diff options
Diffstat (limited to 'third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c')
-rw-r--r-- | third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c | 70 |
1 files changed, 29 insertions, 41 deletions
diff --git a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c index dea113a29..b9689202a 100644 --- a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c +++ b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c @@ -110,7 +110,7 @@ static INLINE void quantize(const __m256i *qp, __m256i *c, } void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, - int skip_block, const int16_t *zbin_ptr, + const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, @@ -120,12 +120,23 @@ void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, (void)scan; const unsigned int step = 8; - if (LIKELY(!skip_block)) { - __m256i qp[5], coeff; - init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp); - coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); + __m256i qp[5], coeff; + init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp); + coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); + + __m256i eob = _mm256_setzero_si256(); + quantize(qp, &coeff, iscan, qcoeff_ptr, dqcoeff_ptr, &eob); + + coeff_ptr += step; + qcoeff_ptr += step; + dqcoeff_ptr += step; + iscan += step; + n_coeffs -= step; + + update_qp(qp); - __m256i eob = _mm256_setzero_si256(); + while (n_coeffs > 0) { + coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); quantize(qp, &coeff, iscan, qcoeff_ptr, dqcoeff_ptr, &eob); coeff_ptr += step; @@ -133,40 +144,17 @@ void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, dqcoeff_ptr += step; iscan += step; n_coeffs -= step; - - update_qp(qp); - - while (n_coeffs > 0) { - coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr); - quantize(qp, &coeff, iscan, qcoeff_ptr, dqcoeff_ptr, &eob); - - coeff_ptr += step; - qcoeff_ptr += step; - dqcoeff_ptr += step; - iscan += step; - n_coeffs -= step; - } - { - __m256i eob_s; - eob_s = _mm256_shuffle_epi32(eob, 0xe); - eob = _mm256_max_epi16(eob, eob_s); - eob_s = _mm256_shufflelo_epi16(eob, 0xe); - eob = _mm256_max_epi16(eob, eob_s); - eob_s = _mm256_shufflelo_epi16(eob, 1); - eob = _mm256_max_epi16(eob, eob_s); - const __m128i final_eob = _mm_max_epi16(_mm256_castsi256_si128(eob), - _mm256_extractf128_si256(eob, 1)); - *eob_ptr = _mm_extract_epi16(final_eob, 0); - } - } else { - do { - const __m256i zero = _mm256_setzero_si256(); - _mm256_storeu_si256((__m256i *)qcoeff_ptr, zero); - _mm256_storeu_si256((__m256i *)dqcoeff_ptr, zero); - qcoeff_ptr += step; - dqcoeff_ptr += step; - n_coeffs -= step; - } while (n_coeffs > 0); - *eob_ptr = 0; + } + { + __m256i eob_s; + eob_s = _mm256_shuffle_epi32(eob, 0xe); + eob = _mm256_max_epi16(eob, eob_s); + eob_s = _mm256_shufflelo_epi16(eob, 0xe); + eob = _mm256_max_epi16(eob, eob_s); + eob_s = _mm256_shufflelo_epi16(eob, 1); + eob = _mm256_max_epi16(eob, eob_s); + const __m128i final_eob = _mm_max_epi16(_mm256_castsi256_si128(eob), + _mm256_extractf128_si256(eob, 1)); + *eob_ptr = _mm_extract_epi16(final_eob, 0); } } |