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.
16 #include "./vp9_rtcd.h"
17 #include "./vpx_config.h"
18 #include "./vpx_dsp_rtcd.h"
19 #include "./vpx_scale_rtcd.h"
20 #include "vpx_dsp/psnr.h"
21 #include "vpx_dsp/vpx_dsp_common.h"
22 #include "vpx_dsp/vpx_filter.h"
23 #if CONFIG_INTERNAL_STATS
24 #include "vpx_dsp/ssim.h"
26 #include "vpx_ports/mem.h"
27 #include "vpx_ports/system_state.h"
28 #include "vpx_ports/vpx_timer.h"
29 #if CONFIG_BITSTREAM_DEBUG || CONFIG_MISMATCH_DEBUG
30 #include "vpx_util/vpx_debug_util.h"
31 #endif // CONFIG_BITSTREAM_DEBUG || CONFIG_MISMATCH_DEBUG
33 #include "vp9/common/vp9_alloccommon.h"
34 #include "vp9/common/vp9_filter.h"
35 #include "vp9/common/vp9_idct.h"
36 #if CONFIG_NON_GREEDY_MV
37 #include "vp9/common/vp9_mvref_common.h"
39 #if CONFIG_VP9_POSTPROC
40 #include "vp9/common/vp9_postproc.h"
42 #include "vp9/common/vp9_reconinter.h"
43 #include "vp9/common/vp9_reconintra.h"
44 #include "vp9/common/vp9_tile_common.h"
45 #include "vp9/common/vp9_scan.h"
47 #if !CONFIG_REALTIME_ONLY
48 #include "vp9/encoder/vp9_alt_ref_aq.h"
49 #include "vp9/encoder/vp9_aq_360.h"
50 #include "vp9/encoder/vp9_aq_complexity.h"
52 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
53 #if !CONFIG_REALTIME_ONLY
54 #include "vp9/encoder/vp9_aq_variance.h"
56 #include "vp9/encoder/vp9_bitstream.h"
57 #if CONFIG_INTERNAL_STATS
58 #include "vp9/encoder/vp9_blockiness.h"
60 #include "vp9/encoder/vp9_context_tree.h"
61 #include "vp9/encoder/vp9_encodeframe.h"
62 #include "vp9/encoder/vp9_encodemb.h"
63 #include "vp9/encoder/vp9_encodemv.h"
64 #include "vp9/encoder/vp9_encoder.h"
65 #include "vp9/encoder/vp9_ethread.h"
66 #include "vp9/encoder/vp9_extend.h"
67 #include "vp9/encoder/vp9_firstpass.h"
68 #include "vp9/encoder/vp9_mbgraph.h"
69 #if CONFIG_NON_GREEDY_MV
70 #include "vp9/encoder/vp9_mcomp.h"
72 #include "vp9/encoder/vp9_multi_thread.h"
73 #include "vp9/encoder/vp9_noise_estimate.h"
74 #include "vp9/encoder/vp9_picklpf.h"
75 #include "vp9/encoder/vp9_ratectrl.h"
76 #include "vp9/encoder/vp9_rd.h"
77 #include "vp9/encoder/vp9_resize.h"
78 #include "vp9/encoder/vp9_segmentation.h"
79 #include "vp9/encoder/vp9_skin_detection.h"
80 #include "vp9/encoder/vp9_speed_features.h"
81 #include "vp9/encoder/vp9_svc_layercontext.h"
82 #include "vp9/encoder/vp9_temporal_filter.h"
84 #define AM_SEGMENT_ID_INACTIVE 7
85 #define AM_SEGMENT_ID_ACTIVE 0
87 // Whether to use high precision mv for altref computation.
88 #define ALTREF_HIGH_PRECISION_MV 1
90 // Q threshold for high precision mv. Choose a very high value for now so that
91 // HIGH_PRECISION is always chosen.
92 #define HIGH_PRECISION_MV_QTHRESH 200
94 #define FRAME_SIZE_FACTOR 128 // empirical params for context model threshold
95 #define FRAME_RATE_FACTOR 8
97 #ifdef OUTPUT_YUV_DENOISED
98 FILE *yuv_denoised_file = NULL;
100 #ifdef OUTPUT_YUV_SKINMAP
101 static FILE *yuv_skinmap_file = NULL;
103 #ifdef OUTPUT_YUV_REC
106 #ifdef OUTPUT_YUV_SVC_SRC
107 FILE *yuv_svc_src[3] = { NULL, NULL, NULL };
116 #ifdef ENABLE_KF_DENOISE
117 // Test condition for spatial denoise of source.
118 static int is_spatial_denoise_enabled(VP9_COMP *cpi) {
119 VP9_COMMON *const cm = &cpi->common;
120 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
122 return (oxcf->pass != 1) && !is_lossless_requested(&cpi->oxcf) &&
123 frame_is_intra_only(cm);
127 #if CONFIG_VP9_HIGHBITDEPTH
128 void highbd_wht_fwd_txfm(int16_t *src_diff, int bw, tran_low_t *coeff,
131 void wht_fwd_txfm(int16_t *src_diff, int bw, tran_low_t *coeff,
134 #if !CONFIG_REALTIME_ONLY
135 // compute adaptive threshold for skip recoding
136 static int compute_context_model_thresh(const VP9_COMP *const cpi) {
137 const VP9_COMMON *const cm = &cpi->common;
138 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
139 const int frame_size = (cm->width * cm->height) >> 10;
140 const int bitrate = (int)(oxcf->target_bandwidth >> 10);
141 const int qindex_factor = cm->base_qindex + (MAXQ >> 1);
143 // This equation makes the threshold adaptive to frame size.
144 // Coding gain obtained by recoding comes from alternate frames of large
145 // content change. We skip recoding if the difference of previous and current
146 // frame context probability model is less than a certain threshold.
147 // The first component is the most critical part to guarantee adaptivity.
148 // Other parameters are estimated based on normal setting of hd resolution
149 // parameters. e.g frame_size = 1920x1080, bitrate = 8000, qindex_factor < 50
151 ((FRAME_SIZE_FACTOR * frame_size - FRAME_RATE_FACTOR * bitrate) *
158 // compute the total cost difference between current
159 // and previous frame context prob model.
160 static int compute_context_model_diff(const VP9_COMMON *const cm) {
161 const FRAME_CONTEXT *const pre_fc =
162 &cm->frame_contexts[cm->frame_context_idx];
163 const FRAME_CONTEXT *const cur_fc = cm->fc;
164 const FRAME_COUNTS *counts = &cm->counts;
165 vpx_prob pre_last_prob, cur_last_prob;
167 int i, j, k, l, m, n;
170 for (i = 0; i < BLOCK_SIZE_GROUPS; ++i) {
171 for (j = 0; j < INTRA_MODES - 1; ++j) {
172 diff += (int)counts->y_mode[i][j] *
173 (pre_fc->y_mode_prob[i][j] - cur_fc->y_mode_prob[i][j]);
175 pre_last_prob = MAX_PROB - pre_fc->y_mode_prob[i][INTRA_MODES - 2];
176 cur_last_prob = MAX_PROB - cur_fc->y_mode_prob[i][INTRA_MODES - 2];
178 diff += (int)counts->y_mode[i][INTRA_MODES - 1] *
179 (pre_last_prob - cur_last_prob);
183 for (i = 0; i < INTRA_MODES; ++i) {
184 for (j = 0; j < INTRA_MODES - 1; ++j) {
185 diff += (int)counts->uv_mode[i][j] *
186 (pre_fc->uv_mode_prob[i][j] - cur_fc->uv_mode_prob[i][j]);
188 pre_last_prob = MAX_PROB - pre_fc->uv_mode_prob[i][INTRA_MODES - 2];
189 cur_last_prob = MAX_PROB - cur_fc->uv_mode_prob[i][INTRA_MODES - 2];
191 diff += (int)counts->uv_mode[i][INTRA_MODES - 1] *
192 (pre_last_prob - cur_last_prob);
196 for (i = 0; i < PARTITION_CONTEXTS; ++i) {
197 for (j = 0; j < PARTITION_TYPES - 1; ++j) {
198 diff += (int)counts->partition[i][j] *
199 (pre_fc->partition_prob[i][j] - cur_fc->partition_prob[i][j]);
201 pre_last_prob = MAX_PROB - pre_fc->partition_prob[i][PARTITION_TYPES - 2];
202 cur_last_prob = MAX_PROB - cur_fc->partition_prob[i][PARTITION_TYPES - 2];
204 diff += (int)counts->partition[i][PARTITION_TYPES - 1] *
205 (pre_last_prob - cur_last_prob);
209 for (i = 0; i < TX_SIZES; ++i) {
210 for (j = 0; j < PLANE_TYPES; ++j) {
211 for (k = 0; k < REF_TYPES; ++k) {
212 for (l = 0; l < COEF_BANDS; ++l) {
213 for (m = 0; m < BAND_COEFF_CONTEXTS(l); ++m) {
214 for (n = 0; n < UNCONSTRAINED_NODES; ++n) {
215 diff += (int)counts->coef[i][j][k][l][m][n] *
216 (pre_fc->coef_probs[i][j][k][l][m][n] -
217 cur_fc->coef_probs[i][j][k][l][m][n]);
222 pre_fc->coef_probs[i][j][k][l][m][UNCONSTRAINED_NODES - 1];
225 cur_fc->coef_probs[i][j][k][l][m][UNCONSTRAINED_NODES - 1];
227 diff += (int)counts->coef[i][j][k][l][m][UNCONSTRAINED_NODES] *
228 (pre_last_prob - cur_last_prob);
235 // switchable_interp_prob
236 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) {
237 for (j = 0; j < SWITCHABLE_FILTERS - 1; ++j) {
238 diff += (int)counts->switchable_interp[i][j] *
239 (pre_fc->switchable_interp_prob[i][j] -
240 cur_fc->switchable_interp_prob[i][j]);
243 MAX_PROB - pre_fc->switchable_interp_prob[i][SWITCHABLE_FILTERS - 2];
245 MAX_PROB - cur_fc->switchable_interp_prob[i][SWITCHABLE_FILTERS - 2];
247 diff += (int)counts->switchable_interp[i][SWITCHABLE_FILTERS - 1] *
248 (pre_last_prob - cur_last_prob);
252 for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
253 for (j = 0; j < INTER_MODES - 1; ++j) {
254 diff += (int)counts->inter_mode[i][j] *
255 (pre_fc->inter_mode_probs[i][j] - cur_fc->inter_mode_probs[i][j]);
257 pre_last_prob = MAX_PROB - pre_fc->inter_mode_probs[i][INTER_MODES - 2];
258 cur_last_prob = MAX_PROB - cur_fc->inter_mode_probs[i][INTER_MODES - 2];
260 diff += (int)counts->inter_mode[i][INTER_MODES - 1] *
261 (pre_last_prob - cur_last_prob);
265 for (i = 0; i < INTRA_INTER_CONTEXTS; ++i) {
266 diff += (int)counts->intra_inter[i][0] *
267 (pre_fc->intra_inter_prob[i] - cur_fc->intra_inter_prob[i]);
269 pre_last_prob = MAX_PROB - pre_fc->intra_inter_prob[i];
270 cur_last_prob = MAX_PROB - cur_fc->intra_inter_prob[i];
272 diff += (int)counts->intra_inter[i][1] * (pre_last_prob - cur_last_prob);
276 for (i = 0; i < COMP_INTER_CONTEXTS; ++i) {
277 diff += (int)counts->comp_inter[i][0] *
278 (pre_fc->comp_inter_prob[i] - cur_fc->comp_inter_prob[i]);
280 pre_last_prob = MAX_PROB - pre_fc->comp_inter_prob[i];
281 cur_last_prob = MAX_PROB - cur_fc->comp_inter_prob[i];
283 diff += (int)counts->comp_inter[i][1] * (pre_last_prob - cur_last_prob);
287 for (i = 0; i < REF_CONTEXTS; ++i) {
288 for (j = 0; j < 2; ++j) {
289 diff += (int)counts->single_ref[i][j][0] *
290 (pre_fc->single_ref_prob[i][j] - cur_fc->single_ref_prob[i][j]);
292 pre_last_prob = MAX_PROB - pre_fc->single_ref_prob[i][j];
293 cur_last_prob = MAX_PROB - cur_fc->single_ref_prob[i][j];
296 (int)counts->single_ref[i][j][1] * (pre_last_prob - cur_last_prob);
301 for (i = 0; i < REF_CONTEXTS; ++i) {
302 diff += (int)counts->comp_ref[i][0] *
303 (pre_fc->comp_ref_prob[i] - cur_fc->comp_ref_prob[i]);
305 pre_last_prob = MAX_PROB - pre_fc->comp_ref_prob[i];
306 cur_last_prob = MAX_PROB - cur_fc->comp_ref_prob[i];
308 diff += (int)counts->comp_ref[i][1] * (pre_last_prob - cur_last_prob);
312 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
314 for (j = 0; j < TX_SIZES - 1; ++j) {
315 diff += (int)counts->tx.p32x32[i][j] *
316 (pre_fc->tx_probs.p32x32[i][j] - cur_fc->tx_probs.p32x32[i][j]);
318 pre_last_prob = MAX_PROB - pre_fc->tx_probs.p32x32[i][TX_SIZES - 2];
319 cur_last_prob = MAX_PROB - cur_fc->tx_probs.p32x32[i][TX_SIZES - 2];
321 diff += (int)counts->tx.p32x32[i][TX_SIZES - 1] *
322 (pre_last_prob - cur_last_prob);
325 for (j = 0; j < TX_SIZES - 2; ++j) {
326 diff += (int)counts->tx.p16x16[i][j] *
327 (pre_fc->tx_probs.p16x16[i][j] - cur_fc->tx_probs.p16x16[i][j]);
329 pre_last_prob = MAX_PROB - pre_fc->tx_probs.p16x16[i][TX_SIZES - 3];
330 cur_last_prob = MAX_PROB - cur_fc->tx_probs.p16x16[i][TX_SIZES - 3];
332 diff += (int)counts->tx.p16x16[i][TX_SIZES - 2] *
333 (pre_last_prob - cur_last_prob);
336 for (j = 0; j < TX_SIZES - 3; ++j) {
337 diff += (int)counts->tx.p8x8[i][j] *
338 (pre_fc->tx_probs.p8x8[i][j] - cur_fc->tx_probs.p8x8[i][j]);
340 pre_last_prob = MAX_PROB - pre_fc->tx_probs.p8x8[i][TX_SIZES - 4];
341 cur_last_prob = MAX_PROB - cur_fc->tx_probs.p8x8[i][TX_SIZES - 4];
344 (int)counts->tx.p8x8[i][TX_SIZES - 3] * (pre_last_prob - cur_last_prob);
348 for (i = 0; i < SKIP_CONTEXTS; ++i) {
349 diff += (int)counts->skip[i][0] *
350 (pre_fc->skip_probs[i] - cur_fc->skip_probs[i]);
352 pre_last_prob = MAX_PROB - pre_fc->skip_probs[i];
353 cur_last_prob = MAX_PROB - cur_fc->skip_probs[i];
355 diff += (int)counts->skip[i][1] * (pre_last_prob - cur_last_prob);
359 for (i = 0; i < MV_JOINTS - 1; ++i) {
360 diff += (int)counts->mv.joints[i] *
361 (pre_fc->nmvc.joints[i] - cur_fc->nmvc.joints[i]);
363 pre_last_prob = MAX_PROB - pre_fc->nmvc.joints[MV_JOINTS - 2];
364 cur_last_prob = MAX_PROB - cur_fc->nmvc.joints[MV_JOINTS - 2];
367 (int)counts->mv.joints[MV_JOINTS - 1] * (pre_last_prob - cur_last_prob);
369 for (i = 0; i < 2; ++i) {
370 const nmv_component_counts *nmv_count = &counts->mv.comps[i];
371 const nmv_component *pre_nmv_prob = &pre_fc->nmvc.comps[i];
372 const nmv_component *cur_nmv_prob = &cur_fc->nmvc.comps[i];
375 diff += (int)nmv_count->sign[0] * (pre_nmv_prob->sign - cur_nmv_prob->sign);
377 pre_last_prob = MAX_PROB - pre_nmv_prob->sign;
378 cur_last_prob = MAX_PROB - cur_nmv_prob->sign;
380 diff += (int)nmv_count->sign[1] * (pre_last_prob - cur_last_prob);
383 for (j = 0; j < MV_CLASSES - 1; ++j) {
384 diff += (int)nmv_count->classes[j] *
385 (pre_nmv_prob->classes[j] - cur_nmv_prob->classes[j]);
387 pre_last_prob = MAX_PROB - pre_nmv_prob->classes[MV_CLASSES - 2];
388 cur_last_prob = MAX_PROB - cur_nmv_prob->classes[MV_CLASSES - 2];
390 diff += (int)nmv_count->classes[MV_CLASSES - 1] *
391 (pre_last_prob - cur_last_prob);
394 for (j = 0; j < CLASS0_SIZE - 1; ++j) {
395 diff += (int)nmv_count->class0[j] *
396 (pre_nmv_prob->class0[j] - cur_nmv_prob->class0[j]);
398 pre_last_prob = MAX_PROB - pre_nmv_prob->class0[CLASS0_SIZE - 2];
399 cur_last_prob = MAX_PROB - cur_nmv_prob->class0[CLASS0_SIZE - 2];
401 diff += (int)nmv_count->class0[CLASS0_SIZE - 1] *
402 (pre_last_prob - cur_last_prob);
405 for (j = 0; j < MV_OFFSET_BITS; ++j) {
406 diff += (int)nmv_count->bits[j][0] *
407 (pre_nmv_prob->bits[j] - cur_nmv_prob->bits[j]);
409 pre_last_prob = MAX_PROB - pre_nmv_prob->bits[j];
410 cur_last_prob = MAX_PROB - cur_nmv_prob->bits[j];
412 diff += (int)nmv_count->bits[j][1] * (pre_last_prob - cur_last_prob);
416 for (j = 0; j < CLASS0_SIZE; ++j) {
417 for (k = 0; k < MV_FP_SIZE - 1; ++k) {
418 diff += (int)nmv_count->class0_fp[j][k] *
419 (pre_nmv_prob->class0_fp[j][k] - cur_nmv_prob->class0_fp[j][k]);
421 pre_last_prob = MAX_PROB - pre_nmv_prob->class0_fp[j][MV_FP_SIZE - 2];
422 cur_last_prob = MAX_PROB - cur_nmv_prob->class0_fp[j][MV_FP_SIZE - 2];
424 diff += (int)nmv_count->class0_fp[j][MV_FP_SIZE - 1] *
425 (pre_last_prob - cur_last_prob);
429 for (j = 0; j < MV_FP_SIZE - 1; ++j) {
431 (int)nmv_count->fp[j] * (pre_nmv_prob->fp[j] - cur_nmv_prob->fp[j]);
433 pre_last_prob = MAX_PROB - pre_nmv_prob->fp[MV_FP_SIZE - 2];
434 cur_last_prob = MAX_PROB - cur_nmv_prob->fp[MV_FP_SIZE - 2];
437 (int)nmv_count->fp[MV_FP_SIZE - 1] * (pre_last_prob - cur_last_prob);
440 diff += (int)nmv_count->class0_hp[0] *
441 (pre_nmv_prob->class0_hp - cur_nmv_prob->class0_hp);
443 pre_last_prob = MAX_PROB - pre_nmv_prob->class0_hp;
444 cur_last_prob = MAX_PROB - cur_nmv_prob->class0_hp;
446 diff += (int)nmv_count->class0_hp[1] * (pre_last_prob - cur_last_prob);
449 diff += (int)nmv_count->hp[0] * (pre_nmv_prob->hp - cur_nmv_prob->hp);
451 pre_last_prob = MAX_PROB - pre_nmv_prob->hp;
452 cur_last_prob = MAX_PROB - cur_nmv_prob->hp;
454 diff += (int)nmv_count->hp[1] * (pre_last_prob - cur_last_prob);
459 #endif // !CONFIG_REALTIME_ONLY
461 // Test for whether to calculate metrics for the frame.
462 static int is_psnr_calc_enabled(VP9_COMP *cpi) {
463 VP9_COMMON *const cm = &cpi->common;
464 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
466 return cpi->b_calculate_psnr && (oxcf->pass != 1) && cm->show_frame;
469 /* clang-format off */
470 const Vp9LevelSpec vp9_level_defs[VP9_LEVELS] = {
471 // sample rate size breadth bitrate cpb
472 { LEVEL_1, 829440, 36864, 512, 200, 400, 2, 1, 4, 8 },
473 { LEVEL_1_1, 2764800, 73728, 768, 800, 1000, 2, 1, 4, 8 },
474 { LEVEL_2, 4608000, 122880, 960, 1800, 1500, 2, 1, 4, 8 },
475 { LEVEL_2_1, 9216000, 245760, 1344, 3600, 2800, 2, 2, 4, 8 },
476 { LEVEL_3, 20736000, 552960, 2048, 7200, 6000, 2, 4, 4, 8 },
477 { LEVEL_3_1, 36864000, 983040, 2752, 12000, 10000, 2, 4, 4, 8 },
478 { LEVEL_4, 83558400, 2228224, 4160, 18000, 16000, 4, 4, 4, 8 },
479 { LEVEL_4_1, 160432128, 2228224, 4160, 30000, 18000, 4, 4, 5, 6 },
480 { LEVEL_5, 311951360, 8912896, 8384, 60000, 36000, 6, 8, 6, 4 },
481 { LEVEL_5_1, 588251136, 8912896, 8384, 120000, 46000, 8, 8, 10, 4 },
482 // TODO(huisu): update max_cpb_size for level 5_2 ~ 6_2 when
483 // they are finalized (currently tentative).
484 { LEVEL_5_2, 1176502272, 8912896, 8384, 180000, 90000, 8, 8, 10, 4 },
485 { LEVEL_6, 1176502272, 35651584, 16832, 180000, 90000, 8, 16, 10, 4 },
486 { LEVEL_6_1, 2353004544u, 35651584, 16832, 240000, 180000, 8, 16, 10, 4 },
487 { LEVEL_6_2, 4706009088u, 35651584, 16832, 480000, 360000, 8, 16, 10, 4 },
489 /* clang-format on */
491 static const char *level_fail_messages[TARGET_LEVEL_FAIL_IDS] = {
492 "The average bit-rate is too high.",
493 "The picture size is too large.",
494 "The picture width/height is too large.",
495 "The luma sample rate is too large.",
496 "The CPB size is too large.",
497 "The compression ratio is too small",
498 "Too many column tiles are used.",
499 "The alt-ref distance is too small.",
500 "Too many reference buffers are used."
503 static INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) {
518 assert(mode == ONETWO);
525 // Mark all inactive blocks as active. Other segmentation features may be set
526 // so memset cannot be used, instead only inactive blocks should be reset.
527 static void suppress_active_map(VP9_COMP *cpi) {
528 unsigned char *const seg_map = cpi->segmentation_map;
530 if (cpi->active_map.enabled || cpi->active_map.update) {
531 const int rows = cpi->common.mi_rows;
532 const int cols = cpi->common.mi_cols;
535 for (i = 0; i < rows * cols; ++i)
536 if (seg_map[i] == AM_SEGMENT_ID_INACTIVE)
537 seg_map[i] = AM_SEGMENT_ID_ACTIVE;
541 static void apply_active_map(VP9_COMP *cpi) {
542 struct segmentation *const seg = &cpi->common.seg;
543 unsigned char *const seg_map = cpi->segmentation_map;
544 const unsigned char *const active_map = cpi->active_map.map;
547 assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE);
549 if (frame_is_intra_only(&cpi->common)) {
550 cpi->active_map.enabled = 0;
551 cpi->active_map.update = 1;
554 if (cpi->active_map.update) {
555 if (cpi->active_map.enabled) {
556 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
557 if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i];
558 vp9_enable_segmentation(seg);
559 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
560 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
561 // Setting the data to -MAX_LOOP_FILTER will result in the computed loop
562 // filter level being zero regardless of the value of seg->abs_delta.
563 vp9_set_segdata(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF,
566 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
567 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
569 seg->update_data = 1;
573 cpi->active_map.update = 0;
577 static void apply_roi_map(VP9_COMP *cpi) {
578 VP9_COMMON *cm = &cpi->common;
579 struct segmentation *const seg = &cm->seg;
580 vpx_roi_map_t *roi = &cpi->roi;
581 const int *delta_q = roi->delta_q;
582 const int *delta_lf = roi->delta_lf;
583 const int *skip = roi->skip;
585 int internal_delta_q[MAX_SEGMENTS];
587 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
590 // TODO(jianj): Investigate why ROI not working in speed < 5 or in non
592 if (cpi->oxcf.mode != REALTIME || cpi->oxcf.speed < 5) return;
593 if (!roi->enabled) return;
595 memcpy(&ref_frame, roi->ref_frame, sizeof(ref_frame));
597 vp9_enable_segmentation(seg);
598 vp9_clearall_segfeatures(seg);
599 // Select delta coding method;
600 seg->abs_delta = SEGMENT_DELTADATA;
602 memcpy(cpi->segmentation_map, roi->roi_map, (cm->mi_rows * cm->mi_cols));
604 for (i = 0; i < MAX_SEGMENTS; ++i) {
605 // Translate the external delta q values to internal values.
606 internal_delta_q[i] = vp9_quantizer_to_qindex(abs(delta_q[i]));
607 if (delta_q[i] < 0) internal_delta_q[i] = -internal_delta_q[i];
608 vp9_disable_segfeature(seg, i, SEG_LVL_ALT_Q);
609 vp9_disable_segfeature(seg, i, SEG_LVL_ALT_LF);
610 if (internal_delta_q[i] != 0) {
611 vp9_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
612 vp9_set_segdata(seg, i, SEG_LVL_ALT_Q, internal_delta_q[i]);
614 if (delta_lf[i] != 0) {
615 vp9_enable_segfeature(seg, i, SEG_LVL_ALT_LF);
616 vp9_set_segdata(seg, i, SEG_LVL_ALT_LF, delta_lf[i]);
619 vp9_enable_segfeature(seg, i, SEG_LVL_SKIP);
620 vp9_set_segdata(seg, i, SEG_LVL_SKIP, skip[i]);
622 if (ref_frame[i] >= 0) {
624 // ALTREF is not used as reference for nonrd_pickmode with 0 lag.
625 if (ref_frame[i] == ALTREF_FRAME && cpi->sf.use_nonrd_pick_mode)
627 // If GOLDEN is selected, make sure it's set as reference.
628 if (ref_frame[i] == GOLDEN_FRAME &&
629 !(cpi->ref_frame_flags & flag_list[ref_frame[i]])) {
632 // GOLDEN was updated in previous encoded frame, so GOLDEN and LAST are
634 if (ref_frame[i] == GOLDEN_FRAME && cpi->rc.frames_since_golden == 0)
635 ref_frame[i] = LAST_FRAME;
637 vp9_enable_segfeature(seg, i, SEG_LVL_REF_FRAME);
638 vp9_set_segdata(seg, i, SEG_LVL_REF_FRAME, ref_frame[i]);
645 static void init_level_info(Vp9LevelInfo *level_info) {
646 Vp9LevelStats *const level_stats = &level_info->level_stats;
647 Vp9LevelSpec *const level_spec = &level_info->level_spec;
649 memset(level_stats, 0, sizeof(*level_stats));
650 memset(level_spec, 0, sizeof(*level_spec));
651 level_spec->level = LEVEL_UNKNOWN;
652 level_spec->min_altref_distance = INT_MAX;
655 static int check_seg_range(int seg_data[8], int range) {
656 return !(abs(seg_data[0]) > range || abs(seg_data[1]) > range ||
657 abs(seg_data[2]) > range || abs(seg_data[3]) > range ||
658 abs(seg_data[4]) > range || abs(seg_data[5]) > range ||
659 abs(seg_data[6]) > range || abs(seg_data[7]) > range);
662 VP9_LEVEL vp9_get_level(const Vp9LevelSpec *const level_spec) {
664 const Vp9LevelSpec *this_level;
666 vpx_clear_system_state();
668 for (i = 0; i < VP9_LEVELS; ++i) {
669 this_level = &vp9_level_defs[i];
670 if ((double)level_spec->max_luma_sample_rate >
671 (double)this_level->max_luma_sample_rate *
672 (1 + SAMPLE_RATE_GRACE_P) ||
673 level_spec->max_luma_picture_size > this_level->max_luma_picture_size ||
674 level_spec->max_luma_picture_breadth >
675 this_level->max_luma_picture_breadth ||
676 level_spec->average_bitrate > this_level->average_bitrate ||
677 level_spec->max_cpb_size > this_level->max_cpb_size ||
678 level_spec->compression_ratio < this_level->compression_ratio ||
679 level_spec->max_col_tiles > this_level->max_col_tiles ||
680 level_spec->min_altref_distance < this_level->min_altref_distance ||
681 level_spec->max_ref_frame_buffers > this_level->max_ref_frame_buffers)
685 return (i == VP9_LEVELS) ? LEVEL_UNKNOWN : vp9_level_defs[i].level;
688 int vp9_set_roi_map(VP9_COMP *cpi, unsigned char *map, unsigned int rows,
689 unsigned int cols, int delta_q[8], int delta_lf[8],
690 int skip[8], int ref_frame[8]) {
691 VP9_COMMON *cm = &cpi->common;
692 vpx_roi_map_t *roi = &cpi->roi;
693 const int range = 63;
694 const int ref_frame_range = 3; // Alt-ref
695 const int skip_range = 1;
696 const int frame_rows = cpi->common.mi_rows;
697 const int frame_cols = cpi->common.mi_cols;
699 // Check number of rows and columns match
700 if (frame_rows != (int)rows || frame_cols != (int)cols) {
704 if (!check_seg_range(delta_q, range) || !check_seg_range(delta_lf, range) ||
705 !check_seg_range(ref_frame, ref_frame_range) ||
706 !check_seg_range(skip, skip_range))
709 // Also disable segmentation if no deltas are specified.
711 (!(delta_q[0] | delta_q[1] | delta_q[2] | delta_q[3] | delta_q[4] |
712 delta_q[5] | delta_q[6] | delta_q[7] | delta_lf[0] | delta_lf[1] |
713 delta_lf[2] | delta_lf[3] | delta_lf[4] | delta_lf[5] | delta_lf[6] |
714 delta_lf[7] | skip[0] | skip[1] | skip[2] | skip[3] | skip[4] |
715 skip[5] | skip[6] | skip[7]) &&
716 (ref_frame[0] == -1 && ref_frame[1] == -1 && ref_frame[2] == -1 &&
717 ref_frame[3] == -1 && ref_frame[4] == -1 && ref_frame[5] == -1 &&
718 ref_frame[6] == -1 && ref_frame[7] == -1))) {
719 vp9_disable_segmentation(&cm->seg);
720 cpi->roi.enabled = 0;
725 vpx_free(roi->roi_map);
728 CHECK_MEM_ERROR(cm, roi->roi_map, vpx_malloc(rows * cols));
730 // Copy to ROI sturcture in the compressor.
731 memcpy(roi->roi_map, map, rows * cols);
732 memcpy(&roi->delta_q, delta_q, MAX_SEGMENTS * sizeof(delta_q[0]));
733 memcpy(&roi->delta_lf, delta_lf, MAX_SEGMENTS * sizeof(delta_lf[0]));
734 memcpy(&roi->skip, skip, MAX_SEGMENTS * sizeof(skip[0]));
735 memcpy(&roi->ref_frame, ref_frame, MAX_SEGMENTS * sizeof(ref_frame[0]));
743 int vp9_set_active_map(VP9_COMP *cpi, unsigned char *new_map_16x16, int rows,
745 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
746 unsigned char *const active_map_8x8 = cpi->active_map.map;
747 const int mi_rows = cpi->common.mi_rows;
748 const int mi_cols = cpi->common.mi_cols;
749 cpi->active_map.update = 1;
752 for (r = 0; r < mi_rows; ++r) {
753 for (c = 0; c < mi_cols; ++c) {
754 active_map_8x8[r * mi_cols + c] =
755 new_map_16x16[(r >> 1) * cols + (c >> 1)]
756 ? AM_SEGMENT_ID_ACTIVE
757 : AM_SEGMENT_ID_INACTIVE;
760 cpi->active_map.enabled = 1;
762 cpi->active_map.enabled = 0;
770 int vp9_get_active_map(VP9_COMP *cpi, unsigned char *new_map_16x16, int rows,
772 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols &&
774 unsigned char *const seg_map_8x8 = cpi->segmentation_map;
775 const int mi_rows = cpi->common.mi_rows;
776 const int mi_cols = cpi->common.mi_cols;
777 memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
778 if (cpi->active_map.enabled) {
780 for (r = 0; r < mi_rows; ++r) {
781 for (c = 0; c < mi_cols; ++c) {
782 // Cyclic refresh segments are considered active despite not having
783 // AM_SEGMENT_ID_ACTIVE
784 new_map_16x16[(r >> 1) * cols + (c >> 1)] |=
785 seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE;
795 void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
796 MACROBLOCK *const mb = &cpi->td.mb;
797 cpi->common.allow_high_precision_mv = allow_high_precision_mv;
798 if (cpi->common.allow_high_precision_mv) {
799 mb->mvcost = mb->nmvcost_hp;
800 mb->mvsadcost = mb->nmvsadcost_hp;
802 mb->mvcost = mb->nmvcost;
803 mb->mvsadcost = mb->nmvsadcost;
807 static void setup_frame(VP9_COMP *cpi) {
808 VP9_COMMON *const cm = &cpi->common;
809 // Set up entropy context depending on frame type. The decoder mandates
810 // the use of the default context, index 0, for keyframes and inter
811 // frames where the error_resilient_mode or intra_only flag is set. For
812 // other inter-frames the encoder currently uses only two contexts;
813 // context 1 for ALTREF frames and context 0 for the others.
814 if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
815 vp9_setup_past_independence(cm);
817 if (!cpi->use_svc) cm->frame_context_idx = cpi->refresh_alt_ref_frame;
820 // TODO(jingning): Overwrite the frame_context_idx index in multi-layer ARF
821 // case. Need some further investigation on if we could apply this to single
822 // layer ARF case as well.
823 if (cpi->multi_layer_arf && !cpi->use_svc) {
824 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
825 cm->frame_context_idx = clamp(gf_group->layer_depth[gf_group->index] - 1, 0,
829 if (cm->frame_type == KEY_FRAME) {
830 cpi->refresh_golden_frame = 1;
831 cpi->refresh_alt_ref_frame = 1;
832 vp9_zero(cpi->interp_filter_selected);
834 *cm->fc = cm->frame_contexts[cm->frame_context_idx];
835 vp9_zero(cpi->interp_filter_selected[0]);
839 static void vp9_enc_setup_mi(VP9_COMMON *cm) {
841 cm->mi = cm->mip + cm->mi_stride + 1;
842 memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
843 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
844 // Clear top border row
845 memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride);
846 // Clear left border column
847 for (i = 1; i < cm->mi_rows + 1; ++i)
848 memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip));
850 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
851 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
853 memset(cm->mi_grid_base, 0,
854 cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
857 static int vp9_enc_alloc_mi(VP9_COMMON *cm, int mi_size) {
858 cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
859 if (!cm->mip) return 1;
860 cm->prev_mip = vpx_calloc(mi_size, sizeof(*cm->prev_mip));
861 if (!cm->prev_mip) return 1;
862 cm->mi_alloc_size = mi_size;
864 cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO *));
865 if (!cm->mi_grid_base) return 1;
866 cm->prev_mi_grid_base =
867 (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO *));
868 if (!cm->prev_mi_grid_base) return 1;
873 static void vp9_enc_free_mi(VP9_COMMON *cm) {
876 vpx_free(cm->prev_mip);
878 vpx_free(cm->mi_grid_base);
879 cm->mi_grid_base = NULL;
880 vpx_free(cm->prev_mi_grid_base);
881 cm->prev_mi_grid_base = NULL;
882 cm->mi_alloc_size = 0;
885 static void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
886 // Current mip will be the prev_mip for the next frame.
887 MODE_INFO **temp_base = cm->prev_mi_grid_base;
888 MODE_INFO *temp = cm->prev_mip;
890 // Skip update prev_mi frame in show_existing_frame mode.
891 if (cm->show_existing_frame) return;
893 cm->prev_mip = cm->mip;
896 // Update the upper left visible macroblock ptrs.
897 cm->mi = cm->mip + cm->mi_stride + 1;
898 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
900 cm->prev_mi_grid_base = cm->mi_grid_base;
901 cm->mi_grid_base = temp_base;
902 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
903 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
906 void vp9_initialize_enc(void) {
907 static volatile int init_done = 0;
913 vp9_init_intra_predictors();
915 vp9_rc_init_minq_luts();
916 vp9_entropy_mv_init();
917 #if !CONFIG_REALTIME_ONLY
918 vp9_temporal_filter_init();
924 static void dealloc_compressor_data(VP9_COMP *cpi) {
925 VP9_COMMON *const cm = &cpi->common;
928 vpx_free(cpi->mbmi_ext_base);
929 cpi->mbmi_ext_base = NULL;
931 vpx_free(cpi->tile_data);
932 cpi->tile_data = NULL;
934 vpx_free(cpi->segmentation_map);
935 cpi->segmentation_map = NULL;
936 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
937 cpi->coding_context.last_frame_seg_map_copy = NULL;
939 vpx_free(cpi->nmvcosts[0]);
940 vpx_free(cpi->nmvcosts[1]);
941 cpi->nmvcosts[0] = NULL;
942 cpi->nmvcosts[1] = NULL;
944 vpx_free(cpi->nmvcosts_hp[0]);
945 vpx_free(cpi->nmvcosts_hp[1]);
946 cpi->nmvcosts_hp[0] = NULL;
947 cpi->nmvcosts_hp[1] = NULL;
949 vpx_free(cpi->nmvsadcosts[0]);
950 vpx_free(cpi->nmvsadcosts[1]);
951 cpi->nmvsadcosts[0] = NULL;
952 cpi->nmvsadcosts[1] = NULL;
954 vpx_free(cpi->nmvsadcosts_hp[0]);
955 vpx_free(cpi->nmvsadcosts_hp[1]);
956 cpi->nmvsadcosts_hp[0] = NULL;
957 cpi->nmvsadcosts_hp[1] = NULL;
959 vpx_free(cpi->skin_map);
960 cpi->skin_map = NULL;
962 vpx_free(cpi->prev_partition);
963 cpi->prev_partition = NULL;
965 vpx_free(cpi->svc.prev_partition_svc);
966 cpi->svc.prev_partition_svc = NULL;
968 vpx_free(cpi->prev_segment_id);
969 cpi->prev_segment_id = NULL;
971 vpx_free(cpi->prev_variance_low);
972 cpi->prev_variance_low = NULL;
974 vpx_free(cpi->copied_frame_cnt);
975 cpi->copied_frame_cnt = NULL;
977 vpx_free(cpi->content_state_sb_fd);
978 cpi->content_state_sb_fd = NULL;
980 vpx_free(cpi->count_arf_frame_usage);
981 cpi->count_arf_frame_usage = NULL;
982 vpx_free(cpi->count_lastgolden_frame_usage);
983 cpi->count_lastgolden_frame_usage = NULL;
985 vp9_cyclic_refresh_free(cpi->cyclic_refresh);
986 cpi->cyclic_refresh = NULL;
988 vpx_free(cpi->active_map.map);
989 cpi->active_map.map = NULL;
991 vpx_free(cpi->roi.roi_map);
992 cpi->roi.roi_map = NULL;
994 vpx_free(cpi->consec_zero_mv);
995 cpi->consec_zero_mv = NULL;
997 vpx_free(cpi->mb_wiener_variance);
998 cpi->mb_wiener_variance = NULL;
1000 vpx_free(cpi->mi_ssim_rdmult_scaling_factors);
1001 cpi->mi_ssim_rdmult_scaling_factors = NULL;
1003 vp9_free_ref_frame_buffers(cm->buffer_pool);
1004 #if CONFIG_VP9_POSTPROC
1005 vp9_free_postproc_buffers(cm);
1007 vp9_free_context_buffers(cm);
1009 vpx_free_frame_buffer(&cpi->last_frame_uf);
1010 vpx_free_frame_buffer(&cpi->scaled_source);
1011 vpx_free_frame_buffer(&cpi->scaled_last_source);
1012 vpx_free_frame_buffer(&cpi->alt_ref_buffer);
1013 #ifdef ENABLE_KF_DENOISE
1014 vpx_free_frame_buffer(&cpi->raw_unscaled_source);
1015 vpx_free_frame_buffer(&cpi->raw_scaled_source);
1018 vp9_lookahead_destroy(cpi->lookahead);
1020 vpx_free(cpi->tile_tok[0][0]);
1021 cpi->tile_tok[0][0] = 0;
1023 vpx_free(cpi->tplist[0][0]);
1024 cpi->tplist[0][0] = NULL;
1026 vp9_free_pc_tree(&cpi->td);
1028 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
1029 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
1030 vpx_free(lc->rc_twopass_stats_in.buf);
1031 lc->rc_twopass_stats_in.buf = NULL;
1032 lc->rc_twopass_stats_in.sz = 0;
1035 if (cpi->source_diff_var != NULL) {
1036 vpx_free(cpi->source_diff_var);
1037 cpi->source_diff_var = NULL;
1040 for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
1041 vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]);
1043 memset(&cpi->svc.scaled_frames[0], 0,
1044 MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
1046 vpx_free_frame_buffer(&cpi->svc.scaled_temp);
1047 memset(&cpi->svc.scaled_temp, 0, sizeof(cpi->svc.scaled_temp));
1049 vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
1050 memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
1052 vp9_free_svc_cyclic_refresh(cpi);
1055 static void save_coding_context(VP9_COMP *cpi) {
1056 CODING_CONTEXT *const cc = &cpi->coding_context;
1057 VP9_COMMON *cm = &cpi->common;
1059 // Stores a snapshot of key state variables which can subsequently be
1060 // restored with a call to vp9_restore_coding_context. These functions are
1061 // intended for use in a re-code loop in vp9_compress_frame where the
1062 // quantizer value is adjusted between loop iterations.
1063 vp9_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost);
1065 memcpy(cc->nmvcosts[0], cpi->nmvcosts[0],
1066 MV_VALS * sizeof(*cpi->nmvcosts[0]));
1067 memcpy(cc->nmvcosts[1], cpi->nmvcosts[1],
1068 MV_VALS * sizeof(*cpi->nmvcosts[1]));
1069 memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0],
1070 MV_VALS * sizeof(*cpi->nmvcosts_hp[0]));
1071 memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1],
1072 MV_VALS * sizeof(*cpi->nmvcosts_hp[1]));
1074 vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs);
1076 memcpy(cpi->coding_context.last_frame_seg_map_copy, cm->last_frame_seg_map,
1077 (cm->mi_rows * cm->mi_cols));
1079 vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
1080 vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
1085 static void restore_coding_context(VP9_COMP *cpi) {
1086 CODING_CONTEXT *const cc = &cpi->coding_context;
1087 VP9_COMMON *cm = &cpi->common;
1089 // Restore key state variables to the snapshot state stored in the
1090 // previous call to vp9_save_coding_context.
1091 vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
1093 memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0]));
1094 memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1]));
1095 memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0],
1096 MV_VALS * sizeof(*cc->nmvcosts_hp[0]));
1097 memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1],
1098 MV_VALS * sizeof(*cc->nmvcosts_hp[1]));
1100 vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs);
1102 memcpy(cm->last_frame_seg_map, cpi->coding_context.last_frame_seg_map_copy,
1103 (cm->mi_rows * cm->mi_cols));
1105 vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
1106 vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
1111 #if !CONFIG_REALTIME_ONLY
1112 static void configure_static_seg_features(VP9_COMP *cpi) {
1113 VP9_COMMON *const cm = &cpi->common;
1114 const RATE_CONTROL *const rc = &cpi->rc;
1115 struct segmentation *const seg = &cm->seg;
1117 int high_q = (int)(rc->avg_q > 48.0);
1120 // Disable and clear down for KF
1121 if (cm->frame_type == KEY_FRAME) {
1122 // Clear down the global segmentation map
1123 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
1124 seg->update_map = 0;
1125 seg->update_data = 0;
1126 cpi->static_mb_pct = 0;
1128 // Disable segmentation
1129 vp9_disable_segmentation(seg);
1131 // Clear down the segment features.
1132 vp9_clearall_segfeatures(seg);
1133 } else if (cpi->refresh_alt_ref_frame) {
1134 // If this is an alt ref frame
1135 // Clear down the global segmentation map
1136 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
1137 seg->update_map = 0;
1138 seg->update_data = 0;
1139 cpi->static_mb_pct = 0;
1141 // Disable segmentation and individual segment features by default
1142 vp9_disable_segmentation(seg);
1143 vp9_clearall_segfeatures(seg);
1145 // Scan frames from current to arf frame.
1146 // This function re-enables segmentation if appropriate.
1147 vp9_update_mbgraph_stats(cpi);
1149 // If segmentation was enabled set those features needed for the
1152 seg->update_map = 1;
1153 seg->update_data = 1;
1156 vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875, cm->bit_depth);
1157 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
1158 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
1160 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
1161 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
1163 // Where relevant assume segment data is delta data
1164 seg->abs_delta = SEGMENT_DELTADATA;
1166 } else if (seg->enabled) {
1167 // All other frames if segmentation has been enabled
1169 // First normal frame in a valid gf or alt ref group
1170 if (rc->frames_since_golden == 0) {
1171 // Set up segment features for normal frames in an arf group
1172 if (rc->source_alt_ref_active) {
1173 seg->update_map = 0;
1174 seg->update_data = 1;
1175 seg->abs_delta = SEGMENT_DELTADATA;
1178 vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125, cm->bit_depth);
1179 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
1180 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
1182 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
1183 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
1185 // Segment coding disabled for compred testing
1186 if (high_q || (cpi->static_mb_pct == 100)) {
1187 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
1188 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
1189 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
1192 // Disable segmentation and clear down features if alt ref
1193 // is not active for this group
1195 vp9_disable_segmentation(seg);
1197 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
1199 seg->update_map = 0;
1200 seg->update_data = 0;
1202 vp9_clearall_segfeatures(seg);
1204 } else if (rc->is_src_frame_alt_ref) {
1205 // Special case where we are coding over the top of a previous
1207 // Segment coding disabled for compred testing
1209 // Enable ref frame features for segment 0 as well
1210 vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
1211 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
1213 // All mbs should use ALTREF_FRAME
1214 vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
1215 vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
1216 vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
1217 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
1219 // Skip all MBs if high Q (0,0 mv and skip coeffs)
1221 vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP);
1222 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
1224 // Enable data update
1225 seg->update_data = 1;
1227 // All other frames.
1229 // No updates.. leave things as they are.
1230 seg->update_map = 0;
1231 seg->update_data = 0;
1235 #endif // !CONFIG_REALTIME_ONLY
1237 static void update_reference_segmentation_map(VP9_COMP *cpi) {
1238 VP9_COMMON *const cm = &cpi->common;
1239 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
1240 uint8_t *cache_ptr = cm->last_frame_seg_map;
1243 for (row = 0; row < cm->mi_rows; row++) {
1244 MODE_INFO **mi_8x8 = mi_8x8_ptr;
1245 uint8_t *cache = cache_ptr;
1246 for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
1247 cache[0] = mi_8x8[0]->segment_id;
1248 mi_8x8_ptr += cm->mi_stride;
1249 cache_ptr += cm->mi_cols;
1253 static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
1254 VP9_COMMON *cm = &cpi->common;
1255 const VP9EncoderConfig *oxcf = &cpi->oxcf;
1257 if (!cpi->lookahead)
1258 cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
1259 cm->subsampling_x, cm->subsampling_y,
1260 #if CONFIG_VP9_HIGHBITDEPTH
1261 cm->use_highbitdepth,
1263 oxcf->lag_in_frames);
1264 if (!cpi->lookahead)
1265 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1266 "Failed to allocate lag buffers");
1268 // TODO(agrange) Check if ARF is enabled and skip allocation if not.
1269 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer, oxcf->width, oxcf->height,
1270 cm->subsampling_x, cm->subsampling_y,
1271 #if CONFIG_VP9_HIGHBITDEPTH
1272 cm->use_highbitdepth,
1274 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
1276 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1277 "Failed to allocate altref buffer");
1280 static void alloc_util_frame_buffers(VP9_COMP *cpi) {
1281 VP9_COMMON *const cm = &cpi->common;
1282 if (vpx_realloc_frame_buffer(&cpi->last_frame_uf, cm->width, cm->height,
1283 cm->subsampling_x, cm->subsampling_y,
1284 #if CONFIG_VP9_HIGHBITDEPTH
1285 cm->use_highbitdepth,
1287 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
1289 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1290 "Failed to allocate last frame buffer");
1292 if (vpx_realloc_frame_buffer(&cpi->scaled_source, cm->width, cm->height,
1293 cm->subsampling_x, cm->subsampling_y,
1294 #if CONFIG_VP9_HIGHBITDEPTH
1295 cm->use_highbitdepth,
1297 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
1299 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1300 "Failed to allocate scaled source buffer");
1302 // For 1 pass cbr: allocate scaled_frame that may be used as an intermediate
1303 // buffer for a 2 stage down-sampling: two stages of 1:2 down-sampling for a
1304 // target of 1/4x1/4. number_spatial_layers must be greater than 2.
1305 if (is_one_pass_cbr_svc(cpi) && !cpi->svc.scaled_temp_is_alloc &&
1306 cpi->svc.number_spatial_layers > 2) {
1307 cpi->svc.scaled_temp_is_alloc = 1;
1308 if (vpx_realloc_frame_buffer(
1309 &cpi->svc.scaled_temp, cm->width >> 1, cm->height >> 1,
1310 cm->subsampling_x, cm->subsampling_y,
1311 #if CONFIG_VP9_HIGHBITDEPTH
1312 cm->use_highbitdepth,
1314 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, NULL, NULL, NULL))
1315 vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
1316 "Failed to allocate scaled_frame for svc ");
1319 if (vpx_realloc_frame_buffer(&cpi->scaled_last_source, cm->width, cm->height,
1320 cm->subsampling_x, cm->subsampling_y,
1321 #if CONFIG_VP9_HIGHBITDEPTH
1322 cm->use_highbitdepth,
1324 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
1326 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1327 "Failed to allocate scaled last source buffer");
1328 #ifdef ENABLE_KF_DENOISE
1329 if (vpx_realloc_frame_buffer(&cpi->raw_unscaled_source, cm->width, cm->height,
1330 cm->subsampling_x, cm->subsampling_y,
1331 #if CONFIG_VP9_HIGHBITDEPTH
1332 cm->use_highbitdepth,
1334 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
1336 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1337 "Failed to allocate unscaled raw source frame buffer");
1339 if (vpx_realloc_frame_buffer(&cpi->raw_scaled_source, cm->width, cm->height,
1340 cm->subsampling_x, cm->subsampling_y,
1341 #if CONFIG_VP9_HIGHBITDEPTH
1342 cm->use_highbitdepth,
1344 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
1346 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1347 "Failed to allocate scaled raw source frame buffer");
1351 static int alloc_context_buffers_ext(VP9_COMP *cpi) {
1352 VP9_COMMON *cm = &cpi->common;
1353 int mi_size = cm->mi_cols * cm->mi_rows;
1355 cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
1356 if (!cpi->mbmi_ext_base) return 1;
1361 static void alloc_compressor_data(VP9_COMP *cpi) {
1362 VP9_COMMON *cm = &cpi->common;
1365 vp9_alloc_context_buffers(cm, cm->width, cm->height);
1367 alloc_context_buffers_ext(cpi);
1369 vpx_free(cpi->tile_tok[0][0]);
1372 unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
1373 CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
1374 vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
1377 sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
1378 vpx_free(cpi->tplist[0][0]);
1380 cm, cpi->tplist[0][0],
1381 vpx_calloc(sb_rows * 4 * (1 << 6), sizeof(*cpi->tplist[0][0])));
1383 vp9_setup_pc_tree(&cpi->common, &cpi->td);
1386 void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
1387 cpi->framerate = framerate < 0.1 ? 30 : framerate;
1388 vp9_rc_update_framerate(cpi);
1391 static void set_tile_limits(VP9_COMP *cpi) {
1392 VP9_COMMON *const cm = &cpi->common;
1394 int min_log2_tile_cols, max_log2_tile_cols;
1395 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
1397 cm->log2_tile_cols =
1398 clamp(cpi->oxcf.tile_columns, min_log2_tile_cols, max_log2_tile_cols);
1399 cm->log2_tile_rows = cpi->oxcf.tile_rows;
1401 if (cpi->oxcf.target_level == LEVEL_AUTO) {
1402 const int level_tile_cols =
1403 log_tile_cols_from_picsize_level(cpi->common.width, cpi->common.height);
1404 if (cm->log2_tile_cols > level_tile_cols) {
1405 cm->log2_tile_cols = VPXMAX(level_tile_cols, min_log2_tile_cols);
1410 static void update_frame_size(VP9_COMP *cpi) {
1411 VP9_COMMON *const cm = &cpi->common;
1412 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1414 vp9_set_mb_mi(cm, cm->width, cm->height);
1415 vp9_init_context_buffers(cm);
1416 vp9_init_macroblockd(cm, xd, NULL);
1417 cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base;
1418 memset(cpi->mbmi_ext_base, 0,
1419 cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
1421 set_tile_limits(cpi);
1424 static void init_buffer_indices(VP9_COMP *cpi) {
1427 for (ref_frame = 0; ref_frame < REF_FRAMES; ++ref_frame)
1428 cpi->ref_fb_idx[ref_frame] = ref_frame;
1430 cpi->lst_fb_idx = cpi->ref_fb_idx[LAST_FRAME - 1];
1431 cpi->gld_fb_idx = cpi->ref_fb_idx[GOLDEN_FRAME - 1];
1432 cpi->alt_fb_idx = cpi->ref_fb_idx[ALTREF_FRAME - 1];
1435 static void init_level_constraint(LevelConstraint *lc) {
1436 lc->level_index = -1;
1437 lc->max_cpb_size = INT_MAX;
1438 lc->max_frame_size = INT_MAX;
1439 lc->rc_config_updated = 0;
1443 static void set_level_constraint(LevelConstraint *ls, int8_t level_index) {
1444 vpx_clear_system_state();
1445 ls->level_index = level_index;
1446 if (level_index >= 0) {
1447 ls->max_cpb_size = vp9_level_defs[level_index].max_cpb_size * (double)1000;
1451 static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
1452 VP9_COMMON *const cm = &cpi->common;
1455 cpi->framerate = oxcf->init_framerate;
1456 cm->profile = oxcf->profile;
1457 cm->bit_depth = oxcf->bit_depth;
1458 #if CONFIG_VP9_HIGHBITDEPTH
1459 cm->use_highbitdepth = oxcf->use_highbitdepth;
1461 cm->color_space = oxcf->color_space;
1462 cm->color_range = oxcf->color_range;
1464 cpi->target_level = oxcf->target_level;
1465 cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
1466 set_level_constraint(&cpi->level_constraint,
1467 get_level_index(cpi->target_level));
1469 cm->width = oxcf->width;
1470 cm->height = oxcf->height;
1471 alloc_compressor_data(cpi);
1473 cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode;
1475 // Single thread case: use counts in common.
1476 cpi->td.counts = &cm->counts;
1478 // Spatial scalability.
1479 cpi->svc.number_spatial_layers = oxcf->ss_number_layers;
1480 // Temporal scalability.
1481 cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
1483 if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
1484 ((cpi->svc.number_temporal_layers > 1 ||
1485 cpi->svc.number_spatial_layers > 1) &&
1486 cpi->oxcf.pass != 1)) {
1487 vp9_init_layer_context(cpi);
1490 // change includes all joint functionality
1491 vp9_change_config(cpi, oxcf);
1493 cpi->static_mb_pct = 0;
1494 cpi->ref_frame_flags = 0;
1496 init_buffer_indices(cpi);
1498 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
1501 static void set_rc_buffer_sizes(RATE_CONTROL *rc,
1502 const VP9EncoderConfig *oxcf) {
1503 const int64_t bandwidth = oxcf->target_bandwidth;
1504 const int64_t starting = oxcf->starting_buffer_level_ms;
1505 const int64_t optimal = oxcf->optimal_buffer_level_ms;
1506 const int64_t maximum = oxcf->maximum_buffer_size_ms;
1508 rc->starting_buffer_level = starting * bandwidth / 1000;
1509 rc->optimal_buffer_level =
1510 (optimal == 0) ? bandwidth / 8 : optimal * bandwidth / 1000;
1511 rc->maximum_buffer_size =
1512 (maximum == 0) ? bandwidth / 8 : maximum * bandwidth / 1000;
1515 #if CONFIG_VP9_HIGHBITDEPTH
1516 #define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX4DF) \
1517 cpi->fn_ptr[BT].sdf = SDF; \
1518 cpi->fn_ptr[BT].sdaf = SDAF; \
1519 cpi->fn_ptr[BT].vf = VF; \
1520 cpi->fn_ptr[BT].svf = SVF; \
1521 cpi->fn_ptr[BT].svaf = SVAF; \
1522 cpi->fn_ptr[BT].sdx4df = SDX4DF;
1524 #define MAKE_BFP_SAD_WRAPPER(fnname) \
1525 static unsigned int fnname##_bits8(const uint8_t *src_ptr, \
1526 int source_stride, \
1527 const uint8_t *ref_ptr, int ref_stride) { \
1528 return fnname(src_ptr, source_stride, ref_ptr, ref_stride); \
1530 static unsigned int fnname##_bits10( \
1531 const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, \
1533 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2; \
1535 static unsigned int fnname##_bits12( \
1536 const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, \
1538 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4; \
1541 #define MAKE_BFP_SADAVG_WRAPPER(fnname) \
1542 static unsigned int fnname##_bits8( \
1543 const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, \
1544 int ref_stride, const uint8_t *second_pred) { \
1545 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred); \
1547 static unsigned int fnname##_bits10( \
1548 const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, \
1549 int ref_stride, const uint8_t *second_pred) { \
1550 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred) >> \
1553 static unsigned int fnname##_bits12( \
1554 const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, \
1555 int ref_stride, const uint8_t *second_pred) { \
1556 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred) >> \
1560 #define MAKE_BFP_SAD4D_WRAPPER(fnname) \
1561 static void fnname##_bits8(const uint8_t *src_ptr, int source_stride, \
1562 const uint8_t *const ref_ptr[], int ref_stride, \
1563 unsigned int *sad_array) { \
1564 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1566 static void fnname##_bits10(const uint8_t *src_ptr, int source_stride, \
1567 const uint8_t *const ref_ptr[], int ref_stride, \
1568 unsigned int *sad_array) { \
1570 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1571 for (i = 0; i < 4; i++) sad_array[i] >>= 2; \
1573 static void fnname##_bits12(const uint8_t *src_ptr, int source_stride, \
1574 const uint8_t *const ref_ptr[], int ref_stride, \
1575 unsigned int *sad_array) { \
1577 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \
1578 for (i = 0; i < 4; i++) sad_array[i] >>= 4; \
1581 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16)
1582 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg)
1583 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d)
1584 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32)
1585 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg)
1586 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d)
1587 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32)
1588 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg)
1589 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d)
1590 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64)
1591 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg)
1592 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d)
1593 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32)
1594 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg)
1595 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d)
1596 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64)
1597 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg)
1598 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d)
1599 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16)
1600 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg)
1601 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d)
1602 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8)
1603 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg)
1604 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d)
1605 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16)
1606 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg)
1607 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d)
1608 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8)
1609 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg)
1610 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d)
1611 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4)
1612 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg)
1613 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d)
1614 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8)
1615 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg)
1616 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d)
1617 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4)
1618 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg)
1619 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d)
1621 static void highbd_set_var_fns(VP9_COMP *const cpi) {
1622 VP9_COMMON *const cm = &cpi->common;
1623 if (cm->use_highbitdepth) {
1624 switch (cm->bit_depth) {
1626 HIGHBD_BFP(BLOCK_32X16, vpx_highbd_sad32x16_bits8,
1627 vpx_highbd_sad32x16_avg_bits8, vpx_highbd_8_variance32x16,
1628 vpx_highbd_8_sub_pixel_variance32x16,
1629 vpx_highbd_8_sub_pixel_avg_variance32x16,
1630 vpx_highbd_sad32x16x4d_bits8)
1632 HIGHBD_BFP(BLOCK_16X32, vpx_highbd_sad16x32_bits8,
1633 vpx_highbd_sad16x32_avg_bits8, vpx_highbd_8_variance16x32,
1634 vpx_highbd_8_sub_pixel_variance16x32,
1635 vpx_highbd_8_sub_pixel_avg_variance16x32,
1636 vpx_highbd_sad16x32x4d_bits8)
1638 HIGHBD_BFP(BLOCK_64X32, vpx_highbd_sad64x32_bits8,
1639 vpx_highbd_sad64x32_avg_bits8, vpx_highbd_8_variance64x32,
1640 vpx_highbd_8_sub_pixel_variance64x32,
1641 vpx_highbd_8_sub_pixel_avg_variance64x32,
1642 vpx_highbd_sad64x32x4d_bits8)
1644 HIGHBD_BFP(BLOCK_32X64, vpx_highbd_sad32x64_bits8,
1645 vpx_highbd_sad32x64_avg_bits8, vpx_highbd_8_variance32x64,
1646 vpx_highbd_8_sub_pixel_variance32x64,
1647 vpx_highbd_8_sub_pixel_avg_variance32x64,
1648 vpx_highbd_sad32x64x4d_bits8)
1650 HIGHBD_BFP(BLOCK_32X32, vpx_highbd_sad32x32_bits8,
1651 vpx_highbd_sad32x32_avg_bits8, vpx_highbd_8_variance32x32,
1652 vpx_highbd_8_sub_pixel_variance32x32,
1653 vpx_highbd_8_sub_pixel_avg_variance32x32,
1654 vpx_highbd_sad32x32x4d_bits8)
1656 HIGHBD_BFP(BLOCK_64X64, vpx_highbd_sad64x64_bits8,
1657 vpx_highbd_sad64x64_avg_bits8, vpx_highbd_8_variance64x64,
1658 vpx_highbd_8_sub_pixel_variance64x64,
1659 vpx_highbd_8_sub_pixel_avg_variance64x64,
1660 vpx_highbd_sad64x64x4d_bits8)
1662 HIGHBD_BFP(BLOCK_16X16, vpx_highbd_sad16x16_bits8,
1663 vpx_highbd_sad16x16_avg_bits8, vpx_highbd_8_variance16x16,
1664 vpx_highbd_8_sub_pixel_variance16x16,
1665 vpx_highbd_8_sub_pixel_avg_variance16x16,
1666 vpx_highbd_sad16x16x4d_bits8)
1668 HIGHBD_BFP(BLOCK_16X8, vpx_highbd_sad16x8_bits8,
1669 vpx_highbd_sad16x8_avg_bits8, vpx_highbd_8_variance16x8,
1670 vpx_highbd_8_sub_pixel_variance16x8,
1671 vpx_highbd_8_sub_pixel_avg_variance16x8,
1672 vpx_highbd_sad16x8x4d_bits8)
1674 HIGHBD_BFP(BLOCK_8X16, vpx_highbd_sad8x16_bits8,
1675 vpx_highbd_sad8x16_avg_bits8, vpx_highbd_8_variance8x16,
1676 vpx_highbd_8_sub_pixel_variance8x16,
1677 vpx_highbd_8_sub_pixel_avg_variance8x16,
1678 vpx_highbd_sad8x16x4d_bits8)
1681 BLOCK_8X8, vpx_highbd_sad8x8_bits8, vpx_highbd_sad8x8_avg_bits8,
1682 vpx_highbd_8_variance8x8, vpx_highbd_8_sub_pixel_variance8x8,
1683 vpx_highbd_8_sub_pixel_avg_variance8x8, vpx_highbd_sad8x8x4d_bits8)
1686 BLOCK_8X4, vpx_highbd_sad8x4_bits8, vpx_highbd_sad8x4_avg_bits8,
1687 vpx_highbd_8_variance8x4, vpx_highbd_8_sub_pixel_variance8x4,
1688 vpx_highbd_8_sub_pixel_avg_variance8x4, vpx_highbd_sad8x4x4d_bits8)
1691 BLOCK_4X8, vpx_highbd_sad4x8_bits8, vpx_highbd_sad4x8_avg_bits8,
1692 vpx_highbd_8_variance4x8, vpx_highbd_8_sub_pixel_variance4x8,
1693 vpx_highbd_8_sub_pixel_avg_variance4x8, vpx_highbd_sad4x8x4d_bits8)
1696 BLOCK_4X4, vpx_highbd_sad4x4_bits8, vpx_highbd_sad4x4_avg_bits8,
1697 vpx_highbd_8_variance4x4, vpx_highbd_8_sub_pixel_variance4x4,
1698 vpx_highbd_8_sub_pixel_avg_variance4x4, vpx_highbd_sad4x4x4d_bits8)
1702 HIGHBD_BFP(BLOCK_32X16, vpx_highbd_sad32x16_bits10,
1703 vpx_highbd_sad32x16_avg_bits10, vpx_highbd_10_variance32x16,
1704 vpx_highbd_10_sub_pixel_variance32x16,
1705 vpx_highbd_10_sub_pixel_avg_variance32x16,
1706 vpx_highbd_sad32x16x4d_bits10)
1708 HIGHBD_BFP(BLOCK_16X32, vpx_highbd_sad16x32_bits10,
1709 vpx_highbd_sad16x32_avg_bits10, vpx_highbd_10_variance16x32,
1710 vpx_highbd_10_sub_pixel_variance16x32,
1711 vpx_highbd_10_sub_pixel_avg_variance16x32,
1712 vpx_highbd_sad16x32x4d_bits10)
1714 HIGHBD_BFP(BLOCK_64X32, vpx_highbd_sad64x32_bits10,
1715 vpx_highbd_sad64x32_avg_bits10, vpx_highbd_10_variance64x32,
1716 vpx_highbd_10_sub_pixel_variance64x32,
1717 vpx_highbd_10_sub_pixel_avg_variance64x32,
1718 vpx_highbd_sad64x32x4d_bits10)
1720 HIGHBD_BFP(BLOCK_32X64, vpx_highbd_sad32x64_bits10,
1721 vpx_highbd_sad32x64_avg_bits10, vpx_highbd_10_variance32x64,
1722 vpx_highbd_10_sub_pixel_variance32x64,
1723 vpx_highbd_10_sub_pixel_avg_variance32x64,
1724 vpx_highbd_sad32x64x4d_bits10)
1726 HIGHBD_BFP(BLOCK_32X32, vpx_highbd_sad32x32_bits10,
1727 vpx_highbd_sad32x32_avg_bits10, vpx_highbd_10_variance32x32,
1728 vpx_highbd_10_sub_pixel_variance32x32,
1729 vpx_highbd_10_sub_pixel_avg_variance32x32,
1730 vpx_highbd_sad32x32x4d_bits10)
1732 HIGHBD_BFP(BLOCK_64X64, vpx_highbd_sad64x64_bits10,
1733 vpx_highbd_sad64x64_avg_bits10, vpx_highbd_10_variance64x64,
1734 vpx_highbd_10_sub_pixel_variance64x64,
1735 vpx_highbd_10_sub_pixel_avg_variance64x64,
1736 vpx_highbd_sad64x64x4d_bits10)
1738 HIGHBD_BFP(BLOCK_16X16, vpx_highbd_sad16x16_bits10,
1739 vpx_highbd_sad16x16_avg_bits10, vpx_highbd_10_variance16x16,
1740 vpx_highbd_10_sub_pixel_variance16x16,
1741 vpx_highbd_10_sub_pixel_avg_variance16x16,
1742 vpx_highbd_sad16x16x4d_bits10)
1744 HIGHBD_BFP(BLOCK_16X8, vpx_highbd_sad16x8_bits10,
1745 vpx_highbd_sad16x8_avg_bits10, vpx_highbd_10_variance16x8,
1746 vpx_highbd_10_sub_pixel_variance16x8,
1747 vpx_highbd_10_sub_pixel_avg_variance16x8,
1748 vpx_highbd_sad16x8x4d_bits10)
1750 HIGHBD_BFP(BLOCK_8X16, vpx_highbd_sad8x16_bits10,
1751 vpx_highbd_sad8x16_avg_bits10, vpx_highbd_10_variance8x16,
1752 vpx_highbd_10_sub_pixel_variance8x16,
1753 vpx_highbd_10_sub_pixel_avg_variance8x16,
1754 vpx_highbd_sad8x16x4d_bits10)
1756 HIGHBD_BFP(BLOCK_8X8, vpx_highbd_sad8x8_bits10,
1757 vpx_highbd_sad8x8_avg_bits10, vpx_highbd_10_variance8x8,
1758 vpx_highbd_10_sub_pixel_variance8x8,
1759 vpx_highbd_10_sub_pixel_avg_variance8x8,
1760 vpx_highbd_sad8x8x4d_bits10)
1762 HIGHBD_BFP(BLOCK_8X4, vpx_highbd_sad8x4_bits10,
1763 vpx_highbd_sad8x4_avg_bits10, vpx_highbd_10_variance8x4,
1764 vpx_highbd_10_sub_pixel_variance8x4,
1765 vpx_highbd_10_sub_pixel_avg_variance8x4,
1766 vpx_highbd_sad8x4x4d_bits10)
1768 HIGHBD_BFP(BLOCK_4X8, vpx_highbd_sad4x8_bits10,
1769 vpx_highbd_sad4x8_avg_bits10, vpx_highbd_10_variance4x8,
1770 vpx_highbd_10_sub_pixel_variance4x8,
1771 vpx_highbd_10_sub_pixel_avg_variance4x8,
1772 vpx_highbd_sad4x8x4d_bits10)
1774 HIGHBD_BFP(BLOCK_4X4, vpx_highbd_sad4x4_bits10,
1775 vpx_highbd_sad4x4_avg_bits10, vpx_highbd_10_variance4x4,
1776 vpx_highbd_10_sub_pixel_variance4x4,
1777 vpx_highbd_10_sub_pixel_avg_variance4x4,
1778 vpx_highbd_sad4x4x4d_bits10)
1782 assert(cm->bit_depth == VPX_BITS_12);
1783 HIGHBD_BFP(BLOCK_32X16, vpx_highbd_sad32x16_bits12,
1784 vpx_highbd_sad32x16_avg_bits12, vpx_highbd_12_variance32x16,
1785 vpx_highbd_12_sub_pixel_variance32x16,
1786 vpx_highbd_12_sub_pixel_avg_variance32x16,
1787 vpx_highbd_sad32x16x4d_bits12)
1789 HIGHBD_BFP(BLOCK_16X32, vpx_highbd_sad16x32_bits12,
1790 vpx_highbd_sad16x32_avg_bits12, vpx_highbd_12_variance16x32,
1791 vpx_highbd_12_sub_pixel_variance16x32,
1792 vpx_highbd_12_sub_pixel_avg_variance16x32,
1793 vpx_highbd_sad16x32x4d_bits12)
1795 HIGHBD_BFP(BLOCK_64X32, vpx_highbd_sad64x32_bits12,
1796 vpx_highbd_sad64x32_avg_bits12, vpx_highbd_12_variance64x32,
1797 vpx_highbd_12_sub_pixel_variance64x32,
1798 vpx_highbd_12_sub_pixel_avg_variance64x32,
1799 vpx_highbd_sad64x32x4d_bits12)
1801 HIGHBD_BFP(BLOCK_32X64, vpx_highbd_sad32x64_bits12,
1802 vpx_highbd_sad32x64_avg_bits12, vpx_highbd_12_variance32x64,
1803 vpx_highbd_12_sub_pixel_variance32x64,
1804 vpx_highbd_12_sub_pixel_avg_variance32x64,
1805 vpx_highbd_sad32x64x4d_bits12)
1807 HIGHBD_BFP(BLOCK_32X32, vpx_highbd_sad32x32_bits12,
1808 vpx_highbd_sad32x32_avg_bits12, vpx_highbd_12_variance32x32,
1809 vpx_highbd_12_sub_pixel_variance32x32,
1810 vpx_highbd_12_sub_pixel_avg_variance32x32,
1811 vpx_highbd_sad32x32x4d_bits12)
1813 HIGHBD_BFP(BLOCK_64X64, vpx_highbd_sad64x64_bits12,
1814 vpx_highbd_sad64x64_avg_bits12, vpx_highbd_12_variance64x64,
1815 vpx_highbd_12_sub_pixel_variance64x64,
1816 vpx_highbd_12_sub_pixel_avg_variance64x64,
1817 vpx_highbd_sad64x64x4d_bits12)
1819 HIGHBD_BFP(BLOCK_16X16, vpx_highbd_sad16x16_bits12,
1820 vpx_highbd_sad16x16_avg_bits12, vpx_highbd_12_variance16x16,
1821 vpx_highbd_12_sub_pixel_variance16x16,
1822 vpx_highbd_12_sub_pixel_avg_variance16x16,
1823 vpx_highbd_sad16x16x4d_bits12)
1825 HIGHBD_BFP(BLOCK_16X8, vpx_highbd_sad16x8_bits12,
1826 vpx_highbd_sad16x8_avg_bits12, vpx_highbd_12_variance16x8,
1827 vpx_highbd_12_sub_pixel_variance16x8,
1828 vpx_highbd_12_sub_pixel_avg_variance16x8,
1829 vpx_highbd_sad16x8x4d_bits12)
1831 HIGHBD_BFP(BLOCK_8X16, vpx_highbd_sad8x16_bits12,
1832 vpx_highbd_sad8x16_avg_bits12, vpx_highbd_12_variance8x16,
1833 vpx_highbd_12_sub_pixel_variance8x16,
1834 vpx_highbd_12_sub_pixel_avg_variance8x16,
1835 vpx_highbd_sad8x16x4d_bits12)
1837 HIGHBD_BFP(BLOCK_8X8, vpx_highbd_sad8x8_bits12,
1838 vpx_highbd_sad8x8_avg_bits12, vpx_highbd_12_variance8x8,
1839 vpx_highbd_12_sub_pixel_variance8x8,
1840 vpx_highbd_12_sub_pixel_avg_variance8x8,
1841 vpx_highbd_sad8x8x4d_bits12)
1843 HIGHBD_BFP(BLOCK_8X4, vpx_highbd_sad8x4_bits12,
1844 vpx_highbd_sad8x4_avg_bits12, vpx_highbd_12_variance8x4,
1845 vpx_highbd_12_sub_pixel_variance8x4,
1846 vpx_highbd_12_sub_pixel_avg_variance8x4,
1847 vpx_highbd_sad8x4x4d_bits12)
1849 HIGHBD_BFP(BLOCK_4X8, vpx_highbd_sad4x8_bits12,
1850 vpx_highbd_sad4x8_avg_bits12, vpx_highbd_12_variance4x8,
1851 vpx_highbd_12_sub_pixel_variance4x8,
1852 vpx_highbd_12_sub_pixel_avg_variance4x8,
1853 vpx_highbd_sad4x8x4d_bits12)
1855 HIGHBD_BFP(BLOCK_4X4, vpx_highbd_sad4x4_bits12,
1856 vpx_highbd_sad4x4_avg_bits12, vpx_highbd_12_variance4x4,
1857 vpx_highbd_12_sub_pixel_variance4x4,
1858 vpx_highbd_12_sub_pixel_avg_variance4x4,
1859 vpx_highbd_sad4x4x4d_bits12)
1864 #endif // CONFIG_VP9_HIGHBITDEPTH
1866 static void realloc_segmentation_maps(VP9_COMP *cpi) {
1867 VP9_COMMON *const cm = &cpi->common;
1869 // Create the encoder segmentation map and set all entries to 0
1870 vpx_free(cpi->segmentation_map);
1871 CHECK_MEM_ERROR(cm, cpi->segmentation_map,
1872 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1874 // Create a map used for cyclic background refresh.
1875 if (cpi->cyclic_refresh) vp9_cyclic_refresh_free(cpi->cyclic_refresh);
1876 CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
1877 vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
1879 // Create a map used to mark inactive areas.
1880 vpx_free(cpi->active_map.map);
1881 CHECK_MEM_ERROR(cm, cpi->active_map.map,
1882 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1884 // And a place holder structure is the coding context
1885 // for use if we want to save and restore it
1886 vpx_free(cpi->coding_context.last_frame_seg_map_copy);
1887 CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
1888 vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1891 static void alloc_copy_partition_data(VP9_COMP *cpi) {
1892 VP9_COMMON *const cm = &cpi->common;
1893 if (cpi->prev_partition == NULL) {
1894 CHECK_MEM_ERROR(cm, cpi->prev_partition,
1895 (BLOCK_SIZE *)vpx_calloc(cm->mi_stride * cm->mi_rows,
1896 sizeof(*cpi->prev_partition)));
1898 if (cpi->prev_segment_id == NULL) {
1900 cm, cpi->prev_segment_id,
1901 (int8_t *)vpx_calloc((cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1),
1902 sizeof(*cpi->prev_segment_id)));
1904 if (cpi->prev_variance_low == NULL) {
1905 CHECK_MEM_ERROR(cm, cpi->prev_variance_low,
1906 (uint8_t *)vpx_calloc(
1907 (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1) * 25,
1908 sizeof(*cpi->prev_variance_low)));
1910 if (cpi->copied_frame_cnt == NULL) {
1912 cm, cpi->copied_frame_cnt,
1913 (uint8_t *)vpx_calloc((cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1),
1914 sizeof(*cpi->copied_frame_cnt)));
1918 void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
1919 VP9_COMMON *const cm = &cpi->common;
1920 RATE_CONTROL *const rc = &cpi->rc;
1921 int last_w = cpi->oxcf.width;
1922 int last_h = cpi->oxcf.height;
1924 vp9_init_quantizer(cpi);
1925 if (cm->profile != oxcf->profile) cm->profile = oxcf->profile;
1926 cm->bit_depth = oxcf->bit_depth;
1927 cm->color_space = oxcf->color_space;
1928 cm->color_range = oxcf->color_range;
1930 cpi->target_level = oxcf->target_level;
1931 cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
1932 set_level_constraint(&cpi->level_constraint,
1933 get_level_index(cpi->target_level));
1935 if (cm->profile <= PROFILE_1)
1936 assert(cm->bit_depth == VPX_BITS_8);
1938 assert(cm->bit_depth > VPX_BITS_8);
1941 #if CONFIG_VP9_HIGHBITDEPTH
1942 cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
1943 #endif // CONFIG_VP9_HIGHBITDEPTH
1945 if ((oxcf->pass == 0) && (oxcf->rc_mode == VPX_Q)) {
1946 rc->baseline_gf_interval = FIXED_GF_INTERVAL;
1948 rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
1951 cpi->refresh_golden_frame = 0;
1952 cpi->refresh_last_frame = 1;
1953 cm->refresh_frame_context = 1;
1954 cm->reset_frame_context = 0;
1956 vp9_reset_segment_features(&cm->seg);
1957 vp9_set_high_precision_mv(cpi, 0);
1962 for (i = 0; i < MAX_SEGMENTS; i++)
1963 cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
1965 cpi->encode_breakout = cpi->oxcf.encode_breakout;
1967 set_rc_buffer_sizes(rc, &cpi->oxcf);
1969 // Under a configuration change, where maximum_buffer_size may change,
1970 // keep buffer level clipped to the maximum allowed buffer size.
1971 rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
1972 rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size);
1974 // Set up frame rate and related parameters rate control values.
1975 vp9_new_framerate(cpi, cpi->framerate);
1977 // Set absolute upper and lower quality limits
1978 rc->worst_quality = cpi->oxcf.worst_allowed_q;
1979 rc->best_quality = cpi->oxcf.best_allowed_q;
1981 cm->interp_filter = cpi->sf.default_interp_filter;
1983 if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) {
1984 cm->render_width = cpi->oxcf.render_width;
1985 cm->render_height = cpi->oxcf.render_height;
1987 cm->render_width = cpi->oxcf.width;
1988 cm->render_height = cpi->oxcf.height;
1990 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1991 cm->width = cpi->oxcf.width;
1992 cm->height = cpi->oxcf.height;
1993 cpi->external_resize = 1;
1996 if (cpi->initial_width) {
1997 int new_mi_size = 0;
1998 vp9_set_mb_mi(cm, cm->width, cm->height);
1999 new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
2000 if (cm->mi_alloc_size < new_mi_size) {
2001 vp9_free_context_buffers(cm);
2002 alloc_compressor_data(cpi);
2003 realloc_segmentation_maps(cpi);
2004 cpi->initial_width = cpi->initial_height = 0;
2005 cpi->external_resize = 0;
2006 } else if (cm->mi_alloc_size == new_mi_size &&
2007 (cpi->oxcf.width > last_w || cpi->oxcf.height > last_h)) {
2008 vp9_alloc_loop_filter(cm);
2012 if (cm->current_video_frame == 0 || last_w != cpi->oxcf.width ||
2013 last_h != cpi->oxcf.height)
2014 update_frame_size(cpi);
2016 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
2017 memset(cpi->consec_zero_mv, 0,
2018 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
2019 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
2020 vp9_cyclic_refresh_reset_resize(cpi);
2025 if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
2026 ((cpi->svc.number_temporal_layers > 1 ||
2027 cpi->svc.number_spatial_layers > 1) &&
2028 cpi->oxcf.pass != 1)) {
2029 vp9_update_layer_context_change_config(cpi,
2030 (int)cpi->oxcf.target_bandwidth);
2033 // Check for resetting the rc flags (rc_1_frame, rc_2_frame) if the
2034 // configuration change has a large change in avg_frame_bandwidth.
2035 // For SVC check for resetting based on spatial layer average bandwidth.
2036 // Also reset buffer level to optimal level.
2037 if (cm->current_video_frame > 0) {
2039 vp9_svc_check_reset_layer_rc_flag(cpi);
2041 if (rc->avg_frame_bandwidth > (3 * rc->last_avg_frame_bandwidth >> 1) ||
2042 rc->avg_frame_bandwidth < (rc->last_avg_frame_bandwidth >> 1)) {
2045 rc->bits_off_target = rc->optimal_buffer_level;
2046 rc->buffer_level = rc->optimal_buffer_level;
2051 cpi->alt_ref_source = NULL;
2052 rc->is_src_frame_alt_ref = 0;
2055 // Experimental RD Code
2056 cpi->frame_distortion = 0;
2057 cpi->last_frame_distortion = 0;
2060 set_tile_limits(cpi);
2062 cpi->ext_refresh_frame_flags_pending = 0;
2063 cpi->ext_refresh_frame_context_pending = 0;
2065 #if CONFIG_VP9_HIGHBITDEPTH
2066 highbd_set_var_fns(cpi);
2069 vp9_set_row_mt(cpi);
2073 #define M_LOG2_E 0.693147180559945309417
2075 #define log2f(x) (log(x) / (float)M_LOG2_E)
2077 /***********************************************************************
2078 * Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts' *
2079 ***********************************************************************
2080 * The following 2 functions ('cal_nmvjointsadcost' and *
2081 * 'cal_nmvsadcosts') are used to calculate cost lookup tables *
2082 * used by 'vp9_diamond_search_sad'. The C implementation of the *
2083 * function is generic, but the AVX intrinsics optimised version *
2084 * relies on the following properties of the computed tables: *
2085 * For cal_nmvjointsadcost: *
2086 * - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3] *
2087 * For cal_nmvsadcosts: *
2088 * - For all i: mvsadcost[0][i] == mvsadcost[1][i] *
2089 * (Equal costs for both components) *
2090 * - For all i: mvsadcost[0][i] == mvsadcost[0][-i] *
2091 * (Cost function is even) *
2092 * If these do not hold, then the AVX optimised version of the *
2093 * 'vp9_diamond_search_sad' function cannot be used as it is, in which *
2094 * case you can revert to using the C function instead. *
2095 ***********************************************************************/
2097 static void cal_nmvjointsadcost(int *mvjointsadcost) {
2098 /*********************************************************************
2099 * Warning: Read the comments above before modifying this function *
2100 *********************************************************************/
2101 mvjointsadcost[0] = 600;
2102 mvjointsadcost[1] = 300;
2103 mvjointsadcost[2] = 300;
2104 mvjointsadcost[3] = 300;
2107 static void cal_nmvsadcosts(int *mvsadcost[2]) {
2108 /*********************************************************************
2109 * Warning: Read the comments above before modifying this function *
2110 *********************************************************************/
2113 mvsadcost[0][0] = 0;
2114 mvsadcost[1][0] = 0;
2117 double z = 256 * (2 * (log2f(8 * i) + .6));
2118 mvsadcost[0][i] = (int)z;
2119 mvsadcost[1][i] = (int)z;
2120 mvsadcost[0][-i] = (int)z;
2121 mvsadcost[1][-i] = (int)z;
2122 } while (++i <= MV_MAX);
2125 static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
2128 mvsadcost[0][0] = 0;
2129 mvsadcost[1][0] = 0;
2132 double z = 256 * (2 * (log2f(8 * i) + .6));
2133 mvsadcost[0][i] = (int)z;
2134 mvsadcost[1][i] = (int)z;
2135 mvsadcost[0][-i] = (int)z;
2136 mvsadcost[1][-i] = (int)z;
2137 } while (++i <= MV_MAX);
2140 VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
2141 BufferPool *const pool) {
2143 VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
2144 VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
2146 if (!cm) return NULL;
2150 if (setjmp(cm->error.jmp)) {
2151 cm->error.setjmp = 0;
2152 vp9_remove_compressor(cpi);
2156 cm->error.setjmp = 1;
2157 cm->alloc_mi = vp9_enc_alloc_mi;
2158 cm->free_mi = vp9_enc_free_mi;
2159 cm->setup_mi = vp9_enc_setup_mi;
2161 CHECK_MEM_ERROR(cm, cm->fc, (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
2163 cm, cm->frame_contexts,
2164 (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS, sizeof(*cm->frame_contexts)));
2167 cpi->resize_state = ORIG;
2168 cpi->external_resize = 0;
2169 cpi->resize_avg_qp = 0;
2170 cpi->resize_buffer_underflow = 0;
2171 cpi->use_skin_detection = 0;
2172 cpi->common.buffer_pool = pool;
2174 cpi->force_update_segmentation = 0;
2176 init_config(cpi, oxcf);
2177 vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
2179 cm->current_video_frame = 0;
2180 cpi->partition_search_skippable_frame = 0;
2181 cpi->tile_data = NULL;
2183 realloc_segmentation_maps(cpi);
2187 vpx_calloc(cm->mi_rows * cm->mi_cols, sizeof(cpi->skin_map[0])));
2189 #if !CONFIG_REALTIME_ONLY
2190 CHECK_MEM_ERROR(cm, cpi->alt_ref_aq, vp9_alt_ref_aq_create());
2194 cm, cpi->consec_zero_mv,
2195 vpx_calloc(cm->mi_rows * cm->mi_cols, sizeof(*cpi->consec_zero_mv)));
2197 CHECK_MEM_ERROR(cm, cpi->nmvcosts[0],
2198 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0])));
2199 CHECK_MEM_ERROR(cm, cpi->nmvcosts[1],
2200 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1])));
2201 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0],
2202 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0])));
2203 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1],
2204 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1])));
2205 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0],
2206 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0])));
2207 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1],
2208 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1])));
2209 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0],
2210 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0])));
2211 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
2212 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
2214 for (i = 0; i < (sizeof(cpi->mbgraph_stats) / sizeof(cpi->mbgraph_stats[0]));
2217 cm, cpi->mbgraph_stats[i].mb_stats,
2218 vpx_calloc(cm->MBs * sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
2221 #if CONFIG_FP_MB_STATS
2222 cpi->use_fp_mb_stats = 0;
2223 if (cpi->use_fp_mb_stats) {
2224 // a place holder used to store the first pass mb stats in the first pass
2225 CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
2226 vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
2228 cpi->twopass.frame_mb_stats_buf = NULL;
2232 cpi->refresh_alt_ref_frame = 0;
2233 cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
2235 init_level_info(&cpi->level_info);
2236 init_level_constraint(&cpi->level_constraint);
2238 #if CONFIG_INTERNAL_STATS
2239 cpi->b_calculate_blockiness = 1;
2240 cpi->b_calculate_consistency = 1;
2241 cpi->total_inconsistency = 0;
2242 cpi->psnr.worst = 100.0;
2243 cpi->worst_ssim = 100.0;
2248 if (cpi->b_calculate_psnr) {
2249 cpi->total_sq_error = 0;
2250 cpi->total_samples = 0;
2252 cpi->totalp_sq_error = 0;
2253 cpi->totalp_samples = 0;
2255 cpi->tot_recode_hits = 0;
2256 cpi->summed_quality = 0;
2257 cpi->summed_weights = 0;
2258 cpi->summedp_quality = 0;
2259 cpi->summedp_weights = 0;
2262 cpi->fastssim.worst = 100.0;
2264 cpi->psnrhvs.worst = 100.0;
2266 if (cpi->b_calculate_blockiness) {
2267 cpi->total_blockiness = 0;
2268 cpi->worst_blockiness = 0.0;
2271 if (cpi->b_calculate_consistency) {
2272 CHECK_MEM_ERROR(cm, cpi->ssim_vars,
2273 vpx_calloc(cpi->common.mi_rows * cpi->common.mi_cols,
2274 sizeof(*cpi->ssim_vars) * 4));
2275 cpi->worst_consistency = 100.0;
2277 cpi->ssim_vars = NULL;
2282 cpi->first_time_stamp_ever = INT64_MAX;
2284 /*********************************************************************
2285 * Warning: Read the comments around 'cal_nmvjointsadcost' and *
2286 * 'cal_nmvsadcosts' before modifying how these tables are computed. *
2287 *********************************************************************/
2288 cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
2289 cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
2290 cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
2291 cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
2292 cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
2293 cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
2295 cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
2296 cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
2297 cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
2298 cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
2299 cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
2301 #if CONFIG_VP9_TEMPORAL_DENOISING
2302 #ifdef OUTPUT_YUV_DENOISED
2303 yuv_denoised_file = fopen("denoised.yuv", "ab");
2306 #ifdef OUTPUT_YUV_SKINMAP
2307 yuv_skinmap_file = fopen("skinmap.yuv", "wb");
2309 #ifdef OUTPUT_YUV_REC
2310 yuv_rec_file = fopen("rec.yuv", "wb");
2312 #ifdef OUTPUT_YUV_SVC_SRC
2313 yuv_svc_src[0] = fopen("svc_src_0.yuv", "wb");
2314 yuv_svc_src[1] = fopen("svc_src_1.yuv", "wb");
2315 yuv_svc_src[2] = fopen("svc_src_2.yuv", "wb");
2319 framepsnr = fopen("framepsnr.stt", "a");
2320 kf_list = fopen("kf_list.stt", "w");
2323 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
2325 #if !CONFIG_REALTIME_ONLY
2326 if (oxcf->pass == 1) {
2327 vp9_init_first_pass(cpi);
2328 } else if (oxcf->pass == 2) {
2329 const size_t packet_sz = sizeof(FIRSTPASS_STATS);
2330 const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
2332 if (cpi->svc.number_spatial_layers > 1 ||
2333 cpi->svc.number_temporal_layers > 1) {
2334 FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf;
2335 FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = { 0 };
2338 for (i = 0; i < oxcf->ss_number_layers; ++i) {
2339 FIRSTPASS_STATS *const last_packet_for_layer =
2340 &stats[packets - oxcf->ss_number_layers + i];
2341 const int layer_id = (int)last_packet_for_layer->spatial_layer_id;
2342 const int packets_in_layer = (int)last_packet_for_layer->count + 1;
2343 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) {
2344 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id];
2346 vpx_free(lc->rc_twopass_stats_in.buf);
2348 lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz;
2349 CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf,
2350 vpx_malloc(lc->rc_twopass_stats_in.sz));
2351 lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf;
2352 lc->twopass.stats_in = lc->twopass.stats_in_start;
2353 lc->twopass.stats_in_end =
2354 lc->twopass.stats_in_start + packets_in_layer - 1;
2355 stats_copy[layer_id] = lc->rc_twopass_stats_in.buf;
2359 for (i = 0; i < packets; ++i) {
2360 const int layer_id = (int)stats[i].spatial_layer_id;
2361 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers &&
2362 stats_copy[layer_id] != NULL) {
2363 *stats_copy[layer_id] = stats[i];
2364 ++stats_copy[layer_id];
2368 vp9_init_second_pass_spatial_svc(cpi);
2370 #if CONFIG_FP_MB_STATS
2371 if (cpi->use_fp_mb_stats) {
2372 const size_t psz = cpi->common.MBs * sizeof(uint8_t);
2373 const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
2375 cpi->twopass.firstpass_mb_stats.mb_stats_start =
2376 oxcf->firstpass_mb_stats_in.buf;
2377 cpi->twopass.firstpass_mb_stats.mb_stats_end =
2378 cpi->twopass.firstpass_mb_stats.mb_stats_start +
2379 (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
2383 cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
2384 cpi->twopass.stats_in = cpi->twopass.stats_in_start;
2385 cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
2387 vp9_init_second_pass(cpi);
2390 #endif // !CONFIG_REALTIME_ONLY
2392 cpi->mb_wiener_var_cols = 0;
2393 cpi->mb_wiener_var_rows = 0;
2394 cpi->mb_wiener_variance = NULL;
2396 vp9_set_speed_features_framesize_independent(cpi, oxcf->speed);
2397 vp9_set_speed_features_framesize_dependent(cpi, oxcf->speed);
2400 const int bsize = BLOCK_16X16;
2401 const int w = num_8x8_blocks_wide_lookup[bsize];
2402 const int h = num_8x8_blocks_high_lookup[bsize];
2403 const int num_cols = (cm->mi_cols + w - 1) / w;
2404 const int num_rows = (cm->mi_rows + h - 1) / h;
2405 CHECK_MEM_ERROR(cm, cpi->mi_ssim_rdmult_scaling_factors,
2406 vpx_calloc(num_rows * num_cols,
2407 sizeof(*cpi->mi_ssim_rdmult_scaling_factors)));
2410 cpi->kmeans_data_arr_alloc = 0;
2411 #if CONFIG_NON_GREEDY_MV
2412 cpi->feature_score_loc_alloc = 0;
2414 #endif // CONFIG_NON_GREEDY_MV
2415 for (i = 0; i < MAX_ARF_GOP_SIZE; ++i) cpi->tpl_stats[i].tpl_stats_ptr = NULL;
2417 // Allocate memory to store variances for a frame.
2418 CHECK_MEM_ERROR(cm, cpi->source_diff_var, vpx_calloc(cm->MBs, sizeof(diff)));
2419 cpi->source_var_thresh = 0;
2420 cpi->frames_till_next_var_check = 0;
2422 #define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX4DF) \
2423 cpi->fn_ptr[BT].sdf = SDF; \
2424 cpi->fn_ptr[BT].sdaf = SDAF; \
2425 cpi->fn_ptr[BT].vf = VF; \
2426 cpi->fn_ptr[BT].svf = SVF; \
2427 cpi->fn_ptr[BT].svaf = SVAF; \
2428 cpi->fn_ptr[BT].sdx4df = SDX4DF;
2430 BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg, vpx_variance32x16,
2431 vpx_sub_pixel_variance32x16, vpx_sub_pixel_avg_variance32x16,
2434 BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg, vpx_variance16x32,
2435 vpx_sub_pixel_variance16x32, vpx_sub_pixel_avg_variance16x32,
2438 BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg, vpx_variance64x32,
2439 vpx_sub_pixel_variance64x32, vpx_sub_pixel_avg_variance64x32,
2442 BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg, vpx_variance32x64,
2443 vpx_sub_pixel_variance32x64, vpx_sub_pixel_avg_variance32x64,
2446 BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg, vpx_variance32x32,
2447 vpx_sub_pixel_variance32x32, vpx_sub_pixel_avg_variance32x32,
2450 BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg, vpx_variance64x64,
2451 vpx_sub_pixel_variance64x64, vpx_sub_pixel_avg_variance64x64,
2454 BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg, vpx_variance16x16,
2455 vpx_sub_pixel_variance16x16, vpx_sub_pixel_avg_variance16x16,
2458 BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg, vpx_variance16x8,
2459 vpx_sub_pixel_variance16x8, vpx_sub_pixel_avg_variance16x8,
2462 BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg, vpx_variance8x16,
2463 vpx_sub_pixel_variance8x16, vpx_sub_pixel_avg_variance8x16,
2466 BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg, vpx_variance8x8,
2467 vpx_sub_pixel_variance8x8, vpx_sub_pixel_avg_variance8x8, vpx_sad8x8x4d)
2469 BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg, vpx_variance8x4,
2470 vpx_sub_pixel_variance8x4, vpx_sub_pixel_avg_variance8x4, vpx_sad8x4x4d)
2472 BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg, vpx_variance4x8,
2473 vpx_sub_pixel_variance4x8, vpx_sub_pixel_avg_variance4x8, vpx_sad4x8x4d)
2475 BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg, vpx_variance4x4,
2476 vpx_sub_pixel_variance4x4, vpx_sub_pixel_avg_variance4x4, vpx_sad4x4x4d)
2478 #if CONFIG_VP9_HIGHBITDEPTH
2479 highbd_set_var_fns(cpi);
2482 /* vp9_init_quantizer() is first called here. Add check in
2483 * vp9_frame_init_quantizer() so that vp9_init_quantizer is only
2484 * called later when needed. This will avoid unnecessary calls of
2485 * vp9_init_quantizer() for every frame.
2487 vp9_init_quantizer(cpi);
2489 vp9_loop_filter_init(cm);
2491 // Set up the unit scaling factor used during motion search.
2492 #if CONFIG_VP9_HIGHBITDEPTH
2493 vp9_setup_scale_factors_for_frame(&cpi->me_sf, cm->width, cm->height,
2494 cm->width, cm->height,
2495 cm->use_highbitdepth);
2497 vp9_setup_scale_factors_for_frame(&cpi->me_sf, cm->width, cm->height,
2498 cm->width, cm->height);
2499 #endif // CONFIG_VP9_HIGHBITDEPTH
2500 cpi->td.mb.me_sf = &cpi->me_sf;
2502 cm->error.setjmp = 0;
2507 #if CONFIG_INTERNAL_STATS
2508 #define SNPRINT(H, T) snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
2510 #define SNPRINT2(H, T, V) \
2511 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
2512 #endif // CONFIG_INTERNAL_STATS
2514 void vp9_remove_compressor(VP9_COMP *cpi) {
2516 unsigned int i, frame;
2521 #if CONFIG_INTERNAL_STATS
2522 vpx_free(cpi->ssim_vars);
2526 if (cm->current_video_frame > 0) {
2527 #if CONFIG_INTERNAL_STATS
2528 vpx_clear_system_state();
2530 if (cpi->oxcf.pass != 1) {
2531 char headings[512] = { 0 };
2532 char results[512] = { 0 };
2533 FILE *f = fopen("opsnr.stt", "a");
2534 double time_encoded =
2535 (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
2537 double total_encode_time =
2538 (cpi->time_receive_data + cpi->time_compress_data) / 1000.000;
2540 (double)cpi->bytes * (double)8 / (double)1000 / time_encoded;
2541 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
2542 const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
2543 const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
2545 if (cpi->b_calculate_psnr) {
2546 const double total_psnr = vpx_sse_to_psnr(
2547 (double)cpi->total_samples, peak, (double)cpi->total_sq_error);
2548 const double totalp_psnr = vpx_sse_to_psnr(
2549 (double)cpi->totalp_samples, peak, (double)cpi->totalp_sq_error);
2550 const double total_ssim =
2551 100 * pow(cpi->summed_quality / cpi->summed_weights, 8.0);
2552 const double totalp_ssim =
2553 100 * pow(cpi->summedp_quality / cpi->summedp_weights, 8.0);
2555 snprintf(headings, sizeof(headings),
2556 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
2557 "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
2558 "WstPsnr\tWstSsim\tWstFast\tWstHVS\t"
2559 "AVPsnrY\tAPsnrCb\tAPsnrCr");
2560 snprintf(results, sizeof(results),
2561 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2562 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2563 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
2564 "%7.3f\t%7.3f\t%7.3f",
2565 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
2566 cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr, total_ssim,
2567 totalp_ssim, cpi->fastssim.stat[ALL] / cpi->count,
2568 cpi->psnrhvs.stat[ALL] / cpi->count, cpi->psnr.worst,
2569 cpi->worst_ssim, cpi->fastssim.worst, cpi->psnrhvs.worst,
2570 cpi->psnr.stat[Y] / cpi->count, cpi->psnr.stat[U] / cpi->count,
2571 cpi->psnr.stat[V] / cpi->count);
2573 if (cpi->b_calculate_blockiness) {
2574 SNPRINT(headings, "\t Block\tWstBlck");
2575 SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
2576 SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
2579 if (cpi->b_calculate_consistency) {
2580 double consistency =
2581 vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
2582 (double)cpi->total_inconsistency);
2584 SNPRINT(headings, "\tConsist\tWstCons");
2585 SNPRINT2(results, "\t%7.3f", consistency);
2586 SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
2589 fprintf(f, "%s\t Time\tRcErr\tAbsErr\n", headings);
2590 fprintf(f, "%s\t%8.0f\t%7.2f\t%7.2f\n", results, total_encode_time,
2591 rate_err, fabs(rate_err));
2600 printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
2601 printf("\n_frames recive_data encod_mb_row compress_frame Total\n");
2602 printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
2603 cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
2604 cpi->time_compress_data / 1000,
2605 (cpi->time_receive_data + cpi->time_compress_data) / 1000);
2610 #if CONFIG_VP9_TEMPORAL_DENOISING
2611 vp9_denoiser_free(&(cpi->denoiser));
2614 if (cpi->kmeans_data_arr_alloc) {
2615 #if CONFIG_MULTITHREAD
2616 pthread_mutex_destroy(&cpi->kmeans_mutex);
2618 vpx_free(cpi->kmeans_data_arr);
2621 #if CONFIG_NON_GREEDY_MV
2622 vpx_free(cpi->feature_score_loc_arr);
2623 vpx_free(cpi->feature_score_loc_sort);
2624 vpx_free(cpi->feature_score_loc_heap);
2625 vpx_free(cpi->select_mv_arr);
2627 for (frame = 0; frame < MAX_ARF_GOP_SIZE; ++frame) {
2628 #if CONFIG_NON_GREEDY_MV
2630 for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
2632 for (sqr_bsize = 0; sqr_bsize < SQUARE_BLOCK_SIZES; ++sqr_bsize) {
2633 vpx_free(cpi->tpl_stats[frame].pyramid_mv_arr[rf_idx][sqr_bsize]);
2635 vpx_free(cpi->tpl_stats[frame].mv_mode_arr[rf_idx]);
2636 vpx_free(cpi->tpl_stats[frame].rd_diff_arr[rf_idx]);
2639 vpx_free(cpi->tpl_stats[frame].tpl_stats_ptr);
2640 cpi->tpl_stats[frame].is_valid = 0;
2643 for (t = 0; t < cpi->num_workers; ++t) {
2644 VPxWorker *const worker = &cpi->workers[t];
2645 EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
2647 // Deallocate allocated threads.
2648 vpx_get_worker_interface()->end(worker);
2650 // Deallocate allocated thread data.
2651 if (t < cpi->num_workers - 1) {
2652 vpx_free(thread_data->td->counts);
2653 vp9_free_pc_tree(thread_data->td);
2654 vpx_free(thread_data->td);
2657 vpx_free(cpi->tile_thr_data);
2658 vpx_free(cpi->workers);
2659 vp9_row_mt_mem_dealloc(cpi);
2661 if (cpi->num_workers > 1) {
2662 vp9_loop_filter_dealloc(&cpi->lf_row_sync);
2663 vp9_bitstream_encode_tiles_buffer_dealloc(cpi);
2666 #if !CONFIG_REALTIME_ONLY
2667 vp9_alt_ref_aq_destroy(cpi->alt_ref_aq);
2670 dealloc_compressor_data(cpi);
2672 for (i = 0; i < sizeof(cpi->mbgraph_stats) / sizeof(cpi->mbgraph_stats[0]);
2674 vpx_free(cpi->mbgraph_stats[i].mb_stats);
2677 #if CONFIG_FP_MB_STATS
2678 if (cpi->use_fp_mb_stats) {
2679 vpx_free(cpi->twopass.frame_mb_stats_buf);
2680 cpi->twopass.frame_mb_stats_buf = NULL;
2684 vp9_remove_common(cm);
2685 vp9_free_ref_frame_buffers(cm->buffer_pool);
2686 #if CONFIG_VP9_POSTPROC
2687 vp9_free_postproc_buffers(cm);
2691 #if CONFIG_VP9_TEMPORAL_DENOISING
2692 #ifdef OUTPUT_YUV_DENOISED
2693 fclose(yuv_denoised_file);
2696 #ifdef OUTPUT_YUV_SKINMAP
2697 fclose(yuv_skinmap_file);
2699 #ifdef OUTPUT_YUV_REC
2700 fclose(yuv_rec_file);
2702 #ifdef OUTPUT_YUV_SVC_SRC
2703 fclose(yuv_svc_src[0]);
2704 fclose(yuv_svc_src[1]);
2705 fclose(yuv_svc_src[2]);
2722 static void generate_psnr_packet(VP9_COMP *cpi) {
2723 struct vpx_codec_cx_pkt pkt;
2726 #if CONFIG_VP9_HIGHBITDEPTH
2727 vpx_calc_highbd_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, &psnr,
2728 cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
2730 vpx_calc_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, &psnr);
2733 for (i = 0; i < 4; ++i) {
2734 pkt.data.psnr.samples[i] = psnr.samples[i];
2735 pkt.data.psnr.sse[i] = psnr.sse[i];
2736 pkt.data.psnr.psnr[i] = psnr.psnr[i];
2738 pkt.kind = VPX_CODEC_PSNR_PKT;
2741 .layer_context[cpi->svc.spatial_layer_id *
2742 cpi->svc.number_temporal_layers]
2743 .psnr_pkt = pkt.data.psnr;
2745 vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
2748 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) {
2749 if (ref_frame_flags > 7) return -1;
2751 cpi->ref_frame_flags = ref_frame_flags;
2755 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) {
2756 cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0;
2757 cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0;
2758 cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0;
2759 cpi->ext_refresh_frame_flags_pending = 1;
2762 static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(
2763 VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag) {
2764 MV_REFERENCE_FRAME ref_frame = NONE;
2765 if (ref_frame_flag == VP9_LAST_FLAG)
2766 ref_frame = LAST_FRAME;
2767 else if (ref_frame_flag == VP9_GOLD_FLAG)
2768 ref_frame = GOLDEN_FRAME;
2769 else if (ref_frame_flag == VP9_ALT_FLAG)
2770 ref_frame = ALTREF_FRAME;
2772 return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame);
2775 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2776 YV12_BUFFER_CONFIG *sd) {
2777 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2779 vpx_yv12_copy_frame(cfg, sd);
2786 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2787 YV12_BUFFER_CONFIG *sd) {
2788 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2790 vpx_yv12_copy_frame(sd, cfg);
2797 int vp9_update_entropy(VP9_COMP *cpi, int update) {
2798 cpi->ext_refresh_frame_context = update;
2799 cpi->ext_refresh_frame_context_pending = 1;
2803 #ifdef OUTPUT_YUV_REC
2804 void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
2805 YV12_BUFFER_CONFIG *s = cm->frame_to_show;
2806 uint8_t *src = s->y_buffer;
2809 #if CONFIG_VP9_HIGHBITDEPTH
2810 if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
2811 uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
2814 fwrite(src16, s->y_width, 2, yuv_rec_file);
2815 src16 += s->y_stride;
2818 src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
2822 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2823 src16 += s->uv_stride;
2826 src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
2830 fwrite(src16, s->uv_width, 2, yuv_rec_file);
2831 src16 += s->uv_stride;
2834 fflush(yuv_rec_file);
2837 #endif // CONFIG_VP9_HIGHBITDEPTH
2840 fwrite(src, s->y_width, 1, yuv_rec_file);
2848 fwrite(src, s->uv_width, 1, yuv_rec_file);
2849 src += s->uv_stride;
2856 fwrite(src, s->uv_width, 1, yuv_rec_file);
2857 src += s->uv_stride;
2860 fflush(yuv_rec_file);
2864 #if CONFIG_VP9_HIGHBITDEPTH
2865 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2866 YV12_BUFFER_CONFIG *dst,
2869 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2870 YV12_BUFFER_CONFIG *dst) {
2871 #endif // CONFIG_VP9_HIGHBITDEPTH
2872 // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
2874 const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,
2876 const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
2877 const int src_widths[3] = { src->y_crop_width, src->uv_crop_width,
2878 src->uv_crop_width };
2879 const int src_heights[3] = { src->y_crop_height, src->uv_crop_height,
2880 src->uv_crop_height };
2881 uint8_t *const dsts[3] = { dst->y_buffer, dst->u_buffer, dst->v_buffer };
2882 const int dst_strides[3] = { dst->y_stride, dst->uv_stride, dst->uv_stride };
2883 const int dst_widths[3] = { dst->y_crop_width, dst->uv_crop_width,
2884 dst->uv_crop_width };
2885 const int dst_heights[3] = { dst->y_crop_height, dst->uv_crop_height,
2886 dst->uv_crop_height };
2888 for (i = 0; i < MAX_MB_PLANE; ++i) {
2889 #if CONFIG_VP9_HIGHBITDEPTH
2890 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2891 vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
2892 src_strides[i], dsts[i], dst_heights[i],
2893 dst_widths[i], dst_strides[i], bd);
2895 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2896 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2899 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2900 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2901 #endif // CONFIG_VP9_HIGHBITDEPTH
2903 vpx_extend_frame_borders(dst);
2906 #if CONFIG_VP9_HIGHBITDEPTH
2907 static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
2908 YV12_BUFFER_CONFIG *dst, int bd,
2909 INTERP_FILTER filter_type,
2911 const int src_w = src->y_crop_width;
2912 const int src_h = src->y_crop_height;
2913 const int dst_w = dst->y_crop_width;
2914 const int dst_h = dst->y_crop_height;
2915 const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,
2917 const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
2918 uint8_t *const dsts[3] = { dst->y_buffer, dst->u_buffer, dst->v_buffer };
2919 const int dst_strides[3] = { dst->y_stride, dst->uv_stride, dst->uv_stride };
2920 const InterpKernel *const kernel = vp9_filter_kernels[filter_type];
2923 for (i = 0; i < MAX_MB_PLANE; ++i) {
2924 const int factor = (i == 0 || i == 3 ? 1 : 2);
2925 const int src_stride = src_strides[i];
2926 const int dst_stride = dst_strides[i];
2927 for (y = 0; y < dst_h; y += 16) {
2928 const int y_q4 = y * (16 / factor) * src_h / dst_h + phase_scaler;
2929 for (x = 0; x < dst_w; x += 16) {
2930 const int x_q4 = x * (16 / factor) * src_w / dst_w + phase_scaler;
2931 const uint8_t *src_ptr = srcs[i] +
2932 (y / factor) * src_h / dst_h * src_stride +
2933 (x / factor) * src_w / dst_w;
2934 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2936 if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2937 vpx_highbd_convolve8(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
2938 CONVERT_TO_SHORTPTR(dst_ptr), dst_stride, kernel,
2939 x_q4 & 0xf, 16 * src_w / dst_w, y_q4 & 0xf,
2940 16 * src_h / dst_h, 16 / factor, 16 / factor,
2943 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride, kernel,
2944 x_q4 & 0xf, 16 * src_w / dst_w, y_q4 & 0xf,
2945 16 * src_h / dst_h, 16 / factor, 16 / factor);
2951 vpx_extend_frame_borders(dst);
2953 #endif // CONFIG_VP9_HIGHBITDEPTH
2955 #if !CONFIG_REALTIME_ONLY
2956 static int scale_down(VP9_COMP *cpi, int q) {
2957 RATE_CONTROL *const rc = &cpi->rc;
2958 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2960 assert(frame_is_kf_gf_arf(cpi));
2962 if (rc->frame_size_selector == UNSCALED &&
2963 q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
2964 const int max_size_thresh =
2965 (int)(rate_thresh_mult[SCALE_STEP1] *
2966 VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
2967 scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
2972 static int big_rate_miss_high_threshold(VP9_COMP *cpi) {
2973 const RATE_CONTROL *const rc = &cpi->rc;
2976 if (frame_is_kf_gf_arf(cpi))
2977 big_miss_high = rc->this_frame_target * 3 / 2;
2979 big_miss_high = rc->this_frame_target * 2;
2981 return big_miss_high;
2984 static int big_rate_miss(VP9_COMP *cpi) {
2985 const RATE_CONTROL *const rc = &cpi->rc;
2989 // Ignore for overlay frames
2990 if (rc->is_src_frame_alt_ref) {
2993 big_miss_low = (rc->this_frame_target / 2);
2994 big_miss_high = big_rate_miss_high_threshold(cpi);
2996 return (rc->projected_frame_size > big_miss_high) ||
2997 (rc->projected_frame_size < big_miss_low);
3001 // test in two pass for the first
3002 static int two_pass_first_group_inter(VP9_COMP *cpi) {
3003 if (cpi->oxcf.pass == 2) {
3004 TWO_PASS *const twopass = &cpi->twopass;
3005 GF_GROUP *const gf_group = &twopass->gf_group;
3006 const int gfg_index = gf_group->index;
3008 if (gfg_index == 0) return gf_group->update_type[gfg_index] == LF_UPDATE;
3009 return gf_group->update_type[gfg_index - 1] != LF_UPDATE &&
3010 gf_group->update_type[gfg_index] == LF_UPDATE;
3016 // Function to test for conditions that indicate we should loop
3017 // back and recode a frame.
3018 static int recode_loop_test(VP9_COMP *cpi, int high_limit, int low_limit, int q,
3019 int maxq, int minq) {
3020 const RATE_CONTROL *const rc = &cpi->rc;
3021 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3022 const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
3023 int force_recode = 0;
3025 if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
3026 big_rate_miss(cpi) || (cpi->sf.recode_loop == ALLOW_RECODE) ||
3027 (two_pass_first_group_inter(cpi) &&
3028 (cpi->sf.recode_loop == ALLOW_RECODE_FIRST)) ||
3029 (frame_is_kfgfarf && (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF))) {
3030 if (frame_is_kfgfarf && (oxcf->resize_mode == RESIZE_DYNAMIC) &&
3031 scale_down(cpi, q)) {
3032 // Code this group at a lower resolution.
3033 cpi->resize_pending = 1;
3037 // Force recode for extreme overshoot.
3038 if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
3039 (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF &&
3040 rc->projected_frame_size >= big_rate_miss_high_threshold(cpi))) {
3044 // TODO(agrange) high_limit could be greater than the scale-down threshold.
3045 if ((rc->projected_frame_size > high_limit && q < maxq) ||
3046 (rc->projected_frame_size < low_limit && q > minq)) {
3048 } else if (cpi->oxcf.rc_mode == VPX_CQ) {
3049 // Deal with frame undershoot and whether or not we are
3050 // below the automatically set cq level.
3051 if (q > oxcf->cq_level &&
3052 rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
3057 return force_recode;
3059 #endif // !CONFIG_REALTIME_ONLY
3061 static void update_ref_frames(VP9_COMP *cpi) {
3062 VP9_COMMON *const cm = &cpi->common;
3063 BufferPool *const pool = cm->buffer_pool;
3064 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
3066 if (cpi->rc.show_arf_as_gld) {
3067 int tmp = cpi->alt_fb_idx;
3068 cpi->alt_fb_idx = cpi->gld_fb_idx;
3069 cpi->gld_fb_idx = tmp;
3070 } else if (cm->show_existing_frame) {
3072 cpi->lst_fb_idx = cpi->alt_fb_idx;
3074 stack_pop(gf_group->arf_index_stack, gf_group->stack_size);
3075 --gf_group->stack_size;
3078 // At this point the new frame has been encoded.
3079 // If any buffer copy / swapping is signaled it should be done here.
3080 if (cm->frame_type == KEY_FRAME) {
3081 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
3083 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx],
3085 } else if (vp9_preserve_existing_gf(cpi)) {
3086 // We have decided to preserve the previously existing golden frame as our
3087 // new ARF frame. However, in the short term in function
3088 // vp9_get_refresh_mask() we left it in the GF slot and, if
3089 // we're updating the GF with the current decoded frame, we save it to the
3090 // ARF slot instead.
3091 // We now have to update the ARF with the current frame and swap gld_fb_idx
3092 // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
3093 // slot and, if we're updating the GF, the current frame becomes the new GF.
3096 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx],
3099 tmp = cpi->alt_fb_idx;
3100 cpi->alt_fb_idx = cpi->gld_fb_idx;
3101 cpi->gld_fb_idx = tmp;
3102 } else { /* For non key/golden frames */
3103 if (cpi->refresh_alt_ref_frame) {
3104 int arf_idx = gf_group->top_arf_idx;
3106 // Push new ARF into stack.
3107 stack_push(gf_group->arf_index_stack, cpi->alt_fb_idx,
3108 gf_group->stack_size);
3109 ++gf_group->stack_size;
3111 assert(arf_idx < REF_FRAMES);
3113 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
3114 memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
3115 cpi->interp_filter_selected[0],
3116 sizeof(cpi->interp_filter_selected[0]));
3118 cpi->alt_fb_idx = arf_idx;
3121 if (cpi->refresh_golden_frame) {
3122 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
3124 if (!cpi->rc.is_src_frame_alt_ref)
3125 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
3126 cpi->interp_filter_selected[0],
3127 sizeof(cpi->interp_filter_selected[0]));
3129 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
3130 cpi->interp_filter_selected[ALTREF_FRAME],
3131 sizeof(cpi->interp_filter_selected[ALTREF_FRAME]));
3135 if (cpi->refresh_last_frame) {
3136 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
3138 if (!cpi->rc.is_src_frame_alt_ref)
3139 memcpy(cpi->interp_filter_selected[LAST_FRAME],
3140 cpi->interp_filter_selected[0],
3141 sizeof(cpi->interp_filter_selected[0]));
3144 if (gf_group->update_type[gf_group->index] == MID_OVERLAY_UPDATE) {
3146 stack_pop(gf_group->arf_index_stack, gf_group->stack_size);
3147 --gf_group->stack_size;
3151 void vp9_update_reference_frames(VP9_COMP *cpi) {
3152 update_ref_frames(cpi);
3154 #if CONFIG_VP9_TEMPORAL_DENOISING
3155 vp9_denoiser_update_ref_frame(cpi);
3158 if (is_one_pass_cbr_svc(cpi)) vp9_svc_update_ref_frame(cpi);
3161 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
3162 MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
3163 struct loopfilter *lf = &cm->lf;
3164 int is_reference_frame =
3165 (cm->frame_type == KEY_FRAME || cpi->refresh_last_frame ||
3166 cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame);
3168 cpi->svc.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS)
3169 is_reference_frame = !cpi->svc.non_reference_frame;
3171 // Skip loop filter in show_existing_frame mode.
3172 if (cm->show_existing_frame) {
3173 lf->filter_level = 0;
3178 lf->filter_level = 0;
3179 lf->last_filt_level = 0;
3181 struct vpx_usec_timer timer;
3183 vpx_clear_system_state();
3185 vpx_usec_timer_start(&timer);
3187 if (!cpi->rc.is_src_frame_alt_ref) {
3188 if ((cpi->common.frame_type == KEY_FRAME) &&
3189 (!cpi->rc.this_key_frame_forced)) {
3190 lf->last_filt_level = 0;
3192 vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick);
3193 lf->last_filt_level = lf->filter_level;
3195 lf->filter_level = 0;
3198 vpx_usec_timer_mark(&timer);
3199 cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
3202 if (lf->filter_level > 0 && is_reference_frame) {
3203 vp9_build_mask_frame(cm, lf->filter_level, 0);
3205 if (cpi->num_workers > 1)
3206 vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
3207 lf->filter_level, 0, 0, cpi->workers,
3208 cpi->num_workers, &cpi->lf_row_sync);
3210 vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
3213 vpx_extend_frame_inner_borders(cm->frame_to_show);
3216 static INLINE void alloc_frame_mvs(VP9_COMMON *const cm, int buffer_idx) {
3217 RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
3218 if (new_fb_ptr->mvs == NULL || new_fb_ptr->mi_rows < cm->mi_rows ||
3219 new_fb_ptr->mi_cols < cm->mi_cols) {
3220 vpx_free(new_fb_ptr->mvs);
3221 CHECK_MEM_ERROR(cm, new_fb_ptr->mvs,
3222 (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
3223 sizeof(*new_fb_ptr->mvs)));
3224 new_fb_ptr->mi_rows = cm->mi_rows;
3225 new_fb_ptr->mi_cols = cm->mi_cols;
3229 void vp9_scale_references(VP9_COMP *cpi) {
3230 VP9_COMMON *cm = &cpi->common;
3231 MV_REFERENCE_FRAME ref_frame;
3232 const VP9_REFFRAME ref_mask[3] = { VP9_LAST_FLAG, VP9_GOLD_FLAG,
3235 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3236 // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
3237 if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
3238 BufferPool *const pool = cm->buffer_pool;
3239 const YV12_BUFFER_CONFIG *const ref =
3240 get_ref_frame_buffer(cpi, ref_frame);
3243 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3247 #if CONFIG_VP9_HIGHBITDEPTH
3248 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
3249 RefCntBuffer *new_fb_ptr = NULL;
3250 int force_scaling = 0;
3251 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
3252 if (new_fb == INVALID_IDX) {
3253 new_fb = get_free_fb(cm);
3256 if (new_fb == INVALID_IDX) return;
3257 new_fb_ptr = &pool->frame_bufs[new_fb];
3258 if (force_scaling || new_fb_ptr->buf.y_crop_width != cm->width ||
3259 new_fb_ptr->buf.y_crop_height != cm->height) {
3260 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
3261 cm->subsampling_x, cm->subsampling_y,
3262 cm->use_highbitdepth,
3263 VP9_ENC_BORDER_IN_PIXELS,
3264 cm->byte_alignment, NULL, NULL, NULL))
3265 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3266 "Failed to allocate frame buffer");
3267 scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth,
3269 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
3270 alloc_frame_mvs(cm, new_fb);
3273 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
3274 RefCntBuffer *new_fb_ptr = NULL;
3275 int force_scaling = 0;
3276 int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
3277 if (new_fb == INVALID_IDX) {
3278 new_fb = get_free_fb(cm);
3281 if (new_fb == INVALID_IDX) return;
3282 new_fb_ptr = &pool->frame_bufs[new_fb];
3283 if (force_scaling || new_fb_ptr->buf.y_crop_width != cm->width ||
3284 new_fb_ptr->buf.y_crop_height != cm->height) {
3285 if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
3286 cm->subsampling_x, cm->subsampling_y,
3287 VP9_ENC_BORDER_IN_PIXELS,
3288 cm->byte_alignment, NULL, NULL, NULL))
3289 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3290 "Failed to allocate frame buffer");
3291 vp9_scale_and_extend_frame(ref, &new_fb_ptr->buf, EIGHTTAP, 0);
3292 cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
3293 alloc_frame_mvs(cm, new_fb);
3295 #endif // CONFIG_VP9_HIGHBITDEPTH
3298 RefCntBuffer *buf = NULL;
3299 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3300 // Check for release of scaled reference.
3301 buf_idx = cpi->scaled_ref_idx[ref_frame - 1];
3302 if (buf_idx != INVALID_IDX) {
3303 buf = &pool->frame_bufs[buf_idx];
3305 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3308 buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3309 buf = &pool->frame_bufs[buf_idx];
3310 buf->buf.y_crop_width = ref->y_crop_width;
3311 buf->buf.y_crop_height = ref->y_crop_height;
3312 cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
3316 if (cpi->oxcf.pass != 0 || cpi->use_svc)
3317 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
3322 static void release_scaled_references(VP9_COMP *cpi) {
3323 VP9_COMMON *cm = &cpi->common;
3325 if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
3326 // Only release scaled references under certain conditions:
3327 // if reference will be updated, or if scaled reference has same resolution.
3329 refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
3330 refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
3331 refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
3332 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3333 const int idx = cpi->scaled_ref_idx[i - 1];
3334 if (idx != INVALID_IDX) {
3335 RefCntBuffer *const buf = &cm->buffer_pool->frame_bufs[idx];
3336 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
3337 if (refresh[i - 1] || (buf->buf.y_crop_width == ref->y_crop_width &&
3338 buf->buf.y_crop_height == ref->y_crop_height)) {
3340 cpi->scaled_ref_idx[i - 1] = INVALID_IDX;
3345 for (i = 0; i < REFS_PER_FRAME; ++i) {
3346 const int idx = cpi->scaled_ref_idx[i];
3347 if (idx != INVALID_IDX) {
3348 RefCntBuffer *const buf = &cm->buffer_pool->frame_bufs[idx];
3350 cpi->scaled_ref_idx[i] = INVALID_IDX;
3356 static void full_to_model_count(unsigned int *model_count,
3357 unsigned int *full_count) {
3359 model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
3360 model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
3361 model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
3362 for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
3363 model_count[TWO_TOKEN] += full_count[n];
3364 model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
3367 static void full_to_model_counts(vp9_coeff_count_model *model_count,
3368 vp9_coeff_count *full_count) {
3371 for (i = 0; i < PLANE_TYPES; ++i)
3372 for (j = 0; j < REF_TYPES; ++j)
3373 for (k = 0; k < COEF_BANDS; ++k)
3374 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
3375 full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
3378 #if 0 && CONFIG_INTERNAL_STATS
3379 static void output_frame_level_debug_stats(VP9_COMP *cpi) {
3380 VP9_COMMON *const cm = &cpi->common;
3381 FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
3384 vpx_clear_system_state();
3386 #if CONFIG_VP9_HIGHBITDEPTH
3387 if (cm->use_highbitdepth) {
3388 recon_err = vpx_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3390 recon_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3393 recon_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3394 #endif // CONFIG_VP9_HIGHBITDEPTH
3397 if (cpi->twopass.total_left_stats.coded_error != 0.0) {
3398 double dc_quant_devisor;
3399 #if CONFIG_VP9_HIGHBITDEPTH
3400 switch (cm->bit_depth) {
3402 dc_quant_devisor = 4.0;
3405 dc_quant_devisor = 16.0;
3408 assert(cm->bit_depth == VPX_BITS_12);
3409 dc_quant_devisor = 64.0;
3413 dc_quant_devisor = 4.0;
3416 if (!cm->current_video_frame) {
3417 fprintf(f, "frame, width, height, last ts, last end ts, "
3418 "source_alt_ref_pending, source_alt_ref_active, "
3419 "this_frame_target, projected_frame_size, "
3420 "projected_frame_size / MBs, "
3421 "projected_frame_size - this_frame_target, "
3422 "vbr_bits_off_target, vbr_bits_off_target_fast, "
3423 "twopass.extend_minq, twopass.extend_minq_fast, "
3424 "total_target_vs_actual, "
3425 "starting_buffer_level - bits_off_target, "
3426 "total_actual_bits, base_qindex, q for base_qindex, "
3427 "dc quant, q for active_worst_quality, avg_q, q for oxcf.cq_level, "
3428 "refresh_last_frame, refresh_golden_frame, refresh_alt_ref_frame, "
3429 "frame_type, gfu_boost, "
3430 "twopass.bits_left, "
3431 "twopass.total_left_stats.coded_error, "
3432 "twopass.bits_left / (1 + twopass.total_left_stats.coded_error), "
3433 "tot_recode_hits, recon_err, kf_boost, "
3434 "twopass.kf_zeromotion_pct, twopass.fr_content_type, "
3435 "filter_level, seg.aq_av_offset\n");
3438 fprintf(f, "%10u, %d, %d, %10"PRId64", %10"PRId64", %d, %d, %10d, %10d, "
3439 "%10d, %10d, %10"PRId64", %10"PRId64", %5d, %5d, %10"PRId64", "
3440 "%10"PRId64", %10"PRId64", %10d, %7.2lf, %7.2lf, %7.2lf, %7.2lf, "
3441 "%7.2lf, %6d, %6d, %5d, %5d, %5d, %10"PRId64", %10.3lf, %10lf, %8u, "
3442 "%10"PRId64", %10d, %10d, %10d, %10d, %10d\n",
3443 cpi->common.current_video_frame,
3444 cm->width, cm->height,
3445 cpi->last_time_stamp_seen,
3446 cpi->last_end_time_stamp_seen,
3447 cpi->rc.source_alt_ref_pending,
3448 cpi->rc.source_alt_ref_active,
3449 cpi->rc.this_frame_target,
3450 cpi->rc.projected_frame_size,
3451 cpi->rc.projected_frame_size / cpi->common.MBs,
3452 (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
3453 cpi->rc.vbr_bits_off_target,
3454 cpi->rc.vbr_bits_off_target_fast,
3455 cpi->twopass.extend_minq,
3456 cpi->twopass.extend_minq_fast,
3457 cpi->rc.total_target_vs_actual,
3458 (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
3459 cpi->rc.total_actual_bits, cm->base_qindex,
3460 vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
3461 (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) /
3463 vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality,
3466 vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
3467 cpi->refresh_last_frame, cpi->refresh_golden_frame,
3468 cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
3469 cpi->twopass.bits_left,
3470 cpi->twopass.total_left_stats.coded_error,
3471 cpi->twopass.bits_left /
3472 (1 + cpi->twopass.total_left_stats.coded_error),
3473 cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
3474 cpi->twopass.kf_zeromotion_pct,
3475 cpi->twopass.fr_content_type,
3476 cm->lf.filter_level,
3477 cm->seg.aq_av_offset);
3482 FILE *const fmodes = fopen("Modes.stt", "a");
3485 fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
3486 cm->frame_type, cpi->refresh_golden_frame,
3487 cpi->refresh_alt_ref_frame);
3489 for (i = 0; i < MAX_MODES; ++i)
3490 fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
3492 fprintf(fmodes, "\n");
3499 static void set_mv_search_params(VP9_COMP *cpi) {
3500 const VP9_COMMON *const cm = &cpi->common;
3501 const unsigned int max_mv_def = VPXMIN(cm->width, cm->height);
3503 // Default based on max resolution.
3504 cpi->mv_step_param = vp9_init_search_range(max_mv_def);
3506 if (cpi->sf.mv.auto_mv_step_size) {
3507 if (frame_is_intra_only(cm)) {
3508 // Initialize max_mv_magnitude for use in the first INTER frame
3509 // after a key/intra-only frame.
3510 cpi->max_mv_magnitude = max_mv_def;
3512 if (cm->show_frame) {
3513 // Allow mv_steps to correspond to twice the max mv magnitude found
3514 // in the previous frame, capped by the default max_mv_magnitude based
3516 cpi->mv_step_param = vp9_init_search_range(
3517 VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
3519 cpi->max_mv_magnitude = 0;
3524 static void set_size_independent_vars(VP9_COMP *cpi) {
3525 vp9_set_speed_features_framesize_independent(cpi, cpi->oxcf.speed);
3526 vp9_set_rd_speed_thresholds(cpi);
3527 vp9_set_rd_speed_thresholds_sub8x8(cpi);
3528 cpi->common.interp_filter = cpi->sf.default_interp_filter;
3531 static void set_size_dependent_vars(VP9_COMP *cpi, int *q, int *bottom_index,
3533 VP9_COMMON *const cm = &cpi->common;
3535 // Setup variables that depend on the dimensions of the frame.
3536 vp9_set_speed_features_framesize_dependent(cpi, cpi->oxcf.speed);
3538 // Decide q and q bounds.
3539 *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
3541 if (cpi->oxcf.rc_mode == VPX_CBR && cpi->rc.force_max_q) {
3542 *q = cpi->rc.worst_quality;
3543 cpi->rc.force_max_q = 0;
3546 if (!frame_is_intra_only(cm)) {
3547 vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
3550 #if !CONFIG_REALTIME_ONLY
3551 // Configure experimental use of segmentation for enhanced coding of
3552 // static regions if indicated.
3553 // Only allowed in the second pass of a two pass encode, as it requires
3554 // lagged coding, and if the relevant speed feature flag is set.
3555 if (cpi->oxcf.pass == 2 && cpi->sf.static_segmentation)
3556 configure_static_seg_features(cpi);
3557 #endif // !CONFIG_REALTIME_ONLY
3559 #if CONFIG_VP9_POSTPROC && !(CONFIG_VP9_TEMPORAL_DENOISING)
3560 if (cpi->oxcf.noise_sensitivity > 0) {
3562 switch (cpi->oxcf.noise_sensitivity) {
3563 case 1: l = 20; break;
3564 case 2: l = 40; break;
3565 case 3: l = 60; break;
3567 case 5: l = 100; break;
3568 case 6: l = 150; break;
3570 if (!cpi->common.postproc_state.limits) {
3571 cpi->common.postproc_state.limits =
3572 vpx_calloc(cpi->un_scaled_source->y_width,
3573 sizeof(*cpi->common.postproc_state.limits));
3575 vp9_denoise(&cpi->common, cpi->Source, cpi->Source, l,
3576 cpi->common.postproc_state.limits);
3578 #endif // CONFIG_VP9_POSTPROC
3581 #if CONFIG_VP9_TEMPORAL_DENOISING
3582 static void setup_denoiser_buffer(VP9_COMP *cpi) {
3583 VP9_COMMON *const cm = &cpi->common;
3584 if (cpi->oxcf.noise_sensitivity > 0 &&
3585 !cpi->denoiser.frame_buffer_initialized) {
3586 if (vp9_denoiser_alloc(cm, &cpi->svc, &cpi->denoiser, cpi->use_svc,
3587 cpi->oxcf.noise_sensitivity, cm->width, cm->height,
3588 cm->subsampling_x, cm->subsampling_y,
3589 #if CONFIG_VP9_HIGHBITDEPTH
3590 cm->use_highbitdepth,
3592 VP9_ENC_BORDER_IN_PIXELS))
3593 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3594 "Failed to allocate denoiser");
3599 static void init_motion_estimation(VP9_COMP *cpi) {
3600 int y_stride = cpi->scaled_source.y_stride;
3602 if (cpi->sf.mv.search_method == NSTEP) {
3603 vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
3604 } else if (cpi->sf.mv.search_method == DIAMOND) {
3605 vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
3609 static void set_frame_size(VP9_COMP *cpi) {
3611 VP9_COMMON *const cm = &cpi->common;
3612 VP9EncoderConfig *const oxcf = &cpi->oxcf;
3613 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
3615 #if !CONFIG_REALTIME_ONLY
3616 if (oxcf->pass == 2 && oxcf->rc_mode == VPX_VBR &&
3617 ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
3618 (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
3619 calculate_coded_size(cpi, &oxcf->scaled_frame_width,
3620 &oxcf->scaled_frame_height);
3622 // There has been a change in frame size.
3623 vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
3624 oxcf->scaled_frame_height);
3626 #endif // !CONFIG_REALTIME_ONLY
3628 if (oxcf->pass == 0 && oxcf->rc_mode == VPX_CBR && !cpi->use_svc &&
3629 oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending != 0) {
3630 oxcf->scaled_frame_width =
3631 (oxcf->width * cpi->resize_scale_num) / cpi->resize_scale_den;
3632 oxcf->scaled_frame_height =
3633 (oxcf->height * cpi->resize_scale_num) / cpi->resize_scale_den;
3634 // There has been a change in frame size.
3635 vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
3636 oxcf->scaled_frame_height);
3638 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3639 set_mv_search_params(cpi);
3641 vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
3642 #if CONFIG_VP9_TEMPORAL_DENOISING
3643 // Reset the denoiser on the resized frame.
3644 if (cpi->oxcf.noise_sensitivity > 0) {
3645 vp9_denoiser_free(&(cpi->denoiser));
3646 setup_denoiser_buffer(cpi);
3647 // Dynamic resize is only triggered for non-SVC, so we can force
3648 // golden frame update here as temporary fix to denoiser.
3649 cpi->refresh_golden_frame = 1;
3654 if ((oxcf->pass == 2) && !cpi->use_svc) {
3655 vp9_set_target_rate(cpi);
3658 alloc_frame_mvs(cm, cm->new_fb_idx);
3660 // Reset the frame pointers to the current frame size.
3661 if (vpx_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
3662 cm->subsampling_x, cm->subsampling_y,
3663 #if CONFIG_VP9_HIGHBITDEPTH
3664 cm->use_highbitdepth,
3666 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
3668 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
3669 "Failed to allocate frame buffer");
3671 alloc_util_frame_buffers(cpi);
3672 init_motion_estimation(cpi);
3674 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3675 RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
3676 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3678 ref_buf->idx = buf_idx;
3680 if (buf_idx != INVALID_IDX) {
3681 YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
3683 #if CONFIG_VP9_HIGHBITDEPTH
3684 vp9_setup_scale_factors_for_frame(
3685 &ref_buf->sf, buf->y_crop_width, buf->y_crop_height, cm->width,
3686 cm->height, (buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0);
3688 vp9_setup_scale_factors_for_frame(&ref_buf->sf, buf->y_crop_width,
3689 buf->y_crop_height, cm->width,
3691 #endif // CONFIG_VP9_HIGHBITDEPTH
3692 if (vp9_is_scaled(&ref_buf->sf)) vpx_extend_frame_borders(buf);
3694 ref_buf->buf = NULL;
3698 set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
3701 #if CONFIG_CONSISTENT_RECODE
3702 static void save_encode_params(VP9_COMP *cpi) {
3703 VP9_COMMON *const cm = &cpi->common;
3704 const int tile_cols = 1 << cm->log2_tile_cols;
3705 const int tile_rows = 1 << cm->log2_tile_rows;
3706 int tile_col, tile_row;
3708 RD_OPT *rd_opt = &cpi->rd;
3709 for (i = 0; i < MAX_REF_FRAMES; i++) {
3710 for (j = 0; j < REFERENCE_MODES; j++)
3711 rd_opt->prediction_type_threshes_prev[i][j] =
3712 rd_opt->prediction_type_threshes[i][j];
3714 for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; j++)
3715 rd_opt->filter_threshes_prev[i][j] = rd_opt->filter_threshes[i][j];
3718 if (cpi->tile_data != NULL) {
3719 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3720 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3721 TileDataEnc *tile_data =
3722 &cpi->tile_data[tile_row * tile_cols + tile_col];
3723 for (i = 0; i < BLOCK_SIZES; ++i) {
3724 for (j = 0; j < MAX_MODES; ++j) {
3725 tile_data->thresh_freq_fact_prev[i][j] =
3726 tile_data->thresh_freq_fact[i][j];
3734 static INLINE void set_raw_source_frame(VP9_COMP *cpi) {
3735 #ifdef ENABLE_KF_DENOISE
3736 if (is_spatial_denoise_enabled(cpi)) {
3737 cpi->raw_source_frame = vp9_scale_if_required(
3738 cm, &cpi->raw_unscaled_source, &cpi->raw_scaled_source,
3739 (oxcf->pass == 0), EIGHTTAP, 0);
3741 cpi->raw_source_frame = cpi->Source;
3744 cpi->raw_source_frame = cpi->Source;
3748 static int encode_without_recode_loop(VP9_COMP *cpi, size_t *size,
3750 VP9_COMMON *const cm = &cpi->common;
3751 SVC *const svc = &cpi->svc;
3752 int q = 0, bottom_index = 0, top_index = 0;
3753 int no_drop_scene_change = 0;
3754 const INTERP_FILTER filter_scaler =
3755 (is_one_pass_cbr_svc(cpi))
3756 ? svc->downsample_filter_type[svc->spatial_layer_id]
3758 const int phase_scaler =
3759 (is_one_pass_cbr_svc(cpi))
3760 ? svc->downsample_filter_phase[svc->spatial_layer_id]
3763 if (cm->show_existing_frame) {
3764 cpi->rc.this_frame_target = 0;
3765 if (is_psnr_calc_enabled(cpi)) set_raw_source_frame(cpi);
3769 svc->time_stamp_prev[svc->spatial_layer_id] = svc->time_stamp_superframe;
3771 // Flag to check if its valid to compute the source sad (used for
3772 // scene detection and for superblock content state in CBR mode).
3773 // The flag may get reset below based on SVC or resizing state.
3774 cpi->compute_source_sad_onepass = cpi->oxcf.mode == REALTIME;
3776 vpx_clear_system_state();
3778 set_frame_size(cpi);
3780 if (is_one_pass_cbr_svc(cpi) &&
3781 cpi->un_scaled_source->y_width == cm->width << 2 &&
3782 cpi->un_scaled_source->y_height == cm->height << 2 &&
3783 svc->scaled_temp.y_width == cm->width << 1 &&
3784 svc->scaled_temp.y_height == cm->height << 1) {
3785 // For svc, if it is a 1/4x1/4 downscaling, do a two-stage scaling to take
3786 // advantage of the 1:2 optimized scaler. In the process, the 1/2x1/2
3787 // result will be saved in scaled_temp and might be used later.
3788 const INTERP_FILTER filter_scaler2 = svc->downsample_filter_type[1];
3789 const int phase_scaler2 = svc->downsample_filter_phase[1];
3790 cpi->Source = vp9_svc_twostage_scale(
3791 cm, cpi->un_scaled_source, &cpi->scaled_source, &svc->scaled_temp,
3792 filter_scaler, phase_scaler, filter_scaler2, phase_scaler2);
3793 svc->scaled_one_half = 1;
3794 } else if (is_one_pass_cbr_svc(cpi) &&
3795 cpi->un_scaled_source->y_width == cm->width << 1 &&
3796 cpi->un_scaled_source->y_height == cm->height << 1 &&
3797 svc->scaled_one_half) {
3798 // If the spatial layer is 1/2x1/2 and the scaling is already done in the
3799 // two-stage scaling, use the result directly.
3800 cpi->Source = &svc->scaled_temp;
3801 svc->scaled_one_half = 0;
3803 cpi->Source = vp9_scale_if_required(
3804 cm, cpi->un_scaled_source, &cpi->scaled_source, (cpi->oxcf.pass == 0),
3805 filter_scaler, phase_scaler);
3807 #ifdef OUTPUT_YUV_SVC_SRC
3808 // Write out at most 3 spatial layers.
3809 if (is_one_pass_cbr_svc(cpi) && svc->spatial_layer_id < 3) {
3810 vpx_write_yuv_frame(yuv_svc_src[svc->spatial_layer_id], cpi->Source);
3813 // Unfiltered raw source used in metrics calculation if the source
3814 // has been filtered.
3815 if (is_psnr_calc_enabled(cpi)) {
3816 #ifdef ENABLE_KF_DENOISE
3817 if (is_spatial_denoise_enabled(cpi)) {
3818 cpi->raw_source_frame = vp9_scale_if_required(
3819 cm, &cpi->raw_unscaled_source, &cpi->raw_scaled_source,
3820 (cpi->oxcf.pass == 0), EIGHTTAP, phase_scaler);
3822 cpi->raw_source_frame = cpi->Source;
3825 cpi->raw_source_frame = cpi->Source;
3829 if ((cpi->use_svc &&
3830 (svc->spatial_layer_id < svc->number_spatial_layers - 1 ||
3831 svc->temporal_layer_id < svc->number_temporal_layers - 1 ||
3832 svc->current_superframe < 1)) ||
3833 cpi->resize_pending || cpi->resize_state || cpi->external_resize ||
3834 cpi->resize_state != ORIG) {
3835 cpi->compute_source_sad_onepass = 0;
3836 if (cpi->content_state_sb_fd != NULL)
3837 memset(cpi->content_state_sb_fd, 0,
3838 (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1) *
3839 sizeof(*cpi->content_state_sb_fd));
3842 // Avoid scaling last_source unless its needed.
3843 // Last source is needed if avg_source_sad() is used, or if
3844 // partition_search_type == SOURCE_VAR_BASED_PARTITION, or if noise
3845 // estimation is enabled.
3846 if (cpi->unscaled_last_source != NULL &&
3847 (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3848 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_VBR &&
3849 cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5) ||
3850 cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION ||
3851 (cpi->noise_estimate.enabled && !cpi->oxcf.noise_sensitivity) ||
3852 cpi->compute_source_sad_onepass))
3853 cpi->Last_Source = vp9_scale_if_required(
3854 cm, cpi->unscaled_last_source, &cpi->scaled_last_source,
3855 (cpi->oxcf.pass == 0), EIGHTTAP, 0);
3857 if (cpi->Last_Source == NULL ||
3858 cpi->Last_Source->y_width != cpi->Source->y_width ||
3859 cpi->Last_Source->y_height != cpi->Source->y_height)
3860 cpi->compute_source_sad_onepass = 0;
3862 if (frame_is_intra_only(cm) || cpi->resize_pending != 0) {
3863 memset(cpi->consec_zero_mv, 0,
3864 cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
3867 #if CONFIG_VP9_TEMPORAL_DENOISING
3868 if (cpi->oxcf.noise_sensitivity > 0 && cpi->use_svc)
3869 vp9_denoiser_reset_on_first_frame(cpi);
3872 // Scene detection is always used for VBR mode or screen-content case.
3873 // For other cases (e.g., CBR mode) use it for 5 <= speed < 8 for now
3874 // (need to check encoding time cost for doing this for speed 8).
3875 cpi->rc.high_source_sad = 0;
3876 cpi->rc.hybrid_intra_scene_change = 0;
3877 cpi->rc.re_encode_maxq_scene_change = 0;
3878 if (cm->show_frame && cpi->oxcf.mode == REALTIME &&
3879 (cpi->oxcf.rc_mode == VPX_VBR ||
3880 cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3881 (cpi->oxcf.speed >= 5 && cpi->oxcf.speed < 8)))
3882 vp9_scene_detection_onepass(cpi);
3884 if (svc->spatial_layer_id == svc->first_spatial_layer_to_encode) {
3885 svc->high_source_sad_superframe = cpi->rc.high_source_sad;
3886 svc->high_num_blocks_with_motion = cpi->rc.high_num_blocks_with_motion;
3887 // On scene change reset temporal layer pattern to TL0.
3888 // Note that if the base/lower spatial layers are skipped: instead of
3889 // inserting base layer here, we force max-q for the next superframe
3890 // with lower spatial layers: this is done in vp9_encodedframe_overshoot()
3891 // when max-q is decided for the current layer.
3892 // Only do this reset for bypass/flexible mode.
3893 if (svc->high_source_sad_superframe && svc->temporal_layer_id > 0 &&
3894 svc->temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS) {
3895 // rc->high_source_sad will get reset so copy it to restore it.
3896 int tmp_high_source_sad = cpi->rc.high_source_sad;
3897 vp9_svc_reset_temporal_layers(cpi, cm->frame_type == KEY_FRAME);
3898 cpi->rc.high_source_sad = tmp_high_source_sad;
3902 vp9_update_noise_estimate(cpi);
3904 // For 1 pass CBR, check if we are dropping this frame.
3905 // Never drop on key frame, if base layer is key for svc,
3906 // on scene change, or if superframe has layer sync.
3907 if ((cpi->rc.high_source_sad || svc->high_source_sad_superframe) &&
3908 !(cpi->rc.use_post_encode_drop && svc->last_layer_dropped[0]))
3909 no_drop_scene_change = 1;
3910 if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR &&
3911 !frame_is_intra_only(cm) && !no_drop_scene_change &&
3912 !svc->superframe_has_layer_sync &&
3914 !svc->layer_context[svc->temporal_layer_id].is_key_frame)) {
3915 if (vp9_rc_drop_frame(cpi)) return 0;
3918 // For 1 pass CBR SVC, only ZEROMV is allowed for spatial reference frame
3919 // when svc->force_zero_mode_spatial_ref = 1. Under those conditions we can
3920 // avoid this frame-level upsampling (for non intra_only frames).
3921 if (frame_is_intra_only(cm) == 0 &&
3922 !(is_one_pass_cbr_svc(cpi) && svc->force_zero_mode_spatial_ref)) {
3923 vp9_scale_references(cpi);
3926 set_size_independent_vars(cpi);
3927 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3929 // search method and step parameter might be changed in speed settings.
3930 init_motion_estimation(cpi);
3932 if (cpi->sf.copy_partition_flag) alloc_copy_partition_data(cpi);
3934 if (cpi->sf.svc_use_lowres_part &&
3935 svc->spatial_layer_id == svc->number_spatial_layers - 2) {
3936 if (svc->prev_partition_svc == NULL) {
3938 cm, svc->prev_partition_svc,
3939 (BLOCK_SIZE *)vpx_calloc(cm->mi_stride * cm->mi_rows,
3940 sizeof(*svc->prev_partition_svc)));
3944 // TODO(jianj): Look into issue of skin detection with high bitdepth.
3945 if (cm->bit_depth == 8 && cpi->oxcf.speed >= 5 && cpi->oxcf.pass == 0 &&
3946 cpi->oxcf.rc_mode == VPX_CBR &&
3947 cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
3948 cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3949 cpi->use_skin_detection = 1;
3952 // Enable post encode frame dropping for CBR on non key frame, when
3953 // ext_use_post_encode_drop is specified by user.
3954 cpi->rc.use_post_encode_drop = cpi->rc.ext_use_post_encode_drop &&
3955 cpi->oxcf.rc_mode == VPX_CBR &&
3956 cm->frame_type != KEY_FRAME;
3958 vp9_set_quantizer(cm, q);
3959 vp9_set_variance_partition_thresholds(cpi, q, 0);
3963 suppress_active_map(cpi);
3966 // On non-zero spatial layer, check for disabling inter-layer
3968 if (svc->spatial_layer_id > 0) vp9_svc_constrain_inter_layer_pred(cpi);
3969 vp9_svc_assert_constraints_pattern(cpi);
3972 if (cpi->rc.last_post_encode_dropped_scene_change) {
3973 cpi->rc.high_source_sad = 1;
3974 svc->high_source_sad_superframe = 1;
3975 // For now disable use_source_sad since Last_Source will not be the previous
3976 // encoded but the dropped one.
3977 cpi->sf.use_source_sad = 0;
3978 cpi->rc.last_post_encode_dropped_scene_change = 0;
3980 // Check if this high_source_sad (scene/slide change) frame should be
3981 // encoded at high/max QP, and if so, set the q and adjust some rate
3982 // control parameters.
3983 if (cpi->sf.overshoot_detection_cbr_rt == FAST_DETECTION_MAXQ &&
3984 (cpi->rc.high_source_sad ||
3985 (cpi->use_svc && svc->high_source_sad_superframe))) {
3986 if (vp9_encodedframe_overshoot(cpi, -1, &q)) {
3987 vp9_set_quantizer(cm, q);
3988 vp9_set_variance_partition_thresholds(cpi, q, 0);
3992 #if !CONFIG_REALTIME_ONLY
3993 // Variance adaptive and in frame q adjustment experiments are mutually
3995 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3996 vp9_vaq_frame_setup(cpi);
3997 } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3998 vp9_360aq_frame_setup(cpi);
3999 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
4000 vp9_setup_in_frame_q_adj(cpi);
4001 } else if (cpi->oxcf.aq_mode == LOOKAHEAD_AQ) {
4002 // it may be pretty bad for rate-control,
4003 // and I should handle it somehow
4004 vp9_alt_ref_aq_setup_map(cpi->alt_ref_aq, cpi);
4007 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
4008 vp9_cyclic_refresh_setup(cpi);
4009 } else if (cpi->roi.enabled && !frame_is_intra_only(cm)) {
4012 #if !CONFIG_REALTIME_ONLY
4016 apply_active_map(cpi);
4018 vp9_encode_frame(cpi);
4020 // Check if we should re-encode this frame at high Q because of high
4021 // overshoot based on the encoded frame size. Only for frames where
4022 // high temporal-source SAD is detected.
4023 // For SVC: all spatial layers are checked for re-encoding.
4024 if (cpi->sf.overshoot_detection_cbr_rt == RE_ENCODE_MAXQ &&
4025 (cpi->rc.high_source_sad ||
4026 (cpi->use_svc && svc->high_source_sad_superframe))) {
4028 // Get an estimate of the encoded frame size.
4029 save_coding_context(cpi);
4030 vp9_pack_bitstream(cpi, dest, size);
4031 restore_coding_context(cpi);
4032 frame_size = (int)(*size) << 3;
4033 // Check if encoded frame will overshoot too much, and if so, set the q and
4034 // adjust some rate control parameters, and return to re-encode the frame.
4035 if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) {
4036 vpx_clear_system_state();
4037 vp9_set_quantizer(cm, q);
4038 vp9_set_variance_partition_thresholds(cpi, q, 0);
4039 suppress_active_map(cpi);
4040 // Turn-off cyclic refresh for re-encoded frame.
4041 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
4042 CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
4043 unsigned char *const seg_map = cpi->segmentation_map;
4044 memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
4045 memset(cr->last_coded_q_map, MAXQ,
4046 cm->mi_rows * cm->mi_cols * sizeof(*cr->last_coded_q_map));
4048 vp9_disable_segmentation(&cm->seg);
4050 apply_active_map(cpi);
4051 vp9_encode_frame(cpi);
4055 // Update some stats from cyclic refresh, and check for golden frame update.
4056 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
4057 !frame_is_intra_only(cm))
4058 vp9_cyclic_refresh_postencode(cpi);
4060 // Update the skip mb flag probabilities based on the distribution
4061 // seen in the last encoder iteration.
4062 // update_base_skip_probs(cpi);
4063 vpx_clear_system_state();
4067 #if !CONFIG_REALTIME_ONLY
4068 #define MAX_QSTEP_ADJ 4
4069 static int get_qstep_adj(int rate_excess, int rate_limit) {
4071 rate_limit ? ((rate_excess + rate_limit / 2) / rate_limit) : INT_MAX;
4072 return VPXMIN(qstep, MAX_QSTEP_ADJ);
4075 static void encode_with_recode_loop(VP9_COMP *cpi, size_t *size,
4077 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4078 VP9_COMMON *const cm = &cpi->common;
4079 RATE_CONTROL *const rc = &cpi->rc;
4080 int bottom_index, top_index;
4082 int loop_at_this_size = 0;
4084 int overshoot_seen = 0;
4085 int undershoot_seen = 0;
4086 int frame_over_shoot_limit;
4087 int frame_under_shoot_limit;
4088 int q = 0, q_low = 0, q_high = 0;
4090 #ifdef AGGRESSIVE_VBR
4094 if (cm->show_existing_frame) {
4095 rc->this_frame_target = 0;
4096 if (is_psnr_calc_enabled(cpi)) set_raw_source_frame(cpi);
4100 set_size_independent_vars(cpi);
4102 enable_acl = cpi->sf.allow_acl ? (cm->frame_type == KEY_FRAME) ||
4103 (cpi->twopass.gf_group.index == 1)
4107 vpx_clear_system_state();
4109 set_frame_size(cpi);
4111 if (loop_count == 0 || cpi->resize_pending != 0) {
4112 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
4114 #ifdef AGGRESSIVE_VBR
4115 if (two_pass_first_group_inter(cpi)) {
4116 // Adjustment limits for min and max q
4117 qrange_adj = VPXMAX(1, (top_index - bottom_index) / 2);
4120 VPXMAX(bottom_index - qrange_adj / 2, oxcf->best_allowed_q);
4121 top_index = VPXMIN(oxcf->worst_allowed_q, top_index + qrange_adj / 2);
4124 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
4125 set_mv_search_params(cpi);
4127 // Reset the loop state for new frame size.
4129 undershoot_seen = 0;
4131 // Reconfiguration for change in frame size has concluded.
4132 cpi->resize_pending = 0;
4134 q_low = bottom_index;
4137 loop_at_this_size = 0;
4140 // Decide frame size bounds first time through.
4141 if (loop_count == 0) {
4142 vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
4143 &frame_under_shoot_limit,
4144 &frame_over_shoot_limit);
4148 vp9_scale_if_required(cm, cpi->un_scaled_source, &cpi->scaled_source,
4149 (oxcf->pass == 0), EIGHTTAP, 0);
4151 // Unfiltered raw source used in metrics calculation if the source
4152 // has been filtered.
4153 if (is_psnr_calc_enabled(cpi)) {
4154 #ifdef ENABLE_KF_DENOISE
4155 if (is_spatial_denoise_enabled(cpi)) {
4156 cpi->raw_source_frame = vp9_scale_if_required(
4157 cm, &cpi->raw_unscaled_source, &cpi->raw_scaled_source,
4158 (oxcf->pass == 0), EIGHTTAP, 0);
4160 cpi->raw_source_frame = cpi->Source;
4163 cpi->raw_source_frame = cpi->Source;
4167 if (cpi->unscaled_last_source != NULL)
4168 cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
4169 &cpi->scaled_last_source,
4170 (oxcf->pass == 0), EIGHTTAP, 0);
4172 if (frame_is_intra_only(cm) == 0) {
4173 if (loop_count > 0) {
4174 release_scaled_references(cpi);
4176 vp9_scale_references(cpi);
4179 vp9_set_quantizer(cm, q);
4181 if (loop_count == 0) setup_frame(cpi);
4183 // Variance adaptive and in frame q adjustment experiments are mutually
4185 if (oxcf->aq_mode == VARIANCE_AQ) {
4186 vp9_vaq_frame_setup(cpi);
4187 } else if (oxcf->aq_mode == EQUATOR360_AQ) {
4188 vp9_360aq_frame_setup(cpi);
4189 } else if (oxcf->aq_mode == COMPLEXITY_AQ) {
4190 vp9_setup_in_frame_q_adj(cpi);
4191 } else if (oxcf->aq_mode == LOOKAHEAD_AQ) {
4192 vp9_alt_ref_aq_setup_map(cpi->alt_ref_aq, cpi);
4193 } else if (oxcf->aq_mode == PSNR_AQ) {
4194 vp9_psnr_aq_mode_setup(&cm->seg);
4197 vp9_encode_frame(cpi);
4199 // Update the skip mb flag probabilities based on the distribution
4200 // seen in the last encoder iteration.
4201 // update_base_skip_probs(cpi);
4203 vpx_clear_system_state();
4205 // Dummy pack of the bitstream using up to date stats to get an
4206 // accurate estimate of output frame size to determine if we need
4208 if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
4209 save_coding_context(cpi);
4210 if (!cpi->sf.use_nonrd_pick_mode) vp9_pack_bitstream(cpi, dest, size);
4212 rc->projected_frame_size = (int)(*size) << 3;
4214 if (frame_over_shoot_limit == 0) frame_over_shoot_limit = 1;
4217 if (oxcf->rc_mode == VPX_Q) {
4220 if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced &&
4221 (rc->projected_frame_size < rc->max_frame_bandwidth)) {
4225 int64_t high_err_target = cpi->ambient_err;
4226 int64_t low_err_target = cpi->ambient_err >> 1;
4228 #if CONFIG_VP9_HIGHBITDEPTH
4229 if (cm->use_highbitdepth) {
4230 kf_err = vpx_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4232 kf_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4235 kf_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
4236 #endif // CONFIG_VP9_HIGHBITDEPTH
4238 // Prevent possible divide by zero error below for perfect KF
4241 // The key frame is not good enough or we can afford
4242 // to make it better without undue risk of popping.
4243 if ((kf_err > high_err_target &&
4244 rc->projected_frame_size <= frame_over_shoot_limit) ||
4245 (kf_err > low_err_target &&
4246 rc->projected_frame_size <= frame_under_shoot_limit)) {
4248 q_high = q > q_low ? q - 1 : q_low;
4251 q = (int)((q * high_err_target) / kf_err);
4252 q = VPXMIN(q, (q_high + q_low) >> 1);
4253 } else if (kf_err < low_err_target &&
4254 rc->projected_frame_size >= frame_under_shoot_limit) {
4255 // The key frame is much better than the previous frame
4257 q_low = q < q_high ? q + 1 : q_high;
4260 q = (int)((q * low_err_target) / kf_err);
4261 q = VPXMIN(q, (q_high + q_low + 1) >> 1);
4264 // Clamp Q to upper and lower limits:
4265 q = clamp(q, q_low, q_high);
4268 } else if (recode_loop_test(cpi, frame_over_shoot_limit,
4269 frame_under_shoot_limit, q,
4270 VPXMAX(q_high, top_index), bottom_index)) {
4271 // Is the projected frame size out of range and are we allowed
4272 // to attempt to recode.
4277 if (cpi->resize_pending == 1) {
4278 // Change in frame size so go back around the recode loop.
4279 cpi->rc.frame_size_selector =
4280 SCALE_STEP1 - cpi->rc.frame_size_selector;
4281 cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
4283 #if CONFIG_INTERNAL_STATS
4284 ++cpi->tot_recode_hits;
4291 // Frame size out of permitted range:
4292 // Update correction factor & compute new Q to try...
4294 // Frame is too large
4295 if (rc->projected_frame_size > rc->this_frame_target) {
4296 // Special case if the projected size is > the max allowed.
4297 if ((q == q_high) &&
4298 ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
4299 (!rc->is_src_frame_alt_ref &&
4300 (rc->projected_frame_size >=
4301 big_rate_miss_high_threshold(cpi))))) {
4302 int max_rate = VPXMAX(1, VPXMIN(rc->max_frame_bandwidth,
4303 big_rate_miss_high_threshold(cpi)));
4305 q_val_high = vp9_convert_qindex_to_q(q_high, cm->bit_depth);
4307 q_val_high * ((double)rc->projected_frame_size / max_rate);
4308 q_high = vp9_convert_q_to_qindex(q_val_high, cm->bit_depth);
4309 q_high = clamp(q_high, rc->best_quality, rc->worst_quality);
4312 // Raise Qlow as to at least the current value
4314 get_qstep_adj(rc->projected_frame_size, rc->this_frame_target);
4315 q_low = VPXMIN(q + qstep, q_high);
4317 if (undershoot_seen || loop_at_this_size > 1) {
4318 // Update rate_correction_factor unless
4319 vp9_rc_update_rate_correction_factors(cpi);
4321 q = (q_high + q_low + 1) / 2;
4323 // Update rate_correction_factor unless
4324 vp9_rc_update_rate_correction_factors(cpi);
4326 q = vp9_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
4327 VPXMAX(q_high, top_index));
4329 while (q < q_low && retries < 10) {
4330 vp9_rc_update_rate_correction_factors(cpi);
4331 q = vp9_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
4332 VPXMAX(q_high, top_index));
4339 // Frame is too small
4341 get_qstep_adj(rc->this_frame_target, rc->projected_frame_size);
4342 q_high = VPXMAX(q - qstep, q_low);
4344 if (overshoot_seen || loop_at_this_size > 1) {
4345 vp9_rc_update_rate_correction_factors(cpi);
4346 q = (q_high + q_low) / 2;
4348 vp9_rc_update_rate_correction_factors(cpi);
4349 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
4350 VPXMIN(q_low, bottom_index), top_index);
4351 // Special case reset for qlow for constrained quality.
4352 // This should only trigger where there is very substantial
4353 // undershoot on a frame and the auto cq level is above
4354 // the user passsed in value.
4355 if (oxcf->rc_mode == VPX_CQ && q < q_low) {
4359 while (q > q_high && retries < 10) {
4360 vp9_rc_update_rate_correction_factors(cpi);
4361 q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
4362 VPXMIN(q_low, bottom_index), top_index);
4366 undershoot_seen = 1;
4369 // Clamp Q to upper and lower limits:
4370 q = clamp(q, q_low, q_high);
4372 loop = (q != last_q);
4378 // Special case for overlay frame.
4379 if (rc->is_src_frame_alt_ref &&
4380 rc->projected_frame_size < rc->max_frame_bandwidth)
4385 ++loop_at_this_size;
4387 #if CONFIG_INTERNAL_STATS
4388 ++cpi->tot_recode_hits;
4392 if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF)
4393 if (loop || !enable_acl) restore_coding_context(cpi);
4396 #ifdef AGGRESSIVE_VBR
4397 if (two_pass_first_group_inter(cpi)) {
4398 cpi->twopass.active_worst_quality =
4399 VPXMIN(q + qrange_adj, oxcf->worst_allowed_q);
4400 } else if (!frame_is_kf_gf_arf(cpi)) {
4402 if (!frame_is_kf_gf_arf(cpi)) {
4404 // Have we been forced to adapt Q outside the expected range by an extreme
4405 // rate miss. If so adjust the active maxQ for the subsequent frames.
4406 if (!rc->is_src_frame_alt_ref && (q > cpi->twopass.active_worst_quality)) {
4407 cpi->twopass.active_worst_quality = q;
4408 } else if (oxcf->vbr_corpus_complexity && q == q_low &&
4409 rc->projected_frame_size < rc->this_frame_target) {
4410 cpi->twopass.active_worst_quality =
4411 VPXMAX(q, cpi->twopass.active_worst_quality - 1);
4416 // Skip recoding, if model diff is below threshold
4417 const int thresh = compute_context_model_thresh(cpi);
4418 const int diff = compute_context_model_diff(cm);
4419 if (diff < thresh) {
4420 vpx_clear_system_state();
4421 restore_coding_context(cpi);
4425 vp9_encode_frame(cpi);
4426 vpx_clear_system_state();
4427 restore_coding_context(cpi);
4430 #endif // !CONFIG_REALTIME_ONLY
4432 static int get_ref_frame_flags(const VP9_COMP *cpi) {
4433 const int *const map = cpi->common.ref_frame_map;
4434 const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
4435 const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
4436 const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
4437 int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
4439 if (gold_is_last) flags &= ~VP9_GOLD_FLAG;
4441 if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
4442 (cpi->svc.number_temporal_layers == 1 &&
4443 cpi->svc.number_spatial_layers == 1))
4444 flags &= ~VP9_GOLD_FLAG;
4446 if (alt_is_last) flags &= ~VP9_ALT_FLAG;
4448 if (gold_is_alt) flags &= ~VP9_ALT_FLAG;
4453 static void set_ext_overrides(VP9_COMP *cpi) {
4454 // Overrides the defaults with the externally supplied values with
4455 // vp9_update_reference() and vp9_update_entropy() calls
4456 // Note: The overrides are valid only for the next frame passed
4457 // to encode_frame_to_data_rate() function
4458 if (cpi->ext_refresh_frame_context_pending) {
4459 cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
4460 cpi->ext_refresh_frame_context_pending = 0;
4462 if (cpi->ext_refresh_frame_flags_pending) {
4463 cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
4464 cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
4465 cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
4469 YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(
4470 VP9_COMMON *cm, YV12_BUFFER_CONFIG *unscaled, YV12_BUFFER_CONFIG *scaled,
4471 YV12_BUFFER_CONFIG *scaled_temp, INTERP_FILTER filter_type,
4472 int phase_scaler, INTERP_FILTER filter_type2, int phase_scaler2) {
4473 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
4474 cm->mi_rows * MI_SIZE != unscaled->y_height) {
4475 #if CONFIG_VP9_HIGHBITDEPTH
4476 if (cm->bit_depth == VPX_BITS_8) {
4477 vp9_scale_and_extend_frame(unscaled, scaled_temp, filter_type2,
4479 vp9_scale_and_extend_frame(scaled_temp, scaled, filter_type,
4482 scale_and_extend_frame(unscaled, scaled_temp, (int)cm->bit_depth,
4483 filter_type2, phase_scaler2);
4484 scale_and_extend_frame(scaled_temp, scaled, (int)cm->bit_depth,
4485 filter_type, phase_scaler);
4488 vp9_scale_and_extend_frame(unscaled, scaled_temp, filter_type2,
4490 vp9_scale_and_extend_frame(scaled_temp, scaled, filter_type, phase_scaler);
4491 #endif // CONFIG_VP9_HIGHBITDEPTH
4498 YV12_BUFFER_CONFIG *vp9_scale_if_required(
4499 VP9_COMMON *cm, YV12_BUFFER_CONFIG *unscaled, YV12_BUFFER_CONFIG *scaled,
4500 int use_normative_scaler, INTERP_FILTER filter_type, int phase_scaler) {
4501 if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
4502 cm->mi_rows * MI_SIZE != unscaled->y_height) {
4503 #if CONFIG_VP9_HIGHBITDEPTH
4504 if (use_normative_scaler && unscaled->y_width <= (scaled->y_width << 1) &&
4505 unscaled->y_height <= (scaled->y_height << 1))
4506 if (cm->bit_depth == VPX_BITS_8)
4507 vp9_scale_and_extend_frame(unscaled, scaled, filter_type, phase_scaler);
4509 scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth,
4510 filter_type, phase_scaler);
4512 scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
4514 if (use_normative_scaler && unscaled->y_width <= (scaled->y_width << 1) &&
4515 unscaled->y_height <= (scaled->y_height << 1))
4516 vp9_scale_and_extend_frame(unscaled, scaled, filter_type, phase_scaler);
4518 scale_and_extend_frame_nonnormative(unscaled, scaled);
4519 #endif // CONFIG_VP9_HIGHBITDEPTH
4526 static void set_ref_sign_bias(VP9_COMP *cpi) {
4527 VP9_COMMON *const cm = &cpi->common;
4528 RefCntBuffer *const ref_buffer = get_ref_cnt_buffer(cm, cm->new_fb_idx);
4529 const int cur_frame_index = ref_buffer->frame_index;
4530 MV_REFERENCE_FRAME ref_frame;
4532 for (ref_frame = LAST_FRAME; ref_frame < MAX_REF_FRAMES; ++ref_frame) {
4533 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
4534 const RefCntBuffer *const ref_cnt_buf =
4535 get_ref_cnt_buffer(&cpi->common, buf_idx);
4537 cm->ref_frame_sign_bias[ref_frame] =
4538 cur_frame_index < ref_cnt_buf->frame_index;
4543 static int setup_interp_filter_search_mask(VP9_COMP *cpi) {
4544 INTERP_FILTER ifilter;
4545 int ref_total[MAX_REF_FRAMES] = { 0 };
4546 MV_REFERENCE_FRAME ref;
4548 if (cpi->common.last_frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame)
4550 for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
4551 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter)
4552 ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
4554 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) {
4555 if ((ref_total[LAST_FRAME] &&
4556 cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
4557 (ref_total[GOLDEN_FRAME] == 0 ||
4558 cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50 <
4559 ref_total[GOLDEN_FRAME]) &&
4560 (ref_total[ALTREF_FRAME] == 0 ||
4561 cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50 <
4562 ref_total[ALTREF_FRAME]))
4563 mask |= 1 << ifilter;
4568 #ifdef ENABLE_KF_DENOISE
4569 // Baseline Kernal weights for denoise
4570 static uint8_t dn_kernal_3[9] = { 1, 2, 1, 2, 4, 2, 1, 2, 1 };
4571 static uint8_t dn_kernal_5[25] = { 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 2, 4,
4572 2, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1 };
4574 static INLINE void add_denoise_point(int centre_val, int data_val, int thresh,
4575 uint8_t point_weight, int *sum_val,
4577 if (abs(centre_val - data_val) <= thresh) {
4578 *sum_weight += point_weight;
4579 *sum_val += (int)data_val * (int)point_weight;
4583 static void spatial_denoise_point(uint8_t *src_ptr, const int stride,
4584 const int strength) {
4587 int thresh = strength;
4588 int kernal_size = 5;
4589 int half_k_size = 2;
4593 uint8_t *kernal_ptr;
4595 // Find the maximum deviation from the source point in the locale.
4596 tmp_ptr = src_ptr - (stride * (half_k_size + 1)) - (half_k_size + 1);
4597 for (i = 0; i < kernal_size + 2; ++i) {
4598 for (j = 0; j < kernal_size + 2; ++j) {
4599 max_diff = VPXMAX(max_diff, abs((int)*src_ptr - (int)tmp_ptr[j]));
4604 // Select the kernal size.
4605 if (max_diff > (strength + (strength >> 1))) {
4608 thresh = thresh >> 1;
4610 kernal_ptr = (kernal_size == 3) ? dn_kernal_3 : dn_kernal_5;
4613 tmp_ptr = src_ptr - (stride * half_k_size) - half_k_size;
4614 for (i = 0; i < kernal_size; ++i) {
4615 for (j = 0; j < kernal_size; ++j) {
4616 add_denoise_point((int)*src_ptr, (int)tmp_ptr[j], thresh, *kernal_ptr,
4617 &sum_val, &sum_weight);
4623 // Update the source value with the new filtered value
4624 *src_ptr = (uint8_t)((sum_val + (sum_weight >> 1)) / sum_weight);
4627 #if CONFIG_VP9_HIGHBITDEPTH
4628 static void highbd_spatial_denoise_point(uint16_t *src_ptr, const int stride,
4629 const int strength) {
4632 int thresh = strength;
4633 int kernal_size = 5;
4634 int half_k_size = 2;
4638 uint8_t *kernal_ptr;
4640 // Find the maximum deviation from the source point in the locale.
4641 tmp_ptr = src_ptr - (stride * (half_k_size + 1)) - (half_k_size + 1);
4642 for (i = 0; i < kernal_size + 2; ++i) {
4643 for (j = 0; j < kernal_size + 2; ++j) {
4644 max_diff = VPXMAX(max_diff, abs((int)src_ptr - (int)tmp_ptr[j]));
4649 // Select the kernal size.
4650 if (max_diff > (strength + (strength >> 1))) {
4653 thresh = thresh >> 1;
4655 kernal_ptr = (kernal_size == 3) ? dn_kernal_3 : dn_kernal_5;
4658 tmp_ptr = src_ptr - (stride * half_k_size) - half_k_size;
4659 for (i = 0; i < kernal_size; ++i) {
4660 for (j = 0; j < kernal_size; ++j) {
4661 add_denoise_point((int)*src_ptr, (int)tmp_ptr[j], thresh, *kernal_ptr,
4662 &sum_val, &sum_weight);
4668 // Update the source value with the new filtered value
4669 *src_ptr = (uint16_t)((sum_val + (sum_weight >> 1)) / sum_weight);
4671 #endif // CONFIG_VP9_HIGHBITDEPTH
4673 // Apply thresholded spatial noise supression to a given buffer.
4674 static void spatial_denoise_buffer(VP9_COMP *cpi, uint8_t *buffer,
4675 const int stride, const int width,
4676 const int height, const int strength) {
4677 VP9_COMMON *const cm = &cpi->common;
4678 uint8_t *src_ptr = buffer;
4682 for (row = 0; row < height; ++row) {
4683 for (col = 0; col < width; ++col) {
4684 #if CONFIG_VP9_HIGHBITDEPTH
4685 if (cm->use_highbitdepth)
4686 highbd_spatial_denoise_point(CONVERT_TO_SHORTPTR(&src_ptr[col]), stride,
4689 spatial_denoise_point(&src_ptr[col], stride, strength);
4691 spatial_denoise_point(&src_ptr[col], stride, strength);
4692 #endif // CONFIG_VP9_HIGHBITDEPTH
4698 // Apply thresholded spatial noise supression to source.
4699 static void spatial_denoise_frame(VP9_COMP *cpi) {
4700 YV12_BUFFER_CONFIG *src = cpi->Source;
4701 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4702 TWO_PASS *const twopass = &cpi->twopass;
4703 VP9_COMMON *const cm = &cpi->common;
4705 // Base the filter strength on the current active max Q.
4706 const int q = (int)(vp9_convert_qindex_to_q(twopass->active_worst_quality,
4709 VPXMAX(oxcf->arnr_strength >> 2, VPXMIN(oxcf->arnr_strength, (q >> 4)));
4711 // Denoise each of Y,U and V buffers.
4712 spatial_denoise_buffer(cpi, src->y_buffer, src->y_stride, src->y_width,
4713 src->y_height, strength);
4715 strength += (strength >> 1);
4716 spatial_denoise_buffer(cpi, src->u_buffer, src->uv_stride, src->uv_width,
4717 src->uv_height, strength << 1);
4719 spatial_denoise_buffer(cpi, src->v_buffer, src->uv_stride, src->uv_width,
4720 src->uv_height, strength << 1);
4722 #endif // ENABLE_KF_DENOISE
4724 #if !CONFIG_REALTIME_ONLY
4725 static void vp9_try_disable_lookahead_aq(VP9_COMP *cpi, size_t *size,
4727 if (cpi->common.seg.enabled)
4728 if (ALT_REF_AQ_PROTECT_GAIN) {
4729 size_t nsize = *size;
4732 // TODO(yuryg): optimize this, as
4733 // we don't really need to repack
4735 save_coding_context(cpi);
4736 vp9_disable_segmentation(&cpi->common.seg);
4737 vp9_pack_bitstream(cpi, dest, &nsize);
4738 restore_coding_context(cpi);
4740 overhead = (int)*size - (int)nsize;
4742 if (vp9_alt_ref_aq_disable_if(cpi->alt_ref_aq, overhead, (int)*size))
4743 vp9_encode_frame(cpi);
4745 vp9_enable_segmentation(&cpi->common.seg);
4750 static void set_frame_index(VP9_COMP *cpi, VP9_COMMON *cm) {
4751 RefCntBuffer *const ref_buffer = get_ref_cnt_buffer(cm, cm->new_fb_idx);
4754 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4755 ref_buffer->frame_index =
4756 cm->current_video_frame + gf_group->arf_src_offset[gf_group->index];
4760 // Implementation and modifications of C. Yeo, H. L. Tan, and Y. H. Tan, "On
4761 // rate distortion optimization using SSIM," Circuits and Systems for Video
4762 // Technology, IEEE Transactions on, vol. 23, no. 7, pp. 1170-1181, 2013.
4763 // SSIM_VAR_SCALE defines the strength of the bias towards SSIM in RDO.
4764 // Some sample values are:
4765 // (for midres test set)
4766 // SSIM_VAR_SCALE avg_psnr ssim ms_ssim
4767 // 8.0 9.421 -5.537 -6.898
4768 // 16.0 4.703 -5.378 -6.238
4769 // 32.0 1.929 -4.308 -4.807
4770 #define SSIM_VAR_SCALE 16.0
4771 static void set_mb_ssim_rdmult_scaling(VP9_COMP *cpi) {
4772 VP9_COMMON *cm = &cpi->common;
4773 ThreadData *td = &cpi->td;
4774 MACROBLOCK *x = &td->mb;
4775 MACROBLOCKD *xd = &x->e_mbd;
4776 uint8_t *y_buffer = cpi->Source->y_buffer;
4777 const int y_stride = cpi->Source->y_stride;
4778 const int block_size = BLOCK_16X16;
4780 const int num_8x8_w = num_8x8_blocks_wide_lookup[block_size];
4781 const int num_8x8_h = num_8x8_blocks_high_lookup[block_size];
4782 const int num_cols = (cm->mi_cols + num_8x8_w - 1) / num_8x8_w;
4783 const int num_rows = (cm->mi_rows + num_8x8_h - 1) / num_8x8_h;
4784 double log_sum = 0.0;
4787 const double c2 = 58.5225 * SSIM_VAR_SCALE; // 58.5225 = (.03*255)^2
4789 // Loop through each 64x64 block.
4790 for (row = 0; row < num_rows; ++row) {
4791 for (col = 0; col < num_cols; ++col) {
4793 double var = 0.0, num_of_var = 0.0;
4794 const int index = row * num_cols + col;
4796 for (mi_row = row * num_8x8_h;
4797 mi_row < cm->mi_rows && mi_row < (row + 1) * num_8x8_h; ++mi_row) {
4798 for (mi_col = col * num_8x8_w;
4799 mi_col < cm->mi_cols && mi_col < (col + 1) * num_8x8_w; ++mi_col) {
4801 const int row_offset_y = mi_row << 3;
4802 const int col_offset_y = mi_col << 3;
4804 buf.buf = y_buffer + row_offset_y * y_stride + col_offset_y;
4805 buf.stride = y_stride;
4807 // In order to make SSIM_VAR_SCALE in a same scale for both 8 bit
4808 // and high bit videos, the variance needs to be divided by 2.0 or
4810 // TODO(sdeng): need to tune for 12bit videos.
4811 #if CONFIG_VP9_HIGHBITDEPTH
4812 if (cpi->Source->flags & YV12_FLAG_HIGHBITDEPTH)
4813 var += vp9_high_get_sby_variance(cpi, &buf, BLOCK_8X8, xd->bd);
4816 var += vp9_get_sby_variance(cpi, &buf, BLOCK_8X8);
4821 var = var / num_of_var / 64.0;
4822 var = 2.0 * var + c2;
4823 cpi->mi_ssim_rdmult_scaling_factors[index] = var;
4824 log_sum += log(var);
4827 log_sum = exp(log_sum / (double)(num_rows * num_cols));
4829 for (row = 0; row < num_rows; ++row) {
4830 for (col = 0; col < num_cols; ++col) {
4831 const int index = row * num_cols + col;
4832 cpi->mi_ssim_rdmult_scaling_factors[index] /= log_sum;
4839 // Process the wiener variance in 16x16 block basis.
4840 static int qsort_comp(const void *elem1, const void *elem2) {
4841 int a = *((const int *)elem1);
4842 int b = *((const int *)elem2);
4843 if (a > b) return 1;
4844 if (a < b) return -1;
4848 static void init_mb_wiener_var_buffer(VP9_COMP *cpi) {
4849 VP9_COMMON *cm = &cpi->common;
4851 if (cpi->mb_wiener_variance && cpi->mb_wiener_var_rows >= cm->mb_rows &&
4852 cpi->mb_wiener_var_cols >= cm->mb_cols)
4855 vpx_free(cpi->mb_wiener_variance);
4856 cpi->mb_wiener_variance = NULL;
4859 cm, cpi->mb_wiener_variance,
4860 vpx_calloc(cm->mb_rows * cm->mb_cols, sizeof(*cpi->mb_wiener_variance)));
4861 cpi->mb_wiener_var_rows = cm->mb_rows;
4862 cpi->mb_wiener_var_cols = cm->mb_cols;
4865 static void set_mb_wiener_variance(VP9_COMP *cpi) {
4866 VP9_COMMON *cm = &cpi->common;
4867 uint8_t *buffer = cpi->Source->y_buffer;
4868 int buf_stride = cpi->Source->y_stride;
4870 #if CONFIG_VP9_HIGHBITDEPTH
4871 ThreadData *td = &cpi->td;
4872 MACROBLOCK *x = &td->mb;
4873 MACROBLOCKD *xd = &x->e_mbd;
4874 DECLARE_ALIGNED(16, uint16_t, zero_pred16[32 * 32]);
4875 DECLARE_ALIGNED(16, uint8_t, zero_pred8[32 * 32]);
4878 DECLARE_ALIGNED(16, uint8_t, zero_pred[32 * 32]);
4881 DECLARE_ALIGNED(16, int16_t, src_diff[32 * 32]);
4882 DECLARE_ALIGNED(16, tran_low_t, coeff[32 * 32]);
4884 int mb_row, mb_col, count = 0;
4885 // Hard coded operating block size
4886 const int block_size = 16;
4887 const int coeff_count = block_size * block_size;
4888 const TX_SIZE tx_size = TX_16X16;
4890 #if CONFIG_VP9_HIGHBITDEPTH
4891 xd->cur_buf = cpi->Source;
4892 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4893 zero_pred = CONVERT_TO_BYTEPTR(zero_pred16);
4894 memset(zero_pred16, 0, sizeof(*zero_pred16) * coeff_count);
4896 zero_pred = zero_pred8;
4897 memset(zero_pred8, 0, sizeof(*zero_pred8) * coeff_count);
4900 memset(zero_pred, 0, sizeof(*zero_pred) * coeff_count);
4903 cpi->norm_wiener_variance = 0;
4905 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
4906 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
4908 int16_t median_val = 0;
4909 uint8_t *mb_buffer =
4910 buffer + mb_row * block_size * buf_stride + mb_col * block_size;
4911 int64_t wiener_variance = 0;
4913 #if CONFIG_VP9_HIGHBITDEPTH
4914 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4915 vpx_highbd_subtract_block(block_size, block_size, src_diff, block_size,
4916 mb_buffer, buf_stride, zero_pred, block_size,
4918 highbd_wht_fwd_txfm(src_diff, block_size, coeff, tx_size);
4920 vpx_subtract_block(block_size, block_size, src_diff, block_size,
4921 mb_buffer, buf_stride, zero_pred, block_size);
4922 wht_fwd_txfm(src_diff, block_size, coeff, tx_size);
4925 vpx_subtract_block(block_size, block_size, src_diff, block_size,
4926 mb_buffer, buf_stride, zero_pred, block_size);
4927 wht_fwd_txfm(src_diff, block_size, coeff, tx_size);
4928 #endif // CONFIG_VP9_HIGHBITDEPTH
4931 for (idx = 1; idx < coeff_count; ++idx) coeff[idx] = abs(coeff[idx]);
4933 qsort(coeff, coeff_count - 1, sizeof(*coeff), qsort_comp);
4935 // Noise level estimation
4936 median_val = coeff[coeff_count / 2];
4939 for (idx = 1; idx < coeff_count; ++idx) {
4940 int64_t sqr_coeff = (int64_t)coeff[idx] * coeff[idx];
4941 int64_t tmp_coeff = (int64_t)coeff[idx];
4943 tmp_coeff = (sqr_coeff * coeff[idx]) /
4944 (sqr_coeff + (int64_t)median_val * median_val);
4946 wiener_variance += tmp_coeff * tmp_coeff;
4948 cpi->mb_wiener_variance[mb_row * cm->mb_cols + mb_col] =
4949 wiener_variance / coeff_count;
4950 cpi->norm_wiener_variance +=
4951 cpi->mb_wiener_variance[mb_row * cm->mb_cols + mb_col];
4956 if (count) cpi->norm_wiener_variance /= count;
4957 cpi->norm_wiener_variance = VPXMAX(1, cpi->norm_wiener_variance);
4960 static void encode_frame_to_data_rate(VP9_COMP *cpi, size_t *size,
4962 unsigned int *frame_flags) {
4963 VP9_COMMON *const cm = &cpi->common;
4964 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4965 struct segmentation *const seg = &cm->seg;
4968 // SVC: skip encoding of enhancement layer if the layer target bandwidth = 0.
4969 // If in constrained layer drop mode (svc.framedrop_mode != LAYER_DROP) and
4970 // base spatial layer was dropped, no need to set svc.skip_enhancement_layer,
4971 // as whole superframe will be dropped.
4972 if (cpi->use_svc && cpi->svc.spatial_layer_id > 0 &&
4973 cpi->oxcf.target_bandwidth == 0 &&
4974 !(cpi->svc.framedrop_mode != LAYER_DROP &&
4975 cpi->svc.drop_spatial_layer[0])) {
4976 cpi->svc.skip_enhancement_layer = 1;
4977 vp9_rc_postencode_update_drop_frame(cpi);
4978 cpi->ext_refresh_frame_flags_pending = 0;
4979 cpi->last_frame_dropped = 1;
4980 cpi->svc.last_layer_dropped[cpi->svc.spatial_layer_id] = 1;
4981 cpi->svc.drop_spatial_layer[cpi->svc.spatial_layer_id] = 1;
4982 if (cpi->svc.framedrop_mode == LAYER_DROP ||
4983 cpi->svc.drop_spatial_layer[0] == 0) {
4984 // For the case of constrained drop mode where the base is dropped
4985 // (drop_spatial_layer[0] == 1), which means full superframe dropped,
4986 // we don't increment the svc frame counters. In particular temporal
4987 // layer counter (which is incremented in vp9_inc_frame_in_layer())
4988 // won't be incremented, so on a dropped frame we try the same
4989 // temporal_layer_id on next incoming frame. This is to avoid an
4990 // issue with temporal alignement with full superframe dropping.
4991 vp9_inc_frame_in_layer(cpi);
4996 set_ext_overrides(cpi);
4997 vpx_clear_system_state();
4999 #ifdef ENABLE_KF_DENOISE
5000 // Spatial denoise of key frame.
5001 if (is_spatial_denoise_enabled(cpi)) spatial_denoise_frame(cpi);
5004 if (cm->show_existing_frame == 0) {
5005 // Update frame index
5006 set_frame_index(cpi, cm);
5008 // Set the arf sign bias for this frame.
5009 set_ref_sign_bias(cpi);
5012 // Set default state for segment based loop filter update flags.
5013 cm->lf.mode_ref_delta_update = 0;
5015 if (cpi->oxcf.pass == 2 && cpi->sf.adaptive_interp_filter_search)
5016 cpi->sf.interp_filter_search_mask = setup_interp_filter_search_mask(cpi);
5018 // Set various flags etc to special state if it is a key frame.
5019 if (frame_is_intra_only(cm)) {
5020 // Reset the loop filter deltas and segmentation map.
5021 vp9_reset_segment_features(&cm->seg);
5023 // If segmentation is enabled force a map update for key frames.
5025 seg->update_map = 1;
5026 seg->update_data = 1;
5029 // The alternate reference frame cannot be active for a key frame.
5030 cpi->rc.source_alt_ref_active = 0;
5032 cm->error_resilient_mode = oxcf->error_resilient_mode;
5033 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
5035 // By default, encoder assumes decoder can use prev_mi.
5036 if (cm->error_resilient_mode) {
5037 cm->frame_parallel_decoding_mode = 1;
5038 cm->reset_frame_context = 0;
5039 cm->refresh_frame_context = 0;
5040 } else if (cm->intra_only) {
5041 // Only reset the current context.
5042 cm->reset_frame_context = 2;
5046 if (oxcf->tuning == VP8_TUNE_SSIM) set_mb_ssim_rdmult_scaling(cpi);
5048 if (oxcf->aq_mode == PERCEPTUAL_AQ) {
5049 init_mb_wiener_var_buffer(cpi);
5050 set_mb_wiener_variance(cpi);
5053 vpx_clear_system_state();
5055 #if CONFIG_INTERNAL_STATS
5056 memset(cpi->mode_chosen_counts, 0,
5057 MAX_MODES * sizeof(*cpi->mode_chosen_counts));
5059 #if CONFIG_CONSISTENT_RECODE
5060 // Backup to ensure consistency between recodes
5061 save_encode_params(cpi);
5064 if (cpi->sf.recode_loop == DISALLOW_RECODE) {
5065 if (!encode_without_recode_loop(cpi, size, dest)) return;
5067 #if !CONFIG_REALTIME_ONLY
5068 encode_with_recode_loop(cpi, size, dest);
5072 // TODO(jingning): When using show existing frame mode, we assume that the
5073 // current ARF will be directly used as the final reconstructed frame. This is
5074 // an encoder control scheme. One could in principle explore other
5075 // possibilities to arrange the reference frame buffer and their coding order.
5076 if (cm->show_existing_frame) {
5077 ref_cnt_fb(cm->buffer_pool->frame_bufs, &cm->new_fb_idx,
5078 cm->ref_frame_map[cpi->alt_fb_idx]);
5081 #if !CONFIG_REALTIME_ONLY
5082 // Disable segmentation if it decrease rate/distortion ratio
5083 if (cpi->oxcf.aq_mode == LOOKAHEAD_AQ)
5084 vp9_try_disable_lookahead_aq(cpi, size, dest);
5087 #if CONFIG_VP9_TEMPORAL_DENOISING
5088 #ifdef OUTPUT_YUV_DENOISED
5089 if (oxcf->noise_sensitivity > 0 && denoise_svc(cpi)) {
5090 vpx_write_yuv_frame(yuv_denoised_file,
5091 &cpi->denoiser.running_avg_y[INTRA_FRAME]);
5095 #ifdef OUTPUT_YUV_SKINMAP
5096 if (cpi->common.current_video_frame > 1) {
5097 vp9_output_skin_map(cpi, yuv_skinmap_file);
5101 // Special case code to reduce pulsing when key frames are forced at a
5102 // fixed interval. Note the reconstruction error if it is the frame before
5103 // the force key frame
5104 if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
5105 #if CONFIG_VP9_HIGHBITDEPTH
5106 if (cm->use_highbitdepth) {
5108 vpx_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
5110 cpi->ambient_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
5113 cpi->ambient_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
5114 #endif // CONFIG_VP9_HIGHBITDEPTH
5117 // If the encoder forced a KEY_FRAME decision
5118 if (cm->frame_type == KEY_FRAME) cpi->refresh_last_frame = 1;
5120 cm->frame_to_show = get_frame_new_buffer(cm);
5121 cm->frame_to_show->color_space = cm->color_space;
5122 cm->frame_to_show->color_range = cm->color_range;
5123 cm->frame_to_show->render_width = cm->render_width;
5124 cm->frame_to_show->render_height = cm->render_height;
5126 // Pick the loop filter level for the frame.
5127 loopfilter_frame(cpi, cm);
5129 if (cpi->rc.use_post_encode_drop) save_coding_context(cpi);
5131 // build the bitstream
5132 vp9_pack_bitstream(cpi, dest, size);
5134 if (cpi->rc.use_post_encode_drop && cm->base_qindex < cpi->rc.worst_quality &&
5135 cpi->svc.spatial_layer_id == 0 && post_encode_drop_cbr(cpi, size)) {
5136 restore_coding_context(cpi);
5140 cpi->last_frame_dropped = 0;
5141 cpi->svc.last_layer_dropped[cpi->svc.spatial_layer_id] = 0;
5142 if (cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)
5143 cpi->svc.num_encoded_top_layer++;
5145 // Keep track of the frame buffer index updated/refreshed for the
5146 // current encoded TL0 superframe.
5147 if (cpi->svc.temporal_layer_id == 0) {
5148 if (cpi->refresh_last_frame)
5149 cpi->svc.fb_idx_upd_tl0[cpi->svc.spatial_layer_id] = cpi->lst_fb_idx;
5150 else if (cpi->refresh_golden_frame)
5151 cpi->svc.fb_idx_upd_tl0[cpi->svc.spatial_layer_id] = cpi->gld_fb_idx;
5152 else if (cpi->refresh_alt_ref_frame)
5153 cpi->svc.fb_idx_upd_tl0[cpi->svc.spatial_layer_id] = cpi->alt_fb_idx;
5156 if (cm->seg.update_map) update_reference_segmentation_map(cpi);
5158 if (frame_is_intra_only(cm) == 0) {
5159 release_scaled_references(cpi);
5161 vp9_update_reference_frames(cpi);
5163 if (!cm->show_existing_frame) {
5164 for (t = TX_4X4; t <= TX_32X32; ++t) {
5165 full_to_model_counts(cpi->td.counts->coef[t],
5166 cpi->td.rd_counts.coef_counts[t]);
5169 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
5170 if (!frame_is_intra_only(cm)) {
5171 vp9_adapt_mode_probs(cm);
5172 vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
5174 vp9_adapt_coef_probs(cm);
5178 cpi->ext_refresh_frame_flags_pending = 0;
5180 if (cpi->refresh_golden_frame == 1)
5181 cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
5183 cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
5185 if (cpi->refresh_alt_ref_frame == 1)
5186 cpi->frame_flags |= FRAMEFLAGS_ALTREF;
5188 cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
5190 cpi->ref_frame_flags = get_ref_frame_flags(cpi);
5192 cm->last_frame_type = cm->frame_type;
5194 vp9_rc_postencode_update(cpi, *size);
5196 *size = VPXMAX(1, *size);
5199 output_frame_level_debug_stats(cpi);
5202 if (cm->frame_type == KEY_FRAME) {
5203 // Tell the caller that the frame was coded as a key frame
5204 *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
5206 *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
5209 // Clear the one shot update flags for segmentation map and mode/ref loop
5211 cm->seg.update_map = 0;
5212 cm->seg.update_data = 0;
5213 cm->lf.mode_ref_delta_update = 0;
5215 // keep track of the last coded dimensions
5216 cm->last_width = cm->width;
5217 cm->last_height = cm->height;
5219 // reset to normal state now that we are done.
5220 if (!cm->show_existing_frame) {
5221 cm->last_show_frame = cm->show_frame;
5222 cm->prev_frame = cm->cur_frame;
5225 if (cm->show_frame) {
5226 vp9_swap_mi_and_prev_mi(cm);
5227 // Don't increment frame counters if this was an altref buffer
5228 // update not a real frame
5229 ++cm->current_video_frame;
5230 if (cpi->use_svc) vp9_inc_frame_in_layer(cpi);
5235 .layer_context[cpi->svc.spatial_layer_id *
5236 cpi->svc.number_temporal_layers +
5237 cpi->svc.temporal_layer_id]
5238 .last_frame_type = cm->frame_type;
5239 // Reset layer_sync back to 0 for next frame.
5240 cpi->svc.spatial_layer_sync[cpi->svc.spatial_layer_id] = 0;
5243 cpi->force_update_segmentation = 0;
5245 #if !CONFIG_REALTIME_ONLY
5246 if (cpi->oxcf.aq_mode == LOOKAHEAD_AQ)
5247 vp9_alt_ref_aq_unset_all(cpi->alt_ref_aq, cpi);
5250 cpi->svc.previous_frame_is_intra_only = cm->intra_only;
5251 cpi->svc.set_intra_only_frame = 0;
5254 static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
5255 unsigned int *frame_flags) {
5256 vp9_rc_get_svc_params(cpi);
5257 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
5260 static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
5261 unsigned int *frame_flags) {
5262 if (cpi->oxcf.rc_mode == VPX_CBR) {
5263 vp9_rc_get_one_pass_cbr_params(cpi);
5265 vp9_rc_get_one_pass_vbr_params(cpi);
5267 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
5270 #if !CONFIG_REALTIME_ONLY
5271 static void Pass2Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
5272 unsigned int *frame_flags) {
5273 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
5274 #if CONFIG_MISMATCH_DEBUG
5275 mismatch_move_frame_idx_w();
5277 encode_frame_to_data_rate(cpi, size, dest, frame_flags);
5279 vp9_twopass_postencode_update(cpi);
5281 #endif // !CONFIG_REALTIME_ONLY
5283 static void init_ref_frame_bufs(VP9_COMMON *cm) {
5285 BufferPool *const pool = cm->buffer_pool;
5286 cm->new_fb_idx = INVALID_IDX;
5287 for (i = 0; i < REF_FRAMES; ++i) {
5288 cm->ref_frame_map[i] = INVALID_IDX;
5290 for (i = 0; i < FRAME_BUFFERS; ++i) {
5291 pool->frame_bufs[i].ref_count = 0;
5295 static void check_initial_width(VP9_COMP *cpi,
5296 #if CONFIG_VP9_HIGHBITDEPTH
5297 int use_highbitdepth,
5299 int subsampling_x, int subsampling_y) {
5300 VP9_COMMON *const cm = &cpi->common;
5302 if (!cpi->initial_width ||
5303 #if CONFIG_VP9_HIGHBITDEPTH
5304 cm->use_highbitdepth != use_highbitdepth ||
5306 cm->subsampling_x != subsampling_x ||
5307 cm->subsampling_y != subsampling_y) {
5308 cm->subsampling_x = subsampling_x;
5309 cm->subsampling_y = subsampling_y;
5310 #if CONFIG_VP9_HIGHBITDEPTH
5311 cm->use_highbitdepth = use_highbitdepth;
5314 alloc_raw_frame_buffers(cpi);
5315 init_ref_frame_bufs(cm);
5316 alloc_util_frame_buffers(cpi);
5318 init_motion_estimation(cpi); // TODO(agrange) This can be removed.
5320 cpi->initial_width = cm->width;
5321 cpi->initial_height = cm->height;
5322 cpi->initial_mbs = cm->MBs;
5326 int vp9_receive_raw_frame(VP9_COMP *cpi, vpx_enc_frame_flags_t frame_flags,
5327 YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
5329 VP9_COMMON *const cm = &cpi->common;
5330 struct vpx_usec_timer timer;
5332 const int subsampling_x = sd->subsampling_x;
5333 const int subsampling_y = sd->subsampling_y;
5334 #if CONFIG_VP9_HIGHBITDEPTH
5335 const int use_highbitdepth = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
5338 #if CONFIG_VP9_HIGHBITDEPTH
5339 check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
5341 check_initial_width(cpi, subsampling_x, subsampling_y);
5342 #endif // CONFIG_VP9_HIGHBITDEPTH
5344 #if CONFIG_VP9_HIGHBITDEPTH
5345 // Disable denoiser for high bitdepth since vp9_denoiser_filter only works for
5347 if (cm->bit_depth > 8) cpi->oxcf.noise_sensitivity = 0;
5350 #if CONFIG_VP9_TEMPORAL_DENOISING
5351 setup_denoiser_buffer(cpi);
5353 vpx_usec_timer_start(&timer);
5355 if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
5356 #if CONFIG_VP9_HIGHBITDEPTH
5358 #endif // CONFIG_VP9_HIGHBITDEPTH
5361 vpx_usec_timer_mark(&timer);
5362 cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
5364 if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
5365 (subsampling_x != 1 || subsampling_y != 1)) {
5366 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
5367 "Non-4:2:0 color format requires profile 1 or 3");
5370 if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
5371 (subsampling_x == 1 && subsampling_y == 1)) {
5372 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
5373 "4:2:0 color format requires profile 0 or 2");
5380 static int frame_is_reference(const VP9_COMP *cpi) {
5381 const VP9_COMMON *cm = &cpi->common;
5383 return cm->frame_type == KEY_FRAME || cpi->refresh_last_frame ||
5384 cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame ||
5385 cm->refresh_frame_context || cm->lf.mode_ref_delta_update ||
5386 cm->seg.update_map || cm->seg.update_data;
5389 static void adjust_frame_rate(VP9_COMP *cpi,
5390 const struct lookahead_entry *source) {
5391 int64_t this_duration;
5394 if (source->ts_start == cpi->first_time_stamp_ever) {
5395 this_duration = source->ts_end - source->ts_start;
5398 int64_t last_duration =
5399 cpi->last_end_time_stamp_seen - cpi->last_time_stamp_seen;
5401 this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
5403 // do a step update if the duration changes by 10%
5405 step = (int)((this_duration - last_duration) * 10 / last_duration);
5408 if (this_duration) {
5410 vp9_new_framerate(cpi, 10000000.0 / this_duration);
5412 // Average this frame's rate into the last second's average
5413 // frame rate. If we haven't seen 1 second yet, then average
5414 // over the whole interval seen.
5415 const double interval = VPXMIN(
5416 (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
5417 double avg_duration = 10000000.0 / cpi->framerate;
5418 avg_duration *= (interval - avg_duration + this_duration);
5419 avg_duration /= interval;
5421 vp9_new_framerate(cpi, 10000000.0 / avg_duration);
5424 cpi->last_time_stamp_seen = source->ts_start;
5425 cpi->last_end_time_stamp_seen = source->ts_end;
5428 // Returns 0 if this is not an alt ref else the offset of the source frame
5429 // used as the arf midpoint.
5430 static int get_arf_src_index(VP9_COMP *cpi) {
5431 RATE_CONTROL *const rc = &cpi->rc;
5432 int arf_src_index = 0;
5433 if (is_altref_enabled(cpi)) {
5434 if (cpi->oxcf.pass == 2) {
5435 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
5436 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
5437 arf_src_index = gf_group->arf_src_offset[gf_group->index];
5439 } else if (rc->source_alt_ref_pending) {
5440 arf_src_index = rc->frames_till_gf_update_due;
5443 return arf_src_index;
5446 static void check_src_altref(VP9_COMP *cpi,
5447 const struct lookahead_entry *source) {
5448 RATE_CONTROL *const rc = &cpi->rc;
5450 if (cpi->oxcf.pass == 2) {
5451 const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
5452 rc->is_src_frame_alt_ref =
5453 (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
5455 rc->is_src_frame_alt_ref =
5456 cpi->alt_ref_source && (source == cpi->alt_ref_source);
5459 if (rc->is_src_frame_alt_ref) {
5460 // Current frame is an ARF overlay frame.
5461 cpi->alt_ref_source = NULL;
5463 // Don't refresh the last buffer for an ARF overlay frame. It will
5464 // become the GF so preserve last as an alternative prediction option.
5465 cpi->refresh_last_frame = 0;
5469 #if CONFIG_INTERNAL_STATS
5470 static void adjust_image_stat(double y, double u, double v, double all,
5475 s->stat[ALL] += all;
5476 s->worst = VPXMIN(s->worst, all);
5478 #endif // CONFIG_INTERNAL_STATS
5480 // Adjust the maximum allowable frame size for the target level.
5481 static void level_rc_framerate(VP9_COMP *cpi, int arf_src_index) {
5482 RATE_CONTROL *const rc = &cpi->rc;
5483 LevelConstraint *const ls = &cpi->level_constraint;
5484 VP9_COMMON *const cm = &cpi->common;
5485 const double max_cpb_size = ls->max_cpb_size;
5486 vpx_clear_system_state();
5487 rc->max_frame_bandwidth = VPXMIN(rc->max_frame_bandwidth, ls->max_frame_size);
5488 if (frame_is_intra_only(cm)) {
5489 rc->max_frame_bandwidth =
5490 VPXMIN(rc->max_frame_bandwidth, (int)(max_cpb_size * 0.5));
5491 } else if (arf_src_index > 0) {
5492 rc->max_frame_bandwidth =
5493 VPXMIN(rc->max_frame_bandwidth, (int)(max_cpb_size * 0.4));
5495 rc->max_frame_bandwidth =
5496 VPXMIN(rc->max_frame_bandwidth, (int)(max_cpb_size * 0.2));
5500 static void update_level_info(VP9_COMP *cpi, size_t *size, int arf_src_index) {
5501 VP9_COMMON *const cm = &cpi->common;
5502 Vp9LevelInfo *const level_info = &cpi->level_info;
5503 Vp9LevelSpec *const level_spec = &level_info->level_spec;
5504 Vp9LevelStats *const level_stats = &level_info->level_stats;
5506 uint64_t luma_samples, dur_end;
5507 const uint32_t luma_pic_size = cm->width * cm->height;
5508 const uint32_t luma_pic_breadth = VPXMAX(cm->width, cm->height);
5509 LevelConstraint *const level_constraint = &cpi->level_constraint;
5510 const int8_t level_index = level_constraint->level_index;
5511 double cpb_data_size;
5513 vpx_clear_system_state();
5515 // update level_stats
5516 level_stats->total_compressed_size += *size;
5517 if (cm->show_frame) {
5518 level_stats->total_uncompressed_size +=
5520 2 * (luma_pic_size >> (cm->subsampling_x + cm->subsampling_y));
5521 level_stats->time_encoded =
5522 (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
5523 (double)TICKS_PER_SEC;
5526 if (arf_src_index > 0) {
5527 if (!level_stats->seen_first_altref) {
5528 level_stats->seen_first_altref = 1;
5529 } else if (level_stats->frames_since_last_altref <
5530 level_spec->min_altref_distance) {
5531 level_spec->min_altref_distance = level_stats->frames_since_last_altref;
5533 level_stats->frames_since_last_altref = 0;
5535 ++level_stats->frames_since_last_altref;
5538 if (level_stats->frame_window_buffer.len < FRAME_WINDOW_SIZE - 1) {
5539 idx = (level_stats->frame_window_buffer.start +
5540 level_stats->frame_window_buffer.len++) %
5543 idx = level_stats->frame_window_buffer.start;
5544 level_stats->frame_window_buffer.start = (idx + 1) % FRAME_WINDOW_SIZE;
5546 level_stats->frame_window_buffer.buf[idx].ts = cpi->last_time_stamp_seen;
5547 level_stats->frame_window_buffer.buf[idx].size = (uint32_t)(*size);
5548 level_stats->frame_window_buffer.buf[idx].luma_samples = luma_pic_size;
5550 if (cm->frame_type == KEY_FRAME) {
5551 level_stats->ref_refresh_map = 0;
5554 level_stats->ref_refresh_map |= vp9_get_refresh_mask(cpi);
5555 // Also need to consider the case where the encoder refers to a buffer
5556 // that has been implicitly refreshed after encoding a keyframe.
5557 if (!cm->intra_only) {
5558 level_stats->ref_refresh_map |= (1 << cpi->lst_fb_idx);
5559 level_stats->ref_refresh_map |= (1 << cpi->gld_fb_idx);
5560 level_stats->ref_refresh_map |= (1 << cpi->alt_fb_idx);
5562 for (i = 0; i < REF_FRAMES; ++i) {
5563 count += (level_stats->ref_refresh_map >> i) & 1;
5565 if (count > level_spec->max_ref_frame_buffers) {
5566 level_spec->max_ref_frame_buffers = count;
5570 // update average_bitrate
5571 level_spec->average_bitrate = (double)level_stats->total_compressed_size /
5572 125.0 / level_stats->time_encoded;
5574 // update max_luma_sample_rate
5576 for (i = 0; i < level_stats->frame_window_buffer.len; ++i) {
5577 idx = (level_stats->frame_window_buffer.start +
5578 level_stats->frame_window_buffer.len - 1 - i) %
5581 dur_end = level_stats->frame_window_buffer.buf[idx].ts;
5583 if (dur_end - level_stats->frame_window_buffer.buf[idx].ts >=
5587 luma_samples += level_stats->frame_window_buffer.buf[idx].luma_samples;
5589 if (luma_samples > level_spec->max_luma_sample_rate) {
5590 level_spec->max_luma_sample_rate = luma_samples;
5593 // update max_cpb_size
5595 for (i = 0; i < CPB_WINDOW_SIZE; ++i) {
5596 if (i >= level_stats->frame_window_buffer.len) break;
5597 idx = (level_stats->frame_window_buffer.start +
5598 level_stats->frame_window_buffer.len - 1 - i) %
5600 cpb_data_size += level_stats->frame_window_buffer.buf[idx].size;
5602 cpb_data_size = cpb_data_size / 125.0;
5603 if (cpb_data_size > level_spec->max_cpb_size) {
5604 level_spec->max_cpb_size = cpb_data_size;
5607 // update max_luma_picture_size
5608 if (luma_pic_size > level_spec->max_luma_picture_size) {
5609 level_spec->max_luma_picture_size = luma_pic_size;
5612 // update max_luma_picture_breadth
5613 if (luma_pic_breadth > level_spec->max_luma_picture_breadth) {
5614 level_spec->max_luma_picture_breadth = luma_pic_breadth;
5617 // update compression_ratio
5618 level_spec->compression_ratio = (double)level_stats->total_uncompressed_size *
5620 level_stats->total_compressed_size / 8.0;
5622 // update max_col_tiles
5623 if (level_spec->max_col_tiles < (1 << cm->log2_tile_cols)) {
5624 level_spec->max_col_tiles = (1 << cm->log2_tile_cols);
5627 if (level_index >= 0 && level_constraint->fail_flag == 0) {
5628 if (level_spec->max_luma_picture_size >
5629 vp9_level_defs[level_index].max_luma_picture_size) {
5630 level_constraint->fail_flag |= (1 << LUMA_PIC_SIZE_TOO_LARGE);
5631 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
5632 "Failed to encode to the target level %d. %s",
5633 vp9_level_defs[level_index].level,
5634 level_fail_messages[LUMA_PIC_SIZE_TOO_LARGE]);
5637 if (level_spec->max_luma_picture_breadth >
5638 vp9_level_defs[level_index].max_luma_picture_breadth) {
5639 level_constraint->fail_flag |= (1 << LUMA_PIC_BREADTH_TOO_LARGE);
5640 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
5641 "Failed to encode to the target level %d. %s",
5642 vp9_level_defs[level_index].level,
5643 level_fail_messages[LUMA_PIC_BREADTH_TOO_LARGE]);
5646 if ((double)level_spec->max_luma_sample_rate >
5647 (double)vp9_level_defs[level_index].max_luma_sample_rate *
5648 (1 + SAMPLE_RATE_GRACE_P)) {
5649 level_constraint->fail_flag |= (1 << LUMA_SAMPLE_RATE_TOO_LARGE);
5650 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
5651 "Failed to encode to the target level %d. %s",
5652 vp9_level_defs[level_index].level,
5653 level_fail_messages[LUMA_SAMPLE_RATE_TOO_LARGE]);
5656 if (level_spec->max_col_tiles > vp9_level_defs[level_index].max_col_tiles) {
5657 level_constraint->fail_flag |= (1 << TOO_MANY_COLUMN_TILE);
5658 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
5659 "Failed to encode to the target level %d. %s",
5660 vp9_level_defs[level_index].level,
5661 level_fail_messages[TOO_MANY_COLUMN_TILE]);
5664 if (level_spec->min_altref_distance <
5665 vp9_level_defs[level_index].min_altref_distance) {
5666 level_constraint->fail_flag |= (1 << ALTREF_DIST_TOO_SMALL);
5667 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
5668 "Failed to encode to the target level %d. %s",
5669 vp9_level_defs[level_index].level,
5670 level_fail_messages[ALTREF_DIST_TOO_SMALL]);
5673 if (level_spec->max_ref_frame_buffers >
5674 vp9_level_defs[level_index].max_ref_frame_buffers) {
5675 level_constraint->fail_flag |= (1 << TOO_MANY_REF_BUFFER);
5676 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
5677 "Failed to encode to the target level %d. %s",
5678 vp9_level_defs[level_index].level,
5679 level_fail_messages[TOO_MANY_REF_BUFFER]);
5682 if (level_spec->max_cpb_size > vp9_level_defs[level_index].max_cpb_size) {
5683 level_constraint->fail_flag |= (1 << CPB_TOO_LARGE);
5684 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
5685 "Failed to encode to the target level %d. %s",
5686 vp9_level_defs[level_index].level,
5687 level_fail_messages[CPB_TOO_LARGE]);
5690 // Set an upper bound for the next frame size. It will be used in
5691 // level_rc_framerate() before encoding the next frame.
5693 for (i = 0; i < CPB_WINDOW_SIZE - 1; ++i) {
5694 if (i >= level_stats->frame_window_buffer.len) break;
5695 idx = (level_stats->frame_window_buffer.start +
5696 level_stats->frame_window_buffer.len - 1 - i) %
5698 cpb_data_size += level_stats->frame_window_buffer.buf[idx].size;
5700 cpb_data_size = cpb_data_size / 125.0;
5701 level_constraint->max_frame_size =
5702 (int)((vp9_level_defs[level_index].max_cpb_size - cpb_data_size) *
5704 if (level_stats->frame_window_buffer.len < CPB_WINDOW_SIZE - 1)
5705 level_constraint->max_frame_size >>= 1;
5709 typedef struct GF_PICTURE {
5710 YV12_BUFFER_CONFIG *frame;
5712 FRAME_UPDATE_TYPE update_type;
5715 static void init_gop_frames(VP9_COMP *cpi, GF_PICTURE *gf_picture,
5716 const GF_GROUP *gf_group, int *tpl_group_frames) {
5717 VP9_COMMON *cm = &cpi->common;
5723 int arf_index_stack[MAX_ARF_LAYERS];
5724 int arf_stack_size = 0;
5725 int extend_frame_count = 0;
5726 int pframe_qindex = cpi->tpl_stats[2].base_qindex;
5727 int frame_gop_offset = 0;
5729 RefCntBuffer *frame_bufs = cm->buffer_pool->frame_bufs;
5730 int8_t recon_frame_index[REFS_PER_FRAME + MAX_ARF_LAYERS];
5732 memset(recon_frame_index, -1, sizeof(recon_frame_index));
5733 stack_init(arf_index_stack, MAX_ARF_LAYERS);
5735 // TODO(jingning): To be used later for gf frame type parsing.
5738 for (i = 0; i < FRAME_BUFFERS; ++i) {
5739 if (frame_bufs[i].ref_count == 0) {
5740 alloc_frame_mvs(cm, i);
5741 if (vpx_realloc_frame_buffer(&frame_bufs[i].buf, cm->width, cm->height,
5742 cm->subsampling_x, cm->subsampling_y,
5743 #if CONFIG_VP9_HIGHBITDEPTH
5744 cm->use_highbitdepth,
5746 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
5748 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
5749 "Failed to allocate frame buffer");
5751 recon_frame_index[frame_idx] = i;
5754 if (frame_idx >= REFS_PER_FRAME + cpi->oxcf.enable_auto_arf) break;
5758 for (i = 0; i < REFS_PER_FRAME + 1; ++i) {
5759 assert(recon_frame_index[i] >= 0);
5760 cpi->tpl_recon_frames[i] = &frame_bufs[recon_frame_index[i]].buf;
5763 *tpl_group_frames = 0;
5765 // Initialize Golden reference frame.
5766 gf_picture[0].frame = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
5767 for (i = 0; i < 3; ++i) gf_picture[0].ref_frame[i] = -1;
5768 gf_picture[0].update_type = gf_group->update_type[0];
5770 ++*tpl_group_frames;
5772 // Initialize base layer ARF frame
5773 gf_picture[1].frame = cpi->Source;
5774 gf_picture[1].ref_frame[0] = gld_index;
5775 gf_picture[1].ref_frame[1] = lst_index;
5776 gf_picture[1].ref_frame[2] = alt_index;
5777 gf_picture[1].update_type = gf_group->update_type[1];
5779 ++*tpl_group_frames;
5781 // Initialize P frames
5782 for (frame_idx = 2; frame_idx < MAX_ARF_GOP_SIZE; ++frame_idx) {
5783 struct lookahead_entry *buf;
5784 frame_gop_offset = gf_group->frame_gop_index[frame_idx];
5785 buf = vp9_lookahead_peek(cpi->lookahead, frame_gop_offset - 1);
5787 if (buf == NULL) break;
5789 gf_picture[frame_idx].frame = &buf->img;
5790 gf_picture[frame_idx].ref_frame[0] = gld_index;
5791 gf_picture[frame_idx].ref_frame[1] = lst_index;
5792 gf_picture[frame_idx].ref_frame[2] = alt_index;
5793 gf_picture[frame_idx].update_type = gf_group->update_type[frame_idx];
5795 switch (gf_group->update_type[frame_idx]) {
5797 stack_push(arf_index_stack, alt_index, arf_stack_size);
5799 alt_index = frame_idx;
5801 case LF_UPDATE: lst_index = frame_idx; break;
5802 case OVERLAY_UPDATE:
5803 gld_index = frame_idx;
5804 alt_index = stack_pop(arf_index_stack, arf_stack_size);
5808 lst_index = alt_index;
5809 alt_index = stack_pop(arf_index_stack, arf_stack_size);
5815 ++*tpl_group_frames;
5817 // The length of group of pictures is baseline_gf_interval, plus the
5818 // beginning golden frame from last GOP, plus the last overlay frame in
5820 if (frame_idx == gf_group->gf_group_size) break;
5827 // Extend two frames outside the current gf group.
5828 for (; frame_idx < MAX_LAG_BUFFERS && extend_frame_count < 2; ++frame_idx) {
5829 struct lookahead_entry *buf =
5830 vp9_lookahead_peek(cpi->lookahead, frame_gop_offset - 1);
5832 if (buf == NULL) break;
5834 cpi->tpl_stats[frame_idx].base_qindex = pframe_qindex;
5836 gf_picture[frame_idx].frame = &buf->img;
5837 gf_picture[frame_idx].ref_frame[0] = gld_index;
5838 gf_picture[frame_idx].ref_frame[1] = lst_index;
5839 gf_picture[frame_idx].ref_frame[2] = alt_index;
5840 gf_picture[frame_idx].update_type = LF_UPDATE;
5841 lst_index = frame_idx;
5842 ++*tpl_group_frames;
5843 ++extend_frame_count;
5848 static void init_tpl_stats(VP9_COMP *cpi) {
5850 for (frame_idx = 0; frame_idx < MAX_ARF_GOP_SIZE; ++frame_idx) {
5851 TplDepFrame *tpl_frame = &cpi->tpl_stats[frame_idx];
5852 memset(tpl_frame->tpl_stats_ptr, 0,
5853 tpl_frame->height * tpl_frame->width *
5854 sizeof(*tpl_frame->tpl_stats_ptr));
5855 tpl_frame->is_valid = 0;
5859 #if CONFIG_NON_GREEDY_MV
5860 static uint32_t motion_compensated_prediction(
5861 VP9_COMP *cpi, ThreadData *td, int frame_idx, uint8_t *cur_frame_buf,
5862 uint8_t *ref_frame_buf, int stride, BLOCK_SIZE bsize, int mi_row,
5863 int mi_col, MV *mv, int rf_idx) {
5864 #else // CONFIG_NON_GREEDY_MV
5865 static uint32_t motion_compensated_prediction(VP9_COMP *cpi, ThreadData *td,
5867 uint8_t *cur_frame_buf,
5868 uint8_t *ref_frame_buf,
5869 int stride, BLOCK_SIZE bsize,
5870 int mi_row, int mi_col, MV *mv) {
5871 #endif // CONFIG_NON_GREEDY_MV
5872 MACROBLOCK *const x = &td->mb;
5873 MACROBLOCKD *const xd = &x->e_mbd;
5874 MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
5875 const SEARCH_METHODS search_method = NSTEP;
5877 int sadpb = x->sadperbit16;
5878 uint32_t bestsme = UINT_MAX;
5879 uint32_t distortion;
5882 const MvLimits tmp_mv_limits = x->mv_limits;
5883 #if CONFIG_NON_GREEDY_MV
5884 // lambda is used to adjust the importance of motion vector consitency.
5885 // TODO(angiebird): Figure out lambda's proper value.
5886 const int lambda = cpi->tpl_stats[frame_idx].lambda;
5887 int_mv nb_full_mvs[NB_MVS_NUM];
5890 MV best_ref_mv1 = { 0, 0 };
5891 MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
5893 best_ref_mv1_full.col = best_ref_mv1.col >> 3;
5894 best_ref_mv1_full.row = best_ref_mv1.row >> 3;
5896 // Setup frame pointers
5897 x->plane[0].src.buf = cur_frame_buf;
5898 x->plane[0].src.stride = stride;
5899 xd->plane[0].pre[0].buf = ref_frame_buf;
5900 xd->plane[0].pre[0].stride = stride;
5902 step_param = mv_sf->reduce_first_step_size;
5903 step_param = VPXMIN(step_param, MAX_MVSEARCH_STEPS - 2);
5905 vp9_set_mv_search_range(&x->mv_limits, &best_ref_mv1);
5907 #if CONFIG_NON_GREEDY_MV
5908 (void)search_method;
5910 vp9_prepare_nb_full_mvs(&cpi->tpl_stats[frame_idx], mi_row, mi_col, rf_idx,
5911 bsize, nb_full_mvs);
5912 vp9_full_pixel_diamond_new(cpi, x, &best_ref_mv1_full, step_param, lambda, 1,
5913 &cpi->fn_ptr[bsize], nb_full_mvs, NB_MVS_NUM, mv);
5918 vp9_full_pixel_search(cpi, x, bsize, &best_ref_mv1_full, step_param,
5919 search_method, sadpb, cond_cost_list(cpi, cost_list),
5920 &best_ref_mv1, mv, 0, 0);
5923 /* restore UMV window */
5924 x->mv_limits = tmp_mv_limits;
5926 // TODO(yunqing): may use higher tap interp filter than 2 taps.
5927 // Ignore mv costing by sending NULL pointer instead of cost array
5928 bestsme = cpi->find_fractional_mv_step(
5929 x, mv, &best_ref_mv1, cpi->common.allow_high_precision_mv, x->errorperbit,
5930 &cpi->fn_ptr[bsize], 0, mv_sf->subpel_search_level,
5931 cond_cost_list(cpi, cost_list), NULL, NULL, &distortion, &sse, NULL, 0, 0,
5937 static int get_overlap_area(int grid_pos_row, int grid_pos_col, int ref_pos_row,
5938 int ref_pos_col, int block, BLOCK_SIZE bsize) {
5939 int width = 0, height = 0;
5940 int bw = 4 << b_width_log2_lookup[bsize];
5941 int bh = 4 << b_height_log2_lookup[bsize];
5945 width = grid_pos_col + bw - ref_pos_col;
5946 height = grid_pos_row + bh - ref_pos_row;
5949 width = ref_pos_col + bw - grid_pos_col;
5950 height = grid_pos_row + bh - ref_pos_row;
5953 width = grid_pos_col + bw - ref_pos_col;
5954 height = ref_pos_row + bh - grid_pos_row;
5957 width = ref_pos_col + bw - grid_pos_col;
5958 height = ref_pos_row + bh - grid_pos_row;
5963 return width * height;
5966 static int round_floor(int ref_pos, int bsize_pix) {
5969 round = -(1 + (-ref_pos - 1) / bsize_pix);
5971 round = ref_pos / bsize_pix;
5976 static void tpl_model_store(TplDepStats *tpl_stats, int mi_row, int mi_col,
5977 BLOCK_SIZE bsize, int stride) {
5978 const int mi_height = num_8x8_blocks_high_lookup[bsize];
5979 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
5980 const TplDepStats *src_stats = &tpl_stats[mi_row * stride + mi_col];
5983 for (idy = 0; idy < mi_height; ++idy) {
5984 for (idx = 0; idx < mi_width; ++idx) {
5985 TplDepStats *tpl_ptr = &tpl_stats[(mi_row + idy) * stride + mi_col + idx];
5986 const int64_t mc_flow = tpl_ptr->mc_flow;
5987 const int64_t mc_ref_cost = tpl_ptr->mc_ref_cost;
5988 *tpl_ptr = *src_stats;
5989 tpl_ptr->mc_flow = mc_flow;
5990 tpl_ptr->mc_ref_cost = mc_ref_cost;
5991 tpl_ptr->mc_dep_cost = tpl_ptr->intra_cost + tpl_ptr->mc_flow;
5996 static void tpl_model_update_b(TplDepFrame *tpl_frame, TplDepStats *tpl_stats,
5997 int mi_row, int mi_col, const BLOCK_SIZE bsize) {
5998 TplDepFrame *ref_tpl_frame = &tpl_frame[tpl_stats->ref_frame_index];
5999 TplDepStats *ref_stats = ref_tpl_frame->tpl_stats_ptr;
6000 MV mv = tpl_stats->mv.as_mv;
6001 int mv_row = mv.row >> 3;
6002 int mv_col = mv.col >> 3;
6004 int ref_pos_row = mi_row * MI_SIZE + mv_row;
6005 int ref_pos_col = mi_col * MI_SIZE + mv_col;
6007 const int bw = 4 << b_width_log2_lookup[bsize];
6008 const int bh = 4 << b_height_log2_lookup[bsize];
6009 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6010 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6011 const int pix_num = bw * bh;
6013 // top-left on grid block location in pixel
6014 int grid_pos_row_base = round_floor(ref_pos_row, bh) * bh;
6015 int grid_pos_col_base = round_floor(ref_pos_col, bw) * bw;
6018 for (block = 0; block < 4; ++block) {
6019 int grid_pos_row = grid_pos_row_base + bh * (block >> 1);
6020 int grid_pos_col = grid_pos_col_base + bw * (block & 0x01);
6022 if (grid_pos_row >= 0 && grid_pos_row < ref_tpl_frame->mi_rows * MI_SIZE &&
6023 grid_pos_col >= 0 && grid_pos_col < ref_tpl_frame->mi_cols * MI_SIZE) {
6024 int overlap_area = get_overlap_area(
6025 grid_pos_row, grid_pos_col, ref_pos_row, ref_pos_col, block, bsize);
6026 int ref_mi_row = round_floor(grid_pos_row, bh) * mi_height;
6027 int ref_mi_col = round_floor(grid_pos_col, bw) * mi_width;
6029 int64_t mc_flow = tpl_stats->mc_dep_cost -
6030 (tpl_stats->mc_dep_cost * tpl_stats->inter_cost) /
6031 tpl_stats->intra_cost;
6035 for (idy = 0; idy < mi_height; ++idy) {
6036 for (idx = 0; idx < mi_width; ++idx) {
6037 TplDepStats *des_stats =
6038 &ref_stats[(ref_mi_row + idy) * ref_tpl_frame->stride +
6039 (ref_mi_col + idx)];
6041 des_stats->mc_flow += (mc_flow * overlap_area) / pix_num;
6042 des_stats->mc_ref_cost +=
6043 ((tpl_stats->intra_cost - tpl_stats->inter_cost) * overlap_area) /
6045 assert(overlap_area >= 0);
6052 static void tpl_model_update(TplDepFrame *tpl_frame, TplDepStats *tpl_stats,
6053 int mi_row, int mi_col, const BLOCK_SIZE bsize) {
6055 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6056 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6058 for (idy = 0; idy < mi_height; ++idy) {
6059 for (idx = 0; idx < mi_width; ++idx) {
6060 TplDepStats *tpl_ptr =
6061 &tpl_stats[(mi_row + idy) * tpl_frame->stride + (mi_col + idx)];
6062 tpl_model_update_b(tpl_frame, tpl_ptr, mi_row + idy, mi_col + idx,
6068 static void get_quantize_error(MACROBLOCK *x, int plane, tran_low_t *coeff,
6069 tran_low_t *qcoeff, tran_low_t *dqcoeff,
6070 TX_SIZE tx_size, int64_t *recon_error,
6072 MACROBLOCKD *const xd = &x->e_mbd;
6073 const struct macroblock_plane *const p = &x->plane[plane];
6074 const struct macroblockd_plane *const pd = &xd->plane[plane];
6075 const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
6077 int pix_num = 1 << num_pels_log2_lookup[txsize_to_bsize[tx_size]];
6078 const int shift = tx_size == TX_32X32 ? 0 : 2;
6080 #if CONFIG_VP9_HIGHBITDEPTH
6081 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
6082 vp9_highbd_quantize_fp_32x32(coeff, pix_num, x->skip_block, p->round_fp,
6083 p->quant_fp, qcoeff, dqcoeff, pd->dequant,
6084 &eob, scan_order->scan, scan_order->iscan);
6086 vp9_quantize_fp_32x32(coeff, pix_num, x->skip_block, p->round_fp,
6087 p->quant_fp, qcoeff, dqcoeff, pd->dequant, &eob,
6088 scan_order->scan, scan_order->iscan);
6091 vp9_quantize_fp_32x32(coeff, pix_num, x->skip_block, p->round_fp, p->quant_fp,
6092 qcoeff, dqcoeff, pd->dequant, &eob, scan_order->scan,
6094 #endif // CONFIG_VP9_HIGHBITDEPTH
6096 *recon_error = vp9_block_error(coeff, dqcoeff, pix_num, sse) >> shift;
6097 *recon_error = VPXMAX(*recon_error, 1);
6099 *sse = (*sse) >> shift;
6100 *sse = VPXMAX(*sse, 1);
6103 #if CONFIG_VP9_HIGHBITDEPTH
6104 void highbd_wht_fwd_txfm(int16_t *src_diff, int bw, tran_low_t *coeff,
6106 // TODO(sdeng): Implement SIMD based high bit-depth Hadamard transforms.
6108 case TX_8X8: vpx_highbd_hadamard_8x8(src_diff, bw, coeff); break;
6109 case TX_16X16: vpx_highbd_hadamard_16x16(src_diff, bw, coeff); break;
6110 case TX_32X32: vpx_highbd_hadamard_32x32(src_diff, bw, coeff); break;
6114 #endif // CONFIG_VP9_HIGHBITDEPTH
6116 void wht_fwd_txfm(int16_t *src_diff, int bw, tran_low_t *coeff,
6119 case TX_8X8: vpx_hadamard_8x8(src_diff, bw, coeff); break;
6120 case TX_16X16: vpx_hadamard_16x16(src_diff, bw, coeff); break;
6121 case TX_32X32: vpx_hadamard_32x32(src_diff, bw, coeff); break;
6126 static void set_mv_limits(const VP9_COMMON *cm, MACROBLOCK *x, int mi_row,
6128 x->mv_limits.row_min = -((mi_row * MI_SIZE) + (17 - 2 * VP9_INTERP_EXTEND));
6129 x->mv_limits.row_max =
6130 (cm->mi_rows - 1 - mi_row) * MI_SIZE + (17 - 2 * VP9_INTERP_EXTEND);
6131 x->mv_limits.col_min = -((mi_col * MI_SIZE) + (17 - 2 * VP9_INTERP_EXTEND));
6132 x->mv_limits.col_max =
6133 ((cm->mi_cols - 1 - mi_col) * MI_SIZE) + (17 - 2 * VP9_INTERP_EXTEND);
6136 static void mode_estimation(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
6137 struct scale_factors *sf, GF_PICTURE *gf_picture,
6138 int frame_idx, TplDepFrame *tpl_frame,
6139 int16_t *src_diff, tran_low_t *coeff,
6140 tran_low_t *qcoeff, tran_low_t *dqcoeff, int mi_row,
6141 int mi_col, BLOCK_SIZE bsize, TX_SIZE tx_size,
6142 YV12_BUFFER_CONFIG *ref_frame[], uint8_t *predictor,
6143 int64_t *recon_error, int64_t *sse) {
6144 VP9_COMMON *cm = &cpi->common;
6145 ThreadData *td = &cpi->td;
6147 const int bw = 4 << b_width_log2_lookup[bsize];
6148 const int bh = 4 << b_height_log2_lookup[bsize];
6149 const int pix_num = bw * bh;
6150 int best_rf_idx = -1;
6152 int64_t best_inter_cost = INT64_MAX;
6155 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
6157 int64_t best_intra_cost = INT64_MAX;
6159 PREDICTION_MODE mode;
6160 int mb_y_offset = mi_row * MI_SIZE * xd->cur_buf->y_stride + mi_col * MI_SIZE;
6161 MODE_INFO mi_above, mi_left;
6162 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6163 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6164 TplDepStats *tpl_stats =
6165 &tpl_frame->tpl_stats_ptr[mi_row * tpl_frame->stride + mi_col];
6167 xd->mb_to_top_edge = -((mi_row * MI_SIZE) * 8);
6168 xd->mb_to_bottom_edge = ((cm->mi_rows - 1 - mi_row) * MI_SIZE) * 8;
6169 xd->mb_to_left_edge = -((mi_col * MI_SIZE) * 8);
6170 xd->mb_to_right_edge = ((cm->mi_cols - 1 - mi_col) * MI_SIZE) * 8;
6171 xd->above_mi = (mi_row > 0) ? &mi_above : NULL;
6172 xd->left_mi = (mi_col > 0) ? &mi_left : NULL;
6174 // Intra prediction search
6175 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
6177 int src_stride, dst_stride;
6179 src = xd->cur_buf->y_buffer + mb_y_offset;
6180 src_stride = xd->cur_buf->y_stride;
6182 dst = &predictor[0];
6185 xd->mi[0]->sb_type = bsize;
6186 xd->mi[0]->ref_frame[0] = INTRA_FRAME;
6188 vp9_predict_intra_block(xd, b_width_log2_lookup[bsize], tx_size, mode, src,
6189 src_stride, dst, dst_stride, 0, 0, 0);
6191 #if CONFIG_VP9_HIGHBITDEPTH
6192 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
6193 vpx_highbd_subtract_block(bh, bw, src_diff, bw, src, src_stride, dst,
6194 dst_stride, xd->bd);
6195 highbd_wht_fwd_txfm(src_diff, bw, coeff, tx_size);
6196 intra_cost = vpx_highbd_satd(coeff, pix_num);
6198 vpx_subtract_block(bh, bw, src_diff, bw, src, src_stride, dst,
6200 wht_fwd_txfm(src_diff, bw, coeff, tx_size);
6201 intra_cost = vpx_satd(coeff, pix_num);
6204 vpx_subtract_block(bh, bw, src_diff, bw, src, src_stride, dst, dst_stride);
6205 wht_fwd_txfm(src_diff, bw, coeff, tx_size);
6206 intra_cost = vpx_satd(coeff, pix_num);
6207 #endif // CONFIG_VP9_HIGHBITDEPTH
6209 if (intra_cost < best_intra_cost) best_intra_cost = intra_cost;
6212 // Motion compensated prediction
6215 set_mv_limits(cm, x, mi_row, mi_col);
6217 for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
6219 if (ref_frame[rf_idx] == NULL) continue;
6221 #if CONFIG_NON_GREEDY_MV
6224 get_pyramid_mv(tpl_frame, rf_idx, bsize, mi_row, mi_col)->as_int;
6226 motion_compensated_prediction(
6227 cpi, td, frame_idx, xd->cur_buf->y_buffer + mb_y_offset,
6228 ref_frame[rf_idx]->y_buffer + mb_y_offset, xd->cur_buf->y_stride, bsize,
6229 mi_row, mi_col, &mv.as_mv);
6232 #if CONFIG_VP9_HIGHBITDEPTH
6233 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
6234 vp9_highbd_build_inter_predictor(
6235 CONVERT_TO_SHORTPTR(ref_frame[rf_idx]->y_buffer + mb_y_offset),
6236 ref_frame[rf_idx]->y_stride, CONVERT_TO_SHORTPTR(&predictor[0]), bw,
6237 &mv.as_mv, sf, bw, bh, 0, kernel, MV_PRECISION_Q3, mi_col * MI_SIZE,
6238 mi_row * MI_SIZE, xd->bd);
6239 vpx_highbd_subtract_block(
6240 bh, bw, src_diff, bw, xd->cur_buf->y_buffer + mb_y_offset,
6241 xd->cur_buf->y_stride, &predictor[0], bw, xd->bd);
6242 highbd_wht_fwd_txfm(src_diff, bw, coeff, tx_size);
6243 inter_cost = vpx_highbd_satd(coeff, pix_num);
6245 vp9_build_inter_predictor(
6246 ref_frame[rf_idx]->y_buffer + mb_y_offset,
6247 ref_frame[rf_idx]->y_stride, &predictor[0], bw, &mv.as_mv, sf, bw, bh,
6248 0, kernel, MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE);
6249 vpx_subtract_block(bh, bw, src_diff, bw,
6250 xd->cur_buf->y_buffer + mb_y_offset,
6251 xd->cur_buf->y_stride, &predictor[0], bw);
6252 wht_fwd_txfm(src_diff, bw, coeff, tx_size);
6253 inter_cost = vpx_satd(coeff, pix_num);
6256 vp9_build_inter_predictor(ref_frame[rf_idx]->y_buffer + mb_y_offset,
6257 ref_frame[rf_idx]->y_stride, &predictor[0], bw,
6258 &mv.as_mv, sf, bw, bh, 0, kernel, MV_PRECISION_Q3,
6259 mi_col * MI_SIZE, mi_row * MI_SIZE);
6260 vpx_subtract_block(bh, bw, src_diff, bw,
6261 xd->cur_buf->y_buffer + mb_y_offset,
6262 xd->cur_buf->y_stride, &predictor[0], bw);
6263 wht_fwd_txfm(src_diff, bw, coeff, tx_size);
6264 inter_cost = vpx_satd(coeff, pix_num);
6267 if (inter_cost < best_inter_cost) {
6268 best_rf_idx = rf_idx;
6269 best_inter_cost = inter_cost;
6270 best_mv.as_int = mv.as_int;
6271 get_quantize_error(x, 0, coeff, qcoeff, dqcoeff, tx_size, recon_error,
6275 best_intra_cost = VPXMAX(best_intra_cost, 1);
6276 best_inter_cost = VPXMIN(best_intra_cost, best_inter_cost);
6277 tpl_stats->inter_cost = VPXMAX(
6278 1, (best_inter_cost << TPL_DEP_COST_SCALE_LOG2) / (mi_height * mi_width));
6279 tpl_stats->intra_cost = VPXMAX(
6280 1, (best_intra_cost << TPL_DEP_COST_SCALE_LOG2) / (mi_height * mi_width));
6281 tpl_stats->ref_frame_index = gf_picture[frame_idx].ref_frame[best_rf_idx];
6282 tpl_stats->mv.as_int = best_mv.as_int;
6285 #if CONFIG_NON_GREEDY_MV
6286 static int get_block_src_pred_buf(MACROBLOCKD *xd, GF_PICTURE *gf_picture,
6287 int frame_idx, int rf_idx, int mi_row,
6288 int mi_col, struct buf_2d *src,
6289 struct buf_2d *pre) {
6290 const int mb_y_offset =
6291 mi_row * MI_SIZE * xd->cur_buf->y_stride + mi_col * MI_SIZE;
6292 YV12_BUFFER_CONFIG *ref_frame = NULL;
6293 int ref_frame_idx = gf_picture[frame_idx].ref_frame[rf_idx];
6294 if (ref_frame_idx != -1) {
6295 ref_frame = gf_picture[ref_frame_idx].frame;
6296 src->buf = xd->cur_buf->y_buffer + mb_y_offset;
6297 src->stride = xd->cur_buf->y_stride;
6298 pre->buf = ref_frame->y_buffer + mb_y_offset;
6299 pre->stride = ref_frame->y_stride;
6300 assert(src->stride == pre->stride);
6303 printf("invalid ref_frame_idx");
6304 assert(ref_frame_idx != -1);
6309 #define kMvPreCheckLines 5
6310 #define kMvPreCheckSize 15
6312 #define MV_REF_POS_NUM 3
6313 POSITION mv_ref_pos[MV_REF_POS_NUM] = {
6319 static int_mv *get_select_mv(VP9_COMP *cpi, TplDepFrame *tpl_frame, int mi_row,
6321 return &cpi->select_mv_arr[mi_row * tpl_frame->stride + mi_col];
6324 static int_mv find_ref_mv(int mv_mode, VP9_COMP *cpi, TplDepFrame *tpl_frame,
6325 BLOCK_SIZE bsize, int mi_row, int mi_col) {
6327 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6328 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6329 int_mv nearest_mv, near_mv, invalid_mv;
6330 nearest_mv.as_int = INVALID_MV;
6331 near_mv.as_int = INVALID_MV;
6332 invalid_mv.as_int = INVALID_MV;
6333 for (i = 0; i < MV_REF_POS_NUM; ++i) {
6334 int nb_row = mi_row + mv_ref_pos[i].row * mi_height;
6335 int nb_col = mi_col + mv_ref_pos[i].col * mi_width;
6336 assert(mv_ref_pos[i].row <= 0);
6337 assert(mv_ref_pos[i].col <= 0);
6338 if (nb_row >= 0 && nb_col >= 0) {
6339 if (nearest_mv.as_int == INVALID_MV) {
6340 nearest_mv = *get_select_mv(cpi, tpl_frame, nb_row, nb_col);
6342 int_mv mv = *get_select_mv(cpi, tpl_frame, nb_row, nb_col);
6343 if (mv.as_int == nearest_mv.as_int) {
6352 if (nearest_mv.as_int == INVALID_MV) {
6353 nearest_mv.as_mv.row = 0;
6354 nearest_mv.as_mv.col = 0;
6356 if (near_mv.as_int == INVALID_MV) {
6357 near_mv.as_mv.row = 0;
6358 near_mv.as_mv.col = 0;
6360 if (mv_mode == NEAREST_MV_MODE) {
6363 if (mv_mode == NEAR_MV_MODE) {
6370 static int_mv get_mv_from_mv_mode(int mv_mode, VP9_COMP *cpi,
6371 TplDepFrame *tpl_frame, int rf_idx,
6372 BLOCK_SIZE bsize, int mi_row, int mi_col) {
6380 mv = *get_pyramid_mv(tpl_frame, rf_idx, bsize, mi_row, mi_col);
6382 case NEAREST_MV_MODE:
6383 mv = find_ref_mv(mv_mode, cpi, tpl_frame, bsize, mi_row, mi_col);
6386 mv = find_ref_mv(mv_mode, cpi, tpl_frame, bsize, mi_row, mi_col);
6389 mv.as_int = INVALID_MV;
6396 static double get_mv_dist(int mv_mode, VP9_COMP *cpi, MACROBLOCKD *xd,
6397 GF_PICTURE *gf_picture, int frame_idx,
6398 TplDepFrame *tpl_frame, int rf_idx, BLOCK_SIZE bsize,
6399 int mi_row, int mi_col, int_mv *mv) {
6404 *mv = get_mv_from_mv_mode(mv_mode, cpi, tpl_frame, rf_idx, bsize, mi_row,
6406 full_mv = get_full_mv(&mv->as_mv);
6407 if (get_block_src_pred_buf(xd, gf_picture, frame_idx, rf_idx, mi_row, mi_col,
6409 // TODO(angiebird): Consider subpixel when computing the sse.
6410 cpi->fn_ptr[bsize].vf(src.buf, src.stride, get_buf_from_mv(&pre, &full_mv),
6412 return (double)(sse << VP9_DIST_SCALE_LOG2);
6419 static int get_mv_mode_cost(int mv_mode) {
6420 // TODO(angiebird): The probabilities are roughly inferred from
6421 // default_inter_mode_probs. Check if there is a better way to set the
6423 const int zero_mv_prob = 16;
6424 const int new_mv_prob = 24 * 1;
6425 const int ref_mv_prob = 256 - zero_mv_prob - new_mv_prob;
6426 assert(zero_mv_prob + new_mv_prob + ref_mv_prob == 256);
6428 case ZERO_MV_MODE: return vp9_prob_cost[zero_mv_prob]; break;
6429 case NEW_MV_MODE: return vp9_prob_cost[new_mv_prob]; break;
6430 case NEAREST_MV_MODE: return vp9_prob_cost[ref_mv_prob]; break;
6431 case NEAR_MV_MODE: return vp9_prob_cost[ref_mv_prob]; break;
6432 default: assert(0); return -1;
6436 static INLINE double get_mv_diff_cost(MV *new_mv, MV *ref_mv) {
6437 double mv_diff_cost = log2(1 + abs(new_mv->row - ref_mv->row)) +
6438 log2(1 + abs(new_mv->col - ref_mv->col));
6439 mv_diff_cost *= (1 << VP9_PROB_COST_SHIFT);
6440 return mv_diff_cost;
6442 static double get_mv_cost(int mv_mode, VP9_COMP *cpi, TplDepFrame *tpl_frame,
6443 int rf_idx, BLOCK_SIZE bsize, int mi_row,
6445 double mv_cost = get_mv_mode_cost(mv_mode);
6446 if (mv_mode == NEW_MV_MODE) {
6447 MV new_mv = get_mv_from_mv_mode(mv_mode, cpi, tpl_frame, rf_idx, bsize,
6450 MV nearest_mv = get_mv_from_mv_mode(NEAREST_MV_MODE, cpi, tpl_frame, rf_idx,
6451 bsize, mi_row, mi_col)
6453 MV near_mv = get_mv_from_mv_mode(NEAR_MV_MODE, cpi, tpl_frame, rf_idx,
6454 bsize, mi_row, mi_col)
6456 double nearest_cost = get_mv_diff_cost(&new_mv, &nearest_mv);
6457 double near_cost = get_mv_diff_cost(&new_mv, &near_mv);
6458 mv_cost += nearest_cost < near_cost ? nearest_cost : near_cost;
6463 static double eval_mv_mode(int mv_mode, VP9_COMP *cpi, MACROBLOCK *x,
6464 GF_PICTURE *gf_picture, int frame_idx,
6465 TplDepFrame *tpl_frame, int rf_idx, BLOCK_SIZE bsize,
6466 int mi_row, int mi_col, int_mv *mv) {
6467 MACROBLOCKD *xd = &x->e_mbd;
6468 double mv_dist = get_mv_dist(mv_mode, cpi, xd, gf_picture, frame_idx,
6469 tpl_frame, rf_idx, bsize, mi_row, mi_col, mv);
6471 get_mv_cost(mv_mode, cpi, tpl_frame, rf_idx, bsize, mi_row, mi_col);
6474 return mv_cost + mult * log2f(1 + mv_dist);
6477 static int find_best_ref_mv_mode(VP9_COMP *cpi, MACROBLOCK *x,
6478 GF_PICTURE *gf_picture, int frame_idx,
6479 TplDepFrame *tpl_frame, int rf_idx,
6480 BLOCK_SIZE bsize, int mi_row, int mi_col,
6481 double *rd, int_mv *mv) {
6482 int best_mv_mode = ZERO_MV_MODE;
6486 for (mv_mode = 0; mv_mode < MAX_MV_MODE; ++mv_mode) {
6489 if (mv_mode == NEW_MV_MODE) {
6492 this_rd = eval_mv_mode(mv_mode, cpi, x, gf_picture, frame_idx, tpl_frame,
6493 rf_idx, bsize, mi_row, mi_col, &this_mv);
6497 best_mv_mode = mv_mode;
6500 if (this_rd < *rd) {
6503 best_mv_mode = mv_mode;
6507 return best_mv_mode;
6510 static void predict_mv_mode(VP9_COMP *cpi, MACROBLOCK *x,
6511 GF_PICTURE *gf_picture, int frame_idx,
6512 TplDepFrame *tpl_frame, int rf_idx,
6513 BLOCK_SIZE bsize, int mi_row, int mi_col) {
6514 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6515 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6516 int tmp_mv_mode_arr[kMvPreCheckSize];
6517 int *mv_mode_arr = tpl_frame->mv_mode_arr[rf_idx];
6518 double *rd_diff_arr = tpl_frame->rd_diff_arr[rf_idx];
6519 int_mv *select_mv_arr = cpi->select_mv_arr;
6520 int_mv tmp_select_mv_arr[kMvPreCheckSize];
6521 int stride = tpl_frame->stride;
6522 double new_mv_rd = 0;
6523 double no_new_mv_rd = 0;
6524 double this_new_mv_rd = 0;
6525 double this_no_new_mv_rd = 0;
6528 assert(kMvPreCheckSize == (kMvPreCheckLines * (kMvPreCheckLines + 1)) >> 1);
6531 // diagnal scan order
6533 for (idx = 0; idx < kMvPreCheckLines; ++idx) {
6535 for (r = 0; r <= idx; ++r) {
6537 int nb_row = mi_row + r * mi_height;
6538 int nb_col = mi_col + c * mi_width;
6539 if (nb_row < tpl_frame->mi_rows && nb_col < tpl_frame->mi_cols) {
6541 int_mv *mv = &select_mv_arr[nb_row * stride + nb_col];
6542 mv_mode_arr[nb_row * stride + nb_col] =
6543 find_best_ref_mv_mode(cpi, x, gf_picture, frame_idx, tpl_frame,
6544 rf_idx, bsize, nb_row, nb_col, &this_rd, mv);
6545 if (r == 0 && c == 0) {
6546 this_no_new_mv_rd = this_rd;
6548 no_new_mv_rd += this_rd;
6549 tmp_mv_mode_arr[tmp_idx] = mv_mode_arr[nb_row * stride + nb_col];
6550 tmp_select_mv_arr[tmp_idx] = select_mv_arr[nb_row * stride + nb_col];
6557 mv_mode_arr[mi_row * stride + mi_col] = NEW_MV_MODE;
6558 this_new_mv_rd = eval_mv_mode(NEW_MV_MODE, cpi, x, gf_picture, frame_idx,
6559 tpl_frame, rf_idx, bsize, mi_row, mi_col,
6560 &select_mv_arr[mi_row * stride + mi_col]);
6561 new_mv_rd = this_new_mv_rd;
6562 // We start from idx = 1 because idx = 0 is evaluated as NEW_MV_MODE
6564 for (idx = 1; idx < kMvPreCheckLines; ++idx) {
6566 for (r = 0; r <= idx; ++r) {
6568 int nb_row = mi_row + r * mi_height;
6569 int nb_col = mi_col + c * mi_width;
6570 if (nb_row < tpl_frame->mi_rows && nb_col < tpl_frame->mi_cols) {
6572 int_mv *mv = &select_mv_arr[nb_row * stride + nb_col];
6573 mv_mode_arr[nb_row * stride + nb_col] =
6574 find_best_ref_mv_mode(cpi, x, gf_picture, frame_idx, tpl_frame,
6575 rf_idx, bsize, nb_row, nb_col, &this_rd, mv);
6576 new_mv_rd += this_rd;
6581 // update best_mv_mode
6583 if (no_new_mv_rd < new_mv_rd) {
6584 for (idx = 0; idx < kMvPreCheckLines; ++idx) {
6586 for (r = 0; r <= idx; ++r) {
6588 int nb_row = mi_row + r * mi_height;
6589 int nb_col = mi_col + c * mi_width;
6590 if (nb_row < tpl_frame->mi_rows && nb_col < tpl_frame->mi_cols) {
6591 mv_mode_arr[nb_row * stride + nb_col] = tmp_mv_mode_arr[tmp_idx];
6592 select_mv_arr[nb_row * stride + nb_col] = tmp_select_mv_arr[tmp_idx];
6597 rd_diff_arr[mi_row * stride + mi_col] = 0;
6599 rd_diff_arr[mi_row * stride + mi_col] =
6600 (no_new_mv_rd - this_no_new_mv_rd) - (new_mv_rd - this_new_mv_rd);
6604 static void predict_mv_mode_arr(VP9_COMP *cpi, MACROBLOCK *x,
6605 GF_PICTURE *gf_picture, int frame_idx,
6606 TplDepFrame *tpl_frame, int rf_idx,
6608 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6609 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6610 const int unit_rows = tpl_frame->mi_rows / mi_height;
6611 const int unit_cols = tpl_frame->mi_cols / mi_width;
6612 const int max_diagonal_lines = unit_rows + unit_cols - 1;
6614 for (idx = 0; idx < max_diagonal_lines; ++idx) {
6616 for (r = VPXMAX(idx - unit_cols + 1, 0); r <= VPXMIN(idx, unit_rows - 1);
6619 int mi_row = r * mi_height;
6620 int mi_col = c * mi_width;
6621 assert(c >= 0 && c < unit_cols);
6622 assert(mi_row >= 0 && mi_row < tpl_frame->mi_rows);
6623 assert(mi_col >= 0 && mi_col < tpl_frame->mi_cols);
6624 predict_mv_mode(cpi, x, gf_picture, frame_idx, tpl_frame, rf_idx, bsize,
6630 static double get_feature_score(uint8_t *buf, ptrdiff_t stride, int rows,
6637 vpx_clear_system_state();
6638 for (r = 0; r + 1 < rows; ++r) {
6639 for (c = 0; c + 1 < cols; ++c) {
6640 int diff_x = buf[r * stride + c] - buf[r * stride + c + 1];
6641 int diff_y = buf[r * stride + c] - buf[(r + 1) * stride + c];
6642 IxIx += diff_x * diff_x;
6643 IxIy += diff_x * diff_y;
6644 IyIy += diff_y * diff_y;
6647 IxIx /= (rows - 1) * (cols - 1);
6648 IxIy /= (rows - 1) * (cols - 1);
6649 IyIy /= (rows - 1) * (cols - 1);
6650 score = (IxIx * IyIy - IxIy * IxIy + 0.0001) / (IxIx + IyIy + 0.0001);
6654 static int compare_feature_score(const void *a, const void *b) {
6655 const FEATURE_SCORE_LOC *aa = *(FEATURE_SCORE_LOC *const *)a;
6656 const FEATURE_SCORE_LOC *bb = *(FEATURE_SCORE_LOC *const *)b;
6657 if (aa->feature_score < bb->feature_score) {
6659 } else if (aa->feature_score > bb->feature_score) {
6666 static void do_motion_search(VP9_COMP *cpi, ThreadData *td, int frame_idx,
6667 YV12_BUFFER_CONFIG **ref_frame, BLOCK_SIZE bsize,
6668 int mi_row, int mi_col) {
6669 VP9_COMMON *cm = &cpi->common;
6670 MACROBLOCK *x = &td->mb;
6671 MACROBLOCKD *xd = &x->e_mbd;
6672 TplDepFrame *tpl_frame = &cpi->tpl_stats[frame_idx];
6673 TplDepStats *tpl_stats =
6674 &tpl_frame->tpl_stats_ptr[mi_row * tpl_frame->stride + mi_col];
6675 const int mb_y_offset =
6676 mi_row * MI_SIZE * xd->cur_buf->y_stride + mi_col * MI_SIZE;
6679 set_mv_limits(cm, x, mi_row, mi_col);
6681 for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
6682 int_mv *mv = get_pyramid_mv(tpl_frame, rf_idx, bsize, mi_row, mi_col);
6683 if (ref_frame[rf_idx] == NULL) {
6684 tpl_stats->ready[rf_idx] = 0;
6687 tpl_stats->ready[rf_idx] = 1;
6689 motion_compensated_prediction(
6690 cpi, td, frame_idx, xd->cur_buf->y_buffer + mb_y_offset,
6691 ref_frame[rf_idx]->y_buffer + mb_y_offset, xd->cur_buf->y_stride, bsize,
6692 mi_row, mi_col, &mv->as_mv, rf_idx);
6696 #define CHANGE_MV_SEARCH_ORDER 1
6697 #define USE_PQSORT 1
6699 #if CHANGE_MV_SEARCH_ORDER
6701 static void max_heap_pop(FEATURE_SCORE_LOC **heap, int *size,
6702 FEATURE_SCORE_LOC **output) {
6708 heap[0] = heap[*size];
6713 FEATURE_SCORE_LOC *tmp;
6715 if (r < *size && heap[r]->feature_score > heap[l]->feature_score) {
6718 if (heap[p]->feature_score >= heap[c]->feature_score) {
6734 static void max_heap_push(FEATURE_SCORE_LOC **heap, int *size,
6735 FEATURE_SCORE_LOC *input) {
6737 FEATURE_SCORE_LOC *tmp;
6739 heap[*size] = input;
6743 while (c > 0 && heap[c]->feature_score > heap[p]->feature_score) {
6752 static void add_nb_blocks_to_heap(VP9_COMP *cpi, const TplDepFrame *tpl_frame,
6753 BLOCK_SIZE bsize, int mi_row, int mi_col,
6755 const int mi_unit = num_8x8_blocks_wide_lookup[bsize];
6756 const int dirs[NB_MVS_NUM][2] = { { -1, 0 }, { 0, -1 }, { 1, 0 }, { 0, 1 } };
6758 for (i = 0; i < NB_MVS_NUM; ++i) {
6759 int r = dirs[i][0] * mi_unit;
6760 int c = dirs[i][1] * mi_unit;
6761 if (mi_row + r >= 0 && mi_row + r < tpl_frame->mi_rows && mi_col + c >= 0 &&
6762 mi_col + c < tpl_frame->mi_cols) {
6763 FEATURE_SCORE_LOC *fs_loc =
6764 &cpi->feature_score_loc_arr[(mi_row + r) * tpl_frame->stride +
6766 if (fs_loc->visited == 0) {
6767 max_heap_push(cpi->feature_score_loc_heap, heap_size, fs_loc);
6772 #endif // USE_PQSORT
6773 #endif // CHANGE_MV_SEARCH_ORDER
6775 static void build_motion_field(VP9_COMP *cpi, MACROBLOCKD *xd, int frame_idx,
6776 YV12_BUFFER_CONFIG *ref_frame[3],
6778 VP9_COMMON *cm = &cpi->common;
6779 ThreadData *td = &cpi->td;
6780 TplDepFrame *tpl_frame = &cpi->tpl_stats[frame_idx];
6781 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6782 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6783 const int pw = num_4x4_blocks_wide_lookup[bsize] << 2;
6784 const int ph = num_4x4_blocks_high_lookup[bsize] << 2;
6785 int fs_loc_sort_size;
6786 int fs_loc_heap_size;
6789 tpl_frame->lambda = (pw * ph) >> 2;
6790 assert(pw * ph == tpl_frame->lambda << 2);
6792 fs_loc_sort_size = 0;
6793 for (mi_row = 0; mi_row < cm->mi_rows; mi_row += mi_height) {
6794 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += mi_width) {
6795 const int mb_y_offset =
6796 mi_row * MI_SIZE * xd->cur_buf->y_stride + mi_col * MI_SIZE;
6797 const int bw = 4 << b_width_log2_lookup[bsize];
6798 const int bh = 4 << b_height_log2_lookup[bsize];
6799 TplDepStats *tpl_stats =
6800 &tpl_frame->tpl_stats_ptr[mi_row * tpl_frame->stride + mi_col];
6801 FEATURE_SCORE_LOC *fs_loc =
6802 &cpi->feature_score_loc_arr[mi_row * tpl_frame->stride + mi_col];
6803 tpl_stats->feature_score = get_feature_score(
6804 xd->cur_buf->y_buffer + mb_y_offset, xd->cur_buf->y_stride, bw, bh);
6805 fs_loc->visited = 0;
6806 fs_loc->feature_score = tpl_stats->feature_score;
6807 fs_loc->mi_row = mi_row;
6808 fs_loc->mi_col = mi_col;
6809 cpi->feature_score_loc_sort[fs_loc_sort_size] = fs_loc;
6814 qsort(cpi->feature_score_loc_sort, fs_loc_sort_size,
6815 sizeof(*cpi->feature_score_loc_sort), compare_feature_score);
6817 for (mi_row = 0; mi_row < cm->mi_rows; mi_row += mi_height) {
6818 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += mi_width) {
6820 for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
6821 TplDepStats *tpl_stats =
6822 &tpl_frame->tpl_stats_ptr[mi_row * tpl_frame->stride + mi_col];
6823 tpl_stats->ready[rf_idx] = 0;
6828 #if CHANGE_MV_SEARCH_ORDER
6830 for (i = 0; i < fs_loc_sort_size; ++i) {
6831 FEATURE_SCORE_LOC *fs_loc = cpi->feature_score_loc_sort[i];
6832 do_motion_search(cpi, td, frame_idx, ref_frame, bsize, fs_loc->mi_row,
6835 #else // !USE_PQSORT
6836 fs_loc_heap_size = 0;
6837 max_heap_push(cpi->feature_score_loc_heap, &fs_loc_heap_size,
6838 cpi->feature_score_loc_sort[0]);
6840 while (fs_loc_heap_size > 0) {
6841 FEATURE_SCORE_LOC *fs_loc;
6842 max_heap_pop(cpi->feature_score_loc_heap, &fs_loc_heap_size, &fs_loc);
6844 do_motion_search(cpi, td, frame_idx, ref_frame, bsize, fs_loc->mi_row,
6847 add_nb_blocks_to_heap(cpi, tpl_frame, bsize, fs_loc->mi_row, fs_loc->mi_col,
6850 #endif // !USE_PQSORT
6851 #else // CHANGE_MV_SEARCH_ORDER
6852 for (mi_row = 0; mi_row < cm->mi_rows; mi_row += mi_height) {
6853 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += mi_width) {
6854 do_motion_search(cpi, td, frame_idx, ref_frame, bsize, mi_row, mi_col);
6857 #endif // CHANGE_MV_SEARCH_ORDER
6859 #endif // CONFIG_NON_GREEDY_MV
6861 static void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture,
6862 int frame_idx, BLOCK_SIZE bsize) {
6863 TplDepFrame *tpl_frame = &cpi->tpl_stats[frame_idx];
6864 YV12_BUFFER_CONFIG *this_frame = gf_picture[frame_idx].frame;
6865 YV12_BUFFER_CONFIG *ref_frame[3] = { NULL, NULL, NULL };
6867 VP9_COMMON *cm = &cpi->common;
6868 struct scale_factors sf;
6870 ThreadData *td = &cpi->td;
6871 MACROBLOCK *x = &td->mb;
6872 MACROBLOCKD *xd = &x->e_mbd;
6875 #if CONFIG_VP9_HIGHBITDEPTH
6876 DECLARE_ALIGNED(16, uint16_t, predictor16[32 * 32 * 3]);
6877 DECLARE_ALIGNED(16, uint8_t, predictor8[32 * 32 * 3]);
6880 DECLARE_ALIGNED(16, uint8_t, predictor[32 * 32 * 3]);
6882 DECLARE_ALIGNED(16, int16_t, src_diff[32 * 32]);
6883 DECLARE_ALIGNED(16, tran_low_t, coeff[32 * 32]);
6884 DECLARE_ALIGNED(16, tran_low_t, qcoeff[32 * 32]);
6885 DECLARE_ALIGNED(16, tran_low_t, dqcoeff[32 * 32]);
6887 const TX_SIZE tx_size = max_txsize_lookup[bsize];
6888 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6889 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
6890 int64_t recon_error, sse;
6891 #if CONFIG_NON_GREEDY_MV
6892 int square_block_idx;
6896 // Setup scaling factor
6897 #if CONFIG_VP9_HIGHBITDEPTH
6898 vp9_setup_scale_factors_for_frame(
6899 &sf, this_frame->y_crop_width, this_frame->y_crop_height,
6900 this_frame->y_crop_width, this_frame->y_crop_height,
6901 cpi->common.use_highbitdepth);
6903 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
6904 predictor = CONVERT_TO_BYTEPTR(predictor16);
6906 predictor = predictor8;
6908 vp9_setup_scale_factors_for_frame(
6909 &sf, this_frame->y_crop_width, this_frame->y_crop_height,
6910 this_frame->y_crop_width, this_frame->y_crop_height);
6911 #endif // CONFIG_VP9_HIGHBITDEPTH
6913 // Prepare reference frame pointers. If any reference frame slot is
6914 // unavailable, the pointer will be set to Null.
6915 for (idx = 0; idx < 3; ++idx) {
6916 int rf_idx = gf_picture[frame_idx].ref_frame[idx];
6917 if (rf_idx != -1) ref_frame[idx] = gf_picture[rf_idx].frame;
6920 xd->mi = cm->mi_grid_visible;
6922 xd->cur_buf = this_frame;
6924 // Get rd multiplier set up.
6925 rdmult = vp9_compute_rd_mult_based_on_qindex(cpi, tpl_frame->base_qindex);
6926 set_error_per_bit(&cpi->td.mb, rdmult);
6927 vp9_initialize_me_consts(cpi, &cpi->td.mb, tpl_frame->base_qindex);
6929 tpl_frame->is_valid = 1;
6931 cm->base_qindex = tpl_frame->base_qindex;
6932 vp9_frame_init_quantizer(cpi);
6934 #if CONFIG_NON_GREEDY_MV
6935 for (square_block_idx = 0; square_block_idx < SQUARE_BLOCK_SIZES;
6936 ++square_block_idx) {
6937 BLOCK_SIZE square_bsize = square_block_idx_to_bsize(square_block_idx);
6938 build_motion_field(cpi, xd, frame_idx, ref_frame, square_bsize);
6940 for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
6941 int ref_frame_idx = gf_picture[frame_idx].ref_frame[rf_idx];
6942 if (ref_frame_idx != -1) {
6943 predict_mv_mode_arr(cpi, x, gf_picture, frame_idx, tpl_frame, rf_idx,
6949 for (mi_row = 0; mi_row < cm->mi_rows; mi_row += mi_height) {
6950 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += mi_width) {
6951 mode_estimation(cpi, x, xd, &sf, gf_picture, frame_idx, tpl_frame,
6952 src_diff, coeff, qcoeff, dqcoeff, mi_row, mi_col, bsize,
6953 tx_size, ref_frame, predictor, &recon_error, &sse);
6954 // Motion flow dependency dispenser.
6955 tpl_model_store(tpl_frame->tpl_stats_ptr, mi_row, mi_col, bsize,
6958 tpl_model_update(cpi->tpl_stats, tpl_frame->tpl_stats_ptr, mi_row, mi_col,
6964 #if CONFIG_NON_GREEDY_MV
6965 #define DUMP_TPL_STATS 0
6967 static void dump_buf(uint8_t *buf, int stride, int row, int col, int h, int w) {
6969 printf("%d %d\n", h, w);
6970 for (i = 0; i < h; ++i) {
6971 for (j = 0; j < w; ++j) {
6972 printf("%d ", buf[(row + i) * stride + col + j]);
6978 static void dump_frame_buf(const YV12_BUFFER_CONFIG *frame_buf) {
6979 dump_buf(frame_buf->y_buffer, frame_buf->y_stride, 0, 0, frame_buf->y_height,
6980 frame_buf->y_width);
6981 dump_buf(frame_buf->u_buffer, frame_buf->uv_stride, 0, 0,
6982 frame_buf->uv_height, frame_buf->uv_width);
6983 dump_buf(frame_buf->v_buffer, frame_buf->uv_stride, 0, 0,
6984 frame_buf->uv_height, frame_buf->uv_width);
6987 static void dump_tpl_stats(const VP9_COMP *cpi, int tpl_group_frames,
6988 const GF_GROUP *gf_group,
6989 const GF_PICTURE *gf_picture, BLOCK_SIZE bsize) {
6991 const VP9_COMMON *cm = &cpi->common;
6993 for (frame_idx = 1; frame_idx < tpl_group_frames; ++frame_idx) {
6994 for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
6995 const TplDepFrame *tpl_frame = &cpi->tpl_stats[frame_idx];
6998 const int mi_height = num_8x8_blocks_high_lookup[bsize];
6999 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
7000 ref_frame_idx = gf_picture[frame_idx].ref_frame[rf_idx];
7001 if (ref_frame_idx != -1) {
7002 YV12_BUFFER_CONFIG *ref_frame_buf = gf_picture[ref_frame_idx].frame;
7003 const int gf_frame_offset = gf_group->frame_gop_index[frame_idx];
7004 const int ref_gf_frame_offset =
7005 gf_group->frame_gop_index[ref_frame_idx];
7008 "frame_idx %d mi_rows %d mi_cols %d bsize %d ref_frame_idx %d "
7009 "rf_idx %d gf_frame_offset %d ref_gf_frame_offset %d\n",
7010 frame_idx, cm->mi_rows, cm->mi_cols, mi_width * MI_SIZE,
7011 ref_frame_idx, rf_idx, gf_frame_offset, ref_gf_frame_offset);
7012 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row) {
7013 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
7014 if ((mi_row % mi_height) == 0 && (mi_col % mi_width) == 0) {
7016 *get_pyramid_mv(tpl_frame, rf_idx, bsize, mi_row, mi_col);
7017 printf("%d %d %d %d\n", mi_row, mi_col, mv.as_mv.row,
7022 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row) {
7023 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
7024 if ((mi_row % mi_height) == 0 && (mi_col % mi_width) == 0) {
7025 const TplDepStats *tpl_ptr =
7027 ->tpl_stats_ptr[mi_row * tpl_frame->stride + mi_col];
7028 printf("%f ", tpl_ptr->feature_score);
7034 for (mi_row = 0; mi_row < cm->mi_rows; mi_row += mi_height) {
7035 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += mi_width) {
7038 ->mv_mode_arr[rf_idx][mi_row * tpl_frame->stride + mi_col];
7039 printf("%d ", mv_mode);
7044 dump_frame_buf(gf_picture[frame_idx].frame);
7045 dump_frame_buf(ref_frame_buf);
7050 #endif // DUMP_TPL_STATS
7051 #endif // CONFIG_NON_GREEDY_MV
7053 static void init_tpl_buffer(VP9_COMP *cpi) {
7054 VP9_COMMON *cm = &cpi->common;
7057 const int mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
7058 const int mi_rows = mi_cols_aligned_to_sb(cm->mi_rows);
7059 #if CONFIG_NON_GREEDY_MV
7063 // TODO(angiebird): This probably needs further modifications to support
7064 // frame scaling later on.
7065 if (cpi->feature_score_loc_alloc == 0) {
7066 // The smallest block size of motion field is 4x4, but the mi_unit is 8x8,
7067 // therefore the number of units is "mi_rows * mi_cols * 4" here.
7069 cm, cpi->feature_score_loc_arr,
7070 vpx_calloc(mi_rows * mi_cols * 4, sizeof(*cpi->feature_score_loc_arr)));
7071 CHECK_MEM_ERROR(cm, cpi->feature_score_loc_sort,
7072 vpx_calloc(mi_rows * mi_cols * 4,
7073 sizeof(*cpi->feature_score_loc_sort)));
7074 CHECK_MEM_ERROR(cm, cpi->feature_score_loc_heap,
7075 vpx_calloc(mi_rows * mi_cols * 4,
7076 sizeof(*cpi->feature_score_loc_heap)));
7078 cpi->feature_score_loc_alloc = 1;
7080 vpx_free(cpi->select_mv_arr);
7082 cm, cpi->select_mv_arr,
7083 vpx_calloc(mi_rows * mi_cols * 4, sizeof(*cpi->select_mv_arr)));
7086 // TODO(jingning): Reduce the actual memory use for tpl model build up.
7087 for (frame = 0; frame < MAX_ARF_GOP_SIZE; ++frame) {
7088 if (cpi->tpl_stats[frame].width >= mi_cols &&
7089 cpi->tpl_stats[frame].height >= mi_rows &&
7090 cpi->tpl_stats[frame].tpl_stats_ptr)
7093 #if CONFIG_NON_GREEDY_MV
7094 for (rf_idx = 0; rf_idx < 3; ++rf_idx) {
7095 for (sqr_bsize = 0; sqr_bsize < SQUARE_BLOCK_SIZES; ++sqr_bsize) {
7096 vpx_free(cpi->tpl_stats[frame].pyramid_mv_arr[rf_idx][sqr_bsize]);
7098 cm, cpi->tpl_stats[frame].pyramid_mv_arr[rf_idx][sqr_bsize],
7100 mi_rows * mi_cols * 4,
7102 *cpi->tpl_stats[frame].pyramid_mv_arr[rf_idx][sqr_bsize])));
7104 vpx_free(cpi->tpl_stats[frame].mv_mode_arr[rf_idx]);
7106 cm, cpi->tpl_stats[frame].mv_mode_arr[rf_idx],
7107 vpx_calloc(mi_rows * mi_cols * 4,
7108 sizeof(*cpi->tpl_stats[frame].mv_mode_arr[rf_idx])));
7109 vpx_free(cpi->tpl_stats[frame].rd_diff_arr[rf_idx]);
7111 cm, cpi->tpl_stats[frame].rd_diff_arr[rf_idx],
7112 vpx_calloc(mi_rows * mi_cols * 4,
7113 sizeof(*cpi->tpl_stats[frame].rd_diff_arr[rf_idx])));
7116 vpx_free(cpi->tpl_stats[frame].tpl_stats_ptr);
7117 CHECK_MEM_ERROR(cm, cpi->tpl_stats[frame].tpl_stats_ptr,
7118 vpx_calloc(mi_rows * mi_cols,
7119 sizeof(*cpi->tpl_stats[frame].tpl_stats_ptr)));
7120 cpi->tpl_stats[frame].is_valid = 0;
7121 cpi->tpl_stats[frame].width = mi_cols;
7122 cpi->tpl_stats[frame].height = mi_rows;
7123 cpi->tpl_stats[frame].stride = mi_cols;
7124 cpi->tpl_stats[frame].mi_rows = cm->mi_rows;
7125 cpi->tpl_stats[frame].mi_cols = cm->mi_cols;
7128 for (frame = 0; frame < REF_FRAMES; ++frame) {
7129 cpi->enc_frame_buf[frame].mem_valid = 0;
7130 cpi->enc_frame_buf[frame].released = 1;
7134 static void setup_tpl_stats(VP9_COMP *cpi) {
7135 GF_PICTURE gf_picture[MAX_ARF_GOP_SIZE];
7136 const GF_GROUP *gf_group = &cpi->twopass.gf_group;
7137 int tpl_group_frames = 0;
7139 cpi->tpl_bsize = BLOCK_32X32;
7141 init_gop_frames(cpi, gf_picture, gf_group, &tpl_group_frames);
7143 init_tpl_stats(cpi);
7145 // Backward propagation from tpl_group_frames to 1.
7146 for (frame_idx = tpl_group_frames - 1; frame_idx > 0; --frame_idx) {
7147 if (gf_picture[frame_idx].update_type == USE_BUF_FRAME) continue;
7148 mc_flow_dispenser(cpi, gf_picture, frame_idx, cpi->tpl_bsize);
7150 #if CONFIG_NON_GREEDY_MV
7153 dump_tpl_stats(cpi, tpl_group_frames, gf_group, gf_picture, cpi->tpl_bsize);
7154 #endif // DUMP_TPL_STATS
7155 #endif // CONFIG_NON_GREEDY_MV
7158 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
7159 size_t *size, uint8_t *dest, int64_t *time_stamp,
7160 int64_t *time_end, int flush) {
7161 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
7162 VP9_COMMON *const cm = &cpi->common;
7163 BufferPool *const pool = cm->buffer_pool;
7164 RATE_CONTROL *const rc = &cpi->rc;
7165 struct vpx_usec_timer cmptimer;
7166 YV12_BUFFER_CONFIG *force_src_buffer = NULL;
7167 struct lookahead_entry *last_source = NULL;
7168 struct lookahead_entry *source = NULL;
7170 const int gf_group_index = cpi->twopass.gf_group.index;
7173 if (is_one_pass_cbr_svc(cpi)) {
7174 vp9_one_pass_cbr_svc_start_layer(cpi);
7177 vpx_usec_timer_start(&cmptimer);
7179 vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
7181 // Is multi-arf enabled.
7182 // Note that at the moment multi_arf is only configured for 2 pass VBR and
7183 // will not work properly with svc.
7184 // Enable the Jingning's new "multi_layer_arf" code if "enable_auto_arf"
7185 // is greater than or equal to 2.
7186 if ((oxcf->pass == 2) && !cpi->use_svc && (cpi->oxcf.enable_auto_arf >= 2))
7187 cpi->multi_layer_arf = 1;
7189 cpi->multi_layer_arf = 0;
7192 cm->reset_frame_context = 0;
7193 cm->refresh_frame_context = 1;
7194 if (!is_one_pass_cbr_svc(cpi)) {
7195 cpi->refresh_last_frame = 1;
7196 cpi->refresh_golden_frame = 0;
7197 cpi->refresh_alt_ref_frame = 0;
7200 // Should we encode an arf frame.
7201 arf_src_index = get_arf_src_index(cpi);
7203 if (arf_src_index) {
7204 for (i = 0; i <= arf_src_index; ++i) {
7205 struct lookahead_entry *e = vp9_lookahead_peek(cpi->lookahead, i);
7206 // Avoid creating an alt-ref if there's a forced keyframe pending.
7209 } else if (e->flags == VPX_EFLAG_FORCE_KF) {
7217 // Clear arf index stack before group of pictures processing starts.
7218 if (gf_group_index == 1) {
7219 stack_init(cpi->twopass.gf_group.arf_index_stack, MAX_LAG_BUFFERS * 2);
7220 cpi->twopass.gf_group.stack_size = 0;
7223 if (arf_src_index) {
7224 assert(arf_src_index <= rc->frames_to_key);
7225 if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
7226 cpi->alt_ref_source = source;
7228 #if !CONFIG_REALTIME_ONLY
7229 if ((oxcf->mode != REALTIME) && (oxcf->arnr_max_frames > 0) &&
7230 (oxcf->arnr_strength > 0)) {
7231 int bitrate = cpi->rc.avg_frame_bandwidth / 40;
7232 int not_low_bitrate = bitrate > ALT_REF_AQ_LOW_BITRATE_BOUNDARY;
7234 int not_last_frame = (cpi->lookahead->sz - arf_src_index > 1);
7235 not_last_frame |= ALT_REF_AQ_APPLY_TO_LAST_FRAME;
7237 // Produce the filtered ARF frame.
7238 vp9_temporal_filter(cpi, arf_src_index);
7239 vpx_extend_frame_borders(&cpi->alt_ref_buffer);
7241 // for small bitrates segmentation overhead usually
7242 // eats all bitrate gain from enabling delta quantizers
7243 if (cpi->oxcf.alt_ref_aq != 0 && not_low_bitrate && not_last_frame)
7244 vp9_alt_ref_aq_setup_mode(cpi->alt_ref_aq, cpi);
7246 force_src_buffer = &cpi->alt_ref_buffer;
7251 cpi->refresh_alt_ref_frame = 1;
7252 cpi->refresh_golden_frame = 0;
7253 cpi->refresh_last_frame = 0;
7254 rc->is_src_frame_alt_ref = 0;
7255 rc->source_alt_ref_pending = 0;
7257 rc->source_alt_ref_pending = 0;
7262 // Get last frame source.
7263 if (cm->current_video_frame > 0) {
7264 if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
7268 // Read in the source frame.
7269 if (cpi->use_svc || cpi->svc.set_intra_only_frame)
7270 source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
7272 source = vp9_lookahead_pop(cpi->lookahead, flush);
7274 if (source != NULL) {
7277 // if the flags indicate intra frame, but if the current picture is for
7278 // non-zero spatial layer, it should not be an intra picture.
7279 if ((source->flags & VPX_EFLAG_FORCE_KF) && cpi->use_svc &&
7280 cpi->svc.spatial_layer_id > 0) {
7281 source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF);
7284 // Check to see if the frame should be encoded as an arf overlay.
7285 check_src_altref(cpi, source);
7290 cpi->un_scaled_source = cpi->Source =
7291 force_src_buffer ? force_src_buffer : &source->img;
7293 #ifdef ENABLE_KF_DENOISE
7294 // Copy of raw source for metrics calculation.
7295 if (is_psnr_calc_enabled(cpi))
7296 vp9_copy_and_extend_frame(cpi->Source, &cpi->raw_unscaled_source);
7299 cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
7301 *time_stamp = source->ts_start;
7302 *time_end = source->ts_end;
7303 *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
7306 #if !CONFIG_REALTIME_ONLY
7307 if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
7308 vp9_end_first_pass(cpi); /* get last stats packet */
7309 cpi->twopass.first_pass_done = 1;
7311 #endif // !CONFIG_REALTIME_ONLY
7315 if (source->ts_start < cpi->first_time_stamp_ever) {
7316 cpi->first_time_stamp_ever = source->ts_start;
7317 cpi->last_end_time_stamp_seen = source->ts_start;
7320 // Clear down mmx registers
7321 vpx_clear_system_state();
7323 // adjust frame rates based on timestamps given
7324 if (cm->show_frame) {
7325 if (cpi->use_svc && cpi->svc.use_set_ref_frame_config &&
7326 cpi->svc.duration[cpi->svc.spatial_layer_id] > 0)
7327 vp9_svc_adjust_frame_rate(cpi);
7329 adjust_frame_rate(cpi, source);
7332 if (is_one_pass_cbr_svc(cpi)) {
7333 vp9_update_temporal_layer_framerate(cpi);
7334 vp9_restore_layer_context(cpi);
7337 // Find a free buffer for the new frame, releasing the reference previously
7339 if (cm->new_fb_idx != INVALID_IDX) {
7340 --pool->frame_bufs[cm->new_fb_idx].ref_count;
7342 cm->new_fb_idx = get_free_fb(cm);
7344 if (cm->new_fb_idx == INVALID_IDX) return -1;
7346 cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
7348 // Start with a 0 size frame.
7351 cpi->frame_flags = *frame_flags;
7353 #if !CONFIG_REALTIME_ONLY
7354 if ((oxcf->pass == 2) && !cpi->use_svc) {
7355 vp9_rc_get_second_pass_params(cpi);
7356 } else if (oxcf->pass == 1) {
7357 set_frame_size(cpi);
7359 #endif // !CONFIG_REALTIME_ONLY
7361 if (oxcf->pass != 1 && cpi->level_constraint.level_index >= 0 &&
7362 cpi->level_constraint.fail_flag == 0)
7363 level_rc_framerate(cpi, arf_src_index);
7365 if (cpi->oxcf.pass != 0 || cpi->use_svc || frame_is_intra_only(cm) == 1) {
7366 for (i = 0; i < REFS_PER_FRAME; ++i) cpi->scaled_ref_idx[i] = INVALID_IDX;
7369 if (cpi->kmeans_data_arr_alloc == 0) {
7370 const int mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
7371 const int mi_rows = mi_cols_aligned_to_sb(cm->mi_rows);
7372 #if CONFIG_MULTITHREAD
7373 pthread_mutex_init(&cpi->kmeans_mutex, NULL);
7376 cm, cpi->kmeans_data_arr,
7377 vpx_calloc(mi_rows * mi_cols, sizeof(*cpi->kmeans_data_arr)));
7378 cpi->kmeans_data_stride = mi_cols;
7379 cpi->kmeans_data_arr_alloc = 1;
7382 if (gf_group_index == 1 &&
7383 cpi->twopass.gf_group.update_type[gf_group_index] == ARF_UPDATE &&
7384 cpi->sf.enable_tpl_model) {
7385 init_tpl_buffer(cpi);
7386 vp9_estimate_qp_gop(cpi);
7387 setup_tpl_stats(cpi);
7390 #if CONFIG_BITSTREAM_DEBUG
7391 assert(cpi->oxcf.max_threads == 0 &&
7392 "bitstream debug tool does not support multithreading");
7393 bitstream_queue_record_write();
7395 #if CONFIG_BITSTREAM_DEBUG || CONFIG_MISMATCH_DEBUG
7396 bitstream_queue_set_frame_write(cm->current_video_frame * 2 + cm->show_frame);
7399 cpi->td.mb.fp_src_pred = 0;
7400 #if CONFIG_REALTIME_ONLY
7402 SvcEncode(cpi, size, dest, frame_flags);
7405 Pass0Encode(cpi, size, dest, frame_flags);
7407 #else // !CONFIG_REALTIME_ONLY
7408 if (oxcf->pass == 1 && !cpi->use_svc) {
7409 const int lossless = is_lossless_requested(oxcf);
7410 #if CONFIG_VP9_HIGHBITDEPTH
7411 if (cpi->oxcf.use_highbitdepth)
7412 cpi->td.mb.fwd_txfm4x4 =
7413 lossless ? vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
7415 cpi->td.mb.fwd_txfm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
7416 cpi->td.mb.highbd_inv_txfm_add =
7417 lossless ? vp9_highbd_iwht4x4_add : vp9_highbd_idct4x4_add;
7419 cpi->td.mb.fwd_txfm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
7420 #endif // CONFIG_VP9_HIGHBITDEPTH
7421 cpi->td.mb.inv_txfm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
7422 vp9_first_pass(cpi, source);
7423 } else if (oxcf->pass == 2 && !cpi->use_svc) {
7424 Pass2Encode(cpi, size, dest, frame_flags);
7425 } else if (cpi->use_svc) {
7426 SvcEncode(cpi, size, dest, frame_flags);
7429 Pass0Encode(cpi, size, dest, frame_flags);
7431 #endif // CONFIG_REALTIME_ONLY
7433 if (cm->show_frame) cm->cur_show_frame_fb_idx = cm->new_fb_idx;
7435 if (cm->refresh_frame_context)
7436 cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
7438 // No frame encoded, or frame was dropped, release scaled references.
7439 if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
7440 release_scaled_references(cpi);
7444 cpi->droppable = !frame_is_reference(cpi);
7447 // Save layer specific state.
7448 if (is_one_pass_cbr_svc(cpi) || ((cpi->svc.number_temporal_layers > 1 ||
7449 cpi->svc.number_spatial_layers > 1) &&
7451 vp9_save_layer_context(cpi);
7454 vpx_usec_timer_mark(&cmptimer);
7455 cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
7457 // Should we calculate metrics for the frame.
7458 if (is_psnr_calc_enabled(cpi)) generate_psnr_packet(cpi);
7460 if (cpi->keep_level_stats && oxcf->pass != 1)
7461 update_level_info(cpi, size, arf_src_index);
7463 #if CONFIG_INTERNAL_STATS
7465 if (oxcf->pass != 1) {
7466 double samples = 0.0;
7467 cpi->bytes += (int)(*size);
7469 if (cm->show_frame) {
7470 uint32_t bit_depth = 8;
7471 uint32_t in_bit_depth = 8;
7473 #if CONFIG_VP9_HIGHBITDEPTH
7474 if (cm->use_highbitdepth) {
7475 in_bit_depth = cpi->oxcf.input_bit_depth;
7476 bit_depth = cm->bit_depth;
7480 if (cpi->b_calculate_psnr) {
7481 YV12_BUFFER_CONFIG *orig = cpi->raw_source_frame;
7482 YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
7483 YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer;
7485 #if CONFIG_VP9_HIGHBITDEPTH
7486 vpx_calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd,
7489 vpx_calc_psnr(orig, recon, &psnr);
7490 #endif // CONFIG_VP9_HIGHBITDEPTH
7492 adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3],
7493 psnr.psnr[0], &cpi->psnr);
7494 cpi->total_sq_error += psnr.sse[0];
7495 cpi->total_samples += psnr.samples[0];
7496 samples = psnr.samples[0];
7500 double frame_ssim2 = 0, weight = 0;
7501 #if CONFIG_VP9_POSTPROC
7502 if (vpx_alloc_frame_buffer(
7503 pp, recon->y_crop_width, recon->y_crop_height,
7504 cm->subsampling_x, cm->subsampling_y,
7505 #if CONFIG_VP9_HIGHBITDEPTH
7506 cm->use_highbitdepth,
7508 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment) < 0) {
7509 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
7510 "Failed to allocate post processing buffer");
7513 vp9_ppflags_t ppflags;
7514 ppflags.post_proc_flag = VP9D_DEBLOCK;
7515 ppflags.deblocking_level = 0; // not used in vp9_post_proc_frame()
7516 ppflags.noise_level = 0; // not used in vp9_post_proc_frame()
7517 vp9_post_proc_frame(cm, pp, &ppflags,
7518 cpi->un_scaled_source->y_width);
7521 vpx_clear_system_state();
7523 #if CONFIG_VP9_HIGHBITDEPTH
7524 vpx_calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd,
7525 cpi->oxcf.input_bit_depth);
7527 vpx_calc_psnr(orig, pp, &psnr2);
7528 #endif // CONFIG_VP9_HIGHBITDEPTH
7530 cpi->totalp_sq_error += psnr2.sse[0];
7531 cpi->totalp_samples += psnr2.samples[0];
7532 adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3],
7533 psnr2.psnr[0], &cpi->psnrp);
7535 #if CONFIG_VP9_HIGHBITDEPTH
7536 if (cm->use_highbitdepth) {
7537 frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight, bit_depth,
7540 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
7543 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
7544 #endif // CONFIG_VP9_HIGHBITDEPTH
7546 cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2);
7547 cpi->summed_quality += frame_ssim2 * weight;
7548 cpi->summed_weights += weight;
7550 #if CONFIG_VP9_HIGHBITDEPTH
7551 if (cm->use_highbitdepth) {
7552 frame_ssim2 = vpx_highbd_calc_ssim(orig, pp, &weight, bit_depth,
7555 frame_ssim2 = vpx_calc_ssim(orig, pp, &weight);
7558 frame_ssim2 = vpx_calc_ssim(orig, pp, &weight);
7559 #endif // CONFIG_VP9_HIGHBITDEPTH
7561 cpi->summedp_quality += frame_ssim2 * weight;
7562 cpi->summedp_weights += weight;
7564 if (cm->show_frame) {
7565 FILE *f = fopen("q_used.stt", "a");
7566 fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
7567 cpi->common.current_video_frame, psnr2.psnr[1],
7568 psnr2.psnr[2], psnr2.psnr[3], psnr2.psnr[0], frame_ssim2);
7574 if (cpi->b_calculate_blockiness) {
7575 #if CONFIG_VP9_HIGHBITDEPTH
7576 if (!cm->use_highbitdepth)
7579 double frame_blockiness = vp9_get_blockiness(
7580 cpi->Source->y_buffer, cpi->Source->y_stride,
7581 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
7582 cpi->Source->y_width, cpi->Source->y_height);
7583 cpi->worst_blockiness =
7584 VPXMAX(cpi->worst_blockiness, frame_blockiness);
7585 cpi->total_blockiness += frame_blockiness;
7589 if (cpi->b_calculate_consistency) {
7590 #if CONFIG_VP9_HIGHBITDEPTH
7591 if (!cm->use_highbitdepth)
7594 double this_inconsistency = vpx_get_ssim_metrics(
7595 cpi->Source->y_buffer, cpi->Source->y_stride,
7596 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
7597 cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
7600 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
7601 double consistency =
7602 vpx_sse_to_psnr(samples, peak, (double)cpi->total_inconsistency);
7603 if (consistency > 0.0)
7604 cpi->worst_consistency =
7605 VPXMIN(cpi->worst_consistency, consistency);
7606 cpi->total_inconsistency += this_inconsistency;
7611 double y, u, v, frame_all;
7612 frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
7613 &v, bit_depth, in_bit_depth);
7614 adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
7617 double y, u, v, frame_all;
7618 frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v,
7619 bit_depth, in_bit_depth);
7620 adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
7627 if (is_one_pass_cbr_svc(cpi)) {
7628 if (cm->show_frame) {
7629 ++cpi->svc.spatial_layer_to_encode;
7630 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
7631 cpi->svc.spatial_layer_to_encode = 0;
7635 vpx_clear_system_state();
7639 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
7640 vp9_ppflags_t *flags) {
7641 VP9_COMMON *cm = &cpi->common;
7642 #if !CONFIG_VP9_POSTPROC
7646 if (!cm->show_frame) {
7650 #if CONFIG_VP9_POSTPROC
7651 ret = vp9_post_proc_frame(cm, dest, flags, cpi->un_scaled_source->y_width);
7653 if (cm->frame_to_show) {
7654 *dest = *cm->frame_to_show;
7655 dest->y_width = cm->width;
7656 dest->y_height = cm->height;
7657 dest->uv_width = cm->width >> cm->subsampling_x;
7658 dest->uv_height = cm->height >> cm->subsampling_y;
7663 #endif // !CONFIG_VP9_POSTPROC
7664 vpx_clear_system_state();
7669 int vp9_set_internal_size(VP9_COMP *cpi, VPX_SCALING horiz_mode,
7670 VPX_SCALING vert_mode) {
7671 VP9_COMMON *cm = &cpi->common;
7672 int hr = 0, hs = 0, vr = 0, vs = 0;
7674 if (horiz_mode > ONETWO || vert_mode > ONETWO) return -1;
7676 Scale2Ratio(horiz_mode, &hr, &hs);
7677 Scale2Ratio(vert_mode, &vr, &vs);
7679 // always go to the next whole number
7680 cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
7681 cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
7682 if (cm->current_video_frame) {
7683 assert(cm->width <= cpi->initial_width);
7684 assert(cm->height <= cpi->initial_height);
7687 update_frame_size(cpi);
7692 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
7693 unsigned int height) {
7694 VP9_COMMON *cm = &cpi->common;
7695 #if CONFIG_VP9_HIGHBITDEPTH
7696 check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
7698 check_initial_width(cpi, 1, 1);
7699 #endif // CONFIG_VP9_HIGHBITDEPTH
7701 #if CONFIG_VP9_TEMPORAL_DENOISING
7702 setup_denoiser_buffer(cpi);
7707 if (cm->width > cpi->initial_width) {
7708 cm->width = cpi->initial_width;
7709 printf("Warning: Desired width too large, changed to %d\n", cm->width);
7714 cm->height = height;
7715 if (cm->height > cpi->initial_height) {
7716 cm->height = cpi->initial_height;
7717 printf("Warning: Desired height too large, changed to %d\n", cm->height);
7720 assert(cm->width <= cpi->initial_width);
7721 assert(cm->height <= cpi->initial_height);
7723 update_frame_size(cpi);
7728 void vp9_set_svc(VP9_COMP *cpi, int use_svc) {
7729 cpi->use_svc = use_svc;
7733 int vp9_get_quantizer(VP9_COMP *cpi) { return cpi->common.base_qindex; }
7735 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
7737 (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF)) {
7740 if (flags & VP8_EFLAG_NO_REF_LAST) ref ^= VP9_LAST_FLAG;
7742 if (flags & VP8_EFLAG_NO_REF_GF) ref ^= VP9_GOLD_FLAG;
7744 if (flags & VP8_EFLAG_NO_REF_ARF) ref ^= VP9_ALT_FLAG;
7746 vp9_use_as_reference(cpi, ref);
7750 (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
7751 VP8_EFLAG_FORCE_GF | VP8_EFLAG_FORCE_ARF)) {
7754 if (flags & VP8_EFLAG_NO_UPD_LAST) upd ^= VP9_LAST_FLAG;
7756 if (flags & VP8_EFLAG_NO_UPD_GF) upd ^= VP9_GOLD_FLAG;
7758 if (flags & VP8_EFLAG_NO_UPD_ARF) upd ^= VP9_ALT_FLAG;
7760 vp9_update_reference(cpi, upd);
7763 if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
7764 vp9_update_entropy(cpi, 0);
7768 void vp9_set_row_mt(VP9_COMP *cpi) {
7769 // Enable row based multi-threading for supported modes of encoding
7771 if (((cpi->oxcf.mode == GOOD || cpi->oxcf.mode == BEST) &&
7772 cpi->oxcf.speed < 5 && cpi->oxcf.pass == 1) &&
7773 cpi->oxcf.row_mt && !cpi->use_svc)
7776 if (cpi->oxcf.mode == GOOD && cpi->oxcf.speed < 5 &&
7777 (cpi->oxcf.pass == 0 || cpi->oxcf.pass == 2) && cpi->oxcf.row_mt &&
7781 // In realtime mode, enable row based multi-threading for all the speed levels
7782 // where non-rd path is used.
7783 if (cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5 && cpi->oxcf.row_mt) {
7788 cpi->row_mt_bit_exact = 1;
7790 cpi->row_mt_bit_exact = 0;