libaom/av1/common/x86/wiener_convolve_avx2.c - platform/external/libaom - Git at Google

 /*
  * Copyright (c) 2018, 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 <assert.h>

 #include "config/av1_rtcd.h"

 #include "av1/common/convolve.h"
 #include "aom_dsp/aom_dsp_common.h"
 #include "aom_dsp/aom_filter.h"
 #include "aom_dsp/x86/synonyms.h"
 #include "aom_dsp/x86/synonyms_avx2.h"

 // 128-bit xmmwords are written as [ ... ] with the MSB on the left.
 // 256-bit ymmwords are written as two xmmwords, [ ... ][ ... ] with the MSB
 // on the left.
 // A row of, say, 8-bit pixels with values p0, p1, p2, ..., p30, p31 will be
 // loaded and stored as [ p31 ... p17 p16 ][ p15 ... p1 p0 ].
 void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
                                       uint8_t *dst, ptrdiff_t dst_stride,
                                       const int16_t *filter_x, int x_step_q4,
                                       const int16_t *filter_y, int y_step_q4,
                                       int w, int h,
                                       const ConvolveParams *conv_params) {
   const int bd = 8;
   assert(x_step_q4 == 16 && y_step_q4 == 16);
   assert(!(w & 7));
   (void)x_step_q4;
   (void)y_step_q4;

   DECLARE_ALIGNED(32, uint16_t,
                   temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
   int intermediate_height = h + SUBPEL_TAPS - 2;
   memset(temp + (intermediate_height * MAX_SB_SIZE), 0, MAX_SB_SIZE);
   const int center_tap = ((SUBPEL_TAPS - 1) / 2);
   const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap;

   const __m128i zero_128 = _mm_setzero_si128();
   const __m256i zero_256 = _mm256_setzero_si256();

   // Add an offset to account for the "add_src" part of the convolve function.
   const __m128i offset = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3);

   const __m256i clamp_low = zero_256;
   const __m256i clamp_high =
       _mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1);

   /* Horizontal filter */
   {
     // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
     const __m128i coeffs_x = _mm_add_epi16(xx_loadu_128(filter_x), offset);

     // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
     const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
     // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
     const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);

     // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
     const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
     // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
     const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
     // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
     const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
     // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
     const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);

     // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
     const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
     // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
     const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
     // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
     const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
     // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
     const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);

     const __m256i round_const = _mm256_set1_epi32(
         (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));

     for (int i = 0; i < intermediate_height; ++i) {
       for (int j = 0; j < w; j += 16) {
         const uint8_t *data_ij = src_ptr + i * src_stride + j;

         // Load 8-bit src data
         const __m128i data_0 = xx_loadu_128(data_ij + 0);
         const __m128i data_1 = xx_loadu_128(data_ij + 1);
         const __m128i data_2 = xx_loadu_128(data_ij + 2);
         const __m128i data_3 = xx_loadu_128(data_ij + 3);
         const __m128i data_4 = xx_loadu_128(data_ij + 4);
         const __m128i data_5 = xx_loadu_128(data_ij + 5);
         const __m128i data_6 = xx_loadu_128(data_ij + 6);
         const __m128i data_7 = xx_loadu_128(data_ij + 7);

         // (Zero-)Extend 8-bit data to 16-bit data
         const __m256i src_0 = _mm256_cvtepu8_epi16(data_0);
         const __m256i src_1 = _mm256_cvtepu8_epi16(data_1);
         const __m256i src_2 = _mm256_cvtepu8_epi16(data_2);
         const __m256i src_3 = _mm256_cvtepu8_epi16(data_3);
         const __m256i src_4 = _mm256_cvtepu8_epi16(data_4);
         const __m256i src_5 = _mm256_cvtepu8_epi16(data_5);
         const __m256i src_6 = _mm256_cvtepu8_epi16(data_6);
         const __m256i src_7 = _mm256_cvtepu8_epi16(data_7);

         // Multiply src data by filter coeffs and sum pairs
         const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
         const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
         const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
         const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
         const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
         const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
         const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
         const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);

         // Calculate scalar product for even- and odd-indices separately,
         // increasing to 32-bit precision
         const __m256i res_even_sum = _mm256_add_epi32(
             _mm256_add_epi32(res_0, res_4), _mm256_add_epi32(res_2, res_6));
         const __m256i res_odd_sum = _mm256_add_epi32(
             _mm256_add_epi32(res_1, res_5), _mm256_add_epi32(res_3, res_7));

         const __m256i res_even = _mm256_srai_epi32(
             _mm256_add_epi32(res_even_sum, round_const), conv_params->round_0);
         const __m256i res_odd = _mm256_srai_epi32(
             _mm256_add_epi32(res_odd_sum, round_const), conv_params->round_0);

         // Reduce to 16-bit precision and pack even- and odd-index results
         // back into one register. The _mm256_packs_epi32 intrinsic returns
         // a register with the pixels ordered as follows:
         // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
         const __m256i res = _mm256_packs_epi32(res_even, res_odd);
         const __m256i res_clamped =
             _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high);

         // Store in a temporary array
         yy_storeu_256(temp + i * MAX_SB_SIZE + j, res_clamped);
       }
     }
   }

   /* Vertical filter */
   {
     // coeffs [ g7 g6 g5 g4 g3 g2 g1 g0 ]
     const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset);

     // coeffs [ g3 g2 g3 g2 g1 g0 g1 g0 ]
     const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
     // coeffs [ g7 g6 g7 g6 g5 g4 g5 g4 ]
     const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);

     // coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ]
     const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
     // coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ]
     const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
     // coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ]
     const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
     // coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ]
     const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);

     // coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ][ g1 g0 g1 g0 g1 g0 g1 g0 ]
     const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
     // coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ][ g3 g2 g3 g2 g3 g2 g3 g2 ]
     const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
     // coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ][ g5 g4 g5 g4 g5 g4 g5 g4 ]
     const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
     // coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ][ g7 g6 g7 g6 g7 g6 g7 g6 ]
     const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);

     const __m256i round_const =
         _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
                           (1 << (bd + conv_params->round_1 - 1)));

     for (int i = 0; i < h; ++i) {
       for (int j = 0; j < w; j += 16) {
         const uint16_t *data_ij = temp + i * MAX_SB_SIZE + j;

         // Load 16-bit data from the output of the horizontal filter in
         // which the pixels are ordered as follows:
         // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
         const __m256i data_0 = yy_loadu_256(data_ij + 0 * MAX_SB_SIZE);
         const __m256i data_1 = yy_loadu_256(data_ij + 1 * MAX_SB_SIZE);
         const __m256i data_2 = yy_loadu_256(data_ij + 2 * MAX_SB_SIZE);
         const __m256i data_3 = yy_loadu_256(data_ij + 3 * MAX_SB_SIZE);
         const __m256i data_4 = yy_loadu_256(data_ij + 4 * MAX_SB_SIZE);
         const __m256i data_5 = yy_loadu_256(data_ij + 5 * MAX_SB_SIZE);
         const __m256i data_6 = yy_loadu_256(data_ij + 6 * MAX_SB_SIZE);
         const __m256i data_7 = yy_loadu_256(data_ij + 7 * MAX_SB_SIZE);

         // Filter the even-indices, increasing to 32-bit precision
         const __m256i src_0 = _mm256_unpacklo_epi16(data_0, data_1);
         const __m256i src_2 = _mm256_unpacklo_epi16(data_2, data_3);
         const __m256i src_4 = _mm256_unpacklo_epi16(data_4, data_5);
         const __m256i src_6 = _mm256_unpacklo_epi16(data_6, data_7);

         const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
         const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
         const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
         const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);

         const __m256i res_even = _mm256_add_epi32(
             _mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6));

         // Filter the odd-indices, increasing to 32-bit precision
         const __m256i src_1 = _mm256_unpackhi_epi16(data_0, data_1);
         const __m256i src_3 = _mm256_unpackhi_epi16(data_2, data_3);
         const __m256i src_5 = _mm256_unpackhi_epi16(data_4, data_5);
         const __m256i src_7 = _mm256_unpackhi_epi16(data_6, data_7);

         const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
         const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
         const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
         const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);

         const __m256i res_odd = _mm256_add_epi32(
             _mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7));

         // Pixels are currently in the following order:
         // res_even order: [ 14 12 10 8 ] [ 6 4 2 0 ]
         // res_odd order:  [ 15 13 11 9 ] [ 7 5 3 1 ]
         //
         // Rearrange the pixels into the following order:
         // res_lo order: [ 11 10  9  8 ] [ 3 2 1 0 ]
         // res_hi order: [ 15 14 13 12 ] [ 7 6 5 4 ]
         const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
         const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd);

         const __m256i res_lo_round = _mm256_srai_epi32(
             _mm256_add_epi32(res_lo, round_const), conv_params->round_1);
         const __m256i res_hi_round = _mm256_srai_epi32(
             _mm256_add_epi32(res_hi, round_const), conv_params->round_1);

         // Reduce to 16-bit precision and pack into the correct order:
         // [ 15 14 13 12 11 10 9 8 ][ 7 6 5 4 3 2 1 0 ]
         const __m256i res_16bit =
             _mm256_packs_epi32(res_lo_round, res_hi_round);

         // Reduce to 8-bit precision. This messes up the order:
         // [ - - - - - - - - 15 14 13 12 11 10 9 8 ]
         // [ - - - - - - - - 7 6 5 4 3 2 1 0 ]
         const __m256i res_8bit =
             _mm256_packus_epi16(res_16bit, zero_256 /* don't care value */);

         // Swap the two central 32-bit values to get the order:
         // [ - - - - - - - - - - - - - - - - ]
         // [ 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 ]
         const __m256i res_8bit2 = _mm256_permute4x64_epi64(res_8bit, 0xd8);

         // Store the lower 128-bit lane in the dst array
         xx_storeu_128(dst + i * dst_stride + j,
                       _mm256_castsi256_si128(res_8bit2));
       }
     }
   }
 }
	/*
	* Copyright (c) 2018, 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 <assert.h>

	#include "config/av1_rtcd.h"

	#include "av1/common/convolve.h"
	#include "aom_dsp/aom_dsp_common.h"
	#include "aom_dsp/aom_filter.h"
	#include "aom_dsp/x86/synonyms.h"
	#include "aom_dsp/x86/synonyms_avx2.h"

	// 128-bit xmmwords are written as [ ... ] with the MSB on the left.
	// 256-bit ymmwords are written as two xmmwords, [ ... ][ ... ] with the MSB
	// on the left.
	// A row of, say, 8-bit pixels with values p0, p1, p2, ..., p30, p31 will be
	// loaded and stored as [ p31 ... p17 p16 ][ p15 ... p1 p0 ].
	void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
	uint8_t *dst, ptrdiff_t dst_stride,
	const int16_t *filter_x, int x_step_q4,
	const int16_t *filter_y, int y_step_q4,
	int w, int h,
	const ConvolveParams *conv_params) {
	const int bd = 8;
	assert(x_step_q4 == 16 && y_step_q4 == 16);
	assert(!(w & 7));
	(void)x_step_q4;
	(void)y_step_q4;

	DECLARE_ALIGNED(32, uint16_t,
	temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
	int intermediate_height = h + SUBPEL_TAPS - 2;
	memset(temp + (intermediate_height * MAX_SB_SIZE), 0, MAX_SB_SIZE);
	const int center_tap = ((SUBPEL_TAPS - 1) / 2);
	const uint8_t const src_ptr = src - center_tap src_stride - center_tap;

	const __m128i zero_128 = _mm_setzero_si128();
	const __m256i zero_256 = _mm256_setzero_si256();

	// Add an offset to account for the "add_src" part of the convolve function.
	const __m128i offset = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3);

	const __m256i clamp_low = zero_256;
	const __m256i clamp_high =
	_mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1);

	/* Horizontal filter */
	{
	// coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
	const __m128i coeffs_x = _mm_add_epi16(xx_loadu_128(filter_x), offset);

	// coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
	const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
	// coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
	const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);

	// coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
	const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
	// coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
	const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
	// coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
	const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
	// coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
	const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);

	// coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
	const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
	// coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
	const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
	// coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
	const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
	// coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
	const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);

	const __m256i round_const = _mm256_set1_epi32(
	(1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));

	for (int i = 0; i < intermediate_height; ++i) {
	for (int j = 0; j < w; j += 16) {
	const uint8_t data_ij = src_ptr + i src_stride + j;

	// Load 8-bit src data
	const __m128i data_0 = xx_loadu_128(data_ij + 0);
	const __m128i data_1 = xx_loadu_128(data_ij + 1);
	const __m128i data_2 = xx_loadu_128(data_ij + 2);
	const __m128i data_3 = xx_loadu_128(data_ij + 3);
	const __m128i data_4 = xx_loadu_128(data_ij + 4);
	const __m128i data_5 = xx_loadu_128(data_ij + 5);
	const __m128i data_6 = xx_loadu_128(data_ij + 6);
	const __m128i data_7 = xx_loadu_128(data_ij + 7);

	// (Zero-)Extend 8-bit data to 16-bit data
	const __m256i src_0 = _mm256_cvtepu8_epi16(data_0);
	const __m256i src_1 = _mm256_cvtepu8_epi16(data_1);
	const __m256i src_2 = _mm256_cvtepu8_epi16(data_2);
	const __m256i src_3 = _mm256_cvtepu8_epi16(data_3);
	const __m256i src_4 = _mm256_cvtepu8_epi16(data_4);
	const __m256i src_5 = _mm256_cvtepu8_epi16(data_5);
	const __m256i src_6 = _mm256_cvtepu8_epi16(data_6);
	const __m256i src_7 = _mm256_cvtepu8_epi16(data_7);

	// Multiply src data by filter coeffs and sum pairs
	const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
	const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
	const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
	const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
	const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
	const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
	const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
	const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);

	// Calculate scalar product for even- and odd-indices separately,
	// increasing to 32-bit precision
	const __m256i res_even_sum = _mm256_add_epi32(
	_mm256_add_epi32(res_0, res_4), _mm256_add_epi32(res_2, res_6));
	const __m256i res_odd_sum = _mm256_add_epi32(
	_mm256_add_epi32(res_1, res_5), _mm256_add_epi32(res_3, res_7));

	const __m256i res_even = _mm256_srai_epi32(
	_mm256_add_epi32(res_even_sum, round_const), conv_params->round_0);
	const __m256i res_odd = _mm256_srai_epi32(
	_mm256_add_epi32(res_odd_sum, round_const), conv_params->round_0);

	// Reduce to 16-bit precision and pack even- and odd-index results
	// back into one register. The _mm256_packs_epi32 intrinsic returns
	// a register with the pixels ordered as follows:
	// [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
	const __m256i res = _mm256_packs_epi32(res_even, res_odd);
	const __m256i res_clamped =
	_mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high);

	// Store in a temporary array
	yy_storeu_256(temp + i * MAX_SB_SIZE + j, res_clamped);
	}
	}
	}

	/* Vertical filter */
	{
	// coeffs [ g7 g6 g5 g4 g3 g2 g1 g0 ]
	const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset);

	// coeffs [ g3 g2 g3 g2 g1 g0 g1 g0 ]
	const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
	// coeffs [ g7 g6 g7 g6 g5 g4 g5 g4 ]
	const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);

	// coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ]
	const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
	// coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ]
	const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
	// coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ]
	const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
	// coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ]
	const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);

	// coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ][ g1 g0 g1 g0 g1 g0 g1 g0 ]
	const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
	// coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ][ g3 g2 g3 g2 g3 g2 g3 g2 ]
	const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
	// coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ][ g5 g4 g5 g4 g5 g4 g5 g4 ]
	const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
	// coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ][ g7 g6 g7 g6 g7 g6 g7 g6 ]
	const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);

	const __m256i round_const =
	_mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
	(1 << (bd + conv_params->round_1 - 1)));

	for (int i = 0; i < h; ++i) {
	for (int j = 0; j < w; j += 16) {
	const uint16_t data_ij = temp + i MAX_SB_SIZE + j;

	// Load 16-bit data from the output of the horizontal filter in
	// which the pixels are ordered as follows:
	// [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
	const __m256i data_0 = yy_loadu_256(data_ij + 0 * MAX_SB_SIZE);
	const __m256i data_1 = yy_loadu_256(data_ij + 1 * MAX_SB_SIZE);
	const __m256i data_2 = yy_loadu_256(data_ij + 2 * MAX_SB_SIZE);
	const __m256i data_3 = yy_loadu_256(data_ij + 3 * MAX_SB_SIZE);
	const __m256i data_4 = yy_loadu_256(data_ij + 4 * MAX_SB_SIZE);
	const __m256i data_5 = yy_loadu_256(data_ij + 5 * MAX_SB_SIZE);
	const __m256i data_6 = yy_loadu_256(data_ij + 6 * MAX_SB_SIZE);
	const __m256i data_7 = yy_loadu_256(data_ij + 7 * MAX_SB_SIZE);

	// Filter the even-indices, increasing to 32-bit precision
	const __m256i src_0 = _mm256_unpacklo_epi16(data_0, data_1);
	const __m256i src_2 = _mm256_unpacklo_epi16(data_2, data_3);
	const __m256i src_4 = _mm256_unpacklo_epi16(data_4, data_5);
	const __m256i src_6 = _mm256_unpacklo_epi16(data_6, data_7);

	const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
	const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
	const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
	const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);

	const __m256i res_even = _mm256_add_epi32(
	_mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6));

	// Filter the odd-indices, increasing to 32-bit precision
	const __m256i src_1 = _mm256_unpackhi_epi16(data_0, data_1);
	const __m256i src_3 = _mm256_unpackhi_epi16(data_2, data_3);
	const __m256i src_5 = _mm256_unpackhi_epi16(data_4, data_5);
	const __m256i src_7 = _mm256_unpackhi_epi16(data_6, data_7);

	const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
	const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
	const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
	const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);

	const __m256i res_odd = _mm256_add_epi32(
	_mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7));

	// Pixels are currently in the following order:
	// res_even order: [ 14 12 10 8 ] [ 6 4 2 0 ]
	// res_odd order: [ 15 13 11 9 ] [ 7 5 3 1 ]
	//
	// Rearrange the pixels into the following order:
	// res_lo order: [ 11 10 9 8 ] [ 3 2 1 0 ]
	// res_hi order: [ 15 14 13 12 ] [ 7 6 5 4 ]
	const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
	const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd);

	const __m256i res_lo_round = _mm256_srai_epi32(
	_mm256_add_epi32(res_lo, round_const), conv_params->round_1);
	const __m256i res_hi_round = _mm256_srai_epi32(
	_mm256_add_epi32(res_hi, round_const), conv_params->round_1);

	// Reduce to 16-bit precision and pack into the correct order:
	// [ 15 14 13 12 11 10 9 8 ][ 7 6 5 4 3 2 1 0 ]
	const __m256i res_16bit =
	_mm256_packs_epi32(res_lo_round, res_hi_round);

	// Reduce to 8-bit precision. This messes up the order:
	// [ - - - - - - - - 15 14 13 12 11 10 9 8 ]
	// [ - - - - - - - - 7 6 5 4 3 2 1 0 ]
	const __m256i res_8bit =
	_mm256_packus_epi16(res_16bit, zero_256 /* don't care value */);

	// Swap the two central 32-bit values to get the order:
	// [ - - - - - - - - - - - - - - - - ]
	// [ 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 ]
	const __m256i res_8bit2 = _mm256_permute4x64_epi64(res_8bit, 0xd8);

	// Store the lower 128-bit lane in the dst array
	xx_storeu_128(dst + i * dst_stride + j,
	_mm256_castsi256_si128(res_8bit2));
	}
	}
	}
	}