av1/common/x86/wiener_convolve_avx2.c - aom - 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/convolve_avx2.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 ].

 // Exploiting the range of wiener filter coefficients,
 // horizontal filtering can be done in 16 bit intermediate precision.
 // The details are as follows :
 // Consider the horizontal wiener filter coefficients of the following form :
 //      [C0, C1, C2, 2^(FILTER_BITS) -2 * (C0 + C1 + C2), C2, C1, C0]
 // Subtracting  2^(FILTER_BITS) from the centre tap we get the following  :
 //      [C0, C1, C2,     -2 * (C0 + C1 + C2),             C2, C1, C0]
 // The sum of the product "C0 * p0 + C1 * p1 + C2 * p2 -2 * (C0 + C1 + C2) * p3
 // + C2 * p4 + C1 * p5 + C0 * p6" would be in the range of signed 16 bit
 // precision. Finally, after rounding the above result by round_0, we multiply
 // the centre pixel by 2^(FILTER_BITS - round_0) and add it to get the
 // horizontal filter output.

 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 filter_bits = conv_params->filter_bits;
   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, int16_t, im_block[(MAX_SB_SIZE + SUBPEL_TAPS) * 8]);
   int im_h = h + SUBPEL_TAPS - 2;
   int im_stride = 8;
   memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE);
   int i, j;
   const int center_tap = (SUBPEL_TAPS - 1) / 2;
   const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap;

   __m256i filt[4], coeffs_h[4], coeffs_v[4], filt_center;

   assert(conv_params->round_0 > 0);

   filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
   filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
   filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
   filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);

   filt_center = _mm256_load_si256((__m256i const *)filt_center_global_avx2);

   const __m128i coeffs_x = _mm_loadu_si128((__m128i *)filter_x);
   const __m256i filter_coeffs_x = _mm256_broadcastsi128_si256(coeffs_x);

   // coeffs 0 1 0 1 0 1 0 1
   coeffs_h[0] =
       _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0200u));
   // coeffs 2 3 2 3 2 3 2 3
   coeffs_h[1] =
       _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0604u));
   // coeffs 4 5 4 5 4 5 4 5
   coeffs_h[2] =
       _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0a08u));
   // coeffs 6 7 6 7 6 7 6 7
   coeffs_h[3] =
       _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0e0cu));

   const __m256i round_const_h =
       _mm256_set1_epi16((1 << (conv_params->round_0 - 1)));
   const __m256i round_const_horz =
       _mm256_set1_epi16((1 << (bd + filter_bits - conv_params->round_0 - 1)));
   const __m256i clamp_low = _mm256_setzero_si256();
   const __m256i clamp_high = _mm256_set1_epi16(
       WIENER_CLAMP_LIMIT(filter_bits, conv_params->round_0, bd) - 1);
   const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0);

   // Add an offset to account for the "add_src" part of the convolve function.
   const __m128i zero_128 = _mm_setzero_si128();
   const __m128i offset_0 = _mm_insert_epi16(zero_128, 1 << filter_bits, 3);
   const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset_0);

   const __m256i filter_coeffs_y = _mm256_broadcastsi128_si256(coeffs_y);

   // coeffs 0 1 0 1 0 1 0 1
   coeffs_v[0] = _mm256_shuffle_epi32(filter_coeffs_y, 0x00);
   // coeffs 2 3 2 3 2 3 2 3
   coeffs_v[1] = _mm256_shuffle_epi32(filter_coeffs_y, 0x55);
   // coeffs 4 5 4 5 4 5 4 5
   coeffs_v[2] = _mm256_shuffle_epi32(filter_coeffs_y, 0xaa);
   // coeffs 6 7 6 7 6 7 6 7
   coeffs_v[3] = _mm256_shuffle_epi32(filter_coeffs_y, 0xff);

   const __m256i round_const_v =
       _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
                         (1 << (bd + conv_params->round_1 - 1)));
   const __m128i round_shift_v = _mm_cvtsi32_si128(conv_params->round_1);

   for (j = 0; j < w; j += 8) {
     for (i = 0; i < im_h; i += 2) {
       __m256i data = _mm256_castsi128_si256(
           _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));

       // Load the next line
       if (i + 1 < im_h)
         data = _mm256_inserti128_si256(
             data,
             _mm_loadu_si128(
                 (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),
             1);

       __m256i res = convolve_lowbd_x(data, coeffs_h, filt);

       res =
           _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h);

       __m256i data_0 = _mm256_shuffle_epi8(data, filt_center);

       // multiply the center pixel by 2^(filter_bits - round_0) and add it to
       // the result
       data_0 = _mm256_slli_epi16(data_0, filter_bits - conv_params->round_0);
       res = _mm256_add_epi16(res, data_0);
       res = _mm256_add_epi16(res, round_const_horz);
       const __m256i res_clamped =
           _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high);
       _mm256_store_si256((__m256i *)&im_block[i * im_stride], res_clamped);
     }

     /* Vertical filter */
     {
       __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));
       __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));
       __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));
       __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));
       __m256i src_4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride));
       __m256i src_5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride));

       __m256i s[8];
       s[0] = _mm256_unpacklo_epi16(src_0, src_1);
       s[1] = _mm256_unpacklo_epi16(src_2, src_3);
       s[2] = _mm256_unpacklo_epi16(src_4, src_5);

       s[4] = _mm256_unpackhi_epi16(src_0, src_1);
       s[5] = _mm256_unpackhi_epi16(src_2, src_3);
       s[6] = _mm256_unpackhi_epi16(src_4, src_5);

       for (i = 0; i < h - 1; i += 2) {
         const int16_t *data = &im_block[i * im_stride];

         const __m256i s6 =
             _mm256_loadu_si256((__m256i *)(data + 6 * im_stride));
         const __m256i s7 =
             _mm256_loadu_si256((__m256i *)(data + 7 * im_stride));

         s[3] = _mm256_unpacklo_epi16(s6, s7);
         s[7] = _mm256_unpackhi_epi16(s6, s7);

         __m256i res_a = convolve(s, coeffs_v);
         __m256i res_b = convolve(s + 4, coeffs_v);

         const __m256i res_a_round = _mm256_sra_epi32(
             _mm256_add_epi32(res_a, round_const_v), round_shift_v);
         const __m256i res_b_round = _mm256_sra_epi32(
             _mm256_add_epi32(res_b, round_const_v), round_shift_v);

         /* rounding code */
         // 16 bit conversion
         const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
         // 8 bit conversion and saturation to uint8
         const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);

         const __m128i res_0 = _mm256_castsi256_si128(res_8b);
         const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);

         // Store values into the destination buffer
         __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
         __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];

         _mm_storel_epi64(p_0, res_0);
         _mm_storel_epi64(p_1, res_1);

         s[0] = s[1];
         s[1] = s[2];
         s[2] = s[3];

         s[4] = s[5];
         s[5] = s[6];
         s[6] = s[7];
       }
       if (h - i) {
         s[0] = _mm256_permute2x128_si256(s[0], s[4], 0x20);
         s[1] = _mm256_permute2x128_si256(s[1], s[5], 0x20);
         s[2] = _mm256_permute2x128_si256(s[2], s[6], 0x20);

         const int16_t *data = &im_block[i * im_stride];
         const __m128i s6_ = _mm_loadu_si128((__m128i *)(data + 6 * im_stride));
         const __m128i s7_ = _mm_loadu_si128((__m128i *)(data + 7 * im_stride));

         __m128i s3 = _mm_unpacklo_epi16(s6_, s7_);
         __m128i s7 = _mm_unpackhi_epi16(s6_, s7_);

         s[3] = _mm256_inserti128_si256(_mm256_castsi128_si256(s3), s7, 1);
         __m256i convolveres = convolve(s, coeffs_v);

         const __m256i res_round = _mm256_sra_epi32(
             _mm256_add_epi32(convolveres, round_const_v), round_shift_v);

         /* rounding code */
         // 16 bit conversion
         __m128i reslo = _mm256_castsi256_si128(res_round);
         __m128i reshi = _mm256_extracti128_si256(res_round, 1);
         const __m128i res_16bit = _mm_packus_epi32(reslo, reshi);

         // 8 bit conversion and saturation to uint8
         const __m128i res_8b = _mm_packus_epi16(res_16bit, res_16bit);
         __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
         _mm_storel_epi64(p_0, res_8b);
       }
     }
   }
 }
	/*
	* 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/convolve_avx2.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 ].

	// Exploiting the range of wiener filter coefficients,
	// horizontal filtering can be done in 16 bit intermediate precision.
	// The details are as follows :
	// Consider the horizontal wiener filter coefficients of the following form :
	// [C0, C1, C2, 2^(FILTER_BITS) -2 * (C0 + C1 + C2), C2, C1, C0]
	// Subtracting 2^(FILTER_BITS) from the centre tap we get the following :
	// [C0, C1, C2, -2 * (C0 + C1 + C2), C2, C1, C0]
	// The sum of the product "C0 * p0 + C1 * p1 + C2 * p2 -2 * (C0 + C1 + C2) * p3
	// + C2 * p4 + C1 * p5 + C0 * p6" would be in the range of signed 16 bit
	// precision. Finally, after rounding the above result by round_0, we multiply
	// the centre pixel by 2^(FILTER_BITS - round_0) and add it to get the
	// horizontal filter output.

	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 filter_bits = conv_params->filter_bits;
	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, int16_t, im_block[(MAX_SB_SIZE + SUBPEL_TAPS) * 8]);
	int im_h = h + SUBPEL_TAPS - 2;
	int im_stride = 8;
	memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE);
	int i, j;
	const int center_tap = (SUBPEL_TAPS - 1) / 2;
	const uint8_t const src_ptr = src - center_tap src_stride - center_tap;

	__m256i filt[4], coeffs_h[4], coeffs_v[4], filt_center;

	assert(conv_params->round_0 > 0);

	filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
	filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
	filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
	filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);

	filt_center = _mm256_load_si256((__m256i const *)filt_center_global_avx2);

	const __m128i coeffs_x = _mm_loadu_si128((__m128i *)filter_x);
	const __m256i filter_coeffs_x = _mm256_broadcastsi128_si256(coeffs_x);

	// coeffs 0 1 0 1 0 1 0 1
	coeffs_h[0] =
	_mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0200u));
	// coeffs 2 3 2 3 2 3 2 3
	coeffs_h[1] =
	_mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0604u));
	// coeffs 4 5 4 5 4 5 4 5
	coeffs_h[2] =
	_mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0a08u));
	// coeffs 6 7 6 7 6 7 6 7
	coeffs_h[3] =
	_mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0e0cu));

	const __m256i round_const_h =
	_mm256_set1_epi16((1 << (conv_params->round_0 - 1)));
	const __m256i round_const_horz =
	_mm256_set1_epi16((1 << (bd + filter_bits - conv_params->round_0 - 1)));
	const __m256i clamp_low = _mm256_setzero_si256();
	const __m256i clamp_high = _mm256_set1_epi16(
	WIENER_CLAMP_LIMIT(filter_bits, conv_params->round_0, bd) - 1);
	const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0);

	// Add an offset to account for the "add_src" part of the convolve function.
	const __m128i zero_128 = _mm_setzero_si128();
	const __m128i offset_0 = _mm_insert_epi16(zero_128, 1 << filter_bits, 3);
	const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset_0);

	const __m256i filter_coeffs_y = _mm256_broadcastsi128_si256(coeffs_y);

	// coeffs 0 1 0 1 0 1 0 1
	coeffs_v[0] = _mm256_shuffle_epi32(filter_coeffs_y, 0x00);
	// coeffs 2 3 2 3 2 3 2 3
	coeffs_v[1] = _mm256_shuffle_epi32(filter_coeffs_y, 0x55);
	// coeffs 4 5 4 5 4 5 4 5
	coeffs_v[2] = _mm256_shuffle_epi32(filter_coeffs_y, 0xaa);
	// coeffs 6 7 6 7 6 7 6 7
	coeffs_v[3] = _mm256_shuffle_epi32(filter_coeffs_y, 0xff);

	const __m256i round_const_v =
	_mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
	(1 << (bd + conv_params->round_1 - 1)));
	const __m128i round_shift_v = _mm_cvtsi32_si128(conv_params->round_1);

	for (j = 0; j < w; j += 8) {
	for (i = 0; i < im_h; i += 2) {
	__m256i data = _mm256_castsi128_si256(
	_mm_loadu_si128((__m128i )&src_ptr[(i src_stride) + j]));

	// Load the next line
	if (i + 1 < im_h)
	data = _mm256_inserti128_si256(
	data,
	_mm_loadu_si128(
	(__m128i )&src_ptr[(i src_stride) + j + src_stride]),
	1);

	__m256i res = convolve_lowbd_x(data, coeffs_h, filt);

	res =
	_mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h);

	__m256i data_0 = _mm256_shuffle_epi8(data, filt_center);

	// multiply the center pixel by 2^(filter_bits - round_0) and add it to
	// the result
	data_0 = _mm256_slli_epi16(data_0, filter_bits - conv_params->round_0);
	res = _mm256_add_epi16(res, data_0);
	res = _mm256_add_epi16(res, round_const_horz);
	const __m256i res_clamped =
	_mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high);
	_mm256_store_si256((__m256i )&im_block[i im_stride], res_clamped);
	}

	/* Vertical filter */
	{
	__m256i src_0 = _mm256_loadu_si256((__m256i )(im_block + 0 im_stride));
	__m256i src_1 = _mm256_loadu_si256((__m256i )(im_block + 1 im_stride));
	__m256i src_2 = _mm256_loadu_si256((__m256i )(im_block + 2 im_stride));
	__m256i src_3 = _mm256_loadu_si256((__m256i )(im_block + 3 im_stride));
	__m256i src_4 = _mm256_loadu_si256((__m256i )(im_block + 4 im_stride));
	__m256i src_5 = _mm256_loadu_si256((__m256i )(im_block + 5 im_stride));

	__m256i s[8];
	s[0] = _mm256_unpacklo_epi16(src_0, src_1);
	s[1] = _mm256_unpacklo_epi16(src_2, src_3);
	s[2] = _mm256_unpacklo_epi16(src_4, src_5);

	s[4] = _mm256_unpackhi_epi16(src_0, src_1);
	s[5] = _mm256_unpackhi_epi16(src_2, src_3);
	s[6] = _mm256_unpackhi_epi16(src_4, src_5);

	for (i = 0; i < h - 1; i += 2) {
	const int16_t data = &im_block[i im_stride];

	const __m256i s6 =
	_mm256_loadu_si256((__m256i )(data + 6 im_stride));
	const __m256i s7 =
	_mm256_loadu_si256((__m256i )(data + 7 im_stride));

	s[3] = _mm256_unpacklo_epi16(s6, s7);
	s[7] = _mm256_unpackhi_epi16(s6, s7);

	__m256i res_a = convolve(s, coeffs_v);
	__m256i res_b = convolve(s + 4, coeffs_v);

	const __m256i res_a_round = _mm256_sra_epi32(
	_mm256_add_epi32(res_a, round_const_v), round_shift_v);
	const __m256i res_b_round = _mm256_sra_epi32(
	_mm256_add_epi32(res_b, round_const_v), round_shift_v);

	/* rounding code */
	// 16 bit conversion
	const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
	// 8 bit conversion and saturation to uint8
	const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);

	const __m128i res_0 = _mm256_castsi256_si128(res_8b);
	const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);

	// Store values into the destination buffer
	__m128i const p_0 = (__m128i )&dst[i * dst_stride + j];
	__m128i const p_1 = (__m128i )&dst[i * dst_stride + j + dst_stride];

	_mm_storel_epi64(p_0, res_0);
	_mm_storel_epi64(p_1, res_1);

	s[0] = s[1];
	s[1] = s[2];
	s[2] = s[3];

	s[4] = s[5];
	s[5] = s[6];
	s[6] = s[7];
	}
	if (h - i) {
	s[0] = _mm256_permute2x128_si256(s[0], s[4], 0x20);
	s[1] = _mm256_permute2x128_si256(s[1], s[5], 0x20);
	s[2] = _mm256_permute2x128_si256(s[2], s[6], 0x20);

	const int16_t data = &im_block[i im_stride];
	const __m128i s6_ = _mm_loadu_si128((__m128i )(data + 6 im_stride));
	const __m128i s7_ = _mm_loadu_si128((__m128i )(data + 7 im_stride));

	__m128i s3 = _mm_unpacklo_epi16(s6_, s7_);
	__m128i s7 = _mm_unpackhi_epi16(s6_, s7_);

	s[3] = _mm256_inserti128_si256(_mm256_castsi128_si256(s3), s7, 1);
	__m256i convolveres = convolve(s, coeffs_v);

	const __m256i res_round = _mm256_sra_epi32(
	_mm256_add_epi32(convolveres, round_const_v), round_shift_v);

	/* rounding code */
	// 16 bit conversion
	__m128i reslo = _mm256_castsi256_si128(res_round);
	__m128i reshi = _mm256_extracti128_si256(res_round, 1);
	const __m128i res_16bit = _mm_packus_epi32(reslo, reshi);

	// 8 bit conversion and saturation to uint8
	const __m128i res_8b = _mm_packus_epi16(res_16bit, res_16bit);
	__m128i const p_0 = (__m128i )&dst[i * dst_stride + j];
	_mm_storel_epi64(p_0, res_8b);
	}
	}
	}
	}