blob: f5b5d273e31c169c81f6b5f0412145b528686400 [file] [log] [blame] [edit]
/*
* Copyright (c) 2021, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 3-Clause Clear License
* and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear
* License was not distributed with this source code in the LICENSE file, you
* can obtain it at aomedia.org/license/software-license/bsd-3-c-c/. 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
* aomedia.org/license/patent-license/.
*/
#include <immintrin.h>
#include "config/aom_dsp_rtcd.h"
#include "aom/aom_integer.h"
#include "aom_dsp/x86/bitdepth_conversion_avx2.h"
#include "aom_ports/mem.h"
static void hadamard_col8x2_avx2(__m256i *in, int iter) {
__m256i a0 = in[0];
__m256i a1 = in[1];
__m256i a2 = in[2];
__m256i a3 = in[3];
__m256i a4 = in[4];
__m256i a5 = in[5];
__m256i a6 = in[6];
__m256i a7 = in[7];
__m256i b0 = _mm256_add_epi16(a0, a1);
__m256i b1 = _mm256_sub_epi16(a0, a1);
__m256i b2 = _mm256_add_epi16(a2, a3);
__m256i b3 = _mm256_sub_epi16(a2, a3);
__m256i b4 = _mm256_add_epi16(a4, a5);
__m256i b5 = _mm256_sub_epi16(a4, a5);
__m256i b6 = _mm256_add_epi16(a6, a7);
__m256i b7 = _mm256_sub_epi16(a6, a7);
a0 = _mm256_add_epi16(b0, b2);
a1 = _mm256_add_epi16(b1, b3);
a2 = _mm256_sub_epi16(b0, b2);
a3 = _mm256_sub_epi16(b1, b3);
a4 = _mm256_add_epi16(b4, b6);
a5 = _mm256_add_epi16(b5, b7);
a6 = _mm256_sub_epi16(b4, b6);
a7 = _mm256_sub_epi16(b5, b7);
if (iter == 0) {
b0 = _mm256_add_epi16(a0, a4);
b7 = _mm256_add_epi16(a1, a5);
b3 = _mm256_add_epi16(a2, a6);
b4 = _mm256_add_epi16(a3, a7);
b2 = _mm256_sub_epi16(a0, a4);
b6 = _mm256_sub_epi16(a1, a5);
b1 = _mm256_sub_epi16(a2, a6);
b5 = _mm256_sub_epi16(a3, a7);
a0 = _mm256_unpacklo_epi16(b0, b1);
a1 = _mm256_unpacklo_epi16(b2, b3);
a2 = _mm256_unpackhi_epi16(b0, b1);
a3 = _mm256_unpackhi_epi16(b2, b3);
a4 = _mm256_unpacklo_epi16(b4, b5);
a5 = _mm256_unpacklo_epi16(b6, b7);
a6 = _mm256_unpackhi_epi16(b4, b5);
a7 = _mm256_unpackhi_epi16(b6, b7);
b0 = _mm256_unpacklo_epi32(a0, a1);
b1 = _mm256_unpacklo_epi32(a4, a5);
b2 = _mm256_unpackhi_epi32(a0, a1);
b3 = _mm256_unpackhi_epi32(a4, a5);
b4 = _mm256_unpacklo_epi32(a2, a3);
b5 = _mm256_unpacklo_epi32(a6, a7);
b6 = _mm256_unpackhi_epi32(a2, a3);
b7 = _mm256_unpackhi_epi32(a6, a7);
in[0] = _mm256_unpacklo_epi64(b0, b1);
in[1] = _mm256_unpackhi_epi64(b0, b1);
in[2] = _mm256_unpacklo_epi64(b2, b3);
in[3] = _mm256_unpackhi_epi64(b2, b3);
in[4] = _mm256_unpacklo_epi64(b4, b5);
in[5] = _mm256_unpackhi_epi64(b4, b5);
in[6] = _mm256_unpacklo_epi64(b6, b7);
in[7] = _mm256_unpackhi_epi64(b6, b7);
} else {
in[0] = _mm256_add_epi16(a0, a4);
in[7] = _mm256_add_epi16(a1, a5);
in[3] = _mm256_add_epi16(a2, a6);
in[4] = _mm256_add_epi16(a3, a7);
in[2] = _mm256_sub_epi16(a0, a4);
in[6] = _mm256_sub_epi16(a1, a5);
in[1] = _mm256_sub_epi16(a2, a6);
in[5] = _mm256_sub_epi16(a3, a7);
}
}
static void hadamard_8x8x2_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
int16_t *coeff) {
__m256i src[8];
src[0] = _mm256_loadu_si256((const __m256i *)src_diff);
src[1] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
src[2] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
src[3] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
src[4] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
src[5] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
src[6] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
src[7] = _mm256_loadu_si256((const __m256i *)(src_diff + src_stride));
hadamard_col8x2_avx2(src, 0);
hadamard_col8x2_avx2(src, 1);
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[0], src[1], 0x20));
coeff += 16;
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[2], src[3], 0x20));
coeff += 16;
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[4], src[5], 0x20));
coeff += 16;
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[6], src[7], 0x20));
coeff += 16;
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[0], src[1], 0x31));
coeff += 16;
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[2], src[3], 0x31));
coeff += 16;
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[4], src[5], 0x31));
coeff += 16;
_mm256_storeu_si256((__m256i *)coeff,
_mm256_permute2x128_si256(src[6], src[7], 0x31));
}
static INLINE void hadamard_16x16_avx2(const int16_t *src_diff,
ptrdiff_t src_stride, tran_low_t *coeff,
int is_final) {
DECLARE_ALIGNED(32, int16_t, temp_coeff[16 * 16]);
int16_t *t_coeff = temp_coeff;
int16_t *coeff16 = (int16_t *)coeff;
int idx;
for (idx = 0; idx < 2; ++idx) {
const int16_t *src_ptr = src_diff + idx * 8 * src_stride;
hadamard_8x8x2_avx2(src_ptr, src_stride, t_coeff + (idx * 64 * 2));
}
for (idx = 0; idx < 64; idx += 16) {
const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
__m256i b0 = _mm256_add_epi16(coeff0, coeff1);
__m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
__m256i b2 = _mm256_add_epi16(coeff2, coeff3);
__m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
b0 = _mm256_srai_epi16(b0, 1);
b1 = _mm256_srai_epi16(b1, 1);
b2 = _mm256_srai_epi16(b2, 1);
b3 = _mm256_srai_epi16(b3, 1);
if (is_final) {
store_tran_low(_mm256_add_epi16(b0, b2), coeff);
store_tran_low(_mm256_add_epi16(b1, b3), coeff + 64);
store_tran_low(_mm256_sub_epi16(b0, b2), coeff + 128);
store_tran_low(_mm256_sub_epi16(b1, b3), coeff + 192);
coeff += 16;
} else {
_mm256_storeu_si256((__m256i *)coeff16, _mm256_add_epi16(b0, b2));
_mm256_storeu_si256((__m256i *)(coeff16 + 64), _mm256_add_epi16(b1, b3));
_mm256_storeu_si256((__m256i *)(coeff16 + 128), _mm256_sub_epi16(b0, b2));
_mm256_storeu_si256((__m256i *)(coeff16 + 192), _mm256_sub_epi16(b1, b3));
coeff16 += 16;
}
t_coeff += 16;
}
}
void aom_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
tran_low_t *coeff) {
hadamard_16x16_avx2(src_diff, src_stride, coeff, 1);
}
void aom_hadamard_lp_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
int16_t *coeff) {
int16_t *t_coeff = coeff;
for (int idx = 0; idx < 2; ++idx) {
const int16_t *src_ptr = src_diff + idx * 8 * src_stride;
hadamard_8x8x2_avx2(src_ptr, src_stride, t_coeff + (idx * 64 * 2));
}
for (int idx = 0; idx < 64; idx += 16) {
const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
__m256i b0 = _mm256_add_epi16(coeff0, coeff1);
__m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
__m256i b2 = _mm256_add_epi16(coeff2, coeff3);
__m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
b0 = _mm256_srai_epi16(b0, 1);
b1 = _mm256_srai_epi16(b1, 1);
b2 = _mm256_srai_epi16(b2, 1);
b3 = _mm256_srai_epi16(b3, 1);
_mm256_storeu_si256((__m256i *)coeff, _mm256_add_epi16(b0, b2));
_mm256_storeu_si256((__m256i *)(coeff + 64), _mm256_add_epi16(b1, b3));
_mm256_storeu_si256((__m256i *)(coeff + 128), _mm256_sub_epi16(b0, b2));
_mm256_storeu_si256((__m256i *)(coeff + 192), _mm256_sub_epi16(b1, b3));
coeff += 16;
t_coeff += 16;
}
}
void aom_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
tran_low_t *coeff) {
// For high bitdepths, it is unnecessary to store_tran_low
// (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
// next stage. Output to an intermediate buffer first, then store_tran_low()
// in the final stage.
DECLARE_ALIGNED(32, int16_t, temp_coeff[32 * 32]);
int16_t *t_coeff = temp_coeff;
int idx;
for (idx = 0; idx < 4; ++idx) {
// src_diff: 9 bit, dynamic range [-255, 255]
const int16_t *src_ptr =
src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
hadamard_16x16_avx2(src_ptr, src_stride,
(tran_low_t *)(t_coeff + idx * 256), 0);
}
for (idx = 0; idx < 256; idx += 16) {
const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 256));
const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 512));
const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 768));
__m256i b0 = _mm256_add_epi16(coeff0, coeff1);
__m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
__m256i b2 = _mm256_add_epi16(coeff2, coeff3);
__m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
b0 = _mm256_srai_epi16(b0, 2);
b1 = _mm256_srai_epi16(b1, 2);
b2 = _mm256_srai_epi16(b2, 2);
b3 = _mm256_srai_epi16(b3, 2);
store_tran_low(_mm256_add_epi16(b0, b2), coeff);
store_tran_low(_mm256_add_epi16(b1, b3), coeff + 256);
store_tran_low(_mm256_sub_epi16(b0, b2), coeff + 512);
store_tran_low(_mm256_sub_epi16(b1, b3), coeff + 768);
coeff += 16;
t_coeff += 16;
}
}
static void highbd_hadamard_col8_avx2(__m256i *in, int iter) {
__m256i a0 = in[0];
__m256i a1 = in[1];
__m256i a2 = in[2];
__m256i a3 = in[3];
__m256i a4 = in[4];
__m256i a5 = in[5];
__m256i a6 = in[6];
__m256i a7 = in[7];
__m256i b0 = _mm256_add_epi32(a0, a1);
__m256i b1 = _mm256_sub_epi32(a0, a1);
__m256i b2 = _mm256_add_epi32(a2, a3);
__m256i b3 = _mm256_sub_epi32(a2, a3);
__m256i b4 = _mm256_add_epi32(a4, a5);
__m256i b5 = _mm256_sub_epi32(a4, a5);
__m256i b6 = _mm256_add_epi32(a6, a7);
__m256i b7 = _mm256_sub_epi32(a6, a7);
a0 = _mm256_add_epi32(b0, b2);
a1 = _mm256_add_epi32(b1, b3);
a2 = _mm256_sub_epi32(b0, b2);
a3 = _mm256_sub_epi32(b1, b3);
a4 = _mm256_add_epi32(b4, b6);
a5 = _mm256_add_epi32(b5, b7);
a6 = _mm256_sub_epi32(b4, b6);
a7 = _mm256_sub_epi32(b5, b7);
if (iter == 0) {
b0 = _mm256_add_epi32(a0, a4);
b7 = _mm256_add_epi32(a1, a5);
b3 = _mm256_add_epi32(a2, a6);
b4 = _mm256_add_epi32(a3, a7);
b2 = _mm256_sub_epi32(a0, a4);
b6 = _mm256_sub_epi32(a1, a5);
b1 = _mm256_sub_epi32(a2, a6);
b5 = _mm256_sub_epi32(a3, a7);
a0 = _mm256_unpacklo_epi32(b0, b1);
a1 = _mm256_unpacklo_epi32(b2, b3);
a2 = _mm256_unpackhi_epi32(b0, b1);
a3 = _mm256_unpackhi_epi32(b2, b3);
a4 = _mm256_unpacklo_epi32(b4, b5);
a5 = _mm256_unpacklo_epi32(b6, b7);
a6 = _mm256_unpackhi_epi32(b4, b5);
a7 = _mm256_unpackhi_epi32(b6, b7);
b0 = _mm256_unpacklo_epi64(a0, a1);
b1 = _mm256_unpacklo_epi64(a4, a5);
b2 = _mm256_unpackhi_epi64(a0, a1);
b3 = _mm256_unpackhi_epi64(a4, a5);
b4 = _mm256_unpacklo_epi64(a2, a3);
b5 = _mm256_unpacklo_epi64(a6, a7);
b6 = _mm256_unpackhi_epi64(a2, a3);
b7 = _mm256_unpackhi_epi64(a6, a7);
in[0] = _mm256_permute2x128_si256(b0, b1, 0x20);
in[1] = _mm256_permute2x128_si256(b0, b1, 0x31);
in[2] = _mm256_permute2x128_si256(b2, b3, 0x20);
in[3] = _mm256_permute2x128_si256(b2, b3, 0x31);
in[4] = _mm256_permute2x128_si256(b4, b5, 0x20);
in[5] = _mm256_permute2x128_si256(b4, b5, 0x31);
in[6] = _mm256_permute2x128_si256(b6, b7, 0x20);
in[7] = _mm256_permute2x128_si256(b6, b7, 0x31);
} else {
in[0] = _mm256_add_epi32(a0, a4);
in[7] = _mm256_add_epi32(a1, a5);
in[3] = _mm256_add_epi32(a2, a6);
in[4] = _mm256_add_epi32(a3, a7);
in[2] = _mm256_sub_epi32(a0, a4);
in[6] = _mm256_sub_epi32(a1, a5);
in[1] = _mm256_sub_epi32(a2, a6);
in[5] = _mm256_sub_epi32(a3, a7);
}
}
void aom_highbd_hadamard_8x8_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
tran_low_t *coeff) {
__m128i src16[8];
__m256i src32[8];
src16[0] = _mm_loadu_si128((const __m128i *)src_diff);
src16[1] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
src16[2] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
src16[3] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
src16[4] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
src16[5] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
src16[6] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
src16[7] = _mm_loadu_si128((const __m128i *)(src_diff + src_stride));
src32[0] = _mm256_cvtepi16_epi32(src16[0]);
src32[1] = _mm256_cvtepi16_epi32(src16[1]);
src32[2] = _mm256_cvtepi16_epi32(src16[2]);
src32[3] = _mm256_cvtepi16_epi32(src16[3]);
src32[4] = _mm256_cvtepi16_epi32(src16[4]);
src32[5] = _mm256_cvtepi16_epi32(src16[5]);
src32[6] = _mm256_cvtepi16_epi32(src16[6]);
src32[7] = _mm256_cvtepi16_epi32(src16[7]);
highbd_hadamard_col8_avx2(src32, 0);
highbd_hadamard_col8_avx2(src32, 1);
_mm256_storeu_si256((__m256i *)coeff, src32[0]);
coeff += 8;
_mm256_storeu_si256((__m256i *)coeff, src32[1]);
coeff += 8;
_mm256_storeu_si256((__m256i *)coeff, src32[2]);
coeff += 8;
_mm256_storeu_si256((__m256i *)coeff, src32[3]);
coeff += 8;
_mm256_storeu_si256((__m256i *)coeff, src32[4]);
coeff += 8;
_mm256_storeu_si256((__m256i *)coeff, src32[5]);
coeff += 8;
_mm256_storeu_si256((__m256i *)coeff, src32[6]);
coeff += 8;
_mm256_storeu_si256((__m256i *)coeff, src32[7]);
}
void aom_highbd_hadamard_16x16_avx2(const int16_t *src_diff,
ptrdiff_t src_stride, tran_low_t *coeff) {
int idx;
tran_low_t *t_coeff = coeff;
for (idx = 0; idx < 4; ++idx) {
const int16_t *src_ptr =
src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
aom_highbd_hadamard_8x8_avx2(src_ptr, src_stride, t_coeff + idx * 64);
}
for (idx = 0; idx < 64; idx += 8) {
__m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
__m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
__m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
__m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
__m256i b0 = _mm256_add_epi32(coeff0, coeff1);
__m256i b1 = _mm256_sub_epi32(coeff0, coeff1);
__m256i b2 = _mm256_add_epi32(coeff2, coeff3);
__m256i b3 = _mm256_sub_epi32(coeff2, coeff3);
b0 = _mm256_srai_epi32(b0, 1);
b1 = _mm256_srai_epi32(b1, 1);
b2 = _mm256_srai_epi32(b2, 1);
b3 = _mm256_srai_epi32(b3, 1);
coeff0 = _mm256_add_epi32(b0, b2);
coeff1 = _mm256_add_epi32(b1, b3);
coeff2 = _mm256_sub_epi32(b0, b2);
coeff3 = _mm256_sub_epi32(b1, b3);
_mm256_storeu_si256((__m256i *)coeff, coeff0);
_mm256_storeu_si256((__m256i *)(coeff + 64), coeff1);
_mm256_storeu_si256((__m256i *)(coeff + 128), coeff2);
_mm256_storeu_si256((__m256i *)(coeff + 192), coeff3);
coeff += 8;
t_coeff += 8;
}
}
void aom_highbd_hadamard_32x32_avx2(const int16_t *src_diff,
ptrdiff_t src_stride, tran_low_t *coeff) {
int idx;
tran_low_t *t_coeff = coeff;
for (idx = 0; idx < 4; ++idx) {
const int16_t *src_ptr =
src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
aom_highbd_hadamard_16x16_avx2(src_ptr, src_stride, t_coeff + idx * 256);
}
for (idx = 0; idx < 256; idx += 8) {
__m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
__m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 256));
__m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 512));
__m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 768));
__m256i b0 = _mm256_add_epi32(coeff0, coeff1);
__m256i b1 = _mm256_sub_epi32(coeff0, coeff1);
__m256i b2 = _mm256_add_epi32(coeff2, coeff3);
__m256i b3 = _mm256_sub_epi32(coeff2, coeff3);
b0 = _mm256_srai_epi32(b0, 2);
b1 = _mm256_srai_epi32(b1, 2);
b2 = _mm256_srai_epi32(b2, 2);
b3 = _mm256_srai_epi32(b3, 2);
coeff0 = _mm256_add_epi32(b0, b2);
coeff1 = _mm256_add_epi32(b1, b3);
coeff2 = _mm256_sub_epi32(b0, b2);
coeff3 = _mm256_sub_epi32(b1, b3);
_mm256_storeu_si256((__m256i *)coeff, coeff0);
_mm256_storeu_si256((__m256i *)(coeff + 256), coeff1);
_mm256_storeu_si256((__m256i *)(coeff + 512), coeff2);
_mm256_storeu_si256((__m256i *)(coeff + 768), coeff3);
coeff += 8;
t_coeff += 8;
}
}
int aom_satd_avx2(const tran_low_t *coeff, int length) {
__m256i accum = _mm256_setzero_si256();
int i;
for (i = 0; i < length; i += 8, coeff += 8) {
const __m256i src_line = _mm256_loadu_si256((const __m256i *)coeff);
const __m256i abs = _mm256_abs_epi32(src_line);
accum = _mm256_add_epi32(accum, abs);
}
{ // 32 bit horizontal add
const __m256i a = _mm256_srli_si256(accum, 8);
const __m256i b = _mm256_add_epi32(accum, a);
const __m256i c = _mm256_srli_epi64(b, 32);
const __m256i d = _mm256_add_epi32(b, c);
const __m128i accum_128 = _mm_add_epi32(_mm256_castsi256_si128(d),
_mm256_extractf128_si256(d, 1));
return _mm_cvtsi128_si32(accum_128);
}
}
int aom_satd_lp_avx2(const int16_t *coeff, int length) {
const __m256i one = _mm256_set1_epi16(1);
__m256i accum = _mm256_setzero_si256();
for (int i = 0; i < length; i += 16) {
const __m256i src_line = _mm256_loadu_si256((const __m256i *)coeff);
const __m256i abs = _mm256_abs_epi16(src_line);
const __m256i sum = _mm256_madd_epi16(abs, one);
accum = _mm256_add_epi32(accum, sum);
coeff += 16;
}
{ // 32 bit horizontal add
const __m256i a = _mm256_srli_si256(accum, 8);
const __m256i b = _mm256_add_epi32(accum, a);
const __m256i c = _mm256_srli_epi64(b, 32);
const __m256i d = _mm256_add_epi32(b, c);
const __m128i accum_128 = _mm_add_epi32(_mm256_castsi256_si128(d),
_mm256_extractf128_si256(d, 1));
return _mm_cvtsi128_si32(accum_128);
}
}