av1_temporal_denoiser_sse2.c (14026B)
1 /* 2 * Copyright (c) 2017, 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 <assert.h> 13 #include <emmintrin.h> // SSE2 14 15 #include "aom/aom_integer.h" 16 #include "aom_dsp/x86/mem_sse2.h" 17 18 #include "av1/common/reconinter.h" 19 #include "av1/encoder/context_tree.h" 20 #include "av1/encoder/av1_temporal_denoiser.h" 21 22 // Compute the sum of all pixel differences of this MB. 23 static inline int sum_diff_16x1(__m128i acc_diff) { 24 const __m128i k_1 = _mm_set1_epi16(1); 25 const __m128i acc_diff_lo = 26 _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8); 27 const __m128i acc_diff_hi = 28 _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8); 29 const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi); 30 const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1); 31 const __m128i hgfe_dcba = 32 _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8)); 33 const __m128i hgfedcba = 34 _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4)); 35 return _mm_cvtsi128_si32(hgfedcba); 36 } 37 38 // Denoise a 16x1 vector. 39 static inline __m128i av1_denoiser_16x1_sse2( 40 const uint8_t *sig, const uint8_t *mc_running_avg_y, uint8_t *running_avg_y, 41 const __m128i *k_0, const __m128i *k_4, const __m128i *k_8, 42 const __m128i *k_16, const __m128i *l3, const __m128i *l32, 43 const __m128i *l21, __m128i acc_diff) { 44 // Calculate differences 45 const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0])); 46 const __m128i v_mc_running_avg_y = 47 _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0])); 48 __m128i v_running_avg_y; 49 const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig); 50 const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y); 51 // Obtain the sign. FF if diff is negative. 52 const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, *k_0); 53 // Clamp absolute difference to 16 to be used to get mask. Doing this 54 // allows us to use _mm_cmpgt_epi8, which operates on signed byte. 55 const __m128i clamped_absdiff = 56 _mm_min_epu8(_mm_or_si128(pdiff, ndiff), *k_16); 57 // Get masks for l2 l1 and l0 adjustments. 58 const __m128i mask2 = _mm_cmpgt_epi8(*k_16, clamped_absdiff); 59 const __m128i mask1 = _mm_cmpgt_epi8(*k_8, clamped_absdiff); 60 const __m128i mask0 = _mm_cmpgt_epi8(*k_4, clamped_absdiff); 61 // Get adjustments for l2, l1, and l0. 62 __m128i adj2 = _mm_and_si128(mask2, *l32); 63 const __m128i adj1 = _mm_and_si128(mask1, *l21); 64 const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff); 65 __m128i adj, padj, nadj; 66 67 // Combine the adjustments and get absolute adjustments. 68 adj2 = _mm_add_epi8(adj2, adj1); 69 adj = _mm_sub_epi8(*l3, adj2); 70 adj = _mm_andnot_si128(mask0, adj); 71 adj = _mm_or_si128(adj, adj0); 72 73 // Restore the sign and get positive and negative adjustments. 74 padj = _mm_andnot_si128(diff_sign, adj); 75 nadj = _mm_and_si128(diff_sign, adj); 76 77 // Calculate filtered value. 78 v_running_avg_y = _mm_adds_epu8(v_sig, padj); 79 v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj); 80 _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y); 81 82 // Adjustments <=7, and each element in acc_diff can fit in signed 83 // char. 84 acc_diff = _mm_adds_epi8(acc_diff, padj); 85 acc_diff = _mm_subs_epi8(acc_diff, nadj); 86 return acc_diff; 87 } 88 89 // Denoise a 16x1 vector with a weaker filter. 90 static inline __m128i av1_denoiser_adj_16x1_sse2( 91 const uint8_t *sig, const uint8_t *mc_running_avg_y, uint8_t *running_avg_y, 92 const __m128i k_0, const __m128i k_delta, __m128i acc_diff) { 93 __m128i v_running_avg_y = _mm_loadu_si128((__m128i *)(&running_avg_y[0])); 94 // Calculate differences. 95 const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0])); 96 const __m128i v_mc_running_avg_y = 97 _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0])); 98 const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig); 99 const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y); 100 // Obtain the sign. FF if diff is negative. 101 const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0); 102 // Clamp absolute difference to delta to get the adjustment. 103 const __m128i adj = _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta); 104 // Restore the sign and get positive and negative adjustments. 105 __m128i padj, nadj; 106 padj = _mm_andnot_si128(diff_sign, adj); 107 nadj = _mm_and_si128(diff_sign, adj); 108 // Calculate filtered value. 109 v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj); 110 v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj); 111 _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y); 112 113 // Accumulate the adjustments. 114 acc_diff = _mm_subs_epi8(acc_diff, padj); 115 acc_diff = _mm_adds_epi8(acc_diff, nadj); 116 return acc_diff; 117 } 118 119 // Denoise 8x8 and 8x16 blocks. 120 static int av1_denoiser_NxM_sse2_small(const uint8_t *sig, int sig_stride, 121 const uint8_t *mc_running_avg_y, 122 int mc_avg_y_stride, 123 uint8_t *running_avg_y, int avg_y_stride, 124 int increase_denoising, BLOCK_SIZE bs, 125 int motion_magnitude, int width) { 126 int sum_diff_thresh, r, sum_diff = 0; 127 const int shift_inc = 128 (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) 129 ? 1 130 : 0; 131 uint8_t sig_buffer[8][16], mc_running_buffer[8][16], running_buffer[8][16]; 132 __m128i acc_diff = _mm_setzero_si128(); 133 const __m128i k_0 = _mm_setzero_si128(); 134 const __m128i k_4 = _mm_set1_epi8(4 + shift_inc); 135 const __m128i k_8 = _mm_set1_epi8(8); 136 const __m128i k_16 = _mm_set1_epi8(16); 137 // Modify each level's adjustment according to motion_magnitude. 138 const __m128i l3 = _mm_set1_epi8( 139 (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6); 140 // Difference between level 3 and level 2 is 2. 141 const __m128i l32 = _mm_set1_epi8(2); 142 // Difference between level 2 and level 1 is 1. 143 const __m128i l21 = _mm_set1_epi8(1); 144 const int b_height = block_size_high[bs] >> 1; 145 146 for (r = 0; r < b_height; ++r) { 147 memcpy(sig_buffer[r], sig, width); 148 memcpy(sig_buffer[r] + width, sig + sig_stride, width); 149 memcpy(mc_running_buffer[r], mc_running_avg_y, width); 150 memcpy(mc_running_buffer[r] + width, mc_running_avg_y + mc_avg_y_stride, 151 width); 152 memcpy(running_buffer[r], running_avg_y, width); 153 memcpy(running_buffer[r] + width, running_avg_y + avg_y_stride, width); 154 acc_diff = av1_denoiser_16x1_sse2(sig_buffer[r], mc_running_buffer[r], 155 running_buffer[r], &k_0, &k_4, &k_8, 156 &k_16, &l3, &l32, &l21, acc_diff); 157 memcpy(running_avg_y, running_buffer[r], width); 158 memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, width); 159 // Update pointers for next iteration. 160 sig += (sig_stride << 1); 161 mc_running_avg_y += (mc_avg_y_stride << 1); 162 running_avg_y += (avg_y_stride << 1); 163 } 164 165 { 166 sum_diff = sum_diff_16x1(acc_diff); 167 sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising); 168 if (abs(sum_diff) > sum_diff_thresh) { 169 // Before returning to copy the block (i.e., apply no denoising), 170 // check if we can still apply some (weaker) temporal filtering to 171 // this block, that would otherwise not be denoised at all. Simplest 172 // is to apply an additional adjustment to running_avg_y to bring it 173 // closer to sig. The adjustment is capped by a maximum delta, and 174 // chosen such that in most cases the resulting sum_diff will be 175 // within the acceptable range given by sum_diff_thresh. 176 177 // The delta is set by the excess of absolute pixel diff over the 178 // threshold. 179 const int delta = 180 ((abs(sum_diff) - sum_diff_thresh) >> num_pels_log2_lookup[bs]) + 1; 181 // Only apply the adjustment for max delta up to 3. 182 if (delta < 4) { 183 const __m128i k_delta = _mm_set1_epi8(delta); 184 running_avg_y -= avg_y_stride * (b_height << 1); 185 for (r = 0; r < b_height; ++r) { 186 acc_diff = av1_denoiser_adj_16x1_sse2( 187 sig_buffer[r], mc_running_buffer[r], running_buffer[r], k_0, 188 k_delta, acc_diff); 189 memcpy(running_avg_y, running_buffer[r], width); 190 memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, 191 width); 192 // Update pointers for next iteration. 193 running_avg_y += (avg_y_stride << 1); 194 } 195 sum_diff = sum_diff_16x1(acc_diff); 196 if (abs(sum_diff) > sum_diff_thresh) { 197 return COPY_BLOCK; 198 } 199 } else { 200 return COPY_BLOCK; 201 } 202 } 203 } 204 return FILTER_BLOCK; 205 } 206 207 // Denoise 16x16 to 128x128 blocks. 208 static int av1_denoiser_NxM_sse2_big(const uint8_t *sig, int sig_stride, 209 const uint8_t *mc_running_avg_y, 210 int mc_avg_y_stride, 211 uint8_t *running_avg_y, int avg_y_stride, 212 int increase_denoising, BLOCK_SIZE bs, 213 int motion_magnitude) { 214 int sum_diff_thresh, r, c, sum_diff = 0; 215 const int shift_inc = 216 (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) 217 ? 1 218 : 0; 219 __m128i acc_diff[8][8]; 220 const __m128i k_0 = _mm_setzero_si128(); 221 const __m128i k_4 = _mm_set1_epi8(4 + shift_inc); 222 const __m128i k_8 = _mm_set1_epi8(8); 223 const __m128i k_16 = _mm_set1_epi8(16); 224 // Modify each level's adjustment according to motion_magnitude. 225 const __m128i l3 = _mm_set1_epi8( 226 (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6); 227 // Difference between level 3 and level 2 is 2. 228 const __m128i l32 = _mm_set1_epi8(2); 229 // Difference between level 2 and level 1 is 1. 230 const __m128i l21 = _mm_set1_epi8(1); 231 const int b_width = block_size_wide[bs]; 232 const int b_height = block_size_high[bs]; 233 const int b_width_shift4 = b_width >> 4; 234 235 for (r = 0; r < 8; ++r) { 236 for (c = 0; c < b_width_shift4; ++c) { 237 acc_diff[c][r] = _mm_setzero_si128(); 238 } 239 } 240 241 for (r = 0; r < b_height; ++r) { 242 for (c = 0; c < b_width_shift4; ++c) { 243 acc_diff[c][r >> 4] = av1_denoiser_16x1_sse2( 244 sig, mc_running_avg_y, running_avg_y, &k_0, &k_4, &k_8, &k_16, &l3, 245 &l32, &l21, acc_diff[c][r >> 4]); 246 // Update pointers for next iteration. 247 sig += 16; 248 mc_running_avg_y += 16; 249 running_avg_y += 16; 250 } 251 252 if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) { 253 for (c = 0; c < b_width_shift4; ++c) { 254 sum_diff += sum_diff_16x1(acc_diff[c][r >> 4]); 255 } 256 } 257 258 // Update pointers for next iteration. 259 sig = sig - b_width + sig_stride; 260 mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride; 261 running_avg_y = running_avg_y - b_width + avg_y_stride; 262 } 263 264 { 265 sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising); 266 if (abs(sum_diff) > sum_diff_thresh) { 267 const int delta = 268 ((abs(sum_diff) - sum_diff_thresh) >> num_pels_log2_lookup[bs]) + 1; 269 270 // Only apply the adjustment for max delta up to 3. 271 if (delta < 4) { 272 const __m128i k_delta = _mm_set1_epi8(delta); 273 sig -= sig_stride * b_height; 274 mc_running_avg_y -= mc_avg_y_stride * b_height; 275 running_avg_y -= avg_y_stride * b_height; 276 sum_diff = 0; 277 for (r = 0; r < b_height; ++r) { 278 for (c = 0; c < b_width_shift4; ++c) { 279 acc_diff[c][r >> 4] = 280 av1_denoiser_adj_16x1_sse2(sig, mc_running_avg_y, running_avg_y, 281 k_0, k_delta, acc_diff[c][r >> 4]); 282 // Update pointers for next iteration. 283 sig += 16; 284 mc_running_avg_y += 16; 285 running_avg_y += 16; 286 } 287 288 if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) { 289 for (c = 0; c < b_width_shift4; ++c) { 290 sum_diff += sum_diff_16x1(acc_diff[c][r >> 4]); 291 } 292 } 293 sig = sig - b_width + sig_stride; 294 mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride; 295 running_avg_y = running_avg_y - b_width + avg_y_stride; 296 } 297 if (abs(sum_diff) > sum_diff_thresh) { 298 return COPY_BLOCK; 299 } 300 } else { 301 return COPY_BLOCK; 302 } 303 } 304 } 305 return FILTER_BLOCK; 306 } 307 308 int av1_denoiser_filter_sse2(const uint8_t *sig, int sig_stride, 309 const uint8_t *mc_avg, int mc_avg_stride, 310 uint8_t *avg, int avg_stride, 311 int increase_denoising, BLOCK_SIZE bs, 312 int motion_magnitude) { 313 // Rank by frequency of the block type to have an early termination. 314 if (bs == BLOCK_16X16 || bs == BLOCK_32X32 || bs == BLOCK_64X64 || 315 bs == BLOCK_128X128 || bs == BLOCK_128X64 || bs == BLOCK_64X128 || 316 bs == BLOCK_16X32 || bs == BLOCK_16X8 || bs == BLOCK_32X16 || 317 bs == BLOCK_32X64 || bs == BLOCK_64X32) { 318 return av1_denoiser_NxM_sse2_big(sig, sig_stride, mc_avg, mc_avg_stride, 319 avg, avg_stride, increase_denoising, bs, 320 motion_magnitude); 321 } else if (bs == BLOCK_8X8 || bs == BLOCK_8X16) { 322 return av1_denoiser_NxM_sse2_small(sig, sig_stride, mc_avg, mc_avg_stride, 323 avg, avg_stride, increase_denoising, bs, 324 motion_magnitude, 8); 325 } else { 326 return COPY_BLOCK; 327 } 328 }