Optimize aom_quantize_b_adaptive module Added AVX2 variant for aom_quantize_b_adaptive_c For speed = 3 and 4 presets observed encode time reduction of 0.7% and 0.6%(averaged across multiple test cases). Module gains improved by factor of ~1.82x on average w.r.t SSE2 variant. Change-Id: Ia403ae3547020699d59f009f8938ac13d8a4d7b6
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake index abf6a60..4829ee7 100644 --- a/aom_dsp/aom_dsp.cmake +++ b/aom_dsp/aom_dsp.cmake
@@ -207,6 +207,7 @@ "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_avx2.c" "${AOM_ROOT}/aom_dsp/x86/subtract_avx2.c" "${AOM_ROOT}/aom_dsp/x86/highbd_quantize_intrin_avx2.c" + "${AOM_ROOT}/aom_dsp/x86/adaptive_quantize_avx2.c" "${AOM_ROOT}/aom_dsp/x86/sad4d_avx2.c" "${AOM_ROOT}/aom_dsp/x86/sad_avx2.c" "${AOM_ROOT}/aom_dsp/x86/sad_highbd_avx2.c"
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl index 3be3c31..a0ca999 100755 --- a/aom_dsp/aom_dsp_rtcd_defs.pl +++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -519,7 +519,7 @@ specialize qw/aom_quantize_b sse2/, "$ssse3_x86_64", "$avx_x86_64"; add_proto qw/void aom_quantize_b_adaptive/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; - specialize qw/aom_quantize_b_adaptive sse2/; + specialize qw/aom_quantize_b_adaptive sse2 avx2/; add_proto qw/void aom_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan"; specialize qw/aom_quantize_b_32x32/, "$ssse3_x86_64", "$avx_x86_64";
diff --git a/aom_dsp/x86/adaptive_quantize_avx2.c b/aom_dsp/x86/adaptive_quantize_avx2.c new file mode 100644 index 0000000..3335087 --- /dev/null +++ b/aom_dsp/x86/adaptive_quantize_avx2.c
@@ -0,0 +1,244 @@ +/* + * Copyright (c) 2019, Alliance for Open Media. All rights reserved + * + * This source code is subject to the terms of the BSD 2 Clause License and + * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License + * was not distributed with this source code in the LICENSE file, you can + * obtain it at www.aomedia.org/license/software. If the Alliance for Open + * Media Patent License 1.0 was not distributed with this source code in the + * PATENTS file, you can obtain it at www.aomedia.org/license/patent. + */ + +#include <immintrin.h> +#include "config/aom_dsp_rtcd.h" +#include "aom/aom_integer.h" +#include "av1/encoder/av1_quantize.h" +#include "aom_dsp/x86/quantize_x86.h" + +static INLINE void load_b_values_avx2(const int16_t *zbin_ptr, __m256i *zbin, + const int16_t *round_ptr, __m256i *round, + const int16_t *quant_ptr, __m256i *quant, + const int16_t *dequant_ptr, + __m256i *dequant, + const int16_t *shift_ptr, + __m256i *shift) { + *zbin = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)zbin_ptr)); + *zbin = _mm256_permute4x64_epi64(*zbin, 0x54); + *zbin = _mm256_sub_epi16(*zbin, _mm256_set1_epi16(1)); + *round = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)round_ptr)); + *round = _mm256_permute4x64_epi64(*round, 0x54); + *quant = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)quant_ptr)); + *quant = _mm256_permute4x64_epi64(*quant, 0x54); + *dequant = + _mm256_castsi128_si256(_mm_load_si128((const __m128i *)dequant_ptr)); + *dequant = _mm256_permute4x64_epi64(*dequant, 0x54); + *shift = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)shift_ptr)); + *shift = _mm256_permute4x64_epi64(*shift, 0x54); +} + +static INLINE __m256i load_coefficients_avx2(const tran_low_t *coeff_ptr) { + const __m256i coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr)); + const __m256i coeff2 = _mm256_load_si256((__m256i *)(coeff_ptr + 8)); + return _mm256_packs_epi32(coeff1, coeff2); +} + +static INLINE void update_mask1_avx2(__m256i *cmp_mask, + const int16_t *iscan_ptr, int *is_found, + __m256i *mask) { + __m256i temp_mask = _mm256_setzero_si256(); + if (_mm256_movemask_epi8(*cmp_mask)) { + __m256i iscan = _mm256_loadu_si256((const __m256i *)(iscan_ptr)); + temp_mask = _mm256_and_si256(*cmp_mask, iscan); + *is_found = 1; + } + *mask = _mm256_max_epi16(temp_mask, *mask); +} + +static INLINE void update_mask0_avx2(__m256i *qcoeff, __m256i *threshold, + const int16_t *iscan_ptr, int *is_found, + __m256i *mask) { + __m256i zero = _mm256_setzero_si256(); + __m256i coeff[2], cmp_mask0, cmp_mask1; + coeff[0] = _mm256_unpacklo_epi16(*qcoeff, zero); + coeff[1] = _mm256_unpackhi_epi16(*qcoeff, zero); + coeff[0] = _mm256_slli_epi32(coeff[0], AOM_QM_BITS); + cmp_mask0 = _mm256_cmpgt_epi32(coeff[0], threshold[0]); + coeff[1] = _mm256_slli_epi32(coeff[1], AOM_QM_BITS); + cmp_mask1 = _mm256_cmpgt_epi32(coeff[1], threshold[1]); + cmp_mask0 = + _mm256_permute4x64_epi64(_mm256_packs_epi32(cmp_mask0, cmp_mask1), 0xd8); + update_mask1_avx2(&cmp_mask0, iscan_ptr, is_found, mask); +} + +static INLINE void calculate_qcoeff_avx2(__m256i *coeff, const __m256i *round, + const __m256i *quant, + const __m256i *shift) { + __m256i tmp, qcoeff; + qcoeff = _mm256_adds_epi16(*coeff, *round); + tmp = _mm256_mulhi_epi16(qcoeff, *quant); + qcoeff = _mm256_add_epi16(tmp, qcoeff); + *coeff = _mm256_mulhi_epi16(qcoeff, *shift); +} + +static INLINE __m256i calculate_dqcoeff_avx2(__m256i qcoeff, __m256i dequant) { + return _mm256_mullo_epi16(qcoeff, dequant); +} + +static INLINE void store_coefficients_avx2(__m256i coeff_vals, + tran_low_t *coeff_ptr) { + __m256i coeff_sign = _mm256_srai_epi16(coeff_vals, 15); + __m256i coeff_vals_lo = _mm256_unpacklo_epi16(coeff_vals, coeff_sign); + __m256i coeff_vals_hi = _mm256_unpackhi_epi16(coeff_vals, coeff_sign); + _mm256_store_si256((__m256i *)(coeff_ptr), coeff_vals_lo); + _mm256_store_si256((__m256i *)(coeff_ptr + 8), coeff_vals_hi); +} + +void aom_quantize_b_adaptive_avx2( + const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr, + const int16_t *round_ptr, const int16_t *quant_ptr, + const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, + tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, + const int16_t *scan, const int16_t *iscan) { + int index = 16; + int non_zero_count = 0; + int non_zero_count_prescan_add_zero = 0; + int is_found0 = 0, is_found1 = 0; + int eob = -1; + const __m256i zero = _mm256_setzero_si256(); + __m256i zbin, round, quant, dequant, shift; + __m256i coeff, qcoeff; + __m256i cmp_mask, mask0 = zero, mask1 = zero; + __m128i temp_mask0, temp_mask1; + int prescan_add[2]; + int thresh[2]; + const qm_val_t wt = (1 << AOM_QM_BITS); + for (int i = 0; i < 2; ++i) { + prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7); + thresh[i] = (zbin_ptr[i] * wt + prescan_add[i]) - 1; + } + __m256i threshold[2]; + threshold[0] = _mm256_set1_epi32(thresh[0]); + threshold[1] = _mm256_set1_epi32(thresh[1]); + threshold[0] = _mm256_blend_epi32(threshold[0], threshold[1], 0xfe); + +#if SKIP_EOB_FACTOR_ADJUST + int first = -1; +#endif + + // Setup global values. + load_b_values_avx2(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant, + dequant_ptr, &dequant, quant_shift_ptr, &shift); + + // Do DC and first 15 AC. + coeff = load_coefficients_avx2(coeff_ptr); + qcoeff = _mm256_abs_epi16(coeff); + update_mask0_avx2(&qcoeff, threshold, iscan, &is_found0, &mask0); + __m256i temp0 = _mm256_cmpgt_epi16(qcoeff, zbin); + zbin = _mm256_unpackhi_epi64(zbin, zbin); + cmp_mask = _mm256_permute4x64_epi64(temp0, 0xd8); + update_mask1_avx2(&cmp_mask, iscan, &is_found1, &mask1); + threshold[0] = threshold[1]; + if (_mm256_movemask_epi8(cmp_mask) == 0) { + _mm256_store_si256((__m256i *)(qcoeff_ptr), zero); + _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), zero); + _mm256_store_si256((__m256i *)(dqcoeff_ptr), zero); + _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), zero); + round = _mm256_unpackhi_epi64(round, round); + quant = _mm256_unpackhi_epi64(quant, quant); + shift = _mm256_unpackhi_epi64(shift, shift); + dequant = _mm256_unpackhi_epi64(dequant, dequant); + } else { + calculate_qcoeff_avx2(&qcoeff, &round, &quant, &shift); + round = _mm256_unpackhi_epi64(round, round); + quant = _mm256_unpackhi_epi64(quant, quant); + shift = _mm256_unpackhi_epi64(shift, shift); + // Reinsert signs + qcoeff = _mm256_sign_epi16(qcoeff, coeff); + // Mask out zbin threshold coeffs + qcoeff = _mm256_and_si256(qcoeff, temp0); + store_coefficients_avx2(qcoeff, qcoeff_ptr); + coeff = calculate_dqcoeff_avx2(qcoeff, dequant); + dequant = _mm256_unpackhi_epi64(dequant, dequant); + store_coefficients_avx2(coeff, dqcoeff_ptr); + } + + // AC only loop. + while (index < n_coeffs) { + coeff = load_coefficients_avx2(coeff_ptr + index); + qcoeff = _mm256_abs_epi16(coeff); + update_mask0_avx2(&qcoeff, threshold, iscan + index, &is_found0, &mask0); + temp0 = _mm256_cmpgt_epi16(qcoeff, zbin); + cmp_mask = _mm256_permute4x64_epi64(temp0, 0xd8); + update_mask1_avx2(&cmp_mask, iscan + index, &is_found1, &mask1); + if (_mm256_movemask_epi8(cmp_mask) == 0) { + _mm256_store_si256((__m256i *)(qcoeff_ptr + index), zero); + _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), zero); + _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), zero); + _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), zero); + index += 16; + continue; + } + calculate_qcoeff_avx2(&qcoeff, &round, &quant, &shift); + qcoeff = _mm256_sign_epi16(qcoeff, coeff); + qcoeff = _mm256_and_si256(qcoeff, temp0); + store_coefficients_avx2(qcoeff, qcoeff_ptr + index); + coeff = calculate_dqcoeff_avx2(qcoeff, dequant); + store_coefficients_avx2(coeff, dqcoeff_ptr + index); + index += 16; + } + if (is_found0) { + temp_mask0 = _mm_max_epi16(_mm256_castsi256_si128(mask0), + _mm256_extracti128_si256(mask0, 1)); + non_zero_count = calculate_non_zero_count(temp_mask0); + } + if (is_found1) { + temp_mask1 = _mm_max_epi16(_mm256_castsi256_si128(mask1), + _mm256_extracti128_si256(mask1, 1)); + non_zero_count_prescan_add_zero = calculate_non_zero_count(temp_mask1); + } + + for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) { + const int rc = scan[i]; + qcoeff_ptr[rc] = 0; + dqcoeff_ptr[rc] = 0; + } + + for (int i = non_zero_count - 1; i >= 0; i--) { + const int rc = scan[i]; + if (qcoeff_ptr[rc]) { + eob = i; + break; + } + } + + *eob_ptr = eob + 1; +#if SKIP_EOB_FACTOR_ADJUST + // TODO(Aniket): Experiment the following loop with intrinsic by combining + // with the quantization loop above + for (int i = 0; i < non_zero_count; i++) { + const int rc = scan[i]; + const int qcoeff0 = qcoeff_ptr[rc]; + if (qcoeff0) { + first = i; + break; + } + } + if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) { + const int rc = scan[(*eob_ptr - 1)]; + if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) { + const int coeff0 = coeff_ptr[rc] * wt; + const int coeff_sign = (coeff0 >> 31); + const int abs_coeff = (coeff0 ^ coeff_sign) - coeff_sign; + const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST; + const int prescan_add_val = + ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7); + if (abs_coeff < + (zbin_ptr[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) { + qcoeff_ptr[rc] = 0; + dqcoeff_ptr[rc] = 0; + *eob_ptr = 0; + } + } + } +#endif +}
diff --git a/test/quantize_func_test.cc b/test/quantize_func_test.cc index 51f2db8..b705c79 100644 --- a/test/quantize_func_test.cc +++ b/test/quantize_func_test.cc
@@ -378,6 +378,12 @@ TYPE_B, AOM_BITS_10), make_tuple(&aom_highbd_quantize_b_c, &aom_highbd_quantize_b_avx2, TX_16X16, TYPE_B, AOM_BITS_12), + make_tuple(&aom_quantize_b_adaptive_c, &aom_quantize_b_adaptive_avx2, + TX_16X16, TYPE_B, AOM_BITS_8), + make_tuple(&aom_quantize_b_adaptive_c, &aom_quantize_b_adaptive_avx2, TX_8X8, + TYPE_B, AOM_BITS_8), + make_tuple(&aom_quantize_b_adaptive_c, &aom_quantize_b_adaptive_avx2, TX_4X4, + TYPE_B, AOM_BITS_8) }; INSTANTIATE_TEST_CASE_P(AVX2, QuantizeTest,