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 INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) {
115 // Mark all inactive blocks as active. Other segmentation features may be set
116 // so memset cannot be used, instead only inactive blocks should be reset.
117 static void suppress_active_map(VP9_COMP *cpi) {
118 unsigned char *const seg_map = cpi->segmentation_map;
120 if (cpi->active_map.enabled || cpi->active_map.update)
121 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
122 if (seg_map[i] == AM_SEGMENT_ID_INACTIVE)
123 seg_map[i] = AM_SEGMENT_ID_ACTIVE;
126 static void apply_active_map(VP9_COMP *cpi) {
127 struct segmentation *const seg = &cpi->common.seg;
128 unsigned char *const seg_map = cpi->segmentation_map;
129 const unsigned char *const active_map = cpi->active_map.map;
132 assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE);
134 if (frame_is_intra_only(&cpi->common)) {
135 cpi->active_map.enabled = 0;
136 cpi->active_map.update = 1;
139 if (cpi->active_map.update) {
140 if (cpi->active_map.enabled) {
141 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
142 if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i];
143 vp9_enable_segmentation(seg);
144 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
145 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
146 // Setting the data to -MAX_LOOP_FILTER will result in the computed loop
147 // filter level being zero regardless of the value of seg->abs_delta.
148 vp9_set_segdata(seg, AM_SEGMENT_ID_INACTIVE,
149 SEG_LVL_ALT_LF, -MAX_LOOP_FILTER);
151 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
152 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
154 seg->update_data = 1;
158 cpi->active_map.update = 0;
162 int vp9_set_active_map(VP9_COMP* cpi,
163 unsigned char* new_map_16x16,
166 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
167 unsigned char *const active_map_8x8 = cpi->active_map.map;
168 const int mi_rows = cpi->common.mi_rows;
169 const int mi_cols = cpi->common.mi_cols;
170 cpi->active_map.update = 1;
173 for (r = 0; r < mi_rows; ++r) {
174 for (c = 0; c < mi_cols; ++c) {
175 active_map_8x8[r * mi_cols + c] =
176 new_map_16x16[(r >> 1) * cols + (c >> 1)]
177 ? AM_SEGMENT_ID_ACTIVE
178 : AM_SEGMENT_ID_INACTIVE;
181 cpi->active_map.enabled = 1;
183 cpi->active_map.enabled = 0;
191 int vp9_get_active_map(VP9_COMP* cpi,
192 unsigned char* new_map_16x16,
195 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols &&
197 unsigned char* const seg_map_8x8 = cpi->segmentation_map;
198 const int mi_rows = cpi->common.mi_rows;
199 const int mi_cols = cpi->common.mi_cols;
200 memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
201 if (cpi->active_map.enabled) {
203 for (r = 0; r < mi_rows; ++r) {
204 for (c = 0; c < mi_cols; ++c) {
205 // Cyclic refresh segments are considered active despite not having
206 // AM_SEGMENT_ID_ACTIVE
207 new_map_16x16[(r >> 1) * cols + (c >> 1)] |=
208 seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE;
218 void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
219 MACROBLOCK *const mb = &cpi->td.mb;
220 cpi->common.allow_high_precision_mv = allow_high_precision_mv;
221 if (cpi->common.allow_high_precision_mv) {
222 mb->mvcost = mb->nmvcost_hp;
223 mb->mvsadcost = mb->nmvsadcost_hp;
225 mb->mvcost = mb->nmvcost;
226 mb->mvsadcost = mb->nmvsadcost;
230 static void setup_frame(VP9_COMP *cpi) {
231 VP9_COMMON *const cm = &cpi->common;
232 // Set up entropy context depending on frame type. The decoder mandates
233 // the use of the default context, index 0, for keyframes and inter
234 // frames where the error_resilient_mode or intra_only flag is set. For
235 // other inter-frames the encoder currently uses only two contexts;
236 // context 1 for ALTREF frames and context 0 for the others.
237 if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
238 vp9_setup_past_independence(cm);
241 cm->frame_context_idx = cpi->refresh_alt_ref_frame;
244 if (cm->frame_type == KEY_FRAME) {
245 if (!is_two_pass_svc(cpi))
246 cpi->refresh_golden_frame = 1;
247 cpi->refresh_alt_ref_frame = 1;
248 vp9_zero(cpi->interp_filter_selected);
250 *cm->fc = cm->frame_contexts[cm->frame_context_idx];
251 vp9_zero(cpi->interp_filter_selected[0]);
255 static void vp9_enc_setup_mi(VP9_COMMON *cm) {
257 cm->mi = cm->mip + cm->mi_stride + 1;
258 memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
259 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
260 // Clear top border row
261 memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride);
262 // Clear left border column
263 for (i = 1; i < cm->mi_rows + 1; ++i)
264 memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip));
266 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
267 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
269 memset(cm->mi_grid_base, 0,
270 cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
273 static int vp9_enc_alloc_mi(VP9_COMMON *cm, int mi_size) {
274 cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
277 cm->prev_mip = vpx_calloc(mi_size, sizeof(*cm->prev_mip));
280 cm->mi_alloc_size = mi_size;
282 cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
283 if (!cm->mi_grid_base)
285 cm->prev_mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
286 if (!cm->prev_mi_grid_base)
292 static void vp9_enc_free_mi(VP9_COMMON *cm) {
295 vpx_free(cm->prev_mip);
297 vpx_free(cm->mi_grid_base);
298 cm->mi_grid_base = NULL;
299 vpx_free(cm->prev_mi_grid_base);
300 cm->prev_mi_grid_base = NULL;
303 static void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
304 // Current mip will be the prev_mip for the next frame.
305 MODE_INFO **temp_base = cm->prev_mi_grid_base;
306 MODE_INFO *temp = cm->prev_mip;
307 cm->prev_mip = cm->mip;
310 // Update the upper left visible macroblock ptrs.
311 cm->mi = cm->mip + cm->mi_stride + 1;
312 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
314 cm->prev_mi_grid_base = cm->mi_grid_base;
315 cm->mi_grid_base = temp_base;
316 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
317 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
320 void vp9_initialize_enc(void) {
321 static volatile int init_done = 0;
327 vp9_init_intra_predictors();
329 vp9_rc_init_minq_luts();
330 vp9_entropy_mv_init();
331 vp9_temporal_filter_init();
336 static void dealloc_compressor_data(VP9_COMP *cpi) {
337 VP9_COMMON *const cm = &cpi->common;
340 vpx_free(cpi->mbmi_ext_base);
341 cpi->mbmi_ext_base = NULL;
343 vpx_free(cpi->tile_data);
344 cpi->tile_data = NULL;
346 // Delete sementation map
347 vpx_free(cpi->segmentation_map);
348 cpi->segmentation_map = NULL;
349 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
350 cpi->coding_context.last_frame_seg_map_copy = NULL;
352 vpx_free(cpi->nmvcosts[0]);
353 vpx_free(cpi->nmvcosts[1]);
354 cpi->nmvcosts[0] = NULL;
355 cpi->nmvcosts[1] = NULL;
357 vpx_free(cpi->nmvcosts_hp[0]);
358 vpx_free(cpi->nmvcosts_hp[1]);
359 cpi->nmvcosts_hp[0] = NULL;
360 cpi->nmvcosts_hp[1] = NULL;
362 vpx_free(cpi->nmvsadcosts[0]);
363 vpx_free(cpi->nmvsadcosts[1]);
364 cpi->nmvsadcosts[0] = NULL;
365 cpi->nmvsadcosts[1] = NULL;
367 vpx_free(cpi->nmvsadcosts_hp[0]);
368 vpx_free(cpi->nmvsadcosts_hp[1]);
369 cpi->nmvsadcosts_hp[0] = NULL;
370 cpi->nmvsadcosts_hp[1] = NULL;
372 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
373 cpi->cyclic_refresh = NULL;
375 vpx_free(cpi->active_map.map);
376 cpi->active_map.map = NULL;
378 vpx_free(cpi->consec_zero_mv);
379 cpi->consec_zero_mv = NULL;
381 vp9_free_ref_frame_buffers(cm->buffer_pool);
382 #if CONFIG_VP9_POSTPROC
383 vp9_free_postproc_buffers(cm);
385 vp9_free_context_buffers(cm);
387 vpx_free_frame_buffer(&cpi->last_frame_uf);
388 vpx_free_frame_buffer(&cpi->scaled_source);
389 vpx_free_frame_buffer(&cpi->scaled_last_source);
390 vpx_free_frame_buffer(&cpi->alt_ref_buffer);
391 vp9_lookahead_destroy(cpi->lookahead);
393 vpx_free(cpi->tile_tok[0][0]);
394 cpi->tile_tok[0][0] = 0;
396 vp9_free_pc_tree(&cpi->td);
398 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
399 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
400 vpx_free(lc->rc_twopass_stats_in.buf);
401 lc->rc_twopass_stats_in.buf = NULL;
402 lc->rc_twopass_stats_in.sz = 0;
405 if (cpi->source_diff_var != NULL) {
406 vpx_free(cpi->source_diff_var);
407 cpi->source_diff_var = NULL;
410 for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
411 vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]);
413 memset(&cpi->svc.scaled_frames[0], 0,
414 MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
416 vpx_free_frame_buffer(&cpi->svc.scaled_temp);
417 memset(&cpi->svc.scaled_temp, 0, sizeof(cpi->svc.scaled_temp));
419 vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
420 memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
422 vp9_free_svc_cyclic_refresh(cpi);
425 static void save_coding_context(VP9_COMP *cpi) {
426 CODING_CONTEXT *const cc = &cpi->coding_context;
427 VP9_COMMON *cm = &cpi->common;
429 // Stores a snapshot of key state variables which can subsequently be
430 // restored with a call to vp9_restore_coding_context. These functions are
431 // intended for use in a re-code loop in vp9_compress_frame where the
432 // quantizer value is adjusted between loop iterations.
433 vp9_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost);
435 memcpy(cc->nmvcosts[0], cpi->nmvcosts[0],
436 MV_VALS * sizeof(*cpi->nmvcosts[0]));
437 memcpy(cc->nmvcosts[1], cpi->nmvcosts[1],
438 MV_VALS * sizeof(*cpi->nmvcosts[1]));
439 memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0],
440 MV_VALS * sizeof(*cpi->nmvcosts_hp[0]));
441 memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1],
442 MV_VALS * sizeof(*cpi->nmvcosts_hp[1]));
444 vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs);
446 memcpy(cpi->coding_context.last_frame_seg_map_copy,
447 cm->last_frame_seg_map, (cm->mi_rows * cm->mi_cols));
449 vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
450 vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
455 static void restore_coding_context(VP9_COMP *cpi) {
456 CODING_CONTEXT *const cc = &cpi->coding_context;
457 VP9_COMMON *cm = &cpi->common;
459 // Restore key state variables to the snapshot state stored in the
460 // previous call to vp9_save_coding_context.
461 vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
463 memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0]));
464 memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1]));
465 memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0],
466 MV_VALS * sizeof(*cc->nmvcosts_hp[0]));
467 memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1],
468 MV_VALS * sizeof(*cc->nmvcosts_hp[1]));
470 vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs);
472 memcpy(cm->last_frame_seg_map,
473 cpi->coding_context.last_frame_seg_map_copy,
474 (cm->mi_rows * cm->mi_cols));
476 vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
477 vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
482 static void configure_static_seg_features(VP9_COMP *cpi) {
483 VP9_COMMON *const cm = &cpi->common;
484 const RATE_CONTROL *const rc = &cpi->rc;
485 struct segmentation *const seg = &cm->seg;
487 int high_q = (int)(rc->avg_q > 48.0);
490 // Disable and clear down for KF
491 if (cm->frame_type == KEY_FRAME) {
492 // Clear down the global segmentation map
493 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
495 seg->update_data = 0;
496 cpi->static_mb_pct = 0;
498 // Disable segmentation
499 vp9_disable_segmentation(seg);
501 // Clear down the segment features.
502 vp9_clearall_segfeatures(seg);
503 } else if (cpi->refresh_alt_ref_frame) {
504 // If this is an alt ref frame
505 // Clear down the global segmentation map
506 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
508 seg->update_data = 0;
509 cpi->static_mb_pct = 0;
511 // Disable segmentation and individual segment features by default
512 vp9_disable_segmentation(seg);
513 vp9_clearall_segfeatures(seg);
515 // Scan frames from current to arf frame.
516 // This function re-enables segmentation if appropriate.
517 vp9_update_mbgraph_stats(cpi);
519 // If segmentation was enabled set those features needed for the
523 seg->update_data = 1;
525 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875,
527 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
528 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
530 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
531 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
533 // Where relevant assume segment data is delta data
534 seg->abs_delta = SEGMENT_DELTADATA;
536 } else if (seg->enabled) {
537 // All other frames if segmentation has been enabled
539 // First normal frame in a valid gf or alt ref group
540 if (rc->frames_since_golden == 0) {
541 // Set up segment features for normal frames in an arf group
542 if (rc->source_alt_ref_active) {
544 seg->update_data = 1;
545 seg->abs_delta = SEGMENT_DELTADATA;
547 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125,
549 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
550 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
552 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
553 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
555 // Segment coding disabled for compred testing
556 if (high_q || (cpi->static_mb_pct == 100)) {
557 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
558 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
559 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
562 // Disable segmentation and clear down features if alt ref
563 // is not active for this group
565 vp9_disable_segmentation(seg);
567 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
570 seg->update_data = 0;
572 vp9_clearall_segfeatures(seg);
574 } else if (rc->is_src_frame_alt_ref) {
575 // Special case where we are coding over the top of a previous
577 // Segment coding disabled for compred testing
579 // Enable ref frame features for segment 0 as well
580 vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
581 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
583 // All mbs should use ALTREF_FRAME
584 vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
585 vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
586 vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
587 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
589 // Skip all MBs if high Q (0,0 mv and skip coeffs)
591 vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP);
592 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
594 // Enable data update
595 seg->update_data = 1;
599 // No updates.. leave things as they are.
601 seg->update_data = 0;
606 static void update_reference_segmentation_map(VP9_COMP *cpi) {
607 VP9_COMMON *const cm = &cpi->common;
608 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
609 uint8_t *cache_ptr = cm->last_frame_seg_map;
612 for (row = 0; row < cm->mi_rows; row++) {
613 MODE_INFO **mi_8x8 = mi_8x8_ptr;
614 uint8_t *cache = cache_ptr;
615 for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
616 cache[0] = mi_8x8[0]->segment_id;
617 mi_8x8_ptr += cm->mi_stride;
618 cache_ptr += cm->mi_cols;
622 static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
623 VP9_COMMON *cm = &cpi->common;
624 const VP9EncoderConfig *oxcf = &cpi->oxcf;
627 cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
628 cm->subsampling_x, cm->subsampling_y,
629 #if CONFIG_VP9_HIGHBITDEPTH
630 cm->use_highbitdepth,
632 oxcf->lag_in_frames);
634 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
635 "Failed to allocate lag buffers");
637 // TODO(agrange) Check if ARF is enabled and skip allocation if not.
638 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
639 oxcf->width, oxcf->height,
640 cm->subsampling_x, cm->subsampling_y,
641 #if CONFIG_VP9_HIGHBITDEPTH
642 cm->use_highbitdepth,
644 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
646 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
647 "Failed to allocate altref buffer");
650 static void alloc_util_frame_buffers(VP9_COMP *cpi) {
651 VP9_COMMON *const cm = &cpi->common;
652 if (vpx_realloc_frame_buffer(&cpi->last_frame_uf,
653 cm->width, cm->height,
654 cm->subsampling_x, cm->subsampling_y,
655 #if CONFIG_VP9_HIGHBITDEPTH
656 cm->use_highbitdepth,
658 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
660 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
661 "Failed to allocate last frame buffer");
663 if (vpx_realloc_frame_buffer(&cpi->scaled_source,
664 cm->width, cm->height,
665 cm->subsampling_x, cm->subsampling_y,
666 #if CONFIG_VP9_HIGHBITDEPTH
667 cm->use_highbitdepth,
669 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
671 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
672 "Failed to allocate scaled source buffer");
674 if (vpx_realloc_frame_buffer(&cpi->scaled_last_source,
675 cm->width, cm->height,
676 cm->subsampling_x, cm->subsampling_y,
677 #if CONFIG_VP9_HIGHBITDEPTH
678 cm->use_highbitdepth,
680 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
682 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
683 "Failed to allocate scaled last source buffer");
687 static int alloc_context_buffers_ext(VP9_COMP *cpi) {
688 VP9_COMMON *cm = &cpi->common;
689 int mi_size = cm->mi_cols * cm->mi_rows;
691 cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
692 if (!cpi->mbmi_ext_base)
698 static void alloc_compressor_data(VP9_COMP *cpi) {
699 VP9_COMMON *cm = &cpi->common;
701 vp9_alloc_context_buffers(cm, cm->width, cm->height);
703 alloc_context_buffers_ext(cpi);
705 vpx_free(cpi->tile_tok[0][0]);
708 unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
709 CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
710 vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
713 vp9_setup_pc_tree(&cpi->common, &cpi->td);
716 void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
717 cpi->framerate = framerate < 0.1 ? 30 : framerate;
718 vp9_rc_update_framerate(cpi);
721 static void set_tile_limits(VP9_COMP *cpi) {
722 VP9_COMMON *const cm = &cpi->common;
724 int min_log2_tile_cols, max_log2_tile_cols;
725 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
727 if (is_two_pass_svc(cpi) &&
728 (cpi->svc.encode_empty_frame_state == ENCODING ||
729 cpi->svc.number_spatial_layers > 1)) {
730 cm->log2_tile_cols = 0;
731 cm->log2_tile_rows = 0;
733 cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns,
734 min_log2_tile_cols, max_log2_tile_cols);
735 cm->log2_tile_rows = cpi->oxcf.tile_rows;
739 static void update_frame_size(VP9_COMP *cpi) {
740 VP9_COMMON *const cm = &cpi->common;
741 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
743 vp9_set_mb_mi(cm, cm->width, cm->height);
744 vp9_init_context_buffers(cm);
745 vp9_init_macroblockd(cm, xd, NULL);
746 cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base;
747 memset(cpi->mbmi_ext_base, 0,
748 cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
750 set_tile_limits(cpi);
752 if (is_two_pass_svc(cpi)) {
753 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
754 cm->width, cm->height,
755 cm->subsampling_x, cm->subsampling_y,
756 #if CONFIG_VP9_HIGHBITDEPTH
757 cm->use_highbitdepth,
759 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
761 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
762 "Failed to reallocate alt_ref_buffer");
766 static void init_buffer_indices(VP9_COMP *cpi) {
772 static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
773 VP9_COMMON *const cm = &cpi->common;
776 cpi->framerate = oxcf->init_framerate;
777 cm->profile = oxcf->profile;
778 cm->bit_depth = oxcf->bit_depth;
779 #if CONFIG_VP9_HIGHBITDEPTH
780 cm->use_highbitdepth = oxcf->use_highbitdepth;
782 cm->color_space = oxcf->color_space;
783 cm->color_range = oxcf->color_range;
785 cpi->target_level = oxcf->target_level;
786 cm->keep_level_stats = oxcf->target_level != LEVEL_NOT_CARE;
788 cm->width = oxcf->width;
789 cm->height = oxcf->height;
790 alloc_compressor_data(cpi);
792 cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode;
794 // Single thread case: use counts in common.
795 cpi->td.counts = &cm->counts;
797 // Spatial scalability.
798 cpi->svc.number_spatial_layers = oxcf->ss_number_layers;
799 // Temporal scalability.
800 cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
802 if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
803 ((cpi->svc.number_temporal_layers > 1 ||
804 cpi->svc.number_spatial_layers > 1) &&
805 cpi->oxcf.pass != 1)) {
806 vp9_init_layer_context(cpi);
809 // change includes all joint functionality
810 vp9_change_config(cpi, oxcf);
812 cpi->static_mb_pct = 0;
813 cpi->ref_frame_flags = 0;
815 init_buffer_indices(cpi);
817 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
820 static void set_rc_buffer_sizes(RATE_CONTROL *rc,
821 const VP9EncoderConfig *oxcf) {
822 const int64_t bandwidth = oxcf->target_bandwidth;
823 const int64_t starting = oxcf->starting_buffer_level_ms;
824 const int64_t optimal = oxcf->optimal_buffer_level_ms;
825 const int64_t maximum = oxcf->maximum_buffer_size_ms;
827 rc->starting_buffer_level = starting * bandwidth / 1000;
828 rc->optimal_buffer_level = (optimal == 0) ? bandwidth / 8
829 : optimal * bandwidth / 1000;
830 rc->maximum_buffer_size = (maximum == 0) ? bandwidth / 8
831 : maximum * bandwidth / 1000;
834 #if CONFIG_VP9_HIGHBITDEPTH
835 #define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
836 cpi->fn_ptr[BT].sdf = SDF; \
837 cpi->fn_ptr[BT].sdaf = SDAF; \
838 cpi->fn_ptr[BT].vf = VF; \
839 cpi->fn_ptr[BT].svf = SVF; \
840 cpi->fn_ptr[BT].svaf = SVAF; \
841 cpi->fn_ptr[BT].sdx3f = SDX3F; \
842 cpi->fn_ptr[BT].sdx8f = SDX8F; \
843 cpi->fn_ptr[BT].sdx4df = SDX4DF;
845 #define MAKE_BFP_SAD_WRAPPER(fnname) \
846 static unsigned int fnname##_bits8(const uint8_t *src_ptr, \
848 const uint8_t *ref_ptr, \
850 return fnname(src_ptr, source_stride, ref_ptr, ref_stride); \
852 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
854 const uint8_t *ref_ptr, \
856 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2; \
858 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
860 const uint8_t *ref_ptr, \
862 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4; \
865 #define MAKE_BFP_SADAVG_WRAPPER(fnname) static unsigned int \
866 fnname##_bits8(const uint8_t *src_ptr, \
868 const uint8_t *ref_ptr, \
870 const uint8_t *second_pred) { \
871 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred); \
873 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
875 const uint8_t *ref_ptr, \
877 const uint8_t *second_pred) { \
878 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
881 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
883 const uint8_t *ref_ptr, \
885 const uint8_t *second_pred) { \
886 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
890 #define MAKE_BFP_SAD3_WRAPPER(fnname) \
891 static void fnname##_bits8(const uint8_t *src_ptr, \
893 const uint8_t *ref_ptr, \
895 unsigned int *sad_array) { \
896 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
898 static void fnname##_bits10(const uint8_t *src_ptr, \
900 const uint8_t *ref_ptr, \
902 unsigned int *sad_array) { \
904 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
905 for (i = 0; i < 3; i++) \
906 sad_array[i] >>= 2; \
908 static void fnname##_bits12(const uint8_t *src_ptr, \
910 const uint8_t *ref_ptr, \
912 unsigned int *sad_array) { \
914 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
915 for (i = 0; i < 3; i++) \
916 sad_array[i] >>= 4; \
919 #define MAKE_BFP_SAD8_WRAPPER(fnname) \
920 static void fnname##_bits8(const uint8_t *src_ptr, \
922 const uint8_t *ref_ptr, \
924 unsigned int *sad_array) { \
925 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
927 static void fnname##_bits10(const uint8_t *src_ptr, \
929 const uint8_t *ref_ptr, \
931 unsigned int *sad_array) { \
933 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
934 for (i = 0; i < 8; i++) \
935 sad_array[i] >>= 2; \
937 static void fnname##_bits12(const uint8_t *src_ptr, \
939 const uint8_t *ref_ptr, \
941 unsigned int *sad_array) { \
943 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
944 for (i = 0; i < 8; i++) \
945 sad_array[i] >>= 4; \
947 #define MAKE_BFP_SAD4D_WRAPPER(fnname) \
948 static void fnname##_bits8(const uint8_t *src_ptr, \
950 const uint8_t* const ref_ptr[], \
952 unsigned int *sad_array) { \
953 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
955 static void fnname##_bits10(const uint8_t *src_ptr, \
957 const uint8_t* const ref_ptr[], \
959 unsigned int *sad_array) { \
961 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
962 for (i = 0; i < 4; i++) \
963 sad_array[i] >>= 2; \
965 static void fnname##_bits12(const uint8_t *src_ptr, \
967 const uint8_t* const ref_ptr[], \
969 unsigned int *sad_array) { \
971 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
972 for (i = 0; i < 4; i++) \
973 sad_array[i] >>= 4; \
976 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16)
977 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg)
978 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d)
979 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32)
980 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg)
981 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d)
982 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32)
983 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg)
984 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d)
985 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64)
986 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg)
987 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d)
988 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32)
989 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg)
990 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad32x32x3)
991 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad32x32x8)
992 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d)
993 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64)
994 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg)
995 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad64x64x3)
996 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad64x64x8)
997 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d)
998 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16)
999 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg)
1000 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x16x3)
1001 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x16x8)
1002 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d)
1003 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8)
1004 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg)
1005 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x8x3)
1006 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x8x8)
1007 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d)
1008 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16)
1009 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg)
1010 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x16x3)
1011 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x16x8)
1012 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d)
1013 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8)
1014 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg)
1015 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x8x3)
1016 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x8x8)
1017 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d)
1018 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4)
1019 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg)
1020 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x4x8)
1021 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d)
1022 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8)
1023 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg)
1024 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x8x8)
1025 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d)
1026 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4)
1027 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg)
1028 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad4x4x3)
1029 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x4x8)
1030 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d)
1032 static void highbd_set_var_fns(VP9_COMP *const cpi) {
1033 VP9_COMMON *const cm = &cpi->common;
1034 if (cm->use_highbitdepth) {
1035 switch (cm->bit_depth) {
1037 HIGHBD_BFP(BLOCK_32X16,
1038 vpx_highbd_sad32x16_bits8,
1039 vpx_highbd_sad32x16_avg_bits8,
1040 vpx_highbd_8_variance32x16,
1041 vpx_highbd_8_sub_pixel_variance32x16,
1042 vpx_highbd_8_sub_pixel_avg_variance32x16,
1045 vpx_highbd_sad32x16x4d_bits8)
1047 HIGHBD_BFP(BLOCK_16X32,
1048 vpx_highbd_sad16x32_bits8,
1049 vpx_highbd_sad16x32_avg_bits8,
1050 vpx_highbd_8_variance16x32,
1051 vpx_highbd_8_sub_pixel_variance16x32,
1052 vpx_highbd_8_sub_pixel_avg_variance16x32,
1055 vpx_highbd_sad16x32x4d_bits8)
1057 HIGHBD_BFP(BLOCK_64X32,
1058 vpx_highbd_sad64x32_bits8,
1059 vpx_highbd_sad64x32_avg_bits8,
1060 vpx_highbd_8_variance64x32,
1061 vpx_highbd_8_sub_pixel_variance64x32,
1062 vpx_highbd_8_sub_pixel_avg_variance64x32,
1065 vpx_highbd_sad64x32x4d_bits8)
1067 HIGHBD_BFP(BLOCK_32X64,
1068 vpx_highbd_sad32x64_bits8,
1069 vpx_highbd_sad32x64_avg_bits8,
1070 vpx_highbd_8_variance32x64,
1071 vpx_highbd_8_sub_pixel_variance32x64,
1072 vpx_highbd_8_sub_pixel_avg_variance32x64,
1075 vpx_highbd_sad32x64x4d_bits8)
1077 HIGHBD_BFP(BLOCK_32X32,
1078 vpx_highbd_sad32x32_bits8,
1079 vpx_highbd_sad32x32_avg_bits8,
1080 vpx_highbd_8_variance32x32,
1081 vpx_highbd_8_sub_pixel_variance32x32,
1082 vpx_highbd_8_sub_pixel_avg_variance32x32,
1083 vpx_highbd_sad32x32x3_bits8,
1084 vpx_highbd_sad32x32x8_bits8,
1085 vpx_highbd_sad32x32x4d_bits8)
1087 HIGHBD_BFP(BLOCK_64X64,
1088 vpx_highbd_sad64x64_bits8,
1089 vpx_highbd_sad64x64_avg_bits8,
1090 vpx_highbd_8_variance64x64,
1091 vpx_highbd_8_sub_pixel_variance64x64,
1092 vpx_highbd_8_sub_pixel_avg_variance64x64,
1093 vpx_highbd_sad64x64x3_bits8,
1094 vpx_highbd_sad64x64x8_bits8,
1095 vpx_highbd_sad64x64x4d_bits8)
1097 HIGHBD_BFP(BLOCK_16X16,
1098 vpx_highbd_sad16x16_bits8,
1099 vpx_highbd_sad16x16_avg_bits8,
1100 vpx_highbd_8_variance16x16,
1101 vpx_highbd_8_sub_pixel_variance16x16,
1102 vpx_highbd_8_sub_pixel_avg_variance16x16,
1103 vpx_highbd_sad16x16x3_bits8,
1104 vpx_highbd_sad16x16x8_bits8,
1105 vpx_highbd_sad16x16x4d_bits8)
1107 HIGHBD_BFP(BLOCK_16X8,
1108 vpx_highbd_sad16x8_bits8,
1109 vpx_highbd_sad16x8_avg_bits8,
1110 vpx_highbd_8_variance16x8,
1111 vpx_highbd_8_sub_pixel_variance16x8,
1112 vpx_highbd_8_sub_pixel_avg_variance16x8,
1113 vpx_highbd_sad16x8x3_bits8,
1114 vpx_highbd_sad16x8x8_bits8,
1115 vpx_highbd_sad16x8x4d_bits8)
1117 HIGHBD_BFP(BLOCK_8X16,
1118 vpx_highbd_sad8x16_bits8,
1119 vpx_highbd_sad8x16_avg_bits8,
1120 vpx_highbd_8_variance8x16,
1121 vpx_highbd_8_sub_pixel_variance8x16,
1122 vpx_highbd_8_sub_pixel_avg_variance8x16,
1123 vpx_highbd_sad8x16x3_bits8,
1124 vpx_highbd_sad8x16x8_bits8,
1125 vpx_highbd_sad8x16x4d_bits8)
1127 HIGHBD_BFP(BLOCK_8X8,
1128 vpx_highbd_sad8x8_bits8,
1129 vpx_highbd_sad8x8_avg_bits8,
1130 vpx_highbd_8_variance8x8,
1131 vpx_highbd_8_sub_pixel_variance8x8,
1132 vpx_highbd_8_sub_pixel_avg_variance8x8,
1133 vpx_highbd_sad8x8x3_bits8,
1134 vpx_highbd_sad8x8x8_bits8,
1135 vpx_highbd_sad8x8x4d_bits8)
1137 HIGHBD_BFP(BLOCK_8X4,
1138 vpx_highbd_sad8x4_bits8,
1139 vpx_highbd_sad8x4_avg_bits8,
1140 vpx_highbd_8_variance8x4,
1141 vpx_highbd_8_sub_pixel_variance8x4,
1142 vpx_highbd_8_sub_pixel_avg_variance8x4,
1144 vpx_highbd_sad8x4x8_bits8,
1145 vpx_highbd_sad8x4x4d_bits8)
1147 HIGHBD_BFP(BLOCK_4X8,
1148 vpx_highbd_sad4x8_bits8,
1149 vpx_highbd_sad4x8_avg_bits8,
1150 vpx_highbd_8_variance4x8,
1151 vpx_highbd_8_sub_pixel_variance4x8,
1152 vpx_highbd_8_sub_pixel_avg_variance4x8,
1154 vpx_highbd_sad4x8x8_bits8,
1155 vpx_highbd_sad4x8x4d_bits8)
1157 HIGHBD_BFP(BLOCK_4X4,
1158 vpx_highbd_sad4x4_bits8,
1159 vpx_highbd_sad4x4_avg_bits8,
1160 vpx_highbd_8_variance4x4,
1161 vpx_highbd_8_sub_pixel_variance4x4,
1162 vpx_highbd_8_sub_pixel_avg_variance4x4,
1163 vpx_highbd_sad4x4x3_bits8,
1164 vpx_highbd_sad4x4x8_bits8,
1165 vpx_highbd_sad4x4x4d_bits8)
1169 HIGHBD_BFP(BLOCK_32X16,
1170 vpx_highbd_sad32x16_bits10,
1171 vpx_highbd_sad32x16_avg_bits10,
1172 vpx_highbd_10_variance32x16,
1173 vpx_highbd_10_sub_pixel_variance32x16,
1174 vpx_highbd_10_sub_pixel_avg_variance32x16,
1177 vpx_highbd_sad32x16x4d_bits10)
1179 HIGHBD_BFP(BLOCK_16X32,
1180 vpx_highbd_sad16x32_bits10,
1181 vpx_highbd_sad16x32_avg_bits10,
1182 vpx_highbd_10_variance16x32,
1183 vpx_highbd_10_sub_pixel_variance16x32,
1184 vpx_highbd_10_sub_pixel_avg_variance16x32,
1187 vpx_highbd_sad16x32x4d_bits10)
1189 HIGHBD_BFP(BLOCK_64X32,
1190 vpx_highbd_sad64x32_bits10,
1191 vpx_highbd_sad64x32_avg_bits10,
1192 vpx_highbd_10_variance64x32,
1193 vpx_highbd_10_sub_pixel_variance64x32,
1194 vpx_highbd_10_sub_pixel_avg_variance64x32,
1197 vpx_highbd_sad64x32x4d_bits10)
1199 HIGHBD_BFP(BLOCK_32X64,
1200 vpx_highbd_sad32x64_bits10,
1201 vpx_highbd_sad32x64_avg_bits10,
1202 vpx_highbd_10_variance32x64,
1203 vpx_highbd_10_sub_pixel_variance32x64,
1204 vpx_highbd_10_sub_pixel_avg_variance32x64,
1207 vpx_highbd_sad32x64x4d_bits10)
1209 HIGHBD_BFP(BLOCK_32X32,
1210 vpx_highbd_sad32x32_bits10,
1211 vpx_highbd_sad32x32_avg_bits10,
1212 vpx_highbd_10_variance32x32,
1213 vpx_highbd_10_sub_pixel_variance32x32,
1214 vpx_highbd_10_sub_pixel_avg_variance32x32,
1215 vpx_highbd_sad32x32x3_bits10,
1216 vpx_highbd_sad32x32x8_bits10,
1217 vpx_highbd_sad32x32x4d_bits10)
1219 HIGHBD_BFP(BLOCK_64X64,
1220 vpx_highbd_sad64x64_bits10,
1221 vpx_highbd_sad64x64_avg_bits10,
1222 vpx_highbd_10_variance64x64,
1223 vpx_highbd_10_sub_pixel_variance64x64,
1224 vpx_highbd_10_sub_pixel_avg_variance64x64,
1225 vpx_highbd_sad64x64x3_bits10,
1226 vpx_highbd_sad64x64x8_bits10,
1227 vpx_highbd_sad64x64x4d_bits10)
1229 HIGHBD_BFP(BLOCK_16X16,
1230 vpx_highbd_sad16x16_bits10,
1231 vpx_highbd_sad16x16_avg_bits10,
1232 vpx_highbd_10_variance16x16,
1233 vpx_highbd_10_sub_pixel_variance16x16,
1234 vpx_highbd_10_sub_pixel_avg_variance16x16,
1235 vpx_highbd_sad16x16x3_bits10,
1236 vpx_highbd_sad16x16x8_bits10,
1237 vpx_highbd_sad16x16x4d_bits10)
1239 HIGHBD_BFP(BLOCK_16X8,
1240 vpx_highbd_sad16x8_bits10,
1241 vpx_highbd_sad16x8_avg_bits10,
1242 vpx_highbd_10_variance16x8,
1243 vpx_highbd_10_sub_pixel_variance16x8,
1244 vpx_highbd_10_sub_pixel_avg_variance16x8,
1245 vpx_highbd_sad16x8x3_bits10,
1246 vpx_highbd_sad16x8x8_bits10,
1247 vpx_highbd_sad16x8x4d_bits10)
1249 HIGHBD_BFP(BLOCK_8X16,
1250 vpx_highbd_sad8x16_bits10,
1251 vpx_highbd_sad8x16_avg_bits10,
1252 vpx_highbd_10_variance8x16,
1253 vpx_highbd_10_sub_pixel_variance8x16,
1254 vpx_highbd_10_sub_pixel_avg_variance8x16,
1255 vpx_highbd_sad8x16x3_bits10,
1256 vpx_highbd_sad8x16x8_bits10,
1257 vpx_highbd_sad8x16x4d_bits10)
1259 HIGHBD_BFP(BLOCK_8X8,
1260 vpx_highbd_sad8x8_bits10,
1261 vpx_highbd_sad8x8_avg_bits10,
1262 vpx_highbd_10_variance8x8,
1263 vpx_highbd_10_sub_pixel_variance8x8,
1264 vpx_highbd_10_sub_pixel_avg_variance8x8,
1265 vpx_highbd_sad8x8x3_bits10,
1266 vpx_highbd_sad8x8x8_bits10,
1267 vpx_highbd_sad8x8x4d_bits10)
1269 HIGHBD_BFP(BLOCK_8X4,
1270 vpx_highbd_sad8x4_bits10,
1271 vpx_highbd_sad8x4_avg_bits10,
1272 vpx_highbd_10_variance8x4,
1273 vpx_highbd_10_sub_pixel_variance8x4,
1274 vpx_highbd_10_sub_pixel_avg_variance8x4,
1276 vpx_highbd_sad8x4x8_bits10,
1277 vpx_highbd_sad8x4x4d_bits10)
1279 HIGHBD_BFP(BLOCK_4X8,
1280 vpx_highbd_sad4x8_bits10,
1281 vpx_highbd_sad4x8_avg_bits10,
1282 vpx_highbd_10_variance4x8,
1283 vpx_highbd_10_sub_pixel_variance4x8,
1284 vpx_highbd_10_sub_pixel_avg_variance4x8,
1286 vpx_highbd_sad4x8x8_bits10,
1287 vpx_highbd_sad4x8x4d_bits10)
1289 HIGHBD_BFP(BLOCK_4X4,
1290 vpx_highbd_sad4x4_bits10,
1291 vpx_highbd_sad4x4_avg_bits10,
1292 vpx_highbd_10_variance4x4,
1293 vpx_highbd_10_sub_pixel_variance4x4,
1294 vpx_highbd_10_sub_pixel_avg_variance4x4,
1295 vpx_highbd_sad4x4x3_bits10,
1296 vpx_highbd_sad4x4x8_bits10,
1297 vpx_highbd_sad4x4x4d_bits10)
1301 HIGHBD_BFP(BLOCK_32X16,
1302 vpx_highbd_sad32x16_bits12,
1303 vpx_highbd_sad32x16_avg_bits12,
1304 vpx_highbd_12_variance32x16,
1305 vpx_highbd_12_sub_pixel_variance32x16,
1306 vpx_highbd_12_sub_pixel_avg_variance32x16,
1309 vpx_highbd_sad32x16x4d_bits12)
1311 HIGHBD_BFP(BLOCK_16X32,
1312 vpx_highbd_sad16x32_bits12,
1313 vpx_highbd_sad16x32_avg_bits12,
1314 vpx_highbd_12_variance16x32,
1315 vpx_highbd_12_sub_pixel_variance16x32,
1316 vpx_highbd_12_sub_pixel_avg_variance16x32,
1319 vpx_highbd_sad16x32x4d_bits12)
1321 HIGHBD_BFP(BLOCK_64X32,
1322 vpx_highbd_sad64x32_bits12,
1323 vpx_highbd_sad64x32_avg_bits12,
1324 vpx_highbd_12_variance64x32,
1325 vpx_highbd_12_sub_pixel_variance64x32,
1326 vpx_highbd_12_sub_pixel_avg_variance64x32,
1329 vpx_highbd_sad64x32x4d_bits12)
1331 HIGHBD_BFP(BLOCK_32X64,
1332 vpx_highbd_sad32x64_bits12,
1333 vpx_highbd_sad32x64_avg_bits12,
1334 vpx_highbd_12_variance32x64,
1335 vpx_highbd_12_sub_pixel_variance32x64,
1336 vpx_highbd_12_sub_pixel_avg_variance32x64,
1339 vpx_highbd_sad32x64x4d_bits12)
1341 HIGHBD_BFP(BLOCK_32X32,
1342 vpx_highbd_sad32x32_bits12,
1343 vpx_highbd_sad32x32_avg_bits12,
1344 vpx_highbd_12_variance32x32,
1345 vpx_highbd_12_sub_pixel_variance32x32,
1346 vpx_highbd_12_sub_pixel_avg_variance32x32,
1347 vpx_highbd_sad32x32x3_bits12,
1348 vpx_highbd_sad32x32x8_bits12,
1349 vpx_highbd_sad32x32x4d_bits12)
1351 HIGHBD_BFP(BLOCK_64X64,
1352 vpx_highbd_sad64x64_bits12,
1353 vpx_highbd_sad64x64_avg_bits12,
1354 vpx_highbd_12_variance64x64,
1355 vpx_highbd_12_sub_pixel_variance64x64,
1356 vpx_highbd_12_sub_pixel_avg_variance64x64,
1357 vpx_highbd_sad64x64x3_bits12,
1358 vpx_highbd_sad64x64x8_bits12,
1359 vpx_highbd_sad64x64x4d_bits12)
1361 HIGHBD_BFP(BLOCK_16X16,
1362 vpx_highbd_sad16x16_bits12,
1363 vpx_highbd_sad16x16_avg_bits12,
1364 vpx_highbd_12_variance16x16,
1365 vpx_highbd_12_sub_pixel_variance16x16,
1366 vpx_highbd_12_sub_pixel_avg_variance16x16,
1367 vpx_highbd_sad16x16x3_bits12,
1368 vpx_highbd_sad16x16x8_bits12,
1369 vpx_highbd_sad16x16x4d_bits12)
1371 HIGHBD_BFP(BLOCK_16X8,
1372 vpx_highbd_sad16x8_bits12,
1373 vpx_highbd_sad16x8_avg_bits12,
1374 vpx_highbd_12_variance16x8,
1375 vpx_highbd_12_sub_pixel_variance16x8,
1376 vpx_highbd_12_sub_pixel_avg_variance16x8,
1377 vpx_highbd_sad16x8x3_bits12,
1378 vpx_highbd_sad16x8x8_bits12,
1379 vpx_highbd_sad16x8x4d_bits12)
1381 HIGHBD_BFP(BLOCK_8X16,
1382 vpx_highbd_sad8x16_bits12,
1383 vpx_highbd_sad8x16_avg_bits12,
1384 vpx_highbd_12_variance8x16,
1385 vpx_highbd_12_sub_pixel_variance8x16,
1386 vpx_highbd_12_sub_pixel_avg_variance8x16,
1387 vpx_highbd_sad8x16x3_bits12,
1388 vpx_highbd_sad8x16x8_bits12,
1389 vpx_highbd_sad8x16x4d_bits12)
1391 HIGHBD_BFP(BLOCK_8X8,
1392 vpx_highbd_sad8x8_bits12,
1393 vpx_highbd_sad8x8_avg_bits12,
1394 vpx_highbd_12_variance8x8,
1395 vpx_highbd_12_sub_pixel_variance8x8,
1396 vpx_highbd_12_sub_pixel_avg_variance8x8,
1397 vpx_highbd_sad8x8x3_bits12,
1398 vpx_highbd_sad8x8x8_bits12,
1399 vpx_highbd_sad8x8x4d_bits12)
1401 HIGHBD_BFP(BLOCK_8X4,
1402 vpx_highbd_sad8x4_bits12,
1403 vpx_highbd_sad8x4_avg_bits12,
1404 vpx_highbd_12_variance8x4,
1405 vpx_highbd_12_sub_pixel_variance8x4,
1406 vpx_highbd_12_sub_pixel_avg_variance8x4,
1408 vpx_highbd_sad8x4x8_bits12,
1409 vpx_highbd_sad8x4x4d_bits12)
1411 HIGHBD_BFP(BLOCK_4X8,
1412 vpx_highbd_sad4x8_bits12,
1413 vpx_highbd_sad4x8_avg_bits12,
1414 vpx_highbd_12_variance4x8,
1415 vpx_highbd_12_sub_pixel_variance4x8,
1416 vpx_highbd_12_sub_pixel_avg_variance4x8,
1418 vpx_highbd_sad4x8x8_bits12,
1419 vpx_highbd_sad4x8x4d_bits12)
1421 HIGHBD_BFP(BLOCK_4X4,
1422 vpx_highbd_sad4x4_bits12,
1423 vpx_highbd_sad4x4_avg_bits12,
1424 vpx_highbd_12_variance4x4,
1425 vpx_highbd_12_sub_pixel_variance4x4,
1426 vpx_highbd_12_sub_pixel_avg_variance4x4,
1427 vpx_highbd_sad4x4x3_bits12,
1428 vpx_highbd_sad4x4x8_bits12,
1429 vpx_highbd_sad4x4x4d_bits12)
1433 assert(0 && "cm->bit_depth should be VPX_BITS_8, "
1434 "VPX_BITS_10 or VPX_BITS_12");
1438 #endif // CONFIG_VP9_HIGHBITDEPTH
1440 static void realloc_segmentation_maps(VP9_COMP *cpi) {
1441 VP9_COMMON *const cm = &cpi->common;
1443 // Create the encoder segmentation map and set all entries to 0
1444 vpx_free(cpi->segmentation_map);
1445 CHECK_MEM_ERROR(cm, cpi->segmentation_map,
1446 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1448 // Create a map used for cyclic background refresh.
1449 if (cpi->cyclic_refresh)
1450 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
1451 CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
1452 vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
1454 // Create a map used to mark inactive areas.
1455 vpx_free(cpi->active_map.map);
1456 CHECK_MEM_ERROR(cm, cpi->active_map.map,
1457 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1459 // And a place holder structure is the coding context
1460 // for use if we want to save and restore it
1461 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
1462 CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
1463 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1466 void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
1467 VP9_COMMON *const cm = &cpi->common;
1468 RATE_CONTROL *const rc = &cpi->rc;
1469 int last_w = cpi->oxcf.width;
1470 int last_h = cpi->oxcf.height;
1472 if (cm->profile != oxcf->profile)
1473 cm->profile = oxcf->profile;
1474 cm->bit_depth = oxcf->bit_depth;
1475 cm->color_space = oxcf->color_space;
1476 cm->color_range = oxcf->color_range;
1478 cpi->target_level = oxcf->target_level;
1479 cm->keep_level_stats = oxcf->target_level != LEVEL_NOT_CARE;
1481 if (cm->profile <= PROFILE_1)
1482 assert(cm->bit_depth == VPX_BITS_8);
1484 assert(cm->bit_depth > VPX_BITS_8);
1487 #if CONFIG_VP9_HIGHBITDEPTH
1488 cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
1489 #endif // CONFIG_VP9_HIGHBITDEPTH
1491 if ((oxcf->pass == 0) && (oxcf->rc_mode == VPX_Q)) {
1492 rc->baseline_gf_interval = FIXED_GF_INTERVAL;
1494 rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
1497 cpi->refresh_golden_frame = 0;
1498 cpi->refresh_last_frame = 1;
1499 cm->refresh_frame_context = 1;
1500 cm->reset_frame_context = 0;
1502 vp9_reset_segment_features(&cm->seg);
1503 vp9_set_high_precision_mv(cpi, 0);
1508 for (i = 0; i < MAX_SEGMENTS; i++)
1509 cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
1511 cpi->encode_breakout = cpi->oxcf.encode_breakout;
1513 set_rc_buffer_sizes(rc, &cpi->oxcf);
1515 // Under a configuration change, where maximum_buffer_size may change,
1516 // keep buffer level clipped to the maximum allowed buffer size.
1517 rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
1518 rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size);
1520 // Set up frame rate and related parameters rate control values.
1521 vp9_new_framerate(cpi, cpi->framerate);
1523 // Set absolute upper and lower quality limits
1524 rc->worst_quality = cpi->oxcf.worst_allowed_q;
1525 rc->best_quality = cpi->oxcf.best_allowed_q;
1527 cm->interp_filter = cpi->sf.default_interp_filter;
1529 if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) {
1530 cm->render_width = cpi->oxcf.render_width;
1531 cm->render_height = cpi->oxcf.render_height;
1533 cm->render_width = cpi->oxcf.width;
1534 cm->render_height = cpi->oxcf.height;
1536 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1537 cm->width = cpi->oxcf.width;
1538 cm->height = cpi->oxcf.height;
1539 cpi->external_resize = 1;
1542 if (cpi->initial_width) {
1543 int new_mi_size = 0;
1544 vp9_set_mb_mi(cm, cm->width, cm->height);
1545 new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
1546 if (cm->mi_alloc_size < new_mi_size) {
1547 vp9_free_context_buffers(cm);
1548 alloc_compressor_data(cpi);
1549 realloc_segmentation_maps(cpi);
1550 cpi->initial_width = cpi->initial_height = 0;
1551 cpi->external_resize = 0;
1552 } else if (cm->mi_alloc_size == new_mi_size &&
1553 (cpi->oxcf.width > last_w || cpi->oxcf.height > last_h)) {
1554 vp9_alloc_loop_filter(cm);
1558 update_frame_size(cpi);
1560 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1561 memset(cpi->consec_zero_mv, 0,
1562 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
1563 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
1564 vp9_cyclic_refresh_reset_resize(cpi);
1567 if ((cpi->svc.number_temporal_layers > 1 &&
1568 cpi->oxcf.rc_mode == VPX_CBR) ||
1569 ((cpi->svc.number_temporal_layers > 1 ||
1570 cpi->svc.number_spatial_layers > 1) &&
1571 cpi->oxcf.pass != 1)) {
1572 vp9_update_layer_context_change_config(cpi,
1573 (int)cpi->oxcf.target_bandwidth);
1576 cpi->alt_ref_source = NULL;
1577 rc->is_src_frame_alt_ref = 0;
1580 // Experimental RD Code
1581 cpi->frame_distortion = 0;
1582 cpi->last_frame_distortion = 0;
1585 set_tile_limits(cpi);
1587 cpi->ext_refresh_frame_flags_pending = 0;
1588 cpi->ext_refresh_frame_context_pending = 0;
1590 #if CONFIG_VP9_HIGHBITDEPTH
1591 highbd_set_var_fns(cpi);
1596 #define M_LOG2_E 0.693147180559945309417
1598 #define log2f(x) (log (x) / (float) M_LOG2_E)
1600 /***********************************************************************
1601 * Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts' *
1602 ***********************************************************************
1603 * The following 2 functions ('cal_nmvjointsadcost' and *
1604 * 'cal_nmvsadcosts') are used to calculate cost lookup tables *
1605 * used by 'vp9_diamond_search_sad'. The C implementation of the *
1606 * function is generic, but the AVX intrinsics optimised version *
1607 * relies on the following properties of the computed tables: *
1608 * For cal_nmvjointsadcost: *
1609 * - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] *
1610 * For cal_nmvsadcosts: *
1611 * - For all i: mvsadcost[0][i] == mvsadcost[1][i] *
1612 * (Equal costs for both components) *
1613 * - For all i: mvsadcost[0][i] == mvsadcost[0][-i] *
1614 * (Cost function is even) *
1615 * If these do not hold, then the AVX optimised version of the *
1616 * 'vp9_diamond_search_sad' function cannot be used as it is, in which *
1617 * case you can revert to using the C function instead. *
1618 ***********************************************************************/
1620 static void cal_nmvjointsadcost(int *mvjointsadcost) {
1621 /*********************************************************************
1622 * Warning: Read the comments above before modifying this function *
1623 *********************************************************************/
1624 mvjointsadcost[0] = 600;
1625 mvjointsadcost[1] = 300;
1626 mvjointsadcost[2] = 300;
1627 mvjointsadcost[3] = 300;
1630 static void cal_nmvsadcosts(int *mvsadcost[2]) {
1631 /*********************************************************************
1632 * Warning: Read the comments above before modifying this function *
1633 *********************************************************************/
1636 mvsadcost[0][0] = 0;
1637 mvsadcost[1][0] = 0;
1640 double z = 256 * (2 * (log2f(8 * i) + .6));
1641 mvsadcost[0][i] = (int)z;
1642 mvsadcost[1][i] = (int)z;
1643 mvsadcost[0][-i] = (int)z;
1644 mvsadcost[1][-i] = (int)z;
1645 } while (++i <= MV_MAX);
1648 static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
1651 mvsadcost[0][0] = 0;
1652 mvsadcost[1][0] = 0;
1655 double z = 256 * (2 * (log2f(8 * i) + .6));
1656 mvsadcost[0][i] = (int)z;
1657 mvsadcost[1][i] = (int)z;
1658 mvsadcost[0][-i] = (int)z;
1659 mvsadcost[1][-i] = (int)z;
1660 } while (++i <= MV_MAX);
1664 VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
1665 BufferPool *const pool) {
1667 VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
1668 VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
1675 if (setjmp(cm->error.jmp)) {
1676 cm->error.setjmp = 0;
1677 vp9_remove_compressor(cpi);
1681 cm->error.setjmp = 1;
1682 cm->alloc_mi = vp9_enc_alloc_mi;
1683 cm->free_mi = vp9_enc_free_mi;
1684 cm->setup_mi = vp9_enc_setup_mi;
1686 CHECK_MEM_ERROR(cm, cm->fc,
1687 (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
1688 CHECK_MEM_ERROR(cm, cm->frame_contexts,
1689 (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
1690 sizeof(*cm->frame_contexts)));
1693 cpi->resize_state = 0;
1694 cpi->external_resize = 0;
1695 cpi->resize_avg_qp = 0;
1696 cpi->resize_buffer_underflow = 0;
1697 cpi->use_skin_detection = 0;
1698 cpi->common.buffer_pool = pool;
1700 init_config(cpi, oxcf);
1701 vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
1703 cm->current_video_frame = 0;
1704 cpi->partition_search_skippable_frame = 0;
1705 cpi->tile_data = NULL;
1707 realloc_segmentation_maps(cpi);
1709 CHECK_MEM_ERROR(cm, cpi->consec_zero_mv,
1710 vpx_calloc(cm->mi_rows * cm->mi_cols,
1711 sizeof(*cpi->consec_zero_mv)));
1713 CHECK_MEM_ERROR(cm, cpi->nmvcosts[0],
1714 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0])));
1715 CHECK_MEM_ERROR(cm, cpi->nmvcosts[1],
1716 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1])));
1717 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0],
1718 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0])));
1719 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1],
1720 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1])));
1721 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0],
1722 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0])));
1723 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1],
1724 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1])));
1725 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0],
1726 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0])));
1727 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
1728 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
1730 for (i = 0; i < (sizeof(cpi->mbgraph_stats) /
1731 sizeof(cpi->mbgraph_stats[0])); i++) {
1732 CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats,
1733 vpx_calloc(cm->MBs *
1734 sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
1737 #if CONFIG_FP_MB_STATS
1738 cpi->use_fp_mb_stats = 0;
1739 if (cpi->use_fp_mb_stats) {
1740 // a place holder used to store the first pass mb stats in the first pass
1741 CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
1742 vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
1744 cpi->twopass.frame_mb_stats_buf = NULL;
1748 cpi->refresh_alt_ref_frame = 0;
1749 cpi->multi_arf_last_grp_enabled = 0;
1751 cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
1752 #if CONFIG_INTERNAL_STATS
1753 cpi->b_calculate_ssimg = 0;
1754 cpi->b_calculate_blockiness = 1;
1755 cpi->b_calculate_consistency = 1;
1756 cpi->total_inconsistency = 0;
1757 cpi->psnr.worst = 100.0;
1758 cpi->worst_ssim = 100.0;
1763 if (cpi->b_calculate_psnr) {
1764 cpi->total_sq_error = 0;
1765 cpi->total_samples = 0;
1767 cpi->totalp_sq_error = 0;
1768 cpi->totalp_samples = 0;
1770 cpi->tot_recode_hits = 0;
1771 cpi->summed_quality = 0;
1772 cpi->summed_weights = 0;
1773 cpi->summedp_quality = 0;
1774 cpi->summedp_weights = 0;
1777 if (cpi->b_calculate_ssimg) {
1778 cpi->ssimg.worst= 100.0;
1780 cpi->fastssim.worst = 100.0;
1782 cpi->psnrhvs.worst = 100.0;
1784 if (cpi->b_calculate_blockiness) {
1785 cpi->total_blockiness = 0;
1786 cpi->worst_blockiness = 0.0;
1789 if (cpi->b_calculate_consistency) {
1790 CHECK_MEM_ERROR(cm, cpi->ssim_vars,
1791 vpx_malloc(sizeof(*cpi->ssim_vars) * 4 *
1792 cpi->common.mi_rows * cpi->common.mi_cols));
1793 cpi->worst_consistency = 100.0;
1798 cpi->first_time_stamp_ever = INT64_MAX;
1800 /*********************************************************************
1801 * Warning: Read the comments around 'cal_nmvjointsadcost' and *
1802 * 'cal_nmvsadcosts' before modifying how these tables are computed. *
1803 *********************************************************************/
1804 cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
1805 cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
1806 cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
1807 cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
1808 cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
1809 cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
1811 cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
1812 cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
1813 cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
1814 cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
1815 cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
1817 #if CONFIG_VP9_TEMPORAL_DENOISING
1818 #ifdef OUTPUT_YUV_DENOISED
1819 yuv_denoised_file = fopen("denoised.yuv", "ab");
1822 #ifdef OUTPUT_YUV_SKINMAP
1823 yuv_skinmap_file = fopen("skinmap.yuv", "ab");
1825 #ifdef OUTPUT_YUV_REC
1826 yuv_rec_file = fopen("rec.yuv", "wb");
1830 framepsnr = fopen("framepsnr.stt", "a");
1831 kf_list = fopen("kf_list.stt", "w");
1834 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
1836 if (oxcf->pass == 1) {
1837 vp9_init_first_pass(cpi);
1838 } else if (oxcf->pass == 2) {
1839 const size_t packet_sz = sizeof(FIRSTPASS_STATS);
1840 const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
1842 if (cpi->svc.number_spatial_layers > 1
1843 || cpi->svc.number_temporal_layers > 1) {
1844 FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf;
1845 FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = {0};
1848 for (i = 0; i < oxcf->ss_number_layers; ++i) {
1849 FIRSTPASS_STATS *const last_packet_for_layer =
1850 &stats[packets - oxcf->ss_number_layers + i];
1851 const int layer_id = (int)last_packet_for_layer->spatial_layer_id;
1852 const int packets_in_layer = (int)last_packet_for_layer->count + 1;
1853 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) {
1854 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id];
1856 vpx_free(lc->rc_twopass_stats_in.buf);
1858 lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz;
1859 CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf,
1860 vpx_malloc(lc->rc_twopass_stats_in.sz));
1861 lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf;
1862 lc->twopass.stats_in = lc->twopass.stats_in_start;
1863 lc->twopass.stats_in_end = lc->twopass.stats_in_start
1864 + packets_in_layer - 1;
1865 stats_copy[layer_id] = lc->rc_twopass_stats_in.buf;
1869 for (i = 0; i < packets; ++i) {
1870 const int layer_id = (int)stats[i].spatial_layer_id;
1871 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers
1872 && stats_copy[layer_id] != NULL) {
1873 *stats_copy[layer_id] = stats[i];
1874 ++stats_copy[layer_id];
1878 vp9_init_second_pass_spatial_svc(cpi);
1880 #if CONFIG_FP_MB_STATS
1881 if (cpi->use_fp_mb_stats) {
1882 const size_t psz = cpi->common.MBs * sizeof(uint8_t);
1883 const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
1885 cpi->twopass.firstpass_mb_stats.mb_stats_start =
1886 oxcf->firstpass_mb_stats_in.buf;
1887 cpi->twopass.firstpass_mb_stats.mb_stats_end =
1888 cpi->twopass.firstpass_mb_stats.mb_stats_start +
1889 (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
1893 cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
1894 cpi->twopass.stats_in = cpi->twopass.stats_in_start;
1895 cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
1897 vp9_init_second_pass(cpi);
1901 vp9_set_speed_features_framesize_independent(cpi);
1902 vp9_set_speed_features_framesize_dependent(cpi);
1904 // Allocate memory to store variances for a frame.
1905 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
1906 vpx_calloc(cm->MBs, sizeof(diff)));
1907 cpi->source_var_thresh = 0;
1908 cpi->frames_till_next_var_check = 0;
1910 #define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\
1911 cpi->fn_ptr[BT].sdf = SDF; \
1912 cpi->fn_ptr[BT].sdaf = SDAF; \
1913 cpi->fn_ptr[BT].vf = VF; \
1914 cpi->fn_ptr[BT].svf = SVF; \
1915 cpi->fn_ptr[BT].svaf = SVAF; \
1916 cpi->fn_ptr[BT].sdx3f = SDX3F; \
1917 cpi->fn_ptr[BT].sdx8f = SDX8F; \
1918 cpi->fn_ptr[BT].sdx4df = SDX4DF;
1920 BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg,
1921 vpx_variance32x16, vpx_sub_pixel_variance32x16,
1922 vpx_sub_pixel_avg_variance32x16, NULL, NULL, vpx_sad32x16x4d)
1924 BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg,
1925 vpx_variance16x32, vpx_sub_pixel_variance16x32,
1926 vpx_sub_pixel_avg_variance16x32, NULL, NULL, vpx_sad16x32x4d)
1928 BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg,
1929 vpx_variance64x32, vpx_sub_pixel_variance64x32,
1930 vpx_sub_pixel_avg_variance64x32, NULL, NULL, vpx_sad64x32x4d)
1932 BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg,
1933 vpx_variance32x64, vpx_sub_pixel_variance32x64,
1934 vpx_sub_pixel_avg_variance32x64, NULL, NULL, vpx_sad32x64x4d)
1936 BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg,
1937 vpx_variance32x32, vpx_sub_pixel_variance32x32,
1938 vpx_sub_pixel_avg_variance32x32, vpx_sad32x32x3, vpx_sad32x32x8,
1941 BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg,
1942 vpx_variance64x64, vpx_sub_pixel_variance64x64,
1943 vpx_sub_pixel_avg_variance64x64, vpx_sad64x64x3, vpx_sad64x64x8,
1946 BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg,
1947 vpx_variance16x16, vpx_sub_pixel_variance16x16,
1948 vpx_sub_pixel_avg_variance16x16, vpx_sad16x16x3, vpx_sad16x16x8,
1951 BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg,
1952 vpx_variance16x8, vpx_sub_pixel_variance16x8,
1953 vpx_sub_pixel_avg_variance16x8,
1954 vpx_sad16x8x3, vpx_sad16x8x8, vpx_sad16x8x4d)
1956 BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg,
1957 vpx_variance8x16, vpx_sub_pixel_variance8x16,
1958 vpx_sub_pixel_avg_variance8x16,
1959 vpx_sad8x16x3, vpx_sad8x16x8, vpx_sad8x16x4d)
1961 BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg,
1962 vpx_variance8x8, vpx_sub_pixel_variance8x8,
1963 vpx_sub_pixel_avg_variance8x8,
1964 vpx_sad8x8x3, vpx_sad8x8x8, vpx_sad8x8x4d)
1966 BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg,
1967 vpx_variance8x4, vpx_sub_pixel_variance8x4,
1968 vpx_sub_pixel_avg_variance8x4, NULL, vpx_sad8x4x8, vpx_sad8x4x4d)
1970 BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg,
1971 vpx_variance4x8, vpx_sub_pixel_variance4x8,
1972 vpx_sub_pixel_avg_variance4x8, NULL, vpx_sad4x8x8, vpx_sad4x8x4d)
1974 BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg,
1975 vpx_variance4x4, vpx_sub_pixel_variance4x4,
1976 vpx_sub_pixel_avg_variance4x4,
1977 vpx_sad4x4x3, vpx_sad4x4x8, vpx_sad4x4x4d)
1979 #if CONFIG_VP9_HIGHBITDEPTH
1980 highbd_set_var_fns(cpi);
1983 /* vp9_init_quantizer() is first called here. Add check in
1984 * vp9_frame_init_quantizer() so that vp9_init_quantizer is only
1985 * called later when needed. This will avoid unnecessary calls of
1986 * vp9_init_quantizer() for every frame.
1988 vp9_init_quantizer(cpi);
1990 vp9_loop_filter_init(cm);
1992 cm->error.setjmp = 0;
1997 #if CONFIG_INTERNAL_STATS
1998 #define SNPRINT(H, T) \
1999 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
2001 #define SNPRINT2(H, T, V) \
2002 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
2003 #endif // CONFIG_INTERNAL_STATS
2005 void vp9_remove_compressor(VP9_COMP *cpi) {
2014 if (cm->current_video_frame > 0) {
2015 #if CONFIG_INTERNAL_STATS
2016 vpx_clear_system_state();
2018 if (cpi->oxcf.pass != 1) {
2019 char headings[512] = {0};
2020 char results[512] = {0};
2021 FILE *f = fopen("opsnr.stt", "a");
2022 double time_encoded = (cpi->last_end_time_stamp_seen
2023 - cpi->first_time_stamp_ever) / 10000000.000;
2024 double total_encode_time = (cpi->time_receive_data +
2025 cpi->time_compress_data) / 1000.000;
2027 (double)cpi->bytes * (double) 8 / (double)1000 / time_encoded;
2028 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
2029 const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
2030 const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
2032 if (cpi->b_calculate_psnr) {
2033 const double total_psnr =
2034 vpx_sse_to_psnr((double)cpi->total_samples, peak,
2035 (double)cpi->total_sq_error);
2036 const double totalp_psnr =
2037 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
2038 (double)cpi->totalp_sq_error);
2039 const double total_ssim = 100 * pow(cpi->summed_quality /
2040 cpi->summed_weights, 8.0);
2041 const double totalp_ssim = 100 * pow(cpi->summedp_quality /
2042 cpi->summedp_weights, 8.0);
2044 snprintf(headings, sizeof(headings),
2045 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
2046 "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
2047 "WstPsnr\tWstSsim\tWstFast\tWstHVS");
2048 snprintf(results, sizeof(results),
2049 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2050 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2051 "%7.3f\t%7.3f\t%7.3f\t%7.3f",
2052 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
2053 cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr,
2054 total_ssim, totalp_ssim,
2055 cpi->fastssim.stat[ALL] / cpi->count,
2056 cpi->psnrhvs.stat[ALL] / cpi->count,
2057 cpi->psnr.worst, cpi->worst_ssim, cpi->fastssim.worst,
2058 cpi->psnrhvs.worst);
2060 if (cpi->b_calculate_blockiness) {
2061 SNPRINT(headings, "\t Block\tWstBlck");
2062 SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
2063 SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
2066 if (cpi->b_calculate_consistency) {
2067 double consistency =
2068 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
2069 (double)cpi->total_inconsistency);
2071 SNPRINT(headings, "\tConsist\tWstCons");
2072 SNPRINT2(results, "\t%7.3f", consistency);
2073 SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
2076 if (cpi->b_calculate_ssimg) {
2077 SNPRINT(headings, "\t SSIMG\tWtSSIMG");
2078 SNPRINT2(results, "\t%7.3f", cpi->ssimg.stat[ALL] / cpi->count);
2079 SNPRINT2(results, "\t%7.3f", cpi->ssimg.worst);
2082 fprintf(f, "%s\t Time Rc-Err Abs Err\n", headings);
2083 fprintf(f, "%s\t%8.0f %7.2f %7.2f\n", results,
2084 total_encode_time, rate_err, fabs(rate_err));
2094 printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
2095 printf("\n_frames recive_data encod_mb_row compress_frame Total\n");
2096 printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
2097 cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
2098 cpi->time_compress_data / 1000,
2099 (cpi->time_receive_data + cpi->time_compress_data) / 1000);
2104 #if CONFIG_VP9_TEMPORAL_DENOISING
2105 vp9_denoiser_free(&(cpi->denoiser));
2108 for (t = 0; t < cpi->num_workers; ++t) {
2109 VPxWorker *const worker = &cpi->workers[t];
2110 EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
2112 // Deallocate allocated threads.
2113 vpx_get_worker_interface()->end(worker);
2115 // Deallocate allocated thread data.
2116 if (t < cpi->num_workers - 1) {
2117 vpx_free(thread_data->td->counts);
2118 vp9_free_pc_tree(thread_data->td);
2119 vpx_free(thread_data->td);
2122 vpx_free(cpi->tile_thr_data);
2123 vpx_free(cpi->workers);
2125 if (cpi->num_workers > 1)
2126 vp9_loop_filter_dealloc(&cpi->lf_row_sync);
2128 dealloc_compressor_data(cpi);
2130 for (i = 0; i < sizeof(cpi->mbgraph_stats) /
2131 sizeof(cpi->mbgraph_stats[0]); ++i) {
2132 vpx_free(cpi->mbgraph_stats[i].mb_stats);
2135 #if CONFIG_FP_MB_STATS
2136 if (cpi->use_fp_mb_stats) {
2137 vpx_free(cpi->twopass.frame_mb_stats_buf);
2138 cpi->twopass.frame_mb_stats_buf = NULL;
2142 vp9_remove_common(cm);
2143 vp9_free_ref_frame_buffers(cm->buffer_pool);
2144 #if CONFIG_VP9_POSTPROC
2145 vp9_free_postproc_buffers(cm);
2149 #if CONFIG_VP9_TEMPORAL_DENOISING
2150 #ifdef OUTPUT_YUV_DENOISED
2151 fclose(yuv_denoised_file);
2154 #ifdef OUTPUT_YUV_SKINMAP
2155 fclose(yuv_skinmap_file);
2157 #ifdef OUTPUT_YUV_REC
2158 fclose(yuv_rec_file);
2175 /* TODO(yaowu): The block_variance calls the unoptimized versions of variance()
2176 * and highbd_8_variance(). It should not.
2178 static void encoder_variance(const uint8_t *a, int a_stride,
2179 const uint8_t *b, int b_stride,
2180 int w, int h, unsigned int *sse, int *sum) {
2186 for (i = 0; i < h; i++) {
2187 for (j = 0; j < w; j++) {
2188 const int diff = a[j] - b[j];
2190 *sse += diff * diff;
2198 #if CONFIG_VP9_HIGHBITDEPTH
2199 static void encoder_highbd_variance64(const uint8_t *a8, int a_stride,
2200 const uint8_t *b8, int b_stride,
2201 int w, int h, uint64_t *sse,
2205 uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2206 uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2210 for (i = 0; i < h; i++) {
2211 for (j = 0; j < w; j++) {
2212 const int diff = a[j] - b[j];
2214 *sse += diff * diff;
2221 static void encoder_highbd_8_variance(const uint8_t *a8, int a_stride,
2222 const uint8_t *b8, int b_stride,
2224 unsigned int *sse, int *sum) {
2225 uint64_t sse_long = 0;
2226 uint64_t sum_long = 0;
2227 encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h,
2228 &sse_long, &sum_long);
2229 *sse = (unsigned int)sse_long;
2230 *sum = (int)sum_long;
2232 #endif // CONFIG_VP9_HIGHBITDEPTH
2234 static int64_t get_sse(const uint8_t *a, int a_stride,
2235 const uint8_t *b, int b_stride,
2236 int width, int height) {
2237 const int dw = width % 16;
2238 const int dh = height % 16;
2239 int64_t total_sse = 0;
2240 unsigned int sse = 0;
2245 encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride,
2246 dw, height, &sse, &sum);
2251 encoder_variance(&a[(height - dh) * a_stride], a_stride,
2252 &b[(height - dh) * b_stride], b_stride,
2253 width - dw, dh, &sse, &sum);
2257 for (y = 0; y < height / 16; ++y) {
2258 const uint8_t *pa = a;
2259 const uint8_t *pb = b;
2260 for (x = 0; x < width / 16; ++x) {
2261 vpx_mse16x16(pa, a_stride, pb, b_stride, &sse);
2275 #if CONFIG_VP9_HIGHBITDEPTH
2276 static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride,
2277 const uint8_t *b8, int b_stride,
2278 int width, int height,
2279 unsigned int input_shift) {
2280 const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2281 const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2282 int64_t total_sse = 0;
2284 for (y = 0; y < height; ++y) {
2285 for (x = 0; x < width; ++x) {
2287 diff = (a[x] >> input_shift) - (b[x] >> input_shift);
2288 total_sse += diff * diff;
2296 static int64_t highbd_get_sse(const uint8_t *a, int a_stride,
2297 const uint8_t *b, int b_stride,
2298 int width, int height) {
2299 int64_t total_sse = 0;
2301 const int dw = width % 16;
2302 const int dh = height % 16;
2303 unsigned int sse = 0;
2306 encoder_highbd_8_variance(&a[width - dw], a_stride,
2307 &b[width - dw], b_stride,
2308 dw, height, &sse, &sum);
2312 encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride,
2313 &b[(height - dh) * b_stride], b_stride,
2314 width - dw, dh, &sse, &sum);
2317 for (y = 0; y < height / 16; ++y) {
2318 const uint8_t *pa = a;
2319 const uint8_t *pb = b;
2320 for (x = 0; x < width / 16; ++x) {
2321 vpx_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse);
2331 #endif // CONFIG_VP9_HIGHBITDEPTH
2334 double psnr[4]; // total/y/u/v
2335 uint64_t sse[4]; // total/y/u/v
2336 uint32_t samples[4]; // total/y/u/v
2339 #if CONFIG_VP9_HIGHBITDEPTH
2340 static void calc_highbd_psnr(const YV12_BUFFER_CONFIG *a,
2341 const YV12_BUFFER_CONFIG *b,
2343 unsigned int bit_depth,
2344 unsigned int in_bit_depth) {
2345 const int widths[3] =
2346 {a->y_crop_width, a->uv_crop_width, a->uv_crop_width };
2347 const int heights[3] =
2348 {a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2349 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer };
2350 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2351 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer };
2352 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2354 uint64_t total_sse = 0;
2355 uint32_t total_samples = 0;
2356 const double peak = (double)((1 << in_bit_depth) - 1);
2357 const unsigned int input_shift = bit_depth - in_bit_depth;
2359 for (i = 0; i < 3; ++i) {
2360 const int w = widths[i];
2361 const int h = heights[i];
2362 const uint32_t samples = w * h;
2364 if (a->flags & YV12_FLAG_HIGHBITDEPTH) {
2366 sse = highbd_get_sse_shift(a_planes[i], a_strides[i],
2367 b_planes[i], b_strides[i], w, h,
2370 sse = highbd_get_sse(a_planes[i], a_strides[i],
2371 b_planes[i], b_strides[i], w, h);
2374 sse = get_sse(a_planes[i], a_strides[i],
2375 b_planes[i], b_strides[i],
2378 psnr->sse[1 + i] = sse;
2379 psnr->samples[1 + i] = samples;
2380 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2383 total_samples += samples;
2386 psnr->sse[0] = total_sse;
2387 psnr->samples[0] = total_samples;
2388 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2392 #else // !CONFIG_VP9_HIGHBITDEPTH
2394 static void calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b,
2396 static const double peak = 255.0;
2397 const int widths[3] = {
2398 a->y_crop_width, a->uv_crop_width, a->uv_crop_width};
2399 const int heights[3] = {
2400 a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2401 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer};
2402 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2403 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer};
2404 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2406 uint64_t total_sse = 0;
2407 uint32_t total_samples = 0;
2409 for (i = 0; i < 3; ++i) {
2410 const int w = widths[i];
2411 const int h = heights[i];
2412 const uint32_t samples = w * h;
2413 const uint64_t sse = get_sse(a_planes[i], a_strides[i],
2414 b_planes[i], b_strides[i],
2416 psnr->sse[1 + i] = sse;
2417 psnr->samples[1 + i] = samples;
2418 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2421 total_samples += samples;
2424 psnr->sse[0] = total_sse;
2425 psnr->samples[0] = total_samples;
2426 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2429 #endif // CONFIG_VP9_HIGHBITDEPTH
2431 static void generate_psnr_packet(VP9_COMP *cpi) {
2432 struct vpx_codec_cx_pkt pkt;
2435 #if CONFIG_VP9_HIGHBITDEPTH
2436 calc_highbd_psnr(cpi->Source, cpi->common.frame_to_show, &psnr,
2437 cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
2439 calc_psnr(cpi->Source, cpi->common.frame_to_show, &psnr);
2442 for (i = 0; i < 4; ++i) {
2443 pkt.data.psnr.samples[i] = psnr.samples[i];
2444 pkt.data.psnr.sse[i] = psnr.sse[i];
2445 pkt.data.psnr.psnr[i] = psnr.psnr[i];
2447 pkt.kind = VPX_CODEC_PSNR_PKT;
2449 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
2450 cpi->svc.number_temporal_layers].psnr_pkt = pkt.data.psnr;
2452 vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
2455 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) {
2456 if (ref_frame_flags > 7)
2459 cpi->ref_frame_flags = ref_frame_flags;
2463 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) {
2464 cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0;
2465 cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0;
2466 cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0;
2467 cpi->ext_refresh_frame_flags_pending = 1;
2470 static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(VP9_COMP *cpi,
2471 VP9_REFFRAME ref_frame_flag) {
2472 MV_REFERENCE_FRAME ref_frame = NONE;
2473 if (ref_frame_flag == VP9_LAST_FLAG)
2474 ref_frame = LAST_FRAME;
2475 else if (ref_frame_flag == VP9_GOLD_FLAG)
2476 ref_frame = GOLDEN_FRAME;
2477 else if (ref_frame_flag == VP9_ALT_FLAG)
2478 ref_frame = ALTREF_FRAME;
2480 return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame);
2483 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2484 YV12_BUFFER_CONFIG *sd) {
2485 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2487 vp8_yv12_copy_frame(cfg, sd);
2494 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2495 YV12_BUFFER_CONFIG *sd) {
2496 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2498 vp8_yv12_copy_frame(sd, cfg);
2505 int vp9_update_entropy(VP9_COMP * cpi, int update) {
2506 cpi->ext_refresh_frame_context = update;
2507 cpi->ext_refresh_frame_context_pending = 1;
2511 #if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP)
2512 // The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
2513 // as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
2514 // not denoise the UV channels at this time. If ever we implement UV channel
2515 // denoising we will have to modify this.
2516 void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
2517 uint8_t *src = s->y_buffer;
2518 int h = s->y_height;
2521 fwrite(src, s->y_width, 1, f);
2529 fwrite(src, s->uv_width, 1, f);
2530 src += s->uv_stride;
2537 fwrite(src, s->uv_width, 1, f);
2538 src += s->uv_stride;
2543 #ifdef OUTPUT_YUV_REC
2544 void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
2545 YV12_BUFFER_CONFIG *s = cm->frame_to_show;
2546 uint8_t *src = s->y_buffer;
2549 #if CONFIG_VP9_HIGHBITDEPTH
2550 if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
2551 uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
2554 fwrite(src16, s->y_width, 2, yuv_rec_file);
2555 src16 += s->y_stride;
2558 src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
2562 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2563 src16 += s->uv_stride;
2566 src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
2570 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2571 src16 += s->uv_stride;
2574 fflush(yuv_rec_file);
2577 #endif // CONFIG_VP9_HIGHBITDEPTH
2580 fwrite(src, s->y_width, 1, yuv_rec_file);
2588 fwrite(src, s->uv_width, 1, yuv_rec_file);
2589 src += s->uv_stride;
2596 fwrite(src, s->uv_width, 1, yuv_rec_file);
2597 src += s->uv_stride;
2600 fflush(yuv_rec_file);
2604 #if CONFIG_VP9_HIGHBITDEPTH
2605 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2606 YV12_BUFFER_CONFIG *dst,
2609 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2610 YV12_BUFFER_CONFIG *dst) {
2611 #endif // CONFIG_VP9_HIGHBITDEPTH
2612 // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
2614 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2615 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2616 const int src_widths[3] = {src->y_crop_width, src->uv_crop_width,
2617 src->uv_crop_width };
2618 const int src_heights[3] = {src->y_crop_height, src->uv_crop_height,
2619 src->uv_crop_height};
2620 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2621 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2622 const int dst_widths[3] = {dst->y_crop_width, dst->uv_crop_width,
2623 dst->uv_crop_width};
2624 const int dst_heights[3] = {dst->y_crop_height, dst->uv_crop_height,
2625 dst->uv_crop_height};
2627 for (i = 0; i < MAX_MB_PLANE; ++i) {
2628 #if CONFIG_VP9_HIGHBITDEPTH
2629 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2630 vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
2631 src_strides[i], dsts[i], dst_heights[i],
2632 dst_widths[i], dst_strides[i], bd);
2634 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2635 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2638 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2639 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2640 #endif // CONFIG_VP9_HIGHBITDEPTH
2642 vpx_extend_frame_borders(dst);
2645 #if CONFIG_VP9_HIGHBITDEPTH
2646 static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
2647 YV12_BUFFER_CONFIG *dst, int bd) {
2648 const int src_w = src->y_crop_width;
2649 const int src_h = src->y_crop_height;
2650 const int dst_w = dst->y_crop_width;
2651 const int dst_h = dst->y_crop_height;
2652 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2653 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2654 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2655 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2656 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2659 for (i = 0; i < MAX_MB_PLANE; ++i) {
2660 const int factor = (i == 0 || i == 3 ? 1 : 2);
2661 const int src_stride = src_strides[i];
2662 const int dst_stride = dst_strides[i];
2663 for (y = 0; y < dst_h; y += 16) {
2664 const int y_q4 = y * (16 / factor) * src_h / dst_h;
2665 for (x = 0; x < dst_w; x += 16) {
2666 const int x_q4 = x * (16 / factor) * src_w / dst_w;
2667 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
2668 src_stride + (x / factor) * src_w / dst_w;
2669 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2671 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2672 vpx_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
2673 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2674 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2675 16 / factor, 16 / factor, bd);
2677 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2678 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2679 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2680 16 / factor, 16 / factor);
2686 vpx_extend_frame_borders(dst);
2689 void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
2690 YV12_BUFFER_CONFIG *dst) {
2691 const int src_w = src->y_crop_width;
2692 const int src_h = src->y_crop_height;
2693 const int dst_w = dst->y_crop_width;
2694 const int dst_h = dst->y_crop_height;
2695 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2696 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2697 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2698 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2699 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2702 for (i = 0; i < MAX_MB_PLANE; ++i) {
2703 const int factor = (i == 0 || i == 3 ? 1 : 2);
2704 const int src_stride = src_strides[i];
2705 const int dst_stride = dst_strides[i];
2706 for (y = 0; y < dst_h; y += 16) {
2707 const int y_q4 = y * (16 / factor) * src_h / dst_h;
2708 for (x = 0; x < dst_w; x += 16) {
2709 const int x_q4 = x * (16 / factor) * src_w / dst_w;
2710 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
2711 src_stride + (x / factor) * src_w / dst_w;
2712 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2714 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2715 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2716 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2717 16 / factor, 16 / factor);
2722 vpx_extend_frame_borders(dst);
2724 #endif // CONFIG_VP9_HIGHBITDEPTH
2726 static int scale_down(VP9_COMP *cpi, int q) {
2727 RATE_CONTROL *const rc = &cpi->rc;
2728 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2730 assert(frame_is_kf_gf_arf(cpi));
2732 if (rc->frame_size_selector == UNSCALED &&
2733 q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
2734 const int max_size_thresh = (int)(rate_thresh_mult[SCALE_STEP1]
2735 * VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
2736 scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
2741 static int big_rate_miss(VP9_COMP *cpi, int high_limit, int low_limit) {
2742 const RATE_CONTROL *const rc = &cpi->rc;
2744 return (rc->projected_frame_size > ((high_limit * 3) / 2)) ||
2745 (rc->projected_frame_size < (low_limit / 2));
2748 // Function to test for conditions that indicate we should loop
2749 // back and recode a frame.
2750 static int recode_loop_test(VP9_COMP *cpi,
2751 int high_limit, int low_limit,
2752 int q, int maxq, int minq) {
2753 const RATE_CONTROL *const rc = &cpi->rc;
2754 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2755 const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
2756 int force_recode = 0;
2758 if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
2759 big_rate_miss(cpi, high_limit, low_limit) ||
2760 (cpi->sf.recode_loop == ALLOW_RECODE) ||
2761 (frame_is_kfgfarf &&
2762 (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) {
2763 if (frame_is_kfgfarf &&
2764 (oxcf->resize_mode == RESIZE_DYNAMIC) &&
2765 scale_down(cpi, q)) {
2766 // Code this group at a lower resolution.
2767 cpi->resize_pending = 1;
2771 // TODO(agrange) high_limit could be greater than the scale-down threshold.
2772 if ((rc->projected_frame_size > high_limit && q < maxq) ||
2773 (rc->projected_frame_size < low_limit && q > minq)) {
2775 } else if (cpi->oxcf.rc_mode == VPX_CQ) {
2776 // Deal with frame undershoot and whether or not we are
2777 // below the automatically set cq level.
2778 if (q > oxcf->cq_level &&
2779 rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
2784 return force_recode;
2787 void vp9_update_reference_frames(VP9_COMP *cpi) {
2788 VP9_COMMON * const cm = &cpi->common;
2789 BufferPool *const pool = cm->buffer_pool;
2791 // At this point the new frame has been encoded.
2792 // If any buffer copy / swapping is signaled it should be done here.
2793 if (cm->frame_type == KEY_FRAME) {
2794 ref_cnt_fb(pool->frame_bufs,
2795 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2796 ref_cnt_fb(pool->frame_bufs,
2797 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2798 } else if (vp9_preserve_existing_gf(cpi)) {
2799 // We have decided to preserve the previously existing golden frame as our
2800 // new ARF frame. However, in the short term in function
2801 // vp9_bitstream.c::get_refresh_mask() we left it in the GF slot and, if
2802 // we're updating the GF with the current decoded frame, we save it to the
2803 // ARF slot instead.
2804 // We now have to update the ARF with the current frame and swap gld_fb_idx
2805 // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
2806 // slot and, if we're updating the GF, the current frame becomes the new GF.
2809 ref_cnt_fb(pool->frame_bufs,
2810 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2812 tmp = cpi->alt_fb_idx;
2813 cpi->alt_fb_idx = cpi->gld_fb_idx;
2814 cpi->gld_fb_idx = tmp;
2816 if (is_two_pass_svc(cpi)) {
2817 cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx;
2818 cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx;
2820 } else { /* For non key/golden frames */
2821 if (cpi->refresh_alt_ref_frame) {
2822 int arf_idx = cpi->alt_fb_idx;
2823 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
2824 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2825 arf_idx = gf_group->arf_update_idx[gf_group->index];
2828 ref_cnt_fb(pool->frame_bufs,
2829 &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
2830 memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
2831 cpi->interp_filter_selected[0],
2832 sizeof(cpi->interp_filter_selected[0]));
2835 if (cpi->refresh_golden_frame) {
2836 ref_cnt_fb(pool->frame_bufs,
2837 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2838 if (!cpi->rc.is_src_frame_alt_ref)
2839 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2840 cpi->interp_filter_selected[0],
2841 sizeof(cpi->interp_filter_selected[0]));
2843 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2844 cpi->interp_filter_selected[ALTREF_FRAME],
2845 sizeof(cpi->interp_filter_selected[ALTREF_FRAME]));
2849 if (cpi->refresh_last_frame) {
2850 ref_cnt_fb(pool->frame_bufs,
2851 &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
2852 if (!cpi->rc.is_src_frame_alt_ref)
2853 memcpy(cpi->interp_filter_selected[LAST_FRAME],
2854 cpi->interp_filter_selected[0],
2855 sizeof(cpi->interp_filter_selected[0]));
2857 #if CONFIG_VP9_TEMPORAL_DENOISING
2858 if (cpi->oxcf.noise_sensitivity > 0 &&
2859 cpi->denoiser.denoising_level > kDenLowLow) {
2860 vp9_denoiser_update_frame_info(&cpi->denoiser,
2862 cpi->common.frame_type,
2863 cpi->refresh_alt_ref_frame,
2864 cpi->refresh_golden_frame,
2865 cpi->refresh_last_frame,
2866 cpi->resize_pending);
2869 if (is_one_pass_cbr_svc(cpi)) {
2870 // Keep track of frame index for each reference frame.
2871 SVC *const svc = &cpi->svc;
2872 if (cm->frame_type == KEY_FRAME) {
2873 svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2874 svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2875 svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2877 if (cpi->refresh_last_frame)
2878 svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2879 if (cpi->refresh_golden_frame)
2880 svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2881 if (cpi->refresh_alt_ref_frame)
2882 svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2887 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
2888 MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
2889 struct loopfilter *lf = &cm->lf;
2892 lf->filter_level = 0;
2893 lf->last_filt_level = 0;
2895 struct vpx_usec_timer timer;
2897 vpx_clear_system_state();
2899 vpx_usec_timer_start(&timer);
2901 if (!cpi->rc.is_src_frame_alt_ref) {
2902 if ((cpi->common.frame_type == KEY_FRAME) &&
2903 (!cpi->rc.this_key_frame_forced)) {
2904 lf->last_filt_level = 0;
2906 vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick);
2907 lf->last_filt_level = lf->filter_level;
2909 lf->filter_level = 0;
2912 vpx_usec_timer_mark(&timer);
2913 cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
2916 if (lf->filter_level > 0) {
2917 vp9_build_mask_frame(cm, lf->filter_level, 0);
2919 if (cpi->num_workers > 1)
2920 vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
2921 lf->filter_level, 0, 0,
2922 cpi->workers, cpi->num_workers,
2925 vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
2928 vpx_extend_frame_inner_borders(cm->frame_to_show);
2931 static INLINE void alloc_frame_mvs(VP9_COMMON *const cm,
2933 RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
2934 if (new_fb_ptr->mvs == NULL ||
2935 new_fb_ptr->mi_rows < cm->mi_rows ||
2936 new_fb_ptr->mi_cols < cm->mi_cols) {
2937 vpx_free(new_fb_ptr->mvs);
2938 CHECK_MEM_ERROR(cm, new_fb_ptr->mvs,
2939 (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
2940 sizeof(*new_fb_ptr->mvs)));
2941 new_fb_ptr->mi_rows = cm->mi_rows;
2942 new_fb_ptr->mi_cols = cm->mi_cols;
2946 void vp9_scale_references(VP9_COMP *cpi) {
2947 VP9_COMMON *cm = &cpi->common;
2948 MV_REFERENCE_FRAME ref_frame;
2949 const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
2951 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2952 // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
2953 if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
2954 BufferPool *const pool = cm->buffer_pool;
2955 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi,
2959 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
2963 #if CONFIG_VP9_HIGHBITDEPTH
2964 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
2965 RefCntBuffer *new_fb_ptr = NULL;
2966 int force_scaling = 0;
2967 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
2968 if (new_fb == INVALID_IDX) {
2969 new_fb = get_free_fb(cm);
2972 if (new_fb == INVALID_IDX)
2974 new_fb_ptr = &pool->frame_bufs[new_fb];
2975 if (force_scaling ||
2976 new_fb_ptr->buf.y_crop_width != cm->width ||
2977 new_fb_ptr->buf.y_crop_height != cm->height) {
2978 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
2979 cm->subsampling_x, cm->subsampling_y,
2980 cm->use_highbitdepth,
2981 VP9_ENC_BORDER_IN_PIXELS,
2982 cm->byte_alignment, NULL, NULL, NULL))
2983 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
2984 "Failed to allocate frame buffer");
2985 scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth);
2986 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
2987 alloc_frame_mvs(cm, new_fb);
2990 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
2991 RefCntBuffer *new_fb_ptr = NULL;
2992 int force_scaling = 0;
2993 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
2994 if (new_fb == INVALID_IDX) {
2995 new_fb = get_free_fb(cm);
2998 if (new_fb == INVALID_IDX)
3000 new_fb_ptr = &pool->frame_bufs[new_fb];
3001 if (force_scaling ||
3002 new_fb_ptr->buf.y_crop_width != cm->width ||
3003 new_fb_ptr->buf.y_crop_height != cm->height) {
3004 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
3005 cm->subsampling_x, cm->subsampling_y,
3006 VP9_ENC_BORDER_IN_PIXELS,
3007 cm->byte_alignment, NULL, NULL, NULL))
3008 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3009 "Failed to allocate frame buffer");
3010 vp9_scale_and_extend_frame(ref, &new_fb_ptr->buf);
3011 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
3012 alloc_frame_mvs(cm, new_fb);
3014 #endif // CONFIG_VP9_HIGHBITDEPTH
3017 RefCntBuffer *buf = NULL;
3018 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3019 // Check for release of scaled reference.
3020 buf_idx = cpi->scaled_ref_idx[ref_frame - 1];
3021 buf = (buf_idx != INVALID_IDX) ? &pool->frame_bufs[buf_idx] : NULL;
3024 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3027 buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3028 buf = &pool->frame_bufs[buf_idx];
3029 buf->buf.y_crop_width = ref->y_crop_width;
3030 buf->buf.y_crop_height = ref->y_crop_height;
3031 cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
3035 if (cpi->oxcf.pass != 0 || cpi->use_svc)
3036 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3041 static void release_scaled_references(VP9_COMP *cpi) {
3042 VP9_COMMON *cm = &cpi->common;
3044 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3045 // Only release scaled references under certain conditions:
3046 // if reference will be updated, or if scaled reference has same resolution.
3048 refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
3049 refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
3050 refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
3051 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3052 const int idx = cpi->scaled_ref_idx[i - 1];
3053 RefCntBuffer *const buf = idx != INVALID_IDX ?
3054 &cm->buffer_pool->frame_bufs[idx] : NULL;
3055 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
3058 (buf->buf.y_crop_width == ref->y_crop_width &&
3059 buf->buf.y_crop_height == ref->y_crop_height))) {
3061 cpi->scaled_ref_idx[i -1] = INVALID_IDX;
3065 for (i = 0; i < MAX_REF_FRAMES; ++i) {
3066 const int idx = cpi->scaled_ref_idx[i];
3067 RefCntBuffer *const buf = idx != INVALID_IDX ?
3068 &cm->buffer_pool->frame_bufs[idx] : NULL;
3071 cpi->scaled_ref_idx[i] = INVALID_IDX;
3077 static void full_to_model_count(unsigned int *model_count,
3078 unsigned int *full_count) {
3080 model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
3081 model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
3082 model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
3083 for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
3084 model_count[TWO_TOKEN] += full_count[n];
3085 model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
3088 static void full_to_model_counts(vp9_coeff_count_model *model_count,
3089 vp9_coeff_count *full_count) {
3092 for (i = 0; i < PLANE_TYPES; ++i)
3093 for (j = 0; j < REF_TYPES; ++j)
3094 for (k = 0; k < COEF_BANDS; ++k)
3095 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
3096 full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
3099 #if 0 && CONFIG_INTERNAL_STATS
3100 static void output_frame_level_debug_stats(VP9_COMP *cpi) {
3101 VP9_COMMON *const cm = &cpi->common;
3102 FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
3105 vpx_clear_system_state();
3107 #if CONFIG_VP9_HIGHBITDEPTH
3108 if (cm->use_highbitdepth) {
3109 recon_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3111 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3114 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3115 #endif // CONFIG_VP9_HIGHBITDEPTH
3118 if (cpi->twopass.total_left_stats.coded_error != 0.0) {
3119 double dc_quant_devisor;
3120 #if CONFIG_VP9_HIGHBITDEPTH
3121 switch (cm->bit_depth) {
3123 dc_quant_devisor = 4.0;
3126 dc_quant_devisor = 16.0;
3129 dc_quant_devisor = 64.0;
3132 assert(0 && "bit_depth must be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
3136 dc_quant_devisor = 4.0;
3139 fprintf(f, "%10u %dx%d %10d %10d %d %d %10d %10d %10d %10d"
3140 "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
3141 "%10"PRId64" %10"PRId64" %10d "
3142 "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
3143 "%6d %6d %5d %5d %5d "
3144 "%10"PRId64" %10.3lf"
3145 "%10lf %8u %10"PRId64" %10d %10d %10d %10d %10d\n",
3146 cpi->common.current_video_frame,
3147 cm->width, cm->height,
3148 cpi->td.rd_counts.m_search_count,
3149 cpi->td.rd_counts.ex_search_count,
3150 cpi->rc.source_alt_ref_pending,
3151 cpi->rc.source_alt_ref_active,
3152 cpi->rc.this_frame_target,
3153 cpi->rc.projected_frame_size,
3154 cpi->rc.projected_frame_size / cpi->common.MBs,
3155 (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
3156 cpi->rc.vbr_bits_off_target,
3157 cpi->rc.vbr_bits_off_target_fast,
3158 cpi->twopass.extend_minq,
3159 cpi->twopass.extend_minq_fast,
3160 cpi->rc.total_target_vs_actual,
3161 (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
3162 cpi->rc.total_actual_bits, cm->base_qindex,
3163 vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
3164 (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) /
3166 vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality,
3169 vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
3170 cpi->refresh_last_frame, cpi->refresh_golden_frame,
3171 cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
3172 cpi->twopass.bits_left,
3173 cpi->twopass.total_left_stats.coded_error,
3174 cpi->twopass.bits_left /
3175 (1 + cpi->twopass.total_left_stats.coded_error),
3176 cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
3177 cpi->twopass.kf_zeromotion_pct,
3178 cpi->twopass.fr_content_type,
3179 cm->lf.filter_level,
3180 cm->seg.aq_av_offset);
3185 FILE *const fmodes = fopen("Modes.stt", "a");
3188 fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
3189 cm->frame_type, cpi->refresh_golden_frame,
3190 cpi->refresh_alt_ref_frame);
3192 for (i = 0; i < MAX_MODES; ++i)
3193 fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
3195 fprintf(fmodes, "\n");
3202 static void set_mv_search_params(VP9_COMP *cpi) {
3203 const VP9_COMMON *const cm = &cpi->common;
3204 const unsigned int max_mv_def = VPXMIN(cm->width, cm->height);
3206 // Default based on max resolution.
3207 cpi->mv_step_param = vp9_init_search_range(max_mv_def);
3209 if (cpi->sf.mv.auto_mv_step_size) {
3210 if (frame_is_intra_only(cm)) {
3211 // Initialize max_mv_magnitude for use in the first INTER frame
3212 // after a key/intra-only frame.
3213 cpi->max_mv_magnitude = max_mv_def;
3215 if (cm->show_frame) {
3216 // Allow mv_steps to correspond to twice the max mv magnitude found
3217 // in the previous frame, capped by the default max_mv_magnitude based
3219 cpi->mv_step_param = vp9_init_search_range(
3220 VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
3222 cpi->max_mv_magnitude = 0;
3227 static void set_size_independent_vars(VP9_COMP *cpi) {
3228 vp9_set_speed_features_framesize_independent(cpi);
3229 vp9_set_rd_speed_thresholds(cpi);
3230 vp9_set_rd_speed_thresholds_sub8x8(cpi);
3231 cpi->common.interp_filter = cpi->sf.default_interp_filter;
3234 static void set_size_dependent_vars(VP9_COMP *cpi, int *q,
3235 int *bottom_index, int *top_index) {
3236 VP9_COMMON *const cm = &cpi->common;
3237 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3239 // Setup variables that depend on the dimensions of the frame.
3240 vp9_set_speed_features_framesize_dependent(cpi);
3242 // Decide q and q bounds.
3243 *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
3245 if (!frame_is_intra_only(cm)) {
3246 vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
3249 // Configure experimental use of segmentation for enhanced coding of
3250 // static regions if indicated.
3251 // Only allowed in the second pass of a two pass encode, as it requires
3252 // lagged coding, and if the relevant speed feature flag is set.
3253 if (oxcf->pass == 2 && cpi->sf.static_segmentation)
3254 configure_static_seg_features(cpi);
3256 #if CONFIG_VP9_POSTPROC && !(CONFIG_VP9_TEMPORAL_DENOISING)
3257 if (oxcf->noise_sensitivity > 0) {
3259 switch (oxcf->noise_sensitivity) {
3277 vp9_denoise(cpi->Source, cpi->Source, l);
3279 #endif // CONFIG_VP9_POSTPROC
3282 #if CONFIG_VP9_TEMPORAL_DENOISING
3283 static void setup_denoiser_buffer(VP9_COMP *cpi) {
3284 VP9_COMMON *const cm = &cpi->common;
3285 if (cpi->oxcf.noise_sensitivity > 0 &&
3286 !cpi->denoiser.frame_buffer_initialized) {
3287 if (vp9_denoiser_alloc(&cpi->denoiser, cm->width, cm->height,
3288 cm->subsampling_x, cm->subsampling_y,
3289 #if CONFIG_VP9_HIGHBITDEPTH
3290 cm->use_highbitdepth,
3292 VP9_ENC_BORDER_IN_PIXELS))
3293 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3294 "Failed to allocate denoiser");
3299 static void init_motion_estimation(VP9_COMP *cpi) {
3300 int y_stride = cpi->scaled_source.y_stride;
3302 if (cpi->sf.mv.search_method == NSTEP) {
3303 vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
3304 } else if (cpi->sf.mv.search_method == DIAMOND) {
3305 vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
3309 static void set_frame_size(VP9_COMP *cpi) {
3311 VP9_COMMON *const cm = &cpi->common;
3312 VP9EncoderConfig *const oxcf = &cpi->oxcf;
3313 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
3315 if (oxcf->pass == 2 &&
3316 oxcf->rc_mode == VPX_VBR &&
3317 ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
3318 (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
3319 calculate_coded_size(
3320 cpi, &oxcf->scaled_frame_width, &oxcf->scaled_frame_height);
3322 // There has been a change in frame size.
3323 vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
3324 oxcf->scaled_frame_height);
3327 if (oxcf->pass == 0 &&
3328 oxcf->rc_mode == VPX_CBR &&
3330 oxcf->resize_mode == RESIZE_DYNAMIC &&
3331 cpi->resize_pending != 0) {
3332 oxcf->scaled_frame_width =
3333 (oxcf->width * cpi->resize_scale_num) / cpi->resize_scale_den;
3334 oxcf->scaled_frame_height =
3335 (oxcf->height * cpi->resize_scale_num) /cpi->resize_scale_den;
3336 // There has been a change in frame size.
3337 vp9_set_size_literal(cpi,
3338 oxcf->scaled_frame_width,
3339 oxcf->scaled_frame_height);
3341 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3342 set_mv_search_params(cpi);
3344 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
3345 #if CONFIG_VP9_TEMPORAL_DENOISING
3346 // Reset the denoiser on the resized frame.
3347 if (cpi->oxcf.noise_sensitivity > 0) {
3348 vp9_denoiser_free(&(cpi->denoiser));
3349 setup_denoiser_buffer(cpi);
3350 // Dynamic resize is only triggered for non-SVC, so we can force
3351 // golden frame update here as temporary fix to denoiser.
3352 cpi->refresh_golden_frame = 1;
3357 if ((oxcf->pass == 2) &&
3359 (is_two_pass_svc(cpi) &&
3360 cpi->svc.encode_empty_frame_state != ENCODING))) {
3361 vp9_set_target_rate(cpi);
3364 alloc_frame_mvs(cm, cm->new_fb_idx);
3366 // Reset the frame pointers to the current frame size.
3367 if (vpx_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
3368 cm->subsampling_x, cm->subsampling_y,
3369 #if CONFIG_VP9_HIGHBITDEPTH
3370 cm->use_highbitdepth,
3372 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
3374 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3375 "Failed to allocate frame buffer");
3377 alloc_util_frame_buffers(cpi);
3378 init_motion_estimation(cpi);
3380 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3381 RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
3382 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3384 ref_buf->idx = buf_idx;
3386 if (buf_idx != INVALID_IDX) {
3387 YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
3389 #if CONFIG_VP9_HIGHBITDEPTH
3390 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3391 buf->y_crop_width, buf->y_crop_height,
3392 cm->width, cm->height,
3393 (buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
3396 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3397 buf->y_crop_width, buf->y_crop_height,
3398 cm->width, cm->height);
3399 #endif // CONFIG_VP9_HIGHBITDEPTH
3400 if (vp9_is_scaled(&ref_buf->sf))
3401 vpx_extend_frame_borders(buf);
3403 ref_buf->buf = NULL;
3407 set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
3410 static void encode_without_recode_loop(VP9_COMP *cpi,
3413 VP9_COMMON *const cm = &cpi->common;
3414 int q = 0, bottom_index = 0, top_index = 0; // Dummy variables.
3416 vpx_clear_system_state();
3418 set_frame_size(cpi);
3420 if (is_one_pass_cbr_svc(cpi) &&
3421 cpi->un_scaled_source->y_width == cm->width << 2 &&
3422 cpi->un_scaled_source->y_height == cm->height << 2 &&
3423 cpi->svc.scaled_temp.y_width == cm->width << 1 &&
3424 cpi->svc.scaled_temp.y_height == cm->height << 1) {
3425 cpi->Source = vp9_svc_twostage_scale(cm,
3426 cpi->un_scaled_source,
3427 &cpi->scaled_source,
3428 &cpi->svc.scaled_temp);
3430 cpi->Source = vp9_scale_if_required(cm,
3431 cpi->un_scaled_source,
3432 &cpi->scaled_source,
3433 (cpi->oxcf.pass == 0));
3435 // Avoid scaling last_source unless its needed.
3436 // Last source is needed if vp9_avg_source_sad() is used, or if
3437 // partition_search_type == SOURCE_VAR_BASED_PARTITION, or if noise
3438 // estimation is enabled.
3439 if (cpi->unscaled_last_source != NULL &&
3440 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3441 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_VBR &&
3442 cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5) ||
3443 cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION ||
3444 cpi->noise_estimate.enabled))
3445 cpi->Last_Source = vp9_scale_if_required(cm,
3446 cpi->unscaled_last_source,
3447 &cpi->scaled_last_source,
3448 (cpi->oxcf.pass == 0));
3450 if (cm->frame_type == KEY_FRAME || cpi->resize_pending != 0) {
3451 memset(cpi->consec_zero_mv, 0,
3452 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
3455 vp9_update_noise_estimate(cpi);
3457 if (cpi->oxcf.pass == 0 &&
3458 cpi->oxcf.mode == REALTIME &&
3459 cpi->oxcf.speed >= 5 &&
3460 cpi->resize_state == 0 &&
3461 cm->frame_type != KEY_FRAME &&
3462 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3463 cpi->oxcf.rc_mode == VPX_VBR))
3464 vp9_avg_source_sad(cpi);
3466 // For 1 pass SVC, since only ZEROMV is allowed for upsampled reference
3467 // frame (i.e, svc->force_zero_mode_spatial_ref = 0), we can avoid this
3468 // frame-level upsampling.
3469 if (frame_is_intra_only(cm) == 0 && !is_one_pass_cbr_svc(cpi)) {
3470 vp9_scale_references(cpi);
3473 set_size_independent_vars(cpi);
3474 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3476 if (cpi->oxcf.speed >= 5 &&
3477 cpi->oxcf.pass == 0 &&
3478 cpi->oxcf.rc_mode == VPX_CBR &&
3479 cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
3480 cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3481 cpi->use_skin_detection = 1;
3484 vp9_set_quantizer(cm, q);
3485 vp9_set_variance_partition_thresholds(cpi, q);
3489 suppress_active_map(cpi);
3490 // Variance adaptive and in frame q adjustment experiments are mutually
3492 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3493 vp9_vaq_frame_setup(cpi);
3494 } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3495 vp9_360aq_frame_setup(cpi);
3496 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3497 vp9_setup_in_frame_q_adj(cpi);
3498 } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3499 vp9_cyclic_refresh_setup(cpi);
3501 apply_active_map(cpi);
3503 // transform / motion compensation build reconstruction frame
3504 vp9_encode_frame(cpi);
3506 // Check if we should drop this frame because of high overshoot.
3507 // Only for frames where high temporal-source sad is detected.
3508 if (cpi->oxcf.pass == 0 &&
3509 cpi->oxcf.rc_mode == VPX_CBR &&
3510 cpi->resize_state == 0 &&
3511 cm->frame_type != KEY_FRAME &&
3512 cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
3513 cpi->rc.high_source_sad == 1) {
3515 // Get an estimate of the encoded frame size.
3516 save_coding_context(cpi);
3517 vp9_pack_bitstream(cpi, dest, size);
3518 restore_coding_context(cpi);
3519 frame_size = (int)(*size) << 3;
3520 // Check if encoded frame will overshoot too much, and if so, set the q and
3521 // adjust some rate control parameters, and return to re-encode the frame.
3522 if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) {
3523 vpx_clear_system_state();
3524 vp9_set_quantizer(cm, q);
3525 vp9_set_variance_partition_thresholds(cpi, q);
3526 suppress_active_map(cpi);
3527 // Turn-off cyclic refresh for re-encoded frame.
3528 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3529 unsigned char *const seg_map = cpi->segmentation_map;
3530 memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
3531 vp9_disable_segmentation(&cm->seg);
3533 apply_active_map(cpi);
3534 vp9_encode_frame(cpi);
3538 // Update some stats from cyclic refresh, and check if we should not update
3539 // golden reference, for non-SVC 1 pass CBR.
3540 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
3541 cm->frame_type != KEY_FRAME &&
3543 cpi->ext_refresh_frame_flags_pending == 0 &&
3544 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR))
3545 vp9_cyclic_refresh_check_golden_update(cpi);
3547 // Update the skip mb flag probabilities based on the distribution
3548 // seen in the last encoder iteration.
3549 // update_base_skip_probs(cpi);
3550 vpx_clear_system_state();
3553 static void encode_with_recode_loop(VP9_COMP *cpi,
3556 VP9_COMMON *const cm = &cpi->common;
3557 RATE_CONTROL *const rc = &cpi->rc;
3558 int bottom_index, top_index;
3560 int loop_at_this_size = 0;
3562 int overshoot_seen = 0;
3563 int undershoot_seen = 0;
3564 int frame_over_shoot_limit;
3565 int frame_under_shoot_limit;
3566 int q = 0, q_low = 0, q_high = 0;
3568 set_size_independent_vars(cpi);
3571 vpx_clear_system_state();
3573 set_frame_size(cpi);
3575 if (loop_count == 0 || cpi->resize_pending != 0) {
3576 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3578 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3579 set_mv_search_params(cpi);
3581 // Reset the loop state for new frame size.
3583 undershoot_seen = 0;
3585 // Reconfiguration for change in frame size has concluded.
3586 cpi->resize_pending = 0;
3588 q_low = bottom_index;
3591 loop_at_this_size = 0;
3594 // Decide frame size bounds first time through.
3595 if (loop_count == 0) {
3596 vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
3597 &frame_under_shoot_limit,
3598 &frame_over_shoot_limit);
3601 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
3602 &cpi->scaled_source,
3603 (cpi->oxcf.pass == 0));
3605 if (cpi->unscaled_last_source != NULL)
3606 cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
3607 &cpi->scaled_last_source,
3608 (cpi->oxcf.pass == 0));
3610 if (frame_is_intra_only(cm) == 0) {
3611 if (loop_count > 0) {
3612 release_scaled_references(cpi);
3614 vp9_scale_references(cpi);
3617 vp9_set_quantizer(cm, q);
3619 if (loop_count == 0)
3622 // Variance adaptive and in frame q adjustment experiments are mutually
3624 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3625 vp9_vaq_frame_setup(cpi);
3626 } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3627 vp9_360aq_frame_setup(cpi);
3628 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3629 vp9_setup_in_frame_q_adj(cpi);
3632 // transform / motion compensation build reconstruction frame
3633 vp9_encode_frame(cpi);
3635 // Update the skip mb flag probabilities based on the distribution
3636 // seen in the last encoder iteration.
3637 // update_base_skip_probs(cpi);
3639 vpx_clear_system_state();
3641 // Dummy pack of the bitstream using up to date stats to get an
3642 // accurate estimate of output frame size to determine if we need
3644 if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
3645 save_coding_context(cpi);
3646 if (!cpi->sf.use_nonrd_pick_mode)
3647 vp9_pack_bitstream(cpi, dest, size);
3649 rc->projected_frame_size = (int)(*size) << 3;
3650 restore_coding_context(cpi);
3652 if (frame_over_shoot_limit == 0)
3653 frame_over_shoot_limit = 1;
3656 if (cpi->oxcf.rc_mode == VPX_Q) {
3659 if ((cm->frame_type == KEY_FRAME) &&
3660 rc->this_key_frame_forced &&
3661 (rc->projected_frame_size < rc->max_frame_bandwidth)) {
3665 int64_t high_err_target = cpi->ambient_err;
3666 int64_t low_err_target = cpi->ambient_err >> 1;
3668 #if CONFIG_VP9_HIGHBITDEPTH
3669 if (cm->use_highbitdepth) {
3670 kf_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3672 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3675 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3676 #endif // CONFIG_VP9_HIGHBITDEPTH
3678 // Prevent possible divide by zero error below for perfect KF
3681 // The key frame is not good enough or we can afford
3682 // to make it better without undue risk of popping.
3683 if ((kf_err > high_err_target &&
3684 rc->projected_frame_size <= frame_over_shoot_limit) ||
3685 (kf_err > low_err_target &&
3686 rc->projected_frame_size <= frame_under_shoot_limit)) {
3688 q_high = q > q_low ? q - 1 : q_low;
3691 q = (int)((q * high_err_target) / kf_err);
3692 q = VPXMIN(q, (q_high + q_low) >> 1);
3693 } else if (kf_err < low_err_target &&
3694 rc->projected_frame_size >= frame_under_shoot_limit) {
3695 // The key frame is much better than the previous frame
3697 q_low = q < q_high ? q + 1 : q_high;
3700 q = (int)((q * low_err_target) / kf_err);
3701 q = VPXMIN(q, (q_high + q_low + 1) >> 1);
3704 // Clamp Q to upper and lower limits:
3705 q = clamp(q, q_low, q_high);
3708 } else if (recode_loop_test(
3709 cpi, frame_over_shoot_limit, frame_under_shoot_limit,
3710 q, VPXMAX(q_high, top_index), bottom_index)) {
3711 // Is the projected frame size out of range and are we allowed
3712 // to attempt to recode.
3716 if (cpi->resize_pending == 1) {
3717 // Change in frame size so go back around the recode loop.
3718 cpi->rc.frame_size_selector =
3719 SCALE_STEP1 - cpi->rc.frame_size_selector;
3720 cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
3722 #if CONFIG_INTERNAL_STATS
3723 ++cpi->tot_recode_hits;
3730 // Frame size out of permitted range:
3731 // Update correction factor & compute new Q to try...
3733 // Frame is too large
3734 if (rc->projected_frame_size > rc->this_frame_target) {
3735 // Special case if the projected size is > the max allowed.
3736 if (rc->projected_frame_size >= rc->max_frame_bandwidth)
3737 q_high = rc->worst_quality;
3739 // Raise Qlow as to at least the current value
3740 q_low = q < q_high ? q + 1 : q_high;
3742 if (undershoot_seen || loop_at_this_size > 1) {
3743 // Update rate_correction_factor unless
3744 vp9_rc_update_rate_correction_factors(cpi);
3746 q = (q_high + q_low + 1) / 2;
3748 // Update rate_correction_factor unless
3749 vp9_rc_update_rate_correction_factors(cpi);
3751 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3752 bottom_index, VPXMAX(q_high, top_index));
3754 while (q < q_low && retries < 10) {
3755 vp9_rc_update_rate_correction_factors(cpi);
3756 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3757 bottom_index, VPXMAX(q_high, top_index));
3764 // Frame is too small
3765 q_high = q > q_low ? q - 1 : q_low;
3767 if (overshoot_seen || loop_at_this_size > 1) {
3768 vp9_rc_update_rate_correction_factors(cpi);
3769 q = (q_high + q_low) / 2;
3771 vp9_rc_update_rate_correction_factors(cpi);
3772 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3773 bottom_index, top_index);
3774 // Special case reset for qlow for constrained quality.
3775 // This should only trigger where there is very substantial
3776 // undershoot on a frame and the auto cq level is above
3777 // the user passsed in value.
3778 if (cpi->oxcf.rc_mode == VPX_CQ &&
3783 while (q > q_high && retries < 10) {
3784 vp9_rc_update_rate_correction_factors(cpi);
3785 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3786 bottom_index, top_index);
3791 undershoot_seen = 1;
3794 // Clamp Q to upper and lower limits:
3795 q = clamp(q, q_low, q_high);
3797 loop = (q != last_q);
3803 // Special case for overlay frame.
3804 if (rc->is_src_frame_alt_ref &&
3805 rc->projected_frame_size < rc->max_frame_bandwidth)
3810 ++loop_at_this_size;
3812 #if CONFIG_INTERNAL_STATS
3813 ++cpi->tot_recode_hits;
3819 static int get_ref_frame_flags(const VP9_COMP *cpi) {
3820 const int *const map = cpi->common.ref_frame_map;
3821 const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
3822 const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
3823 const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
3824 int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
3827 flags &= ~VP9_GOLD_FLAG;
3829 if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
3830 (cpi->svc.number_temporal_layers == 1 &&
3831 cpi->svc.number_spatial_layers == 1))
3832 flags &= ~VP9_GOLD_FLAG;
3835 flags &= ~VP9_ALT_FLAG;
3838 flags &= ~VP9_ALT_FLAG;
3843 static void set_ext_overrides(VP9_COMP *cpi) {
3844 // Overrides the defaults with the externally supplied values with
3845 // vp9_update_reference() and vp9_update_entropy() calls
3846 // Note: The overrides are valid only for the next frame passed
3847 // to encode_frame_to_data_rate() function
3848 if (cpi->ext_refresh_frame_context_pending) {
3849 cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
3850 cpi->ext_refresh_frame_context_pending = 0;
3852 if (cpi->ext_refresh_frame_flags_pending) {
3853 cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
3854 cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
3855 cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
3859 YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
3860 YV12_BUFFER_CONFIG *unscaled,
3861 YV12_BUFFER_CONFIG *scaled,
3862 YV12_BUFFER_CONFIG *scaled_temp) {
3863 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3864 cm->mi_rows * MI_SIZE != unscaled->y_height) {
3865 #if CONFIG_VP9_HIGHBITDEPTH
3866 scale_and_extend_frame(unscaled, scaled_temp, (int)cm->bit_depth);
3867 scale_and_extend_frame(scaled_temp, scaled, (int)cm->bit_depth);
3869 vp9_scale_and_extend_frame(unscaled, scaled_temp);
3870 vp9_scale_and_extend_frame(scaled_temp, scaled);
3871 #endif // CONFIG_VP9_HIGHBITDEPTH
3878 YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
3879 YV12_BUFFER_CONFIG *unscaled,
3880 YV12_BUFFER_CONFIG *scaled,
3881 int use_normative_scaler) {
3882 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3883 cm->mi_rows * MI_SIZE != unscaled->y_height) {
3884 #if CONFIG_VP9_HIGHBITDEPTH
3885 if (use_normative_scaler &&
3886 unscaled->y_width <= (scaled->y_width << 1) &&
3887 unscaled->y_height <= (scaled->y_height << 1))
3888 scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth);
3890 scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
3892 if (use_normative_scaler &&
3893 unscaled->y_width <= (scaled->y_width << 1) &&
3894 unscaled->y_height <= (scaled->y_height << 1))
3895 vp9_scale_and_extend_frame(unscaled, scaled);
3897 scale_and_extend_frame_nonnormative(unscaled, scaled);
3898 #endif // CONFIG_VP9_HIGHBITDEPTH
3905 static void set_arf_sign_bias(VP9_COMP *cpi) {
3906 VP9_COMMON *const cm = &cpi->common;
3909 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
3910 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
3911 arf_sign_bias = cpi->rc.source_alt_ref_active &&
3912 (!cpi->refresh_alt_ref_frame ||
3913 (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
3916 (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
3918 cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
3921 static int setup_interp_filter_search_mask(VP9_COMP *cpi) {
3922 INTERP_FILTER ifilter;
3923 int ref_total[MAX_REF_FRAMES] = {0};
3924 MV_REFERENCE_FRAME ref;
3926 if (cpi->common.last_frame_type == KEY_FRAME ||
3927 cpi->refresh_alt_ref_frame)
3929 for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
3930 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter)
3931 ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
3933 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) {
3934 if ((ref_total[LAST_FRAME] &&
3935 cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
3936 (ref_total[GOLDEN_FRAME] == 0 ||
3937 cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50
3938 < ref_total[GOLDEN_FRAME]) &&
3939 (ref_total[ALTREF_FRAME] == 0 ||
3940 cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50
3941 < ref_total[ALTREF_FRAME]))
3942 mask |= 1 << ifilter;
3947 static void encode_frame_to_data_rate(VP9_COMP *cpi,
3950 unsigned int *frame_flags) {
3951 VP9_COMMON *const cm = &cpi->common;
3952 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3953 struct segmentation *const seg = &cm->seg;
3956 set_ext_overrides(cpi);
3957 vpx_clear_system_state();
3959 // Set the arf sign bias for this frame.
3960 set_arf_sign_bias(cpi);
3962 // Set default state for segment based loop filter update flags.
3963 cm->lf.mode_ref_delta_update = 0;
3965 if (cpi->oxcf.pass == 2 &&
3966 cpi->sf.adaptive_interp_filter_search)
3967 cpi->sf.interp_filter_search_mask =
3968 setup_interp_filter_search_mask(cpi);
3970 // Set various flags etc to special state if it is a key frame.
3971 if (frame_is_intra_only(cm)) {
3972 // Reset the loop filter deltas and segmentation map.
3973 vp9_reset_segment_features(&cm->seg);
3975 // If segmentation is enabled force a map update for key frames.
3977 seg->update_map = 1;
3978 seg->update_data = 1;
3981 // The alternate reference frame cannot be active for a key frame.
3982 cpi->rc.source_alt_ref_active = 0;
3984 cm->error_resilient_mode = oxcf->error_resilient_mode;
3985 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
3987 // By default, encoder assumes decoder can use prev_mi.
3988 if (cm->error_resilient_mode) {
3989 cm->frame_parallel_decoding_mode = 1;
3990 cm->reset_frame_context = 0;
3991 cm->refresh_frame_context = 0;
3992 } else if (cm->intra_only) {
3993 // Only reset the current context.
3994 cm->reset_frame_context = 2;
3997 if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) {
3998 // Use context 0 for intra only empty frame, but the last frame context
3999 // for other empty frames.
4000 if (cpi->svc.encode_empty_frame_state == ENCODING) {
4001 if (cpi->svc.encode_intra_empty_frame != 0)
4002 cm->frame_context_idx = 0;
4004 cm->frame_context_idx = FRAME_CONTEXTS - 1;
4006 cm->frame_context_idx =
4007 cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
4008 cpi->svc.temporal_layer_id;
4011 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
4013 // The probs will be updated based on the frame type of its previous
4014 // frame if frame_parallel_decoding_mode is 0. The type may vary for
4015 // the frame after a key frame in base layer since we may drop enhancement
4016 // layers. So set frame_parallel_decoding_mode to 1 in this case.
4017 if (cm->frame_parallel_decoding_mode == 0) {
4018 if (cpi->svc.number_temporal_layers == 1) {
4019 if (cpi->svc.spatial_layer_id == 0 &&
4020 cpi->svc.layer_context[0].last_frame_type == KEY_FRAME)
4021 cm->frame_parallel_decoding_mode = 1;
4022 } else if (cpi->svc.spatial_layer_id == 0) {
4023 // Find the 2nd frame in temporal base layer and 1st frame in temporal
4024 // enhancement layers from the key frame.
4026 for (i = 0; i < cpi->svc.number_temporal_layers; ++i) {
4027 if (cpi->svc.layer_context[0].frames_from_key_frame == 1 << i) {
4028 cm->frame_parallel_decoding_mode = 1;
4036 // For 1 pass CBR, check if we are dropping this frame.
4037 // For spatial layers, for now only check for frame-dropping on first spatial
4038 // layer, and if decision is to drop, we drop whole super-frame.
4039 if (oxcf->pass == 0 &&
4040 oxcf->rc_mode == VPX_CBR &&
4041 cm->frame_type != KEY_FRAME) {
4042 if (vp9_rc_drop_frame(cpi) ||
4043 (is_one_pass_cbr_svc(cpi) && cpi->svc.rc_drop_superframe == 1)) {
4044 vp9_rc_postencode_update_drop_frame(cpi);
4045 ++cm->current_video_frame;
4046 cpi->ext_refresh_frame_flags_pending = 0;
4047 cpi->svc.rc_drop_superframe = 1;
4048 // TODO(marpan): Advancing the svc counters on dropped frames can break
4049 // the referencing scheme for the fixed svc patterns defined in
4050 // vp9_one_pass_cbr_svc_start_layer(). Look into fixing this issue, but
4051 // for now, don't advance the svc frame counters on dropped frame.
4052 // if (cpi->use_svc)
4053 // vp9_inc_frame_in_layer(cpi);
4058 vpx_clear_system_state();
4060 #if CONFIG_INTERNAL_STATS
4061 memset(cpi->mode_chosen_counts, 0,
4062 MAX_MODES * sizeof(*cpi->mode_chosen_counts));
4065 if (cpi->sf.recode_loop == DISALLOW_RECODE) {
4066 encode_without_recode_loop(cpi, size, dest);
4068 encode_with_recode_loop(cpi, size, dest);
4071 #if CONFIG_VP9_TEMPORAL_DENOISING
4072 #ifdef OUTPUT_YUV_DENOISED
4073 if (oxcf->noise_sensitivity > 0) {
4074 vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME],
4079 #ifdef OUTPUT_YUV_SKINMAP
4080 if (cpi->common.current_video_frame > 1) {
4081 vp9_compute_skin_map(cpi, yuv_skinmap_file);
4085 // Special case code to reduce pulsing when key frames are forced at a
4086 // fixed interval. Note the reconstruction error if it is the frame before
4087 // the force key frame
4088 if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
4089 #if CONFIG_VP9_HIGHBITDEPTH
4090 if (cm->use_highbitdepth) {
4091 cpi->ambient_err = vp9_highbd_get_y_sse(cpi->Source,
4092 get_frame_new_buffer(cm));
4094 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4097 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4098 #endif // CONFIG_VP9_HIGHBITDEPTH
4101 // If the encoder forced a KEY_FRAME decision
4102 if (cm->frame_type == KEY_FRAME)
4103 cpi->refresh_last_frame = 1;
4105 cm->frame_to_show = get_frame_new_buffer(cm);
4106 cm->frame_to_show->color_space = cm->color_space;
4107 cm->frame_to_show->color_range = cm->color_range;
4108 cm->frame_to_show->render_width = cm->render_width;
4109 cm->frame_to_show->render_height = cm->render_height;
4111 // Pick the loop filter level for the frame.
4112 loopfilter_frame(cpi, cm);
4114 // build the bitstream
4115 vp9_pack_bitstream(cpi, dest, size);
4117 if (cm->seg.update_map)
4118 update_reference_segmentation_map(cpi);
4120 if (frame_is_intra_only(cm) == 0) {
4121 release_scaled_references(cpi);
4123 vp9_update_reference_frames(cpi);
4125 for (t = TX_4X4; t <= TX_32X32; t++)
4126 full_to_model_counts(cpi->td.counts->coef[t],
4127 cpi->td.rd_counts.coef_counts[t]);
4129 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode)
4130 vp9_adapt_coef_probs(cm);
4132 if (!frame_is_intra_only(cm)) {
4133 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
4134 vp9_adapt_mode_probs(cm);
4135 vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
4139 cpi->ext_refresh_frame_flags_pending = 0;
4141 if (cpi->refresh_golden_frame == 1)
4142 cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
4144 cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
4146 if (cpi->refresh_alt_ref_frame == 1)
4147 cpi->frame_flags |= FRAMEFLAGS_ALTREF;
4149 cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
4151 cpi->ref_frame_flags = get_ref_frame_flags(cpi);
4153 cm->last_frame_type = cm->frame_type;
4155 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
4156 vp9_rc_postencode_update(cpi, *size);
4159 output_frame_level_debug_stats(cpi);
4162 if (cm->frame_type == KEY_FRAME) {
4163 // Tell the caller that the frame was coded as a key frame
4164 *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
4166 *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
4169 // Clear the one shot update flags for segmentation map and mode/ref loop
4171 cm->seg.update_map = 0;
4172 cm->seg.update_data = 0;
4173 cm->lf.mode_ref_delta_update = 0;
4175 // keep track of the last coded dimensions
4176 cm->last_width = cm->width;
4177 cm->last_height = cm->height;
4179 // reset to normal state now that we are done.
4180 if (!cm->show_existing_frame)
4181 cm->last_show_frame = cm->show_frame;
4183 if (cm->show_frame) {
4184 vp9_swap_mi_and_prev_mi(cm);
4185 // Don't increment frame counters if this was an altref buffer
4186 // update not a real frame
4187 ++cm->current_video_frame;
4189 vp9_inc_frame_in_layer(cpi);
4191 cm->prev_frame = cm->cur_frame;
4194 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
4195 cpi->svc.number_temporal_layers +
4196 cpi->svc.temporal_layer_id].last_frame_type =
4200 static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
4201 unsigned int *frame_flags) {
4202 vp9_rc_get_svc_params(cpi);
4203 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4206 static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
4207 unsigned int *frame_flags) {
4208 if (cpi->oxcf.rc_mode == VPX_CBR) {
4209 vp9_rc_get_one_pass_cbr_params(cpi);
4211 vp9_rc_get_one_pass_vbr_params(cpi);
4213 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4216 static void Pass2Encode(VP9_COMP *cpi, size_t *size,
4217 uint8_t *dest, unsigned int *frame_flags) {
4218 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
4219 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4221 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
4222 vp9_twopass_postencode_update(cpi);
4225 static void init_ref_frame_bufs(VP9_COMMON *cm) {
4227 BufferPool *const pool = cm->buffer_pool;
4228 cm->new_fb_idx = INVALID_IDX;
4229 for (i = 0; i < REF_FRAMES; ++i) {
4230 cm->ref_frame_map[i] = INVALID_IDX;
4231 pool->frame_bufs[i].ref_count = 0;
4235 static void check_initial_width(VP9_COMP *cpi,
4236 #if CONFIG_VP9_HIGHBITDEPTH
4237 int use_highbitdepth,
4239 int subsampling_x, int subsampling_y) {
4240 VP9_COMMON *const cm = &cpi->common;
4242 if (!cpi->initial_width ||
4243 #if CONFIG_VP9_HIGHBITDEPTH
4244 cm->use_highbitdepth != use_highbitdepth ||
4246 cm->subsampling_x != subsampling_x ||
4247 cm->subsampling_y != subsampling_y) {
4248 cm->subsampling_x = subsampling_x;
4249 cm->subsampling_y = subsampling_y;
4250 #if CONFIG_VP9_HIGHBITDEPTH
4251 cm->use_highbitdepth = use_highbitdepth;
4254 alloc_raw_frame_buffers(cpi);
4255 init_ref_frame_bufs(cm);
4256 alloc_util_frame_buffers(cpi);
4258 init_motion_estimation(cpi); // TODO(agrange) This can be removed.
4260 cpi->initial_width = cm->width;
4261 cpi->initial_height = cm->height;
4262 cpi->initial_mbs = cm->MBs;
4266 int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
4267 YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
4269 VP9_COMMON *const cm = &cpi->common;
4270 struct vpx_usec_timer timer;
4272 const int subsampling_x = sd->subsampling_x;
4273 const int subsampling_y = sd->subsampling_y;
4274 #if CONFIG_VP9_HIGHBITDEPTH
4275 const int use_highbitdepth = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
4278 #if CONFIG_VP9_HIGHBITDEPTH
4279 check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
4281 check_initial_width(cpi, subsampling_x, subsampling_y);
4282 #endif // CONFIG_VP9_HIGHBITDEPTH
4284 #if CONFIG_VP9_TEMPORAL_DENOISING
4285 setup_denoiser_buffer(cpi);
4287 vpx_usec_timer_start(&timer);
4289 if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
4290 #if CONFIG_VP9_HIGHBITDEPTH
4292 #endif // CONFIG_VP9_HIGHBITDEPTH
4295 vpx_usec_timer_mark(&timer);
4296 cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
4298 if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
4299 (subsampling_x != 1 || subsampling_y != 1)) {
4300 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4301 "Non-4:2:0 color format requires profile 1 or 3");
4304 if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
4305 (subsampling_x == 1 && subsampling_y == 1)) {
4306 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4307 "4:2:0 color format requires profile 0 or 2");
4315 static int frame_is_reference(const VP9_COMP *cpi) {
4316 const VP9_COMMON *cm = &cpi->common;
4318 return cm->frame_type == KEY_FRAME ||
4319 cpi->refresh_last_frame ||
4320 cpi->refresh_golden_frame ||
4321 cpi->refresh_alt_ref_frame ||
4322 cm->refresh_frame_context ||
4323 cm->lf.mode_ref_delta_update ||
4324 cm->seg.update_map ||
4325 cm->seg.update_data;
4328 static void adjust_frame_rate(VP9_COMP *cpi,
4329 const struct lookahead_entry *source) {
4330 int64_t this_duration;
4333 if (source->ts_start == cpi->first_time_stamp_ever) {
4334 this_duration = source->ts_end - source->ts_start;
4337 int64_t last_duration = cpi->last_end_time_stamp_seen
4338 - cpi->last_time_stamp_seen;
4340 this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
4342 // do a step update if the duration changes by 10%
4344 step = (int)((this_duration - last_duration) * 10 / last_duration);
4347 if (this_duration) {
4349 vp9_new_framerate(cpi, 10000000.0 / this_duration);
4351 // Average this frame's rate into the last second's average
4352 // frame rate. If we haven't seen 1 second yet, then average
4353 // over the whole interval seen.
4354 const double interval = VPXMIN(
4355 (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
4356 double avg_duration = 10000000.0 / cpi->framerate;
4357 avg_duration *= (interval - avg_duration + this_duration);
4358 avg_duration /= interval;
4360 vp9_new_framerate(cpi, 10000000.0 / avg_duration);
4363 cpi->last_time_stamp_seen = source->ts_start;
4364 cpi->last_end_time_stamp_seen = source->ts_end;
4367 // Returns 0 if this is not an alt ref else the offset of the source frame
4368 // used as the arf midpoint.
4369 static int get_arf_src_index(VP9_COMP *cpi) {
4370 RATE_CONTROL *const rc = &cpi->rc;
4371 int arf_src_index = 0;
4372 if (is_altref_enabled(cpi)) {
4373 if (cpi->oxcf.pass == 2) {
4374 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4375 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
4376 arf_src_index = gf_group->arf_src_offset[gf_group->index];
4378 } else if (rc->source_alt_ref_pending) {
4379 arf_src_index = rc->frames_till_gf_update_due;
4382 return arf_src_index;
4385 static void check_src_altref(VP9_COMP *cpi,
4386 const struct lookahead_entry *source) {
4387 RATE_CONTROL *const rc = &cpi->rc;
4389 if (cpi->oxcf.pass == 2) {
4390 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4391 rc->is_src_frame_alt_ref =
4392 (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
4394 rc->is_src_frame_alt_ref = cpi->alt_ref_source &&
4395 (source == cpi->alt_ref_source);
4398 if (rc->is_src_frame_alt_ref) {
4399 // Current frame is an ARF overlay frame.
4400 cpi->alt_ref_source = NULL;
4402 // Don't refresh the last buffer for an ARF overlay frame. It will
4403 // become the GF so preserve last as an alternative prediction option.
4404 cpi->refresh_last_frame = 0;
4408 #if CONFIG_INTERNAL_STATS
4409 extern double vp9_get_blockiness(const uint8_t *img1, int img1_pitch,
4410 const uint8_t *img2, int img2_pitch,
4411 int width, int height);
4413 static void adjust_image_stat(double y, double u, double v, double all,
4418 s->stat[ALL] += all;
4419 s->worst = VPXMIN(s->worst, all);
4421 #endif // CONFIG_INTERNAL_STATS
4423 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
4424 size_t *size, uint8_t *dest,
4425 int64_t *time_stamp, int64_t *time_end, int flush) {
4426 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4427 VP9_COMMON *const cm = &cpi->common;
4428 BufferPool *const pool = cm->buffer_pool;
4429 RATE_CONTROL *const rc = &cpi->rc;
4430 struct vpx_usec_timer cmptimer;
4431 YV12_BUFFER_CONFIG *force_src_buffer = NULL;
4432 struct lookahead_entry *last_source = NULL;
4433 struct lookahead_entry *source = NULL;
4437 if (is_two_pass_svc(cpi)) {
4438 #if CONFIG_SPATIAL_SVC
4439 vp9_svc_start_frame(cpi);
4440 // Use a small empty frame instead of a real frame
4441 if (cpi->svc.encode_empty_frame_state == ENCODING)
4442 source = &cpi->svc.empty_frame;
4444 if (oxcf->pass == 2)
4445 vp9_restore_layer_context(cpi);
4446 } else if (is_one_pass_cbr_svc(cpi)) {
4447 vp9_one_pass_cbr_svc_start_layer(cpi);
4450 vpx_usec_timer_start(&cmptimer);
4452 vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
4454 // Is multi-arf enabled.
4455 // Note that at the moment multi_arf is only configured for 2 pass VBR and
4456 // will not work properly with svc.
4457 if ((oxcf->pass == 2) && !cpi->use_svc &&
4458 (cpi->oxcf.enable_auto_arf > 1))
4459 cpi->multi_arf_allowed = 1;
4461 cpi->multi_arf_allowed = 0;
4464 cm->reset_frame_context = 0;
4465 cm->refresh_frame_context = 1;
4466 if (!is_one_pass_cbr_svc(cpi)) {
4467 cpi->refresh_last_frame = 1;
4468 cpi->refresh_golden_frame = 0;
4469 cpi->refresh_alt_ref_frame = 0;
4472 // Should we encode an arf frame.
4473 arf_src_index = get_arf_src_index(cpi);
4475 // Skip alt frame if we encode the empty frame
4476 if (is_two_pass_svc(cpi) && source != NULL)
4479 if (arf_src_index) {
4480 for (i = 0; i <= arf_src_index; ++i) {
4481 struct lookahead_entry *e = vp9_lookahead_peek(cpi->lookahead, i);
4482 // Avoid creating an alt-ref if there's a forced keyframe pending.
4485 } else if (e->flags == VPX_EFLAG_FORCE_KF) {
4493 if (arf_src_index) {
4494 assert(arf_src_index <= rc->frames_to_key);
4496 if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
4497 cpi->alt_ref_source = source;
4499 #if CONFIG_SPATIAL_SVC
4500 if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) {
4502 // Reference a hidden frame from a lower layer
4503 for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) {
4504 if (oxcf->ss_enable_auto_arf[i]) {
4505 cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx;
4510 cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1;
4513 if ((oxcf->arnr_max_frames > 0) && (oxcf->arnr_strength > 0)) {
4514 // Produce the filtered ARF frame.
4515 vp9_temporal_filter(cpi, arf_src_index);
4516 vpx_extend_frame_borders(&cpi->alt_ref_buffer);
4517 force_src_buffer = &cpi->alt_ref_buffer;
4522 cpi->refresh_alt_ref_frame = 1;
4523 cpi->refresh_golden_frame = 0;
4524 cpi->refresh_last_frame = 0;
4525 rc->is_src_frame_alt_ref = 0;
4526 rc->source_alt_ref_pending = 0;
4528 rc->source_alt_ref_pending = 0;
4533 // Get last frame source.
4534 if (cm->current_video_frame > 0) {
4535 if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
4539 // Read in the source frame.
4541 source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
4543 source = vp9_lookahead_pop(cpi->lookahead, flush);
4545 if (source != NULL) {
4548 // if the flags indicate intra frame, but if the current picture is for
4549 // non-zero spatial layer, it should not be an intra picture.
4550 // TODO(Won Kap): this needs to change if per-layer intra frame is
4552 if ((source->flags & VPX_EFLAG_FORCE_KF) &&
4553 cpi->svc.spatial_layer_id > cpi->svc.first_spatial_layer_to_encode) {
4554 source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF);
4557 // Check to see if the frame should be encoded as an arf overlay.
4558 check_src_altref(cpi, source);
4563 cpi->un_scaled_source = cpi->Source = force_src_buffer ? force_src_buffer
4566 cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
4568 *time_stamp = source->ts_start;
4569 *time_end = source->ts_end;
4570 *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
4574 if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
4575 vp9_end_first_pass(cpi); /* get last stats packet */
4576 cpi->twopass.first_pass_done = 1;
4581 if (source->ts_start < cpi->first_time_stamp_ever) {
4582 cpi->first_time_stamp_ever = source->ts_start;
4583 cpi->last_end_time_stamp_seen = source->ts_start;
4586 // Clear down mmx registers
4587 vpx_clear_system_state();
4589 // adjust frame rates based on timestamps given
4590 if (cm->show_frame) {
4591 adjust_frame_rate(cpi, source);
4594 if (is_one_pass_cbr_svc(cpi)) {
4595 vp9_update_temporal_layer_framerate(cpi);
4596 vp9_restore_layer_context(cpi);
4599 // Find a free buffer for the new frame, releasing the reference previously
4601 if (cm->new_fb_idx != INVALID_IDX) {
4602 --pool->frame_bufs[cm->new_fb_idx].ref_count;
4604 cm->new_fb_idx = get_free_fb(cm);
4606 if (cm->new_fb_idx == INVALID_IDX)
4609 cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
4611 if (!cpi->use_svc && cpi->multi_arf_allowed) {
4612 if (cm->frame_type == KEY_FRAME) {
4613 init_buffer_indices(cpi);
4614 } else if (oxcf->pass == 2) {
4615 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4616 cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
4620 // Start with a 0 size frame.
4623 cpi->frame_flags = *frame_flags;
4625 if ((oxcf->pass == 2) &&
4627 (is_two_pass_svc(cpi) &&
4628 cpi->svc.encode_empty_frame_state != ENCODING))) {
4629 vp9_rc_get_second_pass_params(cpi);
4630 } else if (oxcf->pass == 1) {
4631 set_frame_size(cpi);
4634 if (cpi->oxcf.pass != 0 ||
4636 frame_is_intra_only(cm) == 1) {
4637 for (i = 0; i < MAX_REF_FRAMES; ++i)
4638 cpi->scaled_ref_idx[i] = INVALID_IDX;
4641 if (oxcf->pass == 1 &&
4642 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4643 const int lossless = is_lossless_requested(oxcf);
4644 #if CONFIG_VP9_HIGHBITDEPTH
4645 if (cpi->oxcf.use_highbitdepth)
4646 cpi->td.mb.fwd_txm4x4 = lossless ?
4647 vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
4649 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4650 cpi->td.mb.highbd_itxm_add = lossless ? vp9_highbd_iwht4x4_add :
4651 vp9_highbd_idct4x4_add;
4653 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4654 #endif // CONFIG_VP9_HIGHBITDEPTH
4655 cpi->td.mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
4656 vp9_first_pass(cpi, source);
4657 } else if (oxcf->pass == 2 &&
4658 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4659 Pass2Encode(cpi, size, dest, frame_flags);
4660 } else if (cpi->use_svc) {
4661 SvcEncode(cpi, size, dest, frame_flags);
4664 Pass0Encode(cpi, size, dest, frame_flags);
4667 if (cm->refresh_frame_context)
4668 cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
4670 // No frame encoded, or frame was dropped, release scaled references.
4671 if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
4672 release_scaled_references(cpi);
4676 cpi->droppable = !frame_is_reference(cpi);
4679 // Save layer specific state.
4680 if (is_one_pass_cbr_svc(cpi) ||
4681 ((cpi->svc.number_temporal_layers > 1 ||
4682 cpi->svc.number_spatial_layers > 1) &&
4684 vp9_save_layer_context(cpi);
4687 vpx_usec_timer_mark(&cmptimer);
4688 cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
4690 if (cpi->b_calculate_psnr && oxcf->pass != 1 && cm->show_frame)
4691 generate_psnr_packet(cpi);
4693 #if CONFIG_INTERNAL_STATS
4695 if (oxcf->pass != 1) {
4696 double samples = 0.0;
4697 cpi->bytes += (int)(*size);
4699 if (cm->show_frame) {
4702 if (cpi->b_calculate_psnr) {
4703 YV12_BUFFER_CONFIG *orig = cpi->Source;
4704 YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
4705 YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer;
4707 #if CONFIG_VP9_HIGHBITDEPTH
4708 calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd,
4709 cpi->oxcf.input_bit_depth);
4711 calc_psnr(orig, recon, &psnr);
4712 #endif // CONFIG_VP9_HIGHBITDEPTH
4714 adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3],
4715 psnr.psnr[0], &cpi->psnr);
4716 cpi->total_sq_error += psnr.sse[0];
4717 cpi->total_samples += psnr.samples[0];
4718 samples = psnr.samples[0];
4722 double frame_ssim2 = 0, weight = 0;
4723 #if CONFIG_VP9_POSTPROC
4724 if (vpx_alloc_frame_buffer(&cm->post_proc_buffer,
4725 recon->y_crop_width, recon->y_crop_height,
4726 cm->subsampling_x, cm->subsampling_y,
4727 #if CONFIG_VP9_HIGHBITDEPTH
4728 cm->use_highbitdepth,
4730 VP9_ENC_BORDER_IN_PIXELS,
4731 cm->byte_alignment) < 0) {
4732 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
4733 "Failed to allocate post processing buffer");
4736 vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer,
4737 cm->lf.filter_level * 10 / 6);
4739 vpx_clear_system_state();
4741 #if CONFIG_VP9_HIGHBITDEPTH
4742 calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd,
4743 cpi->oxcf.input_bit_depth);
4745 calc_psnr(orig, pp, &psnr2);
4746 #endif // CONFIG_VP9_HIGHBITDEPTH
4748 cpi->totalp_sq_error += psnr2.sse[0];
4749 cpi->totalp_samples += psnr2.samples[0];
4750 adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3],
4751 psnr2.psnr[0], &cpi->psnrp);
4753 #if CONFIG_VP9_HIGHBITDEPTH
4754 if (cm->use_highbitdepth) {
4755 frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight,
4756 (int)cm->bit_depth);
4758 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4761 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4762 #endif // CONFIG_VP9_HIGHBITDEPTH
4764 cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2);
4765 cpi->summed_quality += frame_ssim2 * weight;
4766 cpi->summed_weights += weight;
4768 #if CONFIG_VP9_HIGHBITDEPTH
4769 if (cm->use_highbitdepth) {
4770 frame_ssim2 = vpx_highbd_calc_ssim(
4771 orig, &cm->post_proc_buffer, &weight, (int)cm->bit_depth);
4773 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4776 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4777 #endif // CONFIG_VP9_HIGHBITDEPTH
4779 cpi->summedp_quality += frame_ssim2 * weight;
4780 cpi->summedp_weights += weight;
4783 FILE *f = fopen("q_used.stt", "a");
4784 fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
4785 cpi->common.current_video_frame, y2, u2, v2,
4786 frame_psnr2, frame_ssim2);
4792 if (cpi->b_calculate_blockiness) {
4793 #if CONFIG_VP9_HIGHBITDEPTH
4794 if (!cm->use_highbitdepth)
4797 double frame_blockiness = vp9_get_blockiness(
4798 cpi->Source->y_buffer, cpi->Source->y_stride,
4799 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4800 cpi->Source->y_width, cpi->Source->y_height);
4801 cpi->worst_blockiness =
4802 VPXMAX(cpi->worst_blockiness, frame_blockiness);
4803 cpi->total_blockiness += frame_blockiness;
4807 if (cpi->b_calculate_consistency) {
4808 #if CONFIG_VP9_HIGHBITDEPTH
4809 if (!cm->use_highbitdepth)
4812 double this_inconsistency = vpx_get_ssim_metrics(
4813 cpi->Source->y_buffer, cpi->Source->y_stride,
4814 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4815 cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
4818 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
4819 double consistency = vpx_sse_to_psnr(samples, peak,
4820 (double)cpi->total_inconsistency);
4821 if (consistency > 0.0)
4822 cpi->worst_consistency =
4823 VPXMIN(cpi->worst_consistency, consistency);
4824 cpi->total_inconsistency += this_inconsistency;
4828 if (cpi->b_calculate_ssimg) {
4829 double y, u, v, frame_all;
4830 #if CONFIG_VP9_HIGHBITDEPTH
4831 if (cm->use_highbitdepth) {
4832 frame_all = vpx_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y,
4833 &u, &v, (int)cm->bit_depth);
4835 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u,
4839 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v);
4840 #endif // CONFIG_VP9_HIGHBITDEPTH
4841 adjust_image_stat(y, u, v, frame_all, &cpi->ssimg);
4843 #if CONFIG_VP9_HIGHBITDEPTH
4844 if (!cm->use_highbitdepth)
4847 double y, u, v, frame_all;
4848 frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
4850 adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
4851 /* TODO(JBB): add 10/12 bit support */
4853 #if CONFIG_VP9_HIGHBITDEPTH
4854 if (!cm->use_highbitdepth)
4857 double y, u, v, frame_all;
4858 frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
4859 adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
4866 if (is_two_pass_svc(cpi)) {
4867 if (cpi->svc.encode_empty_frame_state == ENCODING) {
4868 cpi->svc.encode_empty_frame_state = ENCODED;
4869 cpi->svc.encode_intra_empty_frame = 0;
4872 if (cm->show_frame) {
4873 ++cpi->svc.spatial_layer_to_encode;
4874 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
4875 cpi->svc.spatial_layer_to_encode = 0;
4877 // May need the empty frame after an visible frame.
4878 cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE;
4880 } else if (is_one_pass_cbr_svc(cpi)) {
4881 if (cm->show_frame) {
4882 ++cpi->svc.spatial_layer_to_encode;
4883 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
4884 cpi->svc.spatial_layer_to_encode = 0;
4887 vpx_clear_system_state();
4891 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
4892 vp9_ppflags_t *flags) {
4893 VP9_COMMON *cm = &cpi->common;
4894 #if !CONFIG_VP9_POSTPROC
4898 if (!cm->show_frame) {
4902 #if CONFIG_VP9_POSTPROC
4903 ret = vp9_post_proc_frame(cm, dest, flags);
4905 if (cm->frame_to_show) {
4906 *dest = *cm->frame_to_show;
4907 dest->y_width = cm->width;
4908 dest->y_height = cm->height;
4909 dest->uv_width = cm->width >> cm->subsampling_x;
4910 dest->uv_height = cm->height >> cm->subsampling_y;
4915 #endif // !CONFIG_VP9_POSTPROC
4916 vpx_clear_system_state();
4921 int vp9_set_internal_size(VP9_COMP *cpi,
4922 VPX_SCALING horiz_mode, VPX_SCALING vert_mode) {
4923 VP9_COMMON *cm = &cpi->common;
4924 int hr = 0, hs = 0, vr = 0, vs = 0;
4926 if (horiz_mode > ONETWO || vert_mode > ONETWO)
4929 Scale2Ratio(horiz_mode, &hr, &hs);
4930 Scale2Ratio(vert_mode, &vr, &vs);
4932 // always go to the next whole number
4933 cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
4934 cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
4935 if (cm->current_video_frame) {
4936 assert(cm->width <= cpi->initial_width);
4937 assert(cm->height <= cpi->initial_height);
4940 update_frame_size(cpi);
4945 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
4946 unsigned int height) {
4947 VP9_COMMON *cm = &cpi->common;
4948 #if CONFIG_VP9_HIGHBITDEPTH
4949 check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
4951 check_initial_width(cpi, 1, 1);
4952 #endif // CONFIG_VP9_HIGHBITDEPTH
4954 #if CONFIG_VP9_TEMPORAL_DENOISING
4955 setup_denoiser_buffer(cpi);
4960 if (cm->width > cpi->initial_width) {
4961 cm->width = cpi->initial_width;
4962 printf("Warning: Desired width too large, changed to %d\n", cm->width);
4967 cm->height = height;
4968 if (cm->height > cpi->initial_height) {
4969 cm->height = cpi->initial_height;
4970 printf("Warning: Desired height too large, changed to %d\n", cm->height);
4973 assert(cm->width <= cpi->initial_width);
4974 assert(cm->height <= cpi->initial_height);
4976 update_frame_size(cpi);
4981 void vp9_set_svc(VP9_COMP *cpi, int use_svc) {
4982 cpi->use_svc = use_svc;
4986 int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a,
4987 const YV12_BUFFER_CONFIG *b) {
4988 assert(a->y_crop_width == b->y_crop_width);
4989 assert(a->y_crop_height == b->y_crop_height);
4991 return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
4992 a->y_crop_width, a->y_crop_height);
4995 #if CONFIG_VP9_HIGHBITDEPTH
4996 int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
4997 const YV12_BUFFER_CONFIG *b) {
4998 assert(a->y_crop_width == b->y_crop_width);
4999 assert(a->y_crop_height == b->y_crop_height);
5000 assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
5001 assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
5003 return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
5004 a->y_crop_width, a->y_crop_height);
5006 #endif // CONFIG_VP9_HIGHBITDEPTH
5008 int vp9_get_quantizer(VP9_COMP *cpi) {
5009 return cpi->common.base_qindex;
5012 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
5013 if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF |
5014 VP8_EFLAG_NO_REF_ARF)) {
5017 if (flags & VP8_EFLAG_NO_REF_LAST)
5018 ref ^= VP9_LAST_FLAG;
5020 if (flags & VP8_EFLAG_NO_REF_GF)
5021 ref ^= VP9_GOLD_FLAG;
5023 if (flags & VP8_EFLAG_NO_REF_ARF)
5024 ref ^= VP9_ALT_FLAG;
5026 vp9_use_as_reference(cpi, ref);
5029 if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
5030 VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF |
5031 VP8_EFLAG_FORCE_ARF)) {
5034 if (flags & VP8_EFLAG_NO_UPD_LAST)
5035 upd ^= VP9_LAST_FLAG;
5037 if (flags & VP8_EFLAG_NO_UPD_GF)
5038 upd ^= VP9_GOLD_FLAG;
5040 if (flags & VP8_EFLAG_NO_UPD_ARF)
5041 upd ^= VP9_ALT_FLAG;
5043 vp9_update_reference(cpi, upd);
5046 if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
5047 vp9_update_entropy(cpi, 0);