1/* 2 * Copyright (c) 2010 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 <math.h> 12#include <limits.h> 13 14#include "vp9/common/vp9_alloccommon.h" 15#include "vp9/common/vp9_onyxc_int.h" 16#include "vp9/common/vp9_quant_common.h" 17#include "vp9/common/vp9_reconinter.h" 18#include "vp9/common/vp9_systemdependent.h" 19#include "vp9/encoder/vp9_extend.h" 20#include "vp9/encoder/vp9_firstpass.h" 21#include "vp9/encoder/vp9_mcomp.h" 22#include "vp9/encoder/vp9_encoder.h" 23#include "vp9/encoder/vp9_quantize.h" 24#include "vp9/encoder/vp9_ratectrl.h" 25#include "vp9/encoder/vp9_segmentation.h" 26#include "vpx_mem/vpx_mem.h" 27#include "vpx_ports/vpx_timer.h" 28#include "vpx_scale/vpx_scale.h" 29 30static int fixed_divide[512]; 31 32static void temporal_filter_predictors_mb_c(MACROBLOCKD *xd, 33 uint8_t *y_mb_ptr, 34 uint8_t *u_mb_ptr, 35 uint8_t *v_mb_ptr, 36 int stride, 37 int uv_block_width, 38 int uv_block_height, 39 int mv_row, 40 int mv_col, 41 uint8_t *pred, 42 struct scale_factors *scale, 43 int x, int y) { 44 const int which_mv = 0; 45 const MV mv = { mv_row, mv_col }; 46 const InterpKernel *const kernel = 47 vp9_get_interp_kernel(xd->mi[0]->mbmi.interp_filter); 48 49 enum mv_precision mv_precision_uv; 50 int uv_stride; 51 if (uv_block_width == 8) { 52 uv_stride = (stride + 1) >> 1; 53 mv_precision_uv = MV_PRECISION_Q4; 54 } else { 55 uv_stride = stride; 56 mv_precision_uv = MV_PRECISION_Q3; 57 } 58 59 vp9_build_inter_predictor(y_mb_ptr, stride, 60 &pred[0], 16, 61 &mv, 62 scale, 63 16, 16, 64 which_mv, 65 kernel, MV_PRECISION_Q3, x, y); 66 67 vp9_build_inter_predictor(u_mb_ptr, uv_stride, 68 &pred[256], uv_block_width, 69 &mv, 70 scale, 71 uv_block_width, uv_block_height, 72 which_mv, 73 kernel, mv_precision_uv, x, y); 74 75 vp9_build_inter_predictor(v_mb_ptr, uv_stride, 76 &pred[512], uv_block_width, 77 &mv, 78 scale, 79 uv_block_width, uv_block_height, 80 which_mv, 81 kernel, mv_precision_uv, x, y); 82} 83 84void vp9_temporal_filter_init() { 85 int i; 86 87 fixed_divide[0] = 0; 88 for (i = 1; i < 512; ++i) 89 fixed_divide[i] = 0x80000 / i; 90} 91 92void vp9_temporal_filter_apply_c(uint8_t *frame1, 93 unsigned int stride, 94 uint8_t *frame2, 95 unsigned int block_width, 96 unsigned int block_height, 97 int strength, 98 int filter_weight, 99 unsigned int *accumulator, 100 uint16_t *count) { 101 unsigned int i, j, k; 102 int modifier; 103 int byte = 0; 104 const int rounding = strength > 0 ? 1 << (strength - 1) : 0; 105 106 for (i = 0, k = 0; i < block_height; i++) { 107 for (j = 0; j < block_width; j++, k++) { 108 int src_byte = frame1[byte]; 109 int pixel_value = *frame2++; 110 111 modifier = src_byte - pixel_value; 112 // This is an integer approximation of: 113 // float coeff = (3.0 * modifer * modifier) / pow(2, strength); 114 // modifier = (int)roundf(coeff > 16 ? 0 : 16-coeff); 115 modifier *= modifier; 116 modifier *= 3; 117 modifier += rounding; 118 modifier >>= strength; 119 120 if (modifier > 16) 121 modifier = 16; 122 123 modifier = 16 - modifier; 124 modifier *= filter_weight; 125 126 count[k] += modifier; 127 accumulator[k] += modifier * pixel_value; 128 129 byte++; 130 } 131 132 byte += stride - block_width; 133 } 134} 135 136static int temporal_filter_find_matching_mb_c(VP9_COMP *cpi, 137 uint8_t *arf_frame_buf, 138 uint8_t *frame_ptr_buf, 139 int stride) { 140 MACROBLOCK *const x = &cpi->mb; 141 MACROBLOCKD *const xd = &x->e_mbd; 142 const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv; 143 int step_param; 144 int sadpb = x->sadperbit16; 145 int bestsme = INT_MAX; 146 int distortion; 147 unsigned int sse; 148 149 MV best_ref_mv1 = {0, 0}; 150 MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */ 151 MV *ref_mv = &x->e_mbd.mi[0]->bmi[0].as_mv[0].as_mv; 152 153 // Save input state 154 struct buf_2d src = x->plane[0].src; 155 struct buf_2d pre = xd->plane[0].pre[0]; 156 157 best_ref_mv1_full.col = best_ref_mv1.col >> 3; 158 best_ref_mv1_full.row = best_ref_mv1.row >> 3; 159 160 // Setup frame pointers 161 x->plane[0].src.buf = arf_frame_buf; 162 x->plane[0].src.stride = stride; 163 xd->plane[0].pre[0].buf = frame_ptr_buf; 164 xd->plane[0].pre[0].stride = stride; 165 166 step_param = mv_sf->reduce_first_step_size; 167 step_param = MIN(step_param, MAX_MVSEARCH_STEPS - 2); 168 169 // Ignore mv costing by sending NULL pointer instead of cost arrays 170 vp9_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1, 171 &cpi->fn_ptr[BLOCK_16X16], 0, &best_ref_mv1, ref_mv); 172 173 // Ignore mv costing by sending NULL pointer instead of cost array 174 bestsme = cpi->find_fractional_mv_step(x, ref_mv, 175 &best_ref_mv1, 176 cpi->common.allow_high_precision_mv, 177 x->errorperbit, 178 &cpi->fn_ptr[BLOCK_16X16], 179 0, mv_sf->subpel_iters_per_step, 180 NULL, NULL, 181 &distortion, &sse, NULL, 0, 0); 182 183 // Restore input state 184 x->plane[0].src = src; 185 xd->plane[0].pre[0] = pre; 186 187 return bestsme; 188} 189 190static void temporal_filter_iterate_c(VP9_COMP *cpi, 191 int frame_count, 192 int alt_ref_index, 193 int strength, 194 struct scale_factors *scale) { 195 int byte; 196 int frame; 197 int mb_col, mb_row; 198 unsigned int filter_weight; 199 int mb_cols = cpi->common.mb_cols; 200 int mb_rows = cpi->common.mb_rows; 201 int mb_y_offset = 0; 202 int mb_uv_offset = 0; 203 DECLARE_ALIGNED_ARRAY(16, unsigned int, accumulator, 16 * 16 * 3); 204 DECLARE_ALIGNED_ARRAY(16, uint16_t, count, 16 * 16 * 3); 205 MACROBLOCKD *mbd = &cpi->mb.e_mbd; 206 YV12_BUFFER_CONFIG *f = cpi->frames[alt_ref_index]; 207 uint8_t *dst1, *dst2; 208 DECLARE_ALIGNED_ARRAY(16, uint8_t, predictor, 16 * 16 * 3); 209 const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y; 210 const int mb_uv_width = 16 >> mbd->plane[1].subsampling_x; 211 212 // Save input state 213 uint8_t* input_buffer[MAX_MB_PLANE]; 214 int i; 215 216 for (i = 0; i < MAX_MB_PLANE; i++) 217 input_buffer[i] = mbd->plane[i].pre[0].buf; 218 219 for (mb_row = 0; mb_row < mb_rows; mb_row++) { 220 // Source frames are extended to 16 pixels. This is different than 221 // L/A/G reference frames that have a border of 32 (VP9ENCBORDERINPIXELS) 222 // A 6/8 tap filter is used for motion search. This requires 2 pixels 223 // before and 3 pixels after. So the largest Y mv on a border would 224 // then be 16 - VP9_INTERP_EXTEND. The UV blocks are half the size of the 225 // Y and therefore only extended by 8. The largest mv that a UV block 226 // can support is 8 - VP9_INTERP_EXTEND. A UV mv is half of a Y mv. 227 // (16 - VP9_INTERP_EXTEND) >> 1 which is greater than 228 // 8 - VP9_INTERP_EXTEND. 229 // To keep the mv in play for both Y and UV planes the max that it 230 // can be on a border is therefore 16 - (2*VP9_INTERP_EXTEND+1). 231 cpi->mb.mv_row_min = -((mb_row * 16) + (17 - 2 * VP9_INTERP_EXTEND)); 232 cpi->mb.mv_row_max = ((cpi->common.mb_rows - 1 - mb_row) * 16) 233 + (17 - 2 * VP9_INTERP_EXTEND); 234 235 for (mb_col = 0; mb_col < mb_cols; mb_col++) { 236 int i, j, k; 237 int stride; 238 239 vpx_memset(accumulator, 0, 16 * 16 * 3 * sizeof(accumulator[0])); 240 vpx_memset(count, 0, 16 * 16 * 3 * sizeof(count[0])); 241 242 cpi->mb.mv_col_min = -((mb_col * 16) + (17 - 2 * VP9_INTERP_EXTEND)); 243 cpi->mb.mv_col_max = ((cpi->common.mb_cols - 1 - mb_col) * 16) 244 + (17 - 2 * VP9_INTERP_EXTEND); 245 246 for (frame = 0; frame < frame_count; frame++) { 247 const int thresh_low = 10000; 248 const int thresh_high = 20000; 249 250 if (cpi->frames[frame] == NULL) 251 continue; 252 253 mbd->mi[0]->bmi[0].as_mv[0].as_mv.row = 0; 254 mbd->mi[0]->bmi[0].as_mv[0].as_mv.col = 0; 255 256 if (frame == alt_ref_index) { 257 filter_weight = 2; 258 } else { 259 // Find best match in this frame by MC 260 int err = temporal_filter_find_matching_mb_c(cpi, 261 cpi->frames[alt_ref_index]->y_buffer + mb_y_offset, 262 cpi->frames[frame]->y_buffer + mb_y_offset, 263 cpi->frames[frame]->y_stride); 264 265 // Assign higher weight to matching MB if it's error 266 // score is lower. If not applying MC default behavior 267 // is to weight all MBs equal. 268 filter_weight = err < thresh_low 269 ? 2 : err < thresh_high ? 1 : 0; 270 } 271 272 if (filter_weight != 0) { 273 // Construct the predictors 274 temporal_filter_predictors_mb_c(mbd, 275 cpi->frames[frame]->y_buffer + mb_y_offset, 276 cpi->frames[frame]->u_buffer + mb_uv_offset, 277 cpi->frames[frame]->v_buffer + mb_uv_offset, 278 cpi->frames[frame]->y_stride, 279 mb_uv_width, mb_uv_height, 280 mbd->mi[0]->bmi[0].as_mv[0].as_mv.row, 281 mbd->mi[0]->bmi[0].as_mv[0].as_mv.col, 282 predictor, scale, 283 mb_col * 16, mb_row * 16); 284 285 // Apply the filter (YUV) 286 vp9_temporal_filter_apply(f->y_buffer + mb_y_offset, f->y_stride, 287 predictor, 16, 16, 288 strength, filter_weight, 289 accumulator, count); 290 vp9_temporal_filter_apply(f->u_buffer + mb_uv_offset, f->uv_stride, 291 predictor + 256, 292 mb_uv_width, mb_uv_height, strength, 293 filter_weight, accumulator + 256, 294 count + 256); 295 vp9_temporal_filter_apply(f->v_buffer + mb_uv_offset, f->uv_stride, 296 predictor + 512, 297 mb_uv_width, mb_uv_height, strength, 298 filter_weight, accumulator + 512, 299 count + 512); 300 } 301 } 302 303 // Normalize filter output to produce AltRef frame 304 dst1 = cpi->alt_ref_buffer.y_buffer; 305 stride = cpi->alt_ref_buffer.y_stride; 306 byte = mb_y_offset; 307 for (i = 0, k = 0; i < 16; i++) { 308 for (j = 0; j < 16; j++, k++) { 309 unsigned int pval = accumulator[k] + (count[k] >> 1); 310 pval *= fixed_divide[count[k]]; 311 pval >>= 19; 312 313 dst1[byte] = (uint8_t)pval; 314 315 // move to next pixel 316 byte++; 317 } 318 byte += stride - 16; 319 } 320 321 dst1 = cpi->alt_ref_buffer.u_buffer; 322 dst2 = cpi->alt_ref_buffer.v_buffer; 323 stride = cpi->alt_ref_buffer.uv_stride; 324 byte = mb_uv_offset; 325 for (i = 0, k = 256; i < mb_uv_height; i++) { 326 for (j = 0; j < mb_uv_width; j++, k++) { 327 int m = k + 256; 328 329 // U 330 unsigned int pval = accumulator[k] + (count[k] >> 1); 331 pval *= fixed_divide[count[k]]; 332 pval >>= 19; 333 dst1[byte] = (uint8_t)pval; 334 335 // V 336 pval = accumulator[m] + (count[m] >> 1); 337 pval *= fixed_divide[count[m]]; 338 pval >>= 19; 339 dst2[byte] = (uint8_t)pval; 340 341 // move to next pixel 342 byte++; 343 } 344 byte += stride - mb_uv_width; 345 } 346 mb_y_offset += 16; 347 mb_uv_offset += mb_uv_width; 348 } 349 mb_y_offset += 16 * (f->y_stride - mb_cols); 350 mb_uv_offset += mb_uv_height * f->uv_stride - mb_uv_width * mb_cols; 351 } 352 353 // Restore input state 354 for (i = 0; i < MAX_MB_PLANE; i++) 355 mbd->plane[i].pre[0].buf = input_buffer[i]; 356} 357 358// Apply buffer limits and context specific adjustments to arnr filter. 359static void adjust_arnr_filter(VP9_COMP *cpi, 360 int distance, int group_boost) { 361 const int frames_after_arf = 362 vp9_lookahead_depth(cpi->lookahead) - distance - 1; 363 int frames_fwd = (cpi->oxcf.arnr_max_frames - 1) >> 1; 364 int frames_bwd; 365 int q; 366 367 // Define the forward and backwards filter limits for this arnr group. 368 if (frames_fwd > frames_after_arf) 369 frames_fwd = frames_after_arf; 370 if (frames_fwd > distance) 371 frames_fwd = distance; 372 373 frames_bwd = frames_fwd; 374 375 // For even length filter there is one more frame backward 376 // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff. 377 if (frames_bwd < distance) 378 frames_bwd += (cpi->oxcf.arnr_max_frames + 1) & 0x1; 379 380 // Set the baseline active filter size. 381 cpi->active_arnr_frames = frames_bwd + 1 + frames_fwd; 382 383 // Adjust the strength based on active max q. 384 if (cpi->common.current_video_frame > 1) 385 q = ((int)vp9_convert_qindex_to_q( 386 cpi->rc.avg_frame_qindex[INTER_FRAME])); 387 else 388 q = ((int)vp9_convert_qindex_to_q( 389 cpi->rc.avg_frame_qindex[KEY_FRAME])); 390 if (q > 16) { 391 cpi->active_arnr_strength = cpi->oxcf.arnr_strength; 392 } else { 393 cpi->active_arnr_strength = cpi->oxcf.arnr_strength - ((16 - q) / 2); 394 if (cpi->active_arnr_strength < 0) 395 cpi->active_arnr_strength = 0; 396 } 397 398 // Adjust number of frames in filter and strength based on gf boost level. 399 if (cpi->active_arnr_frames > (group_boost / 150)) { 400 cpi->active_arnr_frames = (group_boost / 150); 401 cpi->active_arnr_frames += !(cpi->active_arnr_frames & 1); 402 } 403 if (cpi->active_arnr_strength > (group_boost / 300)) { 404 cpi->active_arnr_strength = (group_boost / 300); 405 } 406 407 // Adjustments for second level arf in multi arf case. 408 if (cpi->oxcf.pass == 2 && cpi->multi_arf_allowed) { 409 const GF_GROUP *const gf_group = &cpi->twopass.gf_group; 410 if (gf_group->rf_level[gf_group->index] != GF_ARF_STD) { 411 cpi->active_arnr_strength >>= 1; 412 } 413 } 414} 415 416void vp9_temporal_filter(VP9_COMP *cpi, int distance) { 417 VP9_COMMON *const cm = &cpi->common; 418 RATE_CONTROL *const rc = &cpi->rc; 419 int frame; 420 int frames_to_blur; 421 int start_frame; 422 int strength; 423 int frames_to_blur_backward; 424 int frames_to_blur_forward; 425 struct scale_factors sf; 426 427 // Apply context specific adjustments to the arnr filter parameters. 428 adjust_arnr_filter(cpi, distance, rc->gfu_boost); 429 strength = cpi->active_arnr_strength; 430 frames_to_blur = cpi->active_arnr_frames; 431 frames_to_blur_backward = (frames_to_blur / 2); 432 frames_to_blur_forward = ((frames_to_blur - 1) / 2); 433 start_frame = distance + frames_to_blur_forward; 434 435 // Setup frame pointers, NULL indicates frame not included in filter. 436 vp9_zero(cpi->frames); 437 for (frame = 0; frame < frames_to_blur; ++frame) { 438 const int which_buffer = start_frame - frame; 439 struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead, 440 which_buffer); 441 cpi->frames[frames_to_blur - 1 - frame] = &buf->img; 442 } 443 444 // Setup scaling factors. Scaling on each of the arnr frames is not supported 445 if (is_spatial_svc(cpi)) { 446 // In spatial svc the scaling factors might be less then 1/2. So we will use 447 // non-normative scaling. 448 int frame_used = 0; 449 vp9_setup_scale_factors_for_frame(&sf, 450 get_frame_new_buffer(cm)->y_crop_width, 451 get_frame_new_buffer(cm)->y_crop_height, 452 get_frame_new_buffer(cm)->y_crop_width, 453 get_frame_new_buffer(cm)->y_crop_height); 454 455 for (frame = 0; frame < frames_to_blur; ++frame) { 456 if (cm->mi_cols * MI_SIZE != cpi->frames[frame]->y_width || 457 cm->mi_rows * MI_SIZE != cpi->frames[frame]->y_height) { 458 if (vp9_realloc_frame_buffer(&cpi->svc.scaled_frames[frame_used], 459 cm->width, cm->height, 460 cm->subsampling_x, cm->subsampling_y, 461 VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, 462 NULL)) 463 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 464 "Failed to reallocate alt_ref_buffer"); 465 466 cpi->frames[frame] = 467 vp9_scale_if_required(cm, cpi->frames[frame], 468 &cpi->svc.scaled_frames[frame_used]); 469 ++frame_used; 470 } 471 } 472 } else { 473 vp9_setup_scale_factors_for_frame(&sf, 474 get_frame_new_buffer(cm)->y_crop_width, 475 get_frame_new_buffer(cm)->y_crop_height, 476 cm->width, cm->height); 477 } 478 479 temporal_filter_iterate_c(cpi, frames_to_blur, frames_to_blur_backward, 480 strength, &sf); 481} 482