1/* 2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include <emmintrin.h> 12 13#include "./vpx_config.h" 14#include "./vp9_rtcd.h" 15 16#include "vpx_ports/emmintrin_compat.h" 17#include "vpx/vpx_integer.h" 18#include "vp9/common/vp9_reconinter.h" 19#include "vp9/encoder/vp9_context_tree.h" 20#include "vp9/encoder/vp9_denoiser.h" 21#include "vpx_mem/vpx_mem.h" 22 23// Compute the sum of all pixel differences of this MB. 24static INLINE int sum_diff_16x1(__m128i acc_diff) { 25 const __m128i k_1 = _mm_set1_epi16(1); 26 const __m128i acc_diff_lo = 27 _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8); 28 const __m128i acc_diff_hi = 29 _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8); 30 const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi); 31 const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1); 32 const __m128i hgfe_dcba = 33 _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8)); 34 const __m128i hgfedcba = 35 _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4)); 36 return _mm_cvtsi128_si32(hgfedcba); 37} 38 39// Denoise a 16x1 vector. 40static INLINE __m128i vp9_denoiser_16x1_sse2( 41 const uint8_t *sig, const uint8_t *mc_running_avg_y, uint8_t *running_avg_y, 42 const __m128i *k_0, const __m128i *k_4, const __m128i *k_8, 43 const __m128i *k_16, const __m128i *l3, const __m128i *l32, 44 const __m128i *l21, __m128i acc_diff) { 45 // Calculate differences 46 const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0])); 47 const __m128i v_mc_running_avg_y = 48 _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0])); 49 __m128i v_running_avg_y; 50 const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig); 51 const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y); 52 // Obtain the sign. FF if diff is negative. 53 const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, *k_0); 54 // Clamp absolute difference to 16 to be used to get mask. Doing this 55 // allows us to use _mm_cmpgt_epi8, which operates on signed byte. 56 const __m128i clamped_absdiff = 57 _mm_min_epu8(_mm_or_si128(pdiff, ndiff), *k_16); 58 // Get masks for l2 l1 and l0 adjustments. 59 const __m128i mask2 = _mm_cmpgt_epi8(*k_16, clamped_absdiff); 60 const __m128i mask1 = _mm_cmpgt_epi8(*k_8, clamped_absdiff); 61 const __m128i mask0 = _mm_cmpgt_epi8(*k_4, clamped_absdiff); 62 // Get adjustments for l2, l1, and l0. 63 __m128i adj2 = _mm_and_si128(mask2, *l32); 64 const __m128i adj1 = _mm_and_si128(mask1, *l21); 65 const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff); 66 __m128i adj, padj, nadj; 67 68 // Combine the adjustments and get absolute adjustments. 69 adj2 = _mm_add_epi8(adj2, adj1); 70 adj = _mm_sub_epi8(*l3, adj2); 71 adj = _mm_andnot_si128(mask0, adj); 72 adj = _mm_or_si128(adj, adj0); 73 74 // Restore the sign and get positive and negative adjustments. 75 padj = _mm_andnot_si128(diff_sign, adj); 76 nadj = _mm_and_si128(diff_sign, adj); 77 78 // Calculate filtered value. 79 v_running_avg_y = _mm_adds_epu8(v_sig, padj); 80 v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj); 81 _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y); 82 83 // Adjustments <=7, and each element in acc_diff can fit in signed 84 // char. 85 acc_diff = _mm_adds_epi8(acc_diff, padj); 86 acc_diff = _mm_subs_epi8(acc_diff, nadj); 87 return acc_diff; 88} 89 90// Denoise a 16x1 vector with a weaker filter. 91static INLINE __m128i vp9_denoiser_adj_16x1_sse2( 92 const uint8_t *sig, const uint8_t *mc_running_avg_y, uint8_t *running_avg_y, 93 const __m128i k_0, const __m128i k_delta, __m128i acc_diff) { 94 __m128i v_running_avg_y = _mm_loadu_si128((__m128i *)(&running_avg_y[0])); 95 // Calculate differences. 96 const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0])); 97 const __m128i v_mc_running_avg_y = 98 _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0])); 99 const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig); 100 const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y); 101 // Obtain the sign. FF if diff is negative. 102 const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0); 103 // Clamp absolute difference to delta to get the adjustment. 104 const __m128i adj = _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta); 105 // Restore the sign and get positive and negative adjustments. 106 __m128i padj, nadj; 107 padj = _mm_andnot_si128(diff_sign, adj); 108 nadj = _mm_and_si128(diff_sign, adj); 109 // Calculate filtered value. 110 v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj); 111 v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj); 112 _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y); 113 114 // Accumulate the adjustments. 115 acc_diff = _mm_subs_epi8(acc_diff, padj); 116 acc_diff = _mm_adds_epi8(acc_diff, nadj); 117 return acc_diff; 118} 119 120// Denoise 8x8 and 8x16 blocks. 121static int vp9_denoiser_NxM_sse2_small(const uint8_t *sig, int sig_stride, 122 const uint8_t *mc_running_avg_y, 123 int mc_avg_y_stride, 124 uint8_t *running_avg_y, int avg_y_stride, 125 int increase_denoising, BLOCK_SIZE bs, 126 int motion_magnitude, int width) { 127 int sum_diff_thresh, r, sum_diff = 0; 128 const int shift_inc = 129 (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) 130 ? 1 131 : 0; 132 uint8_t sig_buffer[8][16], mc_running_buffer[8][16], running_buffer[8][16]; 133 __m128i acc_diff = _mm_setzero_si128(); 134 const __m128i k_0 = _mm_setzero_si128(); 135 const __m128i k_4 = _mm_set1_epi8(4 + shift_inc); 136 const __m128i k_8 = _mm_set1_epi8(8); 137 const __m128i k_16 = _mm_set1_epi8(16); 138 // Modify each level's adjustment according to motion_magnitude. 139 const __m128i l3 = _mm_set1_epi8( 140 (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6); 141 // Difference between level 3 and level 2 is 2. 142 const __m128i l32 = _mm_set1_epi8(2); 143 // Difference between level 2 and level 1 is 1. 144 const __m128i l21 = _mm_set1_epi8(1); 145 const int b_height = (4 << b_height_log2_lookup[bs]) >> 1; 146 147 for (r = 0; r < b_height; ++r) { 148 memcpy(sig_buffer[r], sig, width); 149 memcpy(sig_buffer[r] + width, sig + sig_stride, width); 150 memcpy(mc_running_buffer[r], mc_running_avg_y, width); 151 memcpy(mc_running_buffer[r] + width, mc_running_avg_y + mc_avg_y_stride, 152 width); 153 memcpy(running_buffer[r], running_avg_y, width); 154 memcpy(running_buffer[r] + width, running_avg_y + avg_y_stride, width); 155 acc_diff = vp9_denoiser_16x1_sse2(sig_buffer[r], mc_running_buffer[r], 156 running_buffer[r], &k_0, &k_4, &k_8, 157 &k_16, &l3, &l32, &l21, acc_diff); 158 memcpy(running_avg_y, running_buffer[r], width); 159 memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, width); 160 // Update pointers for next iteration. 161 sig += (sig_stride << 1); 162 mc_running_avg_y += (mc_avg_y_stride << 1); 163 running_avg_y += (avg_y_stride << 1); 164 } 165 166 { 167 sum_diff = sum_diff_16x1(acc_diff); 168 sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising); 169 if (abs(sum_diff) > sum_diff_thresh) { 170 // Before returning to copy the block (i.e., apply no denoising), 171 // check if we can still apply some (weaker) temporal filtering to 172 // this block, that would otherwise not be denoised at all. Simplest 173 // is to apply an additional adjustment to running_avg_y to bring it 174 // closer to sig. The adjustment is capped by a maximum delta, and 175 // chosen such that in most cases the resulting sum_diff will be 176 // within the acceptable range given by sum_diff_thresh. 177 178 // The delta is set by the excess of absolute pixel diff over the 179 // threshold. 180 const int delta = 181 ((abs(sum_diff) - sum_diff_thresh) >> num_pels_log2_lookup[bs]) + 1; 182 // Only apply the adjustment for max delta up to 3. 183 if (delta < 4) { 184 const __m128i k_delta = _mm_set1_epi8(delta); 185 running_avg_y -= avg_y_stride * (b_height << 1); 186 for (r = 0; r < b_height; ++r) { 187 acc_diff = vp9_denoiser_adj_16x1_sse2( 188 sig_buffer[r], mc_running_buffer[r], running_buffer[r], k_0, 189 k_delta, acc_diff); 190 memcpy(running_avg_y, running_buffer[r], width); 191 memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, 192 width); 193 // Update pointers for next iteration. 194 running_avg_y += (avg_y_stride << 1); 195 } 196 sum_diff = sum_diff_16x1(acc_diff); 197 if (abs(sum_diff) > sum_diff_thresh) { 198 return COPY_BLOCK; 199 } 200 } else { 201 return COPY_BLOCK; 202 } 203 } 204 } 205 return FILTER_BLOCK; 206} 207 208// Denoise 16x16, 16x32, 32x16, 32x32, 32x64, 64x32 and 64x64 blocks. 209static int vp9_denoiser_NxM_sse2_big(const uint8_t *sig, int sig_stride, 210 const uint8_t *mc_running_avg_y, 211 int mc_avg_y_stride, 212 uint8_t *running_avg_y, int avg_y_stride, 213 int increase_denoising, BLOCK_SIZE bs, 214 int motion_magnitude) { 215 int sum_diff_thresh, r, c, sum_diff = 0; 216 const int shift_inc = 217 (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) 218 ? 1 219 : 0; 220 __m128i acc_diff[4][4]; 221 const __m128i k_0 = _mm_setzero_si128(); 222 const __m128i k_4 = _mm_set1_epi8(4 + shift_inc); 223 const __m128i k_8 = _mm_set1_epi8(8); 224 const __m128i k_16 = _mm_set1_epi8(16); 225 // Modify each level's adjustment according to motion_magnitude. 226 const __m128i l3 = _mm_set1_epi8( 227 (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6); 228 // Difference between level 3 and level 2 is 2. 229 const __m128i l32 = _mm_set1_epi8(2); 230 // Difference between level 2 and level 1 is 1. 231 const __m128i l21 = _mm_set1_epi8(1); 232 const int b_width = (4 << b_width_log2_lookup[bs]); 233 const int b_height = (4 << b_height_log2_lookup[bs]); 234 const int b_width_shift4 = b_width >> 4; 235 236 for (r = 0; r < 4; ++r) { 237 for (c = 0; c < b_width_shift4; ++c) { 238 acc_diff[c][r] = _mm_setzero_si128(); 239 } 240 } 241 242 for (r = 0; r < b_height; ++r) { 243 for (c = 0; c < b_width_shift4; ++c) { 244 acc_diff[c][r >> 4] = vp9_denoiser_16x1_sse2( 245 sig, mc_running_avg_y, running_avg_y, &k_0, &k_4, &k_8, &k_16, &l3, 246 &l32, &l21, acc_diff[c][r >> 4]); 247 // Update pointers for next iteration. 248 sig += 16; 249 mc_running_avg_y += 16; 250 running_avg_y += 16; 251 } 252 253 if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) { 254 for (c = 0; c < b_width_shift4; ++c) { 255 sum_diff += sum_diff_16x1(acc_diff[c][r >> 4]); 256 } 257 } 258 259 // Update pointers for next iteration. 260 sig = sig - b_width + sig_stride; 261 mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride; 262 running_avg_y = running_avg_y - b_width + avg_y_stride; 263 } 264 265 { 266 sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising); 267 if (abs(sum_diff) > sum_diff_thresh) { 268 const int delta = 269 ((abs(sum_diff) - sum_diff_thresh) >> num_pels_log2_lookup[bs]) + 1; 270 271 // Only apply the adjustment for max delta up to 3. 272 if (delta < 4) { 273 const __m128i k_delta = _mm_set1_epi8(delta); 274 sig -= sig_stride * b_height; 275 mc_running_avg_y -= mc_avg_y_stride * b_height; 276 running_avg_y -= avg_y_stride * b_height; 277 sum_diff = 0; 278 for (r = 0; r < b_height; ++r) { 279 for (c = 0; c < b_width_shift4; ++c) { 280 acc_diff[c][r >> 4] = 281 vp9_denoiser_adj_16x1_sse2(sig, mc_running_avg_y, running_avg_y, 282 k_0, k_delta, acc_diff[c][r >> 4]); 283 // Update pointers for next iteration. 284 sig += 16; 285 mc_running_avg_y += 16; 286 running_avg_y += 16; 287 } 288 289 if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) { 290 for (c = 0; c < b_width_shift4; ++c) { 291 sum_diff += sum_diff_16x1(acc_diff[c][r >> 4]); 292 } 293 } 294 sig = sig - b_width + sig_stride; 295 mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride; 296 running_avg_y = running_avg_y - b_width + avg_y_stride; 297 } 298 if (abs(sum_diff) > sum_diff_thresh) { 299 return COPY_BLOCK; 300 } 301 } else { 302 return COPY_BLOCK; 303 } 304 } 305 } 306 return FILTER_BLOCK; 307} 308 309int vp9_denoiser_filter_sse2(const uint8_t *sig, int sig_stride, 310 const uint8_t *mc_avg, int mc_avg_stride, 311 uint8_t *avg, int avg_stride, 312 int increase_denoising, BLOCK_SIZE bs, 313 int motion_magnitude) { 314 // Rank by frequency of the block type to have an early termination. 315 if (bs == BLOCK_16X16 || bs == BLOCK_32X32 || bs == BLOCK_64X64 || 316 bs == BLOCK_16X32 || bs == BLOCK_16X8 || bs == BLOCK_32X16 || 317 bs == BLOCK_32X64 || bs == BLOCK_64X32) { 318 return vp9_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 vp9_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} 329