2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
15 #include "./vp9_rtcd.h"
16 #include "./vpx_config.h"
17 #include "./vpx_dsp_rtcd.h"
18 #include "./vpx_scale_rtcd.h"
19 #include "vpx/internal/vpx_psnr.h"
20 #include "vpx_dsp/vpx_dsp_common.h"
21 #include "vpx_dsp/vpx_filter.h"
22 #if CONFIG_INTERNAL_STATS
23 #include "vpx_dsp/ssim.h"
25 #include "vpx_ports/mem.h"
26 #include "vpx_ports/system_state.h"
27 #include "vpx_ports/vpx_timer.h"
29 #include "vp9/common/vp9_alloccommon.h"
30 #include "vp9/common/vp9_filter.h"
31 #include "vp9/common/vp9_idct.h"
32 #if CONFIG_VP9_POSTPROC
33 #include "vp9/common/vp9_postproc.h"
35 #include "vp9/common/vp9_reconinter.h"
36 #include "vp9/common/vp9_reconintra.h"
37 #include "vp9/common/vp9_tile_common.h"
39 #include "vp9/encoder/vp9_aq_360.h"
40 #include "vp9/encoder/vp9_aq_complexity.h"
41 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
42 #include "vp9/encoder/vp9_aq_variance.h"
43 #include "vp9/encoder/vp9_bitstream.h"
44 #include "vp9/encoder/vp9_context_tree.h"
45 #include "vp9/encoder/vp9_encodeframe.h"
46 #include "vp9/encoder/vp9_encodemv.h"
47 #include "vp9/encoder/vp9_encoder.h"
48 #include "vp9/encoder/vp9_ethread.h"
49 #include "vp9/encoder/vp9_firstpass.h"
50 #include "vp9/encoder/vp9_mbgraph.h"
51 #include "vp9/encoder/vp9_noise_estimate.h"
52 #include "vp9/encoder/vp9_picklpf.h"
53 #include "vp9/encoder/vp9_ratectrl.h"
54 #include "vp9/encoder/vp9_rd.h"
55 #include "vp9/encoder/vp9_resize.h"
56 #include "vp9/encoder/vp9_segmentation.h"
57 #include "vp9/encoder/vp9_skin_detection.h"
58 #include "vp9/encoder/vp9_speed_features.h"
59 #include "vp9/encoder/vp9_svc_layercontext.h"
60 #include "vp9/encoder/vp9_temporal_filter.h"
62 #define AM_SEGMENT_ID_INACTIVE 7
63 #define AM_SEGMENT_ID_ACTIVE 0
65 #define ALTREF_HIGH_PRECISION_MV 1 // Whether to use high precision mv
66 // for altref computation.
67 #define HIGH_PRECISION_MV_QTHRESH 200 // Q threshold for high precision
68 // mv. Choose a very high value for
69 // now so that HIGH_PRECISION is always
71 // #define OUTPUT_YUV_REC
73 #ifdef OUTPUT_YUV_DENOISED
74 FILE *yuv_denoised_file = NULL;
76 #ifdef OUTPUT_YUV_SKINMAP
77 FILE *yuv_skinmap_file = NULL;
89 static const Vp9LevelSpec vp9_level_defs[VP9_LEVELS] = {
90 {LEVEL_1, 829440, 36864, 200, 400, 2, 1, 4, 8},
91 {LEVEL_1_1, 2764800, 73728, 800, 1000, 2, 1, 4, 8},
92 {LEVEL_2, 4608000, 122880, 1800, 1500, 2, 1, 4, 8},
93 {LEVEL_2_1, 9216000, 245760, 3600, 2800, 2, 2, 4, 8},
94 {LEVEL_3, 20736000, 552960, 7200, 6000, 2, 4, 4, 8},
95 {LEVEL_3_1, 36864000, 983040, 12000, 10000, 2, 4, 4, 8},
96 {LEVEL_4, 83558400, 2228224, 18000, 16000, 4, 4, 4, 8},
97 {LEVEL_4_1, 160432128, 2228224, 30000, 18000, 4, 4, 5, 6},
98 {LEVEL_5, 311951360, 8912896, 60000, 36000, 6, 8, 6, 4},
99 {LEVEL_5_1, 588251136, 8912896, 120000, 46000, 8, 8, 10, 4},
100 // TODO(huisu): update max_cpb_size for level 5_2 ~ 6_2 when
101 // they are finalized (currently TBD).
102 {LEVEL_5_2, 1176502272, 8912896, 180000, 0, 8, 8, 10, 4},
103 {LEVEL_6, 1176502272, 35651584, 180000, 0, 8, 16, 10, 4},
104 {LEVEL_6_1, 2353004544u, 35651584, 240000, 0, 8, 16, 10, 4},
105 {LEVEL_6_2, 4706009088u, 35651584, 480000, 0, 8, 16, 10, 4},
108 static INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) {
134 // Mark all inactive blocks as active. Other segmentation features may be set
135 // so memset cannot be used, instead only inactive blocks should be reset.
136 static void suppress_active_map(VP9_COMP *cpi) {
137 unsigned char *const seg_map = cpi->segmentation_map;
139 if (cpi->active_map.enabled || cpi->active_map.update) {
140 const int rows = cpi->common.mi_rows;
141 const int cols = cpi->common.mi_cols;
144 for (i = 0; i < rows * cols; ++i)
145 if (seg_map[i] == AM_SEGMENT_ID_INACTIVE)
146 seg_map[i] = AM_SEGMENT_ID_ACTIVE;
150 static void apply_active_map(VP9_COMP *cpi) {
151 struct segmentation *const seg = &cpi->common.seg;
152 unsigned char *const seg_map = cpi->segmentation_map;
153 const unsigned char *const active_map = cpi->active_map.map;
156 assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE);
158 if (frame_is_intra_only(&cpi->common)) {
159 cpi->active_map.enabled = 0;
160 cpi->active_map.update = 1;
163 if (cpi->active_map.update) {
164 if (cpi->active_map.enabled) {
165 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
166 if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i];
167 vp9_enable_segmentation(seg);
168 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
169 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
170 // Setting the data to -MAX_LOOP_FILTER will result in the computed loop
171 // filter level being zero regardless of the value of seg->abs_delta.
172 vp9_set_segdata(seg, AM_SEGMENT_ID_INACTIVE,
173 SEG_LVL_ALT_LF, -MAX_LOOP_FILTER);
175 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
176 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
178 seg->update_data = 1;
182 cpi->active_map.update = 0;
186 static void init_level_info(Vp9LevelInfo *level_info) {
187 Vp9LevelStats *const level_stats = &level_info->level_stats;
188 Vp9LevelSpec *const level_spec = &level_info->level_spec;
190 memset(level_stats, 0, sizeof(*level_stats));
191 memset(level_spec, 0, sizeof(*level_spec));
192 level_spec->level = LEVEL_UNKNOWN;
193 level_spec->min_altref_distance = INT_MAX;
196 VP9_LEVEL vp9_get_level(const Vp9LevelSpec * const level_spec) {
198 const Vp9LevelSpec *this_level;
200 vpx_clear_system_state();
202 for (i = 0; i < VP9_LEVELS; ++i) {
203 this_level = &vp9_level_defs[i];
204 if ((double)level_spec->max_luma_sample_rate * (1 + SAMPLE_RATE_GRACE_P) >
205 (double)this_level->max_luma_sample_rate ||
206 level_spec->max_luma_picture_size > this_level->max_luma_picture_size ||
207 level_spec->average_bitrate > this_level->average_bitrate ||
208 level_spec->max_cpb_size > this_level->max_cpb_size ||
209 level_spec->compression_ratio < this_level->compression_ratio ||
210 level_spec->max_col_tiles > this_level->max_col_tiles ||
211 level_spec->min_altref_distance < this_level->min_altref_distance ||
212 level_spec->max_ref_frame_buffers > this_level->max_ref_frame_buffers)
216 return (i == VP9_LEVELS) ? LEVEL_UNKNOWN : vp9_level_defs[i].level;
219 int vp9_set_active_map(VP9_COMP* cpi,
220 unsigned char* new_map_16x16,
223 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
224 unsigned char *const active_map_8x8 = cpi->active_map.map;
225 const int mi_rows = cpi->common.mi_rows;
226 const int mi_cols = cpi->common.mi_cols;
227 cpi->active_map.update = 1;
230 for (r = 0; r < mi_rows; ++r) {
231 for (c = 0; c < mi_cols; ++c) {
232 active_map_8x8[r * mi_cols + c] =
233 new_map_16x16[(r >> 1) * cols + (c >> 1)]
234 ? AM_SEGMENT_ID_ACTIVE
235 : AM_SEGMENT_ID_INACTIVE;
238 cpi->active_map.enabled = 1;
240 cpi->active_map.enabled = 0;
248 int vp9_get_active_map(VP9_COMP* cpi,
249 unsigned char* new_map_16x16,
252 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols &&
254 unsigned char* const seg_map_8x8 = cpi->segmentation_map;
255 const int mi_rows = cpi->common.mi_rows;
256 const int mi_cols = cpi->common.mi_cols;
257 memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
258 if (cpi->active_map.enabled) {
260 for (r = 0; r < mi_rows; ++r) {
261 for (c = 0; c < mi_cols; ++c) {
262 // Cyclic refresh segments are considered active despite not having
263 // AM_SEGMENT_ID_ACTIVE
264 new_map_16x16[(r >> 1) * cols + (c >> 1)] |=
265 seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE;
275 void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
276 MACROBLOCK *const mb = &cpi->td.mb;
277 cpi->common.allow_high_precision_mv = allow_high_precision_mv;
278 if (cpi->common.allow_high_precision_mv) {
279 mb->mvcost = mb->nmvcost_hp;
280 mb->mvsadcost = mb->nmvsadcost_hp;
282 mb->mvcost = mb->nmvcost;
283 mb->mvsadcost = mb->nmvsadcost;
287 static void setup_frame(VP9_COMP *cpi) {
288 VP9_COMMON *const cm = &cpi->common;
289 // Set up entropy context depending on frame type. The decoder mandates
290 // the use of the default context, index 0, for keyframes and inter
291 // frames where the error_resilient_mode or intra_only flag is set. For
292 // other inter-frames the encoder currently uses only two contexts;
293 // context 1 for ALTREF frames and context 0 for the others.
294 if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
295 vp9_setup_past_independence(cm);
298 cm->frame_context_idx = cpi->refresh_alt_ref_frame;
301 if (cm->frame_type == KEY_FRAME) {
302 if (!is_two_pass_svc(cpi))
303 cpi->refresh_golden_frame = 1;
304 cpi->refresh_alt_ref_frame = 1;
305 vp9_zero(cpi->interp_filter_selected);
307 *cm->fc = cm->frame_contexts[cm->frame_context_idx];
308 vp9_zero(cpi->interp_filter_selected[0]);
312 static void vp9_enc_setup_mi(VP9_COMMON *cm) {
314 cm->mi = cm->mip + cm->mi_stride + 1;
315 memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
316 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
317 // Clear top border row
318 memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride);
319 // Clear left border column
320 for (i = 1; i < cm->mi_rows + 1; ++i)
321 memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip));
323 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
324 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
326 memset(cm->mi_grid_base, 0,
327 cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
330 static int vp9_enc_alloc_mi(VP9_COMMON *cm, int mi_size) {
331 cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
334 cm->prev_mip = vpx_calloc(mi_size, sizeof(*cm->prev_mip));
337 cm->mi_alloc_size = mi_size;
339 cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
340 if (!cm->mi_grid_base)
342 cm->prev_mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
343 if (!cm->prev_mi_grid_base)
349 static void vp9_enc_free_mi(VP9_COMMON *cm) {
352 vpx_free(cm->prev_mip);
354 vpx_free(cm->mi_grid_base);
355 cm->mi_grid_base = NULL;
356 vpx_free(cm->prev_mi_grid_base);
357 cm->prev_mi_grid_base = NULL;
360 static void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
361 // Current mip will be the prev_mip for the next frame.
362 MODE_INFO **temp_base = cm->prev_mi_grid_base;
363 MODE_INFO *temp = cm->prev_mip;
364 cm->prev_mip = cm->mip;
367 // Update the upper left visible macroblock ptrs.
368 cm->mi = cm->mip + cm->mi_stride + 1;
369 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
371 cm->prev_mi_grid_base = cm->mi_grid_base;
372 cm->mi_grid_base = temp_base;
373 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
374 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
377 void vp9_initialize_enc(void) {
378 static volatile int init_done = 0;
384 vp9_init_intra_predictors();
386 vp9_rc_init_minq_luts();
387 vp9_entropy_mv_init();
388 vp9_temporal_filter_init();
393 static void dealloc_compressor_data(VP9_COMP *cpi) {
394 VP9_COMMON *const cm = &cpi->common;
397 vpx_free(cpi->mbmi_ext_base);
398 cpi->mbmi_ext_base = NULL;
400 vpx_free(cpi->tile_data);
401 cpi->tile_data = NULL;
403 // Delete sementation map
404 vpx_free(cpi->segmentation_map);
405 cpi->segmentation_map = NULL;
406 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
407 cpi->coding_context.last_frame_seg_map_copy = NULL;
409 vpx_free(cpi->nmvcosts[0]);
410 vpx_free(cpi->nmvcosts[1]);
411 cpi->nmvcosts[0] = NULL;
412 cpi->nmvcosts[1] = NULL;
414 vpx_free(cpi->nmvcosts_hp[0]);
415 vpx_free(cpi->nmvcosts_hp[1]);
416 cpi->nmvcosts_hp[0] = NULL;
417 cpi->nmvcosts_hp[1] = NULL;
419 vpx_free(cpi->nmvsadcosts[0]);
420 vpx_free(cpi->nmvsadcosts[1]);
421 cpi->nmvsadcosts[0] = NULL;
422 cpi->nmvsadcosts[1] = NULL;
424 vpx_free(cpi->nmvsadcosts_hp[0]);
425 vpx_free(cpi->nmvsadcosts_hp[1]);
426 cpi->nmvsadcosts_hp[0] = NULL;
427 cpi->nmvsadcosts_hp[1] = NULL;
429 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
430 cpi->cyclic_refresh = NULL;
432 vpx_free(cpi->active_map.map);
433 cpi->active_map.map = NULL;
435 vpx_free(cpi->consec_zero_mv);
436 cpi->consec_zero_mv = NULL;
438 vp9_free_ref_frame_buffers(cm->buffer_pool);
439 #if CONFIG_VP9_POSTPROC
440 vp9_free_postproc_buffers(cm);
442 vp9_free_context_buffers(cm);
444 vpx_free_frame_buffer(&cpi->last_frame_uf);
445 vpx_free_frame_buffer(&cpi->scaled_source);
446 vpx_free_frame_buffer(&cpi->scaled_last_source);
447 vpx_free_frame_buffer(&cpi->alt_ref_buffer);
448 vp9_lookahead_destroy(cpi->lookahead);
450 vpx_free(cpi->tile_tok[0][0]);
451 cpi->tile_tok[0][0] = 0;
453 vp9_free_pc_tree(&cpi->td);
455 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
456 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
457 vpx_free(lc->rc_twopass_stats_in.buf);
458 lc->rc_twopass_stats_in.buf = NULL;
459 lc->rc_twopass_stats_in.sz = 0;
462 if (cpi->source_diff_var != NULL) {
463 vpx_free(cpi->source_diff_var);
464 cpi->source_diff_var = NULL;
467 for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
468 vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]);
470 memset(&cpi->svc.scaled_frames[0], 0,
471 MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
473 vpx_free_frame_buffer(&cpi->svc.scaled_temp);
474 memset(&cpi->svc.scaled_temp, 0, sizeof(cpi->svc.scaled_temp));
476 vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
477 memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
479 vp9_free_svc_cyclic_refresh(cpi);
482 static void save_coding_context(VP9_COMP *cpi) {
483 CODING_CONTEXT *const cc = &cpi->coding_context;
484 VP9_COMMON *cm = &cpi->common;
486 // Stores a snapshot of key state variables which can subsequently be
487 // restored with a call to vp9_restore_coding_context. These functions are
488 // intended for use in a re-code loop in vp9_compress_frame where the
489 // quantizer value is adjusted between loop iterations.
490 vp9_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost);
492 memcpy(cc->nmvcosts[0], cpi->nmvcosts[0],
493 MV_VALS * sizeof(*cpi->nmvcosts[0]));
494 memcpy(cc->nmvcosts[1], cpi->nmvcosts[1],
495 MV_VALS * sizeof(*cpi->nmvcosts[1]));
496 memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0],
497 MV_VALS * sizeof(*cpi->nmvcosts_hp[0]));
498 memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1],
499 MV_VALS * sizeof(*cpi->nmvcosts_hp[1]));
501 vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs);
503 memcpy(cpi->coding_context.last_frame_seg_map_copy,
504 cm->last_frame_seg_map, (cm->mi_rows * cm->mi_cols));
506 vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
507 vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
512 static void restore_coding_context(VP9_COMP *cpi) {
513 CODING_CONTEXT *const cc = &cpi->coding_context;
514 VP9_COMMON *cm = &cpi->common;
516 // Restore key state variables to the snapshot state stored in the
517 // previous call to vp9_save_coding_context.
518 vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
520 memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0]));
521 memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1]));
522 memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0],
523 MV_VALS * sizeof(*cc->nmvcosts_hp[0]));
524 memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1],
525 MV_VALS * sizeof(*cc->nmvcosts_hp[1]));
527 vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs);
529 memcpy(cm->last_frame_seg_map,
530 cpi->coding_context.last_frame_seg_map_copy,
531 (cm->mi_rows * cm->mi_cols));
533 vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
534 vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
539 static void configure_static_seg_features(VP9_COMP *cpi) {
540 VP9_COMMON *const cm = &cpi->common;
541 const RATE_CONTROL *const rc = &cpi->rc;
542 struct segmentation *const seg = &cm->seg;
544 int high_q = (int)(rc->avg_q > 48.0);
547 // Disable and clear down for KF
548 if (cm->frame_type == KEY_FRAME) {
549 // Clear down the global segmentation map
550 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
552 seg->update_data = 0;
553 cpi->static_mb_pct = 0;
555 // Disable segmentation
556 vp9_disable_segmentation(seg);
558 // Clear down the segment features.
559 vp9_clearall_segfeatures(seg);
560 } else if (cpi->refresh_alt_ref_frame) {
561 // If this is an alt ref frame
562 // Clear down the global segmentation map
563 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
565 seg->update_data = 0;
566 cpi->static_mb_pct = 0;
568 // Disable segmentation and individual segment features by default
569 vp9_disable_segmentation(seg);
570 vp9_clearall_segfeatures(seg);
572 // Scan frames from current to arf frame.
573 // This function re-enables segmentation if appropriate.
574 vp9_update_mbgraph_stats(cpi);
576 // If segmentation was enabled set those features needed for the
580 seg->update_data = 1;
582 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875,
584 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
585 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
587 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
588 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
590 // Where relevant assume segment data is delta data
591 seg->abs_delta = SEGMENT_DELTADATA;
593 } else if (seg->enabled) {
594 // All other frames if segmentation has been enabled
596 // First normal frame in a valid gf or alt ref group
597 if (rc->frames_since_golden == 0) {
598 // Set up segment features for normal frames in an arf group
599 if (rc->source_alt_ref_active) {
601 seg->update_data = 1;
602 seg->abs_delta = SEGMENT_DELTADATA;
604 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125,
606 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
607 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
609 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
610 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
612 // Segment coding disabled for compred testing
613 if (high_q || (cpi->static_mb_pct == 100)) {
614 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
615 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
616 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
619 // Disable segmentation and clear down features if alt ref
620 // is not active for this group
622 vp9_disable_segmentation(seg);
624 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
627 seg->update_data = 0;
629 vp9_clearall_segfeatures(seg);
631 } else if (rc->is_src_frame_alt_ref) {
632 // Special case where we are coding over the top of a previous
634 // Segment coding disabled for compred testing
636 // Enable ref frame features for segment 0 as well
637 vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
638 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
640 // All mbs should use ALTREF_FRAME
641 vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
642 vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
643 vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
644 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
646 // Skip all MBs if high Q (0,0 mv and skip coeffs)
648 vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP);
649 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
651 // Enable data update
652 seg->update_data = 1;
656 // No updates.. leave things as they are.
658 seg->update_data = 0;
663 static void update_reference_segmentation_map(VP9_COMP *cpi) {
664 VP9_COMMON *const cm = &cpi->common;
665 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
666 uint8_t *cache_ptr = cm->last_frame_seg_map;
669 for (row = 0; row < cm->mi_rows; row++) {
670 MODE_INFO **mi_8x8 = mi_8x8_ptr;
671 uint8_t *cache = cache_ptr;
672 for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
673 cache[0] = mi_8x8[0]->segment_id;
674 mi_8x8_ptr += cm->mi_stride;
675 cache_ptr += cm->mi_cols;
679 static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
680 VP9_COMMON *cm = &cpi->common;
681 const VP9EncoderConfig *oxcf = &cpi->oxcf;
684 cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
685 cm->subsampling_x, cm->subsampling_y,
686 #if CONFIG_VP9_HIGHBITDEPTH
687 cm->use_highbitdepth,
689 oxcf->lag_in_frames);
691 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
692 "Failed to allocate lag buffers");
694 // TODO(agrange) Check if ARF is enabled and skip allocation if not.
695 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
696 oxcf->width, oxcf->height,
697 cm->subsampling_x, cm->subsampling_y,
698 #if CONFIG_VP9_HIGHBITDEPTH
699 cm->use_highbitdepth,
701 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
703 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
704 "Failed to allocate altref buffer");
707 static void alloc_util_frame_buffers(VP9_COMP *cpi) {
708 VP9_COMMON *const cm = &cpi->common;
709 if (vpx_realloc_frame_buffer(&cpi->last_frame_uf,
710 cm->width, cm->height,
711 cm->subsampling_x, cm->subsampling_y,
712 #if CONFIG_VP9_HIGHBITDEPTH
713 cm->use_highbitdepth,
715 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
717 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
718 "Failed to allocate last frame buffer");
720 if (vpx_realloc_frame_buffer(&cpi->scaled_source,
721 cm->width, cm->height,
722 cm->subsampling_x, cm->subsampling_y,
723 #if CONFIG_VP9_HIGHBITDEPTH
724 cm->use_highbitdepth,
726 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
728 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
729 "Failed to allocate scaled source buffer");
731 if (vpx_realloc_frame_buffer(&cpi->scaled_last_source,
732 cm->width, cm->height,
733 cm->subsampling_x, cm->subsampling_y,
734 #if CONFIG_VP9_HIGHBITDEPTH
735 cm->use_highbitdepth,
737 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
739 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
740 "Failed to allocate scaled last source buffer");
744 static int alloc_context_buffers_ext(VP9_COMP *cpi) {
745 VP9_COMMON *cm = &cpi->common;
746 int mi_size = cm->mi_cols * cm->mi_rows;
748 cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
749 if (!cpi->mbmi_ext_base)
755 static void alloc_compressor_data(VP9_COMP *cpi) {
756 VP9_COMMON *cm = &cpi->common;
758 vp9_alloc_context_buffers(cm, cm->width, cm->height);
760 alloc_context_buffers_ext(cpi);
762 vpx_free(cpi->tile_tok[0][0]);
765 unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
766 CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
767 vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
770 vp9_setup_pc_tree(&cpi->common, &cpi->td);
773 void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
774 cpi->framerate = framerate < 0.1 ? 30 : framerate;
775 vp9_rc_update_framerate(cpi);
778 static void set_tile_limits(VP9_COMP *cpi) {
779 VP9_COMMON *const cm = &cpi->common;
781 int min_log2_tile_cols, max_log2_tile_cols;
782 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
784 if (is_two_pass_svc(cpi) &&
785 (cpi->svc.encode_empty_frame_state == ENCODING ||
786 cpi->svc.number_spatial_layers > 1)) {
787 cm->log2_tile_cols = 0;
788 cm->log2_tile_rows = 0;
790 cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns,
791 min_log2_tile_cols, max_log2_tile_cols);
792 cm->log2_tile_rows = cpi->oxcf.tile_rows;
796 static void update_frame_size(VP9_COMP *cpi) {
797 VP9_COMMON *const cm = &cpi->common;
798 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
800 vp9_set_mb_mi(cm, cm->width, cm->height);
801 vp9_init_context_buffers(cm);
802 vp9_init_macroblockd(cm, xd, NULL);
803 cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base;
804 memset(cpi->mbmi_ext_base, 0,
805 cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
807 set_tile_limits(cpi);
809 if (is_two_pass_svc(cpi)) {
810 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
811 cm->width, cm->height,
812 cm->subsampling_x, cm->subsampling_y,
813 #if CONFIG_VP9_HIGHBITDEPTH
814 cm->use_highbitdepth,
816 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
818 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
819 "Failed to reallocate alt_ref_buffer");
823 static void init_buffer_indices(VP9_COMP *cpi) {
829 static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
830 VP9_COMMON *const cm = &cpi->common;
833 cpi->framerate = oxcf->init_framerate;
834 cm->profile = oxcf->profile;
835 cm->bit_depth = oxcf->bit_depth;
836 #if CONFIG_VP9_HIGHBITDEPTH
837 cm->use_highbitdepth = oxcf->use_highbitdepth;
839 cm->color_space = oxcf->color_space;
840 cm->color_range = oxcf->color_range;
842 cpi->target_level = oxcf->target_level;
843 cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
845 cm->width = oxcf->width;
846 cm->height = oxcf->height;
847 alloc_compressor_data(cpi);
849 cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode;
851 // Single thread case: use counts in common.
852 cpi->td.counts = &cm->counts;
854 // Spatial scalability.
855 cpi->svc.number_spatial_layers = oxcf->ss_number_layers;
856 // Temporal scalability.
857 cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
859 if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
860 ((cpi->svc.number_temporal_layers > 1 ||
861 cpi->svc.number_spatial_layers > 1) &&
862 cpi->oxcf.pass != 1)) {
863 vp9_init_layer_context(cpi);
866 // change includes all joint functionality
867 vp9_change_config(cpi, oxcf);
869 cpi->static_mb_pct = 0;
870 cpi->ref_frame_flags = 0;
872 init_buffer_indices(cpi);
874 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
877 static void set_rc_buffer_sizes(RATE_CONTROL *rc,
878 const VP9EncoderConfig *oxcf) {
879 const int64_t bandwidth = oxcf->target_bandwidth;
880 const int64_t starting = oxcf->starting_buffer_level_ms;
881 const int64_t optimal = oxcf->optimal_buffer_level_ms;
882 const int64_t maximum = oxcf->maximum_buffer_size_ms;
884 rc->starting_buffer_level = starting * bandwidth / 1000;
885 rc->optimal_buffer_level = (optimal == 0) ? bandwidth / 8
886 : optimal * bandwidth / 1000;
887 rc->maximum_buffer_size = (maximum == 0) ? bandwidth / 8
888 : maximum * bandwidth / 1000;
891 #if CONFIG_VP9_HIGHBITDEPTH
892 #define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
893 cpi->fn_ptr[BT].sdf = SDF; \
894 cpi->fn_ptr[BT].sdaf = SDAF; \
895 cpi->fn_ptr[BT].vf = VF; \
896 cpi->fn_ptr[BT].svf = SVF; \
897 cpi->fn_ptr[BT].svaf = SVAF; \
898 cpi->fn_ptr[BT].sdx3f = SDX3F; \
899 cpi->fn_ptr[BT].sdx8f = SDX8F; \
900 cpi->fn_ptr[BT].sdx4df = SDX4DF;
902 #define MAKE_BFP_SAD_WRAPPER(fnname) \
903 static unsigned int fnname##_bits8(const uint8_t *src_ptr, \
905 const uint8_t *ref_ptr, \
907 return fnname(src_ptr, source_stride, ref_ptr, ref_stride); \
909 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
911 const uint8_t *ref_ptr, \
913 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2; \
915 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
917 const uint8_t *ref_ptr, \
919 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4; \
922 #define MAKE_BFP_SADAVG_WRAPPER(fnname) static unsigned int \
923 fnname##_bits8(const uint8_t *src_ptr, \
925 const uint8_t *ref_ptr, \
927 const uint8_t *second_pred) { \
928 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred); \
930 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
932 const uint8_t *ref_ptr, \
934 const uint8_t *second_pred) { \
935 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
938 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
940 const uint8_t *ref_ptr, \
942 const uint8_t *second_pred) { \
943 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
947 #define MAKE_BFP_SAD3_WRAPPER(fnname) \
948 static void fnname##_bits8(const uint8_t *src_ptr, \
950 const uint8_t *ref_ptr, \
952 unsigned int *sad_array) { \
953 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
955 static void fnname##_bits10(const uint8_t *src_ptr, \
957 const uint8_t *ref_ptr, \
959 unsigned int *sad_array) { \
961 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
962 for (i = 0; i < 3; i++) \
963 sad_array[i] >>= 2; \
965 static void fnname##_bits12(const uint8_t *src_ptr, \
967 const uint8_t *ref_ptr, \
969 unsigned int *sad_array) { \
971 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
972 for (i = 0; i < 3; i++) \
973 sad_array[i] >>= 4; \
976 #define MAKE_BFP_SAD8_WRAPPER(fnname) \
977 static void fnname##_bits8(const uint8_t *src_ptr, \
979 const uint8_t *ref_ptr, \
981 unsigned int *sad_array) { \
982 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
984 static void fnname##_bits10(const uint8_t *src_ptr, \
986 const uint8_t *ref_ptr, \
988 unsigned int *sad_array) { \
990 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
991 for (i = 0; i < 8; i++) \
992 sad_array[i] >>= 2; \
994 static void fnname##_bits12(const uint8_t *src_ptr, \
996 const uint8_t *ref_ptr, \
998 unsigned int *sad_array) { \
1000 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1001 for (i = 0; i < 8; i++) \
1002 sad_array[i] >>= 4; \
1004 #define MAKE_BFP_SAD4D_WRAPPER(fnname) \
1005 static void fnname##_bits8(const uint8_t *src_ptr, \
1006 int source_stride, \
1007 const uint8_t* const ref_ptr[], \
1009 unsigned int *sad_array) { \
1010 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1012 static void fnname##_bits10(const uint8_t *src_ptr, \
1013 int source_stride, \
1014 const uint8_t* const ref_ptr[], \
1016 unsigned int *sad_array) { \
1018 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1019 for (i = 0; i < 4; i++) \
1020 sad_array[i] >>= 2; \
1022 static void fnname##_bits12(const uint8_t *src_ptr, \
1023 int source_stride, \
1024 const uint8_t* const ref_ptr[], \
1026 unsigned int *sad_array) { \
1028 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1029 for (i = 0; i < 4; i++) \
1030 sad_array[i] >>= 4; \
1033 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16)
1034 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg)
1035 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d)
1036 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32)
1037 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg)
1038 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d)
1039 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32)
1040 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg)
1041 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d)
1042 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64)
1043 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg)
1044 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d)
1045 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32)
1046 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg)
1047 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad32x32x3)
1048 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad32x32x8)
1049 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d)
1050 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64)
1051 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg)
1052 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad64x64x3)
1053 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad64x64x8)
1054 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d)
1055 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16)
1056 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg)
1057 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x16x3)
1058 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x16x8)
1059 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d)
1060 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8)
1061 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg)
1062 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x8x3)
1063 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x8x8)
1064 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d)
1065 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16)
1066 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg)
1067 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x16x3)
1068 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x16x8)
1069 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d)
1070 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8)
1071 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg)
1072 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x8x3)
1073 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x8x8)
1074 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d)
1075 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4)
1076 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg)
1077 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x4x8)
1078 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d)
1079 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8)
1080 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg)
1081 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x8x8)
1082 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d)
1083 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4)
1084 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg)
1085 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad4x4x3)
1086 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x4x8)
1087 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d)
1089 static void highbd_set_var_fns(VP9_COMP *const cpi) {
1090 VP9_COMMON *const cm = &cpi->common;
1091 if (cm->use_highbitdepth) {
1092 switch (cm->bit_depth) {
1094 HIGHBD_BFP(BLOCK_32X16,
1095 vpx_highbd_sad32x16_bits8,
1096 vpx_highbd_sad32x16_avg_bits8,
1097 vpx_highbd_8_variance32x16,
1098 vpx_highbd_8_sub_pixel_variance32x16,
1099 vpx_highbd_8_sub_pixel_avg_variance32x16,
1102 vpx_highbd_sad32x16x4d_bits8)
1104 HIGHBD_BFP(BLOCK_16X32,
1105 vpx_highbd_sad16x32_bits8,
1106 vpx_highbd_sad16x32_avg_bits8,
1107 vpx_highbd_8_variance16x32,
1108 vpx_highbd_8_sub_pixel_variance16x32,
1109 vpx_highbd_8_sub_pixel_avg_variance16x32,
1112 vpx_highbd_sad16x32x4d_bits8)
1114 HIGHBD_BFP(BLOCK_64X32,
1115 vpx_highbd_sad64x32_bits8,
1116 vpx_highbd_sad64x32_avg_bits8,
1117 vpx_highbd_8_variance64x32,
1118 vpx_highbd_8_sub_pixel_variance64x32,
1119 vpx_highbd_8_sub_pixel_avg_variance64x32,
1122 vpx_highbd_sad64x32x4d_bits8)
1124 HIGHBD_BFP(BLOCK_32X64,
1125 vpx_highbd_sad32x64_bits8,
1126 vpx_highbd_sad32x64_avg_bits8,
1127 vpx_highbd_8_variance32x64,
1128 vpx_highbd_8_sub_pixel_variance32x64,
1129 vpx_highbd_8_sub_pixel_avg_variance32x64,
1132 vpx_highbd_sad32x64x4d_bits8)
1134 HIGHBD_BFP(BLOCK_32X32,
1135 vpx_highbd_sad32x32_bits8,
1136 vpx_highbd_sad32x32_avg_bits8,
1137 vpx_highbd_8_variance32x32,
1138 vpx_highbd_8_sub_pixel_variance32x32,
1139 vpx_highbd_8_sub_pixel_avg_variance32x32,
1140 vpx_highbd_sad32x32x3_bits8,
1141 vpx_highbd_sad32x32x8_bits8,
1142 vpx_highbd_sad32x32x4d_bits8)
1144 HIGHBD_BFP(BLOCK_64X64,
1145 vpx_highbd_sad64x64_bits8,
1146 vpx_highbd_sad64x64_avg_bits8,
1147 vpx_highbd_8_variance64x64,
1148 vpx_highbd_8_sub_pixel_variance64x64,
1149 vpx_highbd_8_sub_pixel_avg_variance64x64,
1150 vpx_highbd_sad64x64x3_bits8,
1151 vpx_highbd_sad64x64x8_bits8,
1152 vpx_highbd_sad64x64x4d_bits8)
1154 HIGHBD_BFP(BLOCK_16X16,
1155 vpx_highbd_sad16x16_bits8,
1156 vpx_highbd_sad16x16_avg_bits8,
1157 vpx_highbd_8_variance16x16,
1158 vpx_highbd_8_sub_pixel_variance16x16,
1159 vpx_highbd_8_sub_pixel_avg_variance16x16,
1160 vpx_highbd_sad16x16x3_bits8,
1161 vpx_highbd_sad16x16x8_bits8,
1162 vpx_highbd_sad16x16x4d_bits8)
1164 HIGHBD_BFP(BLOCK_16X8,
1165 vpx_highbd_sad16x8_bits8,
1166 vpx_highbd_sad16x8_avg_bits8,
1167 vpx_highbd_8_variance16x8,
1168 vpx_highbd_8_sub_pixel_variance16x8,
1169 vpx_highbd_8_sub_pixel_avg_variance16x8,
1170 vpx_highbd_sad16x8x3_bits8,
1171 vpx_highbd_sad16x8x8_bits8,
1172 vpx_highbd_sad16x8x4d_bits8)
1174 HIGHBD_BFP(BLOCK_8X16,
1175 vpx_highbd_sad8x16_bits8,
1176 vpx_highbd_sad8x16_avg_bits8,
1177 vpx_highbd_8_variance8x16,
1178 vpx_highbd_8_sub_pixel_variance8x16,
1179 vpx_highbd_8_sub_pixel_avg_variance8x16,
1180 vpx_highbd_sad8x16x3_bits8,
1181 vpx_highbd_sad8x16x8_bits8,
1182 vpx_highbd_sad8x16x4d_bits8)
1184 HIGHBD_BFP(BLOCK_8X8,
1185 vpx_highbd_sad8x8_bits8,
1186 vpx_highbd_sad8x8_avg_bits8,
1187 vpx_highbd_8_variance8x8,
1188 vpx_highbd_8_sub_pixel_variance8x8,
1189 vpx_highbd_8_sub_pixel_avg_variance8x8,
1190 vpx_highbd_sad8x8x3_bits8,
1191 vpx_highbd_sad8x8x8_bits8,
1192 vpx_highbd_sad8x8x4d_bits8)
1194 HIGHBD_BFP(BLOCK_8X4,
1195 vpx_highbd_sad8x4_bits8,
1196 vpx_highbd_sad8x4_avg_bits8,
1197 vpx_highbd_8_variance8x4,
1198 vpx_highbd_8_sub_pixel_variance8x4,
1199 vpx_highbd_8_sub_pixel_avg_variance8x4,
1201 vpx_highbd_sad8x4x8_bits8,
1202 vpx_highbd_sad8x4x4d_bits8)
1204 HIGHBD_BFP(BLOCK_4X8,
1205 vpx_highbd_sad4x8_bits8,
1206 vpx_highbd_sad4x8_avg_bits8,
1207 vpx_highbd_8_variance4x8,
1208 vpx_highbd_8_sub_pixel_variance4x8,
1209 vpx_highbd_8_sub_pixel_avg_variance4x8,
1211 vpx_highbd_sad4x8x8_bits8,
1212 vpx_highbd_sad4x8x4d_bits8)
1214 HIGHBD_BFP(BLOCK_4X4,
1215 vpx_highbd_sad4x4_bits8,
1216 vpx_highbd_sad4x4_avg_bits8,
1217 vpx_highbd_8_variance4x4,
1218 vpx_highbd_8_sub_pixel_variance4x4,
1219 vpx_highbd_8_sub_pixel_avg_variance4x4,
1220 vpx_highbd_sad4x4x3_bits8,
1221 vpx_highbd_sad4x4x8_bits8,
1222 vpx_highbd_sad4x4x4d_bits8)
1226 HIGHBD_BFP(BLOCK_32X16,
1227 vpx_highbd_sad32x16_bits10,
1228 vpx_highbd_sad32x16_avg_bits10,
1229 vpx_highbd_10_variance32x16,
1230 vpx_highbd_10_sub_pixel_variance32x16,
1231 vpx_highbd_10_sub_pixel_avg_variance32x16,
1234 vpx_highbd_sad32x16x4d_bits10)
1236 HIGHBD_BFP(BLOCK_16X32,
1237 vpx_highbd_sad16x32_bits10,
1238 vpx_highbd_sad16x32_avg_bits10,
1239 vpx_highbd_10_variance16x32,
1240 vpx_highbd_10_sub_pixel_variance16x32,
1241 vpx_highbd_10_sub_pixel_avg_variance16x32,
1244 vpx_highbd_sad16x32x4d_bits10)
1246 HIGHBD_BFP(BLOCK_64X32,
1247 vpx_highbd_sad64x32_bits10,
1248 vpx_highbd_sad64x32_avg_bits10,
1249 vpx_highbd_10_variance64x32,
1250 vpx_highbd_10_sub_pixel_variance64x32,
1251 vpx_highbd_10_sub_pixel_avg_variance64x32,
1254 vpx_highbd_sad64x32x4d_bits10)
1256 HIGHBD_BFP(BLOCK_32X64,
1257 vpx_highbd_sad32x64_bits10,
1258 vpx_highbd_sad32x64_avg_bits10,
1259 vpx_highbd_10_variance32x64,
1260 vpx_highbd_10_sub_pixel_variance32x64,
1261 vpx_highbd_10_sub_pixel_avg_variance32x64,
1264 vpx_highbd_sad32x64x4d_bits10)
1266 HIGHBD_BFP(BLOCK_32X32,
1267 vpx_highbd_sad32x32_bits10,
1268 vpx_highbd_sad32x32_avg_bits10,
1269 vpx_highbd_10_variance32x32,
1270 vpx_highbd_10_sub_pixel_variance32x32,
1271 vpx_highbd_10_sub_pixel_avg_variance32x32,
1272 vpx_highbd_sad32x32x3_bits10,
1273 vpx_highbd_sad32x32x8_bits10,
1274 vpx_highbd_sad32x32x4d_bits10)
1276 HIGHBD_BFP(BLOCK_64X64,
1277 vpx_highbd_sad64x64_bits10,
1278 vpx_highbd_sad64x64_avg_bits10,
1279 vpx_highbd_10_variance64x64,
1280 vpx_highbd_10_sub_pixel_variance64x64,
1281 vpx_highbd_10_sub_pixel_avg_variance64x64,
1282 vpx_highbd_sad64x64x3_bits10,
1283 vpx_highbd_sad64x64x8_bits10,
1284 vpx_highbd_sad64x64x4d_bits10)
1286 HIGHBD_BFP(BLOCK_16X16,
1287 vpx_highbd_sad16x16_bits10,
1288 vpx_highbd_sad16x16_avg_bits10,
1289 vpx_highbd_10_variance16x16,
1290 vpx_highbd_10_sub_pixel_variance16x16,
1291 vpx_highbd_10_sub_pixel_avg_variance16x16,
1292 vpx_highbd_sad16x16x3_bits10,
1293 vpx_highbd_sad16x16x8_bits10,
1294 vpx_highbd_sad16x16x4d_bits10)
1296 HIGHBD_BFP(BLOCK_16X8,
1297 vpx_highbd_sad16x8_bits10,
1298 vpx_highbd_sad16x8_avg_bits10,
1299 vpx_highbd_10_variance16x8,
1300 vpx_highbd_10_sub_pixel_variance16x8,
1301 vpx_highbd_10_sub_pixel_avg_variance16x8,
1302 vpx_highbd_sad16x8x3_bits10,
1303 vpx_highbd_sad16x8x8_bits10,
1304 vpx_highbd_sad16x8x4d_bits10)
1306 HIGHBD_BFP(BLOCK_8X16,
1307 vpx_highbd_sad8x16_bits10,
1308 vpx_highbd_sad8x16_avg_bits10,
1309 vpx_highbd_10_variance8x16,
1310 vpx_highbd_10_sub_pixel_variance8x16,
1311 vpx_highbd_10_sub_pixel_avg_variance8x16,
1312 vpx_highbd_sad8x16x3_bits10,
1313 vpx_highbd_sad8x16x8_bits10,
1314 vpx_highbd_sad8x16x4d_bits10)
1316 HIGHBD_BFP(BLOCK_8X8,
1317 vpx_highbd_sad8x8_bits10,
1318 vpx_highbd_sad8x8_avg_bits10,
1319 vpx_highbd_10_variance8x8,
1320 vpx_highbd_10_sub_pixel_variance8x8,
1321 vpx_highbd_10_sub_pixel_avg_variance8x8,
1322 vpx_highbd_sad8x8x3_bits10,
1323 vpx_highbd_sad8x8x8_bits10,
1324 vpx_highbd_sad8x8x4d_bits10)
1326 HIGHBD_BFP(BLOCK_8X4,
1327 vpx_highbd_sad8x4_bits10,
1328 vpx_highbd_sad8x4_avg_bits10,
1329 vpx_highbd_10_variance8x4,
1330 vpx_highbd_10_sub_pixel_variance8x4,
1331 vpx_highbd_10_sub_pixel_avg_variance8x4,
1333 vpx_highbd_sad8x4x8_bits10,
1334 vpx_highbd_sad8x4x4d_bits10)
1336 HIGHBD_BFP(BLOCK_4X8,
1337 vpx_highbd_sad4x8_bits10,
1338 vpx_highbd_sad4x8_avg_bits10,
1339 vpx_highbd_10_variance4x8,
1340 vpx_highbd_10_sub_pixel_variance4x8,
1341 vpx_highbd_10_sub_pixel_avg_variance4x8,
1343 vpx_highbd_sad4x8x8_bits10,
1344 vpx_highbd_sad4x8x4d_bits10)
1346 HIGHBD_BFP(BLOCK_4X4,
1347 vpx_highbd_sad4x4_bits10,
1348 vpx_highbd_sad4x4_avg_bits10,
1349 vpx_highbd_10_variance4x4,
1350 vpx_highbd_10_sub_pixel_variance4x4,
1351 vpx_highbd_10_sub_pixel_avg_variance4x4,
1352 vpx_highbd_sad4x4x3_bits10,
1353 vpx_highbd_sad4x4x8_bits10,
1354 vpx_highbd_sad4x4x4d_bits10)
1358 HIGHBD_BFP(BLOCK_32X16,
1359 vpx_highbd_sad32x16_bits12,
1360 vpx_highbd_sad32x16_avg_bits12,
1361 vpx_highbd_12_variance32x16,
1362 vpx_highbd_12_sub_pixel_variance32x16,
1363 vpx_highbd_12_sub_pixel_avg_variance32x16,
1366 vpx_highbd_sad32x16x4d_bits12)
1368 HIGHBD_BFP(BLOCK_16X32,
1369 vpx_highbd_sad16x32_bits12,
1370 vpx_highbd_sad16x32_avg_bits12,
1371 vpx_highbd_12_variance16x32,
1372 vpx_highbd_12_sub_pixel_variance16x32,
1373 vpx_highbd_12_sub_pixel_avg_variance16x32,
1376 vpx_highbd_sad16x32x4d_bits12)
1378 HIGHBD_BFP(BLOCK_64X32,
1379 vpx_highbd_sad64x32_bits12,
1380 vpx_highbd_sad64x32_avg_bits12,
1381 vpx_highbd_12_variance64x32,
1382 vpx_highbd_12_sub_pixel_variance64x32,
1383 vpx_highbd_12_sub_pixel_avg_variance64x32,
1386 vpx_highbd_sad64x32x4d_bits12)
1388 HIGHBD_BFP(BLOCK_32X64,
1389 vpx_highbd_sad32x64_bits12,
1390 vpx_highbd_sad32x64_avg_bits12,
1391 vpx_highbd_12_variance32x64,
1392 vpx_highbd_12_sub_pixel_variance32x64,
1393 vpx_highbd_12_sub_pixel_avg_variance32x64,
1396 vpx_highbd_sad32x64x4d_bits12)
1398 HIGHBD_BFP(BLOCK_32X32,
1399 vpx_highbd_sad32x32_bits12,
1400 vpx_highbd_sad32x32_avg_bits12,
1401 vpx_highbd_12_variance32x32,
1402 vpx_highbd_12_sub_pixel_variance32x32,
1403 vpx_highbd_12_sub_pixel_avg_variance32x32,
1404 vpx_highbd_sad32x32x3_bits12,
1405 vpx_highbd_sad32x32x8_bits12,
1406 vpx_highbd_sad32x32x4d_bits12)
1408 HIGHBD_BFP(BLOCK_64X64,
1409 vpx_highbd_sad64x64_bits12,
1410 vpx_highbd_sad64x64_avg_bits12,
1411 vpx_highbd_12_variance64x64,
1412 vpx_highbd_12_sub_pixel_variance64x64,
1413 vpx_highbd_12_sub_pixel_avg_variance64x64,
1414 vpx_highbd_sad64x64x3_bits12,
1415 vpx_highbd_sad64x64x8_bits12,
1416 vpx_highbd_sad64x64x4d_bits12)
1418 HIGHBD_BFP(BLOCK_16X16,
1419 vpx_highbd_sad16x16_bits12,
1420 vpx_highbd_sad16x16_avg_bits12,
1421 vpx_highbd_12_variance16x16,
1422 vpx_highbd_12_sub_pixel_variance16x16,
1423 vpx_highbd_12_sub_pixel_avg_variance16x16,
1424 vpx_highbd_sad16x16x3_bits12,
1425 vpx_highbd_sad16x16x8_bits12,
1426 vpx_highbd_sad16x16x4d_bits12)
1428 HIGHBD_BFP(BLOCK_16X8,
1429 vpx_highbd_sad16x8_bits12,
1430 vpx_highbd_sad16x8_avg_bits12,
1431 vpx_highbd_12_variance16x8,
1432 vpx_highbd_12_sub_pixel_variance16x8,
1433 vpx_highbd_12_sub_pixel_avg_variance16x8,
1434 vpx_highbd_sad16x8x3_bits12,
1435 vpx_highbd_sad16x8x8_bits12,
1436 vpx_highbd_sad16x8x4d_bits12)
1438 HIGHBD_BFP(BLOCK_8X16,
1439 vpx_highbd_sad8x16_bits12,
1440 vpx_highbd_sad8x16_avg_bits12,
1441 vpx_highbd_12_variance8x16,
1442 vpx_highbd_12_sub_pixel_variance8x16,
1443 vpx_highbd_12_sub_pixel_avg_variance8x16,
1444 vpx_highbd_sad8x16x3_bits12,
1445 vpx_highbd_sad8x16x8_bits12,
1446 vpx_highbd_sad8x16x4d_bits12)
1448 HIGHBD_BFP(BLOCK_8X8,
1449 vpx_highbd_sad8x8_bits12,
1450 vpx_highbd_sad8x8_avg_bits12,
1451 vpx_highbd_12_variance8x8,
1452 vpx_highbd_12_sub_pixel_variance8x8,
1453 vpx_highbd_12_sub_pixel_avg_variance8x8,
1454 vpx_highbd_sad8x8x3_bits12,
1455 vpx_highbd_sad8x8x8_bits12,
1456 vpx_highbd_sad8x8x4d_bits12)
1458 HIGHBD_BFP(BLOCK_8X4,
1459 vpx_highbd_sad8x4_bits12,
1460 vpx_highbd_sad8x4_avg_bits12,
1461 vpx_highbd_12_variance8x4,
1462 vpx_highbd_12_sub_pixel_variance8x4,
1463 vpx_highbd_12_sub_pixel_avg_variance8x4,
1465 vpx_highbd_sad8x4x8_bits12,
1466 vpx_highbd_sad8x4x4d_bits12)
1468 HIGHBD_BFP(BLOCK_4X8,
1469 vpx_highbd_sad4x8_bits12,
1470 vpx_highbd_sad4x8_avg_bits12,
1471 vpx_highbd_12_variance4x8,
1472 vpx_highbd_12_sub_pixel_variance4x8,
1473 vpx_highbd_12_sub_pixel_avg_variance4x8,
1475 vpx_highbd_sad4x8x8_bits12,
1476 vpx_highbd_sad4x8x4d_bits12)
1478 HIGHBD_BFP(BLOCK_4X4,
1479 vpx_highbd_sad4x4_bits12,
1480 vpx_highbd_sad4x4_avg_bits12,
1481 vpx_highbd_12_variance4x4,
1482 vpx_highbd_12_sub_pixel_variance4x4,
1483 vpx_highbd_12_sub_pixel_avg_variance4x4,
1484 vpx_highbd_sad4x4x3_bits12,
1485 vpx_highbd_sad4x4x8_bits12,
1486 vpx_highbd_sad4x4x4d_bits12)
1490 assert(0 && "cm->bit_depth should be VPX_BITS_8, "
1491 "VPX_BITS_10 or VPX_BITS_12");
1495 #endif // CONFIG_VP9_HIGHBITDEPTH
1497 static void realloc_segmentation_maps(VP9_COMP *cpi) {
1498 VP9_COMMON *const cm = &cpi->common;
1500 // Create the encoder segmentation map and set all entries to 0
1501 vpx_free(cpi->segmentation_map);
1502 CHECK_MEM_ERROR(cm, cpi->segmentation_map,
1503 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1505 // Create a map used for cyclic background refresh.
1506 if (cpi->cyclic_refresh)
1507 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
1508 CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
1509 vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
1511 // Create a map used to mark inactive areas.
1512 vpx_free(cpi->active_map.map);
1513 CHECK_MEM_ERROR(cm, cpi->active_map.map,
1514 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1516 // And a place holder structure is the coding context
1517 // for use if we want to save and restore it
1518 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
1519 CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
1520 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1523 void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
1524 VP9_COMMON *const cm = &cpi->common;
1525 RATE_CONTROL *const rc = &cpi->rc;
1526 int last_w = cpi->oxcf.width;
1527 int last_h = cpi->oxcf.height;
1529 if (cm->profile != oxcf->profile)
1530 cm->profile = oxcf->profile;
1531 cm->bit_depth = oxcf->bit_depth;
1532 cm->color_space = oxcf->color_space;
1533 cm->color_range = oxcf->color_range;
1535 cpi->target_level = oxcf->target_level;
1536 cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
1538 if (cm->profile <= PROFILE_1)
1539 assert(cm->bit_depth == VPX_BITS_8);
1541 assert(cm->bit_depth > VPX_BITS_8);
1544 #if CONFIG_VP9_HIGHBITDEPTH
1545 cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
1546 #endif // CONFIG_VP9_HIGHBITDEPTH
1548 if ((oxcf->pass == 0) && (oxcf->rc_mode == VPX_Q)) {
1549 rc->baseline_gf_interval = FIXED_GF_INTERVAL;
1551 rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
1554 cpi->refresh_golden_frame = 0;
1555 cpi->refresh_last_frame = 1;
1556 cm->refresh_frame_context = 1;
1557 cm->reset_frame_context = 0;
1559 vp9_reset_segment_features(&cm->seg);
1560 vp9_set_high_precision_mv(cpi, 0);
1565 for (i = 0; i < MAX_SEGMENTS; i++)
1566 cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
1568 cpi->encode_breakout = cpi->oxcf.encode_breakout;
1570 set_rc_buffer_sizes(rc, &cpi->oxcf);
1572 // Under a configuration change, where maximum_buffer_size may change,
1573 // keep buffer level clipped to the maximum allowed buffer size.
1574 rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
1575 rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size);
1577 // Set up frame rate and related parameters rate control values.
1578 vp9_new_framerate(cpi, cpi->framerate);
1580 // Set absolute upper and lower quality limits
1581 rc->worst_quality = cpi->oxcf.worst_allowed_q;
1582 rc->best_quality = cpi->oxcf.best_allowed_q;
1584 cm->interp_filter = cpi->sf.default_interp_filter;
1586 if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) {
1587 cm->render_width = cpi->oxcf.render_width;
1588 cm->render_height = cpi->oxcf.render_height;
1590 cm->render_width = cpi->oxcf.width;
1591 cm->render_height = cpi->oxcf.height;
1593 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1594 cm->width = cpi->oxcf.width;
1595 cm->height = cpi->oxcf.height;
1596 cpi->external_resize = 1;
1599 if (cpi->initial_width) {
1600 int new_mi_size = 0;
1601 vp9_set_mb_mi(cm, cm->width, cm->height);
1602 new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
1603 if (cm->mi_alloc_size < new_mi_size) {
1604 vp9_free_context_buffers(cm);
1605 alloc_compressor_data(cpi);
1606 realloc_segmentation_maps(cpi);
1607 cpi->initial_width = cpi->initial_height = 0;
1608 cpi->external_resize = 0;
1609 } else if (cm->mi_alloc_size == new_mi_size &&
1610 (cpi->oxcf.width > last_w || cpi->oxcf.height > last_h)) {
1611 vp9_alloc_loop_filter(cm);
1615 update_frame_size(cpi);
1617 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1618 memset(cpi->consec_zero_mv, 0,
1619 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
1620 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
1621 vp9_cyclic_refresh_reset_resize(cpi);
1624 if ((cpi->svc.number_temporal_layers > 1 &&
1625 cpi->oxcf.rc_mode == VPX_CBR) ||
1626 ((cpi->svc.number_temporal_layers > 1 ||
1627 cpi->svc.number_spatial_layers > 1) &&
1628 cpi->oxcf.pass != 1)) {
1629 vp9_update_layer_context_change_config(cpi,
1630 (int)cpi->oxcf.target_bandwidth);
1633 cpi->alt_ref_source = NULL;
1634 rc->is_src_frame_alt_ref = 0;
1637 // Experimental RD Code
1638 cpi->frame_distortion = 0;
1639 cpi->last_frame_distortion = 0;
1642 set_tile_limits(cpi);
1644 cpi->ext_refresh_frame_flags_pending = 0;
1645 cpi->ext_refresh_frame_context_pending = 0;
1647 #if CONFIG_VP9_HIGHBITDEPTH
1648 highbd_set_var_fns(cpi);
1653 #define M_LOG2_E 0.693147180559945309417
1655 #define log2f(x) (log (x) / (float) M_LOG2_E)
1657 /***********************************************************************
1658 * Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts' *
1659 ***********************************************************************
1660 * The following 2 functions ('cal_nmvjointsadcost' and *
1661 * 'cal_nmvsadcosts') are used to calculate cost lookup tables *
1662 * used by 'vp9_diamond_search_sad'. The C implementation of the *
1663 * function is generic, but the AVX intrinsics optimised version *
1664 * relies on the following properties of the computed tables: *
1665 * For cal_nmvjointsadcost: *
1666 * - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] *
1667 * For cal_nmvsadcosts: *
1668 * - For all i: mvsadcost[0][i] == mvsadcost[1][i] *
1669 * (Equal costs for both components) *
1670 * - For all i: mvsadcost[0][i] == mvsadcost[0][-i] *
1671 * (Cost function is even) *
1672 * If these do not hold, then the AVX optimised version of the *
1673 * 'vp9_diamond_search_sad' function cannot be used as it is, in which *
1674 * case you can revert to using the C function instead. *
1675 ***********************************************************************/
1677 static void cal_nmvjointsadcost(int *mvjointsadcost) {
1678 /*********************************************************************
1679 * Warning: Read the comments above before modifying this function *
1680 *********************************************************************/
1681 mvjointsadcost[0] = 600;
1682 mvjointsadcost[1] = 300;
1683 mvjointsadcost[2] = 300;
1684 mvjointsadcost[3] = 300;
1687 static void cal_nmvsadcosts(int *mvsadcost[2]) {
1688 /*********************************************************************
1689 * Warning: Read the comments above before modifying this function *
1690 *********************************************************************/
1693 mvsadcost[0][0] = 0;
1694 mvsadcost[1][0] = 0;
1697 double z = 256 * (2 * (log2f(8 * i) + .6));
1698 mvsadcost[0][i] = (int)z;
1699 mvsadcost[1][i] = (int)z;
1700 mvsadcost[0][-i] = (int)z;
1701 mvsadcost[1][-i] = (int)z;
1702 } while (++i <= MV_MAX);
1705 static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
1708 mvsadcost[0][0] = 0;
1709 mvsadcost[1][0] = 0;
1712 double z = 256 * (2 * (log2f(8 * i) + .6));
1713 mvsadcost[0][i] = (int)z;
1714 mvsadcost[1][i] = (int)z;
1715 mvsadcost[0][-i] = (int)z;
1716 mvsadcost[1][-i] = (int)z;
1717 } while (++i <= MV_MAX);
1720 VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
1721 BufferPool *const pool) {
1723 VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
1724 VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
1731 if (setjmp(cm->error.jmp)) {
1732 cm->error.setjmp = 0;
1733 vp9_remove_compressor(cpi);
1737 cm->error.setjmp = 1;
1738 cm->alloc_mi = vp9_enc_alloc_mi;
1739 cm->free_mi = vp9_enc_free_mi;
1740 cm->setup_mi = vp9_enc_setup_mi;
1742 CHECK_MEM_ERROR(cm, cm->fc,
1743 (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
1744 CHECK_MEM_ERROR(cm, cm->frame_contexts,
1745 (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
1746 sizeof(*cm->frame_contexts)));
1749 cpi->resize_state = 0;
1750 cpi->external_resize = 0;
1751 cpi->resize_avg_qp = 0;
1752 cpi->resize_buffer_underflow = 0;
1753 cpi->use_skin_detection = 0;
1754 cpi->common.buffer_pool = pool;
1756 init_config(cpi, oxcf);
1757 vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
1759 cm->current_video_frame = 0;
1760 cpi->partition_search_skippable_frame = 0;
1761 cpi->tile_data = NULL;
1763 realloc_segmentation_maps(cpi);
1765 CHECK_MEM_ERROR(cm, cpi->consec_zero_mv,
1766 vpx_calloc(cm->mi_rows * cm->mi_cols,
1767 sizeof(*cpi->consec_zero_mv)));
1769 CHECK_MEM_ERROR(cm, cpi->nmvcosts[0],
1770 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0])));
1771 CHECK_MEM_ERROR(cm, cpi->nmvcosts[1],
1772 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1])));
1773 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0],
1774 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0])));
1775 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1],
1776 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1])));
1777 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0],
1778 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0])));
1779 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1],
1780 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1])));
1781 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0],
1782 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0])));
1783 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
1784 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
1786 for (i = 0; i < (sizeof(cpi->mbgraph_stats) /
1787 sizeof(cpi->mbgraph_stats[0])); i++) {
1788 CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats,
1789 vpx_calloc(cm->MBs *
1790 sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
1793 #if CONFIG_FP_MB_STATS
1794 cpi->use_fp_mb_stats = 0;
1795 if (cpi->use_fp_mb_stats) {
1796 // a place holder used to store the first pass mb stats in the first pass
1797 CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
1798 vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
1800 cpi->twopass.frame_mb_stats_buf = NULL;
1804 cpi->refresh_alt_ref_frame = 0;
1805 cpi->multi_arf_last_grp_enabled = 0;
1807 cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
1809 init_level_info(&cpi->level_info);
1811 #if CONFIG_INTERNAL_STATS
1812 cpi->b_calculate_ssimg = 0;
1813 cpi->b_calculate_blockiness = 1;
1814 cpi->b_calculate_consistency = 1;
1815 cpi->total_inconsistency = 0;
1816 cpi->psnr.worst = 100.0;
1817 cpi->worst_ssim = 100.0;
1822 if (cpi->b_calculate_psnr) {
1823 cpi->total_sq_error = 0;
1824 cpi->total_samples = 0;
1826 cpi->totalp_sq_error = 0;
1827 cpi->totalp_samples = 0;
1829 cpi->tot_recode_hits = 0;
1830 cpi->summed_quality = 0;
1831 cpi->summed_weights = 0;
1832 cpi->summedp_quality = 0;
1833 cpi->summedp_weights = 0;
1836 if (cpi->b_calculate_ssimg) {
1837 cpi->ssimg.worst= 100.0;
1839 cpi->fastssim.worst = 100.0;
1841 cpi->psnrhvs.worst = 100.0;
1843 if (cpi->b_calculate_blockiness) {
1844 cpi->total_blockiness = 0;
1845 cpi->worst_blockiness = 0.0;
1848 if (cpi->b_calculate_consistency) {
1849 CHECK_MEM_ERROR(cm, cpi->ssim_vars,
1850 vpx_malloc(sizeof(*cpi->ssim_vars) * 4 *
1851 cpi->common.mi_rows * cpi->common.mi_cols));
1852 cpi->worst_consistency = 100.0;
1857 cpi->first_time_stamp_ever = INT64_MAX;
1859 /*********************************************************************
1860 * Warning: Read the comments around 'cal_nmvjointsadcost' and *
1861 * 'cal_nmvsadcosts' before modifying how these tables are computed. *
1862 *********************************************************************/
1863 cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
1864 cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
1865 cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
1866 cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
1867 cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
1868 cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
1870 cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
1871 cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
1872 cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
1873 cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
1874 cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
1876 #if CONFIG_VP9_TEMPORAL_DENOISING
1877 #ifdef OUTPUT_YUV_DENOISED
1878 yuv_denoised_file = fopen("denoised.yuv", "ab");
1881 #ifdef OUTPUT_YUV_SKINMAP
1882 yuv_skinmap_file = fopen("skinmap.yuv", "ab");
1884 #ifdef OUTPUT_YUV_REC
1885 yuv_rec_file = fopen("rec.yuv", "wb");
1889 framepsnr = fopen("framepsnr.stt", "a");
1890 kf_list = fopen("kf_list.stt", "w");
1893 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
1895 if (oxcf->pass == 1) {
1896 vp9_init_first_pass(cpi);
1897 } else if (oxcf->pass == 2) {
1898 const size_t packet_sz = sizeof(FIRSTPASS_STATS);
1899 const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
1901 if (cpi->svc.number_spatial_layers > 1
1902 || cpi->svc.number_temporal_layers > 1) {
1903 FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf;
1904 FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = {0};
1907 for (i = 0; i < oxcf->ss_number_layers; ++i) {
1908 FIRSTPASS_STATS *const last_packet_for_layer =
1909 &stats[packets - oxcf->ss_number_layers + i];
1910 const int layer_id = (int)last_packet_for_layer->spatial_layer_id;
1911 const int packets_in_layer = (int)last_packet_for_layer->count + 1;
1912 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) {
1913 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id];
1915 vpx_free(lc->rc_twopass_stats_in.buf);
1917 lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz;
1918 CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf,
1919 vpx_malloc(lc->rc_twopass_stats_in.sz));
1920 lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf;
1921 lc->twopass.stats_in = lc->twopass.stats_in_start;
1922 lc->twopass.stats_in_end = lc->twopass.stats_in_start
1923 + packets_in_layer - 1;
1924 stats_copy[layer_id] = lc->rc_twopass_stats_in.buf;
1928 for (i = 0; i < packets; ++i) {
1929 const int layer_id = (int)stats[i].spatial_layer_id;
1930 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers
1931 && stats_copy[layer_id] != NULL) {
1932 *stats_copy[layer_id] = stats[i];
1933 ++stats_copy[layer_id];
1937 vp9_init_second_pass_spatial_svc(cpi);
1939 #if CONFIG_FP_MB_STATS
1940 if (cpi->use_fp_mb_stats) {
1941 const size_t psz = cpi->common.MBs * sizeof(uint8_t);
1942 const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
1944 cpi->twopass.firstpass_mb_stats.mb_stats_start =
1945 oxcf->firstpass_mb_stats_in.buf;
1946 cpi->twopass.firstpass_mb_stats.mb_stats_end =
1947 cpi->twopass.firstpass_mb_stats.mb_stats_start +
1948 (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
1952 cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
1953 cpi->twopass.stats_in = cpi->twopass.stats_in_start;
1954 cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
1956 vp9_init_second_pass(cpi);
1960 vp9_set_speed_features_framesize_independent(cpi);
1961 vp9_set_speed_features_framesize_dependent(cpi);
1963 // Allocate memory to store variances for a frame.
1964 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
1965 vpx_calloc(cm->MBs, sizeof(diff)));
1966 cpi->source_var_thresh = 0;
1967 cpi->frames_till_next_var_check = 0;
1969 #define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\
1970 cpi->fn_ptr[BT].sdf = SDF; \
1971 cpi->fn_ptr[BT].sdaf = SDAF; \
1972 cpi->fn_ptr[BT].vf = VF; \
1973 cpi->fn_ptr[BT].svf = SVF; \
1974 cpi->fn_ptr[BT].svaf = SVAF; \
1975 cpi->fn_ptr[BT].sdx3f = SDX3F; \
1976 cpi->fn_ptr[BT].sdx8f = SDX8F; \
1977 cpi->fn_ptr[BT].sdx4df = SDX4DF;
1979 BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg,
1980 vpx_variance32x16, vpx_sub_pixel_variance32x16,
1981 vpx_sub_pixel_avg_variance32x16, NULL, NULL, vpx_sad32x16x4d)
1983 BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg,
1984 vpx_variance16x32, vpx_sub_pixel_variance16x32,
1985 vpx_sub_pixel_avg_variance16x32, NULL, NULL, vpx_sad16x32x4d)
1987 BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg,
1988 vpx_variance64x32, vpx_sub_pixel_variance64x32,
1989 vpx_sub_pixel_avg_variance64x32, NULL, NULL, vpx_sad64x32x4d)
1991 BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg,
1992 vpx_variance32x64, vpx_sub_pixel_variance32x64,
1993 vpx_sub_pixel_avg_variance32x64, NULL, NULL, vpx_sad32x64x4d)
1995 BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg,
1996 vpx_variance32x32, vpx_sub_pixel_variance32x32,
1997 vpx_sub_pixel_avg_variance32x32, vpx_sad32x32x3, vpx_sad32x32x8,
2000 BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg,
2001 vpx_variance64x64, vpx_sub_pixel_variance64x64,
2002 vpx_sub_pixel_avg_variance64x64, vpx_sad64x64x3, vpx_sad64x64x8,
2005 BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg,
2006 vpx_variance16x16, vpx_sub_pixel_variance16x16,
2007 vpx_sub_pixel_avg_variance16x16, vpx_sad16x16x3, vpx_sad16x16x8,
2010 BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg,
2011 vpx_variance16x8, vpx_sub_pixel_variance16x8,
2012 vpx_sub_pixel_avg_variance16x8,
2013 vpx_sad16x8x3, vpx_sad16x8x8, vpx_sad16x8x4d)
2015 BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg,
2016 vpx_variance8x16, vpx_sub_pixel_variance8x16,
2017 vpx_sub_pixel_avg_variance8x16,
2018 vpx_sad8x16x3, vpx_sad8x16x8, vpx_sad8x16x4d)
2020 BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg,
2021 vpx_variance8x8, vpx_sub_pixel_variance8x8,
2022 vpx_sub_pixel_avg_variance8x8,
2023 vpx_sad8x8x3, vpx_sad8x8x8, vpx_sad8x8x4d)
2025 BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg,
2026 vpx_variance8x4, vpx_sub_pixel_variance8x4,
2027 vpx_sub_pixel_avg_variance8x4, NULL, vpx_sad8x4x8, vpx_sad8x4x4d)
2029 BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg,
2030 vpx_variance4x8, vpx_sub_pixel_variance4x8,
2031 vpx_sub_pixel_avg_variance4x8, NULL, vpx_sad4x8x8, vpx_sad4x8x4d)
2033 BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg,
2034 vpx_variance4x4, vpx_sub_pixel_variance4x4,
2035 vpx_sub_pixel_avg_variance4x4,
2036 vpx_sad4x4x3, vpx_sad4x4x8, vpx_sad4x4x4d)
2038 #if CONFIG_VP9_HIGHBITDEPTH
2039 highbd_set_var_fns(cpi);
2042 /* vp9_init_quantizer() is first called here. Add check in
2043 * vp9_frame_init_quantizer() so that vp9_init_quantizer is only
2044 * called later when needed. This will avoid unnecessary calls of
2045 * vp9_init_quantizer() for every frame.
2047 vp9_init_quantizer(cpi);
2049 vp9_loop_filter_init(cm);
2051 cm->error.setjmp = 0;
2056 #if CONFIG_INTERNAL_STATS
2057 #define SNPRINT(H, T) \
2058 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
2060 #define SNPRINT2(H, T, V) \
2061 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
2062 #endif // CONFIG_INTERNAL_STATS
2064 void vp9_remove_compressor(VP9_COMP *cpi) {
2073 if (cm->current_video_frame > 0) {
2074 #if CONFIG_INTERNAL_STATS
2075 vpx_clear_system_state();
2077 if (cpi->oxcf.pass != 1) {
2078 char headings[512] = {0};
2079 char results[512] = {0};
2080 FILE *f = fopen("opsnr.stt", "a");
2081 double time_encoded = (cpi->last_end_time_stamp_seen
2082 - cpi->first_time_stamp_ever) / 10000000.000;
2083 double total_encode_time = (cpi->time_receive_data +
2084 cpi->time_compress_data) / 1000.000;
2086 (double)cpi->bytes * (double) 8 / (double)1000 / time_encoded;
2087 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
2088 const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
2089 const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
2091 if (cpi->b_calculate_psnr) {
2092 const double total_psnr =
2093 vpx_sse_to_psnr((double)cpi->total_samples, peak,
2094 (double)cpi->total_sq_error);
2095 const double totalp_psnr =
2096 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
2097 (double)cpi->totalp_sq_error);
2098 const double total_ssim = 100 * pow(cpi->summed_quality /
2099 cpi->summed_weights, 8.0);
2100 const double totalp_ssim = 100 * pow(cpi->summedp_quality /
2101 cpi->summedp_weights, 8.0);
2103 snprintf(headings, sizeof(headings),
2104 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
2105 "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
2106 "WstPsnr\tWstSsim\tWstFast\tWstHVS");
2107 snprintf(results, sizeof(results),
2108 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2109 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2110 "%7.3f\t%7.3f\t%7.3f\t%7.3f",
2111 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
2112 cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr,
2113 total_ssim, totalp_ssim,
2114 cpi->fastssim.stat[ALL] / cpi->count,
2115 cpi->psnrhvs.stat[ALL] / cpi->count,
2116 cpi->psnr.worst, cpi->worst_ssim, cpi->fastssim.worst,
2117 cpi->psnrhvs.worst);
2119 if (cpi->b_calculate_blockiness) {
2120 SNPRINT(headings, "\t Block\tWstBlck");
2121 SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
2122 SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
2125 if (cpi->b_calculate_consistency) {
2126 double consistency =
2127 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
2128 (double)cpi->total_inconsistency);
2130 SNPRINT(headings, "\tConsist\tWstCons");
2131 SNPRINT2(results, "\t%7.3f", consistency);
2132 SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
2135 if (cpi->b_calculate_ssimg) {
2136 SNPRINT(headings, "\t SSIMG\tWtSSIMG");
2137 SNPRINT2(results, "\t%7.3f", cpi->ssimg.stat[ALL] / cpi->count);
2138 SNPRINT2(results, "\t%7.3f", cpi->ssimg.worst);
2141 fprintf(f, "%s\t Time Rc-Err Abs Err\n", headings);
2142 fprintf(f, "%s\t%8.0f %7.2f %7.2f\n", results,
2143 total_encode_time, rate_err, fabs(rate_err));
2153 printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
2154 printf("\n_frames recive_data encod_mb_row compress_frame Total\n");
2155 printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
2156 cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
2157 cpi->time_compress_data / 1000,
2158 (cpi->time_receive_data + cpi->time_compress_data) / 1000);
2163 #if CONFIG_VP9_TEMPORAL_DENOISING
2164 vp9_denoiser_free(&(cpi->denoiser));
2167 for (t = 0; t < cpi->num_workers; ++t) {
2168 VPxWorker *const worker = &cpi->workers[t];
2169 EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
2171 // Deallocate allocated threads.
2172 vpx_get_worker_interface()->end(worker);
2174 // Deallocate allocated thread data.
2175 if (t < cpi->num_workers - 1) {
2176 vpx_free(thread_data->td->counts);
2177 vp9_free_pc_tree(thread_data->td);
2178 vpx_free(thread_data->td);
2181 vpx_free(cpi->tile_thr_data);
2182 vpx_free(cpi->workers);
2184 if (cpi->num_workers > 1)
2185 vp9_loop_filter_dealloc(&cpi->lf_row_sync);
2187 dealloc_compressor_data(cpi);
2189 for (i = 0; i < sizeof(cpi->mbgraph_stats) /
2190 sizeof(cpi->mbgraph_stats[0]); ++i) {
2191 vpx_free(cpi->mbgraph_stats[i].mb_stats);
2194 #if CONFIG_FP_MB_STATS
2195 if (cpi->use_fp_mb_stats) {
2196 vpx_free(cpi->twopass.frame_mb_stats_buf);
2197 cpi->twopass.frame_mb_stats_buf = NULL;
2201 vp9_remove_common(cm);
2202 vp9_free_ref_frame_buffers(cm->buffer_pool);
2203 #if CONFIG_VP9_POSTPROC
2204 vp9_free_postproc_buffers(cm);
2208 #if CONFIG_VP9_TEMPORAL_DENOISING
2209 #ifdef OUTPUT_YUV_DENOISED
2210 fclose(yuv_denoised_file);
2213 #ifdef OUTPUT_YUV_SKINMAP
2214 fclose(yuv_skinmap_file);
2216 #ifdef OUTPUT_YUV_REC
2217 fclose(yuv_rec_file);
2234 /* TODO(yaowu): The block_variance calls the unoptimized versions of variance()
2235 * and highbd_8_variance(). It should not.
2237 static void encoder_variance(const uint8_t *a, int a_stride,
2238 const uint8_t *b, int b_stride,
2239 int w, int h, unsigned int *sse, int *sum) {
2245 for (i = 0; i < h; i++) {
2246 for (j = 0; j < w; j++) {
2247 const int diff = a[j] - b[j];
2249 *sse += diff * diff;
2257 #if CONFIG_VP9_HIGHBITDEPTH
2258 static void encoder_highbd_variance64(const uint8_t *a8, int a_stride,
2259 const uint8_t *b8, int b_stride,
2260 int w, int h, uint64_t *sse,
2264 uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2265 uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2269 for (i = 0; i < h; i++) {
2270 for (j = 0; j < w; j++) {
2271 const int diff = a[j] - b[j];
2273 *sse += diff * diff;
2280 static void encoder_highbd_8_variance(const uint8_t *a8, int a_stride,
2281 const uint8_t *b8, int b_stride,
2283 unsigned int *sse, int *sum) {
2284 uint64_t sse_long = 0;
2285 int64_t sum_long = 0;
2286 encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h,
2287 &sse_long, &sum_long);
2288 *sse = (unsigned int)sse_long;
2289 *sum = (int)sum_long;
2291 #endif // CONFIG_VP9_HIGHBITDEPTH
2293 static int64_t get_sse(const uint8_t *a, int a_stride,
2294 const uint8_t *b, int b_stride,
2295 int width, int height) {
2296 const int dw = width % 16;
2297 const int dh = height % 16;
2298 int64_t total_sse = 0;
2299 unsigned int sse = 0;
2304 encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride,
2305 dw, height, &sse, &sum);
2310 encoder_variance(&a[(height - dh) * a_stride], a_stride,
2311 &b[(height - dh) * b_stride], b_stride,
2312 width - dw, dh, &sse, &sum);
2316 for (y = 0; y < height / 16; ++y) {
2317 const uint8_t *pa = a;
2318 const uint8_t *pb = b;
2319 for (x = 0; x < width / 16; ++x) {
2320 vpx_mse16x16(pa, a_stride, pb, b_stride, &sse);
2334 #if CONFIG_VP9_HIGHBITDEPTH
2335 static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride,
2336 const uint8_t *b8, int b_stride,
2337 int width, int height,
2338 unsigned int input_shift) {
2339 const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2340 const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2341 int64_t total_sse = 0;
2343 for (y = 0; y < height; ++y) {
2344 for (x = 0; x < width; ++x) {
2346 diff = (a[x] >> input_shift) - (b[x] >> input_shift);
2347 total_sse += diff * diff;
2355 static int64_t highbd_get_sse(const uint8_t *a, int a_stride,
2356 const uint8_t *b, int b_stride,
2357 int width, int height) {
2358 int64_t total_sse = 0;
2360 const int dw = width % 16;
2361 const int dh = height % 16;
2362 unsigned int sse = 0;
2365 encoder_highbd_8_variance(&a[width - dw], a_stride,
2366 &b[width - dw], b_stride,
2367 dw, height, &sse, &sum);
2371 encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride,
2372 &b[(height - dh) * b_stride], b_stride,
2373 width - dw, dh, &sse, &sum);
2376 for (y = 0; y < height / 16; ++y) {
2377 const uint8_t *pa = a;
2378 const uint8_t *pb = b;
2379 for (x = 0; x < width / 16; ++x) {
2380 vpx_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse);
2390 #endif // CONFIG_VP9_HIGHBITDEPTH
2393 double psnr[4]; // total/y/u/v
2394 uint64_t sse[4]; // total/y/u/v
2395 uint32_t samples[4]; // total/y/u/v
2398 #if CONFIG_VP9_HIGHBITDEPTH
2399 static void calc_highbd_psnr(const YV12_BUFFER_CONFIG *a,
2400 const YV12_BUFFER_CONFIG *b,
2402 unsigned int bit_depth,
2403 unsigned int in_bit_depth) {
2404 const int widths[3] =
2405 {a->y_crop_width, a->uv_crop_width, a->uv_crop_width };
2406 const int heights[3] =
2407 {a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2408 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer };
2409 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2410 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer };
2411 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2413 uint64_t total_sse = 0;
2414 uint32_t total_samples = 0;
2415 const double peak = (double)((1 << in_bit_depth) - 1);
2416 const unsigned int input_shift = bit_depth - in_bit_depth;
2418 for (i = 0; i < 3; ++i) {
2419 const int w = widths[i];
2420 const int h = heights[i];
2421 const uint32_t samples = w * h;
2423 if (a->flags & YV12_FLAG_HIGHBITDEPTH) {
2425 sse = highbd_get_sse_shift(a_planes[i], a_strides[i],
2426 b_planes[i], b_strides[i], w, h,
2429 sse = highbd_get_sse(a_planes[i], a_strides[i],
2430 b_planes[i], b_strides[i], w, h);
2433 sse = get_sse(a_planes[i], a_strides[i],
2434 b_planes[i], b_strides[i],
2437 psnr->sse[1 + i] = sse;
2438 psnr->samples[1 + i] = samples;
2439 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2442 total_samples += samples;
2445 psnr->sse[0] = total_sse;
2446 psnr->samples[0] = total_samples;
2447 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2451 #else // !CONFIG_VP9_HIGHBITDEPTH
2453 static void calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b,
2455 static const double peak = 255.0;
2456 const int widths[3] = {
2457 a->y_crop_width, a->uv_crop_width, a->uv_crop_width};
2458 const int heights[3] = {
2459 a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2460 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer};
2461 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2462 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer};
2463 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2465 uint64_t total_sse = 0;
2466 uint32_t total_samples = 0;
2468 for (i = 0; i < 3; ++i) {
2469 const int w = widths[i];
2470 const int h = heights[i];
2471 const uint32_t samples = w * h;
2472 const uint64_t sse = get_sse(a_planes[i], a_strides[i],
2473 b_planes[i], b_strides[i],
2475 psnr->sse[1 + i] = sse;
2476 psnr->samples[1 + i] = samples;
2477 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2480 total_samples += samples;
2483 psnr->sse[0] = total_sse;
2484 psnr->samples[0] = total_samples;
2485 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2488 #endif // CONFIG_VP9_HIGHBITDEPTH
2490 static void generate_psnr_packet(VP9_COMP *cpi) {
2491 struct vpx_codec_cx_pkt pkt;
2494 #if CONFIG_VP9_HIGHBITDEPTH
2495 calc_highbd_psnr(cpi->Source, cpi->common.frame_to_show, &psnr,
2496 cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
2498 calc_psnr(cpi->Source, cpi->common.frame_to_show, &psnr);
2501 for (i = 0; i < 4; ++i) {
2502 pkt.data.psnr.samples[i] = psnr.samples[i];
2503 pkt.data.psnr.sse[i] = psnr.sse[i];
2504 pkt.data.psnr.psnr[i] = psnr.psnr[i];
2506 pkt.kind = VPX_CODEC_PSNR_PKT;
2508 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
2509 cpi->svc.number_temporal_layers].psnr_pkt = pkt.data.psnr;
2511 vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
2514 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) {
2515 if (ref_frame_flags > 7)
2518 cpi->ref_frame_flags = ref_frame_flags;
2522 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) {
2523 cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0;
2524 cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0;
2525 cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0;
2526 cpi->ext_refresh_frame_flags_pending = 1;
2529 static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(VP9_COMP *cpi,
2530 VP9_REFFRAME ref_frame_flag) {
2531 MV_REFERENCE_FRAME ref_frame = NONE;
2532 if (ref_frame_flag == VP9_LAST_FLAG)
2533 ref_frame = LAST_FRAME;
2534 else if (ref_frame_flag == VP9_GOLD_FLAG)
2535 ref_frame = GOLDEN_FRAME;
2536 else if (ref_frame_flag == VP9_ALT_FLAG)
2537 ref_frame = ALTREF_FRAME;
2539 return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame);
2542 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2543 YV12_BUFFER_CONFIG *sd) {
2544 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2546 vp8_yv12_copy_frame(cfg, sd);
2553 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2554 YV12_BUFFER_CONFIG *sd) {
2555 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2557 vp8_yv12_copy_frame(sd, cfg);
2564 int vp9_update_entropy(VP9_COMP * cpi, int update) {
2565 cpi->ext_refresh_frame_context = update;
2566 cpi->ext_refresh_frame_context_pending = 1;
2570 #if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP)
2571 // The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
2572 // as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
2573 // not denoise the UV channels at this time. If ever we implement UV channel
2574 // denoising we will have to modify this.
2575 void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
2576 uint8_t *src = s->y_buffer;
2577 int h = s->y_height;
2580 fwrite(src, s->y_width, 1, f);
2588 fwrite(src, s->uv_width, 1, f);
2589 src += s->uv_stride;
2596 fwrite(src, s->uv_width, 1, f);
2597 src += s->uv_stride;
2602 #ifdef OUTPUT_YUV_REC
2603 void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
2604 YV12_BUFFER_CONFIG *s = cm->frame_to_show;
2605 uint8_t *src = s->y_buffer;
2608 #if CONFIG_VP9_HIGHBITDEPTH
2609 if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
2610 uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
2613 fwrite(src16, s->y_width, 2, yuv_rec_file);
2614 src16 += s->y_stride;
2617 src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
2621 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2622 src16 += s->uv_stride;
2625 src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
2629 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2630 src16 += s->uv_stride;
2633 fflush(yuv_rec_file);
2636 #endif // CONFIG_VP9_HIGHBITDEPTH
2639 fwrite(src, s->y_width, 1, yuv_rec_file);
2647 fwrite(src, s->uv_width, 1, yuv_rec_file);
2648 src += s->uv_stride;
2655 fwrite(src, s->uv_width, 1, yuv_rec_file);
2656 src += s->uv_stride;
2659 fflush(yuv_rec_file);
2663 #if CONFIG_VP9_HIGHBITDEPTH
2664 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2665 YV12_BUFFER_CONFIG *dst,
2668 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2669 YV12_BUFFER_CONFIG *dst) {
2670 #endif // CONFIG_VP9_HIGHBITDEPTH
2671 // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
2673 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2674 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2675 const int src_widths[3] = {src->y_crop_width, src->uv_crop_width,
2676 src->uv_crop_width };
2677 const int src_heights[3] = {src->y_crop_height, src->uv_crop_height,
2678 src->uv_crop_height};
2679 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2680 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2681 const int dst_widths[3] = {dst->y_crop_width, dst->uv_crop_width,
2682 dst->uv_crop_width};
2683 const int dst_heights[3] = {dst->y_crop_height, dst->uv_crop_height,
2684 dst->uv_crop_height};
2686 for (i = 0; i < MAX_MB_PLANE; ++i) {
2687 #if CONFIG_VP9_HIGHBITDEPTH
2688 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2689 vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
2690 src_strides[i], dsts[i], dst_heights[i],
2691 dst_widths[i], dst_strides[i], bd);
2693 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2694 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2697 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2698 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2699 #endif // CONFIG_VP9_HIGHBITDEPTH
2701 vpx_extend_frame_borders(dst);
2704 #if CONFIG_VP9_HIGHBITDEPTH
2705 static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
2706 YV12_BUFFER_CONFIG *dst, int bd) {
2707 const int src_w = src->y_crop_width;
2708 const int src_h = src->y_crop_height;
2709 const int dst_w = dst->y_crop_width;
2710 const int dst_h = dst->y_crop_height;
2711 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2712 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2713 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2714 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2715 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2718 for (i = 0; i < MAX_MB_PLANE; ++i) {
2719 const int factor = (i == 0 || i == 3 ? 1 : 2);
2720 const int src_stride = src_strides[i];
2721 const int dst_stride = dst_strides[i];
2722 for (y = 0; y < dst_h; y += 16) {
2723 const int y_q4 = y * (16 / factor) * src_h / dst_h;
2724 for (x = 0; x < dst_w; x += 16) {
2725 const int x_q4 = x * (16 / factor) * src_w / dst_w;
2726 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
2727 src_stride + (x / factor) * src_w / dst_w;
2728 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2730 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2731 vpx_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
2732 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2733 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2734 16 / factor, 16 / factor, bd);
2736 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2737 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2738 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2739 16 / factor, 16 / factor);
2745 vpx_extend_frame_borders(dst);
2748 void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
2749 YV12_BUFFER_CONFIG *dst) {
2750 const int src_w = src->y_crop_width;
2751 const int src_h = src->y_crop_height;
2752 const int dst_w = dst->y_crop_width;
2753 const int dst_h = dst->y_crop_height;
2754 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2755 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2756 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2757 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2758 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2761 for (i = 0; i < MAX_MB_PLANE; ++i) {
2762 const int factor = (i == 0 || i == 3 ? 1 : 2);
2763 const int src_stride = src_strides[i];
2764 const int dst_stride = dst_strides[i];
2765 for (y = 0; y < dst_h; y += 16) {
2766 const int y_q4 = y * (16 / factor) * src_h / dst_h;
2767 for (x = 0; x < dst_w; x += 16) {
2768 const int x_q4 = x * (16 / factor) * src_w / dst_w;
2769 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
2770 src_stride + (x / factor) * src_w / dst_w;
2771 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2773 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2774 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2775 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2776 16 / factor, 16 / factor);
2781 vpx_extend_frame_borders(dst);
2783 #endif // CONFIG_VP9_HIGHBITDEPTH
2785 static int scale_down(VP9_COMP *cpi, int q) {
2786 RATE_CONTROL *const rc = &cpi->rc;
2787 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2789 assert(frame_is_kf_gf_arf(cpi));
2791 if (rc->frame_size_selector == UNSCALED &&
2792 q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
2793 const int max_size_thresh = (int)(rate_thresh_mult[SCALE_STEP1]
2794 * VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
2795 scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
2800 static int big_rate_miss(VP9_COMP *cpi, int high_limit, int low_limit) {
2801 const RATE_CONTROL *const rc = &cpi->rc;
2803 return (rc->projected_frame_size > ((high_limit * 3) / 2)) ||
2804 (rc->projected_frame_size < (low_limit / 2));
2807 // Function to test for conditions that indicate we should loop
2808 // back and recode a frame.
2809 static int recode_loop_test(VP9_COMP *cpi,
2810 int high_limit, int low_limit,
2811 int q, int maxq, int minq) {
2812 const RATE_CONTROL *const rc = &cpi->rc;
2813 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2814 const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
2815 int force_recode = 0;
2817 if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
2818 big_rate_miss(cpi, high_limit, low_limit) ||
2819 (cpi->sf.recode_loop == ALLOW_RECODE) ||
2820 (frame_is_kfgfarf &&
2821 (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) {
2822 if (frame_is_kfgfarf &&
2823 (oxcf->resize_mode == RESIZE_DYNAMIC) &&
2824 scale_down(cpi, q)) {
2825 // Code this group at a lower resolution.
2826 cpi->resize_pending = 1;
2830 // TODO(agrange) high_limit could be greater than the scale-down threshold.
2831 if ((rc->projected_frame_size > high_limit && q < maxq) ||
2832 (rc->projected_frame_size < low_limit && q > minq)) {
2834 } else if (cpi->oxcf.rc_mode == VPX_CQ) {
2835 // Deal with frame undershoot and whether or not we are
2836 // below the automatically set cq level.
2837 if (q > oxcf->cq_level &&
2838 rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
2843 return force_recode;
2846 void vp9_update_reference_frames(VP9_COMP *cpi) {
2847 VP9_COMMON * const cm = &cpi->common;
2848 BufferPool *const pool = cm->buffer_pool;
2850 // At this point the new frame has been encoded.
2851 // If any buffer copy / swapping is signaled it should be done here.
2852 if (cm->frame_type == KEY_FRAME) {
2853 ref_cnt_fb(pool->frame_bufs,
2854 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2855 ref_cnt_fb(pool->frame_bufs,
2856 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2857 } else if (vp9_preserve_existing_gf(cpi)) {
2858 // We have decided to preserve the previously existing golden frame as our
2859 // new ARF frame. However, in the short term in function
2860 // vp9_get_refresh_mask() we left it in the GF slot and, if
2861 // we're updating the GF with the current decoded frame, we save it to the
2862 // ARF slot instead.
2863 // We now have to update the ARF with the current frame and swap gld_fb_idx
2864 // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
2865 // slot and, if we're updating the GF, the current frame becomes the new GF.
2868 ref_cnt_fb(pool->frame_bufs,
2869 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2871 tmp = cpi->alt_fb_idx;
2872 cpi->alt_fb_idx = cpi->gld_fb_idx;
2873 cpi->gld_fb_idx = tmp;
2875 if (is_two_pass_svc(cpi)) {
2876 cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx;
2877 cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx;
2879 } else { /* For non key/golden frames */
2880 if (cpi->refresh_alt_ref_frame) {
2881 int arf_idx = cpi->alt_fb_idx;
2882 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
2883 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2884 arf_idx = gf_group->arf_update_idx[gf_group->index];
2887 ref_cnt_fb(pool->frame_bufs,
2888 &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
2889 memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
2890 cpi->interp_filter_selected[0],
2891 sizeof(cpi->interp_filter_selected[0]));
2894 if (cpi->refresh_golden_frame) {
2895 ref_cnt_fb(pool->frame_bufs,
2896 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2897 if (!cpi->rc.is_src_frame_alt_ref)
2898 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2899 cpi->interp_filter_selected[0],
2900 sizeof(cpi->interp_filter_selected[0]));
2902 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2903 cpi->interp_filter_selected[ALTREF_FRAME],
2904 sizeof(cpi->interp_filter_selected[ALTREF_FRAME]));
2908 if (cpi->refresh_last_frame) {
2909 ref_cnt_fb(pool->frame_bufs,
2910 &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
2911 if (!cpi->rc.is_src_frame_alt_ref)
2912 memcpy(cpi->interp_filter_selected[LAST_FRAME],
2913 cpi->interp_filter_selected[0],
2914 sizeof(cpi->interp_filter_selected[0]));
2916 #if CONFIG_VP9_TEMPORAL_DENOISING
2917 if (cpi->oxcf.noise_sensitivity > 0 &&
2918 cpi->denoiser.denoising_level > kDenLowLow) {
2919 vp9_denoiser_update_frame_info(&cpi->denoiser,
2921 cpi->common.frame_type,
2922 cpi->refresh_alt_ref_frame,
2923 cpi->refresh_golden_frame,
2924 cpi->refresh_last_frame,
2925 cpi->resize_pending);
2928 if (is_one_pass_cbr_svc(cpi)) {
2929 // Keep track of frame index for each reference frame.
2930 SVC *const svc = &cpi->svc;
2931 if (cm->frame_type == KEY_FRAME) {
2932 svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2933 svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2934 svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2936 if (cpi->refresh_last_frame)
2937 svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2938 if (cpi->refresh_golden_frame)
2939 svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2940 if (cpi->refresh_alt_ref_frame)
2941 svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2946 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
2947 MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
2948 struct loopfilter *lf = &cm->lf;
2951 lf->filter_level = 0;
2952 lf->last_filt_level = 0;
2954 struct vpx_usec_timer timer;
2956 vpx_clear_system_state();
2958 vpx_usec_timer_start(&timer);
2960 if (!cpi->rc.is_src_frame_alt_ref) {
2961 if ((cpi->common.frame_type == KEY_FRAME) &&
2962 (!cpi->rc.this_key_frame_forced)) {
2963 lf->last_filt_level = 0;
2965 vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick);
2966 lf->last_filt_level = lf->filter_level;
2968 lf->filter_level = 0;
2971 vpx_usec_timer_mark(&timer);
2972 cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
2975 if (lf->filter_level > 0) {
2976 vp9_build_mask_frame(cm, lf->filter_level, 0);
2978 if (cpi->num_workers > 1)
2979 vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
2980 lf->filter_level, 0, 0,
2981 cpi->workers, cpi->num_workers,
2984 vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
2987 vpx_extend_frame_inner_borders(cm->frame_to_show);
2990 static INLINE void alloc_frame_mvs(VP9_COMMON *const cm,
2992 RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
2993 if (new_fb_ptr->mvs == NULL ||
2994 new_fb_ptr->mi_rows < cm->mi_rows ||
2995 new_fb_ptr->mi_cols < cm->mi_cols) {
2996 vpx_free(new_fb_ptr->mvs);
2997 CHECK_MEM_ERROR(cm, new_fb_ptr->mvs,
2998 (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
2999 sizeof(*new_fb_ptr->mvs)));
3000 new_fb_ptr->mi_rows = cm->mi_rows;
3001 new_fb_ptr->mi_cols = cm->mi_cols;
3005 void vp9_scale_references(VP9_COMP *cpi) {
3006 VP9_COMMON *cm = &cpi->common;
3007 MV_REFERENCE_FRAME ref_frame;
3008 const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
3010 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3011 // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
3012 if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
3013 BufferPool *const pool = cm->buffer_pool;
3014 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi,
3018 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3022 #if CONFIG_VP9_HIGHBITDEPTH
3023 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
3024 RefCntBuffer *new_fb_ptr = NULL;
3025 int force_scaling = 0;
3026 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
3027 if (new_fb == INVALID_IDX) {
3028 new_fb = get_free_fb(cm);
3031 if (new_fb == INVALID_IDX)
3033 new_fb_ptr = &pool->frame_bufs[new_fb];
3034 if (force_scaling ||
3035 new_fb_ptr->buf.y_crop_width != cm->width ||
3036 new_fb_ptr->buf.y_crop_height != cm->height) {
3037 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
3038 cm->subsampling_x, cm->subsampling_y,
3039 cm->use_highbitdepth,
3040 VP9_ENC_BORDER_IN_PIXELS,
3041 cm->byte_alignment, NULL, NULL, NULL))
3042 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3043 "Failed to allocate frame buffer");
3044 scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth);
3045 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
3046 alloc_frame_mvs(cm, new_fb);
3049 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
3050 RefCntBuffer *new_fb_ptr = NULL;
3051 int force_scaling = 0;
3052 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
3053 if (new_fb == INVALID_IDX) {
3054 new_fb = get_free_fb(cm);
3057 if (new_fb == INVALID_IDX)
3059 new_fb_ptr = &pool->frame_bufs[new_fb];
3060 if (force_scaling ||
3061 new_fb_ptr->buf.y_crop_width != cm->width ||
3062 new_fb_ptr->buf.y_crop_height != cm->height) {
3063 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
3064 cm->subsampling_x, cm->subsampling_y,
3065 VP9_ENC_BORDER_IN_PIXELS,
3066 cm->byte_alignment, NULL, NULL, NULL))
3067 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3068 "Failed to allocate frame buffer");
3069 vp9_scale_and_extend_frame(ref, &new_fb_ptr->buf);
3070 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
3071 alloc_frame_mvs(cm, new_fb);
3073 #endif // CONFIG_VP9_HIGHBITDEPTH
3076 RefCntBuffer *buf = NULL;
3077 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3078 // Check for release of scaled reference.
3079 buf_idx = cpi->scaled_ref_idx[ref_frame - 1];
3080 buf = (buf_idx != INVALID_IDX) ? &pool->frame_bufs[buf_idx] : NULL;
3083 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3086 buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3087 buf = &pool->frame_bufs[buf_idx];
3088 buf->buf.y_crop_width = ref->y_crop_width;
3089 buf->buf.y_crop_height = ref->y_crop_height;
3090 cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
3094 if (cpi->oxcf.pass != 0 || cpi->use_svc)
3095 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3100 static void release_scaled_references(VP9_COMP *cpi) {
3101 VP9_COMMON *cm = &cpi->common;
3103 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3104 // Only release scaled references under certain conditions:
3105 // if reference will be updated, or if scaled reference has same resolution.
3107 refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
3108 refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
3109 refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
3110 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3111 const int idx = cpi->scaled_ref_idx[i - 1];
3112 RefCntBuffer *const buf = idx != INVALID_IDX ?
3113 &cm->buffer_pool->frame_bufs[idx] : NULL;
3114 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
3117 (buf->buf.y_crop_width == ref->y_crop_width &&
3118 buf->buf.y_crop_height == ref->y_crop_height))) {
3120 cpi->scaled_ref_idx[i -1] = INVALID_IDX;
3124 for (i = 0; i < MAX_REF_FRAMES; ++i) {
3125 const int idx = cpi->scaled_ref_idx[i];
3126 RefCntBuffer *const buf = idx != INVALID_IDX ?
3127 &cm->buffer_pool->frame_bufs[idx] : NULL;
3130 cpi->scaled_ref_idx[i] = INVALID_IDX;
3136 static void full_to_model_count(unsigned int *model_count,
3137 unsigned int *full_count) {
3139 model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
3140 model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
3141 model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
3142 for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
3143 model_count[TWO_TOKEN] += full_count[n];
3144 model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
3147 static void full_to_model_counts(vp9_coeff_count_model *model_count,
3148 vp9_coeff_count *full_count) {
3151 for (i = 0; i < PLANE_TYPES; ++i)
3152 for (j = 0; j < REF_TYPES; ++j)
3153 for (k = 0; k < COEF_BANDS; ++k)
3154 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
3155 full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
3158 #if 0 && CONFIG_INTERNAL_STATS
3159 static void output_frame_level_debug_stats(VP9_COMP *cpi) {
3160 VP9_COMMON *const cm = &cpi->common;
3161 FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
3164 vpx_clear_system_state();
3166 #if CONFIG_VP9_HIGHBITDEPTH
3167 if (cm->use_highbitdepth) {
3168 recon_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3170 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3173 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3174 #endif // CONFIG_VP9_HIGHBITDEPTH
3177 if (cpi->twopass.total_left_stats.coded_error != 0.0) {
3178 double dc_quant_devisor;
3179 #if CONFIG_VP9_HIGHBITDEPTH
3180 switch (cm->bit_depth) {
3182 dc_quant_devisor = 4.0;
3185 dc_quant_devisor = 16.0;
3188 dc_quant_devisor = 64.0;
3191 assert(0 && "bit_depth must be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
3195 dc_quant_devisor = 4.0;
3198 fprintf(f, "%10u %dx%d %10d %10d %d %d %10d %10d %10d %10d"
3199 "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
3200 "%10"PRId64" %10"PRId64" %10d "
3201 "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
3202 "%6d %6d %5d %5d %5d "
3203 "%10"PRId64" %10.3lf"
3204 "%10lf %8u %10"PRId64" %10d %10d %10d %10d %10d\n",
3205 cpi->common.current_video_frame,
3206 cm->width, cm->height,
3207 cpi->td.rd_counts.m_search_count,
3208 cpi->td.rd_counts.ex_search_count,
3209 cpi->rc.source_alt_ref_pending,
3210 cpi->rc.source_alt_ref_active,
3211 cpi->rc.this_frame_target,
3212 cpi->rc.projected_frame_size,
3213 cpi->rc.projected_frame_size / cpi->common.MBs,
3214 (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
3215 cpi->rc.vbr_bits_off_target,
3216 cpi->rc.vbr_bits_off_target_fast,
3217 cpi->twopass.extend_minq,
3218 cpi->twopass.extend_minq_fast,
3219 cpi->rc.total_target_vs_actual,
3220 (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
3221 cpi->rc.total_actual_bits, cm->base_qindex,
3222 vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
3223 (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) /
3225 vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality,
3228 vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
3229 cpi->refresh_last_frame, cpi->refresh_golden_frame,
3230 cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
3231 cpi->twopass.bits_left,
3232 cpi->twopass.total_left_stats.coded_error,
3233 cpi->twopass.bits_left /
3234 (1 + cpi->twopass.total_left_stats.coded_error),
3235 cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
3236 cpi->twopass.kf_zeromotion_pct,
3237 cpi->twopass.fr_content_type,
3238 cm->lf.filter_level,
3239 cm->seg.aq_av_offset);
3244 FILE *const fmodes = fopen("Modes.stt", "a");
3247 fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
3248 cm->frame_type, cpi->refresh_golden_frame,
3249 cpi->refresh_alt_ref_frame);
3251 for (i = 0; i < MAX_MODES; ++i)
3252 fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
3254 fprintf(fmodes, "\n");
3261 static void set_mv_search_params(VP9_COMP *cpi) {
3262 const VP9_COMMON *const cm = &cpi->common;
3263 const unsigned int max_mv_def = VPXMIN(cm->width, cm->height);
3265 // Default based on max resolution.
3266 cpi->mv_step_param = vp9_init_search_range(max_mv_def);
3268 if (cpi->sf.mv.auto_mv_step_size) {
3269 if (frame_is_intra_only(cm)) {
3270 // Initialize max_mv_magnitude for use in the first INTER frame
3271 // after a key/intra-only frame.
3272 cpi->max_mv_magnitude = max_mv_def;
3274 if (cm->show_frame) {
3275 // Allow mv_steps to correspond to twice the max mv magnitude found
3276 // in the previous frame, capped by the default max_mv_magnitude based
3278 cpi->mv_step_param = vp9_init_search_range(
3279 VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
3281 cpi->max_mv_magnitude = 0;
3286 static void set_size_independent_vars(VP9_COMP *cpi) {
3287 vp9_set_speed_features_framesize_independent(cpi);
3288 vp9_set_rd_speed_thresholds(cpi);
3289 vp9_set_rd_speed_thresholds_sub8x8(cpi);
3290 cpi->common.interp_filter = cpi->sf.default_interp_filter;
3293 static void set_size_dependent_vars(VP9_COMP *cpi, int *q,
3294 int *bottom_index, int *top_index) {
3295 VP9_COMMON *const cm = &cpi->common;
3296 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3298 // Setup variables that depend on the dimensions of the frame.
3299 vp9_set_speed_features_framesize_dependent(cpi);
3301 // Decide q and q bounds.
3302 *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
3304 if (!frame_is_intra_only(cm)) {
3305 vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
3308 // Configure experimental use of segmentation for enhanced coding of
3309 // static regions if indicated.
3310 // Only allowed in the second pass of a two pass encode, as it requires
3311 // lagged coding, and if the relevant speed feature flag is set.
3312 if (oxcf->pass == 2 && cpi->sf.static_segmentation)
3313 configure_static_seg_features(cpi);
3315 #if CONFIG_VP9_POSTPROC && !(CONFIG_VP9_TEMPORAL_DENOISING)
3316 if (oxcf->noise_sensitivity > 0) {
3318 switch (oxcf->noise_sensitivity) {
3336 vp9_denoise(cpi->Source, cpi->Source, l);
3338 #endif // CONFIG_VP9_POSTPROC
3341 #if CONFIG_VP9_TEMPORAL_DENOISING
3342 static void setup_denoiser_buffer(VP9_COMP *cpi) {
3343 VP9_COMMON *const cm = &cpi->common;
3344 if (cpi->oxcf.noise_sensitivity > 0 &&
3345 !cpi->denoiser.frame_buffer_initialized) {
3346 if (vp9_denoiser_alloc(&cpi->denoiser, cm->width, cm->height,
3347 cm->subsampling_x, cm->subsampling_y,
3348 #if CONFIG_VP9_HIGHBITDEPTH
3349 cm->use_highbitdepth,
3351 VP9_ENC_BORDER_IN_PIXELS))
3352 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3353 "Failed to allocate denoiser");
3358 static void init_motion_estimation(VP9_COMP *cpi) {
3359 int y_stride = cpi->scaled_source.y_stride;
3361 if (cpi->sf.mv.search_method == NSTEP) {
3362 vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
3363 } else if (cpi->sf.mv.search_method == DIAMOND) {
3364 vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
3368 static void set_frame_size(VP9_COMP *cpi) {
3370 VP9_COMMON *const cm = &cpi->common;
3371 VP9EncoderConfig *const oxcf = &cpi->oxcf;
3372 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
3374 if (oxcf->pass == 2 &&
3375 oxcf->rc_mode == VPX_VBR &&
3376 ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
3377 (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
3378 calculate_coded_size(
3379 cpi, &oxcf->scaled_frame_width, &oxcf->scaled_frame_height);
3381 // There has been a change in frame size.
3382 vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
3383 oxcf->scaled_frame_height);
3386 if (oxcf->pass == 0 &&
3387 oxcf->rc_mode == VPX_CBR &&
3389 oxcf->resize_mode == RESIZE_DYNAMIC &&
3390 cpi->resize_pending != 0) {
3391 oxcf->scaled_frame_width =
3392 (oxcf->width * cpi->resize_scale_num) / cpi->resize_scale_den;
3393 oxcf->scaled_frame_height =
3394 (oxcf->height * cpi->resize_scale_num) /cpi->resize_scale_den;
3395 // There has been a change in frame size.
3396 vp9_set_size_literal(cpi,
3397 oxcf->scaled_frame_width,
3398 oxcf->scaled_frame_height);
3400 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3401 set_mv_search_params(cpi);
3403 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
3404 #if CONFIG_VP9_TEMPORAL_DENOISING
3405 // Reset the denoiser on the resized frame.
3406 if (cpi->oxcf.noise_sensitivity > 0) {
3407 vp9_denoiser_free(&(cpi->denoiser));
3408 setup_denoiser_buffer(cpi);
3409 // Dynamic resize is only triggered for non-SVC, so we can force
3410 // golden frame update here as temporary fix to denoiser.
3411 cpi->refresh_golden_frame = 1;
3416 if ((oxcf->pass == 2) &&
3418 (is_two_pass_svc(cpi) &&
3419 cpi->svc.encode_empty_frame_state != ENCODING))) {
3420 vp9_set_target_rate(cpi);
3423 alloc_frame_mvs(cm, cm->new_fb_idx);
3425 // Reset the frame pointers to the current frame size.
3426 if (vpx_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
3427 cm->subsampling_x, cm->subsampling_y,
3428 #if CONFIG_VP9_HIGHBITDEPTH
3429 cm->use_highbitdepth,
3431 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
3433 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3434 "Failed to allocate frame buffer");
3436 alloc_util_frame_buffers(cpi);
3437 init_motion_estimation(cpi);
3439 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3440 RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
3441 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3443 ref_buf->idx = buf_idx;
3445 if (buf_idx != INVALID_IDX) {
3446 YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
3448 #if CONFIG_VP9_HIGHBITDEPTH
3449 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3450 buf->y_crop_width, buf->y_crop_height,
3451 cm->width, cm->height,
3452 (buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
3455 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3456 buf->y_crop_width, buf->y_crop_height,
3457 cm->width, cm->height);
3458 #endif // CONFIG_VP9_HIGHBITDEPTH
3459 if (vp9_is_scaled(&ref_buf->sf))
3460 vpx_extend_frame_borders(buf);
3462 ref_buf->buf = NULL;
3466 set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
3469 static void encode_without_recode_loop(VP9_COMP *cpi,
3472 VP9_COMMON *const cm = &cpi->common;
3473 int q = 0, bottom_index = 0, top_index = 0; // Dummy variables.
3475 vpx_clear_system_state();
3477 set_frame_size(cpi);
3479 if (is_one_pass_cbr_svc(cpi) &&
3480 cpi->un_scaled_source->y_width == cm->width << 2 &&
3481 cpi->un_scaled_source->y_height == cm->height << 2 &&
3482 cpi->svc.scaled_temp.y_width == cm->width << 1 &&
3483 cpi->svc.scaled_temp.y_height == cm->height << 1) {
3484 cpi->Source = vp9_svc_twostage_scale(cm,
3485 cpi->un_scaled_source,
3486 &cpi->scaled_source,
3487 &cpi->svc.scaled_temp);
3489 cpi->Source = vp9_scale_if_required(cm,
3490 cpi->un_scaled_source,
3491 &cpi->scaled_source,
3492 (cpi->oxcf.pass == 0));
3494 // Avoid scaling last_source unless its needed.
3495 // Last source is needed if vp9_avg_source_sad() is used, or if
3496 // partition_search_type == SOURCE_VAR_BASED_PARTITION, or if noise
3497 // estimation is enabled.
3498 if (cpi->unscaled_last_source != NULL &&
3499 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3500 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_VBR &&
3501 cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5) ||
3502 cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION ||
3503 cpi->noise_estimate.enabled))
3504 cpi->Last_Source = vp9_scale_if_required(cm,
3505 cpi->unscaled_last_source,
3506 &cpi->scaled_last_source,
3507 (cpi->oxcf.pass == 0));
3509 if (cm->frame_type == KEY_FRAME || cpi->resize_pending != 0) {
3510 memset(cpi->consec_zero_mv, 0,
3511 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
3514 vp9_update_noise_estimate(cpi);
3516 if (cpi->oxcf.pass == 0 &&
3517 cpi->oxcf.mode == REALTIME &&
3518 cpi->oxcf.speed >= 5 &&
3519 cpi->resize_state == 0 &&
3520 cm->frame_type != KEY_FRAME &&
3521 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3522 cpi->oxcf.rc_mode == VPX_VBR))
3523 vp9_avg_source_sad(cpi);
3525 // For 1 pass SVC, since only ZEROMV is allowed for upsampled reference
3526 // frame (i.e, svc->force_zero_mode_spatial_ref = 0), we can avoid this
3527 // frame-level upsampling.
3528 if (frame_is_intra_only(cm) == 0 && !is_one_pass_cbr_svc(cpi)) {
3529 vp9_scale_references(cpi);
3532 set_size_independent_vars(cpi);
3533 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3535 if (cpi->oxcf.speed >= 5 &&
3536 cpi->oxcf.pass == 0 &&
3537 cpi->oxcf.rc_mode == VPX_CBR &&
3538 cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
3539 cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3540 cpi->use_skin_detection = 1;
3543 vp9_set_quantizer(cm, q);
3544 vp9_set_variance_partition_thresholds(cpi, q);
3548 suppress_active_map(cpi);
3549 // Variance adaptive and in frame q adjustment experiments are mutually
3551 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3552 vp9_vaq_frame_setup(cpi);
3553 } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3554 vp9_360aq_frame_setup(cpi);
3555 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3556 vp9_setup_in_frame_q_adj(cpi);
3557 } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3558 vp9_cyclic_refresh_setup(cpi);
3560 apply_active_map(cpi);
3562 // transform / motion compensation build reconstruction frame
3563 vp9_encode_frame(cpi);
3565 // Check if we should drop this frame because of high overshoot.
3566 // Only for frames where high temporal-source sad is detected.
3567 if (cpi->oxcf.pass == 0 &&
3568 cpi->oxcf.rc_mode == VPX_CBR &&
3569 cpi->resize_state == 0 &&
3570 cm->frame_type != KEY_FRAME &&
3571 cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
3572 cpi->rc.high_source_sad == 1) {
3574 // Get an estimate of the encoded frame size.
3575 save_coding_context(cpi);
3576 vp9_pack_bitstream(cpi, dest, size);
3577 restore_coding_context(cpi);
3578 frame_size = (int)(*size) << 3;
3579 // Check if encoded frame will overshoot too much, and if so, set the q and
3580 // adjust some rate control parameters, and return to re-encode the frame.
3581 if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) {
3582 vpx_clear_system_state();
3583 vp9_set_quantizer(cm, q);
3584 vp9_set_variance_partition_thresholds(cpi, q);
3585 suppress_active_map(cpi);
3586 // Turn-off cyclic refresh for re-encoded frame.
3587 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3588 unsigned char *const seg_map = cpi->segmentation_map;
3589 memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
3590 vp9_disable_segmentation(&cm->seg);
3592 apply_active_map(cpi);
3593 vp9_encode_frame(cpi);
3597 // Update some stats from cyclic refresh, and check if we should not update
3598 // golden reference, for non-SVC 1 pass CBR.
3599 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
3600 cm->frame_type != KEY_FRAME &&
3602 cpi->ext_refresh_frame_flags_pending == 0 &&
3603 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR))
3604 vp9_cyclic_refresh_check_golden_update(cpi);
3606 // Update the skip mb flag probabilities based on the distribution
3607 // seen in the last encoder iteration.
3608 // update_base_skip_probs(cpi);
3609 vpx_clear_system_state();
3612 static void encode_with_recode_loop(VP9_COMP *cpi,
3615 VP9_COMMON *const cm = &cpi->common;
3616 RATE_CONTROL *const rc = &cpi->rc;
3617 int bottom_index, top_index;
3619 int loop_at_this_size = 0;
3621 int overshoot_seen = 0;
3622 int undershoot_seen = 0;
3623 int frame_over_shoot_limit;
3624 int frame_under_shoot_limit;
3625 int q = 0, q_low = 0, q_high = 0;
3627 set_size_independent_vars(cpi);
3630 vpx_clear_system_state();
3632 set_frame_size(cpi);
3634 if (loop_count == 0 || cpi->resize_pending != 0) {
3635 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3637 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3638 set_mv_search_params(cpi);
3640 // Reset the loop state for new frame size.
3642 undershoot_seen = 0;
3644 // Reconfiguration for change in frame size has concluded.
3645 cpi->resize_pending = 0;
3647 q_low = bottom_index;
3650 loop_at_this_size = 0;
3653 // Decide frame size bounds first time through.
3654 if (loop_count == 0) {
3655 vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
3656 &frame_under_shoot_limit,
3657 &frame_over_shoot_limit);
3660 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
3661 &cpi->scaled_source,
3662 (cpi->oxcf.pass == 0));
3664 if (cpi->unscaled_last_source != NULL)
3665 cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
3666 &cpi->scaled_last_source,
3667 (cpi->oxcf.pass == 0));
3669 if (frame_is_intra_only(cm) == 0) {
3670 if (loop_count > 0) {
3671 release_scaled_references(cpi);
3673 vp9_scale_references(cpi);
3676 vp9_set_quantizer(cm, q);
3678 if (loop_count == 0)
3681 // Variance adaptive and in frame q adjustment experiments are mutually
3683 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3684 vp9_vaq_frame_setup(cpi);
3685 } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3686 vp9_360aq_frame_setup(cpi);
3687 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3688 vp9_setup_in_frame_q_adj(cpi);
3691 // transform / motion compensation build reconstruction frame
3692 vp9_encode_frame(cpi);
3694 // Update the skip mb flag probabilities based on the distribution
3695 // seen in the last encoder iteration.
3696 // update_base_skip_probs(cpi);
3698 vpx_clear_system_state();
3700 // Dummy pack of the bitstream using up to date stats to get an
3701 // accurate estimate of output frame size to determine if we need
3703 if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
3704 save_coding_context(cpi);
3705 if (!cpi->sf.use_nonrd_pick_mode)
3706 vp9_pack_bitstream(cpi, dest, size);
3708 rc->projected_frame_size = (int)(*size) << 3;
3709 restore_coding_context(cpi);
3711 if (frame_over_shoot_limit == 0)
3712 frame_over_shoot_limit = 1;
3715 if (cpi->oxcf.rc_mode == VPX_Q) {
3718 if ((cm->frame_type == KEY_FRAME) &&
3719 rc->this_key_frame_forced &&
3720 (rc->projected_frame_size < rc->max_frame_bandwidth)) {
3724 int64_t high_err_target = cpi->ambient_err;
3725 int64_t low_err_target = cpi->ambient_err >> 1;
3727 #if CONFIG_VP9_HIGHBITDEPTH
3728 if (cm->use_highbitdepth) {
3729 kf_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3731 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3734 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3735 #endif // CONFIG_VP9_HIGHBITDEPTH
3737 // Prevent possible divide by zero error below for perfect KF
3740 // The key frame is not good enough or we can afford
3741 // to make it better without undue risk of popping.
3742 if ((kf_err > high_err_target &&
3743 rc->projected_frame_size <= frame_over_shoot_limit) ||
3744 (kf_err > low_err_target &&
3745 rc->projected_frame_size <= frame_under_shoot_limit)) {
3747 q_high = q > q_low ? q - 1 : q_low;
3750 q = (int)((q * high_err_target) / kf_err);
3751 q = VPXMIN(q, (q_high + q_low) >> 1);
3752 } else if (kf_err < low_err_target &&
3753 rc->projected_frame_size >= frame_under_shoot_limit) {
3754 // The key frame is much better than the previous frame
3756 q_low = q < q_high ? q + 1 : q_high;
3759 q = (int)((q * low_err_target) / kf_err);
3760 q = VPXMIN(q, (q_high + q_low + 1) >> 1);
3763 // Clamp Q to upper and lower limits:
3764 q = clamp(q, q_low, q_high);
3767 } else if (recode_loop_test(
3768 cpi, frame_over_shoot_limit, frame_under_shoot_limit,
3769 q, VPXMAX(q_high, top_index), bottom_index)) {
3770 // Is the projected frame size out of range and are we allowed
3771 // to attempt to recode.
3775 if (cpi->resize_pending == 1) {
3776 // Change in frame size so go back around the recode loop.
3777 cpi->rc.frame_size_selector =
3778 SCALE_STEP1 - cpi->rc.frame_size_selector;
3779 cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
3781 #if CONFIG_INTERNAL_STATS
3782 ++cpi->tot_recode_hits;
3789 // Frame size out of permitted range:
3790 // Update correction factor & compute new Q to try...
3792 // Frame is too large
3793 if (rc->projected_frame_size > rc->this_frame_target) {
3794 // Special case if the projected size is > the max allowed.
3795 if (rc->projected_frame_size >= rc->max_frame_bandwidth)
3796 q_high = rc->worst_quality;
3798 // Raise Qlow as to at least the current value
3799 q_low = q < q_high ? q + 1 : q_high;
3801 if (undershoot_seen || loop_at_this_size > 1) {
3802 // Update rate_correction_factor unless
3803 vp9_rc_update_rate_correction_factors(cpi);
3805 q = (q_high + q_low + 1) / 2;
3807 // Update rate_correction_factor unless
3808 vp9_rc_update_rate_correction_factors(cpi);
3810 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3811 bottom_index, VPXMAX(q_high, top_index));
3813 while (q < q_low && retries < 10) {
3814 vp9_rc_update_rate_correction_factors(cpi);
3815 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3816 bottom_index, VPXMAX(q_high, top_index));
3823 // Frame is too small
3824 q_high = q > q_low ? q - 1 : q_low;
3826 if (overshoot_seen || loop_at_this_size > 1) {
3827 vp9_rc_update_rate_correction_factors(cpi);
3828 q = (q_high + q_low) / 2;
3830 vp9_rc_update_rate_correction_factors(cpi);
3831 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3832 bottom_index, top_index);
3833 // Special case reset for qlow for constrained quality.
3834 // This should only trigger where there is very substantial
3835 // undershoot on a frame and the auto cq level is above
3836 // the user passsed in value.
3837 if (cpi->oxcf.rc_mode == VPX_CQ &&
3842 while (q > q_high && retries < 10) {
3843 vp9_rc_update_rate_correction_factors(cpi);
3844 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3845 bottom_index, top_index);
3850 undershoot_seen = 1;
3853 // Clamp Q to upper and lower limits:
3854 q = clamp(q, q_low, q_high);
3856 loop = (q != last_q);
3862 // Special case for overlay frame.
3863 if (rc->is_src_frame_alt_ref &&
3864 rc->projected_frame_size < rc->max_frame_bandwidth)
3869 ++loop_at_this_size;
3871 #if CONFIG_INTERNAL_STATS
3872 ++cpi->tot_recode_hits;
3878 static int get_ref_frame_flags(const VP9_COMP *cpi) {
3879 const int *const map = cpi->common.ref_frame_map;
3880 const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
3881 const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
3882 const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
3883 int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
3886 flags &= ~VP9_GOLD_FLAG;
3888 if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
3889 (cpi->svc.number_temporal_layers == 1 &&
3890 cpi->svc.number_spatial_layers == 1))
3891 flags &= ~VP9_GOLD_FLAG;
3894 flags &= ~VP9_ALT_FLAG;
3897 flags &= ~VP9_ALT_FLAG;
3902 static void set_ext_overrides(VP9_COMP *cpi) {
3903 // Overrides the defaults with the externally supplied values with
3904 // vp9_update_reference() and vp9_update_entropy() calls
3905 // Note: The overrides are valid only for the next frame passed
3906 // to encode_frame_to_data_rate() function
3907 if (cpi->ext_refresh_frame_context_pending) {
3908 cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
3909 cpi->ext_refresh_frame_context_pending = 0;
3911 if (cpi->ext_refresh_frame_flags_pending) {
3912 cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
3913 cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
3914 cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
3918 YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
3919 YV12_BUFFER_CONFIG *unscaled,
3920 YV12_BUFFER_CONFIG *scaled,
3921 YV12_BUFFER_CONFIG *scaled_temp) {
3922 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3923 cm->mi_rows * MI_SIZE != unscaled->y_height) {
3924 #if CONFIG_VP9_HIGHBITDEPTH
3925 scale_and_extend_frame(unscaled, scaled_temp, (int)cm->bit_depth);
3926 scale_and_extend_frame(scaled_temp, scaled, (int)cm->bit_depth);
3928 vp9_scale_and_extend_frame(unscaled, scaled_temp);
3929 vp9_scale_and_extend_frame(scaled_temp, scaled);
3930 #endif // CONFIG_VP9_HIGHBITDEPTH
3937 YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
3938 YV12_BUFFER_CONFIG *unscaled,
3939 YV12_BUFFER_CONFIG *scaled,
3940 int use_normative_scaler) {
3941 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3942 cm->mi_rows * MI_SIZE != unscaled->y_height) {
3943 #if CONFIG_VP9_HIGHBITDEPTH
3944 if (use_normative_scaler &&
3945 unscaled->y_width <= (scaled->y_width << 1) &&
3946 unscaled->y_height <= (scaled->y_height << 1))
3947 scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth);
3949 scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
3951 if (use_normative_scaler &&
3952 unscaled->y_width <= (scaled->y_width << 1) &&
3953 unscaled->y_height <= (scaled->y_height << 1))
3954 vp9_scale_and_extend_frame(unscaled, scaled);
3956 scale_and_extend_frame_nonnormative(unscaled, scaled);
3957 #endif // CONFIG_VP9_HIGHBITDEPTH
3964 static void set_arf_sign_bias(VP9_COMP *cpi) {
3965 VP9_COMMON *const cm = &cpi->common;
3968 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
3969 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
3970 arf_sign_bias = cpi->rc.source_alt_ref_active &&
3971 (!cpi->refresh_alt_ref_frame ||
3972 (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
3975 (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
3977 cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
3980 static int setup_interp_filter_search_mask(VP9_COMP *cpi) {
3981 INTERP_FILTER ifilter;
3982 int ref_total[MAX_REF_FRAMES] = {0};
3983 MV_REFERENCE_FRAME ref;
3985 if (cpi->common.last_frame_type == KEY_FRAME ||
3986 cpi->refresh_alt_ref_frame)
3988 for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
3989 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter)
3990 ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
3992 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) {
3993 if ((ref_total[LAST_FRAME] &&
3994 cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
3995 (ref_total[GOLDEN_FRAME] == 0 ||
3996 cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50
3997 < ref_total[GOLDEN_FRAME]) &&
3998 (ref_total[ALTREF_FRAME] == 0 ||
3999 cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50
4000 < ref_total[ALTREF_FRAME]))
4001 mask |= 1 << ifilter;
4006 static void encode_frame_to_data_rate(VP9_COMP *cpi,
4009 unsigned int *frame_flags) {
4010 VP9_COMMON *const cm = &cpi->common;
4011 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4012 struct segmentation *const seg = &cm->seg;
4015 set_ext_overrides(cpi);
4016 vpx_clear_system_state();
4018 // Set the arf sign bias for this frame.
4019 set_arf_sign_bias(cpi);
4021 // Set default state for segment based loop filter update flags.
4022 cm->lf.mode_ref_delta_update = 0;
4024 if (cpi->oxcf.pass == 2 &&
4025 cpi->sf.adaptive_interp_filter_search)
4026 cpi->sf.interp_filter_search_mask =
4027 setup_interp_filter_search_mask(cpi);
4029 // Set various flags etc to special state if it is a key frame.
4030 if (frame_is_intra_only(cm)) {
4031 // Reset the loop filter deltas and segmentation map.
4032 vp9_reset_segment_features(&cm->seg);
4034 // If segmentation is enabled force a map update for key frames.
4036 seg->update_map = 1;
4037 seg->update_data = 1;
4040 // The alternate reference frame cannot be active for a key frame.
4041 cpi->rc.source_alt_ref_active = 0;
4043 cm->error_resilient_mode = oxcf->error_resilient_mode;
4044 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
4046 // By default, encoder assumes decoder can use prev_mi.
4047 if (cm->error_resilient_mode) {
4048 cm->frame_parallel_decoding_mode = 1;
4049 cm->reset_frame_context = 0;
4050 cm->refresh_frame_context = 0;
4051 } else if (cm->intra_only) {
4052 // Only reset the current context.
4053 cm->reset_frame_context = 2;
4056 if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) {
4057 // Use context 0 for intra only empty frame, but the last frame context
4058 // for other empty frames.
4059 if (cpi->svc.encode_empty_frame_state == ENCODING) {
4060 if (cpi->svc.encode_intra_empty_frame != 0)
4061 cm->frame_context_idx = 0;
4063 cm->frame_context_idx = FRAME_CONTEXTS - 1;
4065 cm->frame_context_idx =
4066 cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
4067 cpi->svc.temporal_layer_id;
4070 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
4072 // The probs will be updated based on the frame type of its previous
4073 // frame if frame_parallel_decoding_mode is 0. The type may vary for
4074 // the frame after a key frame in base layer since we may drop enhancement
4075 // layers. So set frame_parallel_decoding_mode to 1 in this case.
4076 if (cm->frame_parallel_decoding_mode == 0) {
4077 if (cpi->svc.number_temporal_layers == 1) {
4078 if (cpi->svc.spatial_layer_id == 0 &&
4079 cpi->svc.layer_context[0].last_frame_type == KEY_FRAME)
4080 cm->frame_parallel_decoding_mode = 1;
4081 } else if (cpi->svc.spatial_layer_id == 0) {
4082 // Find the 2nd frame in temporal base layer and 1st frame in temporal
4083 // enhancement layers from the key frame.
4085 for (i = 0; i < cpi->svc.number_temporal_layers; ++i) {
4086 if (cpi->svc.layer_context[0].frames_from_key_frame == 1 << i) {
4087 cm->frame_parallel_decoding_mode = 1;
4095 // For 1 pass CBR, check if we are dropping this frame.
4096 // For spatial layers, for now only check for frame-dropping on first spatial
4097 // layer, and if decision is to drop, we drop whole super-frame.
4098 if (oxcf->pass == 0 &&
4099 oxcf->rc_mode == VPX_CBR &&
4100 cm->frame_type != KEY_FRAME) {
4101 if (vp9_rc_drop_frame(cpi) ||
4102 (is_one_pass_cbr_svc(cpi) && cpi->svc.rc_drop_superframe == 1)) {
4103 vp9_rc_postencode_update_drop_frame(cpi);
4104 ++cm->current_video_frame;
4105 cpi->ext_refresh_frame_flags_pending = 0;
4106 cpi->svc.rc_drop_superframe = 1;
4107 // TODO(marpan): Advancing the svc counters on dropped frames can break
4108 // the referencing scheme for the fixed svc patterns defined in
4109 // vp9_one_pass_cbr_svc_start_layer(). Look into fixing this issue, but
4110 // for now, don't advance the svc frame counters on dropped frame.
4111 // if (cpi->use_svc)
4112 // vp9_inc_frame_in_layer(cpi);
4117 vpx_clear_system_state();
4119 #if CONFIG_INTERNAL_STATS
4120 memset(cpi->mode_chosen_counts, 0,
4121 MAX_MODES * sizeof(*cpi->mode_chosen_counts));
4124 if (cpi->sf.recode_loop == DISALLOW_RECODE) {
4125 encode_without_recode_loop(cpi, size, dest);
4127 encode_with_recode_loop(cpi, size, dest);
4130 #if CONFIG_VP9_TEMPORAL_DENOISING
4131 #ifdef OUTPUT_YUV_DENOISED
4132 if (oxcf->noise_sensitivity > 0) {
4133 vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME],
4138 #ifdef OUTPUT_YUV_SKINMAP
4139 if (cpi->common.current_video_frame > 1) {
4140 vp9_compute_skin_map(cpi, yuv_skinmap_file);
4144 // Special case code to reduce pulsing when key frames are forced at a
4145 // fixed interval. Note the reconstruction error if it is the frame before
4146 // the force key frame
4147 if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
4148 #if CONFIG_VP9_HIGHBITDEPTH
4149 if (cm->use_highbitdepth) {
4150 cpi->ambient_err = vp9_highbd_get_y_sse(cpi->Source,
4151 get_frame_new_buffer(cm));
4153 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4156 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4157 #endif // CONFIG_VP9_HIGHBITDEPTH
4160 // If the encoder forced a KEY_FRAME decision
4161 if (cm->frame_type == KEY_FRAME)
4162 cpi->refresh_last_frame = 1;
4164 cm->frame_to_show = get_frame_new_buffer(cm);
4165 cm->frame_to_show->color_space = cm->color_space;
4166 cm->frame_to_show->color_range = cm->color_range;
4167 cm->frame_to_show->render_width = cm->render_width;
4168 cm->frame_to_show->render_height = cm->render_height;
4170 // Pick the loop filter level for the frame.
4171 loopfilter_frame(cpi, cm);
4173 // build the bitstream
4174 vp9_pack_bitstream(cpi, dest, size);
4176 if (cm->seg.update_map)
4177 update_reference_segmentation_map(cpi);
4179 if (frame_is_intra_only(cm) == 0) {
4180 release_scaled_references(cpi);
4182 vp9_update_reference_frames(cpi);
4184 for (t = TX_4X4; t <= TX_32X32; t++)
4185 full_to_model_counts(cpi->td.counts->coef[t],
4186 cpi->td.rd_counts.coef_counts[t]);
4188 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode)
4189 vp9_adapt_coef_probs(cm);
4191 if (!frame_is_intra_only(cm)) {
4192 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
4193 vp9_adapt_mode_probs(cm);
4194 vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
4198 cpi->ext_refresh_frame_flags_pending = 0;
4200 if (cpi->refresh_golden_frame == 1)
4201 cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
4203 cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
4205 if (cpi->refresh_alt_ref_frame == 1)
4206 cpi->frame_flags |= FRAMEFLAGS_ALTREF;
4208 cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
4210 cpi->ref_frame_flags = get_ref_frame_flags(cpi);
4212 cm->last_frame_type = cm->frame_type;
4214 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
4215 vp9_rc_postencode_update(cpi, *size);
4218 output_frame_level_debug_stats(cpi);
4221 if (cm->frame_type == KEY_FRAME) {
4222 // Tell the caller that the frame was coded as a key frame
4223 *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
4225 *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
4228 // Clear the one shot update flags for segmentation map and mode/ref loop
4230 cm->seg.update_map = 0;
4231 cm->seg.update_data = 0;
4232 cm->lf.mode_ref_delta_update = 0;
4234 // keep track of the last coded dimensions
4235 cm->last_width = cm->width;
4236 cm->last_height = cm->height;
4238 // reset to normal state now that we are done.
4239 if (!cm->show_existing_frame)
4240 cm->last_show_frame = cm->show_frame;
4242 if (cm->show_frame) {
4243 vp9_swap_mi_and_prev_mi(cm);
4244 // Don't increment frame counters if this was an altref buffer
4245 // update not a real frame
4246 ++cm->current_video_frame;
4248 vp9_inc_frame_in_layer(cpi);
4250 cm->prev_frame = cm->cur_frame;
4253 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
4254 cpi->svc.number_temporal_layers +
4255 cpi->svc.temporal_layer_id].last_frame_type =
4259 static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
4260 unsigned int *frame_flags) {
4261 vp9_rc_get_svc_params(cpi);
4262 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4265 static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
4266 unsigned int *frame_flags) {
4267 if (cpi->oxcf.rc_mode == VPX_CBR) {
4268 vp9_rc_get_one_pass_cbr_params(cpi);
4270 vp9_rc_get_one_pass_vbr_params(cpi);
4272 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4275 static void Pass2Encode(VP9_COMP *cpi, size_t *size,
4276 uint8_t *dest, unsigned int *frame_flags) {
4277 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
4278 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4280 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
4281 vp9_twopass_postencode_update(cpi);
4284 static void init_ref_frame_bufs(VP9_COMMON *cm) {
4286 BufferPool *const pool = cm->buffer_pool;
4287 cm->new_fb_idx = INVALID_IDX;
4288 for (i = 0; i < REF_FRAMES; ++i) {
4289 cm->ref_frame_map[i] = INVALID_IDX;
4290 pool->frame_bufs[i].ref_count = 0;
4294 static void check_initial_width(VP9_COMP *cpi,
4295 #if CONFIG_VP9_HIGHBITDEPTH
4296 int use_highbitdepth,
4298 int subsampling_x, int subsampling_y) {
4299 VP9_COMMON *const cm = &cpi->common;
4301 if (!cpi->initial_width ||
4302 #if CONFIG_VP9_HIGHBITDEPTH
4303 cm->use_highbitdepth != use_highbitdepth ||
4305 cm->subsampling_x != subsampling_x ||
4306 cm->subsampling_y != subsampling_y) {
4307 cm->subsampling_x = subsampling_x;
4308 cm->subsampling_y = subsampling_y;
4309 #if CONFIG_VP9_HIGHBITDEPTH
4310 cm->use_highbitdepth = use_highbitdepth;
4313 alloc_raw_frame_buffers(cpi);
4314 init_ref_frame_bufs(cm);
4315 alloc_util_frame_buffers(cpi);
4317 init_motion_estimation(cpi); // TODO(agrange) This can be removed.
4319 cpi->initial_width = cm->width;
4320 cpi->initial_height = cm->height;
4321 cpi->initial_mbs = cm->MBs;
4325 int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
4326 YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
4328 VP9_COMMON *const cm = &cpi->common;
4329 struct vpx_usec_timer timer;
4331 const int subsampling_x = sd->subsampling_x;
4332 const int subsampling_y = sd->subsampling_y;
4333 #if CONFIG_VP9_HIGHBITDEPTH
4334 const int use_highbitdepth = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
4337 #if CONFIG_VP9_HIGHBITDEPTH
4338 check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
4340 check_initial_width(cpi, subsampling_x, subsampling_y);
4341 #endif // CONFIG_VP9_HIGHBITDEPTH
4343 #if CONFIG_VP9_TEMPORAL_DENOISING
4344 setup_denoiser_buffer(cpi);
4346 vpx_usec_timer_start(&timer);
4348 if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
4349 #if CONFIG_VP9_HIGHBITDEPTH
4351 #endif // CONFIG_VP9_HIGHBITDEPTH
4354 vpx_usec_timer_mark(&timer);
4355 cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
4357 if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
4358 (subsampling_x != 1 || subsampling_y != 1)) {
4359 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4360 "Non-4:2:0 color format requires profile 1 or 3");
4363 if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
4364 (subsampling_x == 1 && subsampling_y == 1)) {
4365 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4366 "4:2:0 color format requires profile 0 or 2");
4374 static int frame_is_reference(const VP9_COMP *cpi) {
4375 const VP9_COMMON *cm = &cpi->common;
4377 return cm->frame_type == KEY_FRAME ||
4378 cpi->refresh_last_frame ||
4379 cpi->refresh_golden_frame ||
4380 cpi->refresh_alt_ref_frame ||
4381 cm->refresh_frame_context ||
4382 cm->lf.mode_ref_delta_update ||
4383 cm->seg.update_map ||
4384 cm->seg.update_data;
4387 static void adjust_frame_rate(VP9_COMP *cpi,
4388 const struct lookahead_entry *source) {
4389 int64_t this_duration;
4392 if (source->ts_start == cpi->first_time_stamp_ever) {
4393 this_duration = source->ts_end - source->ts_start;
4396 int64_t last_duration = cpi->last_end_time_stamp_seen
4397 - cpi->last_time_stamp_seen;
4399 this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
4401 // do a step update if the duration changes by 10%
4403 step = (int)((this_duration - last_duration) * 10 / last_duration);
4406 if (this_duration) {
4408 vp9_new_framerate(cpi, 10000000.0 / this_duration);
4410 // Average this frame's rate into the last second's average
4411 // frame rate. If we haven't seen 1 second yet, then average
4412 // over the whole interval seen.
4413 const double interval = VPXMIN(
4414 (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
4415 double avg_duration = 10000000.0 / cpi->framerate;
4416 avg_duration *= (interval - avg_duration + this_duration);
4417 avg_duration /= interval;
4419 vp9_new_framerate(cpi, 10000000.0 / avg_duration);
4422 cpi->last_time_stamp_seen = source->ts_start;
4423 cpi->last_end_time_stamp_seen = source->ts_end;
4426 // Returns 0 if this is not an alt ref else the offset of the source frame
4427 // used as the arf midpoint.
4428 static int get_arf_src_index(VP9_COMP *cpi) {
4429 RATE_CONTROL *const rc = &cpi->rc;
4430 int arf_src_index = 0;
4431 if (is_altref_enabled(cpi)) {
4432 if (cpi->oxcf.pass == 2) {
4433 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4434 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
4435 arf_src_index = gf_group->arf_src_offset[gf_group->index];
4437 } else if (rc->source_alt_ref_pending) {
4438 arf_src_index = rc->frames_till_gf_update_due;
4441 return arf_src_index;
4444 static void check_src_altref(VP9_COMP *cpi,
4445 const struct lookahead_entry *source) {
4446 RATE_CONTROL *const rc = &cpi->rc;
4448 if (cpi->oxcf.pass == 2) {
4449 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4450 rc->is_src_frame_alt_ref =
4451 (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
4453 rc->is_src_frame_alt_ref = cpi->alt_ref_source &&
4454 (source == cpi->alt_ref_source);
4457 if (rc->is_src_frame_alt_ref) {
4458 // Current frame is an ARF overlay frame.
4459 cpi->alt_ref_source = NULL;
4461 // Don't refresh the last buffer for an ARF overlay frame. It will
4462 // become the GF so preserve last as an alternative prediction option.
4463 cpi->refresh_last_frame = 0;
4467 #if CONFIG_INTERNAL_STATS
4468 extern double vp9_get_blockiness(const uint8_t *img1, int img1_pitch,
4469 const uint8_t *img2, int img2_pitch,
4470 int width, int height);
4472 static void adjust_image_stat(double y, double u, double v, double all,
4477 s->stat[ALL] += all;
4478 s->worst = VPXMIN(s->worst, all);
4480 #endif // CONFIG_INTERNAL_STATS
4482 static void update_level_info(VP9_COMP *cpi, size_t *size, int arf_src_index) {
4483 VP9_COMMON *const cm = &cpi->common;
4484 Vp9LevelInfo *const level_info = &cpi->level_info;
4485 Vp9LevelSpec *const level_spec = &level_info->level_spec;
4486 Vp9LevelStats *const level_stats = &level_info->level_stats;
4488 uint64_t luma_samples, dur_end;
4489 const uint32_t luma_pic_size = cm->width * cm->height;
4490 double cpb_data_size;
4492 vpx_clear_system_state();
4494 // update level_stats
4495 level_stats->total_compressed_size += *size;
4496 if (cm->show_frame) {
4497 level_stats->total_uncompressed_size +=
4499 2 * (luma_pic_size >> (cm->subsampling_x + cm->subsampling_y));
4500 level_stats->time_encoded =
4501 (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
4502 (double)TICKS_PER_SEC;
4505 if (arf_src_index > 0) {
4506 if (!level_stats->seen_first_altref) {
4507 level_stats->seen_first_altref = 1;
4508 } else if (level_stats->frames_since_last_altref <
4509 level_spec->min_altref_distance) {
4510 level_spec->min_altref_distance = level_stats->frames_since_last_altref;
4512 level_stats->frames_since_last_altref = 0;
4514 ++level_stats->frames_since_last_altref;
4517 if (level_stats->frame_window_buffer.len < FRAME_WINDOW_SIZE - 1) {
4518 idx = (level_stats->frame_window_buffer.start +
4519 level_stats->frame_window_buffer.len++) % FRAME_WINDOW_SIZE;
4521 idx = level_stats->frame_window_buffer.start;
4522 level_stats->frame_window_buffer.start = (idx + 1) % FRAME_WINDOW_SIZE;
4524 level_stats->frame_window_buffer.buf[idx].ts = cpi->last_time_stamp_seen;
4525 level_stats->frame_window_buffer.buf[idx].size = (uint32_t)(*size);
4526 level_stats->frame_window_buffer.buf[idx].luma_samples = luma_pic_size;
4528 if (cm->frame_type == KEY_FRAME) {
4529 level_stats->ref_refresh_map = 0;
4532 level_stats->ref_refresh_map |= vp9_get_refresh_mask(cpi);
4533 // Also need to consider the case where the encoder refers to a buffer
4534 // that has been implicitly refreshed after encoding a keyframe.
4535 if (!cm->intra_only) {
4536 level_stats->ref_refresh_map |= (1 << cpi->lst_fb_idx);
4537 level_stats->ref_refresh_map |= (1 << cpi->gld_fb_idx);
4538 level_stats->ref_refresh_map |= (1 << cpi->alt_fb_idx);
4540 for (i = 0; i < REF_FRAMES; ++i) {
4541 count += (level_stats->ref_refresh_map >> i) & 1;
4543 if (count > level_spec->max_ref_frame_buffers) {
4544 level_spec->max_ref_frame_buffers = count;
4548 // update average_bitrate
4549 level_spec->average_bitrate =
4550 (double)level_stats->total_compressed_size / 125.0 /
4551 level_stats->time_encoded;
4553 // update max_luma_sample_rate
4555 for (i = 0; i < level_stats->frame_window_buffer.len; ++i) {
4556 idx = (level_stats->frame_window_buffer.start +
4557 level_stats->frame_window_buffer.len - 1 - i) % FRAME_WINDOW_SIZE;
4559 dur_end = level_stats->frame_window_buffer.buf[idx].ts;
4561 if (dur_end - level_stats->frame_window_buffer.buf[idx].ts >=
4565 luma_samples += level_stats->frame_window_buffer.buf[idx].luma_samples;
4567 if (luma_samples > level_spec->max_luma_sample_rate) {
4568 level_spec->max_luma_sample_rate = luma_samples;
4571 // update max_cpb_size
4573 for (i = 0; i < CPB_WINDOW_SIZE; ++i) {
4574 if (i >= level_stats->frame_window_buffer.len) break;
4575 idx = (level_stats->frame_window_buffer.start +
4576 level_stats->frame_window_buffer.len - 1 - i) % FRAME_WINDOW_SIZE;
4577 cpb_data_size += level_stats->frame_window_buffer.buf[idx].size;
4579 cpb_data_size = cpb_data_size / 125.0;
4580 if (cpb_data_size > level_spec->max_cpb_size) {
4581 level_spec->max_cpb_size = cpb_data_size;
4584 // update max_luma_picture_size
4585 if (luma_pic_size > level_spec->max_luma_picture_size) {
4586 level_spec->max_luma_picture_size = luma_pic_size;
4589 // update compression_ratio
4590 level_spec->compression_ratio =
4591 (double)level_stats->total_uncompressed_size * cm->bit_depth /
4592 level_stats->total_compressed_size / 8.0;
4594 // update max_col_tiles
4595 if (level_spec->max_col_tiles < (1 << cm->log2_tile_cols)) {
4596 level_spec->max_col_tiles = (1 << cm->log2_tile_cols);
4600 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
4601 size_t *size, uint8_t *dest,
4602 int64_t *time_stamp, int64_t *time_end, int flush) {
4603 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4604 VP9_COMMON *const cm = &cpi->common;
4605 BufferPool *const pool = cm->buffer_pool;
4606 RATE_CONTROL *const rc = &cpi->rc;
4607 struct vpx_usec_timer cmptimer;
4608 YV12_BUFFER_CONFIG *force_src_buffer = NULL;
4609 struct lookahead_entry *last_source = NULL;
4610 struct lookahead_entry *source = NULL;
4614 if (is_two_pass_svc(cpi)) {
4615 #if CONFIG_SPATIAL_SVC
4616 vp9_svc_start_frame(cpi);
4617 // Use a small empty frame instead of a real frame
4618 if (cpi->svc.encode_empty_frame_state == ENCODING)
4619 source = &cpi->svc.empty_frame;
4621 if (oxcf->pass == 2)
4622 vp9_restore_layer_context(cpi);
4623 } else if (is_one_pass_cbr_svc(cpi)) {
4624 vp9_one_pass_cbr_svc_start_layer(cpi);
4627 vpx_usec_timer_start(&cmptimer);
4629 vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
4631 // Is multi-arf enabled.
4632 // Note that at the moment multi_arf is only configured for 2 pass VBR and
4633 // will not work properly with svc.
4634 if ((oxcf->pass == 2) && !cpi->use_svc &&
4635 (cpi->oxcf.enable_auto_arf > 1))
4636 cpi->multi_arf_allowed = 1;
4638 cpi->multi_arf_allowed = 0;
4641 cm->reset_frame_context = 0;
4642 cm->refresh_frame_context = 1;
4643 if (!is_one_pass_cbr_svc(cpi)) {
4644 cpi->refresh_last_frame = 1;
4645 cpi->refresh_golden_frame = 0;
4646 cpi->refresh_alt_ref_frame = 0;
4649 // Should we encode an arf frame.
4650 arf_src_index = get_arf_src_index(cpi);
4652 // Skip alt frame if we encode the empty frame
4653 if (is_two_pass_svc(cpi) && source != NULL)
4656 if (arf_src_index) {
4657 for (i = 0; i <= arf_src_index; ++i) {
4658 struct lookahead_entry *e = vp9_lookahead_peek(cpi->lookahead, i);
4659 // Avoid creating an alt-ref if there's a forced keyframe pending.
4662 } else if (e->flags == VPX_EFLAG_FORCE_KF) {
4670 if (arf_src_index) {
4671 assert(arf_src_index <= rc->frames_to_key);
4673 if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
4674 cpi->alt_ref_source = source;
4676 #if CONFIG_SPATIAL_SVC
4677 if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) {
4679 // Reference a hidden frame from a lower layer
4680 for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) {
4681 if (oxcf->ss_enable_auto_arf[i]) {
4682 cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx;
4687 cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1;
4690 if ((oxcf->arnr_max_frames > 0) && (oxcf->arnr_strength > 0)) {
4691 // Produce the filtered ARF frame.
4692 vp9_temporal_filter(cpi, arf_src_index);
4693 vpx_extend_frame_borders(&cpi->alt_ref_buffer);
4694 force_src_buffer = &cpi->alt_ref_buffer;
4699 cpi->refresh_alt_ref_frame = 1;
4700 cpi->refresh_golden_frame = 0;
4701 cpi->refresh_last_frame = 0;
4702 rc->is_src_frame_alt_ref = 0;
4703 rc->source_alt_ref_pending = 0;
4705 rc->source_alt_ref_pending = 0;
4710 // Get last frame source.
4711 if (cm->current_video_frame > 0) {
4712 if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
4716 // Read in the source frame.
4718 source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
4720 source = vp9_lookahead_pop(cpi->lookahead, flush);
4722 if (source != NULL) {
4725 // if the flags indicate intra frame, but if the current picture is for
4726 // non-zero spatial layer, it should not be an intra picture.
4727 // TODO(Won Kap): this needs to change if per-layer intra frame is
4729 if ((source->flags & VPX_EFLAG_FORCE_KF) &&
4730 cpi->svc.spatial_layer_id > cpi->svc.first_spatial_layer_to_encode) {
4731 source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF);
4734 // Check to see if the frame should be encoded as an arf overlay.
4735 check_src_altref(cpi, source);
4740 cpi->un_scaled_source = cpi->Source = force_src_buffer ? force_src_buffer
4743 cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
4745 *time_stamp = source->ts_start;
4746 *time_end = source->ts_end;
4747 *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
4751 if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
4752 vp9_end_first_pass(cpi); /* get last stats packet */
4753 cpi->twopass.first_pass_done = 1;
4758 if (source->ts_start < cpi->first_time_stamp_ever) {
4759 cpi->first_time_stamp_ever = source->ts_start;
4760 cpi->last_end_time_stamp_seen = source->ts_start;
4763 // Clear down mmx registers
4764 vpx_clear_system_state();
4766 // adjust frame rates based on timestamps given
4767 if (cm->show_frame) {
4768 adjust_frame_rate(cpi, source);
4771 if (is_one_pass_cbr_svc(cpi)) {
4772 vp9_update_temporal_layer_framerate(cpi);
4773 vp9_restore_layer_context(cpi);
4776 // Find a free buffer for the new frame, releasing the reference previously
4778 if (cm->new_fb_idx != INVALID_IDX) {
4779 --pool->frame_bufs[cm->new_fb_idx].ref_count;
4781 cm->new_fb_idx = get_free_fb(cm);
4783 if (cm->new_fb_idx == INVALID_IDX)
4786 cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
4788 if (!cpi->use_svc && cpi->multi_arf_allowed) {
4789 if (cm->frame_type == KEY_FRAME) {
4790 init_buffer_indices(cpi);
4791 } else if (oxcf->pass == 2) {
4792 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4793 cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
4797 // Start with a 0 size frame.
4800 cpi->frame_flags = *frame_flags;
4802 if ((oxcf->pass == 2) &&
4804 (is_two_pass_svc(cpi) &&
4805 cpi->svc.encode_empty_frame_state != ENCODING))) {
4806 vp9_rc_get_second_pass_params(cpi);
4807 } else if (oxcf->pass == 1) {
4808 set_frame_size(cpi);
4811 if (cpi->oxcf.pass != 0 ||
4813 frame_is_intra_only(cm) == 1) {
4814 for (i = 0; i < MAX_REF_FRAMES; ++i)
4815 cpi->scaled_ref_idx[i] = INVALID_IDX;
4818 if (oxcf->pass == 1 &&
4819 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4820 const int lossless = is_lossless_requested(oxcf);
4821 #if CONFIG_VP9_HIGHBITDEPTH
4822 if (cpi->oxcf.use_highbitdepth)
4823 cpi->td.mb.fwd_txm4x4 = lossless ?
4824 vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
4826 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4827 cpi->td.mb.highbd_itxm_add = lossless ? vp9_highbd_iwht4x4_add :
4828 vp9_highbd_idct4x4_add;
4830 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4831 #endif // CONFIG_VP9_HIGHBITDEPTH
4832 cpi->td.mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
4833 vp9_first_pass(cpi, source);
4834 } else if (oxcf->pass == 2 &&
4835 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4836 Pass2Encode(cpi, size, dest, frame_flags);
4837 } else if (cpi->use_svc) {
4838 SvcEncode(cpi, size, dest, frame_flags);
4841 Pass0Encode(cpi, size, dest, frame_flags);
4844 if (cm->refresh_frame_context)
4845 cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
4847 // No frame encoded, or frame was dropped, release scaled references.
4848 if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
4849 release_scaled_references(cpi);
4853 cpi->droppable = !frame_is_reference(cpi);
4856 // Save layer specific state.
4857 if (is_one_pass_cbr_svc(cpi) ||
4858 ((cpi->svc.number_temporal_layers > 1 ||
4859 cpi->svc.number_spatial_layers > 1) &&
4861 vp9_save_layer_context(cpi);
4864 vpx_usec_timer_mark(&cmptimer);
4865 cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
4867 if (cpi->b_calculate_psnr && oxcf->pass != 1 && cm->show_frame)
4868 generate_psnr_packet(cpi);
4870 if (cpi->keep_level_stats && oxcf->pass != 1)
4871 update_level_info(cpi, size, arf_src_index);
4873 #if CONFIG_INTERNAL_STATS
4875 if (oxcf->pass != 1) {
4876 double samples = 0.0;
4877 cpi->bytes += (int)(*size);
4879 if (cm->show_frame) {
4882 if (cpi->b_calculate_psnr) {
4883 YV12_BUFFER_CONFIG *orig = cpi->Source;
4884 YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
4885 YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer;
4887 #if CONFIG_VP9_HIGHBITDEPTH
4888 calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd,
4889 cpi->oxcf.input_bit_depth);
4891 calc_psnr(orig, recon, &psnr);
4892 #endif // CONFIG_VP9_HIGHBITDEPTH
4894 adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3],
4895 psnr.psnr[0], &cpi->psnr);
4896 cpi->total_sq_error += psnr.sse[0];
4897 cpi->total_samples += psnr.samples[0];
4898 samples = psnr.samples[0];
4902 double frame_ssim2 = 0, weight = 0;
4903 #if CONFIG_VP9_POSTPROC
4904 if (vpx_alloc_frame_buffer(&cm->post_proc_buffer,
4905 recon->y_crop_width, recon->y_crop_height,
4906 cm->subsampling_x, cm->subsampling_y,
4907 #if CONFIG_VP9_HIGHBITDEPTH
4908 cm->use_highbitdepth,
4910 VP9_ENC_BORDER_IN_PIXELS,
4911 cm->byte_alignment) < 0) {
4912 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
4913 "Failed to allocate post processing buffer");
4916 vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer,
4917 cm->lf.filter_level * 10 / 6);
4919 vpx_clear_system_state();
4921 #if CONFIG_VP9_HIGHBITDEPTH
4922 calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd,
4923 cpi->oxcf.input_bit_depth);
4925 calc_psnr(orig, pp, &psnr2);
4926 #endif // CONFIG_VP9_HIGHBITDEPTH
4928 cpi->totalp_sq_error += psnr2.sse[0];
4929 cpi->totalp_samples += psnr2.samples[0];
4930 adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3],
4931 psnr2.psnr[0], &cpi->psnrp);
4933 #if CONFIG_VP9_HIGHBITDEPTH
4934 if (cm->use_highbitdepth) {
4935 frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight,
4936 (int)cm->bit_depth);
4938 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4941 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4942 #endif // CONFIG_VP9_HIGHBITDEPTH
4944 cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2);
4945 cpi->summed_quality += frame_ssim2 * weight;
4946 cpi->summed_weights += weight;
4948 #if CONFIG_VP9_HIGHBITDEPTH
4949 if (cm->use_highbitdepth) {
4950 frame_ssim2 = vpx_highbd_calc_ssim(
4951 orig, &cm->post_proc_buffer, &weight, (int)cm->bit_depth);
4953 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4956 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4957 #endif // CONFIG_VP9_HIGHBITDEPTH
4959 cpi->summedp_quality += frame_ssim2 * weight;
4960 cpi->summedp_weights += weight;
4963 FILE *f = fopen("q_used.stt", "a");
4964 fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
4965 cpi->common.current_video_frame, y2, u2, v2,
4966 frame_psnr2, frame_ssim2);
4972 if (cpi->b_calculate_blockiness) {
4973 #if CONFIG_VP9_HIGHBITDEPTH
4974 if (!cm->use_highbitdepth)
4977 double frame_blockiness = vp9_get_blockiness(
4978 cpi->Source->y_buffer, cpi->Source->y_stride,
4979 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4980 cpi->Source->y_width, cpi->Source->y_height);
4981 cpi->worst_blockiness =
4982 VPXMAX(cpi->worst_blockiness, frame_blockiness);
4983 cpi->total_blockiness += frame_blockiness;
4987 if (cpi->b_calculate_consistency) {
4988 #if CONFIG_VP9_HIGHBITDEPTH
4989 if (!cm->use_highbitdepth)
4992 double this_inconsistency = vpx_get_ssim_metrics(
4993 cpi->Source->y_buffer, cpi->Source->y_stride,
4994 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4995 cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
4998 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
4999 double consistency = vpx_sse_to_psnr(samples, peak,
5000 (double)cpi->total_inconsistency);
5001 if (consistency > 0.0)
5002 cpi->worst_consistency =
5003 VPXMIN(cpi->worst_consistency, consistency);
5004 cpi->total_inconsistency += this_inconsistency;
5008 if (cpi->b_calculate_ssimg) {
5009 double y, u, v, frame_all;
5010 #if CONFIG_VP9_HIGHBITDEPTH
5011 if (cm->use_highbitdepth) {
5012 frame_all = vpx_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y,
5013 &u, &v, (int)cm->bit_depth);
5015 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u,
5019 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v);
5020 #endif // CONFIG_VP9_HIGHBITDEPTH
5021 adjust_image_stat(y, u, v, frame_all, &cpi->ssimg);
5023 #if CONFIG_VP9_HIGHBITDEPTH
5024 if (!cm->use_highbitdepth)
5027 double y, u, v, frame_all;
5028 frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
5030 adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
5031 /* TODO(JBB): add 10/12 bit support */
5033 #if CONFIG_VP9_HIGHBITDEPTH
5034 if (!cm->use_highbitdepth)
5037 double y, u, v, frame_all;
5038 frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
5039 adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
5046 if (is_two_pass_svc(cpi)) {
5047 if (cpi->svc.encode_empty_frame_state == ENCODING) {
5048 cpi->svc.encode_empty_frame_state = ENCODED;
5049 cpi->svc.encode_intra_empty_frame = 0;
5052 if (cm->show_frame) {
5053 ++cpi->svc.spatial_layer_to_encode;
5054 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
5055 cpi->svc.spatial_layer_to_encode = 0;
5057 // May need the empty frame after an visible frame.
5058 cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE;
5060 } else if (is_one_pass_cbr_svc(cpi)) {
5061 if (cm->show_frame) {
5062 ++cpi->svc.spatial_layer_to_encode;
5063 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
5064 cpi->svc.spatial_layer_to_encode = 0;
5067 vpx_clear_system_state();
5071 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
5072 vp9_ppflags_t *flags) {
5073 VP9_COMMON *cm = &cpi->common;
5074 #if !CONFIG_VP9_POSTPROC
5078 if (!cm->show_frame) {
5082 #if CONFIG_VP9_POSTPROC
5083 ret = vp9_post_proc_frame(cm, dest, flags);
5085 if (cm->frame_to_show) {
5086 *dest = *cm->frame_to_show;
5087 dest->y_width = cm->width;
5088 dest->y_height = cm->height;
5089 dest->uv_width = cm->width >> cm->subsampling_x;
5090 dest->uv_height = cm->height >> cm->subsampling_y;
5095 #endif // !CONFIG_VP9_POSTPROC
5096 vpx_clear_system_state();
5101 int vp9_set_internal_size(VP9_COMP *cpi,
5102 VPX_SCALING horiz_mode, VPX_SCALING vert_mode) {
5103 VP9_COMMON *cm = &cpi->common;
5104 int hr = 0, hs = 0, vr = 0, vs = 0;
5106 if (horiz_mode > ONETWO || vert_mode > ONETWO)
5109 Scale2Ratio(horiz_mode, &hr, &hs);
5110 Scale2Ratio(vert_mode, &vr, &vs);
5112 // always go to the next whole number
5113 cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
5114 cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
5115 if (cm->current_video_frame) {
5116 assert(cm->width <= cpi->initial_width);
5117 assert(cm->height <= cpi->initial_height);
5120 update_frame_size(cpi);
5125 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
5126 unsigned int height) {
5127 VP9_COMMON *cm = &cpi->common;
5128 #if CONFIG_VP9_HIGHBITDEPTH
5129 check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
5131 check_initial_width(cpi, 1, 1);
5132 #endif // CONFIG_VP9_HIGHBITDEPTH
5134 #if CONFIG_VP9_TEMPORAL_DENOISING
5135 setup_denoiser_buffer(cpi);
5140 if (cm->width > cpi->initial_width) {
5141 cm->width = cpi->initial_width;
5142 printf("Warning: Desired width too large, changed to %d\n", cm->width);
5147 cm->height = height;
5148 if (cm->height > cpi->initial_height) {
5149 cm->height = cpi->initial_height;
5150 printf("Warning: Desired height too large, changed to %d\n", cm->height);
5153 assert(cm->width <= cpi->initial_width);
5154 assert(cm->height <= cpi->initial_height);
5156 update_frame_size(cpi);
5161 void vp9_set_svc(VP9_COMP *cpi, int use_svc) {
5162 cpi->use_svc = use_svc;
5166 int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a,
5167 const YV12_BUFFER_CONFIG *b) {
5168 assert(a->y_crop_width == b->y_crop_width);
5169 assert(a->y_crop_height == b->y_crop_height);
5171 return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
5172 a->y_crop_width, a->y_crop_height);
5175 #if CONFIG_VP9_HIGHBITDEPTH
5176 int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
5177 const YV12_BUFFER_CONFIG *b) {
5178 assert(a->y_crop_width == b->y_crop_width);
5179 assert(a->y_crop_height == b->y_crop_height);
5180 assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
5181 assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
5183 return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
5184 a->y_crop_width, a->y_crop_height);
5186 #endif // CONFIG_VP9_HIGHBITDEPTH
5188 int vp9_get_quantizer(VP9_COMP *cpi) {
5189 return cpi->common.base_qindex;
5192 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
5193 if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF |
5194 VP8_EFLAG_NO_REF_ARF)) {
5197 if (flags & VP8_EFLAG_NO_REF_LAST)
5198 ref ^= VP9_LAST_FLAG;
5200 if (flags & VP8_EFLAG_NO_REF_GF)
5201 ref ^= VP9_GOLD_FLAG;
5203 if (flags & VP8_EFLAG_NO_REF_ARF)
5204 ref ^= VP9_ALT_FLAG;
5206 vp9_use_as_reference(cpi, ref);
5209 if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
5210 VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF |
5211 VP8_EFLAG_FORCE_ARF)) {
5214 if (flags & VP8_EFLAG_NO_UPD_LAST)
5215 upd ^= VP9_LAST_FLAG;
5217 if (flags & VP8_EFLAG_NO_UPD_GF)
5218 upd ^= VP9_GOLD_FLAG;
5220 if (flags & VP8_EFLAG_NO_UPD_ARF)
5221 upd ^= VP9_ALT_FLAG;
5223 vp9_update_reference(cpi, upd);
5226 if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
5227 vp9_update_entropy(cpi, 0);