wiener_convolve_avx2.c (9956B)
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/convolve_avx2.h" 21 #include "aom_dsp/x86/synonyms.h" 22 #include "aom_dsp/x86/synonyms_avx2.h" 23 24 // 128-bit xmmwords are written as [ ... ] with the MSB on the left. 25 // 256-bit ymmwords are written as two xmmwords, [ ... ][ ... ] with the MSB 26 // on the left. 27 // A row of, say, 8-bit pixels with values p0, p1, p2, ..., p30, p31 will be 28 // loaded and stored as [ p31 ... p17 p16 ][ p15 ... p1 p0 ]. 29 30 // Exploiting the range of wiener filter coefficients, 31 // horizontal filtering can be done in 16 bit intermediate precision. 32 // The details are as follows : 33 // Consider the horizontal wiener filter coefficients of the following form : 34 // [C0, C1, C2, 2^(FILTER_BITS) -2 * (C0 + C1 + C2), C2, C1, C0] 35 // Subtracting 2^(FILTER_BITS) from the centre tap we get the following : 36 // [C0, C1, C2, -2 * (C0 + C1 + C2), C2, C1, C0] 37 // The sum of the product "C0 * p0 + C1 * p1 + C2 * p2 -2 * (C0 + C1 + C2) * p3 38 // + C2 * p4 + C1 * p5 + C0 * p6" would be in the range of signed 16 bit 39 // precision. Finally, after rounding the above result by round_0, we multiply 40 // the centre pixel by 2^(FILTER_BITS - round_0) and add it to get the 41 // horizontal filter output. 42 43 void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride, 44 uint8_t *dst, ptrdiff_t dst_stride, 45 const int16_t *filter_x, int x_step_q4, 46 const int16_t *filter_y, int y_step_q4, 47 int w, int h, 48 const WienerConvolveParams *conv_params) { 49 const int bd = 8; 50 assert(x_step_q4 == 16 && y_step_q4 == 16); 51 assert(!(w & 7)); 52 (void)x_step_q4; 53 (void)y_step_q4; 54 55 DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + SUBPEL_TAPS) * 8]); 56 int im_h = h + SUBPEL_TAPS - 2; 57 int im_stride = 8; 58 memset(im_block + (im_h * im_stride), 0, MAX_SB_SIZE); 59 int i, j; 60 const int center_tap = (SUBPEL_TAPS - 1) / 2; 61 const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap; 62 63 __m256i filt[4], coeffs_h[4], coeffs_v[4], filt_center; 64 65 assert(conv_params->round_0 > 0); 66 67 filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2); 68 filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2); 69 filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2); 70 filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2); 71 72 filt_center = _mm256_load_si256((__m256i const *)filt_center_global_avx2); 73 74 const __m128i coeffs_x = _mm_loadu_si128((__m128i *)filter_x); 75 const __m256i filter_coeffs_x = _mm256_broadcastsi128_si256(coeffs_x); 76 77 // coeffs 0 1 0 1 0 1 0 1 78 coeffs_h[0] = 79 _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0200u)); 80 // coeffs 2 3 2 3 2 3 2 3 81 coeffs_h[1] = 82 _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0604u)); 83 // coeffs 4 5 4 5 4 5 4 5 84 coeffs_h[2] = 85 _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0a08u)); 86 // coeffs 6 7 6 7 6 7 6 7 87 coeffs_h[3] = 88 _mm256_shuffle_epi8(filter_coeffs_x, _mm256_set1_epi16(0x0e0cu)); 89 90 const __m256i round_const_h = 91 _mm256_set1_epi16((1 << (conv_params->round_0 - 1))); 92 const __m256i round_const_horz = 93 _mm256_set1_epi16((1 << (bd + FILTER_BITS - conv_params->round_0 - 1))); 94 const __m256i clamp_low = _mm256_setzero_si256(); 95 const __m256i clamp_high = 96 _mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1); 97 const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0); 98 99 // Add an offset to account for the "add_src" part of the convolve function. 100 const __m128i zero_128 = _mm_setzero_si128(); 101 const __m128i offset_0 = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3); 102 const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset_0); 103 104 const __m256i filter_coeffs_y = _mm256_broadcastsi128_si256(coeffs_y); 105 106 // coeffs 0 1 0 1 0 1 0 1 107 coeffs_v[0] = _mm256_shuffle_epi32(filter_coeffs_y, 0x00); 108 // coeffs 2 3 2 3 2 3 2 3 109 coeffs_v[1] = _mm256_shuffle_epi32(filter_coeffs_y, 0x55); 110 // coeffs 4 5 4 5 4 5 4 5 111 coeffs_v[2] = _mm256_shuffle_epi32(filter_coeffs_y, 0xaa); 112 // coeffs 6 7 6 7 6 7 6 7 113 coeffs_v[3] = _mm256_shuffle_epi32(filter_coeffs_y, 0xff); 114 115 const __m256i round_const_v = 116 _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) - 117 (1 << (bd + conv_params->round_1 - 1))); 118 const __m128i round_shift_v = _mm_cvtsi32_si128(conv_params->round_1); 119 120 for (j = 0; j < w; j += 8) { 121 for (i = 0; i < im_h; i += 2) { 122 __m256i data = _mm256_castsi128_si256( 123 _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j])); 124 125 // Load the next line 126 if (i + 1 < im_h) 127 data = _mm256_inserti128_si256( 128 data, 129 _mm_loadu_si128( 130 (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]), 131 1); 132 133 __m256i res = convolve_lowbd_x(data, coeffs_h, filt); 134 135 res = 136 _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); 137 138 __m256i data_0 = _mm256_shuffle_epi8(data, filt_center); 139 140 // multiply the center pixel by 2^(FILTER_BITS - round_0) and add it to 141 // the result 142 data_0 = _mm256_slli_epi16(data_0, FILTER_BITS - conv_params->round_0); 143 res = _mm256_add_epi16(res, data_0); 144 res = _mm256_add_epi16(res, round_const_horz); 145 const __m256i res_clamped = 146 _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high); 147 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res_clamped); 148 } 149 150 /* Vertical filter */ 151 { 152 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride)); 153 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride)); 154 __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride)); 155 __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride)); 156 __m256i src_4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride)); 157 __m256i src_5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride)); 158 159 __m256i s[8]; 160 s[0] = _mm256_unpacklo_epi16(src_0, src_1); 161 s[1] = _mm256_unpacklo_epi16(src_2, src_3); 162 s[2] = _mm256_unpacklo_epi16(src_4, src_5); 163 164 s[4] = _mm256_unpackhi_epi16(src_0, src_1); 165 s[5] = _mm256_unpackhi_epi16(src_2, src_3); 166 s[6] = _mm256_unpackhi_epi16(src_4, src_5); 167 168 for (i = 0; i < h - 1; i += 2) { 169 const int16_t *data = &im_block[i * im_stride]; 170 171 const __m256i s6 = 172 _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); 173 const __m256i s7 = 174 _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); 175 176 s[3] = _mm256_unpacklo_epi16(s6, s7); 177 s[7] = _mm256_unpackhi_epi16(s6, s7); 178 179 __m256i res_a = convolve(s, coeffs_v); 180 __m256i res_b = convolve(s + 4, coeffs_v); 181 182 const __m256i res_a_round = _mm256_sra_epi32( 183 _mm256_add_epi32(res_a, round_const_v), round_shift_v); 184 const __m256i res_b_round = _mm256_sra_epi32( 185 _mm256_add_epi32(res_b, round_const_v), round_shift_v); 186 187 /* rounding code */ 188 // 16 bit conversion 189 const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round); 190 // 8 bit conversion and saturation to uint8 191 const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit); 192 193 const __m128i res_0 = _mm256_castsi256_si128(res_8b); 194 const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1); 195 196 // Store values into the destination buffer 197 __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j]; 198 __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride]; 199 200 _mm_storel_epi64(p_0, res_0); 201 _mm_storel_epi64(p_1, res_1); 202 203 s[0] = s[1]; 204 s[1] = s[2]; 205 s[2] = s[3]; 206 207 s[4] = s[5]; 208 s[5] = s[6]; 209 s[6] = s[7]; 210 } 211 if (h - i) { 212 s[0] = _mm256_permute2x128_si256(s[0], s[4], 0x20); 213 s[1] = _mm256_permute2x128_si256(s[1], s[5], 0x20); 214 s[2] = _mm256_permute2x128_si256(s[2], s[6], 0x20); 215 216 const int16_t *data = &im_block[i * im_stride]; 217 const __m128i s6_ = _mm_loadu_si128((__m128i *)(data + 6 * im_stride)); 218 const __m128i s7_ = _mm_loadu_si128((__m128i *)(data + 7 * im_stride)); 219 220 __m128i s3 = _mm_unpacklo_epi16(s6_, s7_); 221 __m128i s7 = _mm_unpackhi_epi16(s6_, s7_); 222 223 s[3] = _mm256_inserti128_si256(_mm256_castsi128_si256(s3), s7, 1); 224 __m256i convolveres = convolve(s, coeffs_v); 225 226 const __m256i res_round = _mm256_sra_epi32( 227 _mm256_add_epi32(convolveres, round_const_v), round_shift_v); 228 229 /* rounding code */ 230 // 16 bit conversion 231 __m128i reslo = _mm256_castsi256_si128(res_round); 232 __m128i reshi = _mm256_extracti128_si256(res_round, 1); 233 const __m128i res_16bit = _mm_packus_epi32(reslo, reshi); 234 235 // 8 bit conversion and saturation to uint8 236 const __m128i res_8b = _mm_packus_epi16(res_16bit, res_16bit); 237 __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j]; 238 _mm_storel_epi64(p_0, res_8b); 239 } 240 } 241 } 242 }