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_complexity.h"
40 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
41 #include "vp9/encoder/vp9_aq_variance.h"
42 #include "vp9/encoder/vp9_bitstream.h"
43 #include "vp9/encoder/vp9_context_tree.h"
44 #include "vp9/encoder/vp9_encodeframe.h"
45 #include "vp9/encoder/vp9_encodemv.h"
46 #include "vp9/encoder/vp9_encoder.h"
47 #include "vp9/encoder/vp9_ethread.h"
48 #include "vp9/encoder/vp9_firstpass.h"
49 #include "vp9/encoder/vp9_mbgraph.h"
50 #include "vp9/encoder/vp9_picklpf.h"
51 #include "vp9/encoder/vp9_ratectrl.h"
52 #include "vp9/encoder/vp9_rd.h"
53 #include "vp9/encoder/vp9_resize.h"
54 #include "vp9/encoder/vp9_segmentation.h"
55 #include "vp9/encoder/vp9_skin_detection.h"
56 #include "vp9/encoder/vp9_speed_features.h"
57 #include "vp9/encoder/vp9_svc_layercontext.h"
58 #include "vp9/encoder/vp9_temporal_filter.h"
60 #define AM_SEGMENT_ID_INACTIVE 7
61 #define AM_SEGMENT_ID_ACTIVE 0
63 #define SHARP_FILTER_QTHRESH 0 /* Q threshold for 8-tap sharp filter */
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 vp9_free_ref_frame_buffers(cm->buffer_pool);
379 #if CONFIG_VP9_POSTPROC
380 vp9_free_postproc_buffers(cm);
382 vp9_free_context_buffers(cm);
384 vpx_free_frame_buffer(&cpi->last_frame_uf);
385 vpx_free_frame_buffer(&cpi->scaled_source);
386 vpx_free_frame_buffer(&cpi->scaled_last_source);
387 vpx_free_frame_buffer(&cpi->alt_ref_buffer);
388 vp9_lookahead_destroy(cpi->lookahead);
390 vpx_free(cpi->tile_tok[0][0]);
391 cpi->tile_tok[0][0] = 0;
393 vp9_free_pc_tree(&cpi->td);
395 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
396 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
397 vpx_free(lc->rc_twopass_stats_in.buf);
398 lc->rc_twopass_stats_in.buf = NULL;
399 lc->rc_twopass_stats_in.sz = 0;
402 if (cpi->source_diff_var != NULL) {
403 vpx_free(cpi->source_diff_var);
404 cpi->source_diff_var = NULL;
407 for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
408 vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]);
410 memset(&cpi->svc.scaled_frames[0], 0,
411 MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
413 vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
414 memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
416 vp9_free_svc_cyclic_refresh(cpi);
419 static void save_coding_context(VP9_COMP *cpi) {
420 CODING_CONTEXT *const cc = &cpi->coding_context;
421 VP9_COMMON *cm = &cpi->common;
423 // Stores a snapshot of key state variables which can subsequently be
424 // restored with a call to vp9_restore_coding_context. These functions are
425 // intended for use in a re-code loop in vp9_compress_frame where the
426 // quantizer value is adjusted between loop iterations.
427 vp9_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost);
429 memcpy(cc->nmvcosts[0], cpi->nmvcosts[0],
430 MV_VALS * sizeof(*cpi->nmvcosts[0]));
431 memcpy(cc->nmvcosts[1], cpi->nmvcosts[1],
432 MV_VALS * sizeof(*cpi->nmvcosts[1]));
433 memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0],
434 MV_VALS * sizeof(*cpi->nmvcosts_hp[0]));
435 memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1],
436 MV_VALS * sizeof(*cpi->nmvcosts_hp[1]));
438 vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs);
440 memcpy(cpi->coding_context.last_frame_seg_map_copy,
441 cm->last_frame_seg_map, (cm->mi_rows * cm->mi_cols));
443 vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
444 vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
449 static void restore_coding_context(VP9_COMP *cpi) {
450 CODING_CONTEXT *const cc = &cpi->coding_context;
451 VP9_COMMON *cm = &cpi->common;
453 // Restore key state variables to the snapshot state stored in the
454 // previous call to vp9_save_coding_context.
455 vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
457 memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0]));
458 memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1]));
459 memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0],
460 MV_VALS * sizeof(*cc->nmvcosts_hp[0]));
461 memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1],
462 MV_VALS * sizeof(*cc->nmvcosts_hp[1]));
464 vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs);
466 memcpy(cm->last_frame_seg_map,
467 cpi->coding_context.last_frame_seg_map_copy,
468 (cm->mi_rows * cm->mi_cols));
470 vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
471 vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
476 static void configure_static_seg_features(VP9_COMP *cpi) {
477 VP9_COMMON *const cm = &cpi->common;
478 const RATE_CONTROL *const rc = &cpi->rc;
479 struct segmentation *const seg = &cm->seg;
481 int high_q = (int)(rc->avg_q > 48.0);
484 // Disable and clear down for KF
485 if (cm->frame_type == KEY_FRAME) {
486 // Clear down the global segmentation map
487 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
489 seg->update_data = 0;
490 cpi->static_mb_pct = 0;
492 // Disable segmentation
493 vp9_disable_segmentation(seg);
495 // Clear down the segment features.
496 vp9_clearall_segfeatures(seg);
497 } else if (cpi->refresh_alt_ref_frame) {
498 // If this is an alt ref frame
499 // Clear down the global segmentation map
500 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
502 seg->update_data = 0;
503 cpi->static_mb_pct = 0;
505 // Disable segmentation and individual segment features by default
506 vp9_disable_segmentation(seg);
507 vp9_clearall_segfeatures(seg);
509 // Scan frames from current to arf frame.
510 // This function re-enables segmentation if appropriate.
511 vp9_update_mbgraph_stats(cpi);
513 // If segmentation was enabled set those features needed for the
517 seg->update_data = 1;
519 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875,
521 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
522 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
524 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
525 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
527 // Where relevant assume segment data is delta data
528 seg->abs_delta = SEGMENT_DELTADATA;
530 } else if (seg->enabled) {
531 // All other frames if segmentation has been enabled
533 // First normal frame in a valid gf or alt ref group
534 if (rc->frames_since_golden == 0) {
535 // Set up segment features for normal frames in an arf group
536 if (rc->source_alt_ref_active) {
538 seg->update_data = 1;
539 seg->abs_delta = SEGMENT_DELTADATA;
541 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125,
543 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
544 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
546 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
547 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
549 // Segment coding disabled for compred testing
550 if (high_q || (cpi->static_mb_pct == 100)) {
551 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
552 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
553 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
556 // Disable segmentation and clear down features if alt ref
557 // is not active for this group
559 vp9_disable_segmentation(seg);
561 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
564 seg->update_data = 0;
566 vp9_clearall_segfeatures(seg);
568 } else if (rc->is_src_frame_alt_ref) {
569 // Special case where we are coding over the top of a previous
571 // Segment coding disabled for compred testing
573 // Enable ref frame features for segment 0 as well
574 vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
575 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
577 // All mbs should use ALTREF_FRAME
578 vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
579 vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
580 vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
581 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
583 // Skip all MBs if high Q (0,0 mv and skip coeffs)
585 vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP);
586 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
588 // Enable data update
589 seg->update_data = 1;
593 // No updates.. leave things as they are.
595 seg->update_data = 0;
600 static void update_reference_segmentation_map(VP9_COMP *cpi) {
601 VP9_COMMON *const cm = &cpi->common;
602 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
603 uint8_t *cache_ptr = cm->last_frame_seg_map;
606 for (row = 0; row < cm->mi_rows; row++) {
607 MODE_INFO **mi_8x8 = mi_8x8_ptr;
608 uint8_t *cache = cache_ptr;
609 for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
610 cache[0] = mi_8x8[0]->mbmi.segment_id;
611 mi_8x8_ptr += cm->mi_stride;
612 cache_ptr += cm->mi_cols;
616 static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
617 VP9_COMMON *cm = &cpi->common;
618 const VP9EncoderConfig *oxcf = &cpi->oxcf;
621 cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
622 cm->subsampling_x, cm->subsampling_y,
623 #if CONFIG_VP9_HIGHBITDEPTH
624 cm->use_highbitdepth,
626 oxcf->lag_in_frames);
628 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
629 "Failed to allocate lag buffers");
631 // TODO(agrange) Check if ARF is enabled and skip allocation if not.
632 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
633 oxcf->width, oxcf->height,
634 cm->subsampling_x, cm->subsampling_y,
635 #if CONFIG_VP9_HIGHBITDEPTH
636 cm->use_highbitdepth,
638 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
640 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
641 "Failed to allocate altref buffer");
644 static void alloc_util_frame_buffers(VP9_COMP *cpi) {
645 VP9_COMMON *const cm = &cpi->common;
646 if (vpx_realloc_frame_buffer(&cpi->last_frame_uf,
647 cm->width, cm->height,
648 cm->subsampling_x, cm->subsampling_y,
649 #if CONFIG_VP9_HIGHBITDEPTH
650 cm->use_highbitdepth,
652 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
654 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
655 "Failed to allocate last frame buffer");
657 if (vpx_realloc_frame_buffer(&cpi->scaled_source,
658 cm->width, cm->height,
659 cm->subsampling_x, cm->subsampling_y,
660 #if CONFIG_VP9_HIGHBITDEPTH
661 cm->use_highbitdepth,
663 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
665 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
666 "Failed to allocate scaled source buffer");
668 if (vpx_realloc_frame_buffer(&cpi->scaled_last_source,
669 cm->width, cm->height,
670 cm->subsampling_x, cm->subsampling_y,
671 #if CONFIG_VP9_HIGHBITDEPTH
672 cm->use_highbitdepth,
674 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
676 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
677 "Failed to allocate scaled last source buffer");
681 static int alloc_context_buffers_ext(VP9_COMP *cpi) {
682 VP9_COMMON *cm = &cpi->common;
683 int mi_size = cm->mi_cols * cm->mi_rows;
685 cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
686 if (!cpi->mbmi_ext_base)
692 static void alloc_compressor_data(VP9_COMP *cpi) {
693 VP9_COMMON *cm = &cpi->common;
695 vp9_alloc_context_buffers(cm, cm->width, cm->height);
697 alloc_context_buffers_ext(cpi);
699 vpx_free(cpi->tile_tok[0][0]);
702 unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
703 CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
704 vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
707 vp9_setup_pc_tree(&cpi->common, &cpi->td);
710 void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
711 cpi->framerate = framerate < 0.1 ? 30 : framerate;
712 vp9_rc_update_framerate(cpi);
715 static void set_tile_limits(VP9_COMP *cpi) {
716 VP9_COMMON *const cm = &cpi->common;
718 int min_log2_tile_cols, max_log2_tile_cols;
719 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
721 if (is_two_pass_svc(cpi) &&
722 (cpi->svc.encode_empty_frame_state == ENCODING ||
723 cpi->svc.number_spatial_layers > 1)) {
724 cm->log2_tile_cols = 0;
725 cm->log2_tile_rows = 0;
727 cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns,
728 min_log2_tile_cols, max_log2_tile_cols);
729 cm->log2_tile_rows = cpi->oxcf.tile_rows;
733 static void update_frame_size(VP9_COMP *cpi) {
734 VP9_COMMON *const cm = &cpi->common;
735 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
737 vp9_set_mb_mi(cm, cm->width, cm->height);
738 vp9_init_context_buffers(cm);
739 vp9_init_macroblockd(cm, xd, NULL);
740 cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base;
741 memset(cpi->mbmi_ext_base, 0,
742 cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
744 set_tile_limits(cpi);
746 if (is_two_pass_svc(cpi)) {
747 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer,
748 cm->width, cm->height,
749 cm->subsampling_x, cm->subsampling_y,
750 #if CONFIG_VP9_HIGHBITDEPTH
751 cm->use_highbitdepth,
753 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
755 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
756 "Failed to reallocate alt_ref_buffer");
760 static void init_buffer_indices(VP9_COMP *cpi) {
766 static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
767 VP9_COMMON *const cm = &cpi->common;
770 cpi->framerate = oxcf->init_framerate;
772 cm->profile = oxcf->profile;
773 cm->bit_depth = oxcf->bit_depth;
774 #if CONFIG_VP9_HIGHBITDEPTH
775 cm->use_highbitdepth = oxcf->use_highbitdepth;
777 cm->color_space = oxcf->color_space;
778 cm->color_range = oxcf->color_range;
780 cm->width = oxcf->width;
781 cm->height = oxcf->height;
782 alloc_compressor_data(cpi);
784 cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode;
786 // Single thread case: use counts in common.
787 cpi->td.counts = &cm->counts;
789 // Spatial scalability.
790 cpi->svc.number_spatial_layers = oxcf->ss_number_layers;
791 // Temporal scalability.
792 cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
794 if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
795 ((cpi->svc.number_temporal_layers > 1 ||
796 cpi->svc.number_spatial_layers > 1) &&
797 cpi->oxcf.pass != 1)) {
798 vp9_init_layer_context(cpi);
801 // change includes all joint functionality
802 vp9_change_config(cpi, oxcf);
804 cpi->static_mb_pct = 0;
805 cpi->ref_frame_flags = 0;
807 init_buffer_indices(cpi);
810 static void set_rc_buffer_sizes(RATE_CONTROL *rc,
811 const VP9EncoderConfig *oxcf) {
812 const int64_t bandwidth = oxcf->target_bandwidth;
813 const int64_t starting = oxcf->starting_buffer_level_ms;
814 const int64_t optimal = oxcf->optimal_buffer_level_ms;
815 const int64_t maximum = oxcf->maximum_buffer_size_ms;
817 rc->starting_buffer_level = starting * bandwidth / 1000;
818 rc->optimal_buffer_level = (optimal == 0) ? bandwidth / 8
819 : optimal * bandwidth / 1000;
820 rc->maximum_buffer_size = (maximum == 0) ? bandwidth / 8
821 : maximum * bandwidth / 1000;
824 #if CONFIG_VP9_HIGHBITDEPTH
825 #define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
826 cpi->fn_ptr[BT].sdf = SDF; \
827 cpi->fn_ptr[BT].sdaf = SDAF; \
828 cpi->fn_ptr[BT].vf = VF; \
829 cpi->fn_ptr[BT].svf = SVF; \
830 cpi->fn_ptr[BT].svaf = SVAF; \
831 cpi->fn_ptr[BT].sdx3f = SDX3F; \
832 cpi->fn_ptr[BT].sdx8f = SDX8F; \
833 cpi->fn_ptr[BT].sdx4df = SDX4DF;
835 #define MAKE_BFP_SAD_WRAPPER(fnname) \
836 static unsigned int fnname##_bits8(const uint8_t *src_ptr, \
838 const uint8_t *ref_ptr, \
840 return fnname(src_ptr, source_stride, ref_ptr, ref_stride); \
842 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
844 const uint8_t *ref_ptr, \
846 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2; \
848 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
850 const uint8_t *ref_ptr, \
852 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4; \
855 #define MAKE_BFP_SADAVG_WRAPPER(fnname) static unsigned int \
856 fnname##_bits8(const uint8_t *src_ptr, \
858 const uint8_t *ref_ptr, \
860 const uint8_t *second_pred) { \
861 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred); \
863 static unsigned int fnname##_bits10(const uint8_t *src_ptr, \
865 const uint8_t *ref_ptr, \
867 const uint8_t *second_pred) { \
868 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
871 static unsigned int fnname##_bits12(const uint8_t *src_ptr, \
873 const uint8_t *ref_ptr, \
875 const uint8_t *second_pred) { \
876 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \
880 #define MAKE_BFP_SAD3_WRAPPER(fnname) \
881 static void fnname##_bits8(const uint8_t *src_ptr, \
883 const uint8_t *ref_ptr, \
885 unsigned int *sad_array) { \
886 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
888 static void fnname##_bits10(const uint8_t *src_ptr, \
890 const uint8_t *ref_ptr, \
892 unsigned int *sad_array) { \
894 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
895 for (i = 0; i < 3; i++) \
896 sad_array[i] >>= 2; \
898 static void fnname##_bits12(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] >>= 4; \
909 #define MAKE_BFP_SAD8_WRAPPER(fnname) \
910 static void fnname##_bits8(const uint8_t *src_ptr, \
912 const uint8_t *ref_ptr, \
914 unsigned int *sad_array) { \
915 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
917 static void fnname##_bits10(const uint8_t *src_ptr, \
919 const uint8_t *ref_ptr, \
921 unsigned int *sad_array) { \
923 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
924 for (i = 0; i < 8; i++) \
925 sad_array[i] >>= 2; \
927 static void fnname##_bits12(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] >>= 4; \
937 #define MAKE_BFP_SAD4D_WRAPPER(fnname) \
938 static void fnname##_bits8(const uint8_t *src_ptr, \
940 const uint8_t* const ref_ptr[], \
942 unsigned int *sad_array) { \
943 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
945 static void fnname##_bits10(const uint8_t *src_ptr, \
947 const uint8_t* const ref_ptr[], \
949 unsigned int *sad_array) { \
951 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
952 for (i = 0; i < 4; i++) \
953 sad_array[i] >>= 2; \
955 static void fnname##_bits12(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] >>= 4; \
966 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16)
967 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg)
968 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d)
969 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32)
970 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg)
971 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d)
972 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32)
973 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg)
974 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d)
975 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64)
976 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg)
977 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d)
978 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32)
979 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg)
980 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad32x32x3)
981 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad32x32x8)
982 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d)
983 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64)
984 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg)
985 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad64x64x3)
986 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad64x64x8)
987 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d)
988 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16)
989 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg)
990 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x16x3)
991 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x16x8)
992 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d)
993 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8)
994 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg)
995 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x8x3)
996 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x8x8)
997 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d)
998 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16)
999 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg)
1000 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x16x3)
1001 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x16x8)
1002 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d)
1003 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8)
1004 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg)
1005 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x8x3)
1006 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x8x8)
1007 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d)
1008 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4)
1009 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg)
1010 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x4x8)
1011 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d)
1012 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8)
1013 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg)
1014 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x8x8)
1015 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d)
1016 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4)
1017 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg)
1018 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad4x4x3)
1019 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x4x8)
1020 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d)
1022 static void highbd_set_var_fns(VP9_COMP *const cpi) {
1023 VP9_COMMON *const cm = &cpi->common;
1024 if (cm->use_highbitdepth) {
1025 switch (cm->bit_depth) {
1027 HIGHBD_BFP(BLOCK_32X16,
1028 vpx_highbd_sad32x16_bits8,
1029 vpx_highbd_sad32x16_avg_bits8,
1030 vpx_highbd_8_variance32x16,
1031 vpx_highbd_8_sub_pixel_variance32x16,
1032 vpx_highbd_8_sub_pixel_avg_variance32x16,
1035 vpx_highbd_sad32x16x4d_bits8)
1037 HIGHBD_BFP(BLOCK_16X32,
1038 vpx_highbd_sad16x32_bits8,
1039 vpx_highbd_sad16x32_avg_bits8,
1040 vpx_highbd_8_variance16x32,
1041 vpx_highbd_8_sub_pixel_variance16x32,
1042 vpx_highbd_8_sub_pixel_avg_variance16x32,
1045 vpx_highbd_sad16x32x4d_bits8)
1047 HIGHBD_BFP(BLOCK_64X32,
1048 vpx_highbd_sad64x32_bits8,
1049 vpx_highbd_sad64x32_avg_bits8,
1050 vpx_highbd_8_variance64x32,
1051 vpx_highbd_8_sub_pixel_variance64x32,
1052 vpx_highbd_8_sub_pixel_avg_variance64x32,
1055 vpx_highbd_sad64x32x4d_bits8)
1057 HIGHBD_BFP(BLOCK_32X64,
1058 vpx_highbd_sad32x64_bits8,
1059 vpx_highbd_sad32x64_avg_bits8,
1060 vpx_highbd_8_variance32x64,
1061 vpx_highbd_8_sub_pixel_variance32x64,
1062 vpx_highbd_8_sub_pixel_avg_variance32x64,
1065 vpx_highbd_sad32x64x4d_bits8)
1067 HIGHBD_BFP(BLOCK_32X32,
1068 vpx_highbd_sad32x32_bits8,
1069 vpx_highbd_sad32x32_avg_bits8,
1070 vpx_highbd_8_variance32x32,
1071 vpx_highbd_8_sub_pixel_variance32x32,
1072 vpx_highbd_8_sub_pixel_avg_variance32x32,
1073 vpx_highbd_sad32x32x3_bits8,
1074 vpx_highbd_sad32x32x8_bits8,
1075 vpx_highbd_sad32x32x4d_bits8)
1077 HIGHBD_BFP(BLOCK_64X64,
1078 vpx_highbd_sad64x64_bits8,
1079 vpx_highbd_sad64x64_avg_bits8,
1080 vpx_highbd_8_variance64x64,
1081 vpx_highbd_8_sub_pixel_variance64x64,
1082 vpx_highbd_8_sub_pixel_avg_variance64x64,
1083 vpx_highbd_sad64x64x3_bits8,
1084 vpx_highbd_sad64x64x8_bits8,
1085 vpx_highbd_sad64x64x4d_bits8)
1087 HIGHBD_BFP(BLOCK_16X16,
1088 vpx_highbd_sad16x16_bits8,
1089 vpx_highbd_sad16x16_avg_bits8,
1090 vpx_highbd_8_variance16x16,
1091 vpx_highbd_8_sub_pixel_variance16x16,
1092 vpx_highbd_8_sub_pixel_avg_variance16x16,
1093 vpx_highbd_sad16x16x3_bits8,
1094 vpx_highbd_sad16x16x8_bits8,
1095 vpx_highbd_sad16x16x4d_bits8)
1097 HIGHBD_BFP(BLOCK_16X8,
1098 vpx_highbd_sad16x8_bits8,
1099 vpx_highbd_sad16x8_avg_bits8,
1100 vpx_highbd_8_variance16x8,
1101 vpx_highbd_8_sub_pixel_variance16x8,
1102 vpx_highbd_8_sub_pixel_avg_variance16x8,
1103 vpx_highbd_sad16x8x3_bits8,
1104 vpx_highbd_sad16x8x8_bits8,
1105 vpx_highbd_sad16x8x4d_bits8)
1107 HIGHBD_BFP(BLOCK_8X16,
1108 vpx_highbd_sad8x16_bits8,
1109 vpx_highbd_sad8x16_avg_bits8,
1110 vpx_highbd_8_variance8x16,
1111 vpx_highbd_8_sub_pixel_variance8x16,
1112 vpx_highbd_8_sub_pixel_avg_variance8x16,
1113 vpx_highbd_sad8x16x3_bits8,
1114 vpx_highbd_sad8x16x8_bits8,
1115 vpx_highbd_sad8x16x4d_bits8)
1117 HIGHBD_BFP(BLOCK_8X8,
1118 vpx_highbd_sad8x8_bits8,
1119 vpx_highbd_sad8x8_avg_bits8,
1120 vpx_highbd_8_variance8x8,
1121 vpx_highbd_8_sub_pixel_variance8x8,
1122 vpx_highbd_8_sub_pixel_avg_variance8x8,
1123 vpx_highbd_sad8x8x3_bits8,
1124 vpx_highbd_sad8x8x8_bits8,
1125 vpx_highbd_sad8x8x4d_bits8)
1127 HIGHBD_BFP(BLOCK_8X4,
1128 vpx_highbd_sad8x4_bits8,
1129 vpx_highbd_sad8x4_avg_bits8,
1130 vpx_highbd_8_variance8x4,
1131 vpx_highbd_8_sub_pixel_variance8x4,
1132 vpx_highbd_8_sub_pixel_avg_variance8x4,
1134 vpx_highbd_sad8x4x8_bits8,
1135 vpx_highbd_sad8x4x4d_bits8)
1137 HIGHBD_BFP(BLOCK_4X8,
1138 vpx_highbd_sad4x8_bits8,
1139 vpx_highbd_sad4x8_avg_bits8,
1140 vpx_highbd_8_variance4x8,
1141 vpx_highbd_8_sub_pixel_variance4x8,
1142 vpx_highbd_8_sub_pixel_avg_variance4x8,
1144 vpx_highbd_sad4x8x8_bits8,
1145 vpx_highbd_sad4x8x4d_bits8)
1147 HIGHBD_BFP(BLOCK_4X4,
1148 vpx_highbd_sad4x4_bits8,
1149 vpx_highbd_sad4x4_avg_bits8,
1150 vpx_highbd_8_variance4x4,
1151 vpx_highbd_8_sub_pixel_variance4x4,
1152 vpx_highbd_8_sub_pixel_avg_variance4x4,
1153 vpx_highbd_sad4x4x3_bits8,
1154 vpx_highbd_sad4x4x8_bits8,
1155 vpx_highbd_sad4x4x4d_bits8)
1159 HIGHBD_BFP(BLOCK_32X16,
1160 vpx_highbd_sad32x16_bits10,
1161 vpx_highbd_sad32x16_avg_bits10,
1162 vpx_highbd_10_variance32x16,
1163 vpx_highbd_10_sub_pixel_variance32x16,
1164 vpx_highbd_10_sub_pixel_avg_variance32x16,
1167 vpx_highbd_sad32x16x4d_bits10)
1169 HIGHBD_BFP(BLOCK_16X32,
1170 vpx_highbd_sad16x32_bits10,
1171 vpx_highbd_sad16x32_avg_bits10,
1172 vpx_highbd_10_variance16x32,
1173 vpx_highbd_10_sub_pixel_variance16x32,
1174 vpx_highbd_10_sub_pixel_avg_variance16x32,
1177 vpx_highbd_sad16x32x4d_bits10)
1179 HIGHBD_BFP(BLOCK_64X32,
1180 vpx_highbd_sad64x32_bits10,
1181 vpx_highbd_sad64x32_avg_bits10,
1182 vpx_highbd_10_variance64x32,
1183 vpx_highbd_10_sub_pixel_variance64x32,
1184 vpx_highbd_10_sub_pixel_avg_variance64x32,
1187 vpx_highbd_sad64x32x4d_bits10)
1189 HIGHBD_BFP(BLOCK_32X64,
1190 vpx_highbd_sad32x64_bits10,
1191 vpx_highbd_sad32x64_avg_bits10,
1192 vpx_highbd_10_variance32x64,
1193 vpx_highbd_10_sub_pixel_variance32x64,
1194 vpx_highbd_10_sub_pixel_avg_variance32x64,
1197 vpx_highbd_sad32x64x4d_bits10)
1199 HIGHBD_BFP(BLOCK_32X32,
1200 vpx_highbd_sad32x32_bits10,
1201 vpx_highbd_sad32x32_avg_bits10,
1202 vpx_highbd_10_variance32x32,
1203 vpx_highbd_10_sub_pixel_variance32x32,
1204 vpx_highbd_10_sub_pixel_avg_variance32x32,
1205 vpx_highbd_sad32x32x3_bits10,
1206 vpx_highbd_sad32x32x8_bits10,
1207 vpx_highbd_sad32x32x4d_bits10)
1209 HIGHBD_BFP(BLOCK_64X64,
1210 vpx_highbd_sad64x64_bits10,
1211 vpx_highbd_sad64x64_avg_bits10,
1212 vpx_highbd_10_variance64x64,
1213 vpx_highbd_10_sub_pixel_variance64x64,
1214 vpx_highbd_10_sub_pixel_avg_variance64x64,
1215 vpx_highbd_sad64x64x3_bits10,
1216 vpx_highbd_sad64x64x8_bits10,
1217 vpx_highbd_sad64x64x4d_bits10)
1219 HIGHBD_BFP(BLOCK_16X16,
1220 vpx_highbd_sad16x16_bits10,
1221 vpx_highbd_sad16x16_avg_bits10,
1222 vpx_highbd_10_variance16x16,
1223 vpx_highbd_10_sub_pixel_variance16x16,
1224 vpx_highbd_10_sub_pixel_avg_variance16x16,
1225 vpx_highbd_sad16x16x3_bits10,
1226 vpx_highbd_sad16x16x8_bits10,
1227 vpx_highbd_sad16x16x4d_bits10)
1229 HIGHBD_BFP(BLOCK_16X8,
1230 vpx_highbd_sad16x8_bits10,
1231 vpx_highbd_sad16x8_avg_bits10,
1232 vpx_highbd_10_variance16x8,
1233 vpx_highbd_10_sub_pixel_variance16x8,
1234 vpx_highbd_10_sub_pixel_avg_variance16x8,
1235 vpx_highbd_sad16x8x3_bits10,
1236 vpx_highbd_sad16x8x8_bits10,
1237 vpx_highbd_sad16x8x4d_bits10)
1239 HIGHBD_BFP(BLOCK_8X16,
1240 vpx_highbd_sad8x16_bits10,
1241 vpx_highbd_sad8x16_avg_bits10,
1242 vpx_highbd_10_variance8x16,
1243 vpx_highbd_10_sub_pixel_variance8x16,
1244 vpx_highbd_10_sub_pixel_avg_variance8x16,
1245 vpx_highbd_sad8x16x3_bits10,
1246 vpx_highbd_sad8x16x8_bits10,
1247 vpx_highbd_sad8x16x4d_bits10)
1249 HIGHBD_BFP(BLOCK_8X8,
1250 vpx_highbd_sad8x8_bits10,
1251 vpx_highbd_sad8x8_avg_bits10,
1252 vpx_highbd_10_variance8x8,
1253 vpx_highbd_10_sub_pixel_variance8x8,
1254 vpx_highbd_10_sub_pixel_avg_variance8x8,
1255 vpx_highbd_sad8x8x3_bits10,
1256 vpx_highbd_sad8x8x8_bits10,
1257 vpx_highbd_sad8x8x4d_bits10)
1259 HIGHBD_BFP(BLOCK_8X4,
1260 vpx_highbd_sad8x4_bits10,
1261 vpx_highbd_sad8x4_avg_bits10,
1262 vpx_highbd_10_variance8x4,
1263 vpx_highbd_10_sub_pixel_variance8x4,
1264 vpx_highbd_10_sub_pixel_avg_variance8x4,
1266 vpx_highbd_sad8x4x8_bits10,
1267 vpx_highbd_sad8x4x4d_bits10)
1269 HIGHBD_BFP(BLOCK_4X8,
1270 vpx_highbd_sad4x8_bits10,
1271 vpx_highbd_sad4x8_avg_bits10,
1272 vpx_highbd_10_variance4x8,
1273 vpx_highbd_10_sub_pixel_variance4x8,
1274 vpx_highbd_10_sub_pixel_avg_variance4x8,
1276 vpx_highbd_sad4x8x8_bits10,
1277 vpx_highbd_sad4x8x4d_bits10)
1279 HIGHBD_BFP(BLOCK_4X4,
1280 vpx_highbd_sad4x4_bits10,
1281 vpx_highbd_sad4x4_avg_bits10,
1282 vpx_highbd_10_variance4x4,
1283 vpx_highbd_10_sub_pixel_variance4x4,
1284 vpx_highbd_10_sub_pixel_avg_variance4x4,
1285 vpx_highbd_sad4x4x3_bits10,
1286 vpx_highbd_sad4x4x8_bits10,
1287 vpx_highbd_sad4x4x4d_bits10)
1291 HIGHBD_BFP(BLOCK_32X16,
1292 vpx_highbd_sad32x16_bits12,
1293 vpx_highbd_sad32x16_avg_bits12,
1294 vpx_highbd_12_variance32x16,
1295 vpx_highbd_12_sub_pixel_variance32x16,
1296 vpx_highbd_12_sub_pixel_avg_variance32x16,
1299 vpx_highbd_sad32x16x4d_bits12)
1301 HIGHBD_BFP(BLOCK_16X32,
1302 vpx_highbd_sad16x32_bits12,
1303 vpx_highbd_sad16x32_avg_bits12,
1304 vpx_highbd_12_variance16x32,
1305 vpx_highbd_12_sub_pixel_variance16x32,
1306 vpx_highbd_12_sub_pixel_avg_variance16x32,
1309 vpx_highbd_sad16x32x4d_bits12)
1311 HIGHBD_BFP(BLOCK_64X32,
1312 vpx_highbd_sad64x32_bits12,
1313 vpx_highbd_sad64x32_avg_bits12,
1314 vpx_highbd_12_variance64x32,
1315 vpx_highbd_12_sub_pixel_variance64x32,
1316 vpx_highbd_12_sub_pixel_avg_variance64x32,
1319 vpx_highbd_sad64x32x4d_bits12)
1321 HIGHBD_BFP(BLOCK_32X64,
1322 vpx_highbd_sad32x64_bits12,
1323 vpx_highbd_sad32x64_avg_bits12,
1324 vpx_highbd_12_variance32x64,
1325 vpx_highbd_12_sub_pixel_variance32x64,
1326 vpx_highbd_12_sub_pixel_avg_variance32x64,
1329 vpx_highbd_sad32x64x4d_bits12)
1331 HIGHBD_BFP(BLOCK_32X32,
1332 vpx_highbd_sad32x32_bits12,
1333 vpx_highbd_sad32x32_avg_bits12,
1334 vpx_highbd_12_variance32x32,
1335 vpx_highbd_12_sub_pixel_variance32x32,
1336 vpx_highbd_12_sub_pixel_avg_variance32x32,
1337 vpx_highbd_sad32x32x3_bits12,
1338 vpx_highbd_sad32x32x8_bits12,
1339 vpx_highbd_sad32x32x4d_bits12)
1341 HIGHBD_BFP(BLOCK_64X64,
1342 vpx_highbd_sad64x64_bits12,
1343 vpx_highbd_sad64x64_avg_bits12,
1344 vpx_highbd_12_variance64x64,
1345 vpx_highbd_12_sub_pixel_variance64x64,
1346 vpx_highbd_12_sub_pixel_avg_variance64x64,
1347 vpx_highbd_sad64x64x3_bits12,
1348 vpx_highbd_sad64x64x8_bits12,
1349 vpx_highbd_sad64x64x4d_bits12)
1351 HIGHBD_BFP(BLOCK_16X16,
1352 vpx_highbd_sad16x16_bits12,
1353 vpx_highbd_sad16x16_avg_bits12,
1354 vpx_highbd_12_variance16x16,
1355 vpx_highbd_12_sub_pixel_variance16x16,
1356 vpx_highbd_12_sub_pixel_avg_variance16x16,
1357 vpx_highbd_sad16x16x3_bits12,
1358 vpx_highbd_sad16x16x8_bits12,
1359 vpx_highbd_sad16x16x4d_bits12)
1361 HIGHBD_BFP(BLOCK_16X8,
1362 vpx_highbd_sad16x8_bits12,
1363 vpx_highbd_sad16x8_avg_bits12,
1364 vpx_highbd_12_variance16x8,
1365 vpx_highbd_12_sub_pixel_variance16x8,
1366 vpx_highbd_12_sub_pixel_avg_variance16x8,
1367 vpx_highbd_sad16x8x3_bits12,
1368 vpx_highbd_sad16x8x8_bits12,
1369 vpx_highbd_sad16x8x4d_bits12)
1371 HIGHBD_BFP(BLOCK_8X16,
1372 vpx_highbd_sad8x16_bits12,
1373 vpx_highbd_sad8x16_avg_bits12,
1374 vpx_highbd_12_variance8x16,
1375 vpx_highbd_12_sub_pixel_variance8x16,
1376 vpx_highbd_12_sub_pixel_avg_variance8x16,
1377 vpx_highbd_sad8x16x3_bits12,
1378 vpx_highbd_sad8x16x8_bits12,
1379 vpx_highbd_sad8x16x4d_bits12)
1381 HIGHBD_BFP(BLOCK_8X8,
1382 vpx_highbd_sad8x8_bits12,
1383 vpx_highbd_sad8x8_avg_bits12,
1384 vpx_highbd_12_variance8x8,
1385 vpx_highbd_12_sub_pixel_variance8x8,
1386 vpx_highbd_12_sub_pixel_avg_variance8x8,
1387 vpx_highbd_sad8x8x3_bits12,
1388 vpx_highbd_sad8x8x8_bits12,
1389 vpx_highbd_sad8x8x4d_bits12)
1391 HIGHBD_BFP(BLOCK_8X4,
1392 vpx_highbd_sad8x4_bits12,
1393 vpx_highbd_sad8x4_avg_bits12,
1394 vpx_highbd_12_variance8x4,
1395 vpx_highbd_12_sub_pixel_variance8x4,
1396 vpx_highbd_12_sub_pixel_avg_variance8x4,
1398 vpx_highbd_sad8x4x8_bits12,
1399 vpx_highbd_sad8x4x4d_bits12)
1401 HIGHBD_BFP(BLOCK_4X8,
1402 vpx_highbd_sad4x8_bits12,
1403 vpx_highbd_sad4x8_avg_bits12,
1404 vpx_highbd_12_variance4x8,
1405 vpx_highbd_12_sub_pixel_variance4x8,
1406 vpx_highbd_12_sub_pixel_avg_variance4x8,
1408 vpx_highbd_sad4x8x8_bits12,
1409 vpx_highbd_sad4x8x4d_bits12)
1411 HIGHBD_BFP(BLOCK_4X4,
1412 vpx_highbd_sad4x4_bits12,
1413 vpx_highbd_sad4x4_avg_bits12,
1414 vpx_highbd_12_variance4x4,
1415 vpx_highbd_12_sub_pixel_variance4x4,
1416 vpx_highbd_12_sub_pixel_avg_variance4x4,
1417 vpx_highbd_sad4x4x3_bits12,
1418 vpx_highbd_sad4x4x8_bits12,
1419 vpx_highbd_sad4x4x4d_bits12)
1423 assert(0 && "cm->bit_depth should be VPX_BITS_8, "
1424 "VPX_BITS_10 or VPX_BITS_12");
1428 #endif // CONFIG_VP9_HIGHBITDEPTH
1430 static void realloc_segmentation_maps(VP9_COMP *cpi) {
1431 VP9_COMMON *const cm = &cpi->common;
1433 // Create the encoder segmentation map and set all entries to 0
1434 vpx_free(cpi->segmentation_map);
1435 CHECK_MEM_ERROR(cm, cpi->segmentation_map,
1436 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1438 // Create a map used for cyclic background refresh.
1439 if (cpi->cyclic_refresh)
1440 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
1441 CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
1442 vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
1444 // Create a map used to mark inactive areas.
1445 vpx_free(cpi->active_map.map);
1446 CHECK_MEM_ERROR(cm, cpi->active_map.map,
1447 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1449 // And a place holder structure is the coding context
1450 // for use if we want to save and restore it
1451 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
1452 CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
1453 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1456 void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
1457 VP9_COMMON *const cm = &cpi->common;
1458 RATE_CONTROL *const rc = &cpi->rc;
1459 int last_w = cpi->oxcf.width;
1460 int last_h = cpi->oxcf.height;
1462 if (cm->profile != oxcf->profile)
1463 cm->profile = oxcf->profile;
1464 cm->bit_depth = oxcf->bit_depth;
1465 cm->color_space = oxcf->color_space;
1466 cm->color_range = oxcf->color_range;
1468 if (cm->profile <= PROFILE_1)
1469 assert(cm->bit_depth == VPX_BITS_8);
1471 assert(cm->bit_depth > VPX_BITS_8);
1474 #if CONFIG_VP9_HIGHBITDEPTH
1475 cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
1476 #endif // CONFIG_VP9_HIGHBITDEPTH
1478 rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
1480 cpi->refresh_golden_frame = 0;
1481 cpi->refresh_last_frame = 1;
1482 cm->refresh_frame_context = 1;
1483 cm->reset_frame_context = 0;
1485 vp9_reset_segment_features(&cm->seg);
1486 vp9_set_high_precision_mv(cpi, 0);
1491 for (i = 0; i < MAX_SEGMENTS; i++)
1492 cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
1494 cpi->encode_breakout = cpi->oxcf.encode_breakout;
1496 set_rc_buffer_sizes(rc, &cpi->oxcf);
1498 // Under a configuration change, where maximum_buffer_size may change,
1499 // keep buffer level clipped to the maximum allowed buffer size.
1500 rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
1501 rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size);
1503 // Set up frame rate and related parameters rate control values.
1504 vp9_new_framerate(cpi, cpi->framerate);
1506 // Set absolute upper and lower quality limits
1507 rc->worst_quality = cpi->oxcf.worst_allowed_q;
1508 rc->best_quality = cpi->oxcf.best_allowed_q;
1510 cm->interp_filter = cpi->sf.default_interp_filter;
1512 cm->display_width = cpi->oxcf.width;
1513 cm->display_height = cpi->oxcf.height;
1514 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1515 cm->width = cpi->oxcf.width;
1516 cm->height = cpi->oxcf.height;
1519 if (cpi->initial_width) {
1520 if (cm->width > cpi->initial_width || cm->height > cpi->initial_height) {
1521 vp9_free_context_buffers(cm);
1522 alloc_compressor_data(cpi);
1523 realloc_segmentation_maps(cpi);
1524 cpi->initial_width = cpi->initial_height = 0;
1527 update_frame_size(cpi);
1529 if ((cpi->svc.number_temporal_layers > 1 &&
1530 cpi->oxcf.rc_mode == VPX_CBR) ||
1531 ((cpi->svc.number_temporal_layers > 1 ||
1532 cpi->svc.number_spatial_layers > 1) &&
1533 cpi->oxcf.pass != 1)) {
1534 vp9_update_layer_context_change_config(cpi,
1535 (int)cpi->oxcf.target_bandwidth);
1538 cpi->alt_ref_source = NULL;
1539 rc->is_src_frame_alt_ref = 0;
1542 // Experimental RD Code
1543 cpi->frame_distortion = 0;
1544 cpi->last_frame_distortion = 0;
1547 set_tile_limits(cpi);
1549 cpi->ext_refresh_frame_flags_pending = 0;
1550 cpi->ext_refresh_frame_context_pending = 0;
1552 #if CONFIG_VP9_HIGHBITDEPTH
1553 highbd_set_var_fns(cpi);
1558 #define M_LOG2_E 0.693147180559945309417
1560 #define log2f(x) (log (x) / (float) M_LOG2_E)
1562 static void cal_nmvjointsadcost(int *mvjointsadcost) {
1563 mvjointsadcost[0] = 600;
1564 mvjointsadcost[1] = 300;
1565 mvjointsadcost[2] = 300;
1566 mvjointsadcost[3] = 300;
1569 static void cal_nmvsadcosts(int *mvsadcost[2]) {
1572 mvsadcost[0][0] = 0;
1573 mvsadcost[1][0] = 0;
1576 double z = 256 * (2 * (log2f(8 * i) + .6));
1577 mvsadcost[0][i] = (int)z;
1578 mvsadcost[1][i] = (int)z;
1579 mvsadcost[0][-i] = (int)z;
1580 mvsadcost[1][-i] = (int)z;
1581 } while (++i <= MV_MAX);
1584 static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
1587 mvsadcost[0][0] = 0;
1588 mvsadcost[1][0] = 0;
1591 double z = 256 * (2 * (log2f(8 * i) + .6));
1592 mvsadcost[0][i] = (int)z;
1593 mvsadcost[1][i] = (int)z;
1594 mvsadcost[0][-i] = (int)z;
1595 mvsadcost[1][-i] = (int)z;
1596 } while (++i <= MV_MAX);
1600 VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
1601 BufferPool *const pool) {
1603 VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
1604 VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
1611 if (setjmp(cm->error.jmp)) {
1612 cm->error.setjmp = 0;
1613 vp9_remove_compressor(cpi);
1617 cm->error.setjmp = 1;
1618 cm->alloc_mi = vp9_enc_alloc_mi;
1619 cm->free_mi = vp9_enc_free_mi;
1620 cm->setup_mi = vp9_enc_setup_mi;
1622 CHECK_MEM_ERROR(cm, cm->fc,
1623 (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
1624 CHECK_MEM_ERROR(cm, cm->frame_contexts,
1625 (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
1626 sizeof(*cm->frame_contexts)));
1629 cpi->resize_state = 0;
1630 cpi->resize_avg_qp = 0;
1631 cpi->resize_buffer_underflow = 0;
1632 cpi->common.buffer_pool = pool;
1634 cpi->rc.high_source_sad = 0;
1636 init_config(cpi, oxcf);
1637 vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
1639 cm->current_video_frame = 0;
1640 cpi->partition_search_skippable_frame = 0;
1641 cpi->tile_data = NULL;
1643 realloc_segmentation_maps(cpi);
1645 CHECK_MEM_ERROR(cm, cpi->nmvcosts[0],
1646 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0])));
1647 CHECK_MEM_ERROR(cm, cpi->nmvcosts[1],
1648 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1])));
1649 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0],
1650 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0])));
1651 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1],
1652 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1])));
1653 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0],
1654 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0])));
1655 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1],
1656 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1])));
1657 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0],
1658 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0])));
1659 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
1660 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
1662 for (i = 0; i < (sizeof(cpi->mbgraph_stats) /
1663 sizeof(cpi->mbgraph_stats[0])); i++) {
1664 CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats,
1665 vpx_calloc(cm->MBs *
1666 sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
1669 #if CONFIG_FP_MB_STATS
1670 cpi->use_fp_mb_stats = 0;
1671 if (cpi->use_fp_mb_stats) {
1672 // a place holder used to store the first pass mb stats in the first pass
1673 CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
1674 vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
1676 cpi->twopass.frame_mb_stats_buf = NULL;
1680 cpi->refresh_alt_ref_frame = 0;
1681 cpi->multi_arf_last_grp_enabled = 0;
1683 cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
1684 #if CONFIG_INTERNAL_STATS
1685 cpi->b_calculate_ssimg = 0;
1686 cpi->b_calculate_blockiness = 1;
1687 cpi->b_calculate_consistency = 1;
1688 cpi->total_inconsistency = 0;
1689 cpi->psnr.worst = 100.0;
1690 cpi->worst_ssim = 100.0;
1695 if (cpi->b_calculate_psnr) {
1696 cpi->total_sq_error = 0;
1697 cpi->total_samples = 0;
1699 cpi->totalp_sq_error = 0;
1700 cpi->totalp_samples = 0;
1702 cpi->tot_recode_hits = 0;
1703 cpi->summed_quality = 0;
1704 cpi->summed_weights = 0;
1705 cpi->summedp_quality = 0;
1706 cpi->summedp_weights = 0;
1709 if (cpi->b_calculate_ssimg) {
1710 cpi->ssimg.worst= 100.0;
1712 cpi->fastssim.worst = 100.0;
1714 cpi->psnrhvs.worst = 100.0;
1716 if (cpi->b_calculate_blockiness) {
1717 cpi->total_blockiness = 0;
1718 cpi->worst_blockiness = 0.0;
1721 if (cpi->b_calculate_consistency) {
1722 cpi->ssim_vars = vpx_malloc(sizeof(*cpi->ssim_vars) *
1723 4 * cpi->common.mi_rows * cpi->common.mi_cols);
1724 cpi->worst_consistency = 100.0;
1729 cpi->first_time_stamp_ever = INT64_MAX;
1731 cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
1732 cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
1733 cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
1734 cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
1735 cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
1736 cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
1738 cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
1739 cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
1740 cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
1741 cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
1742 cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
1744 #if CONFIG_VP9_TEMPORAL_DENOISING
1745 #ifdef OUTPUT_YUV_DENOISED
1746 yuv_denoised_file = fopen("denoised.yuv", "ab");
1749 #ifdef OUTPUT_YUV_SKINMAP
1750 yuv_skinmap_file = fopen("skinmap.yuv", "ab");
1752 #ifdef OUTPUT_YUV_REC
1753 yuv_rec_file = fopen("rec.yuv", "wb");
1757 framepsnr = fopen("framepsnr.stt", "a");
1758 kf_list = fopen("kf_list.stt", "w");
1761 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
1763 if (oxcf->pass == 1) {
1764 vp9_init_first_pass(cpi);
1765 } else if (oxcf->pass == 2) {
1766 const size_t packet_sz = sizeof(FIRSTPASS_STATS);
1767 const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
1769 if (cpi->svc.number_spatial_layers > 1
1770 || cpi->svc.number_temporal_layers > 1) {
1771 FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf;
1772 FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = {0};
1775 for (i = 0; i < oxcf->ss_number_layers; ++i) {
1776 FIRSTPASS_STATS *const last_packet_for_layer =
1777 &stats[packets - oxcf->ss_number_layers + i];
1778 const int layer_id = (int)last_packet_for_layer->spatial_layer_id;
1779 const int packets_in_layer = (int)last_packet_for_layer->count + 1;
1780 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) {
1781 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id];
1783 vpx_free(lc->rc_twopass_stats_in.buf);
1785 lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz;
1786 CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf,
1787 vpx_malloc(lc->rc_twopass_stats_in.sz));
1788 lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf;
1789 lc->twopass.stats_in = lc->twopass.stats_in_start;
1790 lc->twopass.stats_in_end = lc->twopass.stats_in_start
1791 + packets_in_layer - 1;
1792 stats_copy[layer_id] = lc->rc_twopass_stats_in.buf;
1796 for (i = 0; i < packets; ++i) {
1797 const int layer_id = (int)stats[i].spatial_layer_id;
1798 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers
1799 && stats_copy[layer_id] != NULL) {
1800 *stats_copy[layer_id] = stats[i];
1801 ++stats_copy[layer_id];
1805 vp9_init_second_pass_spatial_svc(cpi);
1807 #if CONFIG_FP_MB_STATS
1808 if (cpi->use_fp_mb_stats) {
1809 const size_t psz = cpi->common.MBs * sizeof(uint8_t);
1810 const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
1812 cpi->twopass.firstpass_mb_stats.mb_stats_start =
1813 oxcf->firstpass_mb_stats_in.buf;
1814 cpi->twopass.firstpass_mb_stats.mb_stats_end =
1815 cpi->twopass.firstpass_mb_stats.mb_stats_start +
1816 (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
1820 cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
1821 cpi->twopass.stats_in = cpi->twopass.stats_in_start;
1822 cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
1824 vp9_init_second_pass(cpi);
1828 vp9_set_speed_features_framesize_independent(cpi);
1829 vp9_set_speed_features_framesize_dependent(cpi);
1831 // Allocate memory to store variances for a frame.
1832 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
1833 vpx_calloc(cm->MBs, sizeof(diff)));
1834 cpi->source_var_thresh = 0;
1835 cpi->frames_till_next_var_check = 0;
1837 #define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\
1838 cpi->fn_ptr[BT].sdf = SDF; \
1839 cpi->fn_ptr[BT].sdaf = SDAF; \
1840 cpi->fn_ptr[BT].vf = VF; \
1841 cpi->fn_ptr[BT].svf = SVF; \
1842 cpi->fn_ptr[BT].svaf = SVAF; \
1843 cpi->fn_ptr[BT].sdx3f = SDX3F; \
1844 cpi->fn_ptr[BT].sdx8f = SDX8F; \
1845 cpi->fn_ptr[BT].sdx4df = SDX4DF;
1847 BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg,
1848 vpx_variance32x16, vpx_sub_pixel_variance32x16,
1849 vpx_sub_pixel_avg_variance32x16, NULL, NULL, vpx_sad32x16x4d)
1851 BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg,
1852 vpx_variance16x32, vpx_sub_pixel_variance16x32,
1853 vpx_sub_pixel_avg_variance16x32, NULL, NULL, vpx_sad16x32x4d)
1855 BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg,
1856 vpx_variance64x32, vpx_sub_pixel_variance64x32,
1857 vpx_sub_pixel_avg_variance64x32, NULL, NULL, vpx_sad64x32x4d)
1859 BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg,
1860 vpx_variance32x64, vpx_sub_pixel_variance32x64,
1861 vpx_sub_pixel_avg_variance32x64, NULL, NULL, vpx_sad32x64x4d)
1863 BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg,
1864 vpx_variance32x32, vpx_sub_pixel_variance32x32,
1865 vpx_sub_pixel_avg_variance32x32, vpx_sad32x32x3, vpx_sad32x32x8,
1868 BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg,
1869 vpx_variance64x64, vpx_sub_pixel_variance64x64,
1870 vpx_sub_pixel_avg_variance64x64, vpx_sad64x64x3, vpx_sad64x64x8,
1873 BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg,
1874 vpx_variance16x16, vpx_sub_pixel_variance16x16,
1875 vpx_sub_pixel_avg_variance16x16, vpx_sad16x16x3, vpx_sad16x16x8,
1878 BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg,
1879 vpx_variance16x8, vpx_sub_pixel_variance16x8,
1880 vpx_sub_pixel_avg_variance16x8,
1881 vpx_sad16x8x3, vpx_sad16x8x8, vpx_sad16x8x4d)
1883 BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg,
1884 vpx_variance8x16, vpx_sub_pixel_variance8x16,
1885 vpx_sub_pixel_avg_variance8x16,
1886 vpx_sad8x16x3, vpx_sad8x16x8, vpx_sad8x16x4d)
1888 BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg,
1889 vpx_variance8x8, vpx_sub_pixel_variance8x8,
1890 vpx_sub_pixel_avg_variance8x8,
1891 vpx_sad8x8x3, vpx_sad8x8x8, vpx_sad8x8x4d)
1893 BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg,
1894 vpx_variance8x4, vpx_sub_pixel_variance8x4,
1895 vpx_sub_pixel_avg_variance8x4, NULL, vpx_sad8x4x8, vpx_sad8x4x4d)
1897 BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg,
1898 vpx_variance4x8, vpx_sub_pixel_variance4x8,
1899 vpx_sub_pixel_avg_variance4x8, NULL, vpx_sad4x8x8, vpx_sad4x8x4d)
1901 BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg,
1902 vpx_variance4x4, vpx_sub_pixel_variance4x4,
1903 vpx_sub_pixel_avg_variance4x4,
1904 vpx_sad4x4x3, vpx_sad4x4x8, vpx_sad4x4x4d)
1906 #if CONFIG_VP9_HIGHBITDEPTH
1907 highbd_set_var_fns(cpi);
1910 /* vp9_init_quantizer() is first called here. Add check in
1911 * vp9_frame_init_quantizer() so that vp9_init_quantizer is only
1912 * called later when needed. This will avoid unnecessary calls of
1913 * vp9_init_quantizer() for every frame.
1915 vp9_init_quantizer(cpi);
1917 vp9_loop_filter_init(cm);
1919 cm->error.setjmp = 0;
1923 #define SNPRINT(H, T) \
1924 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
1926 #define SNPRINT2(H, T, V) \
1927 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
1929 void vp9_remove_compressor(VP9_COMP *cpi) {
1930 VP9_COMMON *const cm = &cpi->common;
1937 if (cpi && (cm->current_video_frame > 0)) {
1938 #if CONFIG_INTERNAL_STATS
1939 vpx_clear_system_state();
1941 if (cpi->oxcf.pass != 1) {
1942 char headings[512] = {0};
1943 char results[512] = {0};
1944 FILE *f = fopen("opsnr.stt", "a");
1945 double time_encoded = (cpi->last_end_time_stamp_seen
1946 - cpi->first_time_stamp_ever) / 10000000.000;
1947 double total_encode_time = (cpi->time_receive_data +
1948 cpi->time_compress_data) / 1000.000;
1950 (double)cpi->bytes * (double) 8 / (double)1000 / time_encoded;
1951 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
1953 if (cpi->b_calculate_psnr) {
1954 const double total_psnr =
1955 vpx_sse_to_psnr((double)cpi->total_samples, peak,
1956 (double)cpi->total_sq_error);
1957 const double totalp_psnr =
1958 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
1959 (double)cpi->totalp_sq_error);
1960 const double total_ssim = 100 * pow(cpi->summed_quality /
1961 cpi->summed_weights, 8.0);
1962 const double totalp_ssim = 100 * pow(cpi->summedp_quality /
1963 cpi->summedp_weights, 8.0);
1965 snprintf(headings, sizeof(headings),
1966 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
1967 "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
1968 "WstPsnr\tWstSsim\tWstFast\tWstHVS");
1969 snprintf(results, sizeof(results),
1970 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
1971 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
1972 "%7.3f\t%7.3f\t%7.3f\t%7.3f",
1973 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
1974 cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr,
1975 total_ssim, totalp_ssim,
1976 cpi->fastssim.stat[ALL] / cpi->count,
1977 cpi->psnrhvs.stat[ALL] / cpi->count,
1978 cpi->psnr.worst, cpi->worst_ssim, cpi->fastssim.worst,
1979 cpi->psnrhvs.worst);
1981 if (cpi->b_calculate_blockiness) {
1982 SNPRINT(headings, "\t Block\tWstBlck");
1983 SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
1984 SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
1987 if (cpi->b_calculate_consistency) {
1988 double consistency =
1989 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
1990 (double)cpi->total_inconsistency);
1992 SNPRINT(headings, "\tConsist\tWstCons");
1993 SNPRINT2(results, "\t%7.3f", consistency);
1994 SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
1997 if (cpi->b_calculate_ssimg) {
1998 SNPRINT(headings, "\t SSIMG\tWtSSIMG");
1999 SNPRINT2(results, "\t%7.3f", cpi->ssimg.stat[ALL] / cpi->count);
2000 SNPRINT2(results, "\t%7.3f", cpi->ssimg.worst);
2003 fprintf(f, "%s\t Time\n", headings);
2004 fprintf(f, "%s\t%8.0f\n", results, total_encode_time);
2014 printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
2015 printf("\n_frames recive_data encod_mb_row compress_frame Total\n");
2016 printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
2017 cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
2018 cpi->time_compress_data / 1000,
2019 (cpi->time_receive_data + cpi->time_compress_data) / 1000);
2024 #if CONFIG_VP9_TEMPORAL_DENOISING
2025 vp9_denoiser_free(&(cpi->denoiser));
2028 for (t = 0; t < cpi->num_workers; ++t) {
2029 VPxWorker *const worker = &cpi->workers[t];
2030 EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
2032 // Deallocate allocated threads.
2033 vpx_get_worker_interface()->end(worker);
2035 // Deallocate allocated thread data.
2036 if (t < cpi->num_workers - 1) {
2037 vpx_free(thread_data->td->counts);
2038 vp9_free_pc_tree(thread_data->td);
2039 vpx_free(thread_data->td);
2042 vpx_free(cpi->tile_thr_data);
2043 vpx_free(cpi->workers);
2045 if (cpi->num_workers > 1)
2046 vp9_loop_filter_dealloc(&cpi->lf_row_sync);
2048 dealloc_compressor_data(cpi);
2050 for (i = 0; i < sizeof(cpi->mbgraph_stats) /
2051 sizeof(cpi->mbgraph_stats[0]); ++i) {
2052 vpx_free(cpi->mbgraph_stats[i].mb_stats);
2055 #if CONFIG_FP_MB_STATS
2056 if (cpi->use_fp_mb_stats) {
2057 vpx_free(cpi->twopass.frame_mb_stats_buf);
2058 cpi->twopass.frame_mb_stats_buf = NULL;
2062 vp9_remove_common(cm);
2063 vp9_free_ref_frame_buffers(cm->buffer_pool);
2064 #if CONFIG_VP9_POSTPROC
2065 vp9_free_postproc_buffers(cm);
2069 #if CONFIG_VP9_TEMPORAL_DENOISING
2070 #ifdef OUTPUT_YUV_DENOISED
2071 fclose(yuv_denoised_file);
2074 #ifdef OUTPUT_YUV_SKINMAP
2075 fclose(yuv_skinmap_file);
2077 #ifdef OUTPUT_YUV_REC
2078 fclose(yuv_rec_file);
2095 /* TODO(yaowu): The block_variance calls the unoptimized versions of variance()
2096 * and highbd_8_variance(). It should not.
2098 static void encoder_variance(const uint8_t *a, int a_stride,
2099 const uint8_t *b, int b_stride,
2100 int w, int h, unsigned int *sse, int *sum) {
2106 for (i = 0; i < h; i++) {
2107 for (j = 0; j < w; j++) {
2108 const int diff = a[j] - b[j];
2110 *sse += diff * diff;
2118 #if CONFIG_VP9_HIGHBITDEPTH
2119 static void encoder_highbd_variance64(const uint8_t *a8, int a_stride,
2120 const uint8_t *b8, int b_stride,
2121 int w, int h, uint64_t *sse,
2125 uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2126 uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2130 for (i = 0; i < h; i++) {
2131 for (j = 0; j < w; j++) {
2132 const int diff = a[j] - b[j];
2134 *sse += diff * diff;
2141 static void encoder_highbd_8_variance(const uint8_t *a8, int a_stride,
2142 const uint8_t *b8, int b_stride,
2144 unsigned int *sse, int *sum) {
2145 uint64_t sse_long = 0;
2146 uint64_t sum_long = 0;
2147 encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h,
2148 &sse_long, &sum_long);
2149 *sse = (unsigned int)sse_long;
2150 *sum = (int)sum_long;
2152 #endif // CONFIG_VP9_HIGHBITDEPTH
2154 static int64_t get_sse(const uint8_t *a, int a_stride,
2155 const uint8_t *b, int b_stride,
2156 int width, int height) {
2157 const int dw = width % 16;
2158 const int dh = height % 16;
2159 int64_t total_sse = 0;
2160 unsigned int sse = 0;
2165 encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride,
2166 dw, height, &sse, &sum);
2171 encoder_variance(&a[(height - dh) * a_stride], a_stride,
2172 &b[(height - dh) * b_stride], b_stride,
2173 width - dw, dh, &sse, &sum);
2177 for (y = 0; y < height / 16; ++y) {
2178 const uint8_t *pa = a;
2179 const uint8_t *pb = b;
2180 for (x = 0; x < width / 16; ++x) {
2181 vpx_mse16x16(pa, a_stride, pb, b_stride, &sse);
2195 #if CONFIG_VP9_HIGHBITDEPTH
2196 static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride,
2197 const uint8_t *b8, int b_stride,
2198 int width, int height,
2199 unsigned int input_shift) {
2200 const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
2201 const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
2202 int64_t total_sse = 0;
2204 for (y = 0; y < height; ++y) {
2205 for (x = 0; x < width; ++x) {
2207 diff = (a[x] >> input_shift) - (b[x] >> input_shift);
2208 total_sse += diff * diff;
2216 static int64_t highbd_get_sse(const uint8_t *a, int a_stride,
2217 const uint8_t *b, int b_stride,
2218 int width, int height) {
2219 int64_t total_sse = 0;
2221 const int dw = width % 16;
2222 const int dh = height % 16;
2223 unsigned int sse = 0;
2226 encoder_highbd_8_variance(&a[width - dw], a_stride,
2227 &b[width - dw], b_stride,
2228 dw, height, &sse, &sum);
2232 encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride,
2233 &b[(height - dh) * b_stride], b_stride,
2234 width - dw, dh, &sse, &sum);
2237 for (y = 0; y < height / 16; ++y) {
2238 const uint8_t *pa = a;
2239 const uint8_t *pb = b;
2240 for (x = 0; x < width / 16; ++x) {
2241 vpx_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse);
2251 #endif // CONFIG_VP9_HIGHBITDEPTH
2254 double psnr[4]; // total/y/u/v
2255 uint64_t sse[4]; // total/y/u/v
2256 uint32_t samples[4]; // total/y/u/v
2259 #if CONFIG_VP9_HIGHBITDEPTH
2260 static void calc_highbd_psnr(const YV12_BUFFER_CONFIG *a,
2261 const YV12_BUFFER_CONFIG *b,
2263 unsigned int bit_depth,
2264 unsigned int in_bit_depth) {
2265 const int widths[3] =
2266 {a->y_crop_width, a->uv_crop_width, a->uv_crop_width };
2267 const int heights[3] =
2268 {a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2269 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer };
2270 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2271 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer };
2272 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2274 uint64_t total_sse = 0;
2275 uint32_t total_samples = 0;
2276 const double peak = (double)((1 << in_bit_depth) - 1);
2277 const unsigned int input_shift = bit_depth - in_bit_depth;
2279 for (i = 0; i < 3; ++i) {
2280 const int w = widths[i];
2281 const int h = heights[i];
2282 const uint32_t samples = w * h;
2284 if (a->flags & YV12_FLAG_HIGHBITDEPTH) {
2286 sse = highbd_get_sse_shift(a_planes[i], a_strides[i],
2287 b_planes[i], b_strides[i], w, h,
2290 sse = highbd_get_sse(a_planes[i], a_strides[i],
2291 b_planes[i], b_strides[i], w, h);
2294 sse = get_sse(a_planes[i], a_strides[i],
2295 b_planes[i], b_strides[i],
2298 psnr->sse[1 + i] = sse;
2299 psnr->samples[1 + i] = samples;
2300 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2303 total_samples += samples;
2306 psnr->sse[0] = total_sse;
2307 psnr->samples[0] = total_samples;
2308 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2312 #else // !CONFIG_VP9_HIGHBITDEPTH
2314 static void calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b,
2316 static const double peak = 255.0;
2317 const int widths[3] = {
2318 a->y_crop_width, a->uv_crop_width, a->uv_crop_width};
2319 const int heights[3] = {
2320 a->y_crop_height, a->uv_crop_height, a->uv_crop_height};
2321 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer};
2322 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride};
2323 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer};
2324 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride};
2326 uint64_t total_sse = 0;
2327 uint32_t total_samples = 0;
2329 for (i = 0; i < 3; ++i) {
2330 const int w = widths[i];
2331 const int h = heights[i];
2332 const uint32_t samples = w * h;
2333 const uint64_t sse = get_sse(a_planes[i], a_strides[i],
2334 b_planes[i], b_strides[i],
2336 psnr->sse[1 + i] = sse;
2337 psnr->samples[1 + i] = samples;
2338 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse);
2341 total_samples += samples;
2344 psnr->sse[0] = total_sse;
2345 psnr->samples[0] = total_samples;
2346 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak,
2349 #endif // CONFIG_VP9_HIGHBITDEPTH
2351 static void generate_psnr_packet(VP9_COMP *cpi) {
2352 struct vpx_codec_cx_pkt pkt;
2355 #if CONFIG_VP9_HIGHBITDEPTH
2356 calc_highbd_psnr(cpi->Source, cpi->common.frame_to_show, &psnr,
2357 cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
2359 calc_psnr(cpi->Source, cpi->common.frame_to_show, &psnr);
2362 for (i = 0; i < 4; ++i) {
2363 pkt.data.psnr.samples[i] = psnr.samples[i];
2364 pkt.data.psnr.sse[i] = psnr.sse[i];
2365 pkt.data.psnr.psnr[i] = psnr.psnr[i];
2367 pkt.kind = VPX_CODEC_PSNR_PKT;
2369 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
2370 cpi->svc.number_temporal_layers].psnr_pkt = pkt.data.psnr;
2372 vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
2375 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) {
2376 if (ref_frame_flags > 7)
2379 cpi->ref_frame_flags = ref_frame_flags;
2383 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) {
2384 cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0;
2385 cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0;
2386 cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0;
2387 cpi->ext_refresh_frame_flags_pending = 1;
2390 static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(VP9_COMP *cpi,
2391 VP9_REFFRAME ref_frame_flag) {
2392 MV_REFERENCE_FRAME ref_frame = NONE;
2393 if (ref_frame_flag == VP9_LAST_FLAG)
2394 ref_frame = LAST_FRAME;
2395 else if (ref_frame_flag == VP9_GOLD_FLAG)
2396 ref_frame = GOLDEN_FRAME;
2397 else if (ref_frame_flag == VP9_ALT_FLAG)
2398 ref_frame = ALTREF_FRAME;
2400 return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame);
2403 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2404 YV12_BUFFER_CONFIG *sd) {
2405 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2407 vp8_yv12_copy_frame(cfg, sd);
2414 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2415 YV12_BUFFER_CONFIG *sd) {
2416 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2418 vp8_yv12_copy_frame(sd, cfg);
2425 int vp9_update_entropy(VP9_COMP * cpi, int update) {
2426 cpi->ext_refresh_frame_context = update;
2427 cpi->ext_refresh_frame_context_pending = 1;
2431 #if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP)
2432 // The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
2433 // as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
2434 // not denoise the UV channels at this time. If ever we implement UV channel
2435 // denoising we will have to modify this.
2436 void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
2437 uint8_t *src = s->y_buffer;
2438 int h = s->y_height;
2441 fwrite(src, s->y_width, 1, f);
2449 fwrite(src, s->uv_width, 1, f);
2450 src += s->uv_stride;
2457 fwrite(src, s->uv_width, 1, f);
2458 src += s->uv_stride;
2463 #ifdef OUTPUT_YUV_REC
2464 void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
2465 YV12_BUFFER_CONFIG *s = cm->frame_to_show;
2466 uint8_t *src = s->y_buffer;
2469 #if CONFIG_VP9_HIGHBITDEPTH
2470 if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
2471 uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
2474 fwrite(src16, s->y_width, 2, yuv_rec_file);
2475 src16 += s->y_stride;
2478 src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
2482 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2483 src16 += s->uv_stride;
2486 src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
2490 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2491 src16 += s->uv_stride;
2494 fflush(yuv_rec_file);
2497 #endif // CONFIG_VP9_HIGHBITDEPTH
2500 fwrite(src, s->y_width, 1, yuv_rec_file);
2508 fwrite(src, s->uv_width, 1, yuv_rec_file);
2509 src += s->uv_stride;
2516 fwrite(src, s->uv_width, 1, yuv_rec_file);
2517 src += s->uv_stride;
2520 fflush(yuv_rec_file);
2524 #if CONFIG_VP9_HIGHBITDEPTH
2525 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2526 YV12_BUFFER_CONFIG *dst,
2529 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2530 YV12_BUFFER_CONFIG *dst) {
2531 #endif // CONFIG_VP9_HIGHBITDEPTH
2532 // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
2534 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2535 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2536 const int src_widths[3] = {src->y_crop_width, src->uv_crop_width,
2537 src->uv_crop_width };
2538 const int src_heights[3] = {src->y_crop_height, src->uv_crop_height,
2539 src->uv_crop_height};
2540 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2541 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2542 const int dst_widths[3] = {dst->y_crop_width, dst->uv_crop_width,
2543 dst->uv_crop_width};
2544 const int dst_heights[3] = {dst->y_crop_height, dst->uv_crop_height,
2545 dst->uv_crop_height};
2547 for (i = 0; i < MAX_MB_PLANE; ++i) {
2548 #if CONFIG_VP9_HIGHBITDEPTH
2549 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2550 vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
2551 src_strides[i], dsts[i], dst_heights[i],
2552 dst_widths[i], dst_strides[i], bd);
2554 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2555 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2558 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2559 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2560 #endif // CONFIG_VP9_HIGHBITDEPTH
2562 vpx_extend_frame_borders(dst);
2565 #if CONFIG_VP9_HIGHBITDEPTH
2566 static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
2567 YV12_BUFFER_CONFIG *dst, int bd) {
2569 static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
2570 YV12_BUFFER_CONFIG *dst) {
2571 #endif // CONFIG_VP9_HIGHBITDEPTH
2572 const int src_w = src->y_crop_width;
2573 const int src_h = src->y_crop_height;
2574 const int dst_w = dst->y_crop_width;
2575 const int dst_h = dst->y_crop_height;
2576 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
2577 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
2578 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
2579 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
2580 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2583 for (y = 0; y < dst_h; y += 16) {
2584 for (x = 0; x < dst_w; x += 16) {
2585 for (i = 0; i < MAX_MB_PLANE; ++i) {
2586 const int factor = (i == 0 || i == 3 ? 1 : 2);
2587 const int x_q4 = x * (16 / factor) * src_w / dst_w;
2588 const int y_q4 = y * (16 / factor) * src_h / dst_h;
2589 const int src_stride = src_strides[i];
2590 const int dst_stride = dst_strides[i];
2591 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h *
2592 src_stride + (x / factor) * src_w / dst_w;
2593 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2595 #if CONFIG_VP9_HIGHBITDEPTH
2596 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2597 vpx_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
2598 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2599 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2600 16 / factor, 16 / factor, bd);
2602 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2603 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2604 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2605 16 / factor, 16 / factor);
2608 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2609 kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2610 kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2611 16 / factor, 16 / factor);
2612 #endif // CONFIG_VP9_HIGHBITDEPTH
2617 vpx_extend_frame_borders(dst);
2620 static int scale_down(VP9_COMP *cpi, int q) {
2621 RATE_CONTROL *const rc = &cpi->rc;
2622 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2624 assert(frame_is_kf_gf_arf(cpi));
2626 if (rc->frame_size_selector == UNSCALED &&
2627 q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
2628 const int max_size_thresh = (int)(rate_thresh_mult[SCALE_STEP1]
2629 * VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
2630 scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
2635 // Function to test for conditions that indicate we should loop
2636 // back and recode a frame.
2637 static int recode_loop_test(VP9_COMP *cpi,
2638 int high_limit, int low_limit,
2639 int q, int maxq, int minq) {
2640 const RATE_CONTROL *const rc = &cpi->rc;
2641 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2642 const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
2643 int force_recode = 0;
2645 if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
2646 (cpi->sf.recode_loop == ALLOW_RECODE) ||
2647 (frame_is_kfgfarf &&
2648 (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) {
2649 if (frame_is_kfgfarf &&
2650 (oxcf->resize_mode == RESIZE_DYNAMIC) &&
2651 scale_down(cpi, q)) {
2652 // Code this group at a lower resolution.
2653 cpi->resize_pending = 1;
2657 // TODO(agrange) high_limit could be greater than the scale-down threshold.
2658 if ((rc->projected_frame_size > high_limit && q < maxq) ||
2659 (rc->projected_frame_size < low_limit && q > minq)) {
2661 } else if (cpi->oxcf.rc_mode == VPX_CQ) {
2662 // Deal with frame undershoot and whether or not we are
2663 // below the automatically set cq level.
2664 if (q > oxcf->cq_level &&
2665 rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
2670 return force_recode;
2673 void vp9_update_reference_frames(VP9_COMP *cpi) {
2674 VP9_COMMON * const cm = &cpi->common;
2675 BufferPool *const pool = cm->buffer_pool;
2677 // At this point the new frame has been encoded.
2678 // If any buffer copy / swapping is signaled it should be done here.
2679 if (cm->frame_type == KEY_FRAME) {
2680 ref_cnt_fb(pool->frame_bufs,
2681 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2682 ref_cnt_fb(pool->frame_bufs,
2683 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2684 } else if (vp9_preserve_existing_gf(cpi)) {
2685 // We have decided to preserve the previously existing golden frame as our
2686 // new ARF frame. However, in the short term in function
2687 // vp9_bitstream.c::get_refresh_mask() we left it in the GF slot and, if
2688 // we're updating the GF with the current decoded frame, we save it to the
2689 // ARF slot instead.
2690 // We now have to update the ARF with the current frame and swap gld_fb_idx
2691 // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
2692 // slot and, if we're updating the GF, the current frame becomes the new GF.
2695 ref_cnt_fb(pool->frame_bufs,
2696 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
2698 tmp = cpi->alt_fb_idx;
2699 cpi->alt_fb_idx = cpi->gld_fb_idx;
2700 cpi->gld_fb_idx = tmp;
2702 if (is_two_pass_svc(cpi)) {
2703 cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx;
2704 cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx;
2706 } else { /* For non key/golden frames */
2707 if (cpi->refresh_alt_ref_frame) {
2708 int arf_idx = cpi->alt_fb_idx;
2709 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
2710 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2711 arf_idx = gf_group->arf_update_idx[gf_group->index];
2714 ref_cnt_fb(pool->frame_bufs,
2715 &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
2716 memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
2717 cpi->interp_filter_selected[0],
2718 sizeof(cpi->interp_filter_selected[0]));
2721 if (cpi->refresh_golden_frame) {
2722 ref_cnt_fb(pool->frame_bufs,
2723 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
2724 if (!cpi->rc.is_src_frame_alt_ref)
2725 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2726 cpi->interp_filter_selected[0],
2727 sizeof(cpi->interp_filter_selected[0]));
2729 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2730 cpi->interp_filter_selected[ALTREF_FRAME],
2731 sizeof(cpi->interp_filter_selected[ALTREF_FRAME]));
2735 if (cpi->refresh_last_frame) {
2736 ref_cnt_fb(pool->frame_bufs,
2737 &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
2738 if (!cpi->rc.is_src_frame_alt_ref)
2739 memcpy(cpi->interp_filter_selected[LAST_FRAME],
2740 cpi->interp_filter_selected[0],
2741 sizeof(cpi->interp_filter_selected[0]));
2743 #if CONFIG_VP9_TEMPORAL_DENOISING
2744 if (cpi->oxcf.noise_sensitivity > 0) {
2745 vp9_denoiser_update_frame_info(&cpi->denoiser,
2747 cpi->common.frame_type,
2748 cpi->refresh_alt_ref_frame,
2749 cpi->refresh_golden_frame,
2750 cpi->refresh_last_frame);
2755 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
2756 MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
2757 struct loopfilter *lf = &cm->lf;
2759 lf->filter_level = 0;
2761 struct vpx_usec_timer timer;
2763 vpx_clear_system_state();
2765 vpx_usec_timer_start(&timer);
2767 vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick);
2769 vpx_usec_timer_mark(&timer);
2770 cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
2773 if (lf->filter_level > 0) {
2774 if (cpi->num_workers > 1)
2775 vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
2776 lf->filter_level, 0, 0,
2777 cpi->workers, cpi->num_workers,
2780 vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
2783 vpx_extend_frame_inner_borders(cm->frame_to_show);
2786 static INLINE void alloc_frame_mvs(const VP9_COMMON *cm,
2788 RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
2789 if (new_fb_ptr->mvs == NULL ||
2790 new_fb_ptr->mi_rows < cm->mi_rows ||
2791 new_fb_ptr->mi_cols < cm->mi_cols) {
2792 vpx_free(new_fb_ptr->mvs);
2794 (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
2795 sizeof(*new_fb_ptr->mvs));
2796 new_fb_ptr->mi_rows = cm->mi_rows;
2797 new_fb_ptr->mi_cols = cm->mi_cols;
2801 void vp9_scale_references(VP9_COMP *cpi) {
2802 VP9_COMMON *cm = &cpi->common;
2803 MV_REFERENCE_FRAME ref_frame;
2804 const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
2806 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2807 // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
2808 if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
2809 BufferPool *const pool = cm->buffer_pool;
2810 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi,
2814 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
2818 #if CONFIG_VP9_HIGHBITDEPTH
2819 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
2820 RefCntBuffer *new_fb_ptr = NULL;
2821 int force_scaling = 0;
2822 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
2823 if (new_fb == INVALID_IDX) {
2824 new_fb = get_free_fb(cm);
2827 if (new_fb == INVALID_IDX)
2829 new_fb_ptr = &pool->frame_bufs[new_fb];
2830 if (force_scaling ||
2831 new_fb_ptr->buf.y_crop_width != cm->width ||
2832 new_fb_ptr->buf.y_crop_height != cm->height) {
2833 vpx_realloc_frame_buffer(&new_fb_ptr->buf,
2834 cm->width, cm->height,
2835 cm->subsampling_x, cm->subsampling_y,
2836 cm->use_highbitdepth,
2837 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
2839 scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth);
2840 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
2841 alloc_frame_mvs(cm, new_fb);
2844 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
2845 RefCntBuffer *new_fb_ptr = NULL;
2846 int force_scaling = 0;
2847 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
2848 if (new_fb == INVALID_IDX) {
2849 new_fb = get_free_fb(cm);
2852 if (new_fb == INVALID_IDX)
2854 new_fb_ptr = &pool->frame_bufs[new_fb];
2855 if (force_scaling ||
2856 new_fb_ptr->buf.y_crop_width != cm->width ||
2857 new_fb_ptr->buf.y_crop_height != cm->height) {
2858 vpx_realloc_frame_buffer(&new_fb_ptr->buf,
2859 cm->width, cm->height,
2860 cm->subsampling_x, cm->subsampling_y,
2861 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
2863 scale_and_extend_frame(ref, &new_fb_ptr->buf);
2864 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
2865 alloc_frame_mvs(cm, new_fb);
2867 #endif // CONFIG_VP9_HIGHBITDEPTH
2869 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
2870 RefCntBuffer *const buf = &pool->frame_bufs[buf_idx];
2871 buf->buf.y_crop_width = ref->y_crop_width;
2872 buf->buf.y_crop_height = ref->y_crop_height;
2873 cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
2877 if (cpi->oxcf.pass != 0 || cpi->use_svc)
2878 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
2883 static void release_scaled_references(VP9_COMP *cpi) {
2884 VP9_COMMON *cm = &cpi->common;
2886 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
2887 // Only release scaled references under certain conditions:
2888 // if reference will be updated, or if scaled reference has same resolution.
2890 refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
2891 refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
2892 refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
2893 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
2894 const int idx = cpi->scaled_ref_idx[i - 1];
2895 RefCntBuffer *const buf = idx != INVALID_IDX ?
2896 &cm->buffer_pool->frame_bufs[idx] : NULL;
2897 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
2900 (buf->buf.y_crop_width == ref->y_crop_width &&
2901 buf->buf.y_crop_height == ref->y_crop_height))) {
2903 cpi->scaled_ref_idx[i -1] = INVALID_IDX;
2907 for (i = 0; i < MAX_REF_FRAMES; ++i) {
2908 const int idx = cpi->scaled_ref_idx[i];
2909 RefCntBuffer *const buf = idx != INVALID_IDX ?
2910 &cm->buffer_pool->frame_bufs[idx] : NULL;
2913 cpi->scaled_ref_idx[i] = INVALID_IDX;
2919 static void full_to_model_count(unsigned int *model_count,
2920 unsigned int *full_count) {
2922 model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
2923 model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
2924 model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
2925 for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
2926 model_count[TWO_TOKEN] += full_count[n];
2927 model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
2930 static void full_to_model_counts(vp9_coeff_count_model *model_count,
2931 vp9_coeff_count *full_count) {
2934 for (i = 0; i < PLANE_TYPES; ++i)
2935 for (j = 0; j < REF_TYPES; ++j)
2936 for (k = 0; k < COEF_BANDS; ++k)
2937 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
2938 full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
2941 #if 0 && CONFIG_INTERNAL_STATS
2942 static void output_frame_level_debug_stats(VP9_COMP *cpi) {
2943 VP9_COMMON *const cm = &cpi->common;
2944 FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
2947 vpx_clear_system_state();
2949 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
2951 if (cpi->twopass.total_left_stats.coded_error != 0.0)
2952 fprintf(f, "%10u %dx%d %d %d %10d %10d %10d %10d"
2953 "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
2954 "%10"PRId64" %10"PRId64" %10d "
2955 "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
2956 "%6d %6d %5d %5d %5d "
2957 "%10"PRId64" %10.3lf"
2958 "%10lf %8u %10"PRId64" %10d %10d %10d\n",
2959 cpi->common.current_video_frame,
2960 cm->width, cm->height,
2961 cpi->rc.source_alt_ref_pending,
2962 cpi->rc.source_alt_ref_active,
2963 cpi->rc.this_frame_target,
2964 cpi->rc.projected_frame_size,
2965 cpi->rc.projected_frame_size / cpi->common.MBs,
2966 (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
2967 cpi->rc.vbr_bits_off_target,
2968 cpi->rc.vbr_bits_off_target_fast,
2969 cpi->twopass.extend_minq,
2970 cpi->twopass.extend_minq_fast,
2971 cpi->rc.total_target_vs_actual,
2972 (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
2973 cpi->rc.total_actual_bits, cm->base_qindex,
2974 vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
2975 (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) / 4.0,
2976 vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality,
2979 vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
2980 cpi->refresh_last_frame, cpi->refresh_golden_frame,
2981 cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
2982 cpi->twopass.bits_left,
2983 cpi->twopass.total_left_stats.coded_error,
2984 cpi->twopass.bits_left /
2985 (1 + cpi->twopass.total_left_stats.coded_error),
2986 cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
2987 cpi->twopass.kf_zeromotion_pct,
2988 cpi->twopass.fr_content_type);
2993 FILE *const fmodes = fopen("Modes.stt", "a");
2996 fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
2997 cm->frame_type, cpi->refresh_golden_frame,
2998 cpi->refresh_alt_ref_frame);
3000 for (i = 0; i < MAX_MODES; ++i)
3001 fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
3003 fprintf(fmodes, "\n");
3010 static void set_mv_search_params(VP9_COMP *cpi) {
3011 const VP9_COMMON *const cm = &cpi->common;
3012 const unsigned int max_mv_def = VPXMIN(cm->width, cm->height);
3014 // Default based on max resolution.
3015 cpi->mv_step_param = vp9_init_search_range(max_mv_def);
3017 if (cpi->sf.mv.auto_mv_step_size) {
3018 if (frame_is_intra_only(cm)) {
3019 // Initialize max_mv_magnitude for use in the first INTER frame
3020 // after a key/intra-only frame.
3021 cpi->max_mv_magnitude = max_mv_def;
3023 if (cm->show_frame) {
3024 // Allow mv_steps to correspond to twice the max mv magnitude found
3025 // in the previous frame, capped by the default max_mv_magnitude based
3027 cpi->mv_step_param = vp9_init_search_range(
3028 VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
3030 cpi->max_mv_magnitude = 0;
3035 static void set_size_independent_vars(VP9_COMP *cpi) {
3036 vp9_set_speed_features_framesize_independent(cpi);
3037 vp9_set_rd_speed_thresholds(cpi);
3038 vp9_set_rd_speed_thresholds_sub8x8(cpi);
3039 cpi->common.interp_filter = cpi->sf.default_interp_filter;
3042 static void set_size_dependent_vars(VP9_COMP *cpi, int *q,
3043 int *bottom_index, int *top_index) {
3044 VP9_COMMON *const cm = &cpi->common;
3045 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3047 // Setup variables that depend on the dimensions of the frame.
3048 vp9_set_speed_features_framesize_dependent(cpi);
3050 // Decide q and q bounds.
3051 *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
3053 if (!frame_is_intra_only(cm)) {
3054 vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
3057 // Configure experimental use of segmentation for enhanced coding of
3058 // static regions if indicated.
3059 // Only allowed in the second pass of a two pass encode, as it requires
3060 // lagged coding, and if the relevant speed feature flag is set.
3061 if (oxcf->pass == 2 && cpi->sf.static_segmentation)
3062 configure_static_seg_features(cpi);
3064 #if CONFIG_VP9_POSTPROC
3065 if (oxcf->noise_sensitivity > 0) {
3067 switch (oxcf->noise_sensitivity) {
3085 vp9_denoise(cpi->Source, cpi->Source, l);
3087 #endif // CONFIG_VP9_POSTPROC
3090 static void init_motion_estimation(VP9_COMP *cpi) {
3091 int y_stride = cpi->scaled_source.y_stride;
3093 if (cpi->sf.mv.search_method == NSTEP) {
3094 vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
3095 } else if (cpi->sf.mv.search_method == DIAMOND) {
3096 vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
3100 static void set_frame_size(VP9_COMP *cpi) {
3102 VP9_COMMON *const cm = &cpi->common;
3103 VP9EncoderConfig *const oxcf = &cpi->oxcf;
3104 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
3106 if (oxcf->pass == 2 &&
3107 oxcf->rc_mode == VPX_VBR &&
3108 ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
3109 (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
3110 calculate_coded_size(
3111 cpi, &oxcf->scaled_frame_width, &oxcf->scaled_frame_height);
3113 // There has been a change in frame size.
3114 vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
3115 oxcf->scaled_frame_height);
3118 if (oxcf->pass == 0 &&
3119 oxcf->rc_mode == VPX_CBR &&
3121 oxcf->resize_mode == RESIZE_DYNAMIC) {
3122 if (cpi->resize_pending == 1) {
3123 oxcf->scaled_frame_width =
3124 (cm->width * cpi->resize_scale_num) / cpi->resize_scale_den;
3125 oxcf->scaled_frame_height =
3126 (cm->height * cpi->resize_scale_num) /cpi->resize_scale_den;
3127 } else if (cpi->resize_pending == -1) {
3128 // Go back up to original size.
3129 oxcf->scaled_frame_width = oxcf->width;
3130 oxcf->scaled_frame_height = oxcf->height;
3132 if (cpi->resize_pending != 0) {
3133 // There has been a change in frame size.
3134 vp9_set_size_literal(cpi,
3135 oxcf->scaled_frame_width,
3136 oxcf->scaled_frame_height);
3138 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3139 set_mv_search_params(cpi);
3143 if ((oxcf->pass == 2) &&
3145 (is_two_pass_svc(cpi) &&
3146 cpi->svc.encode_empty_frame_state != ENCODING))) {
3147 vp9_set_target_rate(cpi);
3150 alloc_frame_mvs(cm, cm->new_fb_idx);
3152 // Reset the frame pointers to the current frame size.
3153 vpx_realloc_frame_buffer(get_frame_new_buffer(cm),
3154 cm->width, cm->height,
3155 cm->subsampling_x, cm->subsampling_y,
3156 #if CONFIG_VP9_HIGHBITDEPTH
3157 cm->use_highbitdepth,
3159 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
3162 alloc_util_frame_buffers(cpi);
3163 init_motion_estimation(cpi);
3165 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3166 RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
3167 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3169 ref_buf->idx = buf_idx;
3171 if (buf_idx != INVALID_IDX) {
3172 YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
3174 #if CONFIG_VP9_HIGHBITDEPTH
3175 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3176 buf->y_crop_width, buf->y_crop_height,
3177 cm->width, cm->height,
3178 (buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
3181 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
3182 buf->y_crop_width, buf->y_crop_height,
3183 cm->width, cm->height);
3184 #endif // CONFIG_VP9_HIGHBITDEPTH
3185 if (vp9_is_scaled(&ref_buf->sf))
3186 vpx_extend_frame_borders(buf);
3188 ref_buf->buf = NULL;
3192 set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
3195 static void encode_without_recode_loop(VP9_COMP *cpi,
3198 VP9_COMMON *const cm = &cpi->common;
3199 int q = 0, bottom_index = 0, top_index = 0; // Dummy variables.
3201 vpx_clear_system_state();
3203 set_frame_size(cpi);
3205 cpi->Source = vp9_scale_if_required(cm,
3206 cpi->un_scaled_source,
3207 &cpi->scaled_source,
3208 (cpi->oxcf.pass == 0));
3210 // Avoid scaling last_source unless its needed.
3211 // Last source is currently only used for screen-content mode,
3212 // or if partition_search_type == SOURCE_VAR_BASED_PARTITION.
3213 if (cpi->unscaled_last_source != NULL &&
3214 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3215 cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION))
3216 cpi->Last_Source = vp9_scale_if_required(cm,
3217 cpi->unscaled_last_source,
3218 &cpi->scaled_last_source,
3219 (cpi->oxcf.pass == 0));
3221 if (cpi->oxcf.pass == 0 &&
3222 cpi->oxcf.rc_mode == VPX_CBR &&
3223 cpi->resize_state == 0 &&
3224 cm->frame_type != KEY_FRAME &&
3225 cpi->oxcf.content == VP9E_CONTENT_SCREEN)
3226 vp9_avg_source_sad(cpi);
3228 if (frame_is_intra_only(cm) == 0) {
3229 vp9_scale_references(cpi);
3232 set_size_independent_vars(cpi);
3233 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3235 vp9_set_quantizer(cm, q);
3236 vp9_set_variance_partition_thresholds(cpi, q);
3240 suppress_active_map(cpi);
3241 // Variance adaptive and in frame q adjustment experiments are mutually
3243 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3244 vp9_vaq_frame_setup(cpi);
3245 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3246 vp9_setup_in_frame_q_adj(cpi);
3247 } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3248 vp9_cyclic_refresh_setup(cpi);
3250 apply_active_map(cpi);
3252 // transform / motion compensation build reconstruction frame
3253 vp9_encode_frame(cpi);
3255 // Check if we should drop this frame because of high overshoot.
3256 // Only for frames where high temporal-source sad is detected.
3257 if (cpi->oxcf.pass == 0 &&
3258 cpi->oxcf.rc_mode == VPX_CBR &&
3259 cpi->resize_state == 0 &&
3260 cm->frame_type != KEY_FRAME &&
3261 cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
3262 cpi->rc.high_source_sad == 1) {
3264 // Get an estimate of the encoded frame size.
3265 save_coding_context(cpi);
3266 vp9_pack_bitstream(cpi, dest, size);
3267 restore_coding_context(cpi);
3268 frame_size = (int)(*size) << 3;
3269 // Check if encoded frame will overshoot too much, and if so, set the q and
3270 // adjust some rate control parameters, and return to re-encode the frame.
3271 if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) {
3272 vpx_clear_system_state();
3273 vp9_set_quantizer(cm, q);
3274 vp9_set_variance_partition_thresholds(cpi, q);
3275 suppress_active_map(cpi);
3276 // Turn-off cyclic refresh for re-encoded frame.
3277 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3278 unsigned char *const seg_map = cpi->segmentation_map;
3279 memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
3280 vp9_disable_segmentation(&cm->seg);
3282 apply_active_map(cpi);
3283 vp9_encode_frame(cpi);
3287 // Update some stats from cyclic refresh, and check if we should not update
3288 // golden reference, for non-SVC 1 pass CBR.
3289 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
3290 cm->frame_type != KEY_FRAME &&
3292 cpi->ext_refresh_frame_flags_pending == 0 &&
3293 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR))
3294 vp9_cyclic_refresh_check_golden_update(cpi);
3296 // Update the skip mb flag probabilities based on the distribution
3297 // seen in the last encoder iteration.
3298 // update_base_skip_probs(cpi);
3299 vpx_clear_system_state();
3302 static void encode_with_recode_loop(VP9_COMP *cpi,
3305 VP9_COMMON *const cm = &cpi->common;
3306 RATE_CONTROL *const rc = &cpi->rc;
3307 int bottom_index, top_index;
3309 int loop_at_this_size = 0;
3311 int overshoot_seen = 0;
3312 int undershoot_seen = 0;
3313 int frame_over_shoot_limit;
3314 int frame_under_shoot_limit;
3315 int q = 0, q_low = 0, q_high = 0;
3317 set_size_independent_vars(cpi);
3320 vpx_clear_system_state();
3322 set_frame_size(cpi);
3324 if (loop_count == 0 || cpi->resize_pending != 0) {
3325 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3327 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3328 set_mv_search_params(cpi);
3330 // Reset the loop state for new frame size.
3332 undershoot_seen = 0;
3334 // Reconfiguration for change in frame size has concluded.
3335 cpi->resize_pending = 0;
3337 q_low = bottom_index;
3340 loop_at_this_size = 0;
3343 // Decide frame size bounds first time through.
3344 if (loop_count == 0) {
3345 vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
3346 &frame_under_shoot_limit,
3347 &frame_over_shoot_limit);
3350 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
3351 &cpi->scaled_source,
3352 (cpi->oxcf.pass == 0));
3354 if (cpi->unscaled_last_source != NULL)
3355 cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
3356 &cpi->scaled_last_source,
3357 (cpi->oxcf.pass == 0));
3359 if (frame_is_intra_only(cm) == 0) {
3360 if (loop_count > 0) {
3361 release_scaled_references(cpi);
3363 vp9_scale_references(cpi);
3366 vp9_set_quantizer(cm, q);
3368 if (loop_count == 0)
3371 // Variance adaptive and in frame q adjustment experiments are mutually
3373 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3374 vp9_vaq_frame_setup(cpi);
3375 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3376 vp9_setup_in_frame_q_adj(cpi);
3379 // transform / motion compensation build reconstruction frame
3380 vp9_encode_frame(cpi);
3382 // Update the skip mb flag probabilities based on the distribution
3383 // seen in the last encoder iteration.
3384 // update_base_skip_probs(cpi);
3386 vpx_clear_system_state();
3388 // Dummy pack of the bitstream using up to date stats to get an
3389 // accurate estimate of output frame size to determine if we need
3391 if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
3392 save_coding_context(cpi);
3393 if (!cpi->sf.use_nonrd_pick_mode)
3394 vp9_pack_bitstream(cpi, dest, size);
3396 rc->projected_frame_size = (int)(*size) << 3;
3397 restore_coding_context(cpi);
3399 if (frame_over_shoot_limit == 0)
3400 frame_over_shoot_limit = 1;
3403 if (cpi->oxcf.rc_mode == VPX_Q) {
3406 if ((cm->frame_type == KEY_FRAME) &&
3407 rc->this_key_frame_forced &&
3408 (rc->projected_frame_size < rc->max_frame_bandwidth)) {
3412 int64_t high_err_target = cpi->ambient_err;
3413 int64_t low_err_target = cpi->ambient_err >> 1;
3415 #if CONFIG_VP9_HIGHBITDEPTH
3416 if (cm->use_highbitdepth) {
3417 kf_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3419 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3422 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3423 #endif // CONFIG_VP9_HIGHBITDEPTH
3425 // Prevent possible divide by zero error below for perfect KF
3428 // The key frame is not good enough or we can afford
3429 // to make it better without undue risk of popping.
3430 if ((kf_err > high_err_target &&
3431 rc->projected_frame_size <= frame_over_shoot_limit) ||
3432 (kf_err > low_err_target &&
3433 rc->projected_frame_size <= frame_under_shoot_limit)) {
3435 q_high = q > q_low ? q - 1 : q_low;
3438 q = (int)((q * high_err_target) / kf_err);
3439 q = VPXMIN(q, (q_high + q_low) >> 1);
3440 } else if (kf_err < low_err_target &&
3441 rc->projected_frame_size >= frame_under_shoot_limit) {
3442 // The key frame is much better than the previous frame
3444 q_low = q < q_high ? q + 1 : q_high;
3447 q = (int)((q * low_err_target) / kf_err);
3448 q = VPXMIN(q, (q_high + q_low + 1) >> 1);
3451 // Clamp Q to upper and lower limits:
3452 q = clamp(q, q_low, q_high);
3455 } else if (recode_loop_test(
3456 cpi, frame_over_shoot_limit, frame_under_shoot_limit,
3457 q, VPXMAX(q_high, top_index), bottom_index)) {
3458 // Is the projected frame size out of range and are we allowed
3459 // to attempt to recode.
3463 if (cpi->resize_pending == 1) {
3464 // Change in frame size so go back around the recode loop.
3465 cpi->rc.frame_size_selector =
3466 SCALE_STEP1 - cpi->rc.frame_size_selector;
3467 cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
3469 #if CONFIG_INTERNAL_STATS
3470 ++cpi->tot_recode_hits;
3477 // Frame size out of permitted range:
3478 // Update correction factor & compute new Q to try...
3480 // Frame is too large
3481 if (rc->projected_frame_size > rc->this_frame_target) {
3482 // Special case if the projected size is > the max allowed.
3483 if (rc->projected_frame_size >= rc->max_frame_bandwidth)
3484 q_high = rc->worst_quality;
3486 // Raise Qlow as to at least the current value
3487 q_low = q < q_high ? q + 1 : q_high;
3489 if (undershoot_seen || loop_at_this_size > 1) {
3490 // Update rate_correction_factor unless
3491 vp9_rc_update_rate_correction_factors(cpi);
3493 q = (q_high + q_low + 1) / 2;
3495 // Update rate_correction_factor unless
3496 vp9_rc_update_rate_correction_factors(cpi);
3498 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3499 bottom_index, VPXMAX(q_high, top_index));
3501 while (q < q_low && retries < 10) {
3502 vp9_rc_update_rate_correction_factors(cpi);
3503 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3504 bottom_index, VPXMAX(q_high, top_index));
3511 // Frame is too small
3512 q_high = q > q_low ? q - 1 : q_low;
3514 if (overshoot_seen || loop_at_this_size > 1) {
3515 vp9_rc_update_rate_correction_factors(cpi);
3516 q = (q_high + q_low) / 2;
3518 vp9_rc_update_rate_correction_factors(cpi);
3519 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3520 bottom_index, top_index);
3521 // Special case reset for qlow for constrained quality.
3522 // This should only trigger where there is very substantial
3523 // undershoot on a frame and the auto cq level is above
3524 // the user passsed in value.
3525 if (cpi->oxcf.rc_mode == VPX_CQ &&
3530 while (q > q_high && retries < 10) {
3531 vp9_rc_update_rate_correction_factors(cpi);
3532 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
3533 bottom_index, top_index);
3538 undershoot_seen = 1;
3541 // Clamp Q to upper and lower limits:
3542 q = clamp(q, q_low, q_high);
3544 loop = (q != last_q);
3550 // Special case for overlay frame.
3551 if (rc->is_src_frame_alt_ref &&
3552 rc->projected_frame_size < rc->max_frame_bandwidth)
3557 ++loop_at_this_size;
3559 #if CONFIG_INTERNAL_STATS
3560 ++cpi->tot_recode_hits;
3566 static int get_ref_frame_flags(const VP9_COMP *cpi) {
3567 const int *const map = cpi->common.ref_frame_map;
3568 const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
3569 const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
3570 const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
3571 int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
3574 flags &= ~VP9_GOLD_FLAG;
3576 if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
3577 (cpi->svc.number_temporal_layers == 1 &&
3578 cpi->svc.number_spatial_layers == 1))
3579 flags &= ~VP9_GOLD_FLAG;
3582 flags &= ~VP9_ALT_FLAG;
3585 flags &= ~VP9_ALT_FLAG;
3590 static void set_ext_overrides(VP9_COMP *cpi) {
3591 // Overrides the defaults with the externally supplied values with
3592 // vp9_update_reference() and vp9_update_entropy() calls
3593 // Note: The overrides are valid only for the next frame passed
3594 // to encode_frame_to_data_rate() function
3595 if (cpi->ext_refresh_frame_context_pending) {
3596 cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
3597 cpi->ext_refresh_frame_context_pending = 0;
3599 if (cpi->ext_refresh_frame_flags_pending) {
3600 cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
3601 cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
3602 cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
3606 YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
3607 YV12_BUFFER_CONFIG *unscaled,
3608 YV12_BUFFER_CONFIG *scaled,
3609 int use_normative_scaler) {
3610 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3611 cm->mi_rows * MI_SIZE != unscaled->y_height) {
3612 #if CONFIG_VP9_HIGHBITDEPTH
3613 if (use_normative_scaler)
3614 scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth);
3616 scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
3618 if (use_normative_scaler)
3619 scale_and_extend_frame(unscaled, scaled);
3621 scale_and_extend_frame_nonnormative(unscaled, scaled);
3622 #endif // CONFIG_VP9_HIGHBITDEPTH
3629 static void set_arf_sign_bias(VP9_COMP *cpi) {
3630 VP9_COMMON *const cm = &cpi->common;
3633 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
3634 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
3635 arf_sign_bias = cpi->rc.source_alt_ref_active &&
3636 (!cpi->refresh_alt_ref_frame ||
3637 (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
3640 (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
3642 cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
3645 static int setup_interp_filter_search_mask(VP9_COMP *cpi) {
3646 INTERP_FILTER ifilter;
3647 int ref_total[MAX_REF_FRAMES] = {0};
3648 MV_REFERENCE_FRAME ref;
3650 if (cpi->common.last_frame_type == KEY_FRAME ||
3651 cpi->refresh_alt_ref_frame)
3653 for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
3654 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter)
3655 ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
3657 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) {
3658 if ((ref_total[LAST_FRAME] &&
3659 cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
3660 (ref_total[GOLDEN_FRAME] == 0 ||
3661 cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50
3662 < ref_total[GOLDEN_FRAME]) &&
3663 (ref_total[ALTREF_FRAME] == 0 ||
3664 cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50
3665 < ref_total[ALTREF_FRAME]))
3666 mask |= 1 << ifilter;
3671 static void encode_frame_to_data_rate(VP9_COMP *cpi,
3674 unsigned int *frame_flags) {
3675 VP9_COMMON *const cm = &cpi->common;
3676 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3677 struct segmentation *const seg = &cm->seg;
3680 set_ext_overrides(cpi);
3681 vpx_clear_system_state();
3683 // Set the arf sign bias for this frame.
3684 set_arf_sign_bias(cpi);
3686 // Set default state for segment based loop filter update flags.
3687 cm->lf.mode_ref_delta_update = 0;
3689 if (cpi->oxcf.pass == 2 &&
3690 cpi->sf.adaptive_interp_filter_search)
3691 cpi->sf.interp_filter_search_mask =
3692 setup_interp_filter_search_mask(cpi);
3694 // Set various flags etc to special state if it is a key frame.
3695 if (frame_is_intra_only(cm)) {
3696 // Reset the loop filter deltas and segmentation map.
3697 vp9_reset_segment_features(&cm->seg);
3699 // If segmentation is enabled force a map update for key frames.
3701 seg->update_map = 1;
3702 seg->update_data = 1;
3705 // The alternate reference frame cannot be active for a key frame.
3706 cpi->rc.source_alt_ref_active = 0;
3708 cm->error_resilient_mode = oxcf->error_resilient_mode;
3709 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
3711 // By default, encoder assumes decoder can use prev_mi.
3712 if (cm->error_resilient_mode) {
3713 cm->frame_parallel_decoding_mode = 1;
3714 cm->reset_frame_context = 0;
3715 cm->refresh_frame_context = 0;
3716 } else if (cm->intra_only) {
3717 // Only reset the current context.
3718 cm->reset_frame_context = 2;
3721 if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) {
3722 // Use context 0 for intra only empty frame, but the last frame context
3723 // for other empty frames.
3724 if (cpi->svc.encode_empty_frame_state == ENCODING) {
3725 if (cpi->svc.encode_intra_empty_frame != 0)
3726 cm->frame_context_idx = 0;
3728 cm->frame_context_idx = FRAME_CONTEXTS - 1;
3730 cm->frame_context_idx =
3731 cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
3732 cpi->svc.temporal_layer_id;
3735 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
3737 // The probs will be updated based on the frame type of its previous
3738 // frame if frame_parallel_decoding_mode is 0. The type may vary for
3739 // the frame after a key frame in base layer since we may drop enhancement
3740 // layers. So set frame_parallel_decoding_mode to 1 in this case.
3741 if (cm->frame_parallel_decoding_mode == 0) {
3742 if (cpi->svc.number_temporal_layers == 1) {
3743 if (cpi->svc.spatial_layer_id == 0 &&
3744 cpi->svc.layer_context[0].last_frame_type == KEY_FRAME)
3745 cm->frame_parallel_decoding_mode = 1;
3746 } else if (cpi->svc.spatial_layer_id == 0) {
3747 // Find the 2nd frame in temporal base layer and 1st frame in temporal
3748 // enhancement layers from the key frame.
3750 for (i = 0; i < cpi->svc.number_temporal_layers; ++i) {
3751 if (cpi->svc.layer_context[0].frames_from_key_frame == 1 << i) {
3752 cm->frame_parallel_decoding_mode = 1;
3760 // For 1 pass CBR, check if we are dropping this frame.
3761 // Never drop on key frame.
3762 if (oxcf->pass == 0 &&
3763 oxcf->rc_mode == VPX_CBR &&
3764 cm->frame_type != KEY_FRAME) {
3765 if (vp9_rc_drop_frame(cpi)) {
3766 vp9_rc_postencode_update_drop_frame(cpi);
3767 ++cm->current_video_frame;
3768 cpi->ext_refresh_frame_flags_pending = 0;
3773 vpx_clear_system_state();
3775 #if CONFIG_INTERNAL_STATS
3776 memset(cpi->mode_chosen_counts, 0,
3777 MAX_MODES * sizeof(*cpi->mode_chosen_counts));
3780 if (cpi->sf.recode_loop == DISALLOW_RECODE) {
3781 encode_without_recode_loop(cpi, size, dest);
3783 encode_with_recode_loop(cpi, size, dest);
3786 #if CONFIG_VP9_TEMPORAL_DENOISING
3787 #ifdef OUTPUT_YUV_DENOISED
3788 if (oxcf->noise_sensitivity > 0) {
3789 vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME],
3794 #ifdef OUTPUT_YUV_SKINMAP
3795 if (cpi->common.current_video_frame > 1) {
3796 vp9_compute_skin_map(cpi, yuv_skinmap_file);
3800 // Special case code to reduce pulsing when key frames are forced at a
3801 // fixed interval. Note the reconstruction error if it is the frame before
3802 // the force key frame
3803 if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
3804 #if CONFIG_VP9_HIGHBITDEPTH
3805 if (cm->use_highbitdepth) {
3806 cpi->ambient_err = vp9_highbd_get_y_sse(cpi->Source,
3807 get_frame_new_buffer(cm));
3809 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3812 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3813 #endif // CONFIG_VP9_HIGHBITDEPTH
3816 // If the encoder forced a KEY_FRAME decision
3817 if (cm->frame_type == KEY_FRAME)
3818 cpi->refresh_last_frame = 1;
3820 cm->frame_to_show = get_frame_new_buffer(cm);
3821 cm->frame_to_show->color_space = cm->color_space;
3822 cm->frame_to_show->color_range = cm->color_range;
3824 // Pick the loop filter level for the frame.
3825 loopfilter_frame(cpi, cm);
3827 // build the bitstream
3828 vp9_pack_bitstream(cpi, dest, size);
3830 if (cm->seg.update_map)
3831 update_reference_segmentation_map(cpi);
3833 if (frame_is_intra_only(cm) == 0) {
3834 release_scaled_references(cpi);
3836 vp9_update_reference_frames(cpi);
3838 for (t = TX_4X4; t <= TX_32X32; t++)
3839 full_to_model_counts(cpi->td.counts->coef[t],
3840 cpi->td.rd_counts.coef_counts[t]);
3842 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode)
3843 vp9_adapt_coef_probs(cm);
3845 if (!frame_is_intra_only(cm)) {
3846 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
3847 vp9_adapt_mode_probs(cm);
3848 vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
3852 cpi->ext_refresh_frame_flags_pending = 0;
3854 if (cpi->refresh_golden_frame == 1)
3855 cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
3857 cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
3859 if (cpi->refresh_alt_ref_frame == 1)
3860 cpi->frame_flags |= FRAMEFLAGS_ALTREF;
3862 cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
3864 cpi->ref_frame_flags = get_ref_frame_flags(cpi);
3866 cm->last_frame_type = cm->frame_type;
3868 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
3869 vp9_rc_postencode_update(cpi, *size);
3872 output_frame_level_debug_stats(cpi);
3875 if (cm->frame_type == KEY_FRAME) {
3876 // Tell the caller that the frame was coded as a key frame
3877 *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
3879 *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
3882 // Clear the one shot update flags for segmentation map and mode/ref loop
3884 cm->seg.update_map = 0;
3885 cm->seg.update_data = 0;
3886 cm->lf.mode_ref_delta_update = 0;
3888 // keep track of the last coded dimensions
3889 cm->last_width = cm->width;
3890 cm->last_height = cm->height;
3892 // reset to normal state now that we are done.
3893 if (!cm->show_existing_frame)
3894 cm->last_show_frame = cm->show_frame;
3896 if (cm->show_frame) {
3897 vp9_swap_mi_and_prev_mi(cm);
3898 // Don't increment frame counters if this was an altref buffer
3899 // update not a real frame
3900 ++cm->current_video_frame;
3902 vp9_inc_frame_in_layer(cpi);
3904 cm->prev_frame = cm->cur_frame;
3907 cpi->svc.layer_context[cpi->svc.spatial_layer_id *
3908 cpi->svc.number_temporal_layers +
3909 cpi->svc.temporal_layer_id].last_frame_type =
3913 static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
3914 unsigned int *frame_flags) {
3915 vp9_rc_get_svc_params(cpi);
3916 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
3919 static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
3920 unsigned int *frame_flags) {
3921 if (cpi->oxcf.rc_mode == VPX_CBR) {
3922 vp9_rc_get_one_pass_cbr_params(cpi);
3924 vp9_rc_get_one_pass_vbr_params(cpi);
3926 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
3929 static void Pass2Encode(VP9_COMP *cpi, size_t *size,
3930 uint8_t *dest, unsigned int *frame_flags) {
3931 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
3932 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
3934 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
3935 vp9_twopass_postencode_update(cpi);
3938 static void init_ref_frame_bufs(VP9_COMMON *cm) {
3940 BufferPool *const pool = cm->buffer_pool;
3941 cm->new_fb_idx = INVALID_IDX;
3942 for (i = 0; i < REF_FRAMES; ++i) {
3943 cm->ref_frame_map[i] = INVALID_IDX;
3944 pool->frame_bufs[i].ref_count = 0;
3948 static void check_initial_width(VP9_COMP *cpi,
3949 #if CONFIG_VP9_HIGHBITDEPTH
3950 int use_highbitdepth,
3952 int subsampling_x, int subsampling_y) {
3953 VP9_COMMON *const cm = &cpi->common;
3955 if (!cpi->initial_width ||
3956 #if CONFIG_VP9_HIGHBITDEPTH
3957 cm->use_highbitdepth != use_highbitdepth ||
3959 cm->subsampling_x != subsampling_x ||
3960 cm->subsampling_y != subsampling_y) {
3961 cm->subsampling_x = subsampling_x;
3962 cm->subsampling_y = subsampling_y;
3963 #if CONFIG_VP9_HIGHBITDEPTH
3964 cm->use_highbitdepth = use_highbitdepth;
3967 alloc_raw_frame_buffers(cpi);
3968 init_ref_frame_bufs(cm);
3969 alloc_util_frame_buffers(cpi);
3971 init_motion_estimation(cpi); // TODO(agrange) This can be removed.
3973 cpi->initial_width = cm->width;
3974 cpi->initial_height = cm->height;
3975 cpi->initial_mbs = cm->MBs;
3979 #if CONFIG_VP9_TEMPORAL_DENOISING
3980 static void setup_denoiser_buffer(VP9_COMP *cpi) {
3981 VP9_COMMON *const cm = &cpi->common;
3982 if (cpi->oxcf.noise_sensitivity > 0 &&
3983 !cpi->denoiser.frame_buffer_initialized) {
3984 vp9_denoiser_alloc(&(cpi->denoiser), cm->width, cm->height,
3985 cm->subsampling_x, cm->subsampling_y,
3986 #if CONFIG_VP9_HIGHBITDEPTH
3987 cm->use_highbitdepth,
3989 VP9_ENC_BORDER_IN_PIXELS);
3994 int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
3995 YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
3997 VP9_COMMON *cm = &cpi->common;
3998 struct vpx_usec_timer timer;
4000 const int subsampling_x = sd->subsampling_x;
4001 const int subsampling_y = sd->subsampling_y;
4002 #if CONFIG_VP9_HIGHBITDEPTH
4003 const int use_highbitdepth = sd->flags & YV12_FLAG_HIGHBITDEPTH;
4004 check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
4006 check_initial_width(cpi, subsampling_x, subsampling_y);
4007 #endif // CONFIG_VP9_HIGHBITDEPTH
4009 #if CONFIG_VP9_TEMPORAL_DENOISING
4010 setup_denoiser_buffer(cpi);
4012 vpx_usec_timer_start(&timer);
4014 if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
4015 #if CONFIG_VP9_HIGHBITDEPTH
4017 #endif // CONFIG_VP9_HIGHBITDEPTH
4020 vpx_usec_timer_mark(&timer);
4021 cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
4023 if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
4024 (subsampling_x != 1 || subsampling_y != 1)) {
4025 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4026 "Non-4:2:0 color format requires profile 1 or 3");
4029 if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
4030 (subsampling_x == 1 && subsampling_y == 1)) {
4031 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4032 "4:2:0 color format requires profile 0 or 2");
4040 static int frame_is_reference(const VP9_COMP *cpi) {
4041 const VP9_COMMON *cm = &cpi->common;
4043 return cm->frame_type == KEY_FRAME ||
4044 cpi->refresh_last_frame ||
4045 cpi->refresh_golden_frame ||
4046 cpi->refresh_alt_ref_frame ||
4047 cm->refresh_frame_context ||
4048 cm->lf.mode_ref_delta_update ||
4049 cm->seg.update_map ||
4050 cm->seg.update_data;
4053 static void adjust_frame_rate(VP9_COMP *cpi,
4054 const struct lookahead_entry *source) {
4055 int64_t this_duration;
4058 if (source->ts_start == cpi->first_time_stamp_ever) {
4059 this_duration = source->ts_end - source->ts_start;
4062 int64_t last_duration = cpi->last_end_time_stamp_seen
4063 - cpi->last_time_stamp_seen;
4065 this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
4067 // do a step update if the duration changes by 10%
4069 step = (int)((this_duration - last_duration) * 10 / last_duration);
4072 if (this_duration) {
4074 vp9_new_framerate(cpi, 10000000.0 / this_duration);
4076 // Average this frame's rate into the last second's average
4077 // frame rate. If we haven't seen 1 second yet, then average
4078 // over the whole interval seen.
4079 const double interval = VPXMIN(
4080 (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
4081 double avg_duration = 10000000.0 / cpi->framerate;
4082 avg_duration *= (interval - avg_duration + this_duration);
4083 avg_duration /= interval;
4085 vp9_new_framerate(cpi, 10000000.0 / avg_duration);
4088 cpi->last_time_stamp_seen = source->ts_start;
4089 cpi->last_end_time_stamp_seen = source->ts_end;
4092 // Returns 0 if this is not an alt ref else the offset of the source frame
4093 // used as the arf midpoint.
4094 static int get_arf_src_index(VP9_COMP *cpi) {
4095 RATE_CONTROL *const rc = &cpi->rc;
4096 int arf_src_index = 0;
4097 if (is_altref_enabled(cpi)) {
4098 if (cpi->oxcf.pass == 2) {
4099 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4100 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
4101 arf_src_index = gf_group->arf_src_offset[gf_group->index];
4103 } else if (rc->source_alt_ref_pending) {
4104 arf_src_index = rc->frames_till_gf_update_due;
4107 return arf_src_index;
4110 static void check_src_altref(VP9_COMP *cpi,
4111 const struct lookahead_entry *source) {
4112 RATE_CONTROL *const rc = &cpi->rc;
4114 if (cpi->oxcf.pass == 2) {
4115 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4116 rc->is_src_frame_alt_ref =
4117 (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
4119 rc->is_src_frame_alt_ref = cpi->alt_ref_source &&
4120 (source == cpi->alt_ref_source);
4123 if (rc->is_src_frame_alt_ref) {
4124 // Current frame is an ARF overlay frame.
4125 cpi->alt_ref_source = NULL;
4127 // Don't refresh the last buffer for an ARF overlay frame. It will
4128 // become the GF so preserve last as an alternative prediction option.
4129 cpi->refresh_last_frame = 0;
4133 #if CONFIG_INTERNAL_STATS
4134 extern double vp9_get_blockiness(const uint8_t *img1, int img1_pitch,
4135 const uint8_t *img2, int img2_pitch,
4136 int width, int height);
4138 static void adjust_image_stat(double y, double u, double v, double all,
4143 s->stat[ALL] += all;
4144 s->worst = VPXMIN(s->worst, all);
4146 #endif // CONFIG_INTERNAL_STATS
4148 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
4149 size_t *size, uint8_t *dest,
4150 int64_t *time_stamp, int64_t *time_end, int flush) {
4151 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4152 VP9_COMMON *const cm = &cpi->common;
4153 BufferPool *const pool = cm->buffer_pool;
4154 RATE_CONTROL *const rc = &cpi->rc;
4155 struct vpx_usec_timer cmptimer;
4156 YV12_BUFFER_CONFIG *force_src_buffer = NULL;
4157 struct lookahead_entry *last_source = NULL;
4158 struct lookahead_entry *source = NULL;
4162 if (is_two_pass_svc(cpi)) {
4163 #if CONFIG_SPATIAL_SVC
4164 vp9_svc_start_frame(cpi);
4165 // Use a small empty frame instead of a real frame
4166 if (cpi->svc.encode_empty_frame_state == ENCODING)
4167 source = &cpi->svc.empty_frame;
4169 if (oxcf->pass == 2)
4170 vp9_restore_layer_context(cpi);
4171 } else if (is_one_pass_cbr_svc(cpi)) {
4172 vp9_one_pass_cbr_svc_start_layer(cpi);
4175 vpx_usec_timer_start(&cmptimer);
4177 vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
4179 // Is multi-arf enabled.
4180 // Note that at the moment multi_arf is only configured for 2 pass VBR and
4181 // will not work properly with svc.
4182 if ((oxcf->pass == 2) && !cpi->use_svc &&
4183 (cpi->oxcf.enable_auto_arf > 1))
4184 cpi->multi_arf_allowed = 1;
4186 cpi->multi_arf_allowed = 0;
4189 cm->reset_frame_context = 0;
4190 cm->refresh_frame_context = 1;
4191 if (!is_one_pass_cbr_svc(cpi)) {
4192 cpi->refresh_last_frame = 1;
4193 cpi->refresh_golden_frame = 0;
4194 cpi->refresh_alt_ref_frame = 0;
4197 // Should we encode an arf frame.
4198 arf_src_index = get_arf_src_index(cpi);
4200 // Skip alt frame if we encode the empty frame
4201 if (is_two_pass_svc(cpi) && source != NULL)
4204 if (arf_src_index) {
4205 assert(arf_src_index <= rc->frames_to_key);
4207 if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
4208 cpi->alt_ref_source = source;
4210 #if CONFIG_SPATIAL_SVC
4211 if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) {
4213 // Reference a hidden frame from a lower layer
4214 for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) {
4215 if (oxcf->ss_enable_auto_arf[i]) {
4216 cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx;
4221 cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1;
4224 if (oxcf->arnr_max_frames > 0) {
4225 // Produce the filtered ARF frame.
4226 vp9_temporal_filter(cpi, arf_src_index);
4227 vpx_extend_frame_borders(&cpi->alt_ref_buffer);
4228 force_src_buffer = &cpi->alt_ref_buffer;
4233 cpi->refresh_alt_ref_frame = 1;
4234 cpi->refresh_golden_frame = 0;
4235 cpi->refresh_last_frame = 0;
4236 rc->is_src_frame_alt_ref = 0;
4237 rc->source_alt_ref_pending = 0;
4239 rc->source_alt_ref_pending = 0;
4244 // Get last frame source.
4245 if (cm->current_video_frame > 0) {
4246 if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
4250 // Read in the source frame.
4252 source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
4254 source = vp9_lookahead_pop(cpi->lookahead, flush);
4256 if (source != NULL) {
4259 // if the flags indicate intra frame, but if the current picture is for
4260 // non-zero spatial layer, it should not be an intra picture.
4261 // TODO(Won Kap): this needs to change if per-layer intra frame is
4263 if ((source->flags & VPX_EFLAG_FORCE_KF) && cpi->svc.spatial_layer_id) {
4264 source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF);
4267 // Check to see if the frame should be encoded as an arf overlay.
4268 check_src_altref(cpi, source);
4273 cpi->un_scaled_source = cpi->Source = force_src_buffer ? force_src_buffer
4276 cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
4278 *time_stamp = source->ts_start;
4279 *time_end = source->ts_end;
4280 *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
4284 if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
4285 vp9_end_first_pass(cpi); /* get last stats packet */
4286 cpi->twopass.first_pass_done = 1;
4291 if (source->ts_start < cpi->first_time_stamp_ever) {
4292 cpi->first_time_stamp_ever = source->ts_start;
4293 cpi->last_end_time_stamp_seen = source->ts_start;
4296 // Clear down mmx registers
4297 vpx_clear_system_state();
4299 // adjust frame rates based on timestamps given
4300 if (cm->show_frame) {
4301 adjust_frame_rate(cpi, source);
4304 if (is_one_pass_cbr_svc(cpi)) {
4305 vp9_update_temporal_layer_framerate(cpi);
4306 vp9_restore_layer_context(cpi);
4309 // Find a free buffer for the new frame, releasing the reference previously
4311 if (cm->new_fb_idx != INVALID_IDX) {
4312 --pool->frame_bufs[cm->new_fb_idx].ref_count;
4314 cm->new_fb_idx = get_free_fb(cm);
4316 if (cm->new_fb_idx == INVALID_IDX)
4319 cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
4321 if (!cpi->use_svc && cpi->multi_arf_allowed) {
4322 if (cm->frame_type == KEY_FRAME) {
4323 init_buffer_indices(cpi);
4324 } else if (oxcf->pass == 2) {
4325 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4326 cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
4330 // Start with a 0 size frame.
4333 cpi->frame_flags = *frame_flags;
4335 if ((oxcf->pass == 2) &&
4337 (is_two_pass_svc(cpi) &&
4338 cpi->svc.encode_empty_frame_state != ENCODING))) {
4339 vp9_rc_get_second_pass_params(cpi);
4340 } else if (oxcf->pass == 1) {
4341 set_frame_size(cpi);
4344 if (cpi->oxcf.pass != 0 ||
4346 frame_is_intra_only(cm) == 1) {
4347 for (i = 0; i < MAX_REF_FRAMES; ++i)
4348 cpi->scaled_ref_idx[i] = INVALID_IDX;
4351 if (oxcf->pass == 1 &&
4352 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4353 const int lossless = is_lossless_requested(oxcf);
4354 #if CONFIG_VP9_HIGHBITDEPTH
4355 if (cpi->oxcf.use_highbitdepth)
4356 cpi->td.mb.fwd_txm4x4 = lossless ?
4357 vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
4359 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4360 cpi->td.mb.highbd_itxm_add = lossless ? vp9_highbd_iwht4x4_add :
4361 vp9_highbd_idct4x4_add;
4363 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4364 #endif // CONFIG_VP9_HIGHBITDEPTH
4365 cpi->td.mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
4366 vp9_first_pass(cpi, source);
4367 } else if (oxcf->pass == 2 &&
4368 (!cpi->use_svc || is_two_pass_svc(cpi))) {
4369 Pass2Encode(cpi, size, dest, frame_flags);
4370 } else if (cpi->use_svc) {
4371 SvcEncode(cpi, size, dest, frame_flags);
4374 Pass0Encode(cpi, size, dest, frame_flags);
4377 if (cm->refresh_frame_context)
4378 cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
4380 // No frame encoded, or frame was dropped, release scaled references.
4381 if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
4382 release_scaled_references(cpi);
4386 cpi->droppable = !frame_is_reference(cpi);
4389 // Save layer specific state.
4390 if (is_one_pass_cbr_svc(cpi) ||
4391 ((cpi->svc.number_temporal_layers > 1 ||
4392 cpi->svc.number_spatial_layers > 1) &&
4394 vp9_save_layer_context(cpi);
4397 vpx_usec_timer_mark(&cmptimer);
4398 cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
4400 if (cpi->b_calculate_psnr && oxcf->pass != 1 && cm->show_frame)
4401 generate_psnr_packet(cpi);
4403 #if CONFIG_INTERNAL_STATS
4405 if (oxcf->pass != 1) {
4406 double samples = 0.0;
4407 cpi->bytes += (int)(*size);
4409 if (cm->show_frame) {
4412 if (cpi->b_calculate_psnr) {
4413 YV12_BUFFER_CONFIG *orig = cpi->Source;
4414 YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
4415 YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer;
4417 #if CONFIG_VP9_HIGHBITDEPTH
4418 calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd,
4419 cpi->oxcf.input_bit_depth);
4421 calc_psnr(orig, recon, &psnr);
4422 #endif // CONFIG_VP9_HIGHBITDEPTH
4424 adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3],
4425 psnr.psnr[0], &cpi->psnr);
4426 cpi->total_sq_error += psnr.sse[0];
4427 cpi->total_samples += psnr.samples[0];
4428 samples = psnr.samples[0];
4432 double frame_ssim2 = 0, weight = 0;
4433 #if CONFIG_VP9_POSTPROC
4434 if (vpx_alloc_frame_buffer(&cm->post_proc_buffer,
4435 recon->y_crop_width, recon->y_crop_height,
4436 cm->subsampling_x, cm->subsampling_y,
4437 #if CONFIG_VP9_HIGHBITDEPTH
4438 cm->use_highbitdepth,
4440 VP9_ENC_BORDER_IN_PIXELS,
4441 cm->byte_alignment) < 0) {
4442 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
4443 "Failed to allocate post processing buffer");
4446 vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer,
4447 cm->lf.filter_level * 10 / 6);
4449 vpx_clear_system_state();
4451 #if CONFIG_VP9_HIGHBITDEPTH
4452 calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd,
4453 cpi->oxcf.input_bit_depth);
4455 calc_psnr(orig, pp, &psnr2);
4456 #endif // CONFIG_VP9_HIGHBITDEPTH
4458 cpi->totalp_sq_error += psnr2.sse[0];
4459 cpi->totalp_samples += psnr2.samples[0];
4460 adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3],
4461 psnr2.psnr[0], &cpi->psnrp);
4463 #if CONFIG_VP9_HIGHBITDEPTH
4464 if (cm->use_highbitdepth) {
4465 frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight,
4466 (int)cm->bit_depth);
4468 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4471 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4472 #endif // CONFIG_VP9_HIGHBITDEPTH
4474 cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2);
4475 cpi->summed_quality += frame_ssim2 * weight;
4476 cpi->summed_weights += weight;
4478 #if CONFIG_VP9_HIGHBITDEPTH
4479 if (cm->use_highbitdepth) {
4480 frame_ssim2 = vpx_highbd_calc_ssim(
4481 orig, &cm->post_proc_buffer, &weight, (int)cm->bit_depth);
4483 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4486 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight);
4487 #endif // CONFIG_VP9_HIGHBITDEPTH
4489 cpi->summedp_quality += frame_ssim2 * weight;
4490 cpi->summedp_weights += weight;
4493 FILE *f = fopen("q_used.stt", "a");
4494 fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
4495 cpi->common.current_video_frame, y2, u2, v2,
4496 frame_psnr2, frame_ssim2);
4502 if (cpi->b_calculate_blockiness) {
4503 #if CONFIG_VP9_HIGHBITDEPTH
4504 if (!cm->use_highbitdepth)
4507 double frame_blockiness = vp9_get_blockiness(
4508 cpi->Source->y_buffer, cpi->Source->y_stride,
4509 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4510 cpi->Source->y_width, cpi->Source->y_height);
4511 cpi->worst_blockiness =
4512 VPXMAX(cpi->worst_blockiness, frame_blockiness);
4513 cpi->total_blockiness += frame_blockiness;
4517 if (cpi->b_calculate_consistency) {
4518 #if CONFIG_VP9_HIGHBITDEPTH
4519 if (!cm->use_highbitdepth)
4522 double this_inconsistency = vpx_get_ssim_metrics(
4523 cpi->Source->y_buffer, cpi->Source->y_stride,
4524 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4525 cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
4528 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
4529 double consistency = vpx_sse_to_psnr(samples, peak,
4530 (double)cpi->total_inconsistency);
4531 if (consistency > 0.0)
4532 cpi->worst_consistency =
4533 VPXMIN(cpi->worst_consistency, consistency);
4534 cpi->total_inconsistency += this_inconsistency;
4538 if (cpi->b_calculate_ssimg) {
4539 double y, u, v, frame_all;
4540 #if CONFIG_VP9_HIGHBITDEPTH
4541 if (cm->use_highbitdepth) {
4542 frame_all = vpx_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y,
4543 &u, &v, (int)cm->bit_depth);
4545 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u,
4549 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v);
4550 #endif // CONFIG_VP9_HIGHBITDEPTH
4551 adjust_image_stat(y, u, v, frame_all, &cpi->ssimg);
4553 #if CONFIG_VP9_HIGHBITDEPTH
4554 if (!cm->use_highbitdepth)
4557 double y, u, v, frame_all;
4558 frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
4560 adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
4561 /* TODO(JBB): add 10/12 bit support */
4563 #if CONFIG_VP9_HIGHBITDEPTH
4564 if (!cm->use_highbitdepth)
4567 double y, u, v, frame_all;
4568 frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v);
4569 adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
4576 if (is_two_pass_svc(cpi)) {
4577 if (cpi->svc.encode_empty_frame_state == ENCODING) {
4578 cpi->svc.encode_empty_frame_state = ENCODED;
4579 cpi->svc.encode_intra_empty_frame = 0;
4582 if (cm->show_frame) {
4583 ++cpi->svc.spatial_layer_to_encode;
4584 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
4585 cpi->svc.spatial_layer_to_encode = 0;
4587 // May need the empty frame after an visible frame.
4588 cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE;
4590 } else if (is_one_pass_cbr_svc(cpi)) {
4591 if (cm->show_frame) {
4592 ++cpi->svc.spatial_layer_to_encode;
4593 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
4594 cpi->svc.spatial_layer_to_encode = 0;
4597 vpx_clear_system_state();
4601 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
4602 vp9_ppflags_t *flags) {
4603 VP9_COMMON *cm = &cpi->common;
4604 #if !CONFIG_VP9_POSTPROC
4608 if (!cm->show_frame) {
4612 #if CONFIG_VP9_POSTPROC
4613 ret = vp9_post_proc_frame(cm, dest, flags);
4615 if (cm->frame_to_show) {
4616 *dest = *cm->frame_to_show;
4617 dest->y_width = cm->width;
4618 dest->y_height = cm->height;
4619 dest->uv_width = cm->width >> cm->subsampling_x;
4620 dest->uv_height = cm->height >> cm->subsampling_y;
4625 #endif // !CONFIG_VP9_POSTPROC
4626 vpx_clear_system_state();
4631 int vp9_set_internal_size(VP9_COMP *cpi,
4632 VPX_SCALING horiz_mode, VPX_SCALING vert_mode) {
4633 VP9_COMMON *cm = &cpi->common;
4634 int hr = 0, hs = 0, vr = 0, vs = 0;
4636 if (horiz_mode > ONETWO || vert_mode > ONETWO)
4639 Scale2Ratio(horiz_mode, &hr, &hs);
4640 Scale2Ratio(vert_mode, &vr, &vs);
4642 // always go to the next whole number
4643 cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
4644 cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
4645 assert(cm->width <= cpi->initial_width);
4646 assert(cm->height <= cpi->initial_height);
4648 update_frame_size(cpi);
4653 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
4654 unsigned int height) {
4655 VP9_COMMON *cm = &cpi->common;
4656 #if CONFIG_VP9_HIGHBITDEPTH
4657 check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
4659 check_initial_width(cpi, 1, 1);
4660 #endif // CONFIG_VP9_HIGHBITDEPTH
4662 #if CONFIG_VP9_TEMPORAL_DENOISING
4663 setup_denoiser_buffer(cpi);
4668 if (cm->width > cpi->initial_width) {
4669 cm->width = cpi->initial_width;
4670 printf("Warning: Desired width too large, changed to %d\n", cm->width);
4675 cm->height = height;
4676 if (cm->height > cpi->initial_height) {
4677 cm->height = cpi->initial_height;
4678 printf("Warning: Desired height too large, changed to %d\n", cm->height);
4681 assert(cm->width <= cpi->initial_width);
4682 assert(cm->height <= cpi->initial_height);
4684 update_frame_size(cpi);
4689 void vp9_set_svc(VP9_COMP *cpi, int use_svc) {
4690 cpi->use_svc = use_svc;
4694 int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a,
4695 const YV12_BUFFER_CONFIG *b) {
4696 assert(a->y_crop_width == b->y_crop_width);
4697 assert(a->y_crop_height == b->y_crop_height);
4699 return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
4700 a->y_crop_width, a->y_crop_height);
4703 #if CONFIG_VP9_HIGHBITDEPTH
4704 int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
4705 const YV12_BUFFER_CONFIG *b) {
4706 assert(a->y_crop_width == b->y_crop_width);
4707 assert(a->y_crop_height == b->y_crop_height);
4708 assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
4709 assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
4711 return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
4712 a->y_crop_width, a->y_crop_height);
4714 #endif // CONFIG_VP9_HIGHBITDEPTH
4716 int vp9_get_quantizer(VP9_COMP *cpi) {
4717 return cpi->common.base_qindex;
4720 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
4721 if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF |
4722 VP8_EFLAG_NO_REF_ARF)) {
4725 if (flags & VP8_EFLAG_NO_REF_LAST)
4726 ref ^= VP9_LAST_FLAG;
4728 if (flags & VP8_EFLAG_NO_REF_GF)
4729 ref ^= VP9_GOLD_FLAG;
4731 if (flags & VP8_EFLAG_NO_REF_ARF)
4732 ref ^= VP9_ALT_FLAG;
4734 vp9_use_as_reference(cpi, ref);
4737 if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF |
4738 VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF |
4739 VP8_EFLAG_FORCE_ARF)) {
4742 if (flags & VP8_EFLAG_NO_UPD_LAST)
4743 upd ^= VP9_LAST_FLAG;
4745 if (flags & VP8_EFLAG_NO_UPD_GF)
4746 upd ^= VP9_GOLD_FLAG;
4748 if (flags & VP8_EFLAG_NO_UPD_ARF)
4749 upd ^= VP9_ALT_FLAG;
4751 vp9_update_reference(cpi, upd);
4754 if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
4755 vp9_update_entropy(cpi, 0);