Add AVX2 variant for highbd_quantize_b_adaptive_c
Added AVX2 optimization for aom_highbd_quantize_b_adaptive_c()
Encode Time
Preset Reduction(%)
1 1.87
2 3.82
3 4.68
4 2.42
Module gains improved by factor of ~6.3x w.r.t C module
Change-Id: If8df960f541062d7de87c65127589521ee3e07c4
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake
index a772731..6cfbd32 100644
--- a/aom_dsp/aom_dsp.cmake
+++ b/aom_dsp/aom_dsp.cmake
@@ -207,6 +207,7 @@
"${AOM_ROOT}/aom_dsp/x86/subtract_avx2.c"
"${AOM_ROOT}/aom_dsp/x86/highbd_quantize_intrin_avx2.c"
"${AOM_ROOT}/aom_dsp/x86/adaptive_quantize_avx2.c"
+ "${AOM_ROOT}/aom_dsp/x86/highbd_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 06214ab..55be079 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -538,6 +538,9 @@
add_proto qw/void aom_highbd_quantize_b/, "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_highbd_quantize_b sse2 avx2/;
+ add_proto qw/void aom_highbd_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_highbd_quantize_b_adaptive avx2/;
+
add_proto qw/void aom_highbd_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_highbd_quantize_b_32x32 sse2/;
diff --git a/aom_dsp/x86/highbd_adaptive_quantize_avx2.c b/aom_dsp/x86/highbd_adaptive_quantize_avx2.c
new file mode 100644
index 0000000..8e4d21b
--- /dev/null
+++ b/aom_dsp/x86/highbd_adaptive_quantize_avx2.c
@@ -0,0 +1,265 @@
+/*
+ * 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 "aom_dsp/x86/quantize_x86.h"
+
+#include "av1/encoder/av1_quantize.h"
+
+static INLINE void highbd_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_cvtepi16_epi32(_mm_load_si128((const __m128i *)zbin_ptr));
+ *zbin = _mm256_sub_epi32(*zbin, _mm256_set1_epi32(1));
+ *round = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)round_ptr));
+ *quant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_ptr));
+ *dequant =
+ _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)dequant_ptr));
+ *shift = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)shift_ptr));
+}
+
+static INLINE void highbd_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 highbd_update_mask0_avx2(__m256i *qcoeff0, __m256i *qcoeff1,
+ __m256i *threshold,
+ const int16_t *iscan_ptr,
+ int *is_found, __m256i *mask) {
+ __m256i coeff[2], cmp_mask0, cmp_mask1;
+ coeff[0] = _mm256_slli_epi32(*qcoeff0, AOM_QM_BITS);
+ cmp_mask0 = _mm256_cmpgt_epi32(coeff[0], threshold[0]);
+ coeff[1] = _mm256_slli_epi32(*qcoeff1, 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);
+ highbd_update_mask1_avx2(&cmp_mask0, iscan_ptr, is_found, mask);
+}
+
+static INLINE void highbd_mul_shift_avx2(const __m256i *x, const __m256i *y,
+ __m256i *p, const int shift) {
+ __m256i prod_lo = _mm256_mul_epi32(*x, *y);
+ __m256i prod_hi = _mm256_srli_epi64(*x, 32);
+ const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
+ prod_hi = _mm256_mul_epi32(prod_hi, mult_hi);
+
+ prod_lo = _mm256_srli_epi64(prod_lo, shift);
+ prod_hi = _mm256_srli_epi64(prod_hi, shift);
+
+ prod_hi = _mm256_slli_epi64(prod_hi, 32);
+ *p = _mm256_blend_epi32(prod_lo, prod_hi, 0xaa);
+}
+
+static INLINE void highbd_calculate_qcoeff_avx2(__m256i *coeff,
+ const __m256i *round,
+ const __m256i *quant,
+ const __m256i *shift,
+ const int *log_scale) {
+ __m256i tmp, qcoeff;
+ qcoeff = _mm256_add_epi32(*coeff, *round);
+ highbd_mul_shift_avx2(&qcoeff, quant, &tmp, 16);
+ qcoeff = _mm256_add_epi32(tmp, qcoeff);
+ highbd_mul_shift_avx2(&qcoeff, shift, coeff, 16 - *log_scale);
+}
+
+static INLINE __m256i highbd_calculate_dqcoeff_avx2(__m256i qcoeff,
+ __m256i dequant) {
+ return _mm256_mullo_epi32(qcoeff, dequant);
+}
+
+static INLINE void highbd_store_coefficients_avx2(__m256i coeff0,
+ __m256i coeff1,
+ tran_low_t *coeff_ptr) {
+ _mm256_store_si256((__m256i *)(coeff_ptr), coeff0);
+ _mm256_store_si256((__m256i *)(coeff_ptr + 8), coeff1);
+}
+
+void aom_highbd_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 coeff0, qcoeff0, coeff1, qcoeff1;
+ __m256i cmp_mask, mask0 = zero, mask1 = zero;
+ __m128i temp_mask0, temp_mask1;
+ int prescan_add[2];
+ int thresh[2];
+ const int log_scale = 0;
+ 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.
+ highbd_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.
+ coeff0 = _mm256_load_si256((__m256i *)(coeff_ptr));
+ qcoeff0 = _mm256_abs_epi32(coeff0);
+ coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr + 8));
+ qcoeff1 = _mm256_abs_epi32(coeff1);
+ highbd_update_mask0_avx2(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0,
+ &mask0);
+ __m256i temp0 = _mm256_cmpgt_epi32(qcoeff0, zbin);
+ zbin = _mm256_unpackhi_epi64(zbin, zbin);
+ __m256i temp1 = _mm256_cmpgt_epi32(qcoeff1, zbin);
+ cmp_mask = _mm256_permute4x64_epi64(_mm256_packs_epi32(temp0, temp1), 0xd8);
+ highbd_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 {
+ highbd_calculate_qcoeff_avx2(&qcoeff0, &round, &quant, &shift, &log_scale);
+ round = _mm256_unpackhi_epi64(round, round);
+ quant = _mm256_unpackhi_epi64(quant, quant);
+ shift = _mm256_unpackhi_epi64(shift, shift);
+ highbd_calculate_qcoeff_avx2(&qcoeff1, &round, &quant, &shift, &log_scale);
+ // Reinsert signs
+ qcoeff0 = _mm256_sign_epi32(qcoeff0, coeff0);
+ qcoeff1 = _mm256_sign_epi32(qcoeff1, coeff1);
+ // Mask out zbin threshold coeffs
+ qcoeff0 = _mm256_and_si256(qcoeff0, temp0);
+ qcoeff1 = _mm256_and_si256(qcoeff1, temp1);
+ highbd_store_coefficients_avx2(qcoeff0, qcoeff1, qcoeff_ptr);
+ coeff0 = highbd_calculate_dqcoeff_avx2(qcoeff0, dequant);
+ dequant = _mm256_unpackhi_epi64(dequant, dequant);
+ coeff1 = highbd_calculate_dqcoeff_avx2(qcoeff1, dequant);
+ highbd_store_coefficients_avx2(coeff0, coeff1, dqcoeff_ptr);
+ }
+
+ // AC only loop.
+ while (index < n_coeffs) {
+ coeff0 = _mm256_load_si256((__m256i *)(coeff_ptr + index));
+ qcoeff0 = _mm256_abs_epi32(coeff0);
+ coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr + index + 8));
+ qcoeff1 = _mm256_abs_epi32(coeff1);
+ highbd_update_mask0_avx2(&qcoeff0, &qcoeff1, threshold, iscan + index,
+ &is_found0, &mask0);
+ temp0 = _mm256_cmpgt_epi32(qcoeff0, zbin);
+ temp1 = _mm256_cmpgt_epi32(qcoeff1, zbin);
+ cmp_mask = _mm256_permute4x64_epi64(_mm256_packs_epi32(temp0, temp1), 0xd8);
+ highbd_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;
+ }
+ highbd_calculate_qcoeff_avx2(&qcoeff0, &round, &quant, &shift, &log_scale);
+ highbd_calculate_qcoeff_avx2(&qcoeff1, &round, &quant, &shift, &log_scale);
+ qcoeff0 = _mm256_sign_epi32(qcoeff0, coeff0);
+ qcoeff1 = _mm256_sign_epi32(qcoeff1, coeff1);
+ qcoeff0 = _mm256_and_si256(qcoeff0, temp0);
+ qcoeff1 = _mm256_and_si256(qcoeff1, temp1);
+ highbd_store_coefficients_avx2(qcoeff0, qcoeff1, qcoeff_ptr + index);
+ coeff0 = highbd_calculate_dqcoeff_avx2(qcoeff0, dequant);
+ coeff1 = highbd_calculate_dqcoeff_avx2(qcoeff1, dequant);
+ highbd_store_coefficients_avx2(coeff0, coeff1, 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 qcoeff = qcoeff_ptr[rc];
+ if (qcoeff) {
+ 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 coeff = coeff_ptr[rc] * wt;
+ const int coeff_sign = (coeff >> 31);
+ const int abs_coeff = (coeff ^ 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/av1/encoder/av1_quantize.c b/av1/encoder/av1_quantize.c
index 5091b19..3e19cbd 100644
--- a/av1/encoder/av1_quantize.c
+++ b/av1/encoder/av1_quantize.c
@@ -424,7 +424,7 @@
switch (qparam->log_scale) {
case 0:
if (LIKELY(n_coeffs >= 8)) {
- aom_highbd_quantize_b_adaptive_c(
+ aom_highbd_quantize_b_adaptive(
coeff_ptr, n_coeffs, p->zbin_QTX, p->round_QTX, p->quant_QTX,
p->quant_shift_QTX, qcoeff_ptr, dqcoeff_ptr, p->dequant_QTX,
eob_ptr, sc->scan, sc->iscan);
diff --git a/test/quantize_func_test.cc b/test/quantize_func_test.cc
index 1ab1c60..b2977f2 100644
--- a/test/quantize_func_test.cc
+++ b/test/quantize_func_test.cc
@@ -383,7 +383,16 @@
make_tuple(&aom_quantize_b_adaptive_c, &aom_quantize_b_adaptive_avx2,
static_cast<TX_SIZE>(TX_8X8), TYPE_B, AOM_BITS_8),
make_tuple(&aom_quantize_b_adaptive_c, &aom_quantize_b_adaptive_avx2,
- static_cast<TX_SIZE>(TX_4X4), TYPE_B, AOM_BITS_8)
+ static_cast<TX_SIZE>(TX_4X4), TYPE_B, AOM_BITS_8),
+ make_tuple(&aom_highbd_quantize_b_adaptive_c,
+ &aom_highbd_quantize_b_adaptive_avx2,
+ static_cast<TX_SIZE>(TX_16X16), TYPE_B, AOM_BITS_8),
+ make_tuple(&aom_highbd_quantize_b_adaptive_c,
+ &aom_highbd_quantize_b_adaptive_avx2,
+ static_cast<TX_SIZE>(TX_16X16), TYPE_B, AOM_BITS_10),
+ make_tuple(&aom_highbd_quantize_b_adaptive_c,
+ &aom_highbd_quantize_b_adaptive_avx2,
+ static_cast<TX_SIZE>(TX_16X16), TYPE_B, AOM_BITS_12)
};
INSTANTIATE_TEST_CASE_P(AVX2, QuantizeTest,
