1/* 2 * Copyright (c) 2012 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 <assert.h> 12#include <limits.h> 13#include <math.h> 14 15#include "./vpx_dsp_rtcd.h" 16#include "vpx_dsp/vpx_dsp_common.h" 17#include "vpx_scale/yv12config.h" 18#include "vpx/vpx_integer.h" 19#include "vp9/common/vp9_reconinter.h" 20#include "vp9/encoder/vp9_context_tree.h" 21#include "vp9/encoder/vp9_denoiser.h" 22#include "vp9/encoder/vp9_encoder.h" 23 24#ifdef OUTPUT_YUV_DENOISED 25static void make_grayscale(YV12_BUFFER_CONFIG *yuv); 26#endif 27 28static int absdiff_thresh(BLOCK_SIZE bs, int increase_denoising) { 29 (void)bs; 30 return 3 + (increase_denoising ? 1 : 0); 31} 32 33static int delta_thresh(BLOCK_SIZE bs, int increase_denoising) { 34 (void)bs; 35 (void)increase_denoising; 36 return 4; 37} 38 39static int noise_motion_thresh(BLOCK_SIZE bs, int increase_denoising) { 40 (void)bs; 41 (void)increase_denoising; 42 return 625; 43} 44 45static unsigned int sse_thresh(BLOCK_SIZE bs, int increase_denoising) { 46 return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 80 : 40); 47} 48 49static int sse_diff_thresh(BLOCK_SIZE bs, int increase_denoising, 50 int motion_magnitude) { 51 if (motion_magnitude > noise_motion_thresh(bs, increase_denoising)) { 52 if (increase_denoising) 53 return (1 << num_pels_log2_lookup[bs]) << 2; 54 else 55 return 0; 56 } else { 57 return (1 << num_pels_log2_lookup[bs]) << 4; 58 } 59} 60 61static int total_adj_weak_thresh(BLOCK_SIZE bs, int increase_denoising) { 62 return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 3 : 2); 63} 64 65// TODO(jackychen): If increase_denoising is enabled in the future, 66// we might need to update the code for calculating 'total_adj' in 67// case the C code is not bit-exact with corresponding sse2 code. 68int vp9_denoiser_filter_c(const uint8_t *sig, int sig_stride, 69 const uint8_t *mc_avg, int mc_avg_stride, 70 uint8_t *avg, int avg_stride, int increase_denoising, 71 BLOCK_SIZE bs, int motion_magnitude) { 72 int r, c; 73 const uint8_t *sig_start = sig; 74 const uint8_t *mc_avg_start = mc_avg; 75 uint8_t *avg_start = avg; 76 int diff, adj, absdiff, delta; 77 int adj_val[] = { 3, 4, 6 }; 78 int total_adj = 0; 79 int shift_inc = 1; 80 81 // If motion_magnitude is small, making the denoiser more aggressive by 82 // increasing the adjustment for each level. Add another increment for 83 // blocks that are labeled for increase denoising. 84 if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) { 85 if (increase_denoising) { 86 shift_inc = 2; 87 } 88 adj_val[0] += shift_inc; 89 adj_val[1] += shift_inc; 90 adj_val[2] += shift_inc; 91 } 92 93 // First attempt to apply a strong temporal denoising filter. 94 for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) { 95 for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) { 96 diff = mc_avg[c] - sig[c]; 97 absdiff = abs(diff); 98 99 if (absdiff <= absdiff_thresh(bs, increase_denoising)) { 100 avg[c] = mc_avg[c]; 101 total_adj += diff; 102 } else { 103 switch (absdiff) { 104 case 4: 105 case 5: 106 case 6: 107 case 7: adj = adj_val[0]; break; 108 case 8: 109 case 9: 110 case 10: 111 case 11: 112 case 12: 113 case 13: 114 case 14: 115 case 15: adj = adj_val[1]; break; 116 default: adj = adj_val[2]; 117 } 118 if (diff > 0) { 119 avg[c] = VPXMIN(UINT8_MAX, sig[c] + adj); 120 total_adj += adj; 121 } else { 122 avg[c] = VPXMAX(0, sig[c] - adj); 123 total_adj -= adj; 124 } 125 } 126 } 127 sig += sig_stride; 128 avg += avg_stride; 129 mc_avg += mc_avg_stride; 130 } 131 132 // If the strong filter did not modify the signal too much, we're all set. 133 if (abs(total_adj) <= total_adj_strong_thresh(bs, increase_denoising)) { 134 return FILTER_BLOCK; 135 } 136 137 // Otherwise, we try to dampen the filter if the delta is not too high. 138 delta = ((abs(total_adj) - total_adj_strong_thresh(bs, increase_denoising)) >> 139 num_pels_log2_lookup[bs]) + 140 1; 141 142 if (delta >= delta_thresh(bs, increase_denoising)) { 143 return COPY_BLOCK; 144 } 145 146 mc_avg = mc_avg_start; 147 avg = avg_start; 148 sig = sig_start; 149 for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) { 150 for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) { 151 diff = mc_avg[c] - sig[c]; 152 adj = abs(diff); 153 if (adj > delta) { 154 adj = delta; 155 } 156 if (diff > 0) { 157 // Diff positive means we made positive adjustment above 158 // (in first try/attempt), so now make negative adjustment to bring 159 // denoised signal down. 160 avg[c] = VPXMAX(0, avg[c] - adj); 161 total_adj -= adj; 162 } else { 163 // Diff negative means we made negative adjustment above 164 // (in first try/attempt), so now make positive adjustment to bring 165 // denoised signal up. 166 avg[c] = VPXMIN(UINT8_MAX, avg[c] + adj); 167 total_adj += adj; 168 } 169 } 170 sig += sig_stride; 171 avg += avg_stride; 172 mc_avg += mc_avg_stride; 173 } 174 175 // We can use the filter if it has been sufficiently dampened 176 if (abs(total_adj) <= total_adj_weak_thresh(bs, increase_denoising)) { 177 return FILTER_BLOCK; 178 } 179 return COPY_BLOCK; 180} 181 182static uint8_t *block_start(uint8_t *framebuf, int stride, int mi_row, 183 int mi_col) { 184 return framebuf + (stride * mi_row << 3) + (mi_col << 3); 185} 186 187static VP9_DENOISER_DECISION perform_motion_compensation( 188 VP9_COMMON *const cm, VP9_DENOISER *denoiser, MACROBLOCK *mb, BLOCK_SIZE bs, 189 int increase_denoising, int mi_row, int mi_col, PICK_MODE_CONTEXT *ctx, 190 int motion_magnitude, int is_skin, int *zeromv_filter, int consec_zeromv, 191 int num_spatial_layers, int width, int lst_fb_idx, int gld_fb_idx, 192 int use_svc, int spatial_layer) { 193 const int sse_diff = (ctx->newmv_sse == UINT_MAX) 194 ? 0 195 : ((int)ctx->zeromv_sse - (int)ctx->newmv_sse); 196 int frame; 197 int denoise_layer_idx = 0; 198 MACROBLOCKD *filter_mbd = &mb->e_mbd; 199 MODE_INFO *mi = filter_mbd->mi[0]; 200 MODE_INFO saved_mi; 201 int i; 202 struct buf_2d saved_dst[MAX_MB_PLANE]; 203 struct buf_2d saved_pre[MAX_MB_PLANE]; 204 RefBuffer *saved_block_refs[2]; 205 MV_REFERENCE_FRAME saved_frame; 206 207 frame = ctx->best_reference_frame; 208 209 saved_mi = *mi; 210 211 if (is_skin && (motion_magnitude > 0 || consec_zeromv < 4)) return COPY_BLOCK; 212 213 // Avoid denoising small blocks. When noise > kDenLow or frame width > 480, 214 // denoise 16x16 blocks. 215 if (bs == BLOCK_8X8 || bs == BLOCK_8X16 || bs == BLOCK_16X8 || 216 (bs == BLOCK_16X16 && width > 480 && 217 denoiser->denoising_level <= kDenLow)) 218 return COPY_BLOCK; 219 220 // If the best reference frame uses inter-prediction and there is enough of a 221 // difference in sum-squared-error, use it. 222 if (frame != INTRA_FRAME && frame != ALTREF_FRAME && 223 (frame != GOLDEN_FRAME || num_spatial_layers == 1) && 224 sse_diff > sse_diff_thresh(bs, increase_denoising, motion_magnitude)) { 225 mi->ref_frame[0] = ctx->best_reference_frame; 226 mi->mode = ctx->best_sse_inter_mode; 227 mi->mv[0] = ctx->best_sse_mv; 228 } else { 229 // Otherwise, use the zero reference frame. 230 frame = ctx->best_zeromv_reference_frame; 231 ctx->newmv_sse = ctx->zeromv_sse; 232 // Bias to last reference. 233 if (num_spatial_layers > 1 || frame == ALTREF_FRAME || 234 (frame != LAST_FRAME && 235 ((ctx->zeromv_lastref_sse<(5 * ctx->zeromv_sse)>> 2) || 236 denoiser->denoising_level >= kDenHigh))) { 237 frame = LAST_FRAME; 238 ctx->newmv_sse = ctx->zeromv_lastref_sse; 239 } 240 mi->ref_frame[0] = frame; 241 mi->mode = ZEROMV; 242 mi->mv[0].as_int = 0; 243 ctx->best_sse_inter_mode = ZEROMV; 244 ctx->best_sse_mv.as_int = 0; 245 *zeromv_filter = 1; 246 if (denoiser->denoising_level > kDenMedium) { 247 motion_magnitude = 0; 248 } 249 } 250 251 saved_frame = frame; 252 // When using SVC, we need to map REF_FRAME to the frame buffer index. 253 if (use_svc) { 254 if (frame == LAST_FRAME) 255 frame = lst_fb_idx + 1; 256 else if (frame == GOLDEN_FRAME) 257 frame = gld_fb_idx + 1; 258 // Shift for the second spatial layer. 259 if (num_spatial_layers - spatial_layer == 2) 260 frame = frame + denoiser->num_ref_frames; 261 denoise_layer_idx = num_spatial_layers - spatial_layer - 1; 262 } 263 264 if (ctx->newmv_sse > sse_thresh(bs, increase_denoising)) { 265 // Restore everything to its original state 266 *mi = saved_mi; 267 return COPY_BLOCK; 268 } 269 if (motion_magnitude > (noise_motion_thresh(bs, increase_denoising) << 3)) { 270 // Restore everything to its original state 271 *mi = saved_mi; 272 return COPY_BLOCK; 273 } 274 275 // We will restore these after motion compensation. 276 for (i = 0; i < MAX_MB_PLANE; ++i) { 277 saved_pre[i] = filter_mbd->plane[i].pre[0]; 278 saved_dst[i] = filter_mbd->plane[i].dst; 279 } 280 saved_block_refs[0] = filter_mbd->block_refs[0]; 281 282 // Set the pointers in the MACROBLOCKD to point to the buffers in the denoiser 283 // struct. 284 filter_mbd->plane[0].pre[0].buf = 285 block_start(denoiser->running_avg_y[frame].y_buffer, 286 denoiser->running_avg_y[frame].y_stride, mi_row, mi_col); 287 filter_mbd->plane[0].pre[0].stride = denoiser->running_avg_y[frame].y_stride; 288 filter_mbd->plane[1].pre[0].buf = 289 block_start(denoiser->running_avg_y[frame].u_buffer, 290 denoiser->running_avg_y[frame].uv_stride, mi_row, mi_col); 291 filter_mbd->plane[1].pre[0].stride = denoiser->running_avg_y[frame].uv_stride; 292 filter_mbd->plane[2].pre[0].buf = 293 block_start(denoiser->running_avg_y[frame].v_buffer, 294 denoiser->running_avg_y[frame].uv_stride, mi_row, mi_col); 295 filter_mbd->plane[2].pre[0].stride = denoiser->running_avg_y[frame].uv_stride; 296 297 filter_mbd->plane[0].dst.buf = block_start( 298 denoiser->mc_running_avg_y[denoise_layer_idx].y_buffer, 299 denoiser->mc_running_avg_y[denoise_layer_idx].y_stride, mi_row, mi_col); 300 filter_mbd->plane[0].dst.stride = 301 denoiser->mc_running_avg_y[denoise_layer_idx].y_stride; 302 filter_mbd->plane[1].dst.buf = block_start( 303 denoiser->mc_running_avg_y[denoise_layer_idx].u_buffer, 304 denoiser->mc_running_avg_y[denoise_layer_idx].uv_stride, mi_row, mi_col); 305 filter_mbd->plane[1].dst.stride = 306 denoiser->mc_running_avg_y[denoise_layer_idx].uv_stride; 307 filter_mbd->plane[2].dst.buf = block_start( 308 denoiser->mc_running_avg_y[denoise_layer_idx].v_buffer, 309 denoiser->mc_running_avg_y[denoise_layer_idx].uv_stride, mi_row, mi_col); 310 filter_mbd->plane[2].dst.stride = 311 denoiser->mc_running_avg_y[denoise_layer_idx].uv_stride; 312 313 set_ref_ptrs(cm, filter_mbd, saved_frame, NONE); 314 vp9_build_inter_predictors_sby(filter_mbd, mi_row, mi_col, bs); 315 316 // Restore everything to its original state 317 *mi = saved_mi; 318 filter_mbd->block_refs[0] = saved_block_refs[0]; 319 for (i = 0; i < MAX_MB_PLANE; ++i) { 320 filter_mbd->plane[i].pre[0] = saved_pre[i]; 321 filter_mbd->plane[i].dst = saved_dst[i]; 322 } 323 324 return FILTER_BLOCK; 325} 326 327void vp9_denoiser_denoise(VP9_COMP *cpi, MACROBLOCK *mb, int mi_row, int mi_col, 328 BLOCK_SIZE bs, PICK_MODE_CONTEXT *ctx, 329 VP9_DENOISER_DECISION *denoiser_decision) { 330 int mv_col, mv_row; 331 int motion_magnitude = 0; 332 int zeromv_filter = 0; 333 VP9_DENOISER *denoiser = &cpi->denoiser; 334 VP9_DENOISER_DECISION decision = COPY_BLOCK; 335 336 const int shift = 337 cpi->svc.number_spatial_layers - cpi->svc.spatial_layer_id == 2 338 ? denoiser->num_ref_frames 339 : 0; 340 YV12_BUFFER_CONFIG avg = denoiser->running_avg_y[INTRA_FRAME + shift]; 341 const int denoise_layer_index = 342 cpi->svc.number_spatial_layers - cpi->svc.spatial_layer_id - 1; 343 YV12_BUFFER_CONFIG mc_avg = denoiser->mc_running_avg_y[denoise_layer_index]; 344 uint8_t *avg_start = block_start(avg.y_buffer, avg.y_stride, mi_row, mi_col); 345 346 uint8_t *mc_avg_start = 347 block_start(mc_avg.y_buffer, mc_avg.y_stride, mi_row, mi_col); 348 struct buf_2d src = mb->plane[0].src; 349 int is_skin = 0; 350 int increase_denoising = 0; 351 int consec_zeromv = 0; 352 mv_col = ctx->best_sse_mv.as_mv.col; 353 mv_row = ctx->best_sse_mv.as_mv.row; 354 motion_magnitude = mv_row * mv_row + mv_col * mv_col; 355 356 if (cpi->use_skin_detection && bs <= BLOCK_32X32 && 357 denoiser->denoising_level < kDenHigh) { 358 int motion_level = (motion_magnitude < 16) ? 0 : 1; 359 // If motion for current block is small/zero, compute consec_zeromv for 360 // skin detection (early exit in skin detection is done for large 361 // consec_zeromv when current block has small/zero motion). 362 consec_zeromv = 0; 363 if (motion_level == 0) { 364 VP9_COMMON *const cm = &cpi->common; 365 int j, i; 366 // Loop through the 8x8 sub-blocks. 367 const int bw = num_8x8_blocks_wide_lookup[bs]; 368 const int bh = num_8x8_blocks_high_lookup[bs]; 369 const int xmis = VPXMIN(cm->mi_cols - mi_col, bw); 370 const int ymis = VPXMIN(cm->mi_rows - mi_row, bh); 371 const int block_index = mi_row * cm->mi_cols + mi_col; 372 consec_zeromv = 100; 373 for (i = 0; i < ymis; i++) { 374 for (j = 0; j < xmis; j++) { 375 int bl_index = block_index + i * cm->mi_cols + j; 376 consec_zeromv = VPXMIN(cpi->consec_zero_mv[bl_index], consec_zeromv); 377 // No need to keep checking 8x8 blocks if any of the sub-blocks 378 // has small consec_zeromv (since threshold for no_skin based on 379 // zero/small motion in skin detection is high, i.e, > 4). 380 if (consec_zeromv < 4) { 381 i = ymis; 382 j = xmis; 383 } 384 } 385 } 386 } 387 // TODO(marpan): Compute skin detection over sub-blocks. 388 is_skin = vp9_compute_skin_block( 389 mb->plane[0].src.buf, mb->plane[1].src.buf, mb->plane[2].src.buf, 390 mb->plane[0].src.stride, mb->plane[1].src.stride, bs, consec_zeromv, 391 motion_level); 392 } 393 if (!is_skin && denoiser->denoising_level == kDenHigh) increase_denoising = 1; 394 395 if (denoiser->denoising_level >= kDenLow && !ctx->sb_skip_denoising) 396 decision = perform_motion_compensation( 397 &cpi->common, denoiser, mb, bs, increase_denoising, mi_row, mi_col, ctx, 398 motion_magnitude, is_skin, &zeromv_filter, consec_zeromv, 399 cpi->svc.number_spatial_layers, cpi->Source->y_width, cpi->lst_fb_idx, 400 cpi->gld_fb_idx, cpi->use_svc, cpi->svc.spatial_layer_id); 401 402 if (decision == FILTER_BLOCK) { 403 decision = vp9_denoiser_filter(src.buf, src.stride, mc_avg_start, 404 mc_avg.y_stride, avg_start, avg.y_stride, 405 increase_denoising, bs, motion_magnitude); 406 } 407 408 if (decision == FILTER_BLOCK) { 409 vpx_convolve_copy(avg_start, avg.y_stride, src.buf, src.stride, NULL, 0, 0, 410 0, 0, num_4x4_blocks_wide_lookup[bs] << 2, 411 num_4x4_blocks_high_lookup[bs] << 2); 412 } else { // COPY_BLOCK 413 vpx_convolve_copy(src.buf, src.stride, avg_start, avg.y_stride, NULL, 0, 0, 414 0, 0, num_4x4_blocks_wide_lookup[bs] << 2, 415 num_4x4_blocks_high_lookup[bs] << 2); 416 } 417 *denoiser_decision = decision; 418 if (decision == FILTER_BLOCK && zeromv_filter == 1) 419 *denoiser_decision = FILTER_ZEROMV_BLOCK; 420} 421 422static void copy_frame(YV12_BUFFER_CONFIG *const dest, 423 const YV12_BUFFER_CONFIG *const src) { 424 int r; 425 const uint8_t *srcbuf = src->y_buffer; 426 uint8_t *destbuf = dest->y_buffer; 427 428 assert(dest->y_width == src->y_width); 429 assert(dest->y_height == src->y_height); 430 431 for (r = 0; r < dest->y_height; ++r) { 432 memcpy(destbuf, srcbuf, dest->y_width); 433 destbuf += dest->y_stride; 434 srcbuf += src->y_stride; 435 } 436} 437 438static void swap_frame_buffer(YV12_BUFFER_CONFIG *const dest, 439 YV12_BUFFER_CONFIG *const src) { 440 uint8_t *tmp_buf = dest->y_buffer; 441 assert(dest->y_width == src->y_width); 442 assert(dest->y_height == src->y_height); 443 dest->y_buffer = src->y_buffer; 444 src->y_buffer = tmp_buf; 445} 446 447void vp9_denoiser_update_frame_info( 448 VP9_DENOISER *denoiser, YV12_BUFFER_CONFIG src, FRAME_TYPE frame_type, 449 int refresh_alt_ref_frame, int refresh_golden_frame, int refresh_last_frame, 450 int alt_fb_idx, int gld_fb_idx, int lst_fb_idx, int resized, 451 int svc_base_is_key, int second_spatial_layer) { 452 const int shift = second_spatial_layer ? denoiser->num_ref_frames : 0; 453 // Copy source into denoised reference buffers on KEY_FRAME or 454 // if the just encoded frame was resized. For SVC, copy source if the base 455 // spatial layer was key frame. 456 if (frame_type == KEY_FRAME || resized != 0 || denoiser->reset || 457 svc_base_is_key) { 458 int i; 459 // Start at 1 so as not to overwrite the INTRA_FRAME 460 for (i = 1; i < denoiser->num_ref_frames; ++i) { 461 if (denoiser->running_avg_y[i + shift].buffer_alloc != NULL) 462 copy_frame(&denoiser->running_avg_y[i + shift], &src); 463 } 464 denoiser->reset = 0; 465 return; 466 } 467 468 // If more than one refresh occurs, must copy frame buffer. 469 if ((refresh_alt_ref_frame + refresh_golden_frame + refresh_last_frame) > 1) { 470 if (refresh_alt_ref_frame) { 471 copy_frame(&denoiser->running_avg_y[alt_fb_idx + 1 + shift], 472 &denoiser->running_avg_y[INTRA_FRAME + shift]); 473 } 474 if (refresh_golden_frame) { 475 copy_frame(&denoiser->running_avg_y[gld_fb_idx + 1 + shift], 476 &denoiser->running_avg_y[INTRA_FRAME + shift]); 477 } 478 if (refresh_last_frame) { 479 copy_frame(&denoiser->running_avg_y[lst_fb_idx + 1 + shift], 480 &denoiser->running_avg_y[INTRA_FRAME + shift]); 481 } 482 } else { 483 if (refresh_alt_ref_frame) { 484 swap_frame_buffer(&denoiser->running_avg_y[alt_fb_idx + 1 + shift], 485 &denoiser->running_avg_y[INTRA_FRAME + shift]); 486 } 487 if (refresh_golden_frame) { 488 swap_frame_buffer(&denoiser->running_avg_y[gld_fb_idx + 1 + shift], 489 &denoiser->running_avg_y[INTRA_FRAME + shift]); 490 } 491 if (refresh_last_frame) { 492 swap_frame_buffer(&denoiser->running_avg_y[lst_fb_idx + 1 + shift], 493 &denoiser->running_avg_y[INTRA_FRAME + shift]); 494 } 495 } 496} 497 498void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx) { 499 ctx->zeromv_sse = UINT_MAX; 500 ctx->newmv_sse = UINT_MAX; 501 ctx->zeromv_lastref_sse = UINT_MAX; 502 ctx->best_sse_mv.as_int = 0; 503} 504 505void vp9_denoiser_update_frame_stats(MODE_INFO *mi, unsigned int sse, 506 PREDICTION_MODE mode, 507 PICK_MODE_CONTEXT *ctx) { 508 if (mi->mv[0].as_int == 0 && sse < ctx->zeromv_sse) { 509 ctx->zeromv_sse = sse; 510 ctx->best_zeromv_reference_frame = mi->ref_frame[0]; 511 if (mi->ref_frame[0] == LAST_FRAME) ctx->zeromv_lastref_sse = sse; 512 } 513 514 if (mi->mv[0].as_int != 0 && sse < ctx->newmv_sse) { 515 ctx->newmv_sse = sse; 516 ctx->best_sse_inter_mode = mode; 517 ctx->best_sse_mv = mi->mv[0]; 518 ctx->best_reference_frame = mi->ref_frame[0]; 519 } 520} 521 522static int vp9_denoiser_realloc_svc_helper(VP9_COMMON *cm, 523 VP9_DENOISER *denoiser, int fb_idx) { 524 int fail = 0; 525 if (denoiser->running_avg_y[fb_idx].buffer_alloc == NULL) { 526 fail = 527 vpx_alloc_frame_buffer(&denoiser->running_avg_y[fb_idx], cm->width, 528 cm->height, cm->subsampling_x, cm->subsampling_y, 529#if CONFIG_VP9_HIGHBITDEPTH 530 cm->use_highbitdepth, 531#endif 532 VP9_ENC_BORDER_IN_PIXELS, 0); 533 if (fail) { 534 vp9_denoiser_free(denoiser); 535 return 1; 536 } 537 } 538 return 0; 539} 540 541int vp9_denoiser_realloc_svc(VP9_COMMON *cm, VP9_DENOISER *denoiser, 542 int svc_buf_shift, int refresh_alt, 543 int refresh_gld, int refresh_lst, int alt_fb_idx, 544 int gld_fb_idx, int lst_fb_idx) { 545 int fail = 0; 546 if (refresh_alt) { 547 // Increase the frame buffer index by 1 to map it to the buffer index in the 548 // denoiser. 549 fail = vp9_denoiser_realloc_svc_helper(cm, denoiser, 550 alt_fb_idx + 1 + svc_buf_shift); 551 if (fail) return 1; 552 } 553 if (refresh_gld) { 554 fail = vp9_denoiser_realloc_svc_helper(cm, denoiser, 555 gld_fb_idx + 1 + svc_buf_shift); 556 if (fail) return 1; 557 } 558 if (refresh_lst) { 559 fail = vp9_denoiser_realloc_svc_helper(cm, denoiser, 560 lst_fb_idx + 1 + svc_buf_shift); 561 if (fail) return 1; 562 } 563 return 0; 564} 565 566int vp9_denoiser_alloc(VP9_COMMON *cm, struct SVC *svc, VP9_DENOISER *denoiser, 567 int use_svc, int noise_sen, int width, int height, 568 int ssx, int ssy, 569#if CONFIG_VP9_HIGHBITDEPTH 570 int use_highbitdepth, 571#endif 572 int border) { 573 int i, layer, fail, init_num_ref_frames; 574 const int legacy_byte_alignment = 0; 575 int num_layers = 1; 576 int scaled_width = width; 577 int scaled_height = height; 578 if (use_svc) { 579 LAYER_CONTEXT *lc = &svc->layer_context[svc->spatial_layer_id * 580 svc->number_temporal_layers + 581 svc->temporal_layer_id]; 582 get_layer_resolution(width, height, lc->scaling_factor_num, 583 lc->scaling_factor_den, &scaled_width, &scaled_height); 584 // For SVC: only denoise at most 2 spatial (highest) layers. 585 if (noise_sen >= 2) 586 // Denoise from one spatial layer below the top. 587 svc->first_layer_denoise = VPXMAX(svc->number_spatial_layers - 2, 0); 588 else 589 // Only denoise the top spatial layer. 590 svc->first_layer_denoise = VPXMAX(svc->number_spatial_layers - 1, 0); 591 num_layers = svc->number_spatial_layers - svc->first_layer_denoise; 592 } 593 assert(denoiser != NULL); 594 denoiser->num_ref_frames = use_svc ? SVC_REF_FRAMES : NONSVC_REF_FRAMES; 595 init_num_ref_frames = use_svc ? MAX_REF_FRAMES : NONSVC_REF_FRAMES; 596 denoiser->num_layers = num_layers; 597 CHECK_MEM_ERROR(cm, denoiser->running_avg_y, 598 vpx_calloc(denoiser->num_ref_frames * num_layers, 599 sizeof(denoiser->running_avg_y[0]))); 600 CHECK_MEM_ERROR( 601 cm, denoiser->mc_running_avg_y, 602 vpx_calloc(num_layers, sizeof(denoiser->mc_running_avg_y[0]))); 603 604 for (layer = 0; layer < num_layers; ++layer) { 605 const int denoise_width = (layer == 0) ? width : scaled_width; 606 const int denoise_height = (layer == 0) ? height : scaled_height; 607 for (i = 0; i < init_num_ref_frames; ++i) { 608 fail = vpx_alloc_frame_buffer( 609 &denoiser->running_avg_y[i + denoiser->num_ref_frames * layer], 610 denoise_width, denoise_height, ssx, ssy, 611#if CONFIG_VP9_HIGHBITDEPTH 612 use_highbitdepth, 613#endif 614 border, legacy_byte_alignment); 615 if (fail) { 616 vp9_denoiser_free(denoiser); 617 return 1; 618 } 619#ifdef OUTPUT_YUV_DENOISED 620 make_grayscale(&denoiser->running_avg_y[i]); 621#endif 622 } 623 624 fail = vpx_alloc_frame_buffer(&denoiser->mc_running_avg_y[layer], 625 denoise_width, denoise_height, ssx, ssy, 626#if CONFIG_VP9_HIGHBITDEPTH 627 use_highbitdepth, 628#endif 629 border, legacy_byte_alignment); 630 if (fail) { 631 vp9_denoiser_free(denoiser); 632 return 1; 633 } 634 } 635 636 // denoiser->last_source only used for noise_estimation, so only for top 637 // layer. 638 fail = vpx_alloc_frame_buffer(&denoiser->last_source, width, height, ssx, ssy, 639#if CONFIG_VP9_HIGHBITDEPTH 640 use_highbitdepth, 641#endif 642 border, legacy_byte_alignment); 643 if (fail) { 644 vp9_denoiser_free(denoiser); 645 return 1; 646 } 647#ifdef OUTPUT_YUV_DENOISED 648 make_grayscale(&denoiser->running_avg_y[i]); 649#endif 650 denoiser->frame_buffer_initialized = 1; 651 denoiser->denoising_level = kDenLow; 652 denoiser->prev_denoising_level = kDenLow; 653 denoiser->reset = 0; 654 return 0; 655} 656 657void vp9_denoiser_free(VP9_DENOISER *denoiser) { 658 int i; 659 if (denoiser == NULL) { 660 return; 661 } 662 denoiser->frame_buffer_initialized = 0; 663 for (i = 0; i < denoiser->num_ref_frames * denoiser->num_layers; ++i) { 664 vpx_free_frame_buffer(&denoiser->running_avg_y[i]); 665 } 666 vpx_free(denoiser->running_avg_y); 667 denoiser->running_avg_y = NULL; 668 669 for (i = 0; i < denoiser->num_layers; ++i) { 670 vpx_free_frame_buffer(&denoiser->mc_running_avg_y[i]); 671 } 672 673 vpx_free(denoiser->mc_running_avg_y); 674 denoiser->mc_running_avg_y = NULL; 675 vpx_free_frame_buffer(&denoiser->last_source); 676} 677 678void vp9_denoiser_set_noise_level(VP9_DENOISER *denoiser, int noise_level) { 679 denoiser->denoising_level = noise_level; 680 if (denoiser->denoising_level > kDenLowLow && 681 denoiser->prev_denoising_level == kDenLowLow) 682 denoiser->reset = 1; 683 else 684 denoiser->reset = 0; 685 denoiser->prev_denoising_level = denoiser->denoising_level; 686} 687 688// Scale/increase the partition threshold 689// for denoiser speed-up. 690int64_t vp9_scale_part_thresh(int64_t threshold, VP9_DENOISER_LEVEL noise_level, 691 int content_state, int temporal_layer_id) { 692 if ((content_state == kLowSadLowSumdiff) || 693 (content_state == kHighSadLowSumdiff) || 694 (content_state == kLowVarHighSumdiff) || (noise_level == kDenHigh) || 695 (temporal_layer_id != 0)) { 696 int64_t scaled_thr = 697 (temporal_layer_id < 2) ? (3 * threshold) >> 1 : (7 * threshold) >> 2; 698 return scaled_thr; 699 } else { 700 return (5 * threshold) >> 2; 701 } 702} 703 704// Scale/increase the ac skip threshold for 705// denoiser speed-up. 706int64_t vp9_scale_acskip_thresh(int64_t threshold, 707 VP9_DENOISER_LEVEL noise_level, int abs_sumdiff, 708 int temporal_layer_id) { 709 if (noise_level >= kDenLow && abs_sumdiff < 5) 710 return threshold *= 711 (noise_level == kDenLow) ? 2 : (temporal_layer_id == 2) ? 10 : 6; 712 else 713 return threshold; 714} 715 716#ifdef OUTPUT_YUV_DENOISED 717static void make_grayscale(YV12_BUFFER_CONFIG *yuv) { 718 int r, c; 719 uint8_t *u = yuv->u_buffer; 720 uint8_t *v = yuv->v_buffer; 721 722 for (r = 0; r < yuv->uv_height; ++r) { 723 for (c = 0; c < yuv->uv_width; ++c) { 724 u[c] = UINT8_MAX / 2; 725 v[c] = UINT8_MAX / 2; 726 } 727 u += yuv->uv_stride; 728 v += yuv->uv_stride; 729 } 730} 731#endif 732