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 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
141 if (seg_map[i] == AM_SEGMENT_ID_INACTIVE)
142 seg_map[i] = AM_SEGMENT_ID_ACTIVE;
145 static void apply_active_map(VP9_COMP *cpi) {
146 struct segmentation *const seg = &cpi->common.seg;
147 unsigned char *const seg_map = cpi->segmentation_map;
148 const unsigned char *const active_map = cpi->active_map.map;
151 assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE);
153 if (frame_is_intra_only(&cpi->common)) {
154 cpi->active_map.enabled = 0;
155 cpi->active_map.update = 1;
158 if (cpi->active_map.update) {
159 if (cpi->active_map.enabled) {
160 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
161 if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i];
162 vp9_enable_segmentation(seg);
163 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
164 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
165 // Setting the data to -MAX_LOOP_FILTER will result in the computed loop
166 // filter level being zero regardless of the value of seg->abs_delta.
167 vp9_set_segdata(seg, AM_SEGMENT_ID_INACTIVE,
168 SEG_LVL_ALT_LF, -MAX_LOOP_FILTER);
170 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
171 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
173 seg->update_data = 1;
177 cpi->active_map.update = 0;
181 static void init_level_info(Vp9LevelInfo *level_info) {
182 Vp9LevelStats *const level_stats = &level_info->level_stats;
183 Vp9LevelSpec *const level_spec = &level_info->level_spec;
185 memset(level_stats, 0, sizeof(*level_stats));
186 memset(level_spec, 0, sizeof(*level_spec));
187 level_spec->level = LEVEL_UNKNOWN;
188 level_spec->min_altref_distance = INT_MAX;
191 VP9_LEVEL vp9_get_level(const Vp9LevelSpec * const level_spec) {
193 const Vp9LevelSpec *this_level;
195 vpx_clear_system_state();
197 for (i = 0; i < VP9_LEVELS; ++i) {
198 this_level = &vp9_level_defs[i];
199 if ((double)level_spec->max_luma_sample_rate * (1 + SAMPLE_RATE_GRACE_P) >
200 (double)this_level->max_luma_sample_rate ||
201 level_spec->max_luma_picture_size > this_level->max_luma_picture_size ||
202 level_spec->average_bitrate > this_level->average_bitrate ||
203 level_spec->max_cpb_size > this_level->max_cpb_size ||
204 level_spec->compression_ratio < this_level->compression_ratio ||
205 level_spec->max_col_tiles > this_level->max_col_tiles ||
206 level_spec->min_altref_distance < this_level->min_altref_distance ||
207 level_spec->max_ref_frame_buffers > this_level->max_ref_frame_buffers)
211 return (i == VP9_LEVELS) ? LEVEL_UNKNOWN : vp9_level_defs[i].level;
214 int vp9_set_active_map(VP9_COMP* cpi,
215 unsigned char* new_map_16x16,
218 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
219 unsigned char *const active_map_8x8 = cpi->active_map.map;
220 const int mi_rows = cpi->common.mi_rows;
221 const int mi_cols = cpi->common.mi_cols;
222 cpi->active_map.update = 1;
225 for (r = 0; r < mi_rows; ++r) {
226 for (c = 0; c < mi_cols; ++c) {
227 active_map_8x8[r * mi_cols + c] =
228 new_map_16x16[(r >> 1) * cols + (c >> 1)]
229 ? AM_SEGMENT_ID_ACTIVE
230 : AM_SEGMENT_ID_INACTIVE;
233 cpi->active_map.enabled = 1;
235 cpi->active_map.enabled = 0;
243 int vp9_get_active_map(VP9_COMP* cpi,
244 unsigned char* new_map_16x16,
247 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols &&
249 unsigned char* const seg_map_8x8 = cpi->segmentation_map;
250 const int mi_rows = cpi->common.mi_rows;
251 const int mi_cols = cpi->common.mi_cols;
252 memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
253 if (cpi->active_map.enabled) {
255 for (r = 0; r < mi_rows; ++r) {
256 for (c = 0; c < mi_cols; ++c) {
257 // Cyclic refresh segments are considered active despite not having
258 // AM_SEGMENT_ID_ACTIVE
259 new_map_16x16[(r >> 1) * cols + (c >> 1)] |=
260 seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE;
270 void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
271 MACROBLOCK *const mb = &cpi->td.mb;
272 cpi->common.allow_high_precision_mv = allow_high_precision_mv;
273 if (cpi->common.allow_high_precision_mv) {
274 mb->mvcost = mb->nmvcost_hp;
275 mb->mvsadcost = mb->nmvsadcost_hp;
277 mb->mvcost = mb->nmvcost;
278 mb->mvsadcost = mb->nmvsadcost;
282 static void setup_frame(VP9_COMP *cpi) {
283 VP9_COMMON *const cm = &cpi->common;
284 // Set up entropy context depending on frame type. The decoder mandates
285 // the use of the default context, index 0, for keyframes and inter
286 // frames where the error_resilient_mode or intra_only flag is set. For
287 // other inter-frames the encoder currently uses only two contexts;
288 // context 1 for ALTREF frames and context 0 for the others.
289 if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
290 vp9_setup_past_independence(cm);
293 cm->frame_context_idx = cpi->refresh_alt_ref_frame;
296 if (cm->frame_type == KEY_FRAME) {
297 if (!is_two_pass_svc(cpi))
298 cpi->refresh_golden_frame = 1;
299 cpi->refresh_alt_ref_frame = 1;
300 vp9_zero(cpi->interp_filter_selected);
302 *cm->fc = cm->frame_contexts[cm->frame_context_idx];
303 vp9_zero(cpi->interp_filter_selected[0]);
307 static void vp9_enc_setup_mi(VP9_COMMON *cm) {
309 cm->mi = cm->mip + cm->mi_stride + 1;
310 memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
311 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
312 // Clear top border row
313 memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride);
314 // Clear left border column
315 for (i = 1; i < cm->mi_rows + 1; ++i)
316 memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip));
318 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
319 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
321 memset(cm->mi_grid_base, 0,
322 cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
325 static int vp9_enc_alloc_mi(VP9_COMMON *cm, int mi_size) {
326 cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
329 cm->prev_mip = vpx_calloc(mi_size, sizeof(*cm->prev_mip));
332 cm->mi_alloc_size = mi_size;
334 cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
335 if (!cm->mi_grid_base)
337 cm->prev_mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
338 if (!cm->prev_mi_grid_base)
344 static void vp9_enc_free_mi(VP9_COMMON *cm) {
347 vpx_free(cm->prev_mip);
349 vpx_free(cm->mi_grid_base);
350 cm->mi_grid_base = NULL;
351 vpx_free(cm->prev_mi_grid_base);
352 cm->prev_mi_grid_base = NULL;
355 static void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
356 // Current mip will be the prev_mip for the next frame.
357 MODE_INFO **temp_base = cm->prev_mi_grid_base;
358 MODE_INFO *temp = cm->prev_mip;
359 cm->prev_mip = cm->mip;
362 // Update the upper left visible macroblock ptrs.
363 cm->mi = cm->mip + cm->mi_stride + 1;
364 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
366 cm->prev_mi_grid_base = cm->mi_grid_base;
367 cm->mi_grid_base = temp_base;
368 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
369 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
372 void vp9_initialize_enc(void) {
373 static volatile int init_done = 0;
379 vp9_init_intra_predictors();
381 vp9_rc_init_minq_luts();
382 vp9_entropy_mv_init();
383 vp9_temporal_filter_init();
388 static void dealloc_compressor_data(VP9_COMP *cpi) {
389 VP9_COMMON *const cm = &cpi->common;
392 vpx_free(cpi->mbmi_ext_base);
393 cpi->mbmi_ext_base = NULL;
395 vpx_free(cpi->tile_data);
396 cpi->tile_data = NULL;
398 // Delete sementation map
399 vpx_free(cpi->segmentation_map);
400 cpi->segmentation_map = NULL;
401 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
402 cpi->coding_context.last_frame_seg_map_copy = NULL;
404 vpx_free(cpi->nmvcosts[0]);
405 vpx_free(cpi->nmvcosts[1]);
406 cpi->nmvcosts[0] = NULL;
407 cpi->nmvcosts[1] = NULL;
409 vpx_free(cpi->nmvcosts_hp[0]);
410 vpx_free(cpi->nmvcosts_hp[1]);
411 cpi->nmvcosts_hp[0] = NULL;
412 cpi->nmvcosts_hp[1] = NULL;
414 vpx_free(cpi->nmvsadcosts[0]);
415 vpx_free(cpi->nmvsadcosts[1]);
416 cpi->nmvsadcosts[0] = NULL;
417 cpi->nmvsadcosts[1] = NULL;
419 vpx_free(cpi->nmvsadcosts_hp[0]);
420 vpx_free(cpi->nmvsadcosts_hp[1]);
421 cpi->nmvsadcosts_hp[0] = NULL;
422 cpi->nmvsadcosts_hp[1] = NULL;
424 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
425 cpi->cyclic_refresh = NULL;
427 vpx_free(cpi->active_map.map);
428 cpi->active_map.map = NULL;
430 vpx_free(cpi->consec_zero_mv);
431 cpi->consec_zero_mv = NULL;
433 vp9_free_ref_frame_buffers(cm->buffer_pool);
434 #if CONFIG_VP9_POSTPROC
435 vp9_free_postproc_buffers(cm);
437 vp9_free_context_buffers(cm);
439 vpx_free_frame_buffer(&cpi->last_frame_uf);
440 vpx_free_frame_buffer(&cpi->scaled_source);
441 vpx_free_frame_buffer(&cpi->scaled_last_source);
442 vpx_free_frame_buffer(&cpi->alt_ref_buffer);
443 vp9_lookahead_destroy(cpi->lookahead);
445 vpx_free(cpi->tile_tok[0][0]);
446 cpi->tile_tok[0][0] = 0;
448 vp9_free_pc_tree(&cpi->td);
450 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
451 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
452 vpx_free(lc->rc_twopass_stats_in.buf);
453 lc->rc_twopass_stats_in.buf = NULL;
454 lc->rc_twopass_stats_in.sz = 0;
457 if (cpi->source_diff_var != NULL) {
458 vpx_free(cpi->source_diff_var);
459 cpi->source_diff_var = NULL;
462 for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
463 vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]);
465 memset(&cpi->svc.scaled_frames[0], 0,
466 MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
468 vpx_free_frame_buffer(&cpi->svc.scaled_temp);
469 memset(&cpi->svc.scaled_temp, 0, sizeof(cpi->svc.scaled_temp));
471 vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
472 memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
474 vp9_free_svc_cyclic_refresh(cpi);
477 static void save_coding_context(VP9_COMP *cpi) {
478 CODING_CONTEXT *const cc = &cpi->coding_context;
479 VP9_COMMON *cm = &cpi->common;
481 // Stores a snapshot of key state variables which can subsequently be
482 // restored with a call to vp9_restore_coding_context. These functions are
483 // intended for use in a re-code loop in vp9_compress_frame where the
484 // quantizer value is adjusted between loop iterations.
485 vp9_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost);
487 memcpy(cc->nmvcosts[0], cpi->nmvcosts[0],
488 MV_VALS * sizeof(*cpi->nmvcosts[0]));
489 memcpy(cc->nmvcosts[1], cpi->nmvcosts[1],
490 MV_VALS * sizeof(*cpi->nmvcosts[1]));
491 memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0],
492 MV_VALS * sizeof(*cpi->nmvcosts_hp[0]));
493 memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1],
494 MV_VALS * sizeof(*cpi->nmvcosts_hp[1]));
496 vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs);
498 memcpy(cpi->coding_context.last_frame_seg_map_copy,
499 cm->last_frame_seg_map, (cm->mi_rows * cm->mi_cols));
501 vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
502 vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
507 static void restore_coding_context(VP9_COMP *cpi) {
508 CODING_CONTEXT *const cc = &cpi->coding_context;
509 VP9_COMMON *cm = &cpi->common;
511 // Restore key state variables to the snapshot state stored in the
512 // previous call to vp9_save_coding_context.
513 vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
515 memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0]));
516 memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1]));
517 memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0],
518 MV_VALS * sizeof(*cc->nmvcosts_hp[0]));
519 memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1],
520 MV_VALS * sizeof(*cc->nmvcosts_hp[1]));
522 vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs);
524 memcpy(cm->last_frame_seg_map,
525 cpi->coding_context.last_frame_seg_map_copy,
526 (cm->mi_rows * cm->mi_cols));
528 vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
529 vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
534 static void configure_static_seg_features(VP9_COMP *cpi) {
535 VP9_COMMON *const cm = &cpi->common;
536 const RATE_CONTROL *const rc = &cpi->rc;
537 struct segmentation *const seg = &cm->seg;
539 int high_q = (int)(rc->avg_q > 48.0);
542 // Disable and clear down for KF
543 if (cm->frame_type == KEY_FRAME) {
544 // Clear down the global segmentation map
545 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
547 seg->update_data = 0;
548 cpi->static_mb_pct = 0;
550 // Disable segmentation
551 vp9_disable_segmentation(seg);
553 // Clear down the segment features.
554 vp9_clearall_segfeatures(seg);
555 } else if (cpi->refresh_alt_ref_frame) {
556 // If this is an alt ref frame
557 // Clear down the global segmentation map
558 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
560 seg->update_data = 0;
561 cpi->static_mb_pct = 0;
563 // Disable segmentation and individual segment features by default
564 vp9_disable_segmentation(seg);
565 vp9_clearall_segfeatures(seg);
567 // Scan frames from current to arf frame.
568 // This function re-enables segmentation if appropriate.
569 vp9_update_mbgraph_stats(cpi);
571 // If segmentation was enabled set those features needed for the
575 seg->update_data = 1;
577 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875,
579 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
580 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
582 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
583 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
585 // Where relevant assume segment data is delta data
586 seg->abs_delta = SEGMENT_DELTADATA;
588 } else if (seg->enabled) {
589 // All other frames if segmentation has been enabled
591 // First normal frame in a valid gf or alt ref group
592 if (rc->frames_since_golden == 0) {
593 // Set up segment features for normal frames in an arf group
594 if (rc->source_alt_ref_active) {
596 seg->update_data = 1;
597 seg->abs_delta = SEGMENT_DELTADATA;
599 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125,
601 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
602 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
604 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
605 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
607 // Segment coding disabled for compred testing
608 if (high_q || (cpi->static_mb_pct == 100)) {
609 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
610 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
611 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
614 // Disable segmentation and clear down features if alt ref
615 // is not active for this group
617 vp9_disable_segmentation(seg);
619 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
622 seg->update_data = 0;
624 vp9_clearall_segfeatures(seg);
626 } else if (rc->is_src_frame_alt_ref) {
627 // Special case where we are coding over the top of a previous
629 // Segment coding disabled for compred testing
631 // Enable ref frame features for segment 0 as well
632 vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
633 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
635 // All mbs should use ALTREF_FRAME
636 vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
637 vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
638 vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
639 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
641 // Skip all MBs if high Q (0,0 mv and skip coeffs)
643 vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP);
644 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
646 // Enable data update
647 seg->update_data = 1;
651 // No updates.. leave things as they are.
653 seg->update_data = 0;
658 static void update_reference_segmentation_map(VP9_COMP *cpi) {
659 VP9_COMMON *const cm = &cpi->common;
660 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
661 uint8_t *cache_ptr = cm->last_frame_seg_map;
664 for (row = 0; row < cm->mi_rows; row++) {
665 MODE_INFO **mi_8x8 = mi_8x8_ptr;
666 uint8_t *cache = cache_ptr;
667 for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
668 cache[0] = mi_8x8[0]->segment_id;
669 mi_8x8_ptr += cm->mi_stride;
670 cache_ptr += cm->mi_cols;
674 static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
675 VP9_COMMON *cm = &cpi->common;
676 const VP9EncoderConfig *oxcf = &cpi->oxcf;
679 cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
680 cm->subsampling_x, cm->subsampling_y,
681 #if CONFIG_VP9_HIGHBITDEPTH
682 cm->use_highbitdepth,
684 oxcf->lag_in_frames);
686 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
687 "Failed to allocate lag buffers");
689 // TODO(agrange) Check if ARF is enabled and skip allocation if not.
690 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
691 oxcf->width, oxcf->height,
692 cm->subsampling_x, cm->subsampling_y,
693 #if CONFIG_VP9_HIGHBITDEPTH
694 cm->use_highbitdepth,
696 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
698 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
699 "Failed to allocate altref buffer");
702 static void alloc_util_frame_buffers(VP9_COMP *cpi) {
703 VP9_COMMON *const cm = &cpi->common;
704 if (vpx_realloc_frame_buffer(&cpi->last_frame_uf,
705 cm->width, cm->height,
706 cm->subsampling_x, cm->subsampling_y,
707 #if CONFIG_VP9_HIGHBITDEPTH
708 cm->use_highbitdepth,
710 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
712 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
713 "Failed to allocate last frame buffer");
715 if (vpx_realloc_frame_buffer(&cpi->scaled_source,
716 cm->width, cm->height,
717 cm->subsampling_x, cm->subsampling_y,
718 #if CONFIG_VP9_HIGHBITDEPTH
719 cm->use_highbitdepth,
721 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
723 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
724 "Failed to allocate scaled source buffer");
726 if (vpx_realloc_frame_buffer(&cpi->scaled_last_source,
727 cm->width, cm->height,
728 cm->subsampling_x, cm->subsampling_y,
729 #if CONFIG_VP9_HIGHBITDEPTH
730 cm->use_highbitdepth,
732 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
734 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
735 "Failed to allocate scaled last source buffer");
739 static int alloc_context_buffers_ext(VP9_COMP *cpi) {
740 VP9_COMMON *cm = &cpi->common;
741 int mi_size = cm->mi_cols * cm->mi_rows;
743 cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
744 if (!cpi->mbmi_ext_base)
750 static void alloc_compressor_data(VP9_COMP *cpi) {
751 VP9_COMMON *cm = &cpi->common;
753 vp9_alloc_context_buffers(cm, cm->width, cm->height);
755 alloc_context_buffers_ext(cpi);
757 vpx_free(cpi->tile_tok[0][0]);
760 unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
761 CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
762 vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
765 vp9_setup_pc_tree(&cpi->common, &cpi->td);
768 void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
769 cpi->framerate = framerate < 0.1 ? 30 : framerate;
770 vp9_rc_update_framerate(cpi);
773 static void set_tile_limits(VP9_COMP *cpi) {
774 VP9_COMMON *const cm = &cpi->common;
776 int min_log2_tile_cols, max_log2_tile_cols;
777 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
779 if (is_two_pass_svc(cpi) &&
780 (cpi->svc.encode_empty_frame_state == ENCODING ||
781 cpi->svc.number_spatial_layers > 1)) {
782 cm->log2_tile_cols = 0;
783 cm->log2_tile_rows = 0;
785 cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns,
786 min_log2_tile_cols, max_log2_tile_cols);
787 cm->log2_tile_rows = cpi->oxcf.tile_rows;
791 static void update_frame_size(VP9_COMP *cpi) {
792 VP9_COMMON *const cm = &cpi->common;
793 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
795 vp9_set_mb_mi(cm, cm->width, cm->height);
796 vp9_init_context_buffers(cm);
797 vp9_init_macroblockd(cm, xd, NULL);
798 cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base;
799 memset(cpi->mbmi_ext_base, 0,
800 cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
802 set_tile_limits(cpi);
804 if (is_two_pass_svc(cpi)) {
805 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
806 cm->width, cm->height,
807 cm->subsampling_x, cm->subsampling_y,
808 #if CONFIG_VP9_HIGHBITDEPTH
809 cm->use_highbitdepth,
811 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
813 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
814 "Failed to reallocate alt_ref_buffer");
818 static void init_buffer_indices(VP9_COMP *cpi) {
824 static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
825 VP9_COMMON *const cm = &cpi->common;
828 cpi->framerate = oxcf->init_framerate;
829 cm->profile = oxcf->profile;
830 cm->bit_depth = oxcf->bit_depth;
831 #if CONFIG_VP9_HIGHBITDEPTH
832 cm->use_highbitdepth = oxcf->use_highbitdepth;
834 cm->color_space = oxcf->color_space;
835 cm->color_range = oxcf->color_range;
837 cpi->target_level = oxcf->target_level;
838 cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
840 cm->width = oxcf->width;
841 cm->height = oxcf->height;
842 alloc_compressor_data(cpi);
844 cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode;
846 // Single thread case: use counts in common.
847 cpi->td.counts = &cm->counts;
849 // Spatial scalability.
850 cpi->svc.number_spatial_layers = oxcf->ss_number_layers;
851 // Temporal scalability.
852 cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
854 if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
855 ((cpi->svc.number_temporal_layers > 1 ||
856 cpi->svc.number_spatial_layers > 1) &&
857 cpi->oxcf.pass != 1)) {
858 vp9_init_layer_context(cpi);
861 // change includes all joint functionality
862 vp9_change_config(cpi, oxcf);
864 cpi->static_mb_pct = 0;
865 cpi->ref_frame_flags = 0;
867 init_buffer_indices(cpi);
869 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
872 static void set_rc_buffer_sizes(RATE_CONTROL *rc,
873 const VP9EncoderConfig *oxcf) {
874 const int64_t bandwidth = oxcf->target_bandwidth;
875 const int64_t starting = oxcf->starting_buffer_level_ms;
876 const int64_t optimal = oxcf->optimal_buffer_level_ms;
877 const int64_t maximum = oxcf->maximum_buffer_size_ms;
879 rc->starting_buffer_level = starting * bandwidth / 1000;
880 rc->optimal_buffer_level = (optimal == 0) ? bandwidth / 8
881 : optimal * bandwidth / 1000;
882 rc->maximum_buffer_size = (maximum == 0) ? bandwidth / 8
883 : maximum * bandwidth / 1000;
886 #if CONFIG_VP9_HIGHBITDEPTH
887 #define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
888 cpi->fn_ptr[BT].sdf = SDF; \
889 cpi->fn_ptr[BT].sdaf = SDAF; \
890 cpi->fn_ptr[BT].vf = VF; \
891 cpi->fn_ptr[BT].svf = SVF; \
892 cpi->fn_ptr[BT].svaf = SVAF; \
893 cpi->fn_ptr[BT].sdx3f = SDX3F; \
894 cpi->fn_ptr[BT].sdx8f = SDX8F; \
895 cpi->fn_ptr[BT].sdx4df = SDX4DF;
897 #define MAKE_BFP_SAD_WRAPPER(fnname) \
898 static unsigned int fnname##_bits8(const uint8_t *src_ptr, \
900 const uint8_t *ref_ptr, \
902 return fnname(src_ptr, source_stride, ref_ptr, ref_stride); \
904 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
906 const uint8_t *ref_ptr, \
908 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2; \
910 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
912 const uint8_t *ref_ptr, \
914 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4; \
917 #define MAKE_BFP_SADAVG_WRAPPER(fnname) static unsigned int \
918 fnname##_bits8(const uint8_t *src_ptr, \
920 const uint8_t *ref_ptr, \
922 const uint8_t *second_pred) { \
923 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred); \
925 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
927 const uint8_t *ref_ptr, \
929 const uint8_t *second_pred) { \
930 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
933 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
935 const uint8_t *ref_ptr, \
937 const uint8_t *second_pred) { \
938 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
942 #define MAKE_BFP_SAD3_WRAPPER(fnname) \
943 static void fnname##_bits8(const uint8_t *src_ptr, \
945 const uint8_t *ref_ptr, \
947 unsigned int *sad_array) { \
948 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
950 static void fnname##_bits10(const uint8_t *src_ptr, \
952 const uint8_t *ref_ptr, \
954 unsigned int *sad_array) { \
956 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
957 for (i = 0; i < 3; i++) \
958 sad_array[i] >>= 2; \
960 static void fnname##_bits12(const uint8_t *src_ptr, \
962 const uint8_t *ref_ptr, \
964 unsigned int *sad_array) { \
966 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
967 for (i = 0; i < 3; i++) \
968 sad_array[i] >>= 4; \
971 #define MAKE_BFP_SAD8_WRAPPER(fnname) \
972 static void fnname##_bits8(const uint8_t *src_ptr, \
974 const uint8_t *ref_ptr, \
976 unsigned int *sad_array) { \
977 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
979 static void fnname##_bits10(const uint8_t *src_ptr, \
981 const uint8_t *ref_ptr, \
983 unsigned int *sad_array) { \
985 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
986 for (i = 0; i < 8; i++) \
987 sad_array[i] >>= 2; \
989 static void fnname##_bits12(const uint8_t *src_ptr, \
991 const uint8_t *ref_ptr, \
993 unsigned int *sad_array) { \
995 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
996 for (i = 0; i < 8; i++) \
997 sad_array[i] >>= 4; \
999 #define MAKE_BFP_SAD4D_WRAPPER(fnname) \
1000 static void fnname##_bits8(const uint8_t *src_ptr, \
1001 int source_stride, \
1002 const uint8_t* const ref_ptr[], \
1004 unsigned int *sad_array) { \
1005 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1007 static void fnname##_bits10(const uint8_t *src_ptr, \
1008 int source_stride, \
1009 const uint8_t* const ref_ptr[], \
1011 unsigned int *sad_array) { \
1013 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1014 for (i = 0; i < 4; i++) \
1015 sad_array[i] >>= 2; \
1017 static void fnname##_bits12(const uint8_t *src_ptr, \
1018 int source_stride, \
1019 const uint8_t* const ref_ptr[], \
1021 unsigned int *sad_array) { \
1023 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1024 for (i = 0; i < 4; i++) \
1025 sad_array[i] >>= 4; \
1028 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16)
1029 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg)
1030 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d)
1031 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32)
1032 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg)
1033 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d)
1034 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32)
1035 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg)
1036 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d)
1037 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64)
1038 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg)
1039 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d)
1040 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32)
1041 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg)
1042 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad32x32x3)
1043 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad32x32x8)
1044 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d)
1045 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64)
1046 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg)
1047 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad64x64x3)
1048 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad64x64x8)
1049 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d)
1050 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16)
1051 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg)
1052 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x16x3)
1053 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x16x8)
1054 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d)
1055 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8)
1056 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg)
1057 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x8x3)
1058 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x8x8)
1059 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d)
1060 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16)
1061 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg)
1062 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x16x3)
1063 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x16x8)
1064 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d)
1065 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8)
1066 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg)
1067 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x8x3)
1068 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x8x8)
1069 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d)
1070 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4)
1071 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg)
1072 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x4x8)
1073 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d)
1074 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8)
1075 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg)
1076 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x8x8)
1077 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d)
1078 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4)
1079 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg)
1080 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad4x4x3)
1081 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x4x8)
1082 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d)
1084 static void highbd_set_var_fns(VP9_COMP *const cpi) {
1085 VP9_COMMON *const cm = &cpi->common;
1086 if (cm->use_highbitdepth) {
1087 switch (cm->bit_depth) {
1089 HIGHBD_BFP(BLOCK_32X16,
1090 vpx_highbd_sad32x16_bits8,
1091 vpx_highbd_sad32x16_avg_bits8,
1092 vpx_highbd_8_variance32x16,
1093 vpx_highbd_8_sub_pixel_variance32x16,
1094 vpx_highbd_8_sub_pixel_avg_variance32x16,
1097 vpx_highbd_sad32x16x4d_bits8)
1099 HIGHBD_BFP(BLOCK_16X32,
1100 vpx_highbd_sad16x32_bits8,
1101 vpx_highbd_sad16x32_avg_bits8,
1102 vpx_highbd_8_variance16x32,
1103 vpx_highbd_8_sub_pixel_variance16x32,
1104 vpx_highbd_8_sub_pixel_avg_variance16x32,
1107 vpx_highbd_sad16x32x4d_bits8)
1109 HIGHBD_BFP(BLOCK_64X32,
1110 vpx_highbd_sad64x32_bits8,
1111 vpx_highbd_sad64x32_avg_bits8,
1112 vpx_highbd_8_variance64x32,
1113 vpx_highbd_8_sub_pixel_variance64x32,
1114 vpx_highbd_8_sub_pixel_avg_variance64x32,
1117 vpx_highbd_sad64x32x4d_bits8)
1119 HIGHBD_BFP(BLOCK_32X64,
1120 vpx_highbd_sad32x64_bits8,
1121 vpx_highbd_sad32x64_avg_bits8,
1122 vpx_highbd_8_variance32x64,
1123 vpx_highbd_8_sub_pixel_variance32x64,
1124 vpx_highbd_8_sub_pixel_avg_variance32x64,
1127 vpx_highbd_sad32x64x4d_bits8)
1129 HIGHBD_BFP(BLOCK_32X32,
1130 vpx_highbd_sad32x32_bits8,
1131 vpx_highbd_sad32x32_avg_bits8,
1132 vpx_highbd_8_variance32x32,
1133 vpx_highbd_8_sub_pixel_variance32x32,
1134 vpx_highbd_8_sub_pixel_avg_variance32x32,
1135 vpx_highbd_sad32x32x3_bits8,
1136 vpx_highbd_sad32x32x8_bits8,
1137 vpx_highbd_sad32x32x4d_bits8)
1139 HIGHBD_BFP(BLOCK_64X64,
1140 vpx_highbd_sad64x64_bits8,
1141 vpx_highbd_sad64x64_avg_bits8,
1142 vpx_highbd_8_variance64x64,
1143 vpx_highbd_8_sub_pixel_variance64x64,
1144 vpx_highbd_8_sub_pixel_avg_variance64x64,
1145 vpx_highbd_sad64x64x3_bits8,
1146 vpx_highbd_sad64x64x8_bits8,
1147 vpx_highbd_sad64x64x4d_bits8)
1149 HIGHBD_BFP(BLOCK_16X16,
1150 vpx_highbd_sad16x16_bits8,
1151 vpx_highbd_sad16x16_avg_bits8,
1152 vpx_highbd_8_variance16x16,
1153 vpx_highbd_8_sub_pixel_variance16x16,
1154 vpx_highbd_8_sub_pixel_avg_variance16x16,
1155 vpx_highbd_sad16x16x3_bits8,
1156 vpx_highbd_sad16x16x8_bits8,
1157 vpx_highbd_sad16x16x4d_bits8)
1159 HIGHBD_BFP(BLOCK_16X8,
1160 vpx_highbd_sad16x8_bits8,
1161 vpx_highbd_sad16x8_avg_bits8,
1162 vpx_highbd_8_variance16x8,
1163 vpx_highbd_8_sub_pixel_variance16x8,
1164 vpx_highbd_8_sub_pixel_avg_variance16x8,
1165 vpx_highbd_sad16x8x3_bits8,
1166 vpx_highbd_sad16x8x8_bits8,
1167 vpx_highbd_sad16x8x4d_bits8)
1169 HIGHBD_BFP(BLOCK_8X16,
1170 vpx_highbd_sad8x16_bits8,
1171 vpx_highbd_sad8x16_avg_bits8,
1172 vpx_highbd_8_variance8x16,
1173 vpx_highbd_8_sub_pixel_variance8x16,
1174 vpx_highbd_8_sub_pixel_avg_variance8x16,
1175 vpx_highbd_sad8x16x3_bits8,
1176 vpx_highbd_sad8x16x8_bits8,
1177 vpx_highbd_sad8x16x4d_bits8)
1179 HIGHBD_BFP(BLOCK_8X8,
1180 vpx_highbd_sad8x8_bits8,
1181 vpx_highbd_sad8x8_avg_bits8,
1182 vpx_highbd_8_variance8x8,
1183 vpx_highbd_8_sub_pixel_variance8x8,
1184 vpx_highbd_8_sub_pixel_avg_variance8x8,
1185 vpx_highbd_sad8x8x3_bits8,
1186 vpx_highbd_sad8x8x8_bits8,
1187 vpx_highbd_sad8x8x4d_bits8)
1189 HIGHBD_BFP(BLOCK_8X4,
1190 vpx_highbd_sad8x4_bits8,
1191 vpx_highbd_sad8x4_avg_bits8,
1192 vpx_highbd_8_variance8x4,
1193 vpx_highbd_8_sub_pixel_variance8x4,
1194 vpx_highbd_8_sub_pixel_avg_variance8x4,
1196 vpx_highbd_sad8x4x8_bits8,
1197 vpx_highbd_sad8x4x4d_bits8)
1199 HIGHBD_BFP(BLOCK_4X8,
1200 vpx_highbd_sad4x8_bits8,
1201 vpx_highbd_sad4x8_avg_bits8,
1202 vpx_highbd_8_variance4x8,
1203 vpx_highbd_8_sub_pixel_variance4x8,
1204 vpx_highbd_8_sub_pixel_avg_variance4x8,
1206 vpx_highbd_sad4x8x8_bits8,
1207 vpx_highbd_sad4x8x4d_bits8)
1209 HIGHBD_BFP(BLOCK_4X4,
1210 vpx_highbd_sad4x4_bits8,
1211 vpx_highbd_sad4x4_avg_bits8,
1212 vpx_highbd_8_variance4x4,
1213 vpx_highbd_8_sub_pixel_variance4x4,
1214 vpx_highbd_8_sub_pixel_avg_variance4x4,
1215 vpx_highbd_sad4x4x3_bits8,
1216 vpx_highbd_sad4x4x8_bits8,
1217 vpx_highbd_sad4x4x4d_bits8)
1221 HIGHBD_BFP(BLOCK_32X16,
1222 vpx_highbd_sad32x16_bits10,
1223 vpx_highbd_sad32x16_avg_bits10,
1224 vpx_highbd_10_variance32x16,
1225 vpx_highbd_10_sub_pixel_variance32x16,
1226 vpx_highbd_10_sub_pixel_avg_variance32x16,
1229 vpx_highbd_sad32x16x4d_bits10)
1231 HIGHBD_BFP(BLOCK_16X32,
1232 vpx_highbd_sad16x32_bits10,
1233 vpx_highbd_sad16x32_avg_bits10,
1234 vpx_highbd_10_variance16x32,
1235 vpx_highbd_10_sub_pixel_variance16x32,
1236 vpx_highbd_10_sub_pixel_avg_variance16x32,
1239 vpx_highbd_sad16x32x4d_bits10)
1241 HIGHBD_BFP(BLOCK_64X32,
1242 vpx_highbd_sad64x32_bits10,
1243 vpx_highbd_sad64x32_avg_bits10,
1244 vpx_highbd_10_variance64x32,
1245 vpx_highbd_10_sub_pixel_variance64x32,
1246 vpx_highbd_10_sub_pixel_avg_variance64x32,
1249 vpx_highbd_sad64x32x4d_bits10)
1251 HIGHBD_BFP(BLOCK_32X64,
1252 vpx_highbd_sad32x64_bits10,
1253 vpx_highbd_sad32x64_avg_bits10,
1254 vpx_highbd_10_variance32x64,
1255 vpx_highbd_10_sub_pixel_variance32x64,
1256 vpx_highbd_10_sub_pixel_avg_variance32x64,
1259 vpx_highbd_sad32x64x4d_bits10)
1261 HIGHBD_BFP(BLOCK_32X32,
1262 vpx_highbd_sad32x32_bits10,
1263 vpx_highbd_sad32x32_avg_bits10,
1264 vpx_highbd_10_variance32x32,
1265 vpx_highbd_10_sub_pixel_variance32x32,
1266 vpx_highbd_10_sub_pixel_avg_variance32x32,
1267 vpx_highbd_sad32x32x3_bits10,
1268 vpx_highbd_sad32x32x8_bits10,
1269 vpx_highbd_sad32x32x4d_bits10)
1271 HIGHBD_BFP(BLOCK_64X64,
1272 vpx_highbd_sad64x64_bits10,
1273 vpx_highbd_sad64x64_avg_bits10,
1274 vpx_highbd_10_variance64x64,
1275 vpx_highbd_10_sub_pixel_variance64x64,
1276 vpx_highbd_10_sub_pixel_avg_variance64x64,
1277 vpx_highbd_sad64x64x3_bits10,
1278 vpx_highbd_sad64x64x8_bits10,
1279 vpx_highbd_sad64x64x4d_bits10)
1281 HIGHBD_BFP(BLOCK_16X16,
1282 vpx_highbd_sad16x16_bits10,
1283 vpx_highbd_sad16x16_avg_bits10,
1284 vpx_highbd_10_variance16x16,
1285 vpx_highbd_10_sub_pixel_variance16x16,
1286 vpx_highbd_10_sub_pixel_avg_variance16x16,
1287 vpx_highbd_sad16x16x3_bits10,
1288 vpx_highbd_sad16x16x8_bits10,
1289 vpx_highbd_sad16x16x4d_bits10)
1291 HIGHBD_BFP(BLOCK_16X8,
1292 vpx_highbd_sad16x8_bits10,
1293 vpx_highbd_sad16x8_avg_bits10,
1294 vpx_highbd_10_variance16x8,
1295 vpx_highbd_10_sub_pixel_variance16x8,
1296 vpx_highbd_10_sub_pixel_avg_variance16x8,
1297 vpx_highbd_sad16x8x3_bits10,
1298 vpx_highbd_sad16x8x8_bits10,
1299 vpx_highbd_sad16x8x4d_bits10)
1301 HIGHBD_BFP(BLOCK_8X16,
1302 vpx_highbd_sad8x16_bits10,
1303 vpx_highbd_sad8x16_avg_bits10,
1304 vpx_highbd_10_variance8x16,
1305 vpx_highbd_10_sub_pixel_variance8x16,
1306 vpx_highbd_10_sub_pixel_avg_variance8x16,
1307 vpx_highbd_sad8x16x3_bits10,
1308 vpx_highbd_sad8x16x8_bits10,
1309 vpx_highbd_sad8x16x4d_bits10)
1311 HIGHBD_BFP(BLOCK_8X8,
1312 vpx_highbd_sad8x8_bits10,
1313 vpx_highbd_sad8x8_avg_bits10,
1314 vpx_highbd_10_variance8x8,
1315 vpx_highbd_10_sub_pixel_variance8x8,
1316 vpx_highbd_10_sub_pixel_avg_variance8x8,
1317 vpx_highbd_sad8x8x3_bits10,
1318 vpx_highbd_sad8x8x8_bits10,
1319 vpx_highbd_sad8x8x4d_bits10)
1321 HIGHBD_BFP(BLOCK_8X4,
1322 vpx_highbd_sad8x4_bits10,
1323 vpx_highbd_sad8x4_avg_bits10,
1324 vpx_highbd_10_variance8x4,
1325 vpx_highbd_10_sub_pixel_variance8x4,
1326 vpx_highbd_10_sub_pixel_avg_variance8x4,
1328 vpx_highbd_sad8x4x8_bits10,
1329 vpx_highbd_sad8x4x4d_bits10)
1331 HIGHBD_BFP(BLOCK_4X8,
1332 vpx_highbd_sad4x8_bits10,
1333 vpx_highbd_sad4x8_avg_bits10,
1334 vpx_highbd_10_variance4x8,
1335 vpx_highbd_10_sub_pixel_variance4x8,
1336 vpx_highbd_10_sub_pixel_avg_variance4x8,
1338 vpx_highbd_sad4x8x8_bits10,
1339 vpx_highbd_sad4x8x4d_bits10)
1341 HIGHBD_BFP(BLOCK_4X4,
1342 vpx_highbd_sad4x4_bits10,
1343 vpx_highbd_sad4x4_avg_bits10,
1344 vpx_highbd_10_variance4x4,
1345 vpx_highbd_10_sub_pixel_variance4x4,
1346 vpx_highbd_10_sub_pixel_avg_variance4x4,
1347 vpx_highbd_sad4x4x3_bits10,
1348 vpx_highbd_sad4x4x8_bits10,
1349 vpx_highbd_sad4x4x4d_bits10)
1353 HIGHBD_BFP(BLOCK_32X16,
1354 vpx_highbd_sad32x16_bits12,
1355 vpx_highbd_sad32x16_avg_bits12,
1356 vpx_highbd_12_variance32x16,
1357 vpx_highbd_12_sub_pixel_variance32x16,
1358 vpx_highbd_12_sub_pixel_avg_variance32x16,
1361 vpx_highbd_sad32x16x4d_bits12)
1363 HIGHBD_BFP(BLOCK_16X32,
1364 vpx_highbd_sad16x32_bits12,
1365 vpx_highbd_sad16x32_avg_bits12,
1366 vpx_highbd_12_variance16x32,
1367 vpx_highbd_12_sub_pixel_variance16x32,
1368 vpx_highbd_12_sub_pixel_avg_variance16x32,
1371 vpx_highbd_sad16x32x4d_bits12)
1373 HIGHBD_BFP(BLOCK_64X32,
1374 vpx_highbd_sad64x32_bits12,
1375 vpx_highbd_sad64x32_avg_bits12,
1376 vpx_highbd_12_variance64x32,
1377 vpx_highbd_12_sub_pixel_variance64x32,
1378 vpx_highbd_12_sub_pixel_avg_variance64x32,
1381 vpx_highbd_sad64x32x4d_bits12)
1383 HIGHBD_BFP(BLOCK_32X64,
1384 vpx_highbd_sad32x64_bits12,
1385 vpx_highbd_sad32x64_avg_bits12,
1386 vpx_highbd_12_variance32x64,
1387 vpx_highbd_12_sub_pixel_variance32x64,
1388 vpx_highbd_12_sub_pixel_avg_variance32x64,
1391 vpx_highbd_sad32x64x4d_bits12)
1393 HIGHBD_BFP(BLOCK_32X32,
1394 vpx_highbd_sad32x32_bits12,
1395 vpx_highbd_sad32x32_avg_bits12,
1396 vpx_highbd_12_variance32x32,
1397 vpx_highbd_12_sub_pixel_variance32x32,
1398 vpx_highbd_12_sub_pixel_avg_variance32x32,
1399 vpx_highbd_sad32x32x3_bits12,
1400 vpx_highbd_sad32x32x8_bits12,
1401 vpx_highbd_sad32x32x4d_bits12)
1403 HIGHBD_BFP(BLOCK_64X64,
1404 vpx_highbd_sad64x64_bits12,
1405 vpx_highbd_sad64x64_avg_bits12,
1406 vpx_highbd_12_variance64x64,
1407 vpx_highbd_12_sub_pixel_variance64x64,
1408 vpx_highbd_12_sub_pixel_avg_variance64x64,
1409 vpx_highbd_sad64x64x3_bits12,
1410 vpx_highbd_sad64x64x8_bits12,
1411 vpx_highbd_sad64x64x4d_bits12)
1413 HIGHBD_BFP(BLOCK_16X16,
1414 vpx_highbd_sad16x16_bits12,
1415 vpx_highbd_sad16x16_avg_bits12,
1416 vpx_highbd_12_variance16x16,
1417 vpx_highbd_12_sub_pixel_variance16x16,
1418 vpx_highbd_12_sub_pixel_avg_variance16x16,
1419 vpx_highbd_sad16x16x3_bits12,
1420 vpx_highbd_sad16x16x8_bits12,
1421 vpx_highbd_sad16x16x4d_bits12)
1423 HIGHBD_BFP(BLOCK_16X8,
1424 vpx_highbd_sad16x8_bits12,
1425 vpx_highbd_sad16x8_avg_bits12,
1426 vpx_highbd_12_variance16x8,
1427 vpx_highbd_12_sub_pixel_variance16x8,
1428 vpx_highbd_12_sub_pixel_avg_variance16x8,
1429 vpx_highbd_sad16x8x3_bits12,
1430 vpx_highbd_sad16x8x8_bits12,
1431 vpx_highbd_sad16x8x4d_bits12)
1433 HIGHBD_BFP(BLOCK_8X16,
1434 vpx_highbd_sad8x16_bits12,
1435 vpx_highbd_sad8x16_avg_bits12,
1436 vpx_highbd_12_variance8x16,
1437 vpx_highbd_12_sub_pixel_variance8x16,
1438 vpx_highbd_12_sub_pixel_avg_variance8x16,
1439 vpx_highbd_sad8x16x3_bits12,
1440 vpx_highbd_sad8x16x8_bits12,
1441 vpx_highbd_sad8x16x4d_bits12)
1443 HIGHBD_BFP(BLOCK_8X8,
1444 vpx_highbd_sad8x8_bits12,
1445 vpx_highbd_sad8x8_avg_bits12,
1446 vpx_highbd_12_variance8x8,
1447 vpx_highbd_12_sub_pixel_variance8x8,
1448 vpx_highbd_12_sub_pixel_avg_variance8x8,
1449 vpx_highbd_sad8x8x3_bits12,
1450 vpx_highbd_sad8x8x8_bits12,
1451 vpx_highbd_sad8x8x4d_bits12)
1453 HIGHBD_BFP(BLOCK_8X4,
1454 vpx_highbd_sad8x4_bits12,
1455 vpx_highbd_sad8x4_avg_bits12,
1456 vpx_highbd_12_variance8x4,
1457 vpx_highbd_12_sub_pixel_variance8x4,
1458 vpx_highbd_12_sub_pixel_avg_variance8x4,
1460 vpx_highbd_sad8x4x8_bits12,
1461 vpx_highbd_sad8x4x4d_bits12)
1463 HIGHBD_BFP(BLOCK_4X8,
1464 vpx_highbd_sad4x8_bits12,
1465 vpx_highbd_sad4x8_avg_bits12,
1466 vpx_highbd_12_variance4x8,
1467 vpx_highbd_12_sub_pixel_variance4x8,
1468 vpx_highbd_12_sub_pixel_avg_variance4x8,
1470 vpx_highbd_sad4x8x8_bits12,
1471 vpx_highbd_sad4x8x4d_bits12)
1473 HIGHBD_BFP(BLOCK_4X4,
1474 vpx_highbd_sad4x4_bits12,
1475 vpx_highbd_sad4x4_avg_bits12,
1476 vpx_highbd_12_variance4x4,
1477 vpx_highbd_12_sub_pixel_variance4x4,
1478 vpx_highbd_12_sub_pixel_avg_variance4x4,
1479 vpx_highbd_sad4x4x3_bits12,
1480 vpx_highbd_sad4x4x8_bits12,
1481 vpx_highbd_sad4x4x4d_bits12)
1485 assert(0 && "cm->bit_depth should be VPX_BITS_8, "
1486 "VPX_BITS_10 or VPX_BITS_12");
1490 #endif // CONFIG_VP9_HIGHBITDEPTH
1492 static void realloc_segmentation_maps(VP9_COMP *cpi) {
1493 VP9_COMMON *const cm = &cpi->common;
1495 // Create the encoder segmentation map and set all entries to 0
1496 vpx_free(cpi->segmentation_map);
1497 CHECK_MEM_ERROR(cm, cpi->segmentation_map,
1498 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1500 // Create a map used for cyclic background refresh.
1501 if (cpi->cyclic_refresh)
1502 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
1503 CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
1504 vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
1506 // Create a map used to mark inactive areas.
1507 vpx_free(cpi->active_map.map);
1508 CHECK_MEM_ERROR(cm, cpi->active_map.map,
1509 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1511 // And a place holder structure is the coding context
1512 // for use if we want to save and restore it
1513 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
1514 CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
1515 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1518 void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
1519 VP9_COMMON *const cm = &cpi->common;
1520 RATE_CONTROL *const rc = &cpi->rc;
1521 int last_w = cpi->oxcf.width;
1522 int last_h = cpi->oxcf.height;
1524 if (cm->profile != oxcf->profile)
1525 cm->profile = oxcf->profile;
1526 cm->bit_depth = oxcf->bit_depth;
1527 cm->color_space = oxcf->color_space;
1528 cm->color_range = oxcf->color_range;
1530 cpi->target_level = oxcf->target_level;
1531 cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
1533 if (cm->profile <= PROFILE_1)
1534 assert(cm->bit_depth == VPX_BITS_8);
1536 assert(cm->bit_depth > VPX_BITS_8);
1539 #if CONFIG_VP9_HIGHBITDEPTH
1540 cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
1541 #endif // CONFIG_VP9_HIGHBITDEPTH
1543 if ((oxcf->pass == 0) && (oxcf->rc_mode == VPX_Q)) {
1544 rc->baseline_gf_interval = FIXED_GF_INTERVAL;
1546 rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
1549 cpi->refresh_golden_frame = 0;
1550 cpi->refresh_last_frame = 1;
1551 cm->refresh_frame_context = 1;
1552 cm->reset_frame_context = 0;
1554 vp9_reset_segment_features(&cm->seg);
1555 vp9_set_high_precision_mv(cpi, 0);
1560 for (i = 0; i < MAX_SEGMENTS; i++)
1561 cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
1563 cpi->encode_breakout = cpi->oxcf.encode_breakout;
1565 set_rc_buffer_sizes(rc, &cpi->oxcf);
1567 // Under a configuration change, where maximum_buffer_size may change,
1568 // keep buffer level clipped to the maximum allowed buffer size.
1569 rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
1570 rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size);
1572 // Set up frame rate and related parameters rate control values.
1573 vp9_new_framerate(cpi, cpi->framerate);
1575 // Set absolute upper and lower quality limits
1576 rc->worst_quality = cpi->oxcf.worst_allowed_q;
1577 rc->best_quality = cpi->oxcf.best_allowed_q;
1579 cm->interp_filter = cpi->sf.default_interp_filter;
1581 if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) {
1582 cm->render_width = cpi->oxcf.render_width;
1583 cm->render_height = cpi->oxcf.render_height;
1585 cm->render_width = cpi->oxcf.width;
1586 cm->render_height = cpi->oxcf.height;
1588 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1589 cm->width = cpi->oxcf.width;
1590 cm->height = cpi->oxcf.height;
1591 cpi->external_resize = 1;
1594 if (cpi->initial_width) {
1595 int new_mi_size = 0;
1596 vp9_set_mb_mi(cm, cm->width, cm->height);
1597 new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
1598 if (cm->mi_alloc_size < new_mi_size) {
1599 vp9_free_context_buffers(cm);
1600 alloc_compressor_data(cpi);
1601 realloc_segmentation_maps(cpi);
1602 cpi->initial_width = cpi->initial_height = 0;
1603 cpi->external_resize = 0;
1604 } else if (cm->mi_alloc_size == new_mi_size &&
1605 (cpi->oxcf.width > last_w || cpi->oxcf.height > last_h)) {
1606 vp9_alloc_loop_filter(cm);
1610 update_frame_size(cpi);
1612 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1613 memset(cpi->consec_zero_mv, 0,
1614 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
1615 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
1616 vp9_cyclic_refresh_reset_resize(cpi);
1619 if ((cpi->svc.number_temporal_layers > 1 &&
1620 cpi->oxcf.rc_mode == VPX_CBR) ||
1621 ((cpi->svc.number_temporal_layers > 1 ||
1622 cpi->svc.number_spatial_layers > 1) &&
1623 cpi->oxcf.pass != 1)) {
1624 vp9_update_layer_context_change_config(cpi,
1625 (int)cpi->oxcf.target_bandwidth);
1628 cpi->alt_ref_source = NULL;
1629 rc->is_src_frame_alt_ref = 0;
1632 // Experimental RD Code
1633 cpi->frame_distortion = 0;
1634 cpi->last_frame_distortion = 0;
1637 set_tile_limits(cpi);
1639 cpi->ext_refresh_frame_flags_pending = 0;
1640 cpi->ext_refresh_frame_context_pending = 0;
1642 #if CONFIG_VP9_HIGHBITDEPTH
1643 highbd_set_var_fns(cpi);
1648 #define M_LOG2_E 0.693147180559945309417
1650 #define log2f(x) (log (x) / (float) M_LOG2_E)
1652 /***********************************************************************
1653 * Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts' *
1654 ***********************************************************************
1655 * The following 2 functions ('cal_nmvjointsadcost' and *
1656 * 'cal_nmvsadcosts') are used to calculate cost lookup tables *
1657 * used by 'vp9_diamond_search_sad'. The C implementation of the *
1658 * function is generic, but the AVX intrinsics optimised version *
1659 * relies on the following properties of the computed tables: *
1660 * For cal_nmvjointsadcost: *
1661 * - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] *
1662 * For cal_nmvsadcosts: *
1663 * - For all i: mvsadcost[0][i] == mvsadcost[1][i] *
1664 * (Equal costs for both components) *
1665 * - For all i: mvsadcost[0][i] == mvsadcost[0][-i] *
1666 * (Cost function is even) *
1667 * If these do not hold, then the AVX optimised version of the *
1668 * 'vp9_diamond_search_sad' function cannot be used as it is, in which *
1669 * case you can revert to using the C function instead. *
1670 ***********************************************************************/
1672 static void cal_nmvjointsadcost(int *mvjointsadcost) {
1673 /*********************************************************************
1674 * Warning: Read the comments above before modifying this function *
1675 *********************************************************************/
1676 mvjointsadcost[0] = 600;
1677 mvjointsadcost[1] = 300;
1678 mvjointsadcost[2] = 300;
1679 mvjointsadcost[3] = 300;
1682 static void cal_nmvsadcosts(int *mvsadcost[2]) {
1683 /*********************************************************************
1684 * Warning: Read the comments above before modifying this function *
1685 *********************************************************************/
1688 mvsadcost[0][0] = 0;
1689 mvsadcost[1][0] = 0;
1692 double z = 256 * (2 * (log2f(8 * i) + .6));
1693 mvsadcost[0][i] = (int)z;
1694 mvsadcost[1][i] = (int)z;
1695 mvsadcost[0][-i] = (int)z;
1696 mvsadcost[1][-i] = (int)z;
1697 } while (++i <= MV_MAX);
1700 static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
1703 mvsadcost[0][0] = 0;
1704 mvsadcost[1][0] = 0;
1707 double z = 256 * (2 * (log2f(8 * i) + .6));
1708 mvsadcost[0][i] = (int)z;
1709 mvsadcost[1][i] = (int)z;
1710 mvsadcost[0][-i] = (int)z;
1711 mvsadcost[1][-i] = (int)z;
1712 } while (++i <= MV_MAX);
1715 VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
1716 BufferPool *const pool) {
1718 VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
1719 VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
1726 if (setjmp(cm->error.jmp)) {
1727 cm->error.setjmp = 0;
1728 vp9_remove_compressor(cpi);
1732 cm->error.setjmp = 1;
1733 cm->alloc_mi = vp9_enc_alloc_mi;
1734 cm->free_mi = vp9_enc_free_mi;
1735 cm->setup_mi = vp9_enc_setup_mi;
1737 CHECK_MEM_ERROR(cm, cm->fc,
1738 (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
1739 CHECK_MEM_ERROR(cm, cm->frame_contexts,
1740 (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
1741 sizeof(*cm->frame_contexts)));
1744 cpi->resize_state = 0;
1745 cpi->external_resize = 0;
1746 cpi->resize_avg_qp = 0;
1747 cpi->resize_buffer_underflow = 0;
1748 cpi->use_skin_detection = 0;
1749 cpi->common.buffer_pool = pool;
1751 init_config(cpi, oxcf);
1752 vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
1754 cm->current_video_frame = 0;
1755 cpi->partition_search_skippable_frame = 0;
1756 cpi->tile_data = NULL;
1758 realloc_segmentation_maps(cpi);
1760 CHECK_MEM_ERROR(cm, cpi->consec_zero_mv,
1761 vpx_calloc(cm->mi_rows * cm->mi_cols,
1762 sizeof(*cpi->consec_zero_mv)));
1764 CHECK_MEM_ERROR(cm, cpi->nmvcosts[0],
1765 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0])));
1766 CHECK_MEM_ERROR(cm, cpi->nmvcosts[1],
1767 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1])));
1768 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0],
1769 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0])));
1770 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1],
1771 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1])));
1772 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0],
1773 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0])));
1774 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1],
1775 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1])));
1776 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0],
1777 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0])));
1778 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
1779 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
1781 for (i = 0; i < (sizeof(cpi->mbgraph_stats) /
1782 sizeof(cpi->mbgraph_stats[0])); i++) {
1783 CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats,
1784 vpx_calloc(cm->MBs *
1785 sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
1788 #if CONFIG_FP_MB_STATS
1789 cpi->use_fp_mb_stats = 0;
1790 if (cpi->use_fp_mb_stats) {
1791 // a place holder used to store the first pass mb stats in the first pass
1792 CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
1793 vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
1795 cpi->twopass.frame_mb_stats_buf = NULL;
1799 cpi->refresh_alt_ref_frame = 0;
1800 cpi->multi_arf_last_grp_enabled = 0;
1802 cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
1804 init_level_info(&cpi->level_info);
1806 #if CONFIG_INTERNAL_STATS
1807 cpi->b_calculate_ssimg = 0;
1808 cpi->b_calculate_blockiness = 1;
1809 cpi->b_calculate_consistency = 1;
1810 cpi->total_inconsistency = 0;
1811 cpi->psnr.worst = 100.0;
1812 cpi->worst_ssim = 100.0;
1817 if (cpi->b_calculate_psnr) {
1818 cpi->total_sq_error = 0;
1819 cpi->total_samples = 0;
1821 cpi->totalp_sq_error = 0;
1822 cpi->totalp_samples = 0;
1824 cpi->tot_recode_hits = 0;
1825 cpi->summed_quality = 0;
1826 cpi->summed_weights = 0;
1827 cpi->summedp_quality = 0;
1828 cpi->summedp_weights = 0;
1831 if (cpi->b_calculate_ssimg) {
1832 cpi->ssimg.worst= 100.0;
1834 cpi->fastssim.worst = 100.0;
1836 cpi->psnrhvs.worst = 100.0;
1838 if (cpi->b_calculate_blockiness) {
1839 cpi->total_blockiness = 0;
1840 cpi->worst_blockiness = 0.0;
1843 if (cpi->b_calculate_consistency) {
1844 CHECK_MEM_ERROR(cm, cpi->ssim_vars,
1845 vpx_malloc(sizeof(*cpi->ssim_vars) * 4 *
1846 cpi->common.mi_rows * cpi->common.mi_cols));
1847 cpi->worst_consistency = 100.0;
1852 cpi->first_time_stamp_ever = INT64_MAX;
1854 /*********************************************************************
1855 * Warning: Read the comments around 'cal_nmvjointsadcost' and *
1856 * 'cal_nmvsadcosts' before modifying how these tables are computed. *
1857 *********************************************************************/
1858 cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
1859 cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
1860 cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
1861 cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
1862 cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
1863 cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
1865 cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
1866 cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
1867 cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
1868 cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
1869 cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
1871 #if CONFIG_VP9_TEMPORAL_DENOISING
1872 #ifdef OUTPUT_YUV_DENOISED
1873 yuv_denoised_file = fopen("denoised.yuv", "ab");
1876 #ifdef OUTPUT_YUV_SKINMAP
1877 yuv_skinmap_file = fopen("skinmap.yuv", "ab");
1879 #ifdef OUTPUT_YUV_REC
1880 yuv_rec_file = fopen("rec.yuv", "wb");
1884 framepsnr = fopen("framepsnr.stt", "a");
1885 kf_list = fopen("kf_list.stt", "w");
1888 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
1890 if (oxcf->pass == 1) {
1891 vp9_init_first_pass(cpi);
1892 } else if (oxcf->pass == 2) {
1893 const size_t packet_sz = sizeof(FIRSTPASS_STATS);
1894 const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
1896 if (cpi->svc.number_spatial_layers > 1
1897 || cpi->svc.number_temporal_layers > 1) {
1898 FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf;
1899 FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = {0};
1902 for (i = 0; i < oxcf->ss_number_layers; ++i) {
1903 FIRSTPASS_STATS *const last_packet_for_layer =
1904 &stats[packets - oxcf->ss_number_layers + i];
1905 const int layer_id = (int)last_packet_for_layer->spatial_layer_id;
1906 const int packets_in_layer = (int)last_packet_for_layer->count + 1;
1907 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) {
1908 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id];
1910 vpx_free(lc->rc_twopass_stats_in.buf);
1912 lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz;
1913 CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf,
1914 vpx_malloc(lc->rc_twopass_stats_in.sz));
1915 lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf;
1916 lc->twopass.stats_in = lc->twopass.stats_in_start;
1917 lc->twopass.stats_in_end = lc->twopass.stats_in_start
1918 + packets_in_layer - 1;
1919 stats_copy[layer_id] = lc->rc_twopass_stats_in.buf;
1923 for (i = 0; i < packets; ++i) {
1924 const int layer_id = (int)stats[i].spatial_layer_id;
1925 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers
1926 && stats_copy[layer_id] != NULL) {
1927 *stats_copy[layer_id] = stats[i];
1928 ++stats_copy[layer_id];
1932 vp9_init_second_pass_spatial_svc(cpi);
1934 #if CONFIG_FP_MB_STATS
1935 if (cpi->use_fp_mb_stats) {
1936 const size_t psz = cpi->common.MBs * sizeof(uint8_t);
1937 const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
1939 cpi->twopass.firstpass_mb_stats.mb_stats_start =
1940 oxcf->firstpass_mb_stats_in.buf;
1941 cpi->twopass.firstpass_mb_stats.mb_stats_end =
1942 cpi->twopass.firstpass_mb_stats.mb_stats_start +
1943 (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
1947 cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
1948 cpi->twopass.stats_in = cpi->twopass.stats_in_start;
1949 cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
1951 vp9_init_second_pass(cpi);
1955 vp9_set_speed_features_framesize_independent(cpi);
1956 vp9_set_speed_features_framesize_dependent(cpi);
1958 // Allocate memory to store variances for a frame.
1959 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
1960 vpx_calloc(cm->MBs, sizeof(diff)));
1961 cpi->source_var_thresh = 0;
1962 cpi->frames_till_next_var_check = 0;
1964 #define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\
1965 cpi->fn_ptr[BT].sdf = SDF; \
1966 cpi->fn_ptr[BT].sdaf = SDAF; \
1967 cpi->fn_ptr[BT].vf = VF; \
1968 cpi->fn_ptr[BT].svf = SVF; \
1969 cpi->fn_ptr[BT].svaf = SVAF; \
1970 cpi->fn_ptr[BT].sdx3f = SDX3F; \
1971 cpi->fn_ptr[BT].sdx8f = SDX8F; \
1972 cpi->fn_ptr[BT].sdx4df = SDX4DF;
1974 BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg,
1975 vpx_variance32x16, vpx_sub_pixel_variance32x16,
1976 vpx_sub_pixel_avg_variance32x16, NULL, NULL, vpx_sad32x16x4d)
1978 BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg,
1979 vpx_variance16x32, vpx_sub_pixel_variance16x32,
1980 vpx_sub_pixel_avg_variance16x32, NULL, NULL, vpx_sad16x32x4d)
1982 BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg,
1983 vpx_variance64x32, vpx_sub_pixel_variance64x32,
1984 vpx_sub_pixel_avg_variance64x32, NULL, NULL, vpx_sad64x32x4d)
1986 BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg,
1987 vpx_variance32x64, vpx_sub_pixel_variance32x64,
1988 vpx_sub_pixel_avg_variance32x64, NULL, NULL, vpx_sad32x64x4d)
1990 BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg,
1991 vpx_variance32x32, vpx_sub_pixel_variance32x32,
1992 vpx_sub_pixel_avg_variance32x32, vpx_sad32x32x3, vpx_sad32x32x8,
1995 BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg,
1996 vpx_variance64x64, vpx_sub_pixel_variance64x64,
1997 vpx_sub_pixel_avg_variance64x64, vpx_sad64x64x3, vpx_sad64x64x8,
2000 BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg,
2001 vpx_variance16x16, vpx_sub_pixel_variance16x16,
2002 vpx_sub_pixel_avg_variance16x16, vpx_sad16x16x3, vpx_sad16x16x8,
2005 BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg,
2006 vpx_variance16x8, vpx_sub_pixel_variance16x8,
2007 vpx_sub_pixel_avg_variance16x8,
2008 vpx_sad16x8x3, vpx_sad16x8x8, vpx_sad16x8x4d)
2010 BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg,
2011 vpx_variance8x16, vpx_sub_pixel_variance8x16,
2012 vpx_sub_pixel_avg_variance8x16,
2013 vpx_sad8x16x3, vpx_sad8x16x8, vpx_sad8x16x4d)
2015 BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg,
2016 vpx_variance8x8, vpx_sub_pixel_variance8x8,
2017 vpx_sub_pixel_avg_variance8x8,
2018 vpx_sad8x8x3, vpx_sad8x8x8, vpx_sad8x8x4d)
2020 BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg,
2021 vpx_variance8x4, vpx_sub_pixel_variance8x4,
2022 vpx_sub_pixel_avg_variance8x4, NULL, vpx_sad8x4x8, vpx_sad8x4x4d)
2024 BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg,
2025 vpx_variance4x8, vpx_sub_pixel_variance4x8,
2026 vpx_sub_pixel_avg_variance4x8, NULL, vpx_sad4x8x8, vpx_sad4x8x4d)
2028 BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg,
2029 vpx_variance4x4, vpx_sub_pixel_variance4x4,
2030 vpx_sub_pixel_avg_variance4x4,
2031 vpx_sad4x4x3, vpx_sad4x4x8, vpx_sad4x4x4d)
2033 #if CONFIG_VP9_HIGHBITDEPTH
2034 highbd_set_var_fns(cpi);
2037 /* vp9_init_quantizer() is first called here. Add check in
2038 * vp9_frame_init_quantizer() so that vp9_init_quantizer is only
2039 * called later when needed. This will avoid unnecessary calls of
2040 * vp9_init_quantizer() for every frame.
2042 vp9_init_quantizer(cpi);
2044 vp9_loop_filter_init(cm);
2046 cm->error.setjmp = 0;
2051 #if CONFIG_INTERNAL_STATS
2052 #define SNPRINT(H, T) \
2053 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
2055 #define SNPRINT2(H, T, V) \
2056 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
2057 #endif // CONFIG_INTERNAL_STATS
2059 void vp9_remove_compressor(VP9_COMP *cpi) {
2068 if (cm->current_video_frame > 0) {
2069 #if CONFIG_INTERNAL_STATS
2070 vpx_clear_system_state();
2072 if (cpi->oxcf.pass != 1) {
2073 char headings[512] = {0};
2074 char results[512] = {0};
2075 FILE *f = fopen("opsnr.stt", "a");
2076 double time_encoded = (cpi->last_end_time_stamp_seen
2077 - cpi->first_time_stamp_ever) / 10000000.000;
2078 double total_encode_time = (cpi->time_receive_data +
2079 cpi->time_compress_data) / 1000.000;
2081 (double)cpi->bytes * (double) 8 / (double)1000 / time_encoded;
2082 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
2083 const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
2084 const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
2086 if (cpi->b_calculate_psnr) {
2087 const double total_psnr =
2088 vpx_sse_to_psnr((double)cpi->total_samples, peak,
2089 (double)cpi->total_sq_error);
2090 const double totalp_psnr =
2091 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
2092 (double)cpi->totalp_sq_error);
2093 const double total_ssim = 100 * pow(cpi->summed_quality /
2094 cpi->summed_weights, 8.0);
2095 const double totalp_ssim = 100 * pow(cpi->summedp_quality /
2096 cpi->summedp_weights, 8.0);
2098 snprintf(headings, sizeof(headings),
2099 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
2100 "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
2101 "WstPsnr\tWstSsim\tWstFast\tWstHVS");
2102 snprintf(results, sizeof(results),
2103 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2104 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2105 "%7.3f\t%7.3f\t%7.3f\t%7.3f",
2106 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
2107 cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr,
2108 total_ssim, totalp_ssim,
2109 cpi->fastssim.stat[ALL] / cpi->count,
2110 cpi->psnrhvs.stat[ALL] / cpi->count,
2111 cpi->psnr.worst, cpi->worst_ssim, cpi->fastssim.worst,
2112 cpi->psnrhvs.worst);
2114 if (cpi->b_calculate_blockiness) {
2115 SNPRINT(headings, "\t Block\tWstBlck");
2116 SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
2117 SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
2120 if (cpi->b_calculate_consistency) {
2121 double consistency =
2122 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
2123 (double)cpi->total_inconsistency);
2125 SNPRINT(headings, "\tConsist\tWstCons");
2126 SNPRINT2(results, "\t%7.3f", consistency);
2127 SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
2130 if (cpi->b_calculate_ssimg) {
2131 SNPRINT(headings, "\t SSIMG\tWtSSIMG");
2132 SNPRINT2(results, "\t%7.3f", cpi->ssimg.stat[ALL] / cpi->count);
2133 SNPRINT2(results, "\t%7.3f", cpi->ssimg.worst);
2136 fprintf(f, "%s\t Time Rc-Err Abs Err\n", headings);
2137 fprintf(f, "%s\t%8.0f %7.2f %7.2f\n", results,
2138 total_encode_time, rate_err, fabs(rate_err));
2148 printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
2149 printf("\n_frames recive_data encod_mb_row compress_frame Total\n");
2150 printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
2151 cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
2152 cpi->time_compress_data / 1000,
2153 (cpi->time_receive_data + cpi->time_compress_data) / 1000);
2158 #if CONFIG_VP9_TEMPORAL_DENOISING
2159 vp9_denoiser_free(&(cpi->denoiser));
2162 for (t = 0; t < cpi->num_workers; ++t) {
2163 VPxWorker *const worker = &cpi->workers[t];
2164 EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
2166 // Deallocate allocated threads.
2167 vpx_get_worker_interface()->end(worker);
2169 // Deallocate allocated thread data.
2170 if (t < cpi->num_workers - 1) {
2171 vpx_free(thread_data->td->counts);
2172 vp9_free_pc_tree(thread_data->td);
2173 vpx_free(thread_data->td);
2176 vpx_free(cpi->tile_thr_data);
2177 vpx_free(cpi->workers);
2179 if (cpi->num_workers > 1)
2180 vp9_loop_filter_dealloc(&cpi->lf_row_sync);
2182 dealloc_compressor_data(cpi);
2184 for (i = 0; i < sizeof(cpi->mbgraph_stats) /
2185 sizeof(cpi->mbgraph_stats[0]); ++i) {
2186 vpx_free(cpi->mbgraph_stats[i].mb_stats);
2189 #if CONFIG_FP_MB_STATS
2190 if (cpi->use_fp_mb_stats) {
2191 vpx_free(cpi->twopass.frame_mb_stats_buf);
2192 cpi->twopass.frame_mb_stats_buf = NULL;
2196 vp9_remove_common(cm);
2197 vp9_free_ref_frame_buffers(cm->buffer_pool);
2198 #if CONFIG_VP9_POSTPROC
2199 vp9_free_postproc_buffers(cm);
2203 #if CONFIG_VP9_TEMPORAL_DENOISING
2204 #ifdef OUTPUT_YUV_DENOISED
2205 fclose(yuv_denoised_file);
2208 #ifdef OUTPUT_YUV_SKINMAP
2209 fclose(yuv_skinmap_file);
2211 #ifdef OUTPUT_YUV_REC
2212 fclose(yuv_rec_file);
2229 /* TODO(yaowu): The block_variance calls the unoptimized versions of variance()
2230 * and highbd_8_variance(). It should not.
2232 static void encoder_variance(const uint8_t *a, int a_stride,
2233 const uint8_t *b, int b_stride,
2234 int w, int h, unsigned int *sse, int *sum) {
2240 for (i = 0; i < h; i++) {
2241 for (j = 0; j < w; j++) {
2242 const int diff = a[j] - b[j];
2244 *sse += diff * diff;
2252 #if CONFIG_VP9_HIGHBITDEPTH
2253 static void encoder_highbd_variance64(const uint8_t *a8, int a_stride,
2254 const uint8_t *b8, int b_stride,
2255 int w, int h, uint64_t *sse,
2259 uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2260 uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2264 for (i = 0; i < h; i++) {
2265 for (j = 0; j < w; j++) {
2266 const int diff = a[j] - b[j];
2268 *sse += diff * diff;
2275 static void encoder_highbd_8_variance(const uint8_t *a8, int a_stride,
2276 const uint8_t *b8, int b_stride,
2278 unsigned int *sse, int *sum) {
2279 uint64_t sse_long = 0;
2280 uint64_t sum_long = 0;
2281 encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h,
2282 &sse_long, &sum_long);
2283 *sse = (unsigned int)sse_long;
2284 *sum = (int)sum_long;
2286 #endif // CONFIG_VP9_HIGHBITDEPTH
2288 static int64_t get_sse(const uint8_t *a, int a_stride,
2289 const uint8_t *b, int b_stride,
2290 int width, int height) {
2291 const int dw = width % 16;
2292 const int dh = height % 16;
2293 int64_t total_sse = 0;
2294 unsigned int sse = 0;
2299 encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride,
2300 dw, height, &sse, &sum);
2305 encoder_variance(&a[(height - dh) * a_stride], a_stride,
2306 &b[(height - dh) * b_stride], b_stride,
2307 width - dw, dh, &sse, &sum);
2311 for (y = 0; y < height / 16; ++y) {
2312 const uint8_t *pa = a;
2313 const uint8_t *pb = b;
2314 for (x = 0; x < width / 16; ++x) {
2315 vpx_mse16x16(pa, a_stride, pb, b_stride, &sse);
2329 #if CONFIG_VP9_HIGHBITDEPTH
2330 static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride,
2331 const uint8_t *b8, int b_stride,
2332 int width, int height,
2333 unsigned int input_shift) {
2334 const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2335 const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2336 int64_t total_sse = 0;
2338 for (y = 0; y < height; ++y) {
2339 for (x = 0; x < width; ++x) {
2341 diff = (a[x] >> input_shift) - (b[x] >> input_shift);
2342 total_sse += diff * diff;
2350 static int64_t highbd_get_sse(const uint8_t *a, int a_stride,
2351 const uint8_t *b, int b_stride,
2352 int width, int height) {
2353 int64_t total_sse = 0;
2355 const int dw = width % 16;
2356 const int dh = height % 16;
2357 unsigned int sse = 0;
2360 encoder_highbd_8_variance(&a[width - dw], a_stride,
2361 &b[width - dw], b_stride,
2362 dw, height, &sse, &sum);
2366 encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride,
2367 &b[(height - dh) * b_stride], b_stride,
2368 width - dw, dh, &sse, &sum);
2371 for (y = 0; y < height / 16; ++y) {
2372 const uint8_t *pa = a;
2373 const uint8_t *pb = b;
2374 for (x = 0; x < width / 16; ++x) {
2375 vpx_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse);
2385 #endif // CONFIG_VP9_HIGHBITDEPTH
2388 double psnr[4]; // total/y/u/v
2389 uint64_t sse[4]; // total/y/u/v
2390 uint32_t samples[4]; // total/y/u/v
2393 #if CONFIG_VP9_HIGHBITDEPTH
2394 static void calc_highbd_psnr(const YV12_BUFFER_CONFIG *a,
2395 const YV12_BUFFER_CONFIG *b,
2397 unsigned int bit_depth,
2398 unsigned int in_bit_depth) {
2399 const int widths[3] =
2400 {a->y_crop_width, a->uv_crop_width, a->uv_crop_width };
2401 const int heights[3] =
2402 {a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2403 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer };
2404 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2405 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer };
2406 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2408 uint64_t total_sse = 0;
2409 uint32_t total_samples = 0;
2410 const double peak = (double)((1 << in_bit_depth) - 1);
2411 const unsigned int input_shift = bit_depth - in_bit_depth;
2413 for (i = 0; i < 3; ++i) {
2414 const int w = widths[i];
2415 const int h = heights[i];
2416 const uint32_t samples = w * h;
2418 if (a->flags & YV12_FLAG_HIGHBITDEPTH) {
2420 sse = highbd_get_sse_shift(a_planes[i], a_strides[i],
2421 b_planes[i], b_strides[i], w, h,
2424 sse = highbd_get_sse(a_planes[i], a_strides[i],
2425 b_planes[i], b_strides[i], w, h);
2428 sse = get_sse(a_planes[i], a_strides[i],
2429 b_planes[i], b_strides[i],
2432 psnr->sse[1 + i] = sse;
2433 psnr->samples[1 + i] = samples;
2434 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2437 total_samples += samples;
2440 psnr->sse[0] = total_sse;
2441 psnr->samples[0] = total_samples;
2442 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2446 #else // !CONFIG_VP9_HIGHBITDEPTH
2448 static void calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b,
2450 static const double peak = 255.0;
2451 const int widths[3] = {
2452 a->y_crop_width, a->uv_crop_width, a->uv_crop_width};
2453 const int heights[3] = {
2454 a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2455 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer};
2456 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2457 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer};
2458 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2460 uint64_t total_sse = 0;
2461 uint32_t total_samples = 0;
2463 for (i = 0; i < 3; ++i) {
2464 const int w = widths[i];
2465 const int h = heights[i];
2466 const uint32_t samples = w * h;
2467 const uint64_t sse = get_sse(a_planes[i], a_strides[i],
2468 b_planes[i], b_strides[i],
2470 psnr->sse[1 + i] = sse;
2471 psnr->samples[1 + i] = samples;
2472 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2475 total_samples += samples;
2478 psnr->sse[0] = total_sse;
2479 psnr->samples[0] = total_samples;
2480 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2483 #endif // CONFIG_VP9_HIGHBITDEPTH
2485 static void generate_psnr_packet(VP9_COMP *cpi) {
2486 struct vpx_codec_cx_pkt pkt;
2489 #if CONFIG_VP9_HIGHBITDEPTH
2490 calc_highbd_psnr(cpi->Source, cpi->common.frame_to_show, &psnr,
2491 cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
2493 calc_psnr(cpi->Source, cpi->common.frame_to_show, &psnr);
2496 for (i = 0; i < 4; ++i) {
2497 pkt.data.psnr.samples[i] = psnr.samples[i];
2498 pkt.data.psnr.sse[i] = psnr.sse[i];
2499 pkt.data.psnr.psnr[i] = psnr.psnr[i];
2501 pkt.kind = VPX_CODEC_PSNR_PKT;
2503 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
2504 cpi->svc.number_temporal_layers].psnr_pkt = pkt.data.psnr;
2506 vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
2509 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) {
2510 if (ref_frame_flags > 7)
2513 cpi->ref_frame_flags = ref_frame_flags;
2517 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) {
2518 cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0;
2519 cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0;
2520 cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0;
2521 cpi->ext_refresh_frame_flags_pending = 1;
2524 static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(VP9_COMP *cpi,
2525 VP9_REFFRAME ref_frame_flag) {
2526 MV_REFERENCE_FRAME ref_frame = NONE;
2527 if (ref_frame_flag == VP9_LAST_FLAG)
2528 ref_frame = LAST_FRAME;
2529 else if (ref_frame_flag == VP9_GOLD_FLAG)
2530 ref_frame = GOLDEN_FRAME;
2531 else if (ref_frame_flag == VP9_ALT_FLAG)
2532 ref_frame = ALTREF_FRAME;
2534 return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame);
2537 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2538 YV12_BUFFER_CONFIG *sd) {
2539 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2541 vp8_yv12_copy_frame(cfg, sd);
2548 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2549 YV12_BUFFER_CONFIG *sd) {
2550 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2552 vp8_yv12_copy_frame(sd, cfg);
2559 int vp9_update_entropy(VP9_COMP * cpi, int update) {
2560 cpi->ext_refresh_frame_context = update;
2561 cpi->ext_refresh_frame_context_pending = 1;
2565 #if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP)
2566 // The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
2567 // as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
2568 // not denoise the UV channels at this time. If ever we implement UV channel
2569 // denoising we will have to modify this.
2570 void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
2571 uint8_t *src = s->y_buffer;
2572 int h = s->y_height;
2575 fwrite(src, s->y_width, 1, f);
2583 fwrite(src, s->uv_width, 1, f);
2584 src += s->uv_stride;
2591 fwrite(src, s->uv_width, 1, f);
2592 src += s->uv_stride;
2597 #ifdef OUTPUT_YUV_REC
2598 void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
2599 YV12_BUFFER_CONFIG *s = cm->frame_to_show;
2600 uint8_t *src = s->y_buffer;
2603 #if CONFIG_VP9_HIGHBITDEPTH
2604 if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
2605 uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
2608 fwrite(src16, s->y_width, 2, yuv_rec_file);
2609 src16 += s->y_stride;
2612 src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
2616 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2617 src16 += s->uv_stride;
2620 src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
2624 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2625 src16 += s->uv_stride;
2628 fflush(yuv_rec_file);
2631 #endif // CONFIG_VP9_HIGHBITDEPTH
2634 fwrite(src, s->y_width, 1, yuv_rec_file);
2642 fwrite(src, s->uv_width, 1, yuv_rec_file);
2643 src += s->uv_stride;
2650 fwrite(src, s->uv_width, 1, yuv_rec_file);
2651 src += s->uv_stride;
2654 fflush(yuv_rec_file);
2658 #if CONFIG_VP9_HIGHBITDEPTH
2659 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2660 YV12_BUFFER_CONFIG *dst,
2663 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2664 YV12_BUFFER_CONFIG *dst) {
2665 #endif // CONFIG_VP9_HIGHBITDEPTH
2666 // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
2668 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2669 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2670 const int src_widths[3] = {src->y_crop_width, src->uv_crop_width,
2671 src->uv_crop_width };
2672 const int src_heights[3] = {src->y_crop_height, src->uv_crop_height,
2673 src->uv_crop_height};
2674 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2675 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2676 const int dst_widths[3] = {dst->y_crop_width, dst->uv_crop_width,
2677 dst->uv_crop_width};
2678 const int dst_heights[3] = {dst->y_crop_height, dst->uv_crop_height,
2679 dst->uv_crop_height};
2681 for (i = 0; i < MAX_MB_PLANE; ++i) {
2682 #if CONFIG_VP9_HIGHBITDEPTH
2683 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2684 vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
2685 src_strides[i], dsts[i], dst_heights[i],
2686 dst_widths[i], dst_strides[i], bd);
2688 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2689 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2692 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2693 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2694 #endif // CONFIG_VP9_HIGHBITDEPTH
2696 vpx_extend_frame_borders(dst);
2699 #if CONFIG_VP9_HIGHBITDEPTH
2700 static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
2701 YV12_BUFFER_CONFIG *dst, int bd) {
2702 const int src_w = src->y_crop_width;
2703 const int src_h = src->y_crop_height;
2704 const int dst_w = dst->y_crop_width;
2705 const int dst_h = dst->y_crop_height;
2706 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2707 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2708 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2709 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2710 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2713 for (i = 0; i < MAX_MB_PLANE; ++i) {
2714 const int factor = (i == 0 || i == 3 ? 1 : 2);
2715 const int src_stride = src_strides[i];
2716 const int dst_stride = dst_strides[i];
2717 for (y = 0; y < dst_h; y += 16) {
2718 const int y_q4 = y * (16 / factor) * src_h / dst_h;
2719 for (x = 0; x < dst_w; x += 16) {
2720 const int x_q4 = x * (16 / factor) * src_w / dst_w;
2721 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
2722 src_stride + (x / factor) * src_w / dst_w;
2723 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2725 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2726 vpx_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
2727 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2728 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2729 16 / factor, 16 / factor, bd);
2731 vpx_scaled_2d(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);
2740 vpx_extend_frame_borders(dst);
2743 void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
2744 YV12_BUFFER_CONFIG *dst) {
2745 const int src_w = src->y_crop_width;
2746 const int src_h = src->y_crop_height;
2747 const int dst_w = dst->y_crop_width;
2748 const int dst_h = dst->y_crop_height;
2749 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2750 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2751 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2752 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2753 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2756 for (i = 0; i < MAX_MB_PLANE; ++i) {
2757 const int factor = (i == 0 || i == 3 ? 1 : 2);
2758 const int src_stride = src_strides[i];
2759 const int dst_stride = dst_strides[i];
2760 for (y = 0; y < dst_h; y += 16) {
2761 const int y_q4 = y * (16 / factor) * src_h / dst_h;
2762 for (x = 0; x < dst_w; x += 16) {
2763 const int x_q4 = x * (16 / factor) * src_w / dst_w;
2764 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
2765 src_stride + (x / factor) * src_w / dst_w;
2766 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2768 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2769 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2770 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2771 16 / factor, 16 / factor);
2776 vpx_extend_frame_borders(dst);
2778 #endif // CONFIG_VP9_HIGHBITDEPTH
2780 static int scale_down(VP9_COMP *cpi, int q) {
2781 RATE_CONTROL *const rc = &cpi->rc;
2782 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2784 assert(frame_is_kf_gf_arf(cpi));
2786 if (rc->frame_size_selector == UNSCALED &&
2787 q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
2788 const int max_size_thresh = (int)(rate_thresh_mult[SCALE_STEP1]
2789 * VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
2790 scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
2795 static int big_rate_miss(VP9_COMP *cpi, int high_limit, int low_limit) {
2796 const RATE_CONTROL *const rc = &cpi->rc;
2798 return (rc->projected_frame_size > ((high_limit * 3) / 2)) ||
2799 (rc->projected_frame_size < (low_limit / 2));
2802 // Function to test for conditions that indicate we should loop
2803 // back and recode a frame.
2804 static int recode_loop_test(VP9_COMP *cpi,
2805 int high_limit, int low_limit,
2806 int q, int maxq, int minq) {
2807 const RATE_CONTROL *const rc = &cpi->rc;
2808 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2809 const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
2810 int force_recode = 0;
2812 if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
2813 big_rate_miss(cpi, high_limit, low_limit) ||
2814 (cpi->sf.recode_loop == ALLOW_RECODE) ||
2815 (frame_is_kfgfarf &&
2816 (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) {
2817 if (frame_is_kfgfarf &&
2818 (oxcf->resize_mode == RESIZE_DYNAMIC) &&
2819 scale_down(cpi, q)) {
2820 // Code this group at a lower resolution.
2821 cpi->resize_pending = 1;
2825 // TODO(agrange) high_limit could be greater than the scale-down threshold.
2826 if ((rc->projected_frame_size > high_limit && q < maxq) ||
2827 (rc->projected_frame_size < low_limit && q > minq)) {
2829 } else if (cpi->oxcf.rc_mode == VPX_CQ) {
2830 // Deal with frame undershoot and whether or not we are
2831 // below the automatically set cq level.
2832 if (q > oxcf->cq_level &&
2833 rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
2838 return force_recode;
2841 void vp9_update_reference_frames(VP9_COMP *cpi) {
2842 VP9_COMMON * const cm = &cpi->common;
2843 BufferPool *const pool = cm->buffer_pool;
2845 // At this point the new frame has been encoded.
2846 // If any buffer copy / swapping is signaled it should be done here.
2847 if (cm->frame_type == KEY_FRAME) {
2848 ref_cnt_fb(pool->frame_bufs,
2849 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2850 ref_cnt_fb(pool->frame_bufs,
2851 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2852 } else if (vp9_preserve_existing_gf(cpi)) {
2853 // We have decided to preserve the previously existing golden frame as our
2854 // new ARF frame. However, in the short term in function
2855 // vp9_get_refresh_mask() we left it in the GF slot and, if
2856 // we're updating the GF with the current decoded frame, we save it to the
2857 // ARF slot instead.
2858 // We now have to update the ARF with the current frame and swap gld_fb_idx
2859 // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
2860 // slot and, if we're updating the GF, the current frame becomes the new GF.
2863 ref_cnt_fb(pool->frame_bufs,
2864 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2866 tmp = cpi->alt_fb_idx;
2867 cpi->alt_fb_idx = cpi->gld_fb_idx;
2868 cpi->gld_fb_idx = tmp;
2870 if (is_two_pass_svc(cpi)) {
2871 cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx;
2872 cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx;
2874 } else { /* For non key/golden frames */
2875 if (cpi->refresh_alt_ref_frame) {
2876 int arf_idx = cpi->alt_fb_idx;
2877 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
2878 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2879 arf_idx = gf_group->arf_update_idx[gf_group->index];
2882 ref_cnt_fb(pool->frame_bufs,
2883 &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
2884 memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
2885 cpi->interp_filter_selected[0],
2886 sizeof(cpi->interp_filter_selected[0]));
2889 if (cpi->refresh_golden_frame) {
2890 ref_cnt_fb(pool->frame_bufs,
2891 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2892 if (!cpi->rc.is_src_frame_alt_ref)
2893 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2894 cpi->interp_filter_selected[0],
2895 sizeof(cpi->interp_filter_selected[0]));
2897 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2898 cpi->interp_filter_selected[ALTREF_FRAME],
2899 sizeof(cpi->interp_filter_selected[ALTREF_FRAME]));
2903 if (cpi->refresh_last_frame) {
2904 ref_cnt_fb(pool->frame_bufs,
2905 &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
2906 if (!cpi->rc.is_src_frame_alt_ref)
2907 memcpy(cpi->interp_filter_selected[LAST_FRAME],
2908 cpi->interp_filter_selected[0],
2909 sizeof(cpi->interp_filter_selected[0]));
2911 #if CONFIG_VP9_TEMPORAL_DENOISING
2912 if (cpi->oxcf.noise_sensitivity > 0 &&
2913 cpi->denoiser.denoising_level > kDenLowLow) {
2914 vp9_denoiser_update_frame_info(&cpi->denoiser,
2916 cpi->common.frame_type,
2917 cpi->refresh_alt_ref_frame,
2918 cpi->refresh_golden_frame,
2919 cpi->refresh_last_frame,
2920 cpi->resize_pending);
2923 if (is_one_pass_cbr_svc(cpi)) {
2924 // Keep track of frame index for each reference frame.
2925 SVC *const svc = &cpi->svc;
2926 if (cm->frame_type == KEY_FRAME) {
2927 svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2928 svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2929 svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2931 if (cpi->refresh_last_frame)
2932 svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2933 if (cpi->refresh_golden_frame)
2934 svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2935 if (cpi->refresh_alt_ref_frame)
2936 svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2941 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
2942 MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
2943 struct loopfilter *lf = &cm->lf;
2946 lf->filter_level = 0;
2947 lf->last_filt_level = 0;
2949 struct vpx_usec_timer timer;
2951 vpx_clear_system_state();
2953 vpx_usec_timer_start(&timer);
2955 if (!cpi->rc.is_src_frame_alt_ref) {
2956 if ((cpi->common.frame_type == KEY_FRAME) &&
2957 (!cpi->rc.this_key_frame_forced)) {
2958 lf->last_filt_level = 0;
2960 vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick);
2961 lf->last_filt_level = lf->filter_level;
2963 lf->filter_level = 0;
2966 vpx_usec_timer_mark(&timer);
2967 cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
2970 if (lf->filter_level > 0) {
2971 vp9_build_mask_frame(cm, lf->filter_level, 0);
2973 if (cpi->num_workers > 1)
2974 vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
2975 lf->filter_level, 0, 0,
2976 cpi->workers, cpi->num_workers,
2979 vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
2982 vpx_extend_frame_inner_borders(cm->frame_to_show);
2985 static INLINE void alloc_frame_mvs(VP9_COMMON *const cm,
2987 RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
2988 if (new_fb_ptr->mvs == NULL ||
2989 new_fb_ptr->mi_rows < cm->mi_rows ||
2990 new_fb_ptr->mi_cols < cm->mi_cols) {
2991 vpx_free(new_fb_ptr->mvs);
2992 CHECK_MEM_ERROR(cm, new_fb_ptr->mvs,
2993 (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
2994 sizeof(*new_fb_ptr->mvs)));
2995 new_fb_ptr->mi_rows = cm->mi_rows;
2996 new_fb_ptr->mi_cols = cm->mi_cols;
3000 void vp9_scale_references(VP9_COMP *cpi) {
3001 VP9_COMMON *cm = &cpi->common;
3002 MV_REFERENCE_FRAME ref_frame;
3003 const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
3005 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3006 // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
3007 if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
3008 BufferPool *const pool = cm->buffer_pool;
3009 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi,
3013 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3017 #if CONFIG_VP9_HIGHBITDEPTH
3018 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
3019 RefCntBuffer *new_fb_ptr = NULL;
3020 int force_scaling = 0;
3021 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
3022 if (new_fb == INVALID_IDX) {
3023 new_fb = get_free_fb(cm);
3026 if (new_fb == INVALID_IDX)
3028 new_fb_ptr = &pool->frame_bufs[new_fb];
3029 if (force_scaling ||
3030 new_fb_ptr->buf.y_crop_width != cm->width ||
3031 new_fb_ptr->buf.y_crop_height != cm->height) {
3032 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
3033 cm->subsampling_x, cm->subsampling_y,
3034 cm->use_highbitdepth,
3035 VP9_ENC_BORDER_IN_PIXELS,
3036 cm->byte_alignment, NULL, NULL, NULL))
3037 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3038 "Failed to allocate frame buffer");
3039 scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth);
3040 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
3041 alloc_frame_mvs(cm, new_fb);
3044 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
3045 RefCntBuffer *new_fb_ptr = NULL;
3046 int force_scaling = 0;
3047 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
3048 if (new_fb == INVALID_IDX) {
3049 new_fb = get_free_fb(cm);
3052 if (new_fb == INVALID_IDX)
3054 new_fb_ptr = &pool->frame_bufs[new_fb];
3055 if (force_scaling ||
3056 new_fb_ptr->buf.y_crop_width != cm->width ||
3057 new_fb_ptr->buf.y_crop_height != cm->height) {
3058 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
3059 cm->subsampling_x, cm->subsampling_y,
3060 VP9_ENC_BORDER_IN_PIXELS,
3061 cm->byte_alignment, NULL, NULL, NULL))
3062 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3063 "Failed to allocate frame buffer");
3064 vp9_scale_and_extend_frame(ref, &new_fb_ptr->buf);
3065 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
3066 alloc_frame_mvs(cm, new_fb);
3068 #endif // CONFIG_VP9_HIGHBITDEPTH
3071 RefCntBuffer *buf = NULL;
3072 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3073 // Check for release of scaled reference.
3074 buf_idx = cpi->scaled_ref_idx[ref_frame - 1];
3075 buf = (buf_idx != INVALID_IDX) ? &pool->frame_bufs[buf_idx] : NULL;
3078 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3081 buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3082 buf = &pool->frame_bufs[buf_idx];
3083 buf->buf.y_crop_width = ref->y_crop_width;
3084 buf->buf.y_crop_height = ref->y_crop_height;
3085 cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
3089 if (cpi->oxcf.pass != 0 || cpi->use_svc)
3090 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3095 static void release_scaled_references(VP9_COMP *cpi) {
3096 VP9_COMMON *cm = &cpi->common;
3098 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3099 // Only release scaled references under certain conditions:
3100 // if reference will be updated, or if scaled reference has same resolution.
3102 refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
3103 refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
3104 refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
3105 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3106 const int idx = cpi->scaled_ref_idx[i - 1];
3107 RefCntBuffer *const buf = idx != INVALID_IDX ?
3108 &cm->buffer_pool->frame_bufs[idx] : NULL;
3109 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
3112 (buf->buf.y_crop_width == ref->y_crop_width &&
3113 buf->buf.y_crop_height == ref->y_crop_height))) {
3115 cpi->scaled_ref_idx[i -1] = INVALID_IDX;
3119 for (i = 0; i < MAX_REF_FRAMES; ++i) {
3120 const int idx = cpi->scaled_ref_idx[i];
3121 RefCntBuffer *const buf = idx != INVALID_IDX ?
3122 &cm->buffer_pool->frame_bufs[idx] : NULL;
3125 cpi->scaled_ref_idx[i] = INVALID_IDX;
3131 static void full_to_model_count(unsigned int *model_count,
3132 unsigned int *full_count) {
3134 model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
3135 model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
3136 model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
3137 for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
3138 model_count[TWO_TOKEN] += full_count[n];
3139 model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
3142 static void full_to_model_counts(vp9_coeff_count_model *model_count,
3143 vp9_coeff_count *full_count) {
3146 for (i = 0; i < PLANE_TYPES; ++i)
3147 for (j = 0; j < REF_TYPES; ++j)
3148 for (k = 0; k < COEF_BANDS; ++k)
3149 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
3150 full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
3153 #if 0 && CONFIG_INTERNAL_STATS
3154 static void output_frame_level_debug_stats(VP9_COMP *cpi) {
3155 VP9_COMMON *const cm = &cpi->common;
3156 FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
3159 vpx_clear_system_state();
3161 #if CONFIG_VP9_HIGHBITDEPTH
3162 if (cm->use_highbitdepth) {
3163 recon_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3165 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3168 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3169 #endif // CONFIG_VP9_HIGHBITDEPTH
3172 if (cpi->twopass.total_left_stats.coded_error != 0.0) {
3173 double dc_quant_devisor;
3174 #if CONFIG_VP9_HIGHBITDEPTH
3175 switch (cm->bit_depth) {
3177 dc_quant_devisor = 4.0;
3180 dc_quant_devisor = 16.0;
3183 dc_quant_devisor = 64.0;
3186 assert(0 && "bit_depth must be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
3190 dc_quant_devisor = 4.0;
3193 fprintf(f, "%10u %dx%d %10d %10d %d %d %10d %10d %10d %10d"
3194 "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
3195 "%10"PRId64" %10"PRId64" %10d "
3196 "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
3197 "%6d %6d %5d %5d %5d "
3198 "%10"PRId64" %10.3lf"
3199 "%10lf %8u %10"PRId64" %10d %10d %10d %10d %10d\n",
3200 cpi->common.current_video_frame,
3201 cm->width, cm->height,
3202 cpi->td.rd_counts.m_search_count,
3203 cpi->td.rd_counts.ex_search_count,
3204 cpi->rc.source_alt_ref_pending,
3205 cpi->rc.source_alt_ref_active,
3206 cpi->rc.this_frame_target,
3207 cpi->rc.projected_frame_size,
3208 cpi->rc.projected_frame_size / cpi->common.MBs,
3209 (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
3210 cpi->rc.vbr_bits_off_target,
3211 cpi->rc.vbr_bits_off_target_fast,
3212 cpi->twopass.extend_minq,
3213 cpi->twopass.extend_minq_fast,
3214 cpi->rc.total_target_vs_actual,
3215 (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
3216 cpi->rc.total_actual_bits, cm->base_qindex,
3217 vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
3218 (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) /
3220 vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality,
3223 vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
3224 cpi->refresh_last_frame, cpi->refresh_golden_frame,
3225 cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
3226 cpi->twopass.bits_left,
3227 cpi->twopass.total_left_stats.coded_error,
3228 cpi->twopass.bits_left /
3229 (1 + cpi->twopass.total_left_stats.coded_error),
3230 cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
3231 cpi->twopass.kf_zeromotion_pct,
3232 cpi->twopass.fr_content_type,
3233 cm->lf.filter_level,
3234 cm->seg.aq_av_offset);
3239 FILE *const fmodes = fopen("Modes.stt", "a");
3242 fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
3243 cm->frame_type, cpi->refresh_golden_frame,
3244 cpi->refresh_alt_ref_frame);
3246 for (i = 0; i < MAX_MODES; ++i)
3247 fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
3249 fprintf(fmodes, "\n");
3256 static void set_mv_search_params(VP9_COMP *cpi) {
3257 const VP9_COMMON *const cm = &cpi->common;
3258 const unsigned int max_mv_def = VPXMIN(cm->width, cm->height);
3260 // Default based on max resolution.
3261 cpi->mv_step_param = vp9_init_search_range(max_mv_def);
3263 if (cpi->sf.mv.auto_mv_step_size) {
3264 if (frame_is_intra_only(cm)) {
3265 // Initialize max_mv_magnitude for use in the first INTER frame
3266 // after a key/intra-only frame.
3267 cpi->max_mv_magnitude = max_mv_def;
3269 if (cm->show_frame) {
3270 // Allow mv_steps to correspond to twice the max mv magnitude found
3271 // in the previous frame, capped by the default max_mv_magnitude based
3273 cpi->mv_step_param = vp9_init_search_range(
3274 VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
3276 cpi->max_mv_magnitude = 0;
3281 static void set_size_independent_vars(VP9_COMP *cpi) {
3282 vp9_set_speed_features_framesize_independent(cpi);
3283 vp9_set_rd_speed_thresholds(cpi);
3284 vp9_set_rd_speed_thresholds_sub8x8(cpi);
3285 cpi->common.interp_filter = cpi->sf.default_interp_filter;
3288 static void set_size_dependent_vars(VP9_COMP *cpi, int *q,
3289 int *bottom_index, int *top_index) {
3290 VP9_COMMON *const cm = &cpi->common;
3291 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3293 // Setup variables that depend on the dimensions of the frame.
3294 vp9_set_speed_features_framesize_dependent(cpi);
3296 // Decide q and q bounds.
3297 *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
3299 if (!frame_is_intra_only(cm)) {
3300 vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
3303 // Configure experimental use of segmentation for enhanced coding of
3304 // static regions if indicated.
3305 // Only allowed in the second pass of a two pass encode, as it requires
3306 // lagged coding, and if the relevant speed feature flag is set.
3307 if (oxcf->pass == 2 && cpi->sf.static_segmentation)
3308 configure_static_seg_features(cpi);
3310 #if CONFIG_VP9_POSTPROC && !(CONFIG_VP9_TEMPORAL_DENOISING)
3311 if (oxcf->noise_sensitivity > 0) {
3313 switch (oxcf->noise_sensitivity) {
3331 vp9_denoise(cpi->Source, cpi->Source, l);
3333 #endif // CONFIG_VP9_POSTPROC
3336 #if CONFIG_VP9_TEMPORAL_DENOISING
3337 static void setup_denoiser_buffer(VP9_COMP *cpi) {
3338 VP9_COMMON *const cm = &cpi->common;
3339 if (cpi->oxcf.noise_sensitivity > 0 &&
3340 !cpi->denoiser.frame_buffer_initialized) {
3341 if (vp9_denoiser_alloc(&cpi->denoiser, cm->width, cm->height,
3342 cm->subsampling_x, cm->subsampling_y,
3343 #if CONFIG_VP9_HIGHBITDEPTH
3344 cm->use_highbitdepth,
3346 VP9_ENC_BORDER_IN_PIXELS))
3347 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3348 "Failed to allocate denoiser");
3353 static void init_motion_estimation(VP9_COMP *cpi) {
3354 int y_stride = cpi->scaled_source.y_stride;
3356 if (cpi->sf.mv.search_method == NSTEP) {
3357 vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
3358 } else if (cpi->sf.mv.search_method == DIAMOND) {
3359 vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
3363 static void set_frame_size(VP9_COMP *cpi) {
3365 VP9_COMMON *const cm = &cpi->common;
3366 VP9EncoderConfig *const oxcf = &cpi->oxcf;
3367 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
3369 if (oxcf->pass == 2 &&
3370 oxcf->rc_mode == VPX_VBR &&
3371 ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
3372 (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
3373 calculate_coded_size(
3374 cpi, &oxcf->scaled_frame_width, &oxcf->scaled_frame_height);
3376 // There has been a change in frame size.
3377 vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
3378 oxcf->scaled_frame_height);
3381 if (oxcf->pass == 0 &&
3382 oxcf->rc_mode == VPX_CBR &&
3384 oxcf->resize_mode == RESIZE_DYNAMIC &&
3385 cpi->resize_pending != 0) {
3386 oxcf->scaled_frame_width =
3387 (oxcf->width * cpi->resize_scale_num) / cpi->resize_scale_den;
3388 oxcf->scaled_frame_height =
3389 (oxcf->height * cpi->resize_scale_num) /cpi->resize_scale_den;
3390 // There has been a change in frame size.
3391 vp9_set_size_literal(cpi,
3392 oxcf->scaled_frame_width,
3393 oxcf->scaled_frame_height);
3395 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3396 set_mv_search_params(cpi);
3398 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
3399 #if CONFIG_VP9_TEMPORAL_DENOISING
3400 // Reset the denoiser on the resized frame.
3401 if (cpi->oxcf.noise_sensitivity > 0) {
3402 vp9_denoiser_free(&(cpi->denoiser));
3403 setup_denoiser_buffer(cpi);
3404 // Dynamic resize is only triggered for non-SVC, so we can force
3405 // golden frame update here as temporary fix to denoiser.
3406 cpi->refresh_golden_frame = 1;
3411 if ((oxcf->pass == 2) &&
3413 (is_two_pass_svc(cpi) &&
3414 cpi->svc.encode_empty_frame_state != ENCODING))) {
3415 vp9_set_target_rate(cpi);
3418 alloc_frame_mvs(cm, cm->new_fb_idx);
3420 // Reset the frame pointers to the current frame size.
3421 if (vpx_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
3422 cm->subsampling_x, cm->subsampling_y,
3423 #if CONFIG_VP9_HIGHBITDEPTH
3424 cm->use_highbitdepth,
3426 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
3428 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3429 "Failed to allocate frame buffer");
3431 alloc_util_frame_buffers(cpi);
3432 init_motion_estimation(cpi);
3434 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3435 RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
3436 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3438 ref_buf->idx = buf_idx;
3440 if (buf_idx != INVALID_IDX) {
3441 YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
3443 #if CONFIG_VP9_HIGHBITDEPTH
3444 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3445 buf->y_crop_width, buf->y_crop_height,
3446 cm->width, cm->height,
3447 (buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
3450 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3451 buf->y_crop_width, buf->y_crop_height,
3452 cm->width, cm->height);
3453 #endif // CONFIG_VP9_HIGHBITDEPTH
3454 if (vp9_is_scaled(&ref_buf->sf))
3455 vpx_extend_frame_borders(buf);
3457 ref_buf->buf = NULL;
3461 set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
3464 static void encode_without_recode_loop(VP9_COMP *cpi,
3467 VP9_COMMON *const cm = &cpi->common;
3468 int q = 0, bottom_index = 0, top_index = 0; // Dummy variables.
3470 vpx_clear_system_state();
3472 set_frame_size(cpi);
3474 if (is_one_pass_cbr_svc(cpi) &&
3475 cpi->un_scaled_source->y_width == cm->width << 2 &&
3476 cpi->un_scaled_source->y_height == cm->height << 2 &&
3477 cpi->svc.scaled_temp.y_width == cm->width << 1 &&
3478 cpi->svc.scaled_temp.y_height == cm->height << 1) {
3479 cpi->Source = vp9_svc_twostage_scale(cm,
3480 cpi->un_scaled_source,
3481 &cpi->scaled_source,
3482 &cpi->svc.scaled_temp);
3484 cpi->Source = vp9_scale_if_required(cm,
3485 cpi->un_scaled_source,
3486 &cpi->scaled_source,
3487 (cpi->oxcf.pass == 0));
3489 // Avoid scaling last_source unless its needed.
3490 // Last source is needed if vp9_avg_source_sad() is used, or if
3491 // partition_search_type == SOURCE_VAR_BASED_PARTITION, or if noise
3492 // estimation is enabled.
3493 if (cpi->unscaled_last_source != NULL &&
3494 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3495 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_VBR &&
3496 cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5) ||
3497 cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION ||
3498 cpi->noise_estimate.enabled))
3499 cpi->Last_Source = vp9_scale_if_required(cm,
3500 cpi->unscaled_last_source,
3501 &cpi->scaled_last_source,
3502 (cpi->oxcf.pass == 0));
3504 if (cm->frame_type == KEY_FRAME || cpi->resize_pending != 0) {
3505 memset(cpi->consec_zero_mv, 0,
3506 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
3509 vp9_update_noise_estimate(cpi);
3511 if (cpi->oxcf.pass == 0 &&
3512 cpi->oxcf.mode == REALTIME &&
3513 cpi->oxcf.speed >= 5 &&
3514 cpi->resize_state == 0 &&
3515 cm->frame_type != KEY_FRAME &&
3516 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3517 cpi->oxcf.rc_mode == VPX_VBR))
3518 vp9_avg_source_sad(cpi);
3520 // For 1 pass SVC, since only ZEROMV is allowed for upsampled reference
3521 // frame (i.e, svc->force_zero_mode_spatial_ref = 0), we can avoid this
3522 // frame-level upsampling.
3523 if (frame_is_intra_only(cm) == 0 && !is_one_pass_cbr_svc(cpi)) {
3524 vp9_scale_references(cpi);
3527 set_size_independent_vars(cpi);
3528 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3530 if (cpi->oxcf.speed >= 5 &&
3531 cpi->oxcf.pass == 0 &&
3532 cpi->oxcf.rc_mode == VPX_CBR &&
3533 cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
3534 cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3535 cpi->use_skin_detection = 1;
3538 vp9_set_quantizer(cm, q);
3539 vp9_set_variance_partition_thresholds(cpi, q);
3543 suppress_active_map(cpi);
3544 // Variance adaptive and in frame q adjustment experiments are mutually
3546 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3547 vp9_vaq_frame_setup(cpi);
3548 } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3549 vp9_360aq_frame_setup(cpi);
3550 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3551 vp9_setup_in_frame_q_adj(cpi);
3552 } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3553 vp9_cyclic_refresh_setup(cpi);
3555 apply_active_map(cpi);
3557 // transform / motion compensation build reconstruction frame
3558 vp9_encode_frame(cpi);
3560 // Check if we should drop this frame because of high overshoot.
3561 // Only for frames where high temporal-source sad is detected.
3562 if (cpi->oxcf.pass == 0 &&
3563 cpi->oxcf.rc_mode == VPX_CBR &&
3564 cpi->resize_state == 0 &&
3565 cm->frame_type != KEY_FRAME &&
3566 cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
3567 cpi->rc.high_source_sad == 1) {
3569 // Get an estimate of the encoded frame size.
3570 save_coding_context(cpi);
3571 vp9_pack_bitstream(cpi, dest, size);
3572 restore_coding_context(cpi);
3573 frame_size = (int)(*size) << 3;
3574 // Check if encoded frame will overshoot too much, and if so, set the q and
3575 // adjust some rate control parameters, and return to re-encode the frame.
3576 if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) {
3577 vpx_clear_system_state();
3578 vp9_set_quantizer(cm, q);
3579 vp9_set_variance_partition_thresholds(cpi, q);
3580 suppress_active_map(cpi);
3581 // Turn-off cyclic refresh for re-encoded frame.
3582 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3583 unsigned char *const seg_map = cpi->segmentation_map;
3584 memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
3585 vp9_disable_segmentation(&cm->seg);
3587 apply_active_map(cpi);
3588 vp9_encode_frame(cpi);
3592 // Update some stats from cyclic refresh, and check if we should not update
3593 // golden reference, for non-SVC 1 pass CBR.
3594 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
3595 cm->frame_type != KEY_FRAME &&
3597 cpi->ext_refresh_frame_flags_pending == 0 &&
3598 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR))
3599 vp9_cyclic_refresh_check_golden_update(cpi);
3601 // Update the skip mb flag probabilities based on the distribution
3602 // seen in the last encoder iteration.
3603 // update_base_skip_probs(cpi);
3604 vpx_clear_system_state();
3607 static void encode_with_recode_loop(VP9_COMP *cpi,
3610 VP9_COMMON *const cm = &cpi->common;
3611 RATE_CONTROL *const rc = &cpi->rc;
3612 int bottom_index, top_index;
3614 int loop_at_this_size = 0;
3616 int overshoot_seen = 0;
3617 int undershoot_seen = 0;
3618 int frame_over_shoot_limit;
3619 int frame_under_shoot_limit;
3620 int q = 0, q_low = 0, q_high = 0;
3622 set_size_independent_vars(cpi);
3625 vpx_clear_system_state();
3627 set_frame_size(cpi);
3629 if (loop_count == 0 || cpi->resize_pending != 0) {
3630 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3632 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3633 set_mv_search_params(cpi);
3635 // Reset the loop state for new frame size.
3637 undershoot_seen = 0;
3639 // Reconfiguration for change in frame size has concluded.
3640 cpi->resize_pending = 0;
3642 q_low = bottom_index;
3645 loop_at_this_size = 0;
3648 // Decide frame size bounds first time through.
3649 if (loop_count == 0) {
3650 vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
3651 &frame_under_shoot_limit,
3652 &frame_over_shoot_limit);
3655 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
3656 &cpi->scaled_source,
3657 (cpi->oxcf.pass == 0));
3659 if (cpi->unscaled_last_source != NULL)
3660 cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
3661 &cpi->scaled_last_source,
3662 (cpi->oxcf.pass == 0));
3664 if (frame_is_intra_only(cm) == 0) {
3665 if (loop_count > 0) {
3666 release_scaled_references(cpi);
3668 vp9_scale_references(cpi);
3671 vp9_set_quantizer(cm, q);
3673 if (loop_count == 0)
3676 // Variance adaptive and in frame q adjustment experiments are mutually
3678 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3679 vp9_vaq_frame_setup(cpi);
3680 } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3681 vp9_360aq_frame_setup(cpi);
3682 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3683 vp9_setup_in_frame_q_adj(cpi);
3686 // transform / motion compensation build reconstruction frame
3687 vp9_encode_frame(cpi);
3689 // Update the skip mb flag probabilities based on the distribution
3690 // seen in the last encoder iteration.
3691 // update_base_skip_probs(cpi);
3693 vpx_clear_system_state();
3695 // Dummy pack of the bitstream using up to date stats to get an
3696 // accurate estimate of output frame size to determine if we need
3698 if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
3699 save_coding_context(cpi);
3700 if (!cpi->sf.use_nonrd_pick_mode)
3701 vp9_pack_bitstream(cpi, dest, size);
3703 rc->projected_frame_size = (int)(*size) << 3;
3704 restore_coding_context(cpi);
3706 if (frame_over_shoot_limit == 0)
3707 frame_over_shoot_limit = 1;
3710 if (cpi->oxcf.rc_mode == VPX_Q) {
3713 if ((cm->frame_type == KEY_FRAME) &&
3714 rc->this_key_frame_forced &&
3715 (rc->projected_frame_size < rc->max_frame_bandwidth)) {
3719 int64_t high_err_target = cpi->ambient_err;
3720 int64_t low_err_target = cpi->ambient_err >> 1;
3722 #if CONFIG_VP9_HIGHBITDEPTH
3723 if (cm->use_highbitdepth) {
3724 kf_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3726 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3729 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3730 #endif // CONFIG_VP9_HIGHBITDEPTH
3732 // Prevent possible divide by zero error below for perfect KF
3735 // The key frame is not good enough or we can afford
3736 // to make it better without undue risk of popping.
3737 if ((kf_err > high_err_target &&
3738 rc->projected_frame_size <= frame_over_shoot_limit) ||
3739 (kf_err > low_err_target &&
3740 rc->projected_frame_size <= frame_under_shoot_limit)) {
3742 q_high = q > q_low ? q - 1 : q_low;
3745 q = (int)((q * high_err_target) / kf_err);
3746 q = VPXMIN(q, (q_high + q_low) >> 1);
3747 } else if (kf_err < low_err_target &&
3748 rc->projected_frame_size >= frame_under_shoot_limit) {
3749 // The key frame is much better than the previous frame
3751 q_low = q < q_high ? q + 1 : q_high;
3754 q = (int)((q * low_err_target) / kf_err);
3755 q = VPXMIN(q, (q_high + q_low + 1) >> 1);
3758 // Clamp Q to upper and lower limits:
3759 q = clamp(q, q_low, q_high);
3762 } else if (recode_loop_test(
3763 cpi, frame_over_shoot_limit, frame_under_shoot_limit,
3764 q, VPXMAX(q_high, top_index), bottom_index)) {
3765 // Is the projected frame size out of range and are we allowed
3766 // to attempt to recode.
3770 if (cpi->resize_pending == 1) {
3771 // Change in frame size so go back around the recode loop.
3772 cpi->rc.frame_size_selector =
3773 SCALE_STEP1 - cpi->rc.frame_size_selector;
3774 cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
3776 #if CONFIG_INTERNAL_STATS
3777 ++cpi->tot_recode_hits;
3784 // Frame size out of permitted range:
3785 // Update correction factor & compute new Q to try...
3787 // Frame is too large
3788 if (rc->projected_frame_size > rc->this_frame_target) {
3789 // Special case if the projected size is > the max allowed.
3790 if (rc->projected_frame_size >= rc->max_frame_bandwidth)
3791 q_high = rc->worst_quality;
3793 // Raise Qlow as to at least the current value
3794 q_low = q < q_high ? q + 1 : q_high;
3796 if (undershoot_seen || loop_at_this_size > 1) {
3797 // Update rate_correction_factor unless
3798 vp9_rc_update_rate_correction_factors(cpi);
3800 q = (q_high + q_low + 1) / 2;
3802 // Update rate_correction_factor unless
3803 vp9_rc_update_rate_correction_factors(cpi);
3805 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3806 bottom_index, VPXMAX(q_high, top_index));
3808 while (q < q_low && retries < 10) {
3809 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));
3818 // Frame is too small
3819 q_high = q > q_low ? q - 1 : q_low;
3821 if (overshoot_seen || loop_at_this_size > 1) {
3822 vp9_rc_update_rate_correction_factors(cpi);
3823 q = (q_high + q_low) / 2;
3825 vp9_rc_update_rate_correction_factors(cpi);
3826 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3827 bottom_index, top_index);
3828 // Special case reset for qlow for constrained quality.
3829 // This should only trigger where there is very substantial
3830 // undershoot on a frame and the auto cq level is above
3831 // the user passsed in value.
3832 if (cpi->oxcf.rc_mode == VPX_CQ &&
3837 while (q > q_high && retries < 10) {
3838 vp9_rc_update_rate_correction_factors(cpi);
3839 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3840 bottom_index, top_index);
3845 undershoot_seen = 1;
3848 // Clamp Q to upper and lower limits:
3849 q = clamp(q, q_low, q_high);
3851 loop = (q != last_q);
3857 // Special case for overlay frame.
3858 if (rc->is_src_frame_alt_ref &&
3859 rc->projected_frame_size < rc->max_frame_bandwidth)
3864 ++loop_at_this_size;
3866 #if CONFIG_INTERNAL_STATS
3867 ++cpi->tot_recode_hits;
3873 static int get_ref_frame_flags(const VP9_COMP *cpi) {
3874 const int *const map = cpi->common.ref_frame_map;
3875 const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
3876 const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
3877 const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
3878 int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
3881 flags &= ~VP9_GOLD_FLAG;
3883 if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
3884 (cpi->svc.number_temporal_layers == 1 &&
3885 cpi->svc.number_spatial_layers == 1))
3886 flags &= ~VP9_GOLD_FLAG;
3889 flags &= ~VP9_ALT_FLAG;
3892 flags &= ~VP9_ALT_FLAG;
3897 static void set_ext_overrides(VP9_COMP *cpi) {
3898 // Overrides the defaults with the externally supplied values with
3899 // vp9_update_reference() and vp9_update_entropy() calls
3900 // Note: The overrides are valid only for the next frame passed
3901 // to encode_frame_to_data_rate() function
3902 if (cpi->ext_refresh_frame_context_pending) {
3903 cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
3904 cpi->ext_refresh_frame_context_pending = 0;
3906 if (cpi->ext_refresh_frame_flags_pending) {
3907 cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
3908 cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
3909 cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
3913 YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
3914 YV12_BUFFER_CONFIG *unscaled,
3915 YV12_BUFFER_CONFIG *scaled,
3916 YV12_BUFFER_CONFIG *scaled_temp) {
3917 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3918 cm->mi_rows * MI_SIZE != unscaled->y_height) {
3919 #if CONFIG_VP9_HIGHBITDEPTH
3920 scale_and_extend_frame(unscaled, scaled_temp, (int)cm->bit_depth);
3921 scale_and_extend_frame(scaled_temp, scaled, (int)cm->bit_depth);
3923 vp9_scale_and_extend_frame(unscaled, scaled_temp);
3924 vp9_scale_and_extend_frame(scaled_temp, scaled);
3925 #endif // CONFIG_VP9_HIGHBITDEPTH
3932 YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
3933 YV12_BUFFER_CONFIG *unscaled,
3934 YV12_BUFFER_CONFIG *scaled,
3935 int use_normative_scaler) {
3936 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3937 cm->mi_rows * MI_SIZE != unscaled->y_height) {
3938 #if CONFIG_VP9_HIGHBITDEPTH
3939 if (use_normative_scaler &&
3940 unscaled->y_width <= (scaled->y_width << 1) &&
3941 unscaled->y_height <= (scaled->y_height << 1))
3942 scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth);
3944 scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
3946 if (use_normative_scaler &&
3947 unscaled->y_width <= (scaled->y_width << 1) &&
3948 unscaled->y_height <= (scaled->y_height << 1))
3949 vp9_scale_and_extend_frame(unscaled, scaled);
3951 scale_and_extend_frame_nonnormative(unscaled, scaled);
3952 #endif // CONFIG_VP9_HIGHBITDEPTH
3959 static void set_arf_sign_bias(VP9_COMP *cpi) {
3960 VP9_COMMON *const cm = &cpi->common;
3963 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
3964 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
3965 arf_sign_bias = cpi->rc.source_alt_ref_active &&
3966 (!cpi->refresh_alt_ref_frame ||
3967 (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
3970 (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
3972 cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
3975 static int setup_interp_filter_search_mask(VP9_COMP *cpi) {
3976 INTERP_FILTER ifilter;
3977 int ref_total[MAX_REF_FRAMES] = {0};
3978 MV_REFERENCE_FRAME ref;
3980 if (cpi->common.last_frame_type == KEY_FRAME ||
3981 cpi->refresh_alt_ref_frame)
3983 for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
3984 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter)
3985 ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
3987 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) {
3988 if ((ref_total[LAST_FRAME] &&
3989 cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
3990 (ref_total[GOLDEN_FRAME] == 0 ||
3991 cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50
3992 < ref_total[GOLDEN_FRAME]) &&
3993 (ref_total[ALTREF_FRAME] == 0 ||
3994 cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50
3995 < ref_total[ALTREF_FRAME]))
3996 mask |= 1 << ifilter;
4001 static void encode_frame_to_data_rate(VP9_COMP *cpi,
4004 unsigned int *frame_flags) {
4005 VP9_COMMON *const cm = &cpi->common;
4006 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4007 struct segmentation *const seg = &cm->seg;
4010 set_ext_overrides(cpi);
4011 vpx_clear_system_state();
4013 // Set the arf sign bias for this frame.
4014 set_arf_sign_bias(cpi);
4016 // Set default state for segment based loop filter update flags.
4017 cm->lf.mode_ref_delta_update = 0;
4019 if (cpi->oxcf.pass == 2 &&
4020 cpi->sf.adaptive_interp_filter_search)
4021 cpi->sf.interp_filter_search_mask =
4022 setup_interp_filter_search_mask(cpi);
4024 // Set various flags etc to special state if it is a key frame.
4025 if (frame_is_intra_only(cm)) {
4026 // Reset the loop filter deltas and segmentation map.
4027 vp9_reset_segment_features(&cm->seg);
4029 // If segmentation is enabled force a map update for key frames.
4031 seg->update_map = 1;
4032 seg->update_data = 1;
4035 // The alternate reference frame cannot be active for a key frame.
4036 cpi->rc.source_alt_ref_active = 0;
4038 cm->error_resilient_mode = oxcf->error_resilient_mode;
4039 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
4041 // By default, encoder assumes decoder can use prev_mi.
4042 if (cm->error_resilient_mode) {
4043 cm->frame_parallel_decoding_mode = 1;
4044 cm->reset_frame_context = 0;
4045 cm->refresh_frame_context = 0;
4046 } else if (cm->intra_only) {
4047 // Only reset the current context.
4048 cm->reset_frame_context = 2;
4051 if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) {
4052 // Use context 0 for intra only empty frame, but the last frame context
4053 // for other empty frames.
4054 if (cpi->svc.encode_empty_frame_state == ENCODING) {
4055 if (cpi->svc.encode_intra_empty_frame != 0)
4056 cm->frame_context_idx = 0;
4058 cm->frame_context_idx = FRAME_CONTEXTS - 1;
4060 cm->frame_context_idx =
4061 cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
4062 cpi->svc.temporal_layer_id;
4065 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
4067 // The probs will be updated based on the frame type of its previous
4068 // frame if frame_parallel_decoding_mode is 0. The type may vary for
4069 // the frame after a key frame in base layer since we may drop enhancement
4070 // layers. So set frame_parallel_decoding_mode to 1 in this case.
4071 if (cm->frame_parallel_decoding_mode == 0) {
4072 if (cpi->svc.number_temporal_layers == 1) {
4073 if (cpi->svc.spatial_layer_id == 0 &&
4074 cpi->svc.layer_context[0].last_frame_type == KEY_FRAME)
4075 cm->frame_parallel_decoding_mode = 1;
4076 } else if (cpi->svc.spatial_layer_id == 0) {
4077 // Find the 2nd frame in temporal base layer and 1st frame in temporal
4078 // enhancement layers from the key frame.
4080 for (i = 0; i < cpi->svc.number_temporal_layers; ++i) {
4081 if (cpi->svc.layer_context[0].frames_from_key_frame == 1 << i) {
4082 cm->frame_parallel_decoding_mode = 1;
4090 // For 1 pass CBR, check if we are dropping this frame.
4091 // For spatial layers, for now only check for frame-dropping on first spatial
4092 // layer, and if decision is to drop, we drop whole super-frame.
4093 if (oxcf->pass == 0 &&
4094 oxcf->rc_mode == VPX_CBR &&
4095 cm->frame_type != KEY_FRAME) {
4096 if (vp9_rc_drop_frame(cpi) ||
4097 (is_one_pass_cbr_svc(cpi) && cpi->svc.rc_drop_superframe == 1)) {
4098 vp9_rc_postencode_update_drop_frame(cpi);
4099 ++cm->current_video_frame;
4100 cpi->ext_refresh_frame_flags_pending = 0;
4101 cpi->svc.rc_drop_superframe = 1;
4102 // TODO(marpan): Advancing the svc counters on dropped frames can break
4103 // the referencing scheme for the fixed svc patterns defined in
4104 // vp9_one_pass_cbr_svc_start_layer(). Look into fixing this issue, but
4105 // for now, don't advance the svc frame counters on dropped frame.
4106 // if (cpi->use_svc)
4107 // vp9_inc_frame_in_layer(cpi);
4112 vpx_clear_system_state();
4114 #if CONFIG_INTERNAL_STATS
4115 memset(cpi->mode_chosen_counts, 0,
4116 MAX_MODES * sizeof(*cpi->mode_chosen_counts));
4119 if (cpi->sf.recode_loop == DISALLOW_RECODE) {
4120 encode_without_recode_loop(cpi, size, dest);
4122 encode_with_recode_loop(cpi, size, dest);
4125 #if CONFIG_VP9_TEMPORAL_DENOISING
4126 #ifdef OUTPUT_YUV_DENOISED
4127 if (oxcf->noise_sensitivity > 0) {
4128 vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME],
4133 #ifdef OUTPUT_YUV_SKINMAP
4134 if (cpi->common.current_video_frame > 1) {
4135 vp9_compute_skin_map(cpi, yuv_skinmap_file);
4139 // Special case code to reduce pulsing when key frames are forced at a
4140 // fixed interval. Note the reconstruction error if it is the frame before
4141 // the force key frame
4142 if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
4143 #if CONFIG_VP9_HIGHBITDEPTH
4144 if (cm->use_highbitdepth) {
4145 cpi->ambient_err = vp9_highbd_get_y_sse(cpi->Source,
4146 get_frame_new_buffer(cm));
4148 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4151 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4152 #endif // CONFIG_VP9_HIGHBITDEPTH
4155 // If the encoder forced a KEY_FRAME decision
4156 if (cm->frame_type == KEY_FRAME)
4157 cpi->refresh_last_frame = 1;
4159 cm->frame_to_show = get_frame_new_buffer(cm);
4160 cm->frame_to_show->color_space = cm->color_space;
4161 cm->frame_to_show->color_range = cm->color_range;
4162 cm->frame_to_show->render_width = cm->render_width;
4163 cm->frame_to_show->render_height = cm->render_height;
4165 // Pick the loop filter level for the frame.
4166 loopfilter_frame(cpi, cm);
4168 // build the bitstream
4169 vp9_pack_bitstream(cpi, dest, size);
4171 if (cm->seg.update_map)
4172 update_reference_segmentation_map(cpi);
4174 if (frame_is_intra_only(cm) == 0) {
4175 release_scaled_references(cpi);
4177 vp9_update_reference_frames(cpi);
4179 for (t = TX_4X4; t <= TX_32X32; t++)
4180 full_to_model_counts(cpi->td.counts->coef[t],
4181 cpi->td.rd_counts.coef_counts[t]);
4183 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode)
4184 vp9_adapt_coef_probs(cm);
4186 if (!frame_is_intra_only(cm)) {
4187 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
4188 vp9_adapt_mode_probs(cm);
4189 vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
4193 cpi->ext_refresh_frame_flags_pending = 0;
4195 if (cpi->refresh_golden_frame == 1)
4196 cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
4198 cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
4200 if (cpi->refresh_alt_ref_frame == 1)
4201 cpi->frame_flags |= FRAMEFLAGS_ALTREF;
4203 cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
4205 cpi->ref_frame_flags = get_ref_frame_flags(cpi);
4207 cm->last_frame_type = cm->frame_type;
4209 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
4210 vp9_rc_postencode_update(cpi, *size);
4213 output_frame_level_debug_stats(cpi);
4216 if (cm->frame_type == KEY_FRAME) {
4217 // Tell the caller that the frame was coded as a key frame
4218 *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
4220 *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
4223 // Clear the one shot update flags for segmentation map and mode/ref loop
4225 cm->seg.update_map = 0;
4226 cm->seg.update_data = 0;
4227 cm->lf.mode_ref_delta_update = 0;
4229 // keep track of the last coded dimensions
4230 cm->last_width = cm->width;
4231 cm->last_height = cm->height;
4233 // reset to normal state now that we are done.
4234 if (!cm->show_existing_frame)
4235 cm->last_show_frame = cm->show_frame;
4237 if (cm->show_frame) {
4238 vp9_swap_mi_and_prev_mi(cm);
4239 // Don't increment frame counters if this was an altref buffer
4240 // update not a real frame
4241 ++cm->current_video_frame;
4243 vp9_inc_frame_in_layer(cpi);
4245 cm->prev_frame = cm->cur_frame;
4248 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
4249 cpi->svc.number_temporal_layers +
4250 cpi->svc.temporal_layer_id].last_frame_type =
4254 static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
4255 unsigned int *frame_flags) {
4256 vp9_rc_get_svc_params(cpi);
4257 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4260 static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
4261 unsigned int *frame_flags) {
4262 if (cpi->oxcf.rc_mode == VPX_CBR) {
4263 vp9_rc_get_one_pass_cbr_params(cpi);
4265 vp9_rc_get_one_pass_vbr_params(cpi);
4267 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4270 static void Pass2Encode(VP9_COMP *cpi, size_t *size,
4271 uint8_t *dest, unsigned int *frame_flags) {
4272 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
4273 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4275 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
4276 vp9_twopass_postencode_update(cpi);
4279 static void init_ref_frame_bufs(VP9_COMMON *cm) {
4281 BufferPool *const pool = cm->buffer_pool;
4282 cm->new_fb_idx = INVALID_IDX;
4283 for (i = 0; i < REF_FRAMES; ++i) {
4284 cm->ref_frame_map[i] = INVALID_IDX;
4285 pool->frame_bufs[i].ref_count = 0;
4289 static void check_initial_width(VP9_COMP *cpi,
4290 #if CONFIG_VP9_HIGHBITDEPTH
4291 int use_highbitdepth,
4293 int subsampling_x, int subsampling_y) {
4294 VP9_COMMON *const cm = &cpi->common;
4296 if (!cpi->initial_width ||
4297 #if CONFIG_VP9_HIGHBITDEPTH
4298 cm->use_highbitdepth != use_highbitdepth ||
4300 cm->subsampling_x != subsampling_x ||
4301 cm->subsampling_y != subsampling_y) {
4302 cm->subsampling_x = subsampling_x;
4303 cm->subsampling_y = subsampling_y;
4304 #if CONFIG_VP9_HIGHBITDEPTH
4305 cm->use_highbitdepth = use_highbitdepth;
4308 alloc_raw_frame_buffers(cpi);
4309 init_ref_frame_bufs(cm);
4310 alloc_util_frame_buffers(cpi);
4312 init_motion_estimation(cpi); // TODO(agrange) This can be removed.
4314 cpi->initial_width = cm->width;
4315 cpi->initial_height = cm->height;
4316 cpi->initial_mbs = cm->MBs;
4320 int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
4321 YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
4323 VP9_COMMON *const cm = &cpi->common;
4324 struct vpx_usec_timer timer;
4326 const int subsampling_x = sd->subsampling_x;
4327 const int subsampling_y = sd->subsampling_y;
4328 #if CONFIG_VP9_HIGHBITDEPTH
4329 const int use_highbitdepth = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
4332 #if CONFIG_VP9_HIGHBITDEPTH
4333 check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
4335 check_initial_width(cpi, subsampling_x, subsampling_y);
4336 #endif // CONFIG_VP9_HIGHBITDEPTH
4338 #if CONFIG_VP9_TEMPORAL_DENOISING
4339 setup_denoiser_buffer(cpi);
4341 vpx_usec_timer_start(&timer);
4343 if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
4344 #if CONFIG_VP9_HIGHBITDEPTH
4346 #endif // CONFIG_VP9_HIGHBITDEPTH
4349 vpx_usec_timer_mark(&timer);
4350 cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
4352 if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
4353 (subsampling_x != 1 || subsampling_y != 1)) {
4354 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4355 "Non-4:2:0 color format requires profile 1 or 3");
4358 if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
4359 (subsampling_x == 1 && subsampling_y == 1)) {
4360 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4361 "4:2:0 color format requires profile 0 or 2");
4369 static int frame_is_reference(const VP9_COMP *cpi) {
4370 const VP9_COMMON *cm = &cpi->common;
4372 return cm->frame_type == KEY_FRAME ||
4373 cpi->refresh_last_frame ||
4374 cpi->refresh_golden_frame ||
4375 cpi->refresh_alt_ref_frame ||
4376 cm->refresh_frame_context ||
4377 cm->lf.mode_ref_delta_update ||
4378 cm->seg.update_map ||
4379 cm->seg.update_data;
4382 static void adjust_frame_rate(VP9_COMP *cpi,
4383 const struct lookahead_entry *source) {
4384 int64_t this_duration;
4387 if (source->ts_start == cpi->first_time_stamp_ever) {
4388 this_duration = source->ts_end - source->ts_start;
4391 int64_t last_duration = cpi->last_end_time_stamp_seen
4392 - cpi->last_time_stamp_seen;
4394 this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
4396 // do a step update if the duration changes by 10%
4398 step = (int)((this_duration - last_duration) * 10 / last_duration);
4401 if (this_duration) {
4403 vp9_new_framerate(cpi, 10000000.0 / this_duration);
4405 // Average this frame's rate into the last second's average
4406 // frame rate. If we haven't seen 1 second yet, then average
4407 // over the whole interval seen.
4408 const double interval = VPXMIN(
4409 (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
4410 double avg_duration = 10000000.0 / cpi->framerate;
4411 avg_duration *= (interval - avg_duration + this_duration);
4412 avg_duration /= interval;
4414 vp9_new_framerate(cpi, 10000000.0 / avg_duration);
4417 cpi->last_time_stamp_seen = source->ts_start;
4418 cpi->last_end_time_stamp_seen = source->ts_end;
4421 // Returns 0 if this is not an alt ref else the offset of the source frame
4422 // used as the arf midpoint.
4423 static int get_arf_src_index(VP9_COMP *cpi) {
4424 RATE_CONTROL *const rc = &cpi->rc;
4425 int arf_src_index = 0;
4426 if (is_altref_enabled(cpi)) {
4427 if (cpi->oxcf.pass == 2) {
4428 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4429 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
4430 arf_src_index = gf_group->arf_src_offset[gf_group->index];
4432 } else if (rc->source_alt_ref_pending) {
4433 arf_src_index = rc->frames_till_gf_update_due;
4436 return arf_src_index;
4439 static void check_src_altref(VP9_COMP *cpi,
4440 const struct lookahead_entry *source) {
4441 RATE_CONTROL *const rc = &cpi->rc;
4443 if (cpi->oxcf.pass == 2) {
4444 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4445 rc->is_src_frame_alt_ref =
4446 (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
4448 rc->is_src_frame_alt_ref = cpi->alt_ref_source &&
4449 (source == cpi->alt_ref_source);
4452 if (rc->is_src_frame_alt_ref) {
4453 // Current frame is an ARF overlay frame.
4454 cpi->alt_ref_source = NULL;
4456 // Don't refresh the last buffer for an ARF overlay frame. It will
4457 // become the GF so preserve last as an alternative prediction option.
4458 cpi->refresh_last_frame = 0;
4462 #if CONFIG_INTERNAL_STATS
4463 extern double vp9_get_blockiness(const uint8_t *img1, int img1_pitch,
4464 const uint8_t *img2, int img2_pitch,
4465 int width, int height);
4467 static void adjust_image_stat(double y, double u, double v, double all,
4472 s->stat[ALL] += all;
4473 s->worst = VPXMIN(s->worst, all);
4475 #endif // CONFIG_INTERNAL_STATS
4477 static void update_level_info(VP9_COMP *cpi, size_t *size, int arf_src_index) {
4478 VP9_COMMON *const cm = &cpi->common;
4479 Vp9LevelInfo *const level_info = &cpi->level_info;
4480 Vp9LevelSpec *const level_spec = &level_info->level_spec;
4481 Vp9LevelStats *const level_stats = &level_info->level_stats;
4483 uint64_t luma_samples, dur_end;
4484 const uint32_t luma_pic_size = cm->width * cm->height;
4485 double cpb_data_size;
4487 vpx_clear_system_state();
4489 // update level_stats
4490 level_stats->total_compressed_size += *size;
4491 if (cm->show_frame) {
4492 level_stats->total_uncompressed_size +=
4494 2 * (luma_pic_size >> (cm->subsampling_x + cm->subsampling_y));
4495 level_stats->time_encoded =
4496 (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
4497 (double)TICKS_PER_SEC;
4500 if (arf_src_index > 0) {
4501 if (!level_stats->seen_first_altref) {
4502 level_stats->seen_first_altref = 1;
4503 } else if (level_stats->frames_since_last_altref <
4504 level_spec->min_altref_distance) {
4505 level_spec->min_altref_distance = level_stats->frames_since_last_altref;
4507 level_stats->frames_since_last_altref = 0;
4509 ++level_stats->frames_since_last_altref;
4512 if (level_stats->frame_window_buffer.len < FRAME_WINDOW_SIZE - 1) {
4513 idx = (level_stats->frame_window_buffer.start +
4514 level_stats->frame_window_buffer.len++) % FRAME_WINDOW_SIZE;
4516 idx = level_stats->frame_window_buffer.start;
4517 level_stats->frame_window_buffer.start = (idx + 1) % FRAME_WINDOW_SIZE;
4519 level_stats->frame_window_buffer.buf[idx].ts = cpi->last_time_stamp_seen;
4520 level_stats->frame_window_buffer.buf[idx].size = (uint32_t)(*size);
4521 level_stats->frame_window_buffer.buf[idx].luma_samples = luma_pic_size;
4523 if (cm->frame_type == KEY_FRAME) {
4524 level_stats->ref_refresh_map = 0;
4527 level_stats->ref_refresh_map |= vp9_get_refresh_mask(cpi);
4528 // Also need to consider the case where the encoder refers to a buffer
4529 // that has been implicitly refreshed after encoding a keyframe.
4530 if (!cm->intra_only) {
4531 level_stats->ref_refresh_map |= (1 << cpi->lst_fb_idx);
4532 level_stats->ref_refresh_map |= (1 << cpi->gld_fb_idx);
4533 level_stats->ref_refresh_map |= (1 << cpi->alt_fb_idx);
4535 for (i = 0; i < REF_FRAMES; ++i) {
4536 count += (level_stats->ref_refresh_map >> i) & 1;
4538 if (count > level_spec->max_ref_frame_buffers) {
4539 level_spec->max_ref_frame_buffers = count;
4543 // update average_bitrate
4544 level_spec->average_bitrate =
4545 (double)level_stats->total_compressed_size / 125.0 /
4546 level_stats->time_encoded;
4548 // update max_luma_sample_rate
4550 for (i = 0; i < level_stats->frame_window_buffer.len; ++i) {
4551 idx = (level_stats->frame_window_buffer.start +
4552 level_stats->frame_window_buffer.len - 1 - i) % FRAME_WINDOW_SIZE;
4554 dur_end = level_stats->frame_window_buffer.buf[idx].ts;
4556 if (dur_end - level_stats->frame_window_buffer.buf[idx].ts >=
4560 luma_samples += level_stats->frame_window_buffer.buf[idx].luma_samples;
4562 if (luma_samples > level_spec->max_luma_sample_rate) {
4563 level_spec->max_luma_sample_rate = luma_samples;
4566 // update max_cpb_size
4568 for (i = 0; i < CPB_WINDOW_SIZE; ++i) {
4569 if (i >= level_stats->frame_window_buffer.len) break;
4570 idx = (level_stats->frame_window_buffer.start +
4571 level_stats->frame_window_buffer.len - 1 - i) % FRAME_WINDOW_SIZE;
4572 cpb_data_size += level_stats->frame_window_buffer.buf[idx].size;
4574 cpb_data_size = cpb_data_size / 125.0;
4575 if (cpb_data_size > level_spec->max_cpb_size) {
4576 level_spec->max_cpb_size = cpb_data_size;
4579 // update max_luma_picture_size
4580 if (luma_pic_size > level_spec->max_luma_picture_size) {
4581 level_spec->max_luma_picture_size = luma_pic_size;
4584 // update compression_ratio
4585 level_spec->compression_ratio =
4586 (double)level_stats->total_uncompressed_size * cm->bit_depth /
4587 level_stats->total_compressed_size / 8.0;
4589 // update max_col_tiles
4590 if (level_spec->max_col_tiles < (1 << cm->log2_tile_cols)) {
4591 level_spec->max_col_tiles = (1 << cm->log2_tile_cols);
4595 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
4596 size_t *size, uint8_t *dest,
4597 int64_t *time_stamp, int64_t *time_end, int flush) {
4598 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4599 VP9_COMMON *const cm = &cpi->common;
4600 BufferPool *const pool = cm->buffer_pool;
4601 RATE_CONTROL *const rc = &cpi->rc;
4602 struct vpx_usec_timer cmptimer;
4603 YV12_BUFFER_CONFIG *force_src_buffer = NULL;
4604 struct lookahead_entry *last_source = NULL;
4605 struct lookahead_entry *source = NULL;
4609 if (is_two_pass_svc(cpi)) {
4610 #if CONFIG_SPATIAL_SVC
4611 vp9_svc_start_frame(cpi);
4612 // Use a small empty frame instead of a real frame
4613 if (cpi->svc.encode_empty_frame_state == ENCODING)
4614 source = &cpi->svc.empty_frame;
4616 if (oxcf->pass == 2)
4617 vp9_restore_layer_context(cpi);
4618 } else if (is_one_pass_cbr_svc(cpi)) {
4619 vp9_one_pass_cbr_svc_start_layer(cpi);
4622 vpx_usec_timer_start(&cmptimer);
4624 vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
4626 // Is multi-arf enabled.
4627 // Note that at the moment multi_arf is only configured for 2 pass VBR and
4628 // will not work properly with svc.
4629 if ((oxcf->pass == 2) && !cpi->use_svc &&
4630 (cpi->oxcf.enable_auto_arf > 1))
4631 cpi->multi_arf_allowed = 1;
4633 cpi->multi_arf_allowed = 0;
4636 cm->reset_frame_context = 0;
4637 cm->refresh_frame_context = 1;
4638 if (!is_one_pass_cbr_svc(cpi)) {
4639 cpi->refresh_last_frame = 1;
4640 cpi->refresh_golden_frame = 0;
4641 cpi->refresh_alt_ref_frame = 0;
4644 // Should we encode an arf frame.
4645 arf_src_index = get_arf_src_index(cpi);
4647 // Skip alt frame if we encode the empty frame
4648 if (is_two_pass_svc(cpi) && source != NULL)
4651 if (arf_src_index) {
4652 for (i = 0; i <= arf_src_index; ++i) {
4653 struct lookahead_entry *e = vp9_lookahead_peek(cpi->lookahead, i);
4654 // Avoid creating an alt-ref if there's a forced keyframe pending.
4657 } else if (e->flags == VPX_EFLAG_FORCE_KF) {
4665 if (arf_src_index) {
4666 assert(arf_src_index <= rc->frames_to_key);
4668 if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
4669 cpi->alt_ref_source = source;
4671 #if CONFIG_SPATIAL_SVC
4672 if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) {
4674 // Reference a hidden frame from a lower layer
4675 for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) {
4676 if (oxcf->ss_enable_auto_arf[i]) {
4677 cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx;
4682 cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1;
4685 if ((oxcf->arnr_max_frames > 0) && (oxcf->arnr_strength > 0)) {
4686 // Produce the filtered ARF frame.
4687 vp9_temporal_filter(cpi, arf_src_index);
4688 vpx_extend_frame_borders(&cpi->alt_ref_buffer);
4689 force_src_buffer = &cpi->alt_ref_buffer;
4694 cpi->refresh_alt_ref_frame = 1;
4695 cpi->refresh_golden_frame = 0;
4696 cpi->refresh_last_frame = 0;
4697 rc->is_src_frame_alt_ref = 0;
4698 rc->source_alt_ref_pending = 0;
4700 rc->source_alt_ref_pending = 0;
4705 // Get last frame source.
4706 if (cm->current_video_frame > 0) {
4707 if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
4711 // Read in the source frame.
4713 source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
4715 source = vp9_lookahead_pop(cpi->lookahead, flush);
4717 if (source != NULL) {
4720 // if the flags indicate intra frame, but if the current picture is for
4721 // non-zero spatial layer, it should not be an intra picture.
4722 // TODO(Won Kap): this needs to change if per-layer intra frame is
4724 if ((source->flags & VPX_EFLAG_FORCE_KF) &&
4725 cpi->svc.spatial_layer_id > cpi->svc.first_spatial_layer_to_encode) {
4726 source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF);
4729 // Check to see if the frame should be encoded as an arf overlay.
4730 check_src_altref(cpi, source);
4735 cpi->un_scaled_source = cpi->Source = force_src_buffer ? force_src_buffer
4738 cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
4740 *time_stamp = source->ts_start;
4741 *time_end = source->ts_end;
4742 *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
4746 if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
4747 vp9_end_first_pass(cpi); /* get last stats packet */
4748 cpi->twopass.first_pass_done = 1;
4753 if (source->ts_start < cpi->first_time_stamp_ever) {
4754 cpi->first_time_stamp_ever = source->ts_start;
4755 cpi->last_end_time_stamp_seen = source->ts_start;
4758 // Clear down mmx registers
4759 vpx_clear_system_state();
4761 // adjust frame rates based on timestamps given
4762 if (cm->show_frame) {
4763 adjust_frame_rate(cpi, source);
4766 if (is_one_pass_cbr_svc(cpi)) {
4767 vp9_update_temporal_layer_framerate(cpi);
4768 vp9_restore_layer_context(cpi);
4771 // Find a free buffer for the new frame, releasing the reference previously
4773 if (cm->new_fb_idx != INVALID_IDX) {
4774 --pool->frame_bufs[cm->new_fb_idx].ref_count;
4776 cm->new_fb_idx = get_free_fb(cm);
4778 if (cm->new_fb_idx == INVALID_IDX)
4781 cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
4783 if (!cpi->use_svc && cpi->multi_arf_allowed) {
4784 if (cm->frame_type == KEY_FRAME) {
4785 init_buffer_indices(cpi);
4786 } else if (oxcf->pass == 2) {
4787 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4788 cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
4792 // Start with a 0 size frame.
4795 cpi->frame_flags = *frame_flags;
4797 if ((oxcf->pass == 2) &&
4799 (is_two_pass_svc(cpi) &&
4800 cpi->svc.encode_empty_frame_state != ENCODING))) {
4801 vp9_rc_get_second_pass_params(cpi);
4802 } else if (oxcf->pass == 1) {
4803 set_frame_size(cpi);
4806 if (cpi->oxcf.pass != 0 ||
4808 frame_is_intra_only(cm) == 1) {
4809 for (i = 0; i < MAX_REF_FRAMES; ++i)
4810 cpi->scaled_ref_idx[i] = INVALID_IDX;
4813 if (oxcf->pass == 1 &&
4814 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4815 const int lossless = is_lossless_requested(oxcf);
4816 #if CONFIG_VP9_HIGHBITDEPTH
4817 if (cpi->oxcf.use_highbitdepth)
4818 cpi->td.mb.fwd_txm4x4 = lossless ?
4819 vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
4821 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4822 cpi->td.mb.highbd_itxm_add = lossless ? vp9_highbd_iwht4x4_add :
4823 vp9_highbd_idct4x4_add;
4825 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4826 #endif // CONFIG_VP9_HIGHBITDEPTH
4827 cpi->td.mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
4828 vp9_first_pass(cpi, source);
4829 } else if (oxcf->pass == 2 &&
4830 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4831 Pass2Encode(cpi, size, dest, frame_flags);
4832 } else if (cpi->use_svc) {
4833 SvcEncode(cpi, size, dest, frame_flags);
4836 Pass0Encode(cpi, size, dest, frame_flags);
4839 if (cm->refresh_frame_context)
4840 cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
4842 // No frame encoded, or frame was dropped, release scaled references.
4843 if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
4844 release_scaled_references(cpi);
4848 cpi->droppable = !frame_is_reference(cpi);
4851 // Save layer specific state.
4852 if (is_one_pass_cbr_svc(cpi) ||
4853 ((cpi->svc.number_temporal_layers > 1 ||
4854 cpi->svc.number_spatial_layers > 1) &&
4856 vp9_save_layer_context(cpi);
4859 vpx_usec_timer_mark(&cmptimer);
4860 cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
4862 if (cpi->b_calculate_psnr && oxcf->pass != 1 && cm->show_frame)
4863 generate_psnr_packet(cpi);
4865 if (cpi->keep_level_stats && oxcf->pass != 1)
4866 update_level_info(cpi, size, arf_src_index);
4868 #if CONFIG_INTERNAL_STATS
4870 if (oxcf->pass != 1) {
4871 double samples = 0.0;
4872 cpi->bytes += (int)(*size);
4874 if (cm->show_frame) {
4877 if (cpi->b_calculate_psnr) {
4878 YV12_BUFFER_CONFIG *orig = cpi->Source;
4879 YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
4880 YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer;
4882 #if CONFIG_VP9_HIGHBITDEPTH
4883 calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd,
4884 cpi->oxcf.input_bit_depth);
4886 calc_psnr(orig, recon, &psnr);
4887 #endif // CONFIG_VP9_HIGHBITDEPTH
4889 adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3],
4890 psnr.psnr[0], &cpi->psnr);
4891 cpi->total_sq_error += psnr.sse[0];
4892 cpi->total_samples += psnr.samples[0];
4893 samples = psnr.samples[0];
4897 double frame_ssim2 = 0, weight = 0;
4898 #if CONFIG_VP9_POSTPROC
4899 if (vpx_alloc_frame_buffer(&cm->post_proc_buffer,
4900 recon->y_crop_width, recon->y_crop_height,
4901 cm->subsampling_x, cm->subsampling_y,
4902 #if CONFIG_VP9_HIGHBITDEPTH
4903 cm->use_highbitdepth,
4905 VP9_ENC_BORDER_IN_PIXELS,
4906 cm->byte_alignment) < 0) {
4907 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
4908 "Failed to allocate post processing buffer");
4911 vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer,
4912 cm->lf.filter_level * 10 / 6);
4914 vpx_clear_system_state();
4916 #if CONFIG_VP9_HIGHBITDEPTH
4917 calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd,
4918 cpi->oxcf.input_bit_depth);
4920 calc_psnr(orig, pp, &psnr2);
4921 #endif // CONFIG_VP9_HIGHBITDEPTH
4923 cpi->totalp_sq_error += psnr2.sse[0];
4924 cpi->totalp_samples += psnr2.samples[0];
4925 adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3],
4926 psnr2.psnr[0], &cpi->psnrp);
4928 #if CONFIG_VP9_HIGHBITDEPTH
4929 if (cm->use_highbitdepth) {
4930 frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight,
4931 (int)cm->bit_depth);
4933 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4936 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4937 #endif // CONFIG_VP9_HIGHBITDEPTH
4939 cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2);
4940 cpi->summed_quality += frame_ssim2 * weight;
4941 cpi->summed_weights += weight;
4943 #if CONFIG_VP9_HIGHBITDEPTH
4944 if (cm->use_highbitdepth) {
4945 frame_ssim2 = vpx_highbd_calc_ssim(
4946 orig, &cm->post_proc_buffer, &weight, (int)cm->bit_depth);
4948 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4951 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4952 #endif // CONFIG_VP9_HIGHBITDEPTH
4954 cpi->summedp_quality += frame_ssim2 * weight;
4955 cpi->summedp_weights += weight;
4958 FILE *f = fopen("q_used.stt", "a");
4959 fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
4960 cpi->common.current_video_frame, y2, u2, v2,
4961 frame_psnr2, frame_ssim2);
4967 if (cpi->b_calculate_blockiness) {
4968 #if CONFIG_VP9_HIGHBITDEPTH
4969 if (!cm->use_highbitdepth)
4972 double frame_blockiness = vp9_get_blockiness(
4973 cpi->Source->y_buffer, cpi->Source->y_stride,
4974 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4975 cpi->Source->y_width, cpi->Source->y_height);
4976 cpi->worst_blockiness =
4977 VPXMAX(cpi->worst_blockiness, frame_blockiness);
4978 cpi->total_blockiness += frame_blockiness;
4982 if (cpi->b_calculate_consistency) {
4983 #if CONFIG_VP9_HIGHBITDEPTH
4984 if (!cm->use_highbitdepth)
4987 double this_inconsistency = vpx_get_ssim_metrics(
4988 cpi->Source->y_buffer, cpi->Source->y_stride,
4989 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4990 cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
4993 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
4994 double consistency = vpx_sse_to_psnr(samples, peak,
4995 (double)cpi->total_inconsistency);
4996 if (consistency > 0.0)
4997 cpi->worst_consistency =
4998 VPXMIN(cpi->worst_consistency, consistency);
4999 cpi->total_inconsistency += this_inconsistency;
5003 if (cpi->b_calculate_ssimg) {
5004 double y, u, v, frame_all;
5005 #if CONFIG_VP9_HIGHBITDEPTH
5006 if (cm->use_highbitdepth) {
5007 frame_all = vpx_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y,
5008 &u, &v, (int)cm->bit_depth);
5010 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u,
5014 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v);
5015 #endif // CONFIG_VP9_HIGHBITDEPTH
5016 adjust_image_stat(y, u, v, frame_all, &cpi->ssimg);
5018 #if CONFIG_VP9_HIGHBITDEPTH
5019 if (!cm->use_highbitdepth)
5022 double y, u, v, frame_all;
5023 frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
5025 adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
5026 /* TODO(JBB): add 10/12 bit support */
5028 #if CONFIG_VP9_HIGHBITDEPTH
5029 if (!cm->use_highbitdepth)
5032 double y, u, v, frame_all;
5033 frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
5034 adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
5041 if (is_two_pass_svc(cpi)) {
5042 if (cpi->svc.encode_empty_frame_state == ENCODING) {
5043 cpi->svc.encode_empty_frame_state = ENCODED;
5044 cpi->svc.encode_intra_empty_frame = 0;
5047 if (cm->show_frame) {
5048 ++cpi->svc.spatial_layer_to_encode;
5049 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
5050 cpi->svc.spatial_layer_to_encode = 0;
5052 // May need the empty frame after an visible frame.
5053 cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE;
5055 } else if (is_one_pass_cbr_svc(cpi)) {
5056 if (cm->show_frame) {
5057 ++cpi->svc.spatial_layer_to_encode;
5058 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
5059 cpi->svc.spatial_layer_to_encode = 0;
5062 vpx_clear_system_state();
5066 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
5067 vp9_ppflags_t *flags) {
5068 VP9_COMMON *cm = &cpi->common;
5069 #if !CONFIG_VP9_POSTPROC
5073 if (!cm->show_frame) {
5077 #if CONFIG_VP9_POSTPROC
5078 ret = vp9_post_proc_frame(cm, dest, flags);
5080 if (cm->frame_to_show) {
5081 *dest = *cm->frame_to_show;
5082 dest->y_width = cm->width;
5083 dest->y_height = cm->height;
5084 dest->uv_width = cm->width >> cm->subsampling_x;
5085 dest->uv_height = cm->height >> cm->subsampling_y;
5090 #endif // !CONFIG_VP9_POSTPROC
5091 vpx_clear_system_state();
5096 int vp9_set_internal_size(VP9_COMP *cpi,
5097 VPX_SCALING horiz_mode, VPX_SCALING vert_mode) {
5098 VP9_COMMON *cm = &cpi->common;
5099 int hr = 0, hs = 0, vr = 0, vs = 0;
5101 if (horiz_mode > ONETWO || vert_mode > ONETWO)
5104 Scale2Ratio(horiz_mode, &hr, &hs);
5105 Scale2Ratio(vert_mode, &vr, &vs);
5107 // always go to the next whole number
5108 cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
5109 cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
5110 if (cm->current_video_frame) {
5111 assert(cm->width <= cpi->initial_width);
5112 assert(cm->height <= cpi->initial_height);
5115 update_frame_size(cpi);
5120 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
5121 unsigned int height) {
5122 VP9_COMMON *cm = &cpi->common;
5123 #if CONFIG_VP9_HIGHBITDEPTH
5124 check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
5126 check_initial_width(cpi, 1, 1);
5127 #endif // CONFIG_VP9_HIGHBITDEPTH
5129 #if CONFIG_VP9_TEMPORAL_DENOISING
5130 setup_denoiser_buffer(cpi);
5135 if (cm->width > cpi->initial_width) {
5136 cm->width = cpi->initial_width;
5137 printf("Warning: Desired width too large, changed to %d\n", cm->width);
5142 cm->height = height;
5143 if (cm->height > cpi->initial_height) {
5144 cm->height = cpi->initial_height;
5145 printf("Warning: Desired height too large, changed to %d\n", cm->height);
5148 assert(cm->width <= cpi->initial_width);
5149 assert(cm->height <= cpi->initial_height);
5151 update_frame_size(cpi);
5156 void vp9_set_svc(VP9_COMP *cpi, int use_svc) {
5157 cpi->use_svc = use_svc;
5161 int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a,
5162 const YV12_BUFFER_CONFIG *b) {
5163 assert(a->y_crop_width == b->y_crop_width);
5164 assert(a->y_crop_height == b->y_crop_height);
5166 return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
5167 a->y_crop_width, a->y_crop_height);
5170 #if CONFIG_VP9_HIGHBITDEPTH
5171 int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
5172 const YV12_BUFFER_CONFIG *b) {
5173 assert(a->y_crop_width == b->y_crop_width);
5174 assert(a->y_crop_height == b->y_crop_height);
5175 assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
5176 assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
5178 return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
5179 a->y_crop_width, a->y_crop_height);
5181 #endif // CONFIG_VP9_HIGHBITDEPTH
5183 int vp9_get_quantizer(VP9_COMP *cpi) {
5184 return cpi->common.base_qindex;
5187 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
5188 if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF |
5189 VP8_EFLAG_NO_REF_ARF)) {
5192 if (flags & VP8_EFLAG_NO_REF_LAST)
5193 ref ^= VP9_LAST_FLAG;
5195 if (flags & VP8_EFLAG_NO_REF_GF)
5196 ref ^= VP9_GOLD_FLAG;
5198 if (flags & VP8_EFLAG_NO_REF_ARF)
5199 ref ^= VP9_ALT_FLAG;
5201 vp9_use_as_reference(cpi, ref);
5204 if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
5205 VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF |
5206 VP8_EFLAG_FORCE_ARF)) {
5209 if (flags & VP8_EFLAG_NO_UPD_LAST)
5210 upd ^= VP9_LAST_FLAG;
5212 if (flags & VP8_EFLAG_NO_UPD_GF)
5213 upd ^= VP9_GOLD_FLAG;
5215 if (flags & VP8_EFLAG_NO_UPD_ARF)
5216 upd ^= VP9_ALT_FLAG;
5218 vp9_update_reference(cpi, upd);
5221 if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
5222 vp9_update_entropy(cpi, 0);