highbd_wiener_convolve_avx2.c (11602B)
1 /* 2 * Copyright (c) 2018, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include <immintrin.h> 13 #include <assert.h> 14 15 #include "config/av1_rtcd.h" 16 17 #include "av1/common/convolve.h" 18 #include "aom_dsp/aom_dsp_common.h" 19 #include "aom_dsp/aom_filter.h" 20 #include "aom_dsp/x86/synonyms.h" 21 #include "aom_dsp/x86/synonyms_avx2.h" 22 23 // 128-bit xmmwords are written as [ ... ] with the MSB on the left. 24 // 256-bit ymmwords are written as two xmmwords, [ ... ][ ... ] with the MSB 25 // on the left. 26 // A row of, say, 16-bit pixels with values p0, p1, p2, ..., p14, p15 will be 27 // loaded and stored as [ p15 ... p9 p8 ][ p7 ... p1 p0 ]. 28 void av1_highbd_wiener_convolve_add_src_avx2( 29 const uint8_t *src8, ptrdiff_t src_stride, uint8_t *dst8, 30 ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, 31 const int16_t *filter_y, int y_step_q4, int w, int h, 32 const WienerConvolveParams *conv_params, int bd) { 33 assert(x_step_q4 == 16 && y_step_q4 == 16); 34 assert(!(w & 7)); 35 assert(bd + FILTER_BITS - conv_params->round_0 + 2 <= 16); 36 (void)x_step_q4; 37 (void)y_step_q4; 38 39 const uint16_t *const src = CONVERT_TO_SHORTPTR(src8); 40 uint16_t *const dst = CONVERT_TO_SHORTPTR(dst8); 41 42 DECLARE_ALIGNED(32, uint16_t, 43 temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]); 44 int intermediate_height = h + SUBPEL_TAPS - 1; 45 const int center_tap = ((SUBPEL_TAPS - 1) / 2); 46 const uint16_t *const src_ptr = src - center_tap * src_stride - center_tap; 47 48 const __m128i zero_128 = _mm_setzero_si128(); 49 const __m256i zero_256 = _mm256_setzero_si256(); 50 51 // Add an offset to account for the "add_src" part of the convolve function. 52 const __m128i offset = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3); 53 54 const __m256i clamp_low = zero_256; 55 56 /* Horizontal filter */ 57 { 58 const __m256i clamp_high_ep = 59 _mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1); 60 61 // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ] 62 const __m128i coeffs_x = _mm_add_epi16(xx_loadu_128(filter_x), offset); 63 64 // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ] 65 const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_x, coeffs_x); 66 // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ] 67 const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_x, coeffs_x); 68 69 // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ] 70 const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123); 71 // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ] 72 const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123); 73 // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ] 74 const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567); 75 // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ] 76 const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567); 77 78 // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ] 79 const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128); 80 // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ] 81 const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128); 82 // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ] 83 const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128); 84 // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ] 85 const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128); 86 87 const __m256i round_const = _mm256_set1_epi32( 88 (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1))); 89 90 for (int i = 0; i < intermediate_height; ++i) { 91 for (int j = 0; j < w; j += 16) { 92 const uint16_t *src_ij = src_ptr + i * src_stride + j; 93 94 // Load 16-bit src data 95 const __m256i src_0 = yy_loadu_256(src_ij + 0); 96 const __m256i src_1 = yy_loadu_256(src_ij + 1); 97 const __m256i src_2 = yy_loadu_256(src_ij + 2); 98 const __m256i src_3 = yy_loadu_256(src_ij + 3); 99 const __m256i src_4 = yy_loadu_256(src_ij + 4); 100 const __m256i src_5 = yy_loadu_256(src_ij + 5); 101 const __m256i src_6 = yy_loadu_256(src_ij + 6); 102 const __m256i src_7 = yy_loadu_256(src_ij + 7); 103 104 // Multiply src data by filter coeffs and sum pairs 105 const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01); 106 const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01); 107 const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23); 108 const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23); 109 const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45); 110 const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45); 111 const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67); 112 const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67); 113 114 // Calculate scalar product for even- and odd-indices separately, 115 // increasing to 32-bit precision 116 const __m256i res_even_sum = _mm256_add_epi32( 117 _mm256_add_epi32(res_0, res_4), _mm256_add_epi32(res_2, res_6)); 118 const __m256i res_even = _mm256_srai_epi32( 119 _mm256_add_epi32(res_even_sum, round_const), conv_params->round_0); 120 121 const __m256i res_odd_sum = _mm256_add_epi32( 122 _mm256_add_epi32(res_1, res_5), _mm256_add_epi32(res_3, res_7)); 123 const __m256i res_odd = _mm256_srai_epi32( 124 _mm256_add_epi32(res_odd_sum, round_const), conv_params->round_0); 125 126 // Reduce to 16-bit precision and pack even- and odd-index results 127 // back into one register. The _mm256_packs_epi32 intrinsic returns 128 // a register with the pixels ordered as follows: 129 // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ] 130 const __m256i res = _mm256_packs_epi32(res_even, res_odd); 131 const __m256i res_clamped = 132 _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high_ep); 133 134 // Store in a temporary array 135 yy_storeu_256(temp + i * MAX_SB_SIZE + j, res_clamped); 136 } 137 } 138 } 139 140 /* Vertical filter */ 141 { 142 const __m256i clamp_high = _mm256_set1_epi16((1 << bd) - 1); 143 144 // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ] 145 const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset); 146 147 // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ] 148 const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_y, coeffs_y); 149 // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ] 150 const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_y, coeffs_y); 151 152 // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ] 153 const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123); 154 // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ] 155 const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123); 156 // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ] 157 const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567); 158 // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ] 159 const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567); 160 161 // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ] 162 const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128); 163 // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ] 164 const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128); 165 // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ] 166 const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128); 167 // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ] 168 const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128); 169 170 const __m256i round_const = 171 _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) - 172 (1 << (bd + conv_params->round_1 - 1))); 173 174 for (int i = 0; i < h; ++i) { 175 for (int j = 0; j < w; j += 16) { 176 const uint16_t *temp_ij = temp + i * MAX_SB_SIZE + j; 177 178 // Load 16-bit data from the output of the horizontal filter in 179 // which the pixels are ordered as follows: 180 // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ] 181 const __m256i data_0 = yy_loadu_256(temp_ij + 0 * MAX_SB_SIZE); 182 const __m256i data_1 = yy_loadu_256(temp_ij + 1 * MAX_SB_SIZE); 183 const __m256i data_2 = yy_loadu_256(temp_ij + 2 * MAX_SB_SIZE); 184 const __m256i data_3 = yy_loadu_256(temp_ij + 3 * MAX_SB_SIZE); 185 const __m256i data_4 = yy_loadu_256(temp_ij + 4 * MAX_SB_SIZE); 186 const __m256i data_5 = yy_loadu_256(temp_ij + 5 * MAX_SB_SIZE); 187 const __m256i data_6 = yy_loadu_256(temp_ij + 6 * MAX_SB_SIZE); 188 const __m256i data_7 = yy_loadu_256(temp_ij + 7 * MAX_SB_SIZE); 189 190 // Filter the even-indices, increasing to 32-bit precision 191 const __m256i src_0 = _mm256_unpacklo_epi16(data_0, data_1); 192 const __m256i src_2 = _mm256_unpacklo_epi16(data_2, data_3); 193 const __m256i src_4 = _mm256_unpacklo_epi16(data_4, data_5); 194 const __m256i src_6 = _mm256_unpacklo_epi16(data_6, data_7); 195 196 const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01); 197 const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23); 198 const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45); 199 const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67); 200 201 const __m256i res_even = _mm256_add_epi32( 202 _mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6)); 203 204 // Filter the odd-indices, increasing to 32-bit precision 205 const __m256i src_1 = _mm256_unpackhi_epi16(data_0, data_1); 206 const __m256i src_3 = _mm256_unpackhi_epi16(data_2, data_3); 207 const __m256i src_5 = _mm256_unpackhi_epi16(data_4, data_5); 208 const __m256i src_7 = _mm256_unpackhi_epi16(data_6, data_7); 209 210 const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01); 211 const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23); 212 const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45); 213 const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67); 214 215 const __m256i res_odd = _mm256_add_epi32( 216 _mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7)); 217 218 // Pixels are currently in the following order: 219 // res_even order: [ 14 12 10 8 ] [ 6 4 2 0 ] 220 // res_odd order: [ 15 13 11 9 ] [ 7 5 3 1 ] 221 // 222 // Rearrange the pixels into the following order: 223 // res_lo order: [ 11 10 9 8 ] [ 3 2 1 0 ] 224 // res_hi order: [ 15 14 13 12 ] [ 7 6 5 4 ] 225 const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd); 226 const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd); 227 228 const __m256i res_lo_round = _mm256_srai_epi32( 229 _mm256_add_epi32(res_lo, round_const), conv_params->round_1); 230 const __m256i res_hi_round = _mm256_srai_epi32( 231 _mm256_add_epi32(res_hi, round_const), conv_params->round_1); 232 233 // Reduce to 16-bit precision and pack into the correct order: 234 // [ 15 14 13 12 11 10 9 8 ][ 7 6 5 4 3 2 1 0 ] 235 const __m256i res_16bit = 236 _mm256_packs_epi32(res_lo_round, res_hi_round); 237 const __m256i res_16bit_clamped = _mm256_min_epi16( 238 _mm256_max_epi16(res_16bit, clamp_low), clamp_high); 239 240 // Store in the dst array 241 yy_storeu_256(dst + i * dst_stride + j, res_16bit_clamped); 242 } 243 } 244 } 245 }