diff options
Diffstat (limited to 'media/libaom/src/aom_dsp/x86/quantize_x86.h')
| -rw-r--r-- | media/libaom/src/aom_dsp/x86/quantize_x86.h | 125 |
1 files changed, 125 insertions, 0 deletions
diff --git a/media/libaom/src/aom_dsp/x86/quantize_x86.h b/media/libaom/src/aom_dsp/x86/quantize_x86.h index 4eed7dd29a..5b040a278a 100644 --- a/media/libaom/src/aom_dsp/x86/quantize_x86.h +++ b/media/libaom/src/aom_dsp/x86/quantize_x86.h @@ -32,6 +32,11 @@ static INLINE __m128i invert_sign_sse2(__m128i a, __m128i sign) { return _mm_sub_epi16(a, sign); } +static INLINE __m128i invert_sign_32_sse2(__m128i a, __m128i sign) { + a = _mm_xor_si128(a, sign); + return _mm_sub_epi32(a, sign); +} + static INLINE void calculate_qcoeff(__m128i *coeff, const __m128i round, const __m128i quant, const __m128i shift) { __m128i tmp, qcoeff; @@ -41,10 +46,53 @@ static INLINE void calculate_qcoeff(__m128i *coeff, const __m128i round, *coeff = _mm_mulhi_epi16(qcoeff, shift); } +static INLINE void calculate_qcoeff_log_scale(__m128i *coeff, + const __m128i round, + const __m128i quant, + const __m128i *shift, + const int *log_scale) { + __m128i tmp, tmp1, qcoeff; + qcoeff = _mm_adds_epi16(*coeff, round); + tmp = _mm_mulhi_epi16(qcoeff, quant); + qcoeff = _mm_add_epi16(tmp, qcoeff); + tmp = _mm_mullo_epi16(qcoeff, *shift); + tmp = _mm_srli_epi16(tmp, (16 - *log_scale)); + tmp1 = _mm_mulhi_epi16(qcoeff, *shift); + tmp1 = _mm_slli_epi16(tmp1, *log_scale); + *coeff = _mm_or_si128(tmp, tmp1); +} + static INLINE __m128i calculate_dqcoeff(__m128i qcoeff, __m128i dequant) { return _mm_mullo_epi16(qcoeff, dequant); } +static INLINE void calculate_dqcoeff_and_store_log_scale(__m128i qcoeff, + __m128i dequant, + const __m128i zero, + tran_low_t *dqcoeff, + const int *log_scale) { + // calculate abs + __m128i coeff_sign = _mm_srai_epi16(qcoeff, 15); + __m128i coeff = invert_sign_sse2(qcoeff, coeff_sign); + + const __m128i sign_0 = _mm_unpacklo_epi16(coeff_sign, zero); + const __m128i sign_1 = _mm_unpackhi_epi16(coeff_sign, zero); + + const __m128i low = _mm_mullo_epi16(coeff, dequant); + const __m128i high = _mm_mulhi_epi16(coeff, dequant); + __m128i dqcoeff32_0 = _mm_unpacklo_epi16(low, high); + __m128i dqcoeff32_1 = _mm_unpackhi_epi16(low, high); + + dqcoeff32_0 = _mm_srli_epi32(dqcoeff32_0, *log_scale); + dqcoeff32_1 = _mm_srli_epi32(dqcoeff32_1, *log_scale); + + dqcoeff32_0 = invert_sign_32_sse2(dqcoeff32_0, sign_0); + dqcoeff32_1 = invert_sign_32_sse2(dqcoeff32_1, sign_1); + + _mm_store_si128((__m128i *)(dqcoeff), dqcoeff32_0); + _mm_store_si128((__m128i *)(dqcoeff + 4), dqcoeff32_1); +} + // Scan 16 values for eob reference in scan_ptr. Use masks (-1) from comparing // to zbin to add 1 to the index in 'scan'. static INLINE __m128i scan_for_eob(__m128i *coeff0, __m128i *coeff1, @@ -75,3 +123,80 @@ static INLINE int16_t accumulate_eob(__m128i eob) { eob = _mm_max_epi16(eob, eob_shuffled); return _mm_extract_epi16(eob, 1); } + +static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) { + assert(sizeof(tran_low_t) == 4); + const __m128i coeff1 = _mm_load_si128((__m128i *)(coeff_ptr)); + const __m128i coeff2 = _mm_load_si128((__m128i *)(coeff_ptr + 4)); + return _mm_packs_epi32(coeff1, coeff2); +} + +static INLINE void store_coefficients(__m128i coeff_vals, + tran_low_t *coeff_ptr) { + assert(sizeof(tran_low_t) == 4); + + __m128i one = _mm_set1_epi16(1); + __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one); + __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one); + __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi); + __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi); + _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1); + _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2); +} + +static INLINE void update_mask1(__m128i *cmp_mask0, __m128i *cmp_mask1, + const int16_t *iscan_ptr, int *is_found, + __m128i *mask) { + __m128i all_zero; + __m128i temp_mask = _mm_setzero_si128(); + all_zero = _mm_or_si128(*cmp_mask0, *cmp_mask1); + if (_mm_movemask_epi8(all_zero)) { + __m128i iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr)); + __m128i mask0 = _mm_and_si128(*cmp_mask0, iscan0); + __m128i iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8)); + __m128i mask1 = _mm_and_si128(*cmp_mask1, iscan1); + temp_mask = _mm_max_epi16(mask0, mask1); + *is_found = 1; + } + *mask = _mm_max_epi16(temp_mask, *mask); +} + +static INLINE void update_mask0(__m128i *qcoeff0, __m128i *qcoeff1, + __m128i *threshold, const int16_t *iscan_ptr, + int *is_found, __m128i *mask) { + __m128i zero = _mm_setzero_si128(); + __m128i coeff[4], cmp_mask0, cmp_mask1, cmp_mask2, cmp_mask3; + + coeff[0] = _mm_unpacklo_epi16(*qcoeff0, zero); + coeff[1] = _mm_unpackhi_epi16(*qcoeff0, zero); + coeff[2] = _mm_unpacklo_epi16(*qcoeff1, zero); + coeff[3] = _mm_unpackhi_epi16(*qcoeff1, zero); + + coeff[0] = _mm_slli_epi32(coeff[0], AOM_QM_BITS); + cmp_mask0 = _mm_cmpgt_epi32(coeff[0], threshold[0]); + coeff[1] = _mm_slli_epi32(coeff[1], AOM_QM_BITS); + cmp_mask1 = _mm_cmpgt_epi32(coeff[1], threshold[1]); + coeff[2] = _mm_slli_epi32(coeff[2], AOM_QM_BITS); + cmp_mask2 = _mm_cmpgt_epi32(coeff[2], threshold[1]); + coeff[3] = _mm_slli_epi32(coeff[3], AOM_QM_BITS); + cmp_mask3 = _mm_cmpgt_epi32(coeff[3], threshold[1]); + + cmp_mask0 = _mm_packs_epi32(cmp_mask0, cmp_mask1); + cmp_mask1 = _mm_packs_epi32(cmp_mask2, cmp_mask3); + + update_mask1(&cmp_mask0, &cmp_mask1, iscan_ptr, is_found, mask); +} + +static INLINE int calculate_non_zero_count(__m128i mask) { + __m128i mask0, mask1; + int non_zero_count = 0; + mask0 = _mm_unpackhi_epi64(mask, mask); + mask1 = _mm_max_epi16(mask0, mask); + mask0 = _mm_shuffle_epi32(mask1, 1); + mask0 = _mm_max_epi16(mask0, mask1); + mask1 = _mm_srli_epi32(mask0, 16); + mask0 = _mm_max_epi16(mask0, mask1); + non_zero_count = _mm_extract_epi16(mask0, 0) + 1; + + return non_zero_count; +} |