable framesize actually works */ -#define OPUS_FRAMESIZE_VARIABLE 5010 /**< Optimize the frame size dynamically */ +/* Encoder/decoder Requests */ #define CELT_SET_PREDICTION_REQUEST 10002 @@ -116,6 +128,9 @@ typedef struct { #define OPUS_SET_ENERGY_MASK_REQUEST 10026 #define OPUS_SET_ENERGY_MASK(x) OPUS_SET_ENERGY_MASK_REQUEST, __opus_check_val16_ptr(x) +#define CELT_SET_SILK_INFO_REQUEST 10028 +#define CELT_SET_SILK_INFO(x) CELT_SET_SILK_INFO_REQUEST, __celt_check_silkinfo_ptr(x) + /* Encoder stuff */ int celt_encoder_get_size(int channels); @@ -194,6 +209,13 @@ static OPUS_INLINE int fromOpus(unsigned char c) extern const signed char tf_select_table[4][8]; +#if defined(ENABLE_HARDENING) || defined(ENABLE_ASSERTIONS) +void validate_celt_decoder(CELTDecoder *st); +#define VALIDATE_CELT_DECODER(st) validate_celt_decoder(st) +#else +#define VALIDATE_CELT_DECODER(st) +#endif + int resampling_factor(opus_int32 rate); void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp, diff --git a/libs/libopus/celt/celt_decoder.c b/libs/libopus/celt/celt_decoder.c index b978bb34d..74ca3b740 100644 --- a/libs/libopus/celt/celt_decoder.c +++ b/libs/libopus/celt/celt_decoder.c @@ -51,6 +51,14 @@ #include "celt_lpc.h" #include "vq.h" +/* The maximum pitch lag to allow in the pitch-based PLC. It's possible to save + CPU time in the PLC pitch search by making this smaller than MAX_PERIOD. The + current value corresponds to a pitch of 66.67 Hz. */ +#define PLC_PITCH_LAG_MAX (720) +/* The minimum pitch lag to allow in the pitch-based PLC. This corresponds to a + pitch of 480 Hz. */ +#define PLC_PITCH_LAG_MIN (100) + #if defined(SMALL_FOOTPRINT) && defined(FIXED_POINT) #define NORM_ALIASING_HACK #endif @@ -73,6 +81,7 @@ struct OpusCustomDecoder { int downsample; int start, end; int signalling; + int disable_inv; int arch; /* Everything beyond this point gets cleared on a reset */ @@ -100,6 +109,44 @@ struct OpusCustomDecoder { /* opus_val16 backgroundLogE[], Size = 2*mode->nbEBands */ }; +#if defined(ENABLE_HARDENING) || defined(ENABLE_ASSERTIONS) +/* Make basic checks on the CELT state to ensure we don't end + up writing all over memory. */ +void validate_celt_decoder(CELTDecoder *st) +{ +#ifndef CUSTOM_MODES + celt_assert(st->mode == opus_custom_mode_create(48000, 960, NULL)); + celt_assert(st->overlap == 120); + celt_assert(st->end <= 21); +#else +/* From Section 4.3 in the spec: "The normal CELT layer uses 21 of those bands, + though Opus Custom (see Section 6.2) may use a different number of bands" + + Check if it's within the maximum number of Bark frequency bands instead */ + celt_assert(st->end <= 25); +#endif + celt_assert(st->channels == 1 || st->channels == 2); + celt_assert(st->stream_channels == 1 || st->stream_channels == 2); + celt_assert(st->downsample > 0); + celt_assert(st->start == 0 || st->start == 17); + celt_assert(st->start < st->end); +#ifdef OPUS_ARCHMASK + celt_assert(st->arch >= 0); + celt_assert(st->arch <= OPUS_ARCHMASK); +#endif + celt_assert(st->last_pitch_index <= PLC_PITCH_LAG_MAX); + celt_assert(st->last_pitch_index >= PLC_PITCH_LAG_MIN || st->last_pitch_index == 0); + celt_assert(st->postfilter_period < MAX_PERIOD); + celt_assert(st->postfilter_period >= COMBFILTER_MINPERIOD || st->postfilter_period == 0); + celt_assert(st->postfilter_period_old < MAX_PERIOD); + celt_assert(st->postfilter_period_old >= COMBFILTER_MINPERIOD || st->postfilter_period_old == 0); + celt_assert(st->postfilter_tapset <= 2); + celt_assert(st->postfilter_tapset >= 0); + celt_assert(st->postfilter_tapset_old <= 2); + celt_assert(st->postfilter_tapset_old >= 0); +} +#endif + int celt_decoder_get_size(int channels) { const CELTMode *mode = opus_custom_mode_create(48000, 960, NULL); @@ -163,6 +210,11 @@ OPUS_CUSTOM_NOSTATIC int opus_custom_decoder_init(CELTDecoder *st, const CELTMod st->start = 0; st->end = st->mode->effEBands; st->signalling = 1; +#ifndef DISABLE_UPDATE_DRAFT + st->disable_inv = channels == 1; +#else + st->disable_inv = 0; +#endif st->arch = opus_select_arch(); opus_custom_decoder_ctl(st, OPUS_RESET_STATE); @@ -177,6 +229,36 @@ void opus_custom_decoder_destroy(CELTDecoder *st) } #endif /* CUSTOM_MODES */ +#ifndef CUSTOM_MODES +/* Special case for stereo with no downsampling and no accumulation. This is + quite common and we can make it faster by processing both channels in the + same loop, reducing overhead due to the dependency loop in the IIR filter. */ +static void deemphasis_stereo_simple(celt_sig *in[], opus_val16 *pcm, int N, const opus_val16 coef0, + celt_sig *mem) +{ + celt_sig * OPUS_RESTRICT x0; + celt_sig * OPUS_RESTRICT x1; + celt_sig m0, m1; + int j; + x0=in[0]; + x1=in[1]; + m0 = mem[0]; + m1 = mem[1]; + for (j=0;j<N;j++) + { + celt_sig tmp0, tmp1; + /* Add VERY_SMALL to x[] first to reduce dependency chain. */ + tmp0 = x0[j] + VERY_SMALL + m0; + tmp1 = x1[j] + VERY_SMALL + m1; + m0 = MULT16_32_Q15(coef0, tmp0); + m1 = MULT16_32_Q15(coef0, tmp1); + pcm[2*j ] = SCALEOUT(SIG2WORD16(tmp0)); + pcm[2*j+1] = SCALEOUT(SIG2WORD16(tmp1)); + } + mem[0] = m0; + mem[1] = m1; +} +#endif #ifndef RESYNTH static @@ -190,6 +272,14 @@ void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, c opus_val16 coef0; VARDECL(celt_sig, scratch); SAVE_STACK; +#ifndef CUSTOM_MODES + /* Short version for common case. */ + if (downsample == 1 && C == 2 && !accum) + { + deemphasis_stereo_simple(in, pcm, N, coef[0], mem); + return; + } +#endif #ifndef FIXED_POINT (void)accum; celt_assert(accum==0); @@ -225,7 +315,7 @@ void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, c /* Shortcut for the standard (non-custom modes) case */ for (j=0;j<N;j++) { - celt_sig tmp = x[j] + m + VERY_SMALL; + celt_sig tmp = x[j] + VERY_SMALL + m; m = MULT16_32_Q15(coef0, tmp); scratch[j] = tmp; } @@ -246,7 +336,7 @@ void deemphasis(celt_sig *in[], opus_val16 *pcm, int N, int C, int downsample, c { for (j=0;j<N;j++) { - celt_sig tmp = x[j] + m + VERY_SMALL; + celt_sig tmp = x[j] + VERY_SMALL + m; m = MULT16_32_Q15(coef0, tmp); y[j*C] = SCALEOUT(SIG2WORD16(tmp)); } @@ -333,7 +423,7 @@ void celt_synthesis(const CELTMode *mode, celt_norm *X, celt_sig * out_syn[], denormalise_bands(mode, X+N, freq2, oldBandE+nbEBands, start, effEnd, M, downsample, silence); for (i=0;i<N;i++) - freq[i] = HALF32(ADD32(freq[i],freq2[i])); + freq[i] = ADD32(HALF32(freq[i]), HALF32(freq2[i])); for (b=0;b<B;b++) clt_mdct_backward(&mode->mdct, &freq[b], out_syn[0]+NB*b, mode->window, overlap, shift, B, arch); } else { @@ -345,6 +435,12 @@ void celt_synthesis(const CELTMode *mode, celt_norm *X, celt_sig * out_syn[], clt_mdct_backward(&mode->mdct, &freq[b], out_syn[c]+NB*b, mode->window, overlap, shift, B, arch); } while (++c<CC); } + /* Saturate IMDCT output so that we can't overflow in the pitch postfilter + or in the */ + c=0; do { + for (i=0;i<N;i++) + out_syn[c][i] = SATURATE(out_syn[c][i], SIG_SAT); + } while (++c<CC); RESTORE_STACK; } @@ -387,14 +483,6 @@ static void tf_decode(int start, int end, int isTransient, int *tf_res, int LM, } } -/* The maximum pitch lag to allow in the pitch-based PLC. It's possible to save - CPU time in the PLC pitch search by making this smaller than MAX_PERIOD. The - current value corresponds to a pitch of 66.67 Hz. */ -#define PLC_PITCH_LAG_MAX (720) -/* The minimum pitch lag to allow in the pitch-based PLC. This corresponds to a - pitch of 480 Hz. */ -#define PLC_PITCH_LAG_MIN (100) - static int celt_plc_pitch_search(celt_sig *decode_mem[2], int C, int arch) { int pitch_index; @@ -504,12 +592,15 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) celt_synthesis(mode, X, out_syn, oldBandE, start, effEnd, C, C, 0, LM, st->downsample, 0, st->arch); } else { + int exc_length; /* Pitch-based PLC */ const opus_val16 *window; + opus_val16 *exc; opus_val16 fade = Q15ONE; int pitch_index; VARDECL(opus_val32, etmp); - VARDECL(opus_val16, exc); + VARDECL(opus_val16, _exc); + VARDECL(opus_val16, fir_tmp); if (loss_count == 0) { @@ -519,8 +610,14 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) fade = QCONST16(.8f,15); } + /* We want the excitation for 2 pitch periods in order to look for a + decaying signal, but we can't get more than MAX_PERIOD. */ + exc_length = IMIN(2*pitch_index, MAX_PERIOD); + ALLOC(etmp, overlap, opus_val32); - ALLOC(exc, MAX_PERIOD, opus_val16); + ALLOC(_exc, MAX_PERIOD+LPC_ORDER, opus_val16); + ALLOC(fir_tmp, exc_length, opus_val16); + exc = _exc+LPC_ORDER; window = mode->window; c=0; do { opus_val16 decay; @@ -529,13 +626,11 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) celt_sig *buf; int extrapolation_offset; int extrapolation_len; - int exc_length; int j; buf = decode_mem[c]; - for (i=0;i<MAX_PERIOD;i++) { - exc[i] = ROUND16(buf[DECODE_BUFFER_SIZE-MAX_PERIOD+i], SIG_SHIFT); - } + for (i=0;i<MAX_PERIOD+LPC_ORDER;i++) + exc[i-LPC_ORDER] = ROUND16(buf[DECODE_BUFFER_SIZE-MAX_PERIOD-LPC_ORDER+i], SIG_SHIFT); if (loss_count == 0) { @@ -561,22 +656,32 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) #endif } _celt_lpc(lpc+c*LPC_ORDER, ac, LPC_ORDER); +#ifdef FIXED_POINT + /* For fixed-point, apply bandwidth expansion until we can guarantee that + no overflow can happen in the IIR filter. This means: + 32768*sum(abs(filter)) < 2^31 */ + while (1) { + opus_val16 tmp=Q15ONE; + opus_val32 sum=QCONST16(1., SIG_SHIFT); + for (i=0;i<LPC_ORDER;i++) + sum += ABS16(lpc[c*LPC_ORDER+i]); + if (sum < 65535) break; + for (i=0;i<LPC_ORDER;i++) + { + tmp = MULT16_16_Q15(QCONST16(.99f,15), tmp); + lpc[c*LPC_ORDER+i] = MULT16_16_Q15(lpc[c*LPC_ORDER+i], tmp); + } + } +#endif } - /* We want the excitation for 2 pitch periods in order to look for a - decaying signal, but we can't get more than MAX_PERIOD. */ - exc_length = IMIN(2*pitch_index, MAX_PERIOD); /* Initialize the LPC history with the samples just before the start of the region for which we're computing the excitation. */ { - opus_val16 lpc_mem[LPC_ORDER]; - for (i=0;i<LPC_ORDER;i++) - { - lpc_mem[i] = - ROUND16(buf[DECODE_BUFFER_SIZE-exc_length-1-i], SIG_SHIFT); - } - /* Compute the excitation for exc_length samples before the loss. */ + /* Compute the excitation for exc_length samples before the loss. We need the copy + because celt_fir() cannot filter in-place. */ celt_fir(exc+MAX_PERIOD-exc_length, lpc+c*LPC_ORDER, - exc+MAX_PERIOD-exc_length, exc_length, LPC_ORDER, lpc_mem, st->arch); + fir_tmp, exc_length, LPC_ORDER, st->arch); + OPUS_COPY(exc+MAX_PERIOD-exc_length, fir_tmp, exc_length); } /* Check if the waveform is decaying, and if so how fast. @@ -630,9 +735,8 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) tmp = ROUND16( buf[DECODE_BUFFER_SIZE-MAX_PERIOD-N+extrapolation_offset+j], SIG_SHIFT); - S1 += SHR32(MULT16_16(tmp, tmp), 8); + S1 += SHR32(MULT16_16(tmp, tmp), 10); } - { opus_val16 lpc_mem[LPC_ORDER]; /* Copy the last decoded samples (prior to the overlap region) to @@ -644,6 +748,10 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) celt_iir(buf+DECODE_BUFFER_SIZE-N, lpc+c*LPC_ORDER, buf+DECODE_BUFFER_SIZE-N, extrapolation_len, LPC_ORDER, lpc_mem, st->arch); +#ifdef FIXED_POINT + for (i=0; i < extrapolation_len; i++) + buf[DECODE_BUFFER_SIZE-N+i] = SATURATE(buf[DECODE_BUFFER_SIZE-N+i], SIG_SAT); +#endif } /* Check if the synthesis energy is higher than expected, which can @@ -654,7 +762,7 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM) for (i=0;i<extrapolation_len;i++) { opus_val16 tmp = ROUND16(buf[DECODE_BUFFER_SIZE-N+i], SIG_SHIFT); - S2 += SHR32(MULT16_16(tmp, tmp), 8); + S2 += SHR32(MULT16_16(tmp, tmp), 10); } /* This checks for an "explosion" in the synthesis. */ #ifdef FIXED_POINT @@ -762,6 +870,7 @@ int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *dat const opus_int16 *eBands; ALLOC_STACK; + VALIDATE_CELT_DECODER(st); mode = st->mode; nbEBands = mode->nbEBands; overlap = mode->overlap; @@ -956,7 +1065,7 @@ int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *dat ALLOC(pulses, nbEBands, int); ALLOC(fine_priority, nbEBands, int); - codedBands = compute_allocation(mode, start, end, offsets, cap, + codedBands = clt_compute_allocation(mode, start, end, offsets, cap, alloc_trim, &intensity, &dual_stereo, bits, &balance, pulses, fine_quant, fine_priority, C, LM, dec, 0, 0, 0); @@ -979,7 +1088,8 @@ int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *dat quant_all_bands(0, mode, start, end, X, C==2 ? X+N : NULL, collapse_masks, NULL, pulses, shortBlocks, spread_decision, dual_stereo, intensity, tf_res, - len*(8<<BITRES)-anti_collapse_rsv, balance, dec, LM, codedBands, &st->rng, st->arch); + len*(8<<BITRES)-anti_collapse_rsv, balance, dec, LM, codedBands, &st->rng, 0, + st->arch, st->disable_inv); if (anti_collapse_rsv > 0) { @@ -1234,6 +1344,26 @@ int opus_custom_decoder_ctl(CELTDecoder * OPUS_RESTRICT st, int request, ...) *value=st->rng; } break; + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 value = va_arg(ap, opus_int32); + if(value<0 || value>1) + { + goto bad_arg; + } + st->disable_inv = value; + } + break; + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 *value = va_arg(ap, opus_int32*); + if (!value) + { + goto bad_arg; + } + *value = st->disable_inv; + } + break; default: goto bad_request; } diff --git a/libs/libopus/celt/celt_encoder.c b/libs/libopus/celt/celt_encoder.c index 3ee7a4d3f..d6f8afc20 100644 --- a/libs/libopus/celt/celt_encoder.c +++ b/libs/libopus/celt/celt_encoder.c @@ -73,8 +73,8 @@ struct OpusCustomEncoder { int constrained_vbr; /* If zero, VBR can do whatever it likes with the rate */ int loss_rate; int lsb_depth; - int variable_duration; int lfe; + int disable_inv; int arch; /* Everything beyond this point gets cleared on a reset */ @@ -98,6 +98,7 @@ struct OpusCustomEncoder { #endif int consec_transient; AnalysisInfo analysis; + SILKInfo silk_info; opus_val32 preemph_memE[2]; opus_val32 preemph_memD[2]; @@ -123,6 +124,7 @@ struct OpusCustomEncoder { /* opus_val16 oldBandE[], Size = channels*mode->nbEBands */ /* opus_val16 oldLogE[], Size = channels*mode->nbEBands */ /* opus_val16 oldLogE2[], Size = channels*mode->nbEBands */ + /* opus_val16 energyError[], Size = channels*mode->nbEBands */ }; int celt_encoder_get_size(int channels) @@ -136,9 +138,10 @@ OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int int size = sizeof(struct CELTEncoder) + (channels*mode->overlap-1)*sizeof(celt_sig) /* celt_sig in_mem[channels*mode->overlap]; */ + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_mem[channels*COMBFILTER_MAXPERIOD]; */ - + 3*channels*mode->nbEBands*sizeof(opus_val16); /* opus_val16 oldBandE[channels*mode->nbEBands]; */ + + 4*channels*mode->nbEBands*sizeof(opus_val16); /* opus_val16 oldBandE[channels*mode->nbEBands]; */ /* opus_val16 oldLogE[channels*mode->nbEBands]; */ /* opus_val16 oldLogE2[channels*mode->nbEBands]; */ + /* opus_val16 energyError[channels*mode->nbEBands]; */ return size; } @@ -178,7 +181,6 @@ static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode, st->start = 0; st->end = st->mode->effEBands; st->signalling = 1; - st->arch = arch; st->constrained_vbr = 1; @@ -223,7 +225,8 @@ void opus_custom_encoder_destroy(CELTEncoder *st) static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C, - opus_val16 *tf_estimate, int *tf_chan) + opus_val16 *tf_estimate, int *tf_chan, int allow_weak_transients, + int *weak_transient) { int i; VARDECL(opus_val16, tmp); @@ -233,6 +236,12 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int int c; opus_val16 tf_max; int len2; + /* Forward masking: 6.7 dB/ms. */ +#ifdef FIXED_POINT + int forward_shift = 4; +#else + opus_val16 forward_decay = QCONST16(.0625f,15); +#endif /* Table of 6*64/x, trained on real data to minimize the average error */ static const unsigned char inv_table[128] = { 255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25, @@ -247,6 +256,19 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int SAVE_STACK; ALLOC(tmp, len, opus_val16); + *weak_transient = 0; + /* For lower bitrates, let's be more conservative and have a forward masking + decay of 3.3 dB/ms. This avoids having to code transients at very low + bitrate (mostly for hybrid), which can result in unstable energy and/or + partial collapse. */ + if (allow_weak_transients) + { +#ifdef FIXED_POINT + forward_shift = 5; +#else + forward_decay = QCONST16(.03125f,15); +#endif + } len2=len/2; for (c=0;c<C;c++) { @@ -269,7 +291,7 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int mem0 = mem1 + y - 2*x; mem1 = x - .5f*y; #endif - tmp[i] = EXTRACT16(SHR32(y,2)); + tmp[i] = SROUND16(y, 2); /*printf("%f ", tmp[i]);*/ } /*printf("\n");*/ @@ -280,7 +302,7 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int /* Normalize tmp to max range */ { int shift=0; - shift = 14-celt_ilog2(1+celt_maxabs16(tmp, len)); + shift = 14-celt_ilog2(MAX16(1, celt_maxabs16(tmp, len))); if (shift!=0) { for (i=0;i<len;i++) @@ -299,9 +321,9 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int mean += x2; #ifdef FIXED_POINT /* FIXME: Use PSHR16() instead */ - tmp[i] = mem0 + PSHR32(x2-mem0,4); + tmp[i] = mem0 + PSHR32(x2-mem0,forward_shift); #else - tmp[i] = mem0 + MULT16_16_P15(QCONST16(.0625f,15),x2-mem0); + tmp[i] = mem0 + MULT16_16_P15(forward_decay,x2-mem0); #endif mem0 = tmp[i]; } @@ -311,6 +333,7 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int /* Backward pass to compute the pre-echo threshold */ for (i=len2-1;i>=0;i--) { + /* Backward masking: 13.9 dB/ms. */ #ifdef FIXED_POINT /* FIXME: Use PSHR16() instead */ tmp[i] = mem0 + PSHR32(tmp[i]-mem0,3); @@ -339,6 +362,12 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int /* Compute harmonic mean discarding the unreliable boundaries The data is smooth, so we only take 1/4th of the samples */ unmask=0; + /* We should never see NaNs here. If we find any, then something really bad happened and we better abort + before it does any damage later on. If these asserts are disabled (no hardening), then the table + lookup a few lines below (id = ...) is likely to crash dur to an out-of-bounds read. DO NOT FIX + that crash on NaN since it could result in a worse issue later on. */ + celt_assert(!celt_isnan(tmp[0])); + celt_assert(!celt_isnan(norm)); for (i=12;i<len2-5;i+=4) { int id; @@ -359,7 +388,12 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int } } is_transient = mask_metric>200; - + /* For low bitrates, define "weak transients" that need to be + handled differently to avoid partial collapse. */ + if (allow_weak_transients && is_transient && mask_metric<600) { + is_transient = 0; + *weak_transient = 1; + } /* Arbitrary metric for VBR boost */ tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42); /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */ @@ -549,7 +583,7 @@ static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias static int tf_analysis(const CELTMode *m, int len, int isTransient, int *tf_res, int lambda, celt_norm *X, int N0, int LM, - int *tf_sum, opus_val16 tf_estimate, int tf_chan) + opus_val16 tf_estimate, int tf_chan, int *importance) { int i; VARDECL(int, metric); @@ -574,7 +608,6 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient, ALLOC(path0, len, int); ALLOC(path1, len, int); - *tf_sum = 0; for (i=0;i<len;i++) { int k, N; @@ -629,27 +662,26 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient, metric[i] = 2*best_level; else metric[i] = -2*best_level; - *tf_sum += (isTransient ? LM : 0) - metric[i]/2; /* For bands that can't be split to -1, set the metric to the half-way point to avoid biasing the decision */ if (narrow && (metric[i]==0 || metric[i]==-2*LM)) metric[i]-=1; - /*printf("%d ", metric[i]);*/ + /*printf("%d ", metric[i]/2 + (!isTransient)*LM);*/ } /*printf("\n");*/ /* Search for the optimal tf resolution, including tf_select */ tf_select = 0; for (sel=0;sel<2;sel++) { - cost0 = 0; - cost1 = isTransient ? 0 : lambda; + cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+0]); + cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+1]) + (isTransient ? 0 : lambda); for (i=1;i<len;i++) { int curr0, curr1; curr0 = IMIN(cost0, cost1 + lambda); curr1 = IMIN(cost0 + lambda, cost1); - cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]); - cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]); + cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]); + cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]); } cost0 = IMIN(cost0, cost1); selcost[sel]=cost0; @@ -658,8 +690,8 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient, * If tests confirm it's useful for non-transients, we could allow it. */ if (selcost[1]<selcost[0] && isTransient) tf_select=1; - cost0 = 0; - cost1 = isTransient ? 0 : lambda; + cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]); + cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]) + (isTransient ? 0 : lambda); /* Viterbi forward pass */ for (i=1;i<len;i++) { @@ -687,8 +719,8 @@ static int tf_analysis(const CELTMode *m, int len, int isTransient, curr1 = from1; path1[i]= 1; } - cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]); - cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]); + cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]); + cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]); } tf_res[len-1] = cost0 < cost1 ? 0 : 1; /* Viterbi backward pass to check the decisions */ @@ -754,7 +786,7 @@ static void tf_encode(int start, int end, int isTransient, int *tf_res, int LM, static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X, const opus_val16 *bandLogE, int end, int LM, int C, int N0, AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate, - int intensity, opus_val16 surround_trim, int arch) + int intensity, opus_val16 surround_trim, opus_int32 equiv_rate, int arch) { int i; opus_val32 diff=0; @@ -762,6 +794,14 @@ static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X, int trim_index; opus_val16 trim = QCONST16(5.f, 8); opus_val16 logXC, logXC2; + /* At low bitrate, reducing the trim seems to help. At higher bitrates, it's less + clear what's best, so we're keeping it as it was before, at least for now. */ + if (equiv_rate < 64000) { + trim = QCONST16(4.f, 8); + } else if (equiv_rate < 80000) { + opus_int32 frac = (equiv_rate-64000) >> 10; + trim = QCONST16(4.f, 8) + QCONST16(1.f/16.f, 8)*frac; + } if (C==2) { opus_val16 sum = 0; /* Q10 */ @@ -809,7 +849,7 @@ static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X, } while (++c<C); diff /= C*(end-1); /*printf("%f\n", diff);*/ - trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), SHR16(diff+QCONST16(1.f, DB_SHIFT),DB_SHIFT-8)/6 )); + trim -= MAX32(-QCONST16(2.f, 8), MIN32(QCONST16(2.f, 8), SHR32(diff+QCONST16(1.f, DB_SHIFT),DB_SHIFT-8)/6 )); trim -= SHR16(surround_trim, DB_SHIFT-8); trim -= 2*SHR16(tf_estimate, 14-8); #ifndef DISABLE_FLOAT_API @@ -930,7 +970,8 @@ static opus_val16 median_of_3(const opus_val16 *x) static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 *bandLogE2, int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, const opus_int16 *logN, int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, int LM, - int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc) + int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc, + AnalysisInfo *analysis, int *importance, int *spread_weight) { int i, c; opus_int32 tot_boost=0; @@ -956,6 +997,42 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 for (i=0;i<end;i++) maxDepth = MAX16(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]); } while (++c<C); + { + /* Compute a really simple masking model to avoid taking into account completely masked + bands when computing the spreading decision. */ + VARDECL(opus_val16, mask); + VARDECL(opus_val16, sig); + ALLOC(mask, nbEBands, opus_val16); + ALLOC(sig, nbEBands, opus_val16); + for (i=0;i<end;i++) + mask[i] = bandLogE[i]-noise_floor[i]; + if (C==2) + { + for (i=0;i<end;i++) + mask[i] = MAX16(mask[i], bandLogE[nbEBands+i]-noise_floor[i]); + } + OPUS_COPY(sig, mask, end); + for (i=1;i<end;i++) + mask[i] = MAX16(mask[i], mask[i-1] - QCONST16(2.f, DB_SHIFT)); + for (i=end-2;i>=0;i--) + mask[i] = MAX16(mask[i], mask[i+1] - QCONST16(3.f, DB_SHIFT)); + for (i=0;i<end;i++) + { + /* Compute SMR: Mask is never more than 72 dB below the peak and never below the noise floor.*/ + opus_val16 smr = sig[i]-MAX16(MAX16(0, maxDepth-QCONST16(12.f, DB_SHIFT)), mask[i]); + /* Clamp SMR to make sure we're not shifting by something negative or too large. */ +#ifdef FIXED_POINT + /* FIXME: Use PSHR16() instead */ + int shift = -PSHR32(MAX16(-QCONST16(5.f, DB_SHIFT), MIN16(0, smr)), DB_SHIFT); +#else + int shift = IMIN(5, IMAX(0, -(int)floor(.5f + smr))); +#endif + spread_weight[i] = 32 >> shift; + } + /*for (i=0;i<end;i++) + printf("%d ", spread_weight[i]); + printf("\n");*/ + } /* Make sure that dynamic allocation can't make us bust the budget */ if (effectiveBytes > 50 && LM>=1 && !lfe) { @@ -1012,6 +1089,14 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 } for (i=start;i<end;i++) follower[i] = MAX16(follower[i], surround_dynalloc[i]); + for (i=start;i<end;i++) + { +#ifdef FIXED_POINT + importance[i] = PSHR32(13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT))), 16); +#else + importance[i] = (int)floor(.5f+13*celt_exp2(MIN16(follower[i], QCONST16(4.f, DB_SHIFT)))); +#endif + } /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */ if ((!vbr || constrained_vbr)&&!isTransient) { @@ -1020,14 +1105,26 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 } for (i=start;i<end;i++) { - int width; - int boost; - int boost_bits; - if (i<8) follower[i] *= 2; if (i>=12) follower[i] = HALF16(follower[i]); + } +#ifdef DISABLE_FLOAT_API + (void)analysis; +#else + if (analysis->valid) + { + for (i=start;i<IMIN(LEAK_BANDS, end);i++) + follower[i] = follower[i] + QCONST16(1.f/64.f, DB_SHIFT)*analysis->leak_boost[i]; + } +#endif + for (i=start;i<end;i++) + { + int width; + int boost; + int boost_bits; + follower[i] = MIN16(follower[i], QCONST16(4, DB_SHIFT)); width = C*(eBands[i+1]-eBands[i])<<LM; @@ -1042,11 +1139,11 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 boost = (int)SHR32(EXTEND32(follower[i])*width/6,DB_SHIFT); boost_bits = boost*6<<BITRES; } - /* For CBR and non-transient CVBR frames, limit dynalloc to 1/4 of the bits */ + /* For CBR and non-transient CVBR frames, limit dynalloc to 2/3 of the bits */ if ((!vbr || (constrained_vbr&&!isTransient)) - && (tot_boost+boost_bits)>>BITRES>>3 > effectiveBytes/4) + && (tot_boost+boost_bits)>>BITRES>>3 > 2*effectiveBytes/3) { - opus_int32 cap = ((effectiveBytes/4)<<BITRES<<3); + opus_int32 cap = ((2*effectiveBytes/3)<<BITRES<<3); offsets[i] = cap-tot_boost; tot_boost = cap; break; @@ -1055,6 +1152,9 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 tot_boost += boost_bits; } } + } else { + for (i=start;i<end;i++) + importance[i] = 13; } *tot_boost_ = tot_boost; RESTORE_STACK; @@ -1063,7 +1163,7 @@ static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, int CC, int N, - int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes) + int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes, AnalysisInfo *analysis) { int c; VARDECL(celt_sig, _pre); @@ -1119,7 +1219,12 @@ static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, gain1 = 0; pitch_index = COMBFILTER_MINPERIOD; } - +#ifndef DISABLE_FLOAT_API + if (analysis->valid) + gain1 = (opus_val16)(gain1 * analysis->max_pitch_ratio); +#else + (void)analysis; +#endif /* Gain threshold for enabling the prefilter/postfilter */ pf_threshold = QCONST16(.2f,15); @@ -1193,7 +1298,7 @@ static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 int LM, opus_int32 bitrate, int lastCodedBands, int C, int intensity, int constrained_vbr, opus_val16 stereo_saving, int tot_boost, opus_val16 tf_estimate, int pitch_change, opus_val16 maxDepth, - int variable_duration, int lfe, int has_surround_mask, opus_val16 surround_masking, + int lfe, int has_surround_mask, opus_val16 surround_masking, opus_val16 temporal_vbr) { /* The target rate in 8th bits per frame */ @@ -1235,10 +1340,9 @@ static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_stereo_dof<<BITRES)),8)); } /* Boost the rate according to dynalloc (minus the dynalloc average for calibration). */ - target += tot_boost-(16<<LM); + target += tot_boost-(19<<LM); /* Apply transient boost, compensating for average boost. */ - tf_calibration = variable_duration==OPUS_FRAMESIZE_VARIABLE ? - QCONST16(0.02f,14) : QCONST16(0.04f,14); + tf_calibration = QCONST16(0.044f,14); target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target),1); #ifndef DISABLE_FLOAT_API @@ -1249,7 +1353,7 @@ static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 float tonal; /* Tonality boost (compensating for the average). */ - tonal = MAX16(0.f,analysis->tonality-.15f)-0.09f; + tonal = MAX16(0.f,analysis->tonality-.15f)-0.12f; tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal); if (pitch_change) tonal_target += (opus_int32)((coded_bins<<BITRES)*.8f); @@ -1279,21 +1383,11 @@ static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 /*printf("%f %d\n", maxDepth, floor_depth);*/ } - if ((!has_surround_mask||lfe) && (constrained_vbr || bitrate<64000)) + /* Make VBR less aggressive for constrained VBR because we can't keep a higher bitrate + for long. Needs tuning. */ + if ((!has_surround_mask||lfe) && constrained_vbr) { - opus_val16 rate_factor = Q15ONE; - if (bitrate < 64000) - { -#ifdef FIXED_POINT - rate_factor = MAX16(0,(bitrate-32000)); -#else - rate_factor = MAX16(0,(1.f/32768)*(bitrate-32000)); -#endif - } - if (constrained_vbr) - rate_factor = MIN16(rate_factor, QCONST16(0.67f, 15)); - target = base_target + (opus_int32)MULT16_32_Q15(rate_factor, target-base_target); - + target = base_target + (opus_int32)MULT16_32_Q15(QCONST16(0.67f, 15), target-base_target); } if (!has_surround_mask && tf_estimate < QCONST16(.2f, 14)) @@ -1327,11 +1421,13 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, VARDECL(int, pulses); VARDECL(int, cap); VARDECL(int, offsets); + VARDECL(int, importance); + VARDECL(int, spread_weight); VARDECL(int, fine_priority); VARDECL(int, tf_res); VARDECL(unsigned char, collapse_masks); celt_sig *prefilter_mem; - opus_val16 *oldBandE, *oldLogE, *oldLogE2; + opus_val16 *oldBandE, *oldLogE, *oldLogE2, *energyError; int shortBlocks=0; int isTransient=0; const int CC = st->channels; @@ -1343,7 +1439,6 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, int end; int effEnd; int codedBands; - int tf_sum; int alloc_trim; int pitch_index=COMBFILTER_MINPERIOD; opus_val16 gain1 = 0; @@ -1355,6 +1450,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, opus_int32 total_boost; opus_int32 balance; opus_int32 tell; + opus_int32 tell0_frac; int prefilter_tapset=0; int pf_on; int anti_collapse_rsv; @@ -1376,7 +1472,10 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, opus_val16 surround_masking=0; opus_val16 temporal_vbr=0; opus_val16 surround_trim = 0; - opus_int32 equiv_rate = 510000; + opus_int32 equiv_rate; + int hybrid; + int weak_transient = 0; + int enable_tf_analysis; VARDECL(opus_val16, surround_dynalloc); ALLOC_STACK; @@ -1386,6 +1485,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, eBands = mode->eBands; start = st->start; end = st->end; + hybrid = start != 0; tf_estimate = 0; if (nbCompressedBytes<2 || pcm==NULL) { @@ -1409,12 +1509,14 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, oldBandE = (opus_val16*)(st->in_mem+CC*(overlap+COMBFILTER_MAXPERIOD)); oldLogE = oldBandE + CC*nbEBands; oldLogE2 = oldLogE + CC*nbEBands; + energyError = oldLogE2 + CC*nbEBands; if (enc==NULL) { - tell=1; + tell0_frac=tell=1; nbFilledBytes=0; } else { + tell0_frac=ec_tell_frac(enc); tell=ec_tell(enc); nbFilledBytes=(tell+4)>>3; } @@ -1467,10 +1569,11 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, if (st->bitrate!=OPUS_BITRATE_MAX) nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes, (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling)); - effectiveBytes = nbCompressedBytes; + effectiveBytes = nbCompressedBytes - nbFilledBytes; } + equiv_rate = ((opus_int32)nbCompressedBytes*8*50 << (3-LM)) - (40*C+20)*((400>>LM) - 50); if (st->bitrate != OPUS_BITRATE_MAX) - equiv_rate = st->bitrate - (40*C+20)*((400>>LM) - 50); + equiv_rate = IMIN(equiv_rate, st->bitrate - (40*C+20)*((400>>LM) - 50)); if (enc==NULL) { @@ -1558,17 +1661,17 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, { int enabled; int qg; - enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && start==0 && !silence && !st->disable_pf - && st->complexity >= 5 && !(st->consec_transient && LM!=3 && st->variable_duration==OPUS_FRAMESIZE_VARIABLE); + enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && !hybrid && !silence && !st->disable_pf + && st->complexity >= 5; prefilter_tapset = st->tapset_decision; - pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes); + pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes, &st->analysis); if ((gain1 > QCONST16(.4f,15) || st->prefilter_gain > QCONST16(.4f,15)) && (!st->analysis.valid || st->analysis.tonality > .3) && (pitch_index > 1.26*st->prefilter_period || pitch_index < .79*st->prefilter_period)) pitch_change = 1; if (pf_on==0) { - if(start==0 && tell+16<=total_bits) + if(!hybrid && tell+16<=total_bits) ec_enc_bit_logp(enc, 0, 1); } else { /*This block is not gated by a total bits check only because @@ -1589,8 +1692,12 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, shortBlocks = 0; if (st->complexity >= 1 && !st->lfe) { + /* Reduces the likelihood of energy instability on fricatives at low bitrate + in hybrid mode. It seems like we still want to have real transients on vowels + though (small SILK quantization offset value). */ + int allow_weak_transients = hybrid && effectiveBytes<15 && st->silk_info.signalType != 2; isTransient = transient_analysis(in, N+overlap, CC, - &tf_estimate, &tf_chan); + &tf_estimate, &tf_chan, allow_weak_transients, &weak_transient); } if (LM>0 && ec_tell(enc)+3<=total_bits) { @@ -1610,16 +1717,19 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, if (secondMdct) { compute_mdcts(mode, 0, in, freq, C, CC, LM, st->upsample, st->arch); - compute_band_energies(mode, freq, bandE, effEnd, C, LM); + compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); amp2Log2(mode, effEnd, end, bandE, bandLogE2, C); for (i=0;i<C*nbEBands;i++) bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT)); } compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch); + /* This should catch any NaN in the CELT input. Since we're not supposed to see any (they're filtered + at the Opus layer), just abort. */ + celt_assert(!celt_isnan(freq[0]) && (C==1 || !celt_isnan(freq[N]))); if (CC==2&&C==1) tf_chan = 0; - compute_band_energies(mode, freq, bandE, effEnd, C, LM); + compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); if (st->lfe) { @@ -1634,7 +1744,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, ALLOC(surround_dynalloc, C*nbEBands, opus_val16); OPUS_CLEAR(surround_dynalloc, end); /* This computes how much masking takes place between surround channels */ - if (start==0&&st->energy_mask&&!st->lfe) + if (!hybrid&&st->energy_mask&&!st->lfe) { int mask_end; int midband; @@ -1736,14 +1846,14 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, /* Last chance to catch any transient we might have missed in the time-domain analysis */ - if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe) + if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe && !hybrid) { if (patch_transient_decision(bandLogE, oldBandE, nbEBands, start, end, C)) { isTransient = 1; shortBlocks = M; compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch); - compute_band_energies(mode, freq, bandE, effEnd, C, LM); + compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); amp2Log2(mode, effEnd, end, bandE, bandLogE, C); /* Compensate for the scaling of short vs long mdcts */ for (i=0;i<C*nbEBands;i++) @@ -1760,31 +1870,59 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, /* Band normalisation */ normalise_bands(mode, freq, X, bandE, effEnd, C, M); + enable_tf_analysis = effectiveBytes>=15*C && !hybrid && st->complexity>=2 && !st->lfe; + + ALLOC(offsets, nbEBands, int); + ALLOC(importance, nbEBands, int); + ALLOC(spread_weight, nbEBands, int); + + maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, start, end, C, offsets, + st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr, + eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc, &st->analysis, importance, spread_weight); + ALLOC(tf_res, nbEBands, int); /* Disable variable tf resolution for hybrid and at very low bitrate */ - if (effectiveBytes>=15*C && start==0 && st->complexity>=2 && !st->lfe) + if (enable_tf_analysis) { int lambda; - if (effectiveBytes<40) - lambda = 12; - else if (effectiveBytes<60) - lambda = 6; - else if (effectiveBytes<100) - lambda = 4; - else - lambda = 3; - lambda*=2; - tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, &tf_sum, tf_estimate, tf_chan); + lambda = IMAX(80, 20480/effectiveBytes + 2); + tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, tf_estimate, tf_chan, importance); for (i=effEnd;i<end;i++) tf_res[i] = tf_res[effEnd-1]; + } else if (hybrid && weak_transient) + { + /* For weak transients, we rely on the fact that improving time resolution using + TF on a long window is imperfect and will not result in an energy collapse at + low bitrate. */ + for (i=0;i<end;i++) + tf_res[i] = 1; + tf_select=0; + } else if (hybrid && effectiveBytes<15 && st->silk_info.signalType != 2) + { + /* For low bitrate hybrid, we force temporal resolution to 5 ms rather than 2.5 ms. */ + for (i=0;i<end;i++) + tf_res[i] = 0; + tf_select=isTransient; } else { - tf_sum = 0; for (i=0;i<end;i++) tf_res[i] = isTransient; tf_select=0; } ALLOC(error, C*nbEBands, opus_val16); + c=0; + do { + for (i=start;i<end;i++) + { + /* When the energy is stable, slightly bias energy quantization towards + the previous error to make the gain more stable (a constant offset is + better than fluctuations). */ + if (ABS32(SUB32(bandLogE[i+c*nbEBands], oldBandE[i+c*nbEBands])) < QCONST16(2.f, DB_SHIFT)) + { + bandLogE[i+c*nbEBands] -= MULT16_16_Q15(energyError[i+c*nbEBands], QCONST16(0.25f, 15)); + } + } + } while (++c < C); quant_coarse_energy(mode, start, end, effEnd, bandLogE, oldBandE, total_bits, error, enc, C, LM, nbAvailableBytes, st->force_intra, @@ -1798,7 +1936,15 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, { st->tapset_decision = 0; st->spread_decision = SPREAD_NORMAL; - } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C || start != 0) + } else if (hybrid) + { + if (st->complexity == 0) + st->spread_decision = SPREAD_NONE; + else if (isTransient) + st->spread_decision = SPREAD_NORMAL; + else + st->spread_decision = SPREAD_AGGRESSIVE; + } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C) { if (st->complexity == 0) st->spread_decision = SPREAD_NONE; @@ -1822,7 +1968,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, { st->spread_decision = spreading_decision(mode, X, &st->tonal_average, st->spread_decision, &st->hf_average, - &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M); + &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M, spread_weight); } /*printf("%d %d\n", st->tapset_decision, st->spread_decision);*/ /*printf("%f %d %f %d\n\n", st->analysis.tonality, st->spread_decision, st->analysis.tonality_slope, st->tapset_decision);*/ @@ -1830,11 +1976,6 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5); } - ALLOC(offsets, nbEBands, int); - - maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, start, end, C, offsets, - st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr, - eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc); /* For LFE, everything interesting is in the first band */ if (st->lfe) offsets[0] = IMIN(8, effectiveBytes/3); @@ -1896,12 +2037,15 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, alloc_trim = 5; if (tell+(6<<BITRES) <= total_bits - total_boost) { - if (st->lfe) + if (start > 0 || st->lfe) + { + st->stereo_saving = 0; alloc_trim = 5; - else + } else { alloc_trim = alloc_trim_analysis(mode, X, bandLogE, end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate, - st->intensity, surround_trim, st->arch); + st->intensity, surround_trim, equiv_rate, st->arch); + } ec_enc_icdf(enc, alloc_trim, trim_icdf, 7); tell = ec_tell_frac(enc); } @@ -1919,17 +2063,36 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20 ms. The CELT allocator will just not be able to use more than that anyway. */ nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM)); - base_target = vbr_rate - ((40*C+20)<<BITRES); + if (!hybrid) + { + base_target = vbr_rate - ((40*C+20)<<BITRES); + } else { + base_target = IMAX(0, vbr_rate - ((9*C+4)<<BITRES)); + } if (st->constrained_vbr) base_target += (st->vbr_offset>>lm_diff); - target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate, + if (!hybrid) + { + target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate, st->lastCodedBands, C, st->intensity, st->constrained_vbr, st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth, - st->variable_duration, st->lfe, st->energy_mask!=NULL, surround_masking, + st->lfe, st->energy_mask!=NULL, surround_masking, temporal_vbr); - + } else { + target = base_target; + /* Tonal frames (offset<100) need more bits than noisy (offset>100) ones. */ + if (st->silk_info.offset < 100) target += 12 << BITRES >> (3-LM); + if (st->silk_info.offset > 100) target -= 18 << BITRES >> (3-LM); + /* Boosting bitrate on transients and vowels with significant temporal + spikes. */ + target += (opus_int32)MULT16_16_Q14(tf_estimate-QCONST16(.25f,14), (50<<BITRES)); + /* If we have a strong transient, let's make sure it has enough bits to code + the first two bands, so that it can use folding rather than noise. */ + if (tf_estimate > QCONST16(.7f,14)) + target = IMAX(target, 50<<BITRES); + } /* The current offset is removed from the target and the space used so far is added*/ target=target+tell; @@ -1937,11 +2100,16 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, result in the encoder running out of bits. The margin of 2 bytes ensures that none of the bust-prevention logic in the decoder will have triggered so far. */ - min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2 - nbFilledBytes; + min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2; + /* Take into account the 37 bits we need to have left in the packet to + signal a redundant frame in hybrid mode. Creating a shorter packet would + create an entropy coder desync. */ + if (hybrid) + min_allowed = IMAX(min_allowed, (tell0_frac+(37<<BITRES)+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)); nbAvailableBytes = (target+(1<<(BITRES+2)))>>(BITRES+3); nbAvailableBytes = IMAX(min_allowed,nbAvailableBytes); - nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes) - nbFilledBytes; + nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes); /* By how much did we "miss" the target on that frame */ delta = target - vbr_rate; @@ -1988,7 +2156,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, st->vbr_reservoir = 0; /*printf ("+%d\n", adjust);*/ } - nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes); + nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes); /*printf("%d\n", nbCompressedBytes*50*8);*/ /* This moves the raw bits to take into account the new compressed size */ ec_enc_shrink(enc, nbCompressedBytes); @@ -2023,7 +2191,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, #endif if (st->lfe) signalBandwidth = 1; - codedBands = compute_allocation(mode, start, end, offsets, cap, + codedBands = clt_compute_allocation(mode, start, end, offsets, cap, alloc_trim, &st->intensity, &dual_stereo, bits, &balance, pulses, fine_quant, fine_priority, C, LM, enc, 1, st->lastCodedBands, signalBandwidth); if (st->lastCodedBands) @@ -2038,7 +2206,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, quant_all_bands(1, mode, start, end, X, C==2 ? X+N : NULL, collapse_masks, bandE, pulses, shortBlocks, st->spread_decision, dual_stereo, st->intensity, tf_res, nbCompressedBytes*(8<<BITRES)-anti_collapse_rsv, - balance, enc, LM, codedBands, &st->rng, st->arch); + balance, enc, LM, codedBands, &st->rng, st->complexity, st->arch, st->disable_inv); if (anti_collapse_rsv > 0) { @@ -2049,6 +2217,14 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, ec_enc_bits(enc, anti_collapse_on, 1); } quant_energy_finalise(mode, start, end, oldBandE, error, fine_quant, fine_priority, nbCompressedBytes*8-ec_tell(enc), enc, C); + OPUS_CLEAR(energyError, nbEBands*CC); + c=0; + do { + for (i=start;i<end;i++) + { + energyError[i+c*nbEBands] = MAX16(-QCONST16(0.5f, 15), MIN16(QCONST16(0.5f, 15), error[i+c*nbEBands])); + } + } while (++c < C); if (silence) { @@ -2321,10 +2497,24 @@ int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...) *value=st->lsb_depth; } break; - case OPUS_SET_EXPERT_FRAME_DURATION_REQUEST: + case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: { opus_int32 value = va_arg(ap, opus_int32); - st->variable_duration = value; + if(value<0 || value>1) + { + goto bad_arg; + } + st->disable_inv = value; + } + break; + case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: + { + opus_int32 *value = va_arg(ap, opus_int32*); + if (!value) + { + goto bad_arg; + } + *value = st->disable_inv; } break; case OPUS_RESET_STATE: @@ -2368,6 +2558,13 @@ int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...) OPUS_COPY(&st->analysis, info, 1); } break; + case CELT_SET_SILK_INFO_REQUEST: + { + SILKInfo *info = va_arg(ap, SILKInfo *); + if (info) + OPUS_COPY(&st->silk_info, info, 1); + } + break; case CELT_GET_MODE_REQUEST: { const CELTMode ** value = va_arg(ap, const CELTMode**); diff --git a/libs/libopus/celt/celt_lpc.c b/libs/libopus/celt/celt_lpc.c index b410a21c5..242e6df55 100644 --- a/libs/libopus/celt/celt_lpc.c +++ b/libs/libopus/celt/celt_lpc.c @@ -50,17 +50,21 @@ int p #endif OPUS_CLEAR(lpc, p); +#ifdef FIXED_POINT if (ac[0] != 0) +#else + if (ac[0] > 1e-10f) +#endif { for (i = 0; i < p; i++) { /* Sum up this iteration's reflection coefficient */ opus_val32 rr = 0; for (j = 0; j < i; j++) rr += MULT32_32_Q31(lpc[j],ac[i - j]); - rr += SHR32(ac[i + 1],3); - r = -frac_div32(SHL32(rr,3), error); + rr += SHR32(ac[i + 1],6); + r = -frac_div32(SHL32(rr,6), error); /* Update LPC coefficients and total error */ - lpc[i] = SHR32(r,3); + lpc[i] = SHR32(r,6); for (j = 0; j < (i+1)>>1; j++) { opus_val32 tmp1, tmp2; @@ -73,74 +77,100 @@ int p error = error - MULT32_32_Q31(MULT32_32_Q31(r,r),error); /* Bail out once we get 30 dB gain */ #ifdef FIXED_POINT - if (error<SHR32(ac[0],10)) + if (error<=SHR32(ac[0],10)) break; #else - if (error<.001f*ac[0]) + if (error<=.001f*ac[0]) break; #endif } } #ifdef FIXED_POINT - for (i=0;i<p;i++) - _lpc[i] = ROUND16(lpc[i],16); + { + /* Convert the int32 lpcs to int16 and ensure there are no wrap-arounds. + This reuses the logic in silk_LPC_fit() and silk_bwexpander_32(). Any bug + fixes should also be applied there. */ + int iter, idx = 0; + opus_val32 maxabs, absval, chirp_Q16, chirp_minus_one_Q16; + + for (iter = 0; iter < 10; iter++) { + maxabs = 0; + for (i = 0; i < p; i++) { + absval = ABS32(lpc[i]); + if (absval > maxabs) { + maxabs = absval; + idx = i; + } + } + maxabs = PSHR32(maxabs, 13); /* Q25->Q12 */ + + if (maxabs > 32767) { + maxabs = MIN32(maxabs, 163838); + chirp_Q16 = QCONST32(0.999, 16) - DIV32(SHL32(maxabs - 32767, 14), + SHR32(MULT32_32_32(maxabs, idx + 1), 2)); + chirp_minus_one_Q16 = chirp_Q16 - 65536; + + /* Apply bandwidth expansion. */ + for (i = 0; i < p - 1; i++) { + lpc[i] = MULT32_32_Q16(chirp_Q16, lpc[i]); + chirp_Q16 += PSHR32(MULT32_32_32(chirp_Q16, chirp_minus_one_Q16), 16); + } + lpc[p - 1] = MULT32_32_Q16(chirp_Q16, lpc[p - 1]); + } else { + break; + } + } + + if (iter == 10) { + /* If the coeffs still do not fit into the 16 bit range after 10 iterations, + fall back to the A(z)=1 filter. */ + OPUS_CLEAR(lpc, p); + _lpc[0] = 4096; /* Q12 */ + } else { + for (i = 0; i < p; i++) { + _lpc[i] = EXTRACT16(PSHR32(lpc[i], 13)); /* Q25->Q12 */ + } + } + } #endif } void celt_fir_c( - const opus_val16 *_x, + const opus_val16 *x, const opus_val16 *num, - opus_val16 *_y, + opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch) { int i,j; VARDECL(opus_val16, rnum); - VARDECL(opus_val16, x); SAVE_STACK; - + celt_assert(x != y); ALLOC(rnum, ord, opus_val16); - ALLOC(x, N+ord, opus_val16); for(i=0;i<ord;i++) rnum[i] = num[ord-i-1]; - for(i=0;i<ord;i++) - x[i] = mem[ord-i-1]; - for (i=0;i<N;i++) - x[i+ord]=_x[i]; - for(i=0;i<ord;i++) - mem[i] = _x[N-i-1]; -#ifdef SMALL_FOOTPRINT - (void)arch; - for (i=0;i<N;i++) - { - opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT); - for (j=0;j<ord;j++) - { - sum = MAC16_16(sum,rnum[j],x[i+j]); - } - _y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT)); - } -#else for (i=0;i<N-3;i+=4) { - opus_val32 sum[4]={0,0,0,0}; - xcorr_kernel(rnum, x+i, sum, ord, arch); - _y[i ] = SATURATE16(ADD32(EXTEND32(_x[i ]), PSHR32(sum[0], SIG_SHIFT))); - _y[i+1] = SATURATE16(ADD32(EXTEND32(_x[i+1]), PSHR32(sum[1], SIG_SHIFT))); - _y[i+2] = SATURATE16(ADD32(EXTEND32(_x[i+2]), PSHR32(sum[2], SIG_SHIFT))); - _y[i+3] = SATURATE16(ADD32(EXTEND32(_x[i+3]), PSHR32(sum[3], SIG_SHIFT))); + opus_val32 sum[4]; + sum[0] = SHL32(EXTEND32(x[i ]), SIG_SHIFT); + sum[1] = SHL32(EXTEND32(x[i+1]), SIG_SHIFT); + sum[2] = SHL32(EXTEND32(x[i+2]), SIG_SHIFT); + sum[3] = SHL32(EXTEND32(x[i+3]), SIG_SHIFT); + xcorr_kernel(rnum, x+i-ord, sum, ord, arch); + y[i ] = ROUND16(sum[0], SIG_SHIFT); + y[i+1] = ROUND16(sum[1], SIG_SHIFT); + y[i+2] = ROUND16(sum[2], SIG_SHIFT); + y[i+3] = ROUND16(sum[3], SIG_SHIFT); } for (;i<N;i++) { - opus_val32 sum = 0; + opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT); for (j=0;j<ord;j++) - sum = MAC16_16(sum,rnum[j],x[i+j]); - _y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT))); + sum = MAC16_16(sum,rnum[j],x[i+j-ord]); + y[i] = ROUND16(sum, SIG_SHIFT); } -#endif RESTORE_STACK; } @@ -166,7 +196,7 @@ void celt_iir(const opus_val32 *_x, { mem[j]=mem[j-1]; } - mem[0] = ROUND16(sum,SIG_SHIFT); + mem[0] = SROUND16(sum, SIG_SHIFT); _y[i] = sum; } #else @@ -195,20 +225,20 @@ void celt_iir(const opus_val32 *_x, xcorr_kernel(rden, y+i, sum, ord, arch); /* Patch up the result to compensate for the fact that this is an IIR */ - y[i+ord ] = -ROUND16(sum[0],SIG_SHIFT); + y[i+ord ] = -SROUND16(sum[0],SIG_SHIFT); _y[i ] = sum[0]; sum[1] = MAC16_16(sum[1], y[i+ord ], den[0]); - y[i+ord+1] = -ROUND16(sum[1],SIG_SHIFT); + y[i+ord+1] = -SROUND16(sum[1],SIG_SHIFT); _y[i+1] = sum[1]; sum[2] = MAC16_16(sum[2], y[i+ord+1], den[0]); sum[2] = MAC16_16(sum[2], y[i+ord ], den[1]); - y[i+ord+2] = -ROUND16(sum[2],SIG_SHIFT); + y[i+ord+2] = -SROUND16(sum[2],SIG_SHIFT); _y[i+2] = sum[2]; sum[3] = MAC16_16(sum[3], y[i+ord+2], den[0]); sum[3] = MAC16_16(sum[3], y[i+ord+1], den[1]); sum[3] = MAC16_16(sum[3], y[i+ord ], den[2]); - y[i+ord+3] = -ROUND16(sum[3],SIG_SHIFT); + y[i+ord+3] = -SROUND16(sum[3],SIG_SHIFT); _y[i+3] = sum[3]; } for (;i<N;i++) @@ -216,7 +246,7 @@ void celt_iir(const opus_val32 *_x, opus_val32 sum = _x[i]; for (j=0;j<ord;j++) sum -= MULT16_16(rden[j],y[i+j]); - y[i+ord] = ROUND16(sum,SIG_SHIFT); + y[i+ord] = SROUND16(sum,SIG_SHIFT); _y[i] = sum; } for(i=0;i<ord;i++) diff --git a/libs/libopus/celt/celt_lpc.h b/libs/libopus/celt/celt_lpc.h index 323459eb1..a4c5fd6ea 100644 --- a/libs/libopus/celt/celt_lpc.h +++ b/libs/libopus/celt/celt_lpc.h @@ -45,12 +45,11 @@ void celt_fir_c( opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch); #if !defined(OVERRIDE_CELT_FIR) -#define celt_fir(x, num, y, N, ord, mem, arch) \ - (celt_fir_c(x, num, y, N, ord, mem, arch)) +#define celt_fir(x, num, y, N, ord, arch) \ + (celt_fir_c(x, num, y, N, ord, arch)) #endif void celt_iir(const opus_val32 *x, diff --git a/libs/libopus/celt/cwrs.c b/libs/libopus/celt/cwrs.c index 9722f0ac8..a552e4f0f 100644 --- a/libs/libopus/celt/cwrs.c +++ b/libs/libopus/celt/cwrs.c @@ -482,7 +482,7 @@ static opus_val32 cwrsi(int _n,int _k,opus_uint32 _i,int *_y){ k0=_k; q=row[_n]; if(q>_i){ - celt_assert(p>q); + celt_sig_assert(p>q); _k=_n; do p=CELT_PVQ_U_ROW[--_k][_n]; while(p>_i); diff --git a/libs/libopus/celt/ecintrin.h b/libs/libopus/celt/ecintrin.h index 2263cff6b..66a4c36ea 100644 --- a/libs/libopus/celt/ecintrin.h +++ b/libs/libopus/celt/ecintrin.h @@ -49,7 +49,11 @@ This macro should only be used for implementing ec_ilog(), if it is defined. All other code should use EC_ILOG() instead.*/ #if defined(_MSC_VER) && (_MSC_VER >= 1400) +#if defined(_MSC_VER) && (_MSC_VER >= 1910) +# include <intrin0.h> /* Improve compiler throughput. */ +#else # include <intrin.h> +#endif /*In _DEBUG mode this is not an intrinsic by default.*/ # pragma intrinsic(_BitScanReverse) diff --git a/libs/libopus/celt/entcode.h b/libs/libopus/celt/entcode.h index 13d6c84ef..3763e3f28 100644 --- a/libs/libopus/celt/entcode.h +++ b/libs/libopus/celt/entcode.h @@ -122,7 +122,7 @@ opus_uint32 ec_tell_frac(ec_ctx *_this); /* Tested exhaustively for all n and for 1<=d<=256 */ static OPUS_INLINE opus_uint32 celt_udiv(opus_uint32 n, opus_uint32 d) { - celt_assert(d>0); + celt_sig_assert(d>0); #ifdef USE_SMALL_DIV_TABLE if (d>256) return n/d; @@ -138,7 +138,7 @@ static OPUS_INLINE opus_uint32 celt_udiv(opus_uint32 n, opus_uint32 d) { } static OPUS_INLINE opus_int32 celt_sudiv(opus_int32 n, opus_int32 d) { - celt_assert(d>0); + celt_sig_assert(d>0); #ifdef USE_SMALL_DIV_TABLE if (n<0) return -(opus_int32)celt_udiv(-n, d); diff --git a/libs/libopus/celt/entdec.h b/libs/libopus/celt/entdec.h index d8ab31873..025fc1870 100644 --- a/libs/libopus/celt/entdec.h +++ b/libs/libopus/celt/entdec.h @@ -85,7 +85,7 @@ int ec_dec_icdf(ec_dec *_this,const unsigned char *_icdf,unsigned _ftb); The bits must have been encoded with ec_enc_uint(). No call to ec_dec_update() is necessary after this call. _ft: The number of integers that can be decoded (one more than the max). - This must be at least one, and no more than 2**32-1. + This must be at least 2, and no more than 2**32-1. Return: The decoded bits.*/ opus_uint32 ec_dec_uint(ec_dec *_this,opus_uint32 _ft); diff --git a/libs/libopus/celt/entenc.h b/libs/libopus/celt/entenc.h index 796bc4d57..f502eaf66 100644 --- a/libs/libopus/celt/entenc.h +++ b/libs/libopus/celt/entenc.h @@ -67,7 +67,7 @@ void ec_enc_icdf(ec_enc *_this,int _s,const unsigned char *_icdf,unsigned _ftb); /*Encodes a raw unsigned integer in the stream. _fl: The integer to encode. _ft: The number of integers that can be encoded (one more than the max). - This must be at least one, and no more than 2**32-1.*/ + This must be at least 2, and no more than 2**32-1.*/ void ec_enc_uint(ec_enc *_this,opus_uint32 _fl,opus_uint32 _ft); /*Encodes a sequence of raw bits in the stream. diff --git a/libs/libopus/celt/fixed_debug.h b/libs/libopus/celt/fixed_debug.h index d28227f5d..3765baa60 100644 --- a/libs/libopus/celt/fixed_debug.h +++ b/libs/libopus/celt/fixed_debug.h @@ -59,6 +59,14 @@ extern opus_int64 celt_mips; #define SHR(a,b) SHR32(a,b) #define PSHR(a,b) PSHR32(a,b) +/** Add two 32-bit values, ignore any overflows */ +#define ADD32_ovflw(a,b) (celt_mips+=2,(opus_val32)((opus_uint32)(a)+(opus_uint32)(b))) +/** Subtract two 32-bit values, ignore any overflows */ +#define SUB32_ovflw(a,b) (celt_mips+=2,(opus_val32)((opus_uint32)(a)-(opus_uint32)(b))) +/* Avoid MSVC warning C4146: unary minus operator applied to unsigned type */ +/** Negate 32-bit value, ignore any overflows */ +#define NEG32_ovflw(a) (celt_mips+=2,(opus_val32)(0-(opus_uint32)(a))) + static OPUS_INLINE short NEG16(int x) { int res; @@ -227,12 +235,11 @@ static OPUS_INLINE int SHL32_(opus_int64 a, int shift, char *file, int line) #define VSHR32(a, shift) (((shift)>0) ? SHR32(a, shift) : SHL32(a, -(shift))) #define ROUND16(x,a) (celt_mips--,EXTRACT16(PSHR32((x),(a)))) +#define SROUND16(x,a) (celt_mips--,EXTRACT16(SATURATE(PSHR32(x,a), 32767))); + #define HALF16(x) (SHR16(x,1)) #define HALF32(x) (SHR32(x,1)) -//#define SHR(a,shift) ((a) >> (shift)) -//#define SHL(a,shift) ((a) << (shift)) - #define ADD16(a, b) ADD16_(a, b, __FILE__, __LINE__) static OPUS_INLINE short ADD16_(int a, int b, char *file, int line) { @@ -403,6 +410,51 @@ static OPUS_INLINE short MULT16_16_16(int a, int b) return res; } +/* result fits in 32 bits */ +static OPUS_INLINE int MULT32_32_32(opus_int64 a, opus_int64 b) +{ + opus_int64 res; + if (!VERIFY_INT(a) || !VERIFY_INT(b)) + { + fprintf (stderr, "MULT32_32_32: inputs are not int: %d %d\n", a, b); +#ifdef FIXED_DEBUG_ASSERT + celt_assert(0); +#endif + } + res = a*b; + if (!VERIFY_INT(res)) + { + fprintf (stderr, "MULT32_32_32: output is not int: %d\n", res); +#ifdef FIXED_DEBUG_ASSERT + celt_assert(0); +#endif + } + celt_mips+=5; + return res; +} + +static OPUS_INLINE int MULT32_32_Q16(opus_int64 a, opus_int64 b) +{ + opus_int64 res; + if (!VERIFY_INT(a) || !VERIFY_INT(b)) + { + fprintf (stderr, "MULT32_32_Q16: inputs are not int: %d %d\n", a, b); +#ifdef FIXED_DEBUG_ASSERT + celt_assert(0); +#endif + } + res = ((opus_int64)(a)*(opus_int64)(b)) >> 16; + if (!VERIFY_INT(res)) + { + fprintf (stderr, "MULT32_32_Q16: output is not int: %d*%d=%d\n", a, b, (int)res); +#ifdef FIXED_DEBUG_ASSERT + celt_assert(0); +#endif + } + celt_mips+=5; + return res; +} + #define MULT16_16(a, b) MULT16_16_(a, b, __FILE__, __LINE__) static OPUS_INLINE int MULT16_16_(int a, int b, char *file, int line) { diff --git a/libs/libopus/celt/fixed_generic.h b/libs/libopus/celt/fixed_generic.h index 1cfd6d698..8f29d46bb 100644 --- a/libs/libopus/celt/fixed_generic.h +++ b/libs/libopus/celt/fixed_generic.h @@ -57,6 +57,13 @@ #define MULT16_32_Q15(a,b) ADD32(SHL(MULT16_16((a),SHR((b),16)),1), SHR(MULT16_16SU((a),((b)&0x0000ffff)),15)) #endif +/** 32x32 multiplication, followed by a 16-bit shift right. Results fits in 32 bits */ +#if OPUS_FAST_INT64 +#define MULT32_32_Q16(a,b) ((opus_val32)SHR((opus_int64)(a)*(opus_int64)(b),16)) +#else +#define MULT32_32_Q16(a,b) (ADD32(ADD32(ADD32((opus_val32)(SHR32(((opus_uint32)((a)&0x0000ffff)*(opus_uint32)((b)&0x0000ffff)),16)), MULT16_16SU(SHR32(a,16),((b)&0x0000ffff))), MULT16_16SU(SHR32(b,16),((a)&0x0000ffff))), SHL32(MULT16_16(SHR32(a,16),SHR32(b,16)),16))) +#endif + /** 32x32 multiplication, followed by a 31-bit shift right. Results fits in 32 bits */ #if OPUS_FAST_INT64 #define MULT32_32_Q31(a,b) ((opus_val32)SHR((opus_int64)(a)*(opus_int64)(b),31)) @@ -102,8 +109,11 @@ #define SATURATE16(x) (EXTRACT16((x)>32767 ? 32767 : (x)<-32768 ? -32768 : (x))) -/** Shift by a and round-to-neareast 32-bit value. Result is a 16-bit value */ +/** Shift by a and round-to-nearest 32-bit value. Result is a 16-bit value */ #define ROUND16(x,a) (EXTRACT16(PSHR32((x),(a)))) +/** Shift by a and round-to-nearest 32-bit value. Result is a saturated 16-bit value */ +#define SROUND16(x,a) EXTRACT16(SATURATE(PSHR32(x,a), 32767)); + /** Divide by two */ #define HALF16(x) (SHR16(x,1)) #define HALF32(x) (SHR32(x,1)) @@ -117,9 +127,20 @@ /** Subtract two 32-bit values */ #define SUB32(a,b) ((opus_val32)(a)-(opus_val32)(b)) +/** Add two 32-bit values, ignore any overflows */ +#define ADD32_ovflw(a,b) ((opus_val32)((opus_uint32)(a)+(opus_uint32)(b))) +/** Subtract two 32-bit values, ignore any overflows */ +#define SUB32_ovflw(a,b) ((opus_val32)((opus_uint32)(a)-(opus_uint32)(b))) +/* Avoid MSVC warning C4146: unary minus operator applied to unsigned type */ +/** Negate 32-bit value, ignore any overflows */ +#define NEG32_ovflw(a) ((opus_val32)(0-(opus_uint32)(a))) + /** 16x16 multiplication where the result fits in 16 bits */ #define MULT16_16_16(a,b) ((((opus_val16)(a))*((opus_val16)(b)))) +/** 32x32 multiplication where the result fits in 32 bits */ +#define MULT32_32_32(a,b) ((((opus_val32)(a))*((opus_val32)(b)))) + /* (opus_val32)(opus_val16) gives TI compiler a hint that it's 16x16->32 multiply */ /** 16x16 multiplication where the result fits in 32 bits */ #define MULT16_16(a,b) (((opus_val32)(opus_val16)(a))*((opus_val32)(opus_val16)(b))) diff --git a/libs/libopus/celt/float_cast.h b/libs/libopus/celt/float_cast.h index ed5a39b54..9d34976ee 100644 --- a/libs/libopus/celt/float_cast.h +++ b/libs/libopus/celt/float_cast.h @@ -61,7 +61,45 @@ ** the config.h file. */ -#if (HAVE_LRINTF) +/* With GCC, when SSE is available, the fastest conversion is cvtss2si. */ +#if defined(__GNUC__) && defined(__SSE__) + +#include <xmmintrin.h> +static OPUS_INLINE opus_int32 float2int(float x) {return _mm_cvt_ss2si(_mm_set_ss(x));} + +#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && (defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)) + + #include <xmmintrin.h> + static OPUS_INLINE opus_int32 float2int(float value) + { + /* _mm_load_ss will generate same code as _mm_set_ss + ** in _MSC_VER >= 1914 /02 so keep __mm_load__ss + ** for backward compatibility. + */ + return _mm_cvtss_si32(_mm_load_ss(&value)); + } + +#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && defined (_M_IX86) + + #include <math.h> + + /* Win32 doesn't seem to have these functions. + ** Therefore implement OPUS_INLINE versions of these functions here. + */ + + static OPUS_INLINE opus_int32 + float2int (float flt) + { int intgr; + + _asm + { fld flt + fistp intgr + } ; + + return intgr ; + } + +#elif defined(HAVE_LRINTF) /* These defines enable functionality introduced with the 1999 ISO C ** standard. They must be defined before the inclusion of math.h to @@ -90,32 +128,6 @@ #include <math.h> #define float2int(x) lrint(x) -#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && defined (_M_X64) - #include <xmmintrin.h> - - __inline long int float2int(float value) - { - return _mm_cvtss_si32(_mm_load_ss(&value)); - } -#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && defined (_M_IX86) - #include <math.h> - - /* Win32 doesn't seem to have these functions. - ** Therefore implement OPUS_INLINE versions of these functions here. - */ - - __inline long int - float2int (float flt) - { int intgr; - - _asm - { fld flt - fistp intgr - } ; - - return intgr ; - } - #else #if (defined(__GNUC__) && defined(__STDC__) && __STDC__ && __STDC_VERSION__ >= 199901L) diff --git a/libs/libopus/celt/kiss_fft.c b/libs/libopus/celt/kiss_fft.c index 1f8fd0532..83775165d 100644 --- a/libs/libopus/celt/kiss_fft.c +++ b/libs/libopus/celt/kiss_fft.c @@ -82,8 +82,8 @@ static void kf_bfly2( C_SUB( Fout2[0] , Fout[0] , t ); C_ADDTO( Fout[0] , t ); - t.r = S_MUL(Fout2[1].r+Fout2[1].i, tw); - t.i = S_MUL(Fout2[1].i-Fout2[1].r, tw); + t.r = S_MUL(ADD32_ovflw(Fout2[1].r, Fout2[1].i), tw); + t.i = S_MUL(SUB32_ovflw(Fout2[1].i, Fout2[1].r), tw); C_SUB( Fout2[1] , Fout[1] , t ); C_ADDTO( Fout[1] , t ); @@ -92,8 +92,8 @@ static void kf_bfly2( C_SUB( Fout2[2] , Fout[2] , t ); C_ADDTO( Fout[2] , t ); - t.r = S_MUL(Fout2[3].i-Fout2[3].r, tw); - t.i = S_MUL(-Fout2[3].i-Fout2[3].r, tw); + t.r = S_MUL(SUB32_ovflw(Fout2[3].i, Fout2[3].r), tw); + t.i = S_MUL(NEG32_ovflw(ADD32_ovflw(Fout2[3].i, Fout2[3].r)), tw); C_SUB( Fout2[3] , Fout[3] , t ); C_ADDTO( Fout[3] , t ); Fout += 8; @@ -126,10 +126,10 @@ static void kf_bfly4( C_ADDTO( *Fout , scratch1 ); C_SUB( scratch1 , Fout[1] , Fout[3] ); - Fout[1].r = scratch0.r + scratch1.i; - Fout[1].i = scratch0.i - scratch1.r; - Fout[3].r = scratch0.r - scratch1.i; - Fout[3].i = scratch0.i + scratch1.r; + Fout[1].r = ADD32_ovflw(scratch0.r, scratch1.i); + Fout[1].i = SUB32_ovflw(scratch0.i, scratch1.r); + Fout[3].r = SUB32_ovflw(scratch0.r, scratch1.i); + Fout[3].i = ADD32_ovflw(scratch0.i, scratch1.r); Fout+=4; } } else { @@ -160,10 +160,10 @@ static void kf_bfly4( tw3 += fstride*3; C_ADDTO( *Fout , scratch[3] ); - Fout[m].r = scratch[5].r + scratch[4].i; - Fout[m].i = scratch[5].i - scratch[4].r; - Fout[m3].r = scratch[5].r - scratch[4].i; - Fout[m3].i = scratch[5].i + scratch[4].r; + Fout[m].r = ADD32_ovflw(scratch[5].r, scratch[4].i); + Fout[m].i = SUB32_ovflw(scratch[5].i, scratch[4].r); + Fout[m3].r = SUB32_ovflw(scratch[5].r, scratch[4].i); + Fout[m3].i = ADD32_ovflw(scratch[5].i, scratch[4].r); ++Fout; } } @@ -212,18 +212,18 @@ static void kf_bfly3( tw1 += fstride; tw2 += fstride*2; - Fout[m].r = Fout->r - HALF_OF(scratch[3].r); - Fout[m].i = Fout->i - HALF_OF(scratch[3].i); + Fout[m].r = SUB32_ovflw(Fout->r, HALF_OF(scratch[3].r)); + Fout[m].i = SUB32_ovflw(Fout->i, HALF_OF(scratch[3].i)); C_MULBYSCALAR( scratch[0] , epi3.i ); C_ADDTO(*Fout,scratch[3]); - Fout[m2].r = Fout[m].r + scratch[0].i; - Fout[m2].i = Fout[m].i - scratch[0].r; + Fout[m2].r = ADD32_ovflw(Fout[m].r, scratch[0].i); + Fout[m2].i = SUB32_ovflw(Fout[m].i, scratch[0].r); - Fout[m].r -= scratch[0].i; - Fout[m].i += scratch[0].r; + Fout[m].r = SUB32_ovflw(Fout[m].r, scratch[0].i); + Fout[m].i = ADD32_ovflw(Fout[m].i, scratch[0].r); ++Fout; } while(--k); @@ -282,22 +282,22 @@ static void kf_bfly5( C_ADD( scratch[8],scratch[2],scratch[3]); C_SUB( scratch[9],scratch[2],scratch[3]); - Fout0->r += scratch[7].r + scratch[8].r; - Fout0->i += scratch[7].i + scratch[8].i; + Fout0->r = ADD32_ovflw(Fout0->r, ADD32_ovflw(scratch[7].r, scratch[8].r)); + Fout0->i = ADD32_ovflw(Fout0->i, ADD32_ovflw(scratch[7].i, scratch[8].i)); - scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r); - scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r); + scratch[5].r = ADD32_ovflw(scratch[0].r, ADD32_ovflw(S_MUL(scratch[7].r,ya.r), S_MUL(scratch[8].r,yb.r))); + scratch[5].i = ADD32_ovflw(scratch[0].i, ADD32_ovflw(S_MUL(scratch[7].i,ya.r), S_MUL(scratch[8].i,yb.r))); - scratch[6].r = S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i); - scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i); + scratch[6].r = ADD32_ovflw(S_MUL(scratch[10].i,ya.i), S_MUL(scratch[9].i,yb.i)); + scratch[6].i = NEG32_ovflw(ADD32_ovflw(S_MUL(scratch[10].r,ya.i), S_MUL(scratch[9].r,yb.i))); C_SUB(*Fout1,scratch[5],scratch[6]); C_ADD(*Fout4,scratch[5],scratch[6]); - scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r); - scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r); - scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i); - scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i); + scratch[11].r = ADD32_ovflw(scratch[0].r, ADD32_ovflw(S_MUL(scratch[7].r,yb.r), S_MUL(scratch[8].r,ya.r))); + scratch[11].i = ADD32_ovflw(scratch[0].i, ADD32_ovflw(S_MUL(scratch[7].i,yb.r), S_MUL(scratch[8].i,ya.r))); + scratch[12].r = SUB32_ovflw(S_MUL(scratch[9].i,ya.i), S_MUL(scratch[10].i,yb.i)); + scratch[12].i = SUB32_ovflw(S_MUL(scratch[10].r,yb.i), S_MUL(scratch[9].r,ya.i)); C_ADD(*Fout2,scratch[11],scratch[12]); C_SUB(*Fout3,scratch[11],scratch[12]); diff --git a/libs/libopus/celt/mathops.c b/libs/libopus/celt/mathops.c index 21a01f52e..6ee9b9e10 100644 --- a/libs/libopus/celt/mathops.c +++ b/libs/libopus/celt/mathops.c @@ -38,7 +38,8 @@ #include "mathops.h" /*Compute floor(sqrt(_val)) with exact arithmetic. - This has been tested on all possible 32-bit inputs.*/ + _val must be greater than 0. + This has been tested on all possible 32-bit inputs greater than 0.*/ unsigned isqrt32(opus_uint32 _val){ unsigned b; unsigned g; @@ -182,7 +183,7 @@ opus_val32 celt_rcp(opus_val32 x) int i; opus_val16 n; opus_val16 r; - celt_assert2(x>0, "celt_rcp() only defined for positive values"); + celt_sig_assert(x>0); i = celt_ilog2(x); /* n is Q15 with range [0,1). */ n = VSHR32(x,i-15)-32768; diff --git a/libs/libopus/celt/mathops.h b/libs/libopus/celt/mathops.h index a0525a961..fe29dac1c 100644 --- a/libs/libopus/celt/mathops.h +++ b/libs/libopus/celt/mathops.h @@ -38,11 +38,44 @@ #include "entcode.h" #include "os_support.h" +#define PI 3.141592653f + /* Multiplies two 16-bit fractional values. Bit-exactness of this macro is important */ #define FRAC_MUL16(a,b) ((16384+((opus_int32)(opus_int16)(a)*(opus_int16)(b)))>>15) unsigned isqrt32(opus_uint32 _val); +/* CELT doesn't need it for fixed-point, by analysis.c does. */ +#if !defined(FIXED_POINT) || defined(ANALYSIS_C) +#define cA 0.43157974f +#define cB 0.67848403f +#define cC 0.08595542f +#define cE ((float)PI/2) +static OPUS_INLINE float fast_atan2f(float y, float x) { + float x2, y2; + x2 = x*x; + y2 = y*y; + /* For very small values, we don't care about the answer, so + we can just return 0. */ + if (x2 + y2 < 1e-18f) + { + return 0; + } + if(x2<y2){ + float den = (y2 + cB*x2) * (y2 + cC*x2); + return -x*y*(y2 + cA*x2) / den + (y<0 ? -cE : cE); + }else{ + float den = (x2 + cB*y2) * (x2 + cC*y2); + return x*y*(x2 + cA*y2) / den + (y<0 ? -cE : cE) - (x*y<0 ? -cE : cE); + } +} +#undef cA +#undef cB +#undef cC +#undef cE +#endif + + #ifndef OVERRIDE_CELT_MAXABS16 static OPUS_INLINE opus_val32 celt_maxabs16(const opus_val16 *x, int len) { @@ -80,7 +113,6 @@ static OPUS_INLINE opus_val32 celt_maxabs32(const opus_val32 *x, int len) #ifndef FIXED_POINT -#define PI 3.141592653f #define celt_sqrt(x) ((float)sqrt(x)) #define celt_rsqrt(x) (1.f/celt_sqrt(x)) #define celt_rsqrt_norm(x) (celt_rsqrt(x)) @@ -105,7 +137,7 @@ static OPUS_INLINE float celt_log2(float x) } in; in.f = x; integer = (in.i>>23)-127; - in.i -= integer<<23; + in.i -= (opus_uint32)integer<<23; frac = in.f - 1.5f; frac = -0.41445418f + frac*(0.95909232f + frac*(-0.33951290f + frac*0.16541097f)); @@ -128,7 +160,7 @@ static OPUS_INLINE float celt_exp2(float x) /* K0 = 1, K1 = log(2), K2 = 3-4*log(2), K3 = 3*log(2) - 2 */ res.f = 0.99992522f + frac * (0.69583354f + frac * (0.22606716f + 0.078024523f*frac)); - res.i = (res.i + (integer<<23)) & 0x7fffffff; + res.i = (res.i + ((opus_uint32)integer<<23)) & 0x7fffffff; return res.f; } @@ -147,7 +179,7 @@ static OPUS_INLINE float celt_exp2(float x) /** Integer log in base2. Undefined for zero and negative numbers */ static OPUS_INLINE opus_int16 celt_ilog2(opus_int32 x) { - celt_assert2(x>0, "celt_ilog2() only defined for strictly positive numbers"); + celt_sig_assert(x>0); return EC_ILOG(x)-1; } #endif diff --git a/libs/libopus/celt/mdct.c b/libs/libopus/celt/mdct.c index 5315ad11a..5c6dab5b7 100644 --- a/libs/libopus/celt/mdct.c +++ b/libs/libopus/celt/mdct.c @@ -270,8 +270,8 @@ void clt_mdct_backward_c(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_sca int rev; kiss_fft_scalar yr, yi; rev = *bitrev++; - yr = S_MUL(*xp2, t[i]) + S_MUL(*xp1, t[N4+i]); - yi = S_MUL(*xp1, t[i]) - S_MUL(*xp2, t[N4+i]); + yr = ADD32_ovflw(S_MUL(*xp2, t[i]), S_MUL(*xp1, t[N4+i])); + yi = SUB32_ovflw(S_MUL(*xp1, t[i]), S_MUL(*xp2, t[N4+i])); /* We swap real and imag because we use an FFT instead of an IFFT. */ yp[2*rev+1] = yr; yp[2*rev] = yi; @@ -301,8 +301,8 @@ void clt_mdct_backward_c(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_sca t0 = t[i]; t1 = t[N4+i]; /* We'd scale up by 2 here, but instead it's done when mixing the windows */ - yr = S_MUL(re,t0) + S_MUL(im,t1); - yi = S_MUL(re,t1) - S_MUL(im,t0); + yr = ADD32_ovflw(S_MUL(re,t0), S_MUL(im,t1)); + yi = SUB32_ovflw(S_MUL(re,t1), S_MUL(im,t0)); /* We swap real and imag because we're using an FFT instead of an IFFT. */ re = yp1[1]; im = yp1[0]; @@ -312,8 +312,8 @@ void clt_mdct_backward_c(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_sca t0 = t[(N4-i-1)]; t1 = t[(N2-i-1)]; /* We'd scale up by 2 here, but instead it's done when mixing the windows */ - yr = S_MUL(re,t0) + S_MUL(im,t1); - yi = S_MUL(re,t1) - S_MUL(im,t0); + yr = ADD32_ovflw(S_MUL(re,t0), S_MUL(im,t1)); + yi = SUB32_ovflw(S_MUL(re,t1), S_MUL(im,t0)); yp1[0] = yr; yp0[1] = yi; yp0 += 2; @@ -333,8 +333,8 @@ void clt_mdct_backward_c(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_sca kiss_fft_scalar x1, x2; x1 = *xp1; x2 = *yp1; - *yp1++ = MULT16_32_Q15(*wp2, x2) - MULT16_32_Q15(*wp1, x1); - *xp1-- = MULT16_32_Q15(*wp1, x2) + MULT16_32_Q15(*wp2, x1); + *yp1++ = SUB32_ovflw(MULT16_32_Q15(*wp2, x2), MULT16_32_Q15(*wp1, x1)); + *xp1-- = ADD32_ovflw(MULT16_32_Q15(*wp1, x2), MULT16_32_Q15(*wp2, x1)); wp1++; wp2--; } diff --git a/libs/libopus/celt/mips/celt_mipsr1.h b/libs/libopus/celt/mips/celt_mipsr1.h index e85661a66..c332fe047 100644 --- a/libs/libopus/celt/mips/celt_mipsr1.h +++ b/libs/libopus/celt/mips/celt_mipsr1.h @@ -53,6 +53,7 @@ #include "celt_lpc.h" #include "vq.h" +#define OVERRIDE_COMB_FILTER_CONST #define OVERRIDE_comb_filter void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N, opus_val16 g0, opus_val16 g1, int tapset0, int tapset1, diff --git a/libs/libopus/celt/mips/vq_mipsr1.h b/libs/libopus/celt/mips/vq_mipsr1.h index 54cef8613..f26a33e75 100644 --- a/libs/libopus/celt/mips/vq_mipsr1.h +++ b/libs/libopus/celt/mips/vq_mipsr1.h @@ -36,11 +36,6 @@ #include "mathops.h" #include "arch.h" -static unsigned extract_collapse_mask(int *iy, int N, int B); -static void normalise_residual(int * OPUS_RESTRICT iy, celt_norm * OPUS_RESTRICT X, int N, opus_val32 Ryy, opus_val16 gain); -static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread); -static void renormalise_vector_mips(celt_norm *X, int N, opus_val16 gain, int arch); - #define OVERRIDE_vq_exp_rotation1 static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s) { @@ -69,11 +64,7 @@ static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_ } #define OVERRIDE_renormalise_vector - -#define renormalise_vector(X, N, gain, arch) \ - (renormalise_vector_mips(X, N, gain, arch)) - -void renormalise_vector_mips(celt_norm *X, int N, opus_val16 gain, int arch) +void renormalise_vector(celt_norm *X, int N, opus_val16 gain, int arch) { int i; #ifdef FIXED_POINT diff --git a/libs/libopus/celt/modes.c b/libs/libopus/celt/modes.c index 911686e90..390c5e8ae 100644 --- a/libs/libopus/celt/modes.c +++ b/libs/libopus/celt/modes.c @@ -427,7 +427,7 @@ void opus_custom_mode_destroy(CELTMode *mode) } #endif /* CUSTOM_MODES_ONLY */ opus_free((opus_int16*)mode->eBands); - opus_free((opus_int16*)mode->allocVectors); + opus_free((unsigned char*)mode->allocVectors); opus_free((opus_val16*)mode->window); opus_free((opus_int16*)mode->logN); diff --git a/libs/libopus/celt/os_support.h b/libs/libopus/celt/os_support.h index a2171971e..009bf861d 100644 --- a/libs/libopus/celt/os_support.h +++ b/libs/libopus/celt/os_support.h @@ -39,7 +39,6 @@ #include "opus_defines.h" #include <string.h> -#include <stdio.h> #include <stdlib.h> /** Opus wrapper for malloc(). To do your own dynamic allocation, all you need to do is replace this function and opus_free */ diff --git a/libs/libopus/celt/pitch.c b/libs/libopus/celt/pitch.c index bf46e7d56..872582a48 100644 --- a/libs/libopus/celt/pitch.c +++ b/libs/libopus/celt/pitch.c @@ -102,11 +102,9 @@ static void find_best_pitch(opus_val32 *xcorr, opus_val16 *y, int len, } } -static void celt_fir5(const opus_val16 *x, +static void celt_fir5(opus_val16 *x, const opus_val16 *num, - opus_val16 *y, - int N, - opus_val16 *mem) + int N) { int i; opus_val16 num0, num1, num2, num3, num4; @@ -116,11 +114,11 @@ static void celt_fir5(const opus_val16 *x, num2=num[2]; num3=num[3]; num4=num[4]; - mem0=mem[0]; - mem1=mem[1]; - mem2=mem[2]; - mem3=mem[3]; - mem4=mem[4]; + mem0=0; + mem1=0; + mem2=0; + mem3=0; + mem4=0; for (i=0;i<N;i++) { opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT); @@ -134,13 +132,8 @@ static void celt_fir5(const opus_val16 *x, mem2 = mem1; mem1 = mem0; mem0 = x[i]; - y[i] = ROUND16(sum, SIG_SHIFT); + x[i] = ROUND16(sum, SIG_SHIFT); } - mem[0]=mem0; - mem[1]=mem1; - mem[2]=mem2; - mem[3]=mem3; - mem[4]=mem4; } @@ -150,7 +143,7 @@ void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x int i; opus_val32 ac[5]; opus_val16 tmp=Q15ONE; - opus_val16 lpc[4], mem[5]={0,0,0,0,0}; + opus_val16 lpc[4]; opus_val16 lpc2[5]; opus_val16 c1 = QCONST16(.8f,15); #ifdef FIXED_POINT @@ -211,7 +204,7 @@ void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x lpc2[2] = lpc[2] + MULT16_16_Q15(c1,lpc[1]); lpc2[3] = lpc[3] + MULT16_16_Q15(c1,lpc[2]); lpc2[4] = MULT16_16_Q15(c1,lpc[3]); - celt_fir5(x_lp, lpc2, x_lp, len>>1, mem); + celt_fir5(x_lp, lpc2, len>>1); } /* Pure C implementation. */ @@ -220,13 +213,8 @@ opus_val32 #else void #endif -#if defined(OVERRIDE_PITCH_XCORR) celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch) -#else -celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y, opus_val32 *xcorr, int len, int max_pitch, int arch) -#endif { #if 0 /* This is a simple version of the pitch correlation that should work @@ -261,15 +249,11 @@ celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y, opus_val32 maxcorr=1; #endif celt_assert(max_pitch>0); - celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0); + celt_sig_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0); for (i=0;i<max_pitch-3;i+=4) { opus_val32 sum[4]={0,0,0,0}; -#if defined(OVERRIDE_PITCH_XCORR) - xcorr_kernel_c(_x, _y+i, sum, len); -#else xcorr_kernel(_x, _y+i, sum, len, arch); -#endif xcorr[i]=sum[0]; xcorr[i+1]=sum[1]; xcorr[i+2]=sum[2]; @@ -285,11 +269,7 @@ celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y, for (;i<max_pitch;i++) { opus_val32 sum; -#if defined(OVERRIDE_PITCH_XCORR) - sum = celt_inner_prod_c(_x, _y+i, len); -#else sum = celt_inner_prod(_x, _y+i, len, arch); -#endif xcorr[i] = sum; #ifdef FIXED_POINT maxcorr = MAX32(maxcorr, sum); @@ -378,7 +358,7 @@ void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTR for (j=0;j<len>>1;j++) sum += SHR32(MULT16_16(x_lp[j],y[i+j]), shift); #else - sum = celt_inner_prod_c(x_lp, y+i, len>>1); + sum = celt_inner_prod(x_lp, y+i, len>>1, arch); #endif xcorr[i] = MAX32(-1, sum); #ifdef FIXED_POINT @@ -424,7 +404,7 @@ static opus_val16 compute_pitch_gain(opus_val32 xy, opus_val32 xx, opus_val32 yy sx = celt_ilog2(xx)-14; sy = celt_ilog2(yy)-14; shift = sx + sy; - x2y2 = MULT16_16_Q14(VSHR32(xx, sx), VSHR32(yy, sy)); + x2y2 = SHR32(MULT16_16(VSHR32(xx, sx), VSHR32(yy, sy)), 14); if (shift & 1) { if (x2y2 < 32768) { diff --git a/libs/libopus/celt/pitch.h b/libs/libopus/celt/pitch.h index d3503532a..e425f56ae 100644 --- a/libs/libopus/celt/pitch.h +++ b/libs/libopus/celt/pitch.h @@ -46,8 +46,7 @@ #include "mips/pitch_mipsr1.h" #endif -#if ((defined(OPUS_ARM_ASM) && defined(FIXED_POINT)) \ - || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) +#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR)) # include "arm/pitch_arm.h" #endif @@ -184,17 +183,10 @@ opus_val32 void #endif celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y, - opus_val32 *xcorr, int len, int max_pitch); - -#if !defined(OVERRIDE_PITCH_XCORR) -#ifdef FIXED_POINT -opus_val32 -#else -void -#endif -celt_pitch_xcorr(const opus_val16 *_x, const opus_val16 *_y, opus_val32 *xcorr, int len, int max_pitch, int arch); +#ifndef OVERRIDE_PITCH_XCORR +# define celt_pitch_xcorr celt_pitch_xcorr_c #endif #endif diff --git a/libs/libopus/celt/quant_bands.c b/libs/libopus/celt/quant_bands.c index 95076e0af..39a221eda 100644 --- a/libs/libopus/celt/quant_bands.c +++ b/libs/libopus/celt/quant_bands.c @@ -418,6 +418,7 @@ void quant_energy_finalise(const CELTMode *m, int start, int end, opus_val16 *ol offset = (q2-.5f)*(1<<(14-fine_quant[i]-1))*(1.f/16384); #endif oldEBands[i+c*m->nbEBands] += offset; + error[i+c*m->nbEBands] -= offset; bits_left--; } while (++c < C); } @@ -456,7 +457,7 @@ void unquant_coarse_energy(const CELTMode *m, int start, int end, opus_val16 *ol /* It would be better to express this invariant as a test on C at function entry, but that isn't enough to make the static analyzer happy. */ - celt_assert(c<2); + celt_sig_assert(c<2); tell = ec_tell(dec); if(budget-tell>=15) { @@ -547,9 +548,15 @@ void amp2Log2(const CELTMode *m, int effEnd, int end, c=0; do { for (i=0;i<effEnd;i++) + { bandLogE[i+c*m->nbEBands] = - celt_log2(SHL32(bandE[i+c*m->nbEBands],2)) + celt_log2(bandE[i+c*m->nbEBands]) - SHL16((opus_val16)eMeans[i],6); +#ifdef FIXED_POINT + /* Compensate for bandE[] being Q12 but celt_log2() taking a Q14 input. */ + bandLogE[i+c*m->nbEBands] += QCONST16(2.f, DB_SHIFT); +#endif + } for (i=effEnd;i<end;i++) bandLogE[c*m->nbEBands+i] = -QCONST16(14.f,DB_SHIFT); } while (++c < C); diff --git a/libs/libopus/celt/rate.c b/libs/libopus/celt/rate.c index 9062ff750..465e1ba26 100644 --- a/libs/libopus/celt/rate.c +++ b/libs/libopus/celt/rate.c @@ -348,12 +348,17 @@ static OPUS_INLINE int interp_bits2pulses(const CELTMode *m, int start, int end, /*This if() block is the only part of the allocation function that is not a mandatory part of the bitstream: any bands we choose to skip here must be explicitly signaled.*/ - /*Choose a threshold with some hysteresis to keep bands from - fluctuating in and out.*/ + int depth_threshold; + /*We choose a threshold with some hysteresis to keep bands from + fluctuating in and out, but we try not to fold below a certain point. */ + if (codedBands > 17) + depth_threshold = j<prev ? 7 : 9; + else + depth_threshold = 0; #ifdef FUZZING if ((rand()&0x1) == 0) #else - if (codedBands<=start+2 || (band_bits > ((j<prev?7:9)*band_width<<LM<<BITRES)>>4 && j<=signalBandwidth)) + if (codedBands<=start+2 || (band_bits > (depth_threshold*band_width<<LM<<BITRES)>>4 && j<=signalBandwidth)) #endif { ec_enc_bit_logp(ec, 1, 1); @@ -524,7 +529,7 @@ static OPUS_INLINE int interp_bits2pulses(const CELTMode *m, int start, int end, return codedBands; } -int compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stereo, +int clt_compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stereo, opus_int32 total, opus_int32 *balance, int *pulses, int *ebits, int *fine_priority, int C, int LM, ec_ctx *ec, int encode, int prev, int signalBandwidth) { int lo, hi, len, j; @@ -636,3 +641,4 @@ int compute_allocation(const CELTMode *m, int start, int end, const int *offsets RESTORE_STACK; return codedBands; } + diff --git a/libs/libopus/celt/rate.h b/libs/libopus/celt/rate.h index 515f7687c..fad5e412d 100644 --- a/libs/libopus/celt/rate.h +++ b/libs/libopus/celt/rate.h @@ -95,7 +95,7 @@ static OPUS_INLINE int pulses2bits(const CELTMode *m, int band, int LM, int puls @param pulses Number of pulses per band (returned) @return Total number of bits allocated */ -int compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stero, +int clt_compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stereo, opus_int32 total, opus_int32 *balance, int *pulses, int *ebits, int *fine_priority, int C, int LM, ec_ctx *ec, int encode, int prev, int signalBandwidth); #endif diff --git a/libs/libopus/celt/stack_alloc.h b/libs/libopus/celt/stack_alloc.h index 2b51c8d80..ae40e2a16 100644 --- a/libs/libopus/celt/stack_alloc.h +++ b/libs/libopus/celt/stack_alloc.h @@ -40,7 +40,7 @@ #endif #ifdef USE_ALLOCA -# ifdef WIN32 +# ifdef _WIN32 # include <malloc.h> # else # ifdef HAVE_ALLOCA_H @@ -102,7 +102,7 @@ #define VARDECL(type, var) type *var -# ifdef WIN32 +# ifdef _WIN32 # define ALLOC(var, size, type) var = ((type*)_alloca(sizeof(type)*(size))) # else # define ALLOC(var, size, type) var = ((type*)alloca(sizeof(type)*(size))) diff --git a/libs/libopus/celt/static_modes_fixed_arm_ne10.h b/libs/libopus/celt/static_modes_fixed_arm_ne10.h index b8ef0cee9..762309219 100644 --- a/libs/libopus/celt/static_modes_fixed_arm_ne10.h +++ b/libs/libopus/celt/static_modes_fixed_arm_ne10.h @@ -1,7 +1,7 @@ /* The contents of this file was automatically generated by * dump_mode_arm_ne10.c with arguments: 48000 960 * It contains static definitions for some pre-defined modes. */ -#include <NE10_init.h> +#include <NE10_types.h> #ifndef NE10_FFT_PARAMS48000_960 #define NE10_FFT_PARAMS48000_960 diff --git a/libs/libopus/celt/static_modes_float_arm_ne10.h b/libs/libopus/celt/static_modes_float_arm_ne10.h index 934a82a42..66e1abb10 100644 --- a/libs/libopus/celt/static_modes_float_arm_ne10.h +++ b/libs/libopus/celt/static_modes_float_arm_ne10.h @@ -1,7 +1,7 @@ /* The contents of this file was automatically generated by * dump_mode_arm_ne10.c with arguments: 48000 960 * It contains static definitions for some pre-defined modes. */ -#include <NE10_init.h> +#include <NE10_types.h> #ifndef NE10_FFT_PARAMS48000_960 #define NE10_FFT_PARAMS48000_960 diff --git a/libs/libopus/celt/vq.c b/libs/libopus/celt/vq.c index d29f38fd8..8011e2254 100644 --- a/libs/libopus/celt/vq.c +++ b/libs/libopus/celt/vq.c @@ -39,6 +39,10 @@ #include "rate.h" #include "pitch.h" +#if defined(MIPSr1_ASM) +#include "mips/vq_mipsr1.h" +#endif + #ifndef OVERRIDE_vq_exp_rotation1 static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s) { @@ -67,7 +71,7 @@ static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_ } #endif /* OVERRIDE_vq_exp_rotation1 */ -static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread) +void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread) { static const int SPREAD_FACTOR[3]={15,10,5}; int i; @@ -158,42 +162,27 @@ static unsigned extract_collapse_mask(int *iy, int N, int B) return collapse_mask; } -unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc -#ifdef RESYNTH - , opus_val16 gain -#endif - ) +opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch) { VARDECL(celt_norm, y); - VARDECL(int, iy); - VARDECL(opus_val16, signx); + VARDECL(int, signx); int i, j; - opus_val16 s; int pulsesLeft; opus_val32 sum; opus_val32 xy; opus_val16 yy; - unsigned collapse_mask; SAVE_STACK; - celt_assert2(K>0, "alg_quant() needs at least one pulse"); - celt_assert2(N>1, "alg_quant() needs at least two dimensions"); - + (void)arch; ALLOC(y, N, celt_norm); - ALLOC(iy, N, int); - ALLOC(signx, N, opus_val16); - - exp_rotation(X, N, 1, B, K, spread); + ALLOC(signx, N, int); /* Get rid of the sign */ sum = 0; j=0; do { - if (X[j]>0) - signx[j]=1; - else { - signx[j]=-1; - X[j]=-X[j]; - } + signx[j] = X[j]<0; + /* OPT: Make sure the compiler doesn't use a branch on ABS16(). */ + X[j] = ABS16(X[j]); iy[j] = 0; y[j] = 0; } while (++j<N); @@ -225,7 +214,12 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc while (++j<N); sum = QCONST16(1.f,14); } - rcp = EXTRACT16(MULT16_32_Q16(K-1, celt_rcp(sum))); +#ifdef FIXED_POINT + rcp = EXTRACT16(MULT16_32_Q16(K, celt_rcp(sum))); +#else + /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */ + rcp = EXTRACT16(MULT16_32_Q16(K+0.8f, celt_rcp(sum))); +#endif j=0; do { #ifdef FIXED_POINT /* It's really important to round *towards zero* here */ @@ -240,12 +234,12 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc pulsesLeft -= iy[j]; } while (++j<N); } - celt_assert2(pulsesLeft>=1, "Allocated too many pulses in the quick pass"); + celt_sig_assert(pulsesLeft>=0); /* This should never happen, but just in case it does (e.g. on silence) we fill the first bin with pulses. */ #ifdef FIXED_POINT_DEBUG - celt_assert2(pulsesLeft<=N+3, "Not enough pulses in the quick pass"); + celt_sig_assert(pulsesLeft<=N+3); #endif if (pulsesLeft > N+3) { @@ -256,12 +250,12 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc pulsesLeft=0; } - s = 1; for (i=0;i<pulsesLeft;i++) { + opus_val16 Rxy, Ryy; int best_id; - opus_val32 best_num = -VERY_LARGE16; - opus_val16 best_den = 0; + opus_val32 best_num; + opus_val16 best_den; #ifdef FIXED_POINT int rshift; #endif @@ -272,9 +266,22 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc /* The squared magnitude term gets added anyway, so we might as well add it outside the loop */ yy = ADD16(yy, 1); - j=0; + + /* Calculations for position 0 are out of the loop, in part to reduce + mispredicted branches (since the if condition is usually false) + in the loop. */ + /* Temporary sums of the new pulse(s) */ + Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[0])),rshift)); + /* We're multiplying y[j] by two so we don't have to do it here */ + Ryy = ADD16(yy, y[0]); + + /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that + Rxy is positive because the sign is pre-computed) */ + Rxy = MULT16_16_Q15(Rxy,Rxy); + best_den = Ryy; + best_num = Rxy; + j=1; do { - opus_val16 Rxy, Ryy; /* Temporary sums of the new pulse(s) */ Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift)); /* We're multiplying y[j] by two so we don't have to do it here */ @@ -285,8 +292,11 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc Rxy = MULT16_16_Q15(Rxy,Rxy); /* The idea is to check for num/den >= best_num/best_den, but that way we can do it without any division */ - /* OPT: Make sure to use conditional moves here */ - if (MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num)) + /* OPT: It's not clear whether a cmov is faster than a branch here + since the condition is more often false than true and using + a cmov introduces data dependencies across iterations. The optimal + choice may be architecture-dependent. */ + if (opus_unlikely(MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num))) { best_den = Ryy; best_num = Rxy; @@ -301,23 +311,47 @@ unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc /* Only now that we've made the final choice, update y/iy */ /* Multiplying y[j] by 2 so we don't have to do it everywhere else */ - y[best_id] += 2*s; + y[best_id] += 2; iy[best_id]++; } /* Put the original sign back */ j=0; do { - X[j] = MULT16_16(signx[j],X[j]); - if (signx[j] < 0) - iy[j] = -iy[j]; + /*iy[j] = signx[j] ? -iy[j] : iy[j];*/ + /* OPT: The is more likely to be compiled without a branch than the code above + but has the same performance otherwise. */ + iy[j] = (iy[j]^-signx[j]) + signx[j]; } while (++j<N); + RESTORE_STACK; + return yy; +} + +unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc, + opus_val16 gain, int resynth, int arch) +{ + VARDECL(int, iy); + opus_val16 yy; + unsigned collapse_mask; + SAVE_STACK; + + celt_assert2(K>0, "alg_quant() needs at least one pulse"); + celt_assert2(N>1, "alg_quant() needs at least two dimensions"); + + /* Covers vectorization by up to 4. */ + ALLOC(iy, N+3, int); + + exp_rotation(X, N, 1, B, K, spread); + + yy = op_pvq_search(X, iy, K, N, arch); + encode_pulses(iy, N, K, enc); -#ifdef RESYNTH - normalise_residual(iy, X, N, yy, gain); - exp_rotation(X, N, -1, B, K, spread); -#endif + if (resynth) + { + normalise_residual(iy, X, N, yy, gain); + exp_rotation(X, N, -1, B, K, spread); + } collapse_mask = extract_collapse_mask(iy, N, B); RESTORE_STACK; @@ -401,7 +435,7 @@ int stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N, int /* 0.63662 = 2/pi */ itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid)); #else - itheta = (int)floor(.5f+16384*0.63662f*atan2(side,mid)); + itheta = (int)floor(.5f+16384*0.63662f*fast_atan2f(side,mid)); #endif return itheta; diff --git a/libs/libopus/celt/vq.h b/libs/libopus/celt/vq.h index 5cfcbe50e..45ec55918 100644 --- a/libs/libopus/celt/vq.h +++ b/libs/libopus/celt/vq.h @@ -37,10 +37,18 @@ #include "entdec.h" #include "modes.h" -#if defined(MIPSr1_ASM) -#include "mips/vq_mipsr1.h" +#if (defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT)) +#include "x86/vq_sse.h" #endif +void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread); + +opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch); + +#if !defined(OVERRIDE_OP_PVQ_SEARCH) +#define op_pvq_search(x, iy, K, N, arch) \ + (op_pvq_search_c(x, iy, K, N, arch)) +#endif /** Algebraic pulse-vector quantiser. The signal x is replaced by the sum of * the pitch and a combination of pulses such that its norm is still equal @@ -51,12 +59,8 @@ * @param enc Entropy encoder state * @ret A mask indicating which blocks in the band received pulses */ -unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, - ec_enc *enc -#ifdef RESYNTH - , opus_val16 gain -#endif - ); +unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc, + opus_val16 gain, int resynth, int arch); /** Algebraic pulse decoder * @param X Decoded normalised spectrum (returned) diff --git a/libs/libopus/celt/x86/celt_lpc_sse.h b/libs/libopus/celt/x86/celt_lpc_sse.h index c5ec796ed..7d1ecf753 100644 --- a/libs/libopus/celt/x86/celt_lpc_sse.h +++ b/libs/libopus/celt/x86/celt_lpc_sse.h @@ -41,12 +41,11 @@ void celt_fir_sse4_1( opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch); #if defined(OPUS_X86_PRESUME_SSE4_1) -#define celt_fir(x, num, y, N, ord, mem, arch) \ - ((void)arch, celt_fir_sse4_1(x, num, y, N, ord, mem, arch)) +#define celt_fir(x, num, y, N, ord, arch) \ + ((void)arch, celt_fir_sse4_1(x, num, y, N, ord, arch)) #else @@ -56,11 +55,10 @@ extern void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])( opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch); -# define celt_fir(x, num, y, N, ord, mem, arch) \ - ((*CELT_FIR_IMPL[(arch) & OPUS_ARCHMASK])(x, num, y, N, ord, mem, arch)) +# define celt_fir(x, num, y, N, ord, arch) \ + ((*CELT_FIR_IMPL[(arch) & OPUS_ARCHMASK])(x, num, y, N, ord, arch)) #endif #endif diff --git a/libs/libopus/celt/x86/celt_lpc_sse.c b/libs/libopus/celt/x86/celt_lpc_sse4_1.c index 67e5592ac..547856884 100644 --- a/libs/libopus/celt/x86/celt_lpc_sse.c +++ b/libs/libopus/celt/x86/celt_lpc_sse4_1.c @@ -40,65 +40,23 @@ #if defined(FIXED_POINT) -void celt_fir_sse4_1(const opus_val16 *_x, +void celt_fir_sse4_1(const opus_val16 *x, const opus_val16 *num, - opus_val16 *_y, + opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch) { int i,j; VARDECL(opus_val16, rnum); - VARDECL(opus_val16, x); __m128i vecNoA; opus_int32 noA ; SAVE_STACK; ALLOC(rnum, ord, opus_val16); - ALLOC(x, N+ord, opus_val16); for(i=0;i<ord;i++) rnum[i] = num[ord-i-1]; - for(i=0;i<ord;i++) - x[i] = mem[ord-i-1]; - - for (i=0;i<N-7;i+=8) - { - x[i+ord ]=_x[i ]; - x[i+ord+1]=_x[i+1]; - x[i+ord+2]=_x[i+2]; - x[i+ord+3]=_x[i+3]; - x[i+ord+4]=_x[i+4]; - x[i+ord+5]=_x[i+5]; - x[i+ord+6]=_x[i+6]; - x[i+ord+7]=_x[i+7]; - } - - for (;i<N-3;i+=4) - { - x[i+ord ]=_x[i ]; - x[i+ord+1]=_x[i+1]; - x[i+ord+2]=_x[i+2]; - x[i+ord+3]=_x[i+3]; - } - - for (;i<N;i++) - x[i+ord]=_x[i]; - - for(i=0;i<ord;i++) - mem[i] = _x[N-i-1]; -#ifdef SMALL_FOOTPRINT - for (i=0;i<N;i++) - { - opus_val32 sum = SHL32(EXTEND32(_x[i]), SIG_SHIFT); - for (j=0;j<ord;j++) - { - sum = MAC16_16(sum,rnum[j],x[i+j]); - } - _y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT)); - } -#else noA = EXTEND32(1) << SIG_SHIFT >> 1; vecNoA = _mm_set_epi32(noA, noA, noA, noA); @@ -107,25 +65,24 @@ void celt_fir_sse4_1(const opus_val16 *_x, opus_val32 sums[4] = {0}; __m128i vecSum, vecX; - xcorr_kernel(rnum, x+i, sums, ord, arch); + xcorr_kernel(rnum, x+i-ord, sums, ord, arch); vecSum = _mm_loadu_si128((__m128i *)sums); vecSum = _mm_add_epi32(vecSum, vecNoA); vecSum = _mm_srai_epi32(vecSum, SIG_SHIFT); - vecX = OP_CVTEPI16_EPI32_M64(_x + i); + vecX = OP_CVTEPI16_EPI32_M64(x + i); vecSum = _mm_add_epi32(vecSum, vecX); vecSum = _mm_packs_epi32(vecSum, vecSum); - _mm_storel_epi64((__m128i *)(_y + i), vecSum); + _mm_storel_epi64((__m128i *)(y + i), vecSum); } for (;i<N;i++) { opus_val32 sum = 0; for (j=0;j<ord;j++) - sum = MAC16_16(sum, rnum[j], x[i + j]); - _y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT))); + sum = MAC16_16(sum, rnum[j], x[i+j-ord]); + y[i] = SATURATE16(ADD32(EXTEND32(x[i]), PSHR32(sum, SIG_SHIFT))); } -#endif RESTORE_STACK; } diff --git a/libs/libopus/celt/x86/pitch_sse.h b/libs/libopus/celt/x86/pitch_sse.h index e5f87ab51..f7a014b6e 100644 --- a/libs/libopus/celt/x86/pitch_sse.h +++ b/libs/libopus/celt/x86/pitch_sse.h @@ -91,7 +91,7 @@ opus_val32 celt_inner_prod_sse2( int N); #endif -#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT) +#if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT) opus_val32 celt_inner_prod_sse( const opus_val16 *x, const opus_val16 *y, diff --git a/libs/libopus/celt/x86/pitch_sse4_1.c b/libs/libopus/celt/x86/pitch_sse4_1.c index a092c68b2..2bc578304 100644 --- a/libs/libopus/celt/x86/pitch_sse4_1.c +++ b/libs/libopus/celt/x86/pitch_sse4_1.c @@ -117,6 +117,14 @@ void xcorr_kernel_sse4_1(const opus_val16 * x, const opus_val16 * y, opus_val32 __m128i sum0, sum1, sum2, sum3, vecSum; __m128i initSum; +#ifdef OPUS_CHECK_ASM + opus_val32 sum_c[4]; + for (j=0;j<4;j++) { + sum_c[j] = sum[j]; + } + xcorr_kernel_c(x, y, sum_c, len); +#endif + celt_assert(len >= 3); sum0 = _mm_setzero_si128(); @@ -177,19 +185,56 @@ void xcorr_kernel_sse4_1(const opus_val16 * x, const opus_val16 * y, opus_val32 vecSum = _mm_add_epi32(vecSum, sum2); } - for (;j<len;j++) + vecX = OP_CVTEPI16_EPI32_M64(&x[len - 4]); + if (len - j == 3) { - vecX = OP_CVTEPI16_EPI32_M64(&x[j + 0]); - vecX0 = _mm_shuffle_epi32(vecX, 0x00); + vecX0 = _mm_shuffle_epi32(vecX, 0x55); + vecX1 = _mm_shuffle_epi32(vecX, 0xaa); + vecX2 = _mm_shuffle_epi32(vecX, 0xff); vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]); + vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]); + vecY2 = OP_CVTEPI16_EPI32_M64(&y[j + 2]); sum0 = _mm_mullo_epi32(vecX0, vecY0); + sum1 = _mm_mullo_epi32(vecX1, vecY1); + sum2 = _mm_mullo_epi32(vecX2, vecY2); + + vecSum = _mm_add_epi32(vecSum, sum0); + vecSum = _mm_add_epi32(vecSum, sum1); + vecSum = _mm_add_epi32(vecSum, sum2); + } + else if (len - j == 2) + { + vecX0 = _mm_shuffle_epi32(vecX, 0xaa); + vecX1 = _mm_shuffle_epi32(vecX, 0xff); + + vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]); + vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]); + + sum0 = _mm_mullo_epi32(vecX0, vecY0); + sum1 = _mm_mullo_epi32(vecX1, vecY1); + + vecSum = _mm_add_epi32(vecSum, sum0); + vecSum = _mm_add_epi32(vecSum, sum1); + } + else if (len - j == 1) + { + vecX0 = _mm_shuffle_epi32(vecX, 0xff); + + vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]); + + sum0 = _mm_mullo_epi32(vecX0, vecY0); + vecSum = _mm_add_epi32(vecSum, sum0); } initSum = _mm_loadu_si128((__m128i *)(&sum[0])); initSum = _mm_add_epi32(initSum, vecSum); _mm_storeu_si128((__m128i *)sum, initSum); + +#ifdef OPUS_CHECK_ASM + celt_assert(!memcmp(sum_c, sum, sizeof(sum_c))); +#endif } #endif diff --git a/libs/libopus/celt/x86/vq_sse.h b/libs/libopus/celt/x86/vq_sse.h new file mode 100644 index 000000000..b4efe8f24 --- /dev/null +++ b/libs/libopus/celt/x86/vq_sse.h @@ -0,0 +1,50 @@ +/* Copyright (c) 2016 Jean-Marc Valin */ +/* + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + + - Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + + - Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER + OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#ifndef VQ_SSE_H +#define VQ_SSE_H + +#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT) +#define OVERRIDE_OP_PVQ_SEARCH + +opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch); + +#if defined(OPUS_X86_PRESUME_SSE2) +#define op_pvq_search(x, iy, K, N, arch) \ + (op_pvq_search_sse2(x, iy, K, N, arch)) + +#else + +extern opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])( + celt_norm *_X, int *iy, int K, int N, int arch); + +# define op_pvq_search(X, iy, K, N, arch) \ + ((*OP_PVQ_SEARCH_IMPL[(arch) & OPUS_ARCHMASK])(X, iy, K, N, arch)) + +#endif +#endif + +#endif diff --git a/libs/libopus/celt/x86/vq_sse2.c b/libs/libopus/celt/x86/vq_sse2.c new file mode 100644 index 000000000..775042860 --- /dev/null +++ b/libs/libopus/celt/x86/vq_sse2.c @@ -0,0 +1,217 @@ +/* Copyright (c) 2007-2008 CSIRO + Copyright (c) 2007-2009 Xiph.Org Foundation + Copyright (c) 2007-2016 Jean-Marc Valin */ +/* + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + + - Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + + - Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER + OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#ifdef HAVE_CONFIG_H +#include "config.h" +#endif + +#include <xmmintrin.h> +#include <emmintrin.h> +#include "celt_lpc.h" +#include "stack_alloc.h" +#include "mathops.h" +#include "vq.h" +#include "x86cpu.h" + + +#ifndef FIXED_POINT + +opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch) +{ + int i, j; + int pulsesLeft; + float xy, yy; + VARDECL(celt_norm, y); + VARDECL(celt_norm, X); + VARDECL(float, signy); + __m128 signmask; + __m128 sums; + __m128i fours; + SAVE_STACK; + + (void)arch; + /* All bits set to zero, except for the sign bit. */ + signmask = _mm_set_ps1(-0.f); + fours = _mm_set_epi32(4, 4, 4, 4); + ALLOC(y, N+3, celt_norm); + ALLOC(X, N+3, celt_norm); + ALLOC(signy, N+3, float); + + OPUS_COPY(X, _X, N); + X[N] = X[N+1] = X[N+2] = 0; + sums = _mm_setzero_ps(); + for (j=0;j<N;j+=4) + { + __m128 x4, s4; + x4 = _mm_loadu_ps(&X[j]); + s4 = _mm_cmplt_ps(x4, _mm_setzero_ps()); + /* Get rid of the sign */ + x4 = _mm_andnot_ps(signmask, x4); + sums = _mm_add_ps(sums, x4); + /* Clear y and iy in case we don't do the projection. */ + _mm_storeu_ps(&y[j], _mm_setzero_ps()); + _mm_storeu_si128((__m128i*)&iy[j], _mm_setzero_si128()); + _mm_storeu_ps(&X[j], x4); + _mm_storeu_ps(&signy[j], s4); + } + sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(1, 0, 3, 2))); + sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(2, 3, 0, 1))); + + xy = yy = 0; + + pulsesLeft = K; + + /* Do a pre-search by projecting on the pyramid */ + if (K > (N>>1)) + { + __m128i pulses_sum; + __m128 yy4, xy4; + __m128 rcp4; + opus_val32 sum = _mm_cvtss_f32(sums); + /* If X is too small, just replace it with a pulse at 0 */ + /* Prevents infinities and NaNs from causing too many pulses + to be allocated. 64 is an approximation of infinity here. */ + if (!(sum > EPSILON && sum < 64)) + { + X[0] = QCONST16(1.f,14); + j=1; do + X[j]=0; + while (++j<N); + sums = _mm_set_ps1(1.f); + } + /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */ + rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K+.8)), _mm_rcp_ps(sums)); + xy4 = yy4 = _mm_setzero_ps(); + pulses_sum = _mm_setzero_si128(); + for (j=0;j<N;j+=4) + { + __m128 rx4, x4, y4; + __m128i iy4; + x4 = _mm_loadu_ps(&X[j]); + rx4 = _mm_mul_ps(x4, rcp4); + iy4 = _mm_cvttps_epi32(rx4); + pulses_sum = _mm_add_epi32(pulses_sum, iy4); + _mm_storeu_si128((__m128i*)&iy[j], iy4); + y4 = _mm_cvtepi32_ps(iy4); + xy4 = _mm_add_ps(xy4, _mm_mul_ps(x4, y4)); + yy4 = _mm_add_ps(yy4, _mm_mul_ps(y4, y4)); + /* double the y[] vector so we don't have to do it in the search loop. */ + _mm_storeu_ps(&y[j], _mm_add_ps(y4, y4)); + } + pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(1, 0, 3, 2))); + pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(2, 3, 0, 1))); + pulsesLeft -= _mm_cvtsi128_si32(pulses_sum); + xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(1, 0, 3, 2))); + xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(2, 3, 0, 1))); + xy = _mm_cvtss_f32(xy4); + yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(1, 0, 3, 2))); + yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(2, 3, 0, 1))); + yy = _mm_cvtss_f32(yy4); + } + X[N] = X[N+1] = X[N+2] = -100; + y[N] = y[N+1] = y[N+2] = 100; + celt_sig_assert(pulsesLeft>=0); + + /* This should never happen, but just in case it does (e.g. on silence) + we fill the first bin with pulses. */ + if (pulsesLeft > N+3) + { + opus_val16 tmp = (opus_val16)pulsesLeft; + yy = MAC16_16(yy, tmp, tmp); + yy = MAC16_16(yy, tmp, y[0]); + iy[0] += pulsesLeft; + pulsesLeft=0; + } + + for (i=0;i<pulsesLeft;i++) + { + int best_id; + __m128 xy4, yy4; + __m128 max, max2; + __m128i count; + __m128i pos; + /* The squared magnitude term gets added anyway, so we might as well + add it outside the loop */ + yy = ADD16(yy, 1); + xy4 = _mm_load1_ps(&xy); + yy4 = _mm_load1_ps(&yy); + max = _mm_setzero_ps(); + pos = _mm_setzero_si128(); + count = _mm_set_epi32(3, 2, 1, 0); + for (j=0;j<N;j+=4) + { + __m128 x4, y4, r4; + x4 = _mm_loadu_ps(&X[j]); + y4 = _mm_loadu_ps(&y[j]); + x4 = _mm_add_ps(x4, xy4); + y4 = _mm_add_ps(y4, yy4); + y4 = _mm_rsqrt_ps(y4); + r4 = _mm_mul_ps(x4, y4); + /* Update the index of the max. */ + pos = _mm_max_epi16(pos, _mm_and_si128(count, _mm_castps_si128(_mm_cmpgt_ps(r4, max)))); + /* Update the max. */ + max = _mm_max_ps(max, r4); + /* Update the indices (+4) */ + count = _mm_add_epi32(count, fours); + } + /* Horizontal max */ + max2 = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(1, 0, 3, 2))); + max2 = _mm_max_ps(max2, _mm_shuffle_ps(max2, max2, _MM_SHUFFLE(2, 3, 0, 1))); + /* Now that max2 contains the max at all positions, look at which value(s) of the + partial max is equal to the global max. */ + pos = _mm_and_si128(pos, _mm_castps_si128(_mm_cmpeq_ps(max, max2))); + pos = _mm_max_epi16(pos, _mm_unpackhi_epi64(pos, pos)); + pos = _mm_max_epi16(pos, _mm_shufflelo_epi16(pos, _MM_SHUFFLE(1, 0, 3, 2))); + best_id = _mm_cvtsi128_si32(pos); + + /* Updating the sums of the new pulse(s) */ + xy = ADD32(xy, EXTEND32(X[best_id])); + /* We're multiplying y[j] by two so we don't have to do it here */ + yy = ADD16(yy, y[best_id]); + + /* Only now that we've made the final choice, update y/iy */ + /* Multiplying y[j] by 2 so we don't have to do it everywhere else */ + y[best_id] += 2; + iy[best_id]++; + } + + /* Put the original sign back */ + for (j=0;j<N;j+=4) + { + __m128i y4; + __m128i s4; + y4 = _mm_loadu_si128((__m128i*)&iy[j]); + s4 = _mm_castps_si128(_mm_loadu_ps(&signy[j])); + y4 = _mm_xor_si128(_mm_add_epi32(y4, s4), s4); + _mm_storeu_si128((__m128i*)&iy[j], y4); + } + RESTORE_STACK; + return yy; +} + +#endif diff --git a/libs/libopus/celt/x86/x86_celt_map.c b/libs/libopus/celt/x86/x86_celt_map.c index 47ba41b9e..d39d88ede 100644 --- a/libs/libopus/celt/x86/x86_celt_map.c +++ b/libs/libopus/celt/x86/x86_celt_map.c @@ -33,6 +33,7 @@ #include "celt_lpc.h" #include "pitch.h" #include "pitch_sse.h" +#include "vq.h" #if defined(OPUS_HAVE_RTCD) @@ -46,7 +47,6 @@ void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])( opus_val16 *y, int N, int ord, - opus_val16 *mem, int arch ) = { celt_fir_c, /* non-sse */ @@ -151,5 +151,17 @@ void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])( #endif +#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2) +opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])( + celt_norm *_X, int *iy, int K, int N, int arch +) = { + op_pvq_search_c, /* non-sse */ + op_pvq_search_c, + MAY_HAVE_SSE2(op_pvq_search), + MAY_HAVE_SSE2(op_pvq_search), + MAY_HAVE_SSE2(op_pvq_search) +}; +#endif + #endif #endif diff --git a/libs/libopus/celt/x86/x86cpu.h b/libs/libopus/celt/x86/x86cpu.h index 04fd48aac..0de8df355 100644 --- a/libs/libopus/celt/x86/x86cpu.h +++ b/libs/libopus/celt/x86/x86cpu.h @@ -56,38 +56,10 @@ int opus_select_arch(void); # endif -/*gcc appears to emit MOVDQA's to load the argument of an _mm_cvtepi8_epi32() - or _mm_cvtepi16_epi32() when optimizations are disabled, even though the - actual PMOVSXWD instruction takes an m32 or m64. Unlike a normal memory - reference, these require 16-byte alignment and load a full 16 bytes (instead - of 4 or 8), possibly reading out of bounds. - - We can insert an explicit MOVD or MOVQ using _mm_cvtsi32_si128() or - _mm_loadl_epi64(), which should have the same semantics as an m32 or m64 - reference in the PMOVSXWD instruction itself, but gcc is not smart enough to - optimize this out when optimizations ARE enabled. - - Clang, in contrast, requires us to do this always for _mm_cvtepi8_epi32 - (which is fair, since technically the compiler is always allowed to do the - dereference before invoking the function implementing the intrinsic). - However, it is smart enough to eliminate the extra MOVD instruction. - For _mm_cvtepi16_epi32, it does the right thing, though does *not* optimize out - the extra MOVQ if it's specified explicitly */ - -# if defined(__clang__) || !defined(__OPTIMIZE__) -# define OP_CVTEPI8_EPI32_M32(x) \ +#define OP_CVTEPI8_EPI32_M32(x) \ (_mm_cvtepi8_epi32(_mm_cvtsi32_si128(*(int *)(x)))) -# else -# define OP_CVTEPI8_EPI32_M32(x) \ - (_mm_cvtepi8_epi32(*(__m128i *)(x))) -#endif -# if !defined(__OPTIMIZE__) -# define OP_CVTEPI16_EPI32_M64(x) \ +#define OP_CVTEPI16_EPI32_M64(x) \ (_mm_cvtepi16_epi32(_mm_loadl_epi64((__m128i *)(x)))) -# else -# define OP_CVTEPI16_EPI32_M64(x) \ - (_mm_cvtepi16_epi32(*(__m128i *)(x))) -# endif #endif |