Make sure webrtc::VideoSourceInterface is released on the main render thread.
[chromium-blink-merge.git] / third_party / libwebp / enc / frame.c
blob2582244c6c4fdf555622e952cfa4d95a021a06e1
1 // Copyright 2011 Google Inc. All Rights Reserved.
2 //
3 // Use of this source code is governed by a BSD-style license
4 // that can be found in the COPYING file in the root of the source
5 // tree. An additional intellectual property rights grant can be found
6 // in the file PATENTS. All contributing project authors may
7 // be found in the AUTHORS file in the root of the source tree.
8 // -----------------------------------------------------------------------------
9 //
10 // frame coding and analysis
12 // Author: Skal (pascal.massimino@gmail.com)
14 #include <assert.h>
15 #include <stdlib.h>
16 #include <string.h>
17 #include <math.h>
19 #include "./vp8enci.h"
20 #include "./cost.h"
21 #include "../webp/format_constants.h" // RIFF constants
23 #define SEGMENT_VISU 0
24 #define DEBUG_SEARCH 0 // useful to track search convergence
26 // On-the-fly info about the current set of residuals. Handy to avoid
27 // passing zillions of params.
28 typedef struct {
29 int first;
30 int last;
31 const int16_t* coeffs;
33 int coeff_type;
34 ProbaArray* prob;
35 StatsArray* stats;
36 CostArray* cost;
37 } VP8Residual;
39 //------------------------------------------------------------------------------
40 // multi-pass convergence
42 #define HEADER_SIZE_ESTIMATE (RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE + \
43 VP8_FRAME_HEADER_SIZE)
44 #define DQ_LIMIT 0.4 // convergence is considered reached if dq < DQ_LIMIT
45 // we allow 2k of extra head-room in PARTITION0 limit.
46 #define PARTITION0_SIZE_LIMIT ((VP8_MAX_PARTITION0_SIZE - 2048ULL) << 11)
48 typedef struct { // struct for organizing convergence in either size or PSNR
49 int is_first;
50 float dq;
51 float q, last_q;
52 double value, last_value; // PSNR or size
53 double target;
54 int do_size_search;
55 } PassStats;
57 static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) {
58 const uint64_t target_size = (uint64_t)enc->config_->target_size;
59 const int do_size_search = (target_size != 0);
60 const float target_PSNR = enc->config_->target_PSNR;
62 s->is_first = 1;
63 s->dq = 10.f;
64 s->q = s->last_q = enc->config_->quality;
65 s->target = do_size_search ? (double)target_size
66 : (target_PSNR > 0.) ? target_PSNR
67 : 40.; // default, just in case
68 s->value = s->last_value = 0.;
69 s->do_size_search = do_size_search;
70 return do_size_search;
73 static float Clamp(float v, float min, float max) {
74 return (v < min) ? min : (v > max) ? max : v;
77 static float ComputeNextQ(PassStats* const s) {
78 float dq;
79 if (s->is_first) {
80 dq = (s->value > s->target) ? -s->dq : s->dq;
81 s->is_first = 0;
82 } else if (s->value != s->last_value) {
83 const double slope = (s->target - s->value) / (s->last_value - s->value);
84 dq = (float)(slope * (s->last_q - s->q));
85 } else {
86 dq = 0.; // we're done?!
88 // Limit variable to avoid large swings.
89 s->dq = Clamp(dq, -30.f, 30.f);
90 s->last_q = s->q;
91 s->last_value = s->value;
92 s->q = Clamp(s->q + s->dq, 0.f, 100.f);
93 return s->q;
96 //------------------------------------------------------------------------------
97 // Tables for level coding
99 const uint8_t VP8EncBands[16 + 1] = {
100 0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
101 0 // sentinel
104 const uint8_t VP8Cat3[] = { 173, 148, 140 };
105 const uint8_t VP8Cat4[] = { 176, 155, 140, 135 };
106 const uint8_t VP8Cat5[] = { 180, 157, 141, 134, 130 };
107 const uint8_t VP8Cat6[] =
108 { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
110 //------------------------------------------------------------------------------
111 // Reset the statistics about: number of skips, token proba, level cost,...
113 static void ResetStats(VP8Encoder* const enc) {
114 VP8Proba* const proba = &enc->proba_;
115 VP8CalculateLevelCosts(proba);
116 proba->nb_skip_ = 0;
119 //------------------------------------------------------------------------------
120 // Skip decision probability
122 #define SKIP_PROBA_THRESHOLD 250 // value below which using skip_proba is OK.
124 static int CalcSkipProba(uint64_t nb, uint64_t total) {
125 return (int)(total ? (total - nb) * 255 / total : 255);
128 // Returns the bit-cost for coding the skip probability.
129 static int FinalizeSkipProba(VP8Encoder* const enc) {
130 VP8Proba* const proba = &enc->proba_;
131 const int nb_mbs = enc->mb_w_ * enc->mb_h_;
132 const int nb_events = proba->nb_skip_;
133 int size;
134 proba->skip_proba_ = CalcSkipProba(nb_events, nb_mbs);
135 proba->use_skip_proba_ = (proba->skip_proba_ < SKIP_PROBA_THRESHOLD);
136 size = 256; // 'use_skip_proba' bit
137 if (proba->use_skip_proba_) {
138 size += nb_events * VP8BitCost(1, proba->skip_proba_)
139 + (nb_mbs - nb_events) * VP8BitCost(0, proba->skip_proba_);
140 size += 8 * 256; // cost of signaling the skip_proba_ itself.
142 return size;
145 //------------------------------------------------------------------------------
146 // Recording of token probabilities.
148 static void ResetTokenStats(VP8Encoder* const enc) {
149 VP8Proba* const proba = &enc->proba_;
150 memset(proba->stats_, 0, sizeof(proba->stats_));
153 // Record proba context used
154 static int Record(int bit, proba_t* const stats) {
155 proba_t p = *stats;
156 if (p >= 0xffff0000u) { // an overflow is inbound.
157 p = ((p + 1u) >> 1) & 0x7fff7fffu; // -> divide the stats by 2.
159 // record bit count (lower 16 bits) and increment total count (upper 16 bits).
160 p += 0x00010000u + bit;
161 *stats = p;
162 return bit;
165 // We keep the table free variant around for reference, in case.
166 #define USE_LEVEL_CODE_TABLE
168 // Simulate block coding, but only record statistics.
169 // Note: no need to record the fixed probas.
170 static int RecordCoeffs(int ctx, const VP8Residual* const res) {
171 int n = res->first;
172 // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
173 proba_t* s = res->stats[n][ctx];
174 if (res->last < 0) {
175 Record(0, s + 0);
176 return 0;
178 while (n <= res->last) {
179 int v;
180 Record(1, s + 0); // order of record doesn't matter
181 while ((v = res->coeffs[n++]) == 0) {
182 Record(0, s + 1);
183 s = res->stats[VP8EncBands[n]][0];
185 Record(1, s + 1);
186 if (!Record(2u < (unsigned int)(v + 1), s + 2)) { // v = -1 or 1
187 s = res->stats[VP8EncBands[n]][1];
188 } else {
189 v = abs(v);
190 #if !defined(USE_LEVEL_CODE_TABLE)
191 if (!Record(v > 4, s + 3)) {
192 if (Record(v != 2, s + 4))
193 Record(v == 4, s + 5);
194 } else if (!Record(v > 10, s + 6)) {
195 Record(v > 6, s + 7);
196 } else if (!Record((v >= 3 + (8 << 2)), s + 8)) {
197 Record((v >= 3 + (8 << 1)), s + 9);
198 } else {
199 Record((v >= 3 + (8 << 3)), s + 10);
201 #else
202 if (v > MAX_VARIABLE_LEVEL)
203 v = MAX_VARIABLE_LEVEL;
206 const int bits = VP8LevelCodes[v - 1][1];
207 int pattern = VP8LevelCodes[v - 1][0];
208 int i;
209 for (i = 0; (pattern >>= 1) != 0; ++i) {
210 const int mask = 2 << i;
211 if (pattern & 1) Record(!!(bits & mask), s + 3 + i);
214 #endif
215 s = res->stats[VP8EncBands[n]][2];
218 if (n < 16) Record(0, s + 0);
219 return 1;
222 // Collect statistics and deduce probabilities for next coding pass.
223 // Return the total bit-cost for coding the probability updates.
224 static int CalcTokenProba(int nb, int total) {
225 assert(nb <= total);
226 return nb ? (255 - nb * 255 / total) : 255;
229 // Cost of coding 'nb' 1's and 'total-nb' 0's using 'proba' probability.
230 static int BranchCost(int nb, int total, int proba) {
231 return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
234 static int FinalizeTokenProbas(VP8Proba* const proba) {
235 int has_changed = 0;
236 int size = 0;
237 int t, b, c, p;
238 for (t = 0; t < NUM_TYPES; ++t) {
239 for (b = 0; b < NUM_BANDS; ++b) {
240 for (c = 0; c < NUM_CTX; ++c) {
241 for (p = 0; p < NUM_PROBAS; ++p) {
242 const proba_t stats = proba->stats_[t][b][c][p];
243 const int nb = (stats >> 0) & 0xffff;
244 const int total = (stats >> 16) & 0xffff;
245 const int update_proba = VP8CoeffsUpdateProba[t][b][c][p];
246 const int old_p = VP8CoeffsProba0[t][b][c][p];
247 const int new_p = CalcTokenProba(nb, total);
248 const int old_cost = BranchCost(nb, total, old_p)
249 + VP8BitCost(0, update_proba);
250 const int new_cost = BranchCost(nb, total, new_p)
251 + VP8BitCost(1, update_proba)
252 + 8 * 256;
253 const int use_new_p = (old_cost > new_cost);
254 size += VP8BitCost(use_new_p, update_proba);
255 if (use_new_p) { // only use proba that seem meaningful enough.
256 proba->coeffs_[t][b][c][p] = new_p;
257 has_changed |= (new_p != old_p);
258 size += 8 * 256;
259 } else {
260 proba->coeffs_[t][b][c][p] = old_p;
266 proba->dirty_ = has_changed;
267 return size;
270 //------------------------------------------------------------------------------
271 // Finalize Segment probability based on the coding tree
273 static int GetProba(int a, int b) {
274 const int total = a + b;
275 return (total == 0) ? 255 // that's the default probability.
276 : (255 * a + total / 2) / total; // rounded proba
279 static void SetSegmentProbas(VP8Encoder* const enc) {
280 int p[NUM_MB_SEGMENTS] = { 0 };
281 int n;
283 for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
284 const VP8MBInfo* const mb = &enc->mb_info_[n];
285 p[mb->segment_]++;
287 if (enc->pic_->stats != NULL) {
288 for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
289 enc->pic_->stats->segment_size[n] = p[n];
292 if (enc->segment_hdr_.num_segments_ > 1) {
293 uint8_t* const probas = enc->proba_.segments_;
294 probas[0] = GetProba(p[0] + p[1], p[2] + p[3]);
295 probas[1] = GetProba(p[0], p[1]);
296 probas[2] = GetProba(p[2], p[3]);
298 enc->segment_hdr_.update_map_ =
299 (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
300 enc->segment_hdr_.size_ =
301 p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
302 p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
303 p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) +
304 p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2]));
305 } else {
306 enc->segment_hdr_.update_map_ = 0;
307 enc->segment_hdr_.size_ = 0;
311 //------------------------------------------------------------------------------
312 // helper functions for residuals struct VP8Residual.
314 static void InitResidual(int first, int coeff_type,
315 VP8Encoder* const enc, VP8Residual* const res) {
316 res->coeff_type = coeff_type;
317 res->prob = enc->proba_.coeffs_[coeff_type];
318 res->stats = enc->proba_.stats_[coeff_type];
319 res->cost = enc->proba_.level_cost_[coeff_type];
320 res->first = first;
323 static void SetResidualCoeffs(const int16_t* const coeffs,
324 VP8Residual* const res) {
325 int n;
326 res->last = -1;
327 for (n = 15; n >= res->first; --n) {
328 if (coeffs[n]) {
329 res->last = n;
330 break;
333 res->coeffs = coeffs;
336 //------------------------------------------------------------------------------
337 // Mode costs
339 static int GetResidualCost(int ctx0, const VP8Residual* const res) {
340 int n = res->first;
341 // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
342 int p0 = res->prob[n][ctx0][0];
343 const uint16_t* t = res->cost[n][ctx0];
344 int cost;
346 if (res->last < 0) {
347 return VP8BitCost(0, p0);
349 cost = VP8BitCost(1, p0);
350 for (; n < res->last; ++n) {
351 const int v = abs(res->coeffs[n]);
352 const int b = VP8EncBands[n + 1];
353 const int ctx = (v >= 2) ? 2 : v;
354 cost += VP8LevelCost(t, v);
355 t = res->cost[b][ctx];
356 // the masking trick is faster than "if (v) cost += ..." with clang
357 cost += (v ? ~0U : 0) & VP8BitCost(1, res->prob[b][ctx][0]);
359 // Last coefficient is always non-zero
361 const int v = abs(res->coeffs[n]);
362 assert(v != 0);
363 cost += VP8LevelCost(t, v);
364 if (n < 15) {
365 const int b = VP8EncBands[n + 1];
366 const int ctx = (v == 1) ? 1 : 2;
367 const int last_p0 = res->prob[b][ctx][0];
368 cost += VP8BitCost(0, last_p0);
371 return cost;
374 int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]) {
375 const int x = (it->i4_ & 3), y = (it->i4_ >> 2);
376 VP8Residual res;
377 VP8Encoder* const enc = it->enc_;
378 int R = 0;
379 int ctx;
381 InitResidual(0, 3, enc, &res);
382 ctx = it->top_nz_[x] + it->left_nz_[y];
383 SetResidualCoeffs(levels, &res);
384 R += GetResidualCost(ctx, &res);
385 return R;
388 int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd) {
389 VP8Residual res;
390 VP8Encoder* const enc = it->enc_;
391 int x, y;
392 int R = 0;
394 VP8IteratorNzToBytes(it); // re-import the non-zero context
396 // DC
397 InitResidual(0, 1, enc, &res);
398 SetResidualCoeffs(rd->y_dc_levels, &res);
399 R += GetResidualCost(it->top_nz_[8] + it->left_nz_[8], &res);
401 // AC
402 InitResidual(1, 0, enc, &res);
403 for (y = 0; y < 4; ++y) {
404 for (x = 0; x < 4; ++x) {
405 const int ctx = it->top_nz_[x] + it->left_nz_[y];
406 SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
407 R += GetResidualCost(ctx, &res);
408 it->top_nz_[x] = it->left_nz_[y] = (res.last >= 0);
411 return R;
414 int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
415 VP8Residual res;
416 VP8Encoder* const enc = it->enc_;
417 int ch, x, y;
418 int R = 0;
420 VP8IteratorNzToBytes(it); // re-import the non-zero context
422 InitResidual(0, 2, enc, &res);
423 for (ch = 0; ch <= 2; ch += 2) {
424 for (y = 0; y < 2; ++y) {
425 for (x = 0; x < 2; ++x) {
426 const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
427 SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
428 R += GetResidualCost(ctx, &res);
429 it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = (res.last >= 0);
433 return R;
436 //------------------------------------------------------------------------------
437 // Coefficient coding
439 static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
440 int n = res->first;
441 // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
442 const uint8_t* p = res->prob[n][ctx];
443 if (!VP8PutBit(bw, res->last >= 0, p[0])) {
444 return 0;
447 while (n < 16) {
448 const int c = res->coeffs[n++];
449 const int sign = c < 0;
450 int v = sign ? -c : c;
451 if (!VP8PutBit(bw, v != 0, p[1])) {
452 p = res->prob[VP8EncBands[n]][0];
453 continue;
455 if (!VP8PutBit(bw, v > 1, p[2])) {
456 p = res->prob[VP8EncBands[n]][1];
457 } else {
458 if (!VP8PutBit(bw, v > 4, p[3])) {
459 if (VP8PutBit(bw, v != 2, p[4]))
460 VP8PutBit(bw, v == 4, p[5]);
461 } else if (!VP8PutBit(bw, v > 10, p[6])) {
462 if (!VP8PutBit(bw, v > 6, p[7])) {
463 VP8PutBit(bw, v == 6, 159);
464 } else {
465 VP8PutBit(bw, v >= 9, 165);
466 VP8PutBit(bw, !(v & 1), 145);
468 } else {
469 int mask;
470 const uint8_t* tab;
471 if (v < 3 + (8 << 1)) { // VP8Cat3 (3b)
472 VP8PutBit(bw, 0, p[8]);
473 VP8PutBit(bw, 0, p[9]);
474 v -= 3 + (8 << 0);
475 mask = 1 << 2;
476 tab = VP8Cat3;
477 } else if (v < 3 + (8 << 2)) { // VP8Cat4 (4b)
478 VP8PutBit(bw, 0, p[8]);
479 VP8PutBit(bw, 1, p[9]);
480 v -= 3 + (8 << 1);
481 mask = 1 << 3;
482 tab = VP8Cat4;
483 } else if (v < 3 + (8 << 3)) { // VP8Cat5 (5b)
484 VP8PutBit(bw, 1, p[8]);
485 VP8PutBit(bw, 0, p[10]);
486 v -= 3 + (8 << 2);
487 mask = 1 << 4;
488 tab = VP8Cat5;
489 } else { // VP8Cat6 (11b)
490 VP8PutBit(bw, 1, p[8]);
491 VP8PutBit(bw, 1, p[10]);
492 v -= 3 + (8 << 3);
493 mask = 1 << 10;
494 tab = VP8Cat6;
496 while (mask) {
497 VP8PutBit(bw, !!(v & mask), *tab++);
498 mask >>= 1;
501 p = res->prob[VP8EncBands[n]][2];
503 VP8PutBitUniform(bw, sign);
504 if (n == 16 || !VP8PutBit(bw, n <= res->last, p[0])) {
505 return 1; // EOB
508 return 1;
511 static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
512 const VP8ModeScore* const rd) {
513 int x, y, ch;
514 VP8Residual res;
515 uint64_t pos1, pos2, pos3;
516 const int i16 = (it->mb_->type_ == 1);
517 const int segment = it->mb_->segment_;
518 VP8Encoder* const enc = it->enc_;
520 VP8IteratorNzToBytes(it);
522 pos1 = VP8BitWriterPos(bw);
523 if (i16) {
524 InitResidual(0, 1, enc, &res);
525 SetResidualCoeffs(rd->y_dc_levels, &res);
526 it->top_nz_[8] = it->left_nz_[8] =
527 PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res);
528 InitResidual(1, 0, enc, &res);
529 } else {
530 InitResidual(0, 3, enc, &res);
533 // luma-AC
534 for (y = 0; y < 4; ++y) {
535 for (x = 0; x < 4; ++x) {
536 const int ctx = it->top_nz_[x] + it->left_nz_[y];
537 SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
538 it->top_nz_[x] = it->left_nz_[y] = PutCoeffs(bw, ctx, &res);
541 pos2 = VP8BitWriterPos(bw);
543 // U/V
544 InitResidual(0, 2, enc, &res);
545 for (ch = 0; ch <= 2; ch += 2) {
546 for (y = 0; y < 2; ++y) {
547 for (x = 0; x < 2; ++x) {
548 const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
549 SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
550 it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
551 PutCoeffs(bw, ctx, &res);
555 pos3 = VP8BitWriterPos(bw);
556 it->luma_bits_ = pos2 - pos1;
557 it->uv_bits_ = pos3 - pos2;
558 it->bit_count_[segment][i16] += it->luma_bits_;
559 it->bit_count_[segment][2] += it->uv_bits_;
560 VP8IteratorBytesToNz(it);
563 // Same as CodeResiduals, but doesn't actually write anything.
564 // Instead, it just records the event distribution.
565 static void RecordResiduals(VP8EncIterator* const it,
566 const VP8ModeScore* const rd) {
567 int x, y, ch;
568 VP8Residual res;
569 VP8Encoder* const enc = it->enc_;
571 VP8IteratorNzToBytes(it);
573 if (it->mb_->type_ == 1) { // i16x16
574 InitResidual(0, 1, enc, &res);
575 SetResidualCoeffs(rd->y_dc_levels, &res);
576 it->top_nz_[8] = it->left_nz_[8] =
577 RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
578 InitResidual(1, 0, enc, &res);
579 } else {
580 InitResidual(0, 3, enc, &res);
583 // luma-AC
584 for (y = 0; y < 4; ++y) {
585 for (x = 0; x < 4; ++x) {
586 const int ctx = it->top_nz_[x] + it->left_nz_[y];
587 SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
588 it->top_nz_[x] = it->left_nz_[y] = RecordCoeffs(ctx, &res);
592 // U/V
593 InitResidual(0, 2, enc, &res);
594 for (ch = 0; ch <= 2; ch += 2) {
595 for (y = 0; y < 2; ++y) {
596 for (x = 0; x < 2; ++x) {
597 const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
598 SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
599 it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
600 RecordCoeffs(ctx, &res);
605 VP8IteratorBytesToNz(it);
608 //------------------------------------------------------------------------------
609 // Token buffer
611 #if !defined(DISABLE_TOKEN_BUFFER)
613 static void RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
614 VP8TBuffer* const tokens) {
615 int x, y, ch;
616 VP8Residual res;
617 VP8Encoder* const enc = it->enc_;
619 VP8IteratorNzToBytes(it);
620 if (it->mb_->type_ == 1) { // i16x16
621 const int ctx = it->top_nz_[8] + it->left_nz_[8];
622 InitResidual(0, 1, enc, &res);
623 SetResidualCoeffs(rd->y_dc_levels, &res);
624 it->top_nz_[8] = it->left_nz_[8] =
625 VP8RecordCoeffTokens(ctx, 1,
626 res.first, res.last, res.coeffs, tokens);
627 RecordCoeffs(ctx, &res);
628 InitResidual(1, 0, enc, &res);
629 } else {
630 InitResidual(0, 3, enc, &res);
633 // luma-AC
634 for (y = 0; y < 4; ++y) {
635 for (x = 0; x < 4; ++x) {
636 const int ctx = it->top_nz_[x] + it->left_nz_[y];
637 SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
638 it->top_nz_[x] = it->left_nz_[y] =
639 VP8RecordCoeffTokens(ctx, res.coeff_type,
640 res.first, res.last, res.coeffs, tokens);
641 RecordCoeffs(ctx, &res);
645 // U/V
646 InitResidual(0, 2, enc, &res);
647 for (ch = 0; ch <= 2; ch += 2) {
648 for (y = 0; y < 2; ++y) {
649 for (x = 0; x < 2; ++x) {
650 const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
651 SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
652 it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
653 VP8RecordCoeffTokens(ctx, 2,
654 res.first, res.last, res.coeffs, tokens);
655 RecordCoeffs(ctx, &res);
659 VP8IteratorBytesToNz(it);
662 #endif // !DISABLE_TOKEN_BUFFER
664 //------------------------------------------------------------------------------
665 // ExtraInfo map / Debug function
667 #if SEGMENT_VISU
668 static void SetBlock(uint8_t* p, int value, int size) {
669 int y;
670 for (y = 0; y < size; ++y) {
671 memset(p, value, size);
672 p += BPS;
675 #endif
677 static void ResetSSE(VP8Encoder* const enc) {
678 enc->sse_[0] = 0;
679 enc->sse_[1] = 0;
680 enc->sse_[2] = 0;
681 // Note: enc->sse_[3] is managed by alpha.c
682 enc->sse_count_ = 0;
685 static void StoreSSE(const VP8EncIterator* const it) {
686 VP8Encoder* const enc = it->enc_;
687 const uint8_t* const in = it->yuv_in_;
688 const uint8_t* const out = it->yuv_out_;
689 // Note: not totally accurate at boundary. And doesn't include in-loop filter.
690 enc->sse_[0] += VP8SSE16x16(in + Y_OFF, out + Y_OFF);
691 enc->sse_[1] += VP8SSE8x8(in + U_OFF, out + U_OFF);
692 enc->sse_[2] += VP8SSE8x8(in + V_OFF, out + V_OFF);
693 enc->sse_count_ += 16 * 16;
696 static void StoreSideInfo(const VP8EncIterator* const it) {
697 VP8Encoder* const enc = it->enc_;
698 const VP8MBInfo* const mb = it->mb_;
699 WebPPicture* const pic = enc->pic_;
701 if (pic->stats != NULL) {
702 StoreSSE(it);
703 enc->block_count_[0] += (mb->type_ == 0);
704 enc->block_count_[1] += (mb->type_ == 1);
705 enc->block_count_[2] += (mb->skip_ != 0);
708 if (pic->extra_info != NULL) {
709 uint8_t* const info = &pic->extra_info[it->x_ + it->y_ * enc->mb_w_];
710 switch (pic->extra_info_type) {
711 case 1: *info = mb->type_; break;
712 case 2: *info = mb->segment_; break;
713 case 3: *info = enc->dqm_[mb->segment_].quant_; break;
714 case 4: *info = (mb->type_ == 1) ? it->preds_[0] : 0xff; break;
715 case 5: *info = mb->uv_mode_; break;
716 case 6: {
717 const int b = (int)((it->luma_bits_ + it->uv_bits_ + 7) >> 3);
718 *info = (b > 255) ? 255 : b; break;
720 case 7: *info = mb->alpha_; break;
721 default: *info = 0; break;
724 #if SEGMENT_VISU // visualize segments and prediction modes
725 SetBlock(it->yuv_out_ + Y_OFF, mb->segment_ * 64, 16);
726 SetBlock(it->yuv_out_ + U_OFF, it->preds_[0] * 64, 8);
727 SetBlock(it->yuv_out_ + V_OFF, mb->uv_mode_ * 64, 8);
728 #endif
731 static double GetPSNR(uint64_t mse, uint64_t size) {
732 return (mse > 0 && size > 0) ? 10. * log10(255. * 255. * size / mse) : 99;
735 //------------------------------------------------------------------------------
736 // StatLoop(): only collect statistics (number of skips, token usage, ...).
737 // This is used for deciding optimal probabilities. It also modifies the
738 // quantizer value if some target (size, PSNR) was specified.
740 static void SetLoopParams(VP8Encoder* const enc, float q) {
741 // Make sure the quality parameter is inside valid bounds
742 q = Clamp(q, 0.f, 100.f);
744 VP8SetSegmentParams(enc, q); // setup segment quantizations and filters
745 SetSegmentProbas(enc); // compute segment probabilities
747 ResetStats(enc);
748 ResetSSE(enc);
751 static uint64_t OneStatPass(VP8Encoder* const enc, VP8RDLevel rd_opt,
752 int nb_mbs, int percent_delta,
753 PassStats* const s) {
754 VP8EncIterator it;
755 uint64_t size = 0;
756 uint64_t size_p0 = 0;
757 uint64_t distortion = 0;
758 const uint64_t pixel_count = nb_mbs * 384;
760 VP8IteratorInit(enc, &it);
761 SetLoopParams(enc, s->q);
762 do {
763 VP8ModeScore info;
764 VP8IteratorImport(&it, NULL);
765 if (VP8Decimate(&it, &info, rd_opt)) {
766 // Just record the number of skips and act like skip_proba is not used.
767 enc->proba_.nb_skip_++;
769 RecordResiduals(&it, &info);
770 size += info.R + info.H;
771 size_p0 += info.H;
772 distortion += info.D;
773 if (percent_delta && !VP8IteratorProgress(&it, percent_delta))
774 return 0;
775 VP8IteratorSaveBoundary(&it);
776 } while (VP8IteratorNext(&it) && --nb_mbs > 0);
778 size_p0 += enc->segment_hdr_.size_;
779 if (s->do_size_search) {
780 size += FinalizeSkipProba(enc);
781 size += FinalizeTokenProbas(&enc->proba_);
782 size = ((size + size_p0 + 1024) >> 11) + HEADER_SIZE_ESTIMATE;
783 s->value = (double)size;
784 } else {
785 s->value = GetPSNR(distortion, pixel_count);
787 return size_p0;
790 static int StatLoop(VP8Encoder* const enc) {
791 const int method = enc->method_;
792 const int do_search = enc->do_search_;
793 const int fast_probe = ((method == 0 || method == 3) && !do_search);
794 int num_pass_left = enc->config_->pass;
795 const int task_percent = 20;
796 const int percent_per_pass =
797 (task_percent + num_pass_left / 2) / num_pass_left;
798 const int final_percent = enc->percent_ + task_percent;
799 const VP8RDLevel rd_opt =
800 (method >= 3 || do_search) ? RD_OPT_BASIC : RD_OPT_NONE;
801 int nb_mbs = enc->mb_w_ * enc->mb_h_;
802 PassStats stats;
804 InitPassStats(enc, &stats);
805 ResetTokenStats(enc);
807 // Fast mode: quick analysis pass over few mbs. Better than nothing.
808 if (fast_probe) {
809 if (method == 3) { // we need more stats for method 3 to be reliable.
810 nb_mbs = (nb_mbs > 200) ? nb_mbs >> 1 : 100;
811 } else {
812 nb_mbs = (nb_mbs > 200) ? nb_mbs >> 2 : 50;
816 while (num_pass_left-- > 0) {
817 const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
818 (num_pass_left == 0) ||
819 (enc->max_i4_header_bits_ == 0);
820 const uint64_t size_p0 =
821 OneStatPass(enc, rd_opt, nb_mbs, percent_per_pass, &stats);
822 if (size_p0 == 0) return 0;
823 #if (DEBUG_SEARCH > 0)
824 printf("#%d value:%.1lf -> %.1lf q:%.2f -> %.2f\n",
825 num_pass_left, stats.last_value, stats.value, stats.last_q, stats.q);
826 #endif
827 if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
828 ++num_pass_left;
829 enc->max_i4_header_bits_ >>= 1; // strengthen header bit limitation...
830 continue; // ...and start over
832 if (is_last_pass) {
833 break;
835 // If no target size: just do several pass without changing 'q'
836 if (do_search) {
837 ComputeNextQ(&stats);
838 if (fabs(stats.dq) <= DQ_LIMIT) break;
841 if (!do_search || !stats.do_size_search) {
842 // Need to finalize probas now, since it wasn't done during the search.
843 FinalizeSkipProba(enc);
844 FinalizeTokenProbas(&enc->proba_);
846 VP8CalculateLevelCosts(&enc->proba_); // finalize costs
847 return WebPReportProgress(enc->pic_, final_percent, &enc->percent_);
850 //------------------------------------------------------------------------------
851 // Main loops
854 static const int kAverageBytesPerMB[8] = { 50, 24, 16, 9, 7, 5, 3, 2 };
856 static int PreLoopInitialize(VP8Encoder* const enc) {
857 int p;
858 int ok = 1;
859 const int average_bytes_per_MB = kAverageBytesPerMB[enc->base_quant_ >> 4];
860 const int bytes_per_parts =
861 enc->mb_w_ * enc->mb_h_ * average_bytes_per_MB / enc->num_parts_;
862 // Initialize the bit-writers
863 for (p = 0; ok && p < enc->num_parts_; ++p) {
864 ok = VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
866 if (!ok) VP8EncFreeBitWriters(enc); // malloc error occurred
867 return ok;
870 static int PostLoopFinalize(VP8EncIterator* const it, int ok) {
871 VP8Encoder* const enc = it->enc_;
872 if (ok) { // Finalize the partitions, check for extra errors.
873 int p;
874 for (p = 0; p < enc->num_parts_; ++p) {
875 VP8BitWriterFinish(enc->parts_ + p);
876 ok &= !enc->parts_[p].error_;
880 if (ok) { // All good. Finish up.
881 if (enc->pic_->stats != NULL) { // finalize byte counters...
882 int i, s;
883 for (i = 0; i <= 2; ++i) {
884 for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
885 enc->residual_bytes_[i][s] = (int)((it->bit_count_[s][i] + 7) >> 3);
889 VP8AdjustFilterStrength(it); // ...and store filter stats.
890 } else {
891 // Something bad happened -> need to do some memory cleanup.
892 VP8EncFreeBitWriters(enc);
894 return ok;
897 //------------------------------------------------------------------------------
898 // VP8EncLoop(): does the final bitstream coding.
900 static void ResetAfterSkip(VP8EncIterator* const it) {
901 if (it->mb_->type_ == 1) {
902 *it->nz_ = 0; // reset all predictors
903 it->left_nz_[8] = 0;
904 } else {
905 *it->nz_ &= (1 << 24); // preserve the dc_nz bit
909 int VP8EncLoop(VP8Encoder* const enc) {
910 VP8EncIterator it;
911 int ok = PreLoopInitialize(enc);
912 if (!ok) return 0;
914 StatLoop(enc); // stats-collection loop
916 VP8IteratorInit(enc, &it);
917 VP8InitFilter(&it);
918 do {
919 VP8ModeScore info;
920 const int dont_use_skip = !enc->proba_.use_skip_proba_;
921 const VP8RDLevel rd_opt = enc->rd_opt_level_;
923 VP8IteratorImport(&it, NULL);
924 // Warning! order is important: first call VP8Decimate() and
925 // *then* decide how to code the skip decision if there's one.
926 if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) {
927 CodeResiduals(it.bw_, &it, &info);
928 } else { // reset predictors after a skip
929 ResetAfterSkip(&it);
931 #ifdef WEBP_EXPERIMENTAL_FEATURES
932 if (enc->use_layer_) {
933 VP8EncCodeLayerBlock(&it);
935 #endif
936 StoreSideInfo(&it);
937 VP8StoreFilterStats(&it);
938 VP8IteratorExport(&it);
939 ok = VP8IteratorProgress(&it, 20);
940 VP8IteratorSaveBoundary(&it);
941 } while (ok && VP8IteratorNext(&it));
943 return PostLoopFinalize(&it, ok);
946 //------------------------------------------------------------------------------
947 // Single pass using Token Buffer.
949 #if !defined(DISABLE_TOKEN_BUFFER)
951 #define MIN_COUNT 96 // minimum number of macroblocks before updating stats
953 int VP8EncTokenLoop(VP8Encoder* const enc) {
954 // Roughly refresh the proba eight times per pass
955 int max_count = (enc->mb_w_ * enc->mb_h_) >> 3;
956 int num_pass_left = enc->config_->pass;
957 const int do_search = enc->do_search_;
958 VP8EncIterator it;
959 VP8Proba* const proba = &enc->proba_;
960 const VP8RDLevel rd_opt = enc->rd_opt_level_;
961 const uint64_t pixel_count = enc->mb_w_ * enc->mb_h_ * 384;
962 PassStats stats;
963 int ok;
965 InitPassStats(enc, &stats);
966 ok = PreLoopInitialize(enc);
967 if (!ok) return 0;
969 if (max_count < MIN_COUNT) max_count = MIN_COUNT;
971 assert(enc->num_parts_ == 1);
972 assert(enc->use_tokens_);
973 assert(proba->use_skip_proba_ == 0);
974 assert(rd_opt >= RD_OPT_BASIC); // otherwise, token-buffer won't be useful
975 assert(num_pass_left > 0);
977 while (ok && num_pass_left-- > 0) {
978 const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
979 (num_pass_left == 0) ||
980 (enc->max_i4_header_bits_ == 0);
981 uint64_t size_p0 = 0;
982 uint64_t distortion = 0;
983 int cnt = max_count;
984 VP8IteratorInit(enc, &it);
985 SetLoopParams(enc, stats.q);
986 if (is_last_pass) {
987 ResetTokenStats(enc);
988 VP8InitFilter(&it); // don't collect stats until last pass (too costly)
990 VP8TBufferClear(&enc->tokens_);
991 do {
992 VP8ModeScore info;
993 VP8IteratorImport(&it, NULL);
994 if (--cnt < 0) {
995 FinalizeTokenProbas(proba);
996 VP8CalculateLevelCosts(proba); // refresh cost tables for rd-opt
997 cnt = max_count;
999 VP8Decimate(&it, &info, rd_opt);
1000 RecordTokens(&it, &info, &enc->tokens_);
1001 size_p0 += info.H;
1002 distortion += info.D;
1003 #ifdef WEBP_EXPERIMENTAL_FEATURES
1004 if (enc->use_layer_) {
1005 VP8EncCodeLayerBlock(&it);
1007 #endif
1008 if (is_last_pass) {
1009 StoreSideInfo(&it);
1010 VP8StoreFilterStats(&it);
1011 VP8IteratorExport(&it);
1012 ok = VP8IteratorProgress(&it, 20);
1014 VP8IteratorSaveBoundary(&it);
1015 } while (ok && VP8IteratorNext(&it));
1016 if (!ok) break;
1018 size_p0 += enc->segment_hdr_.size_;
1019 if (stats.do_size_search) {
1020 uint64_t size = FinalizeTokenProbas(&enc->proba_);
1021 size += VP8EstimateTokenSize(&enc->tokens_,
1022 (const uint8_t*)proba->coeffs_);
1023 size = (size + size_p0 + 1024) >> 11; // -> size in bytes
1024 size += HEADER_SIZE_ESTIMATE;
1025 stats.value = (double)size;
1026 } else { // compute and store PSNR
1027 stats.value = GetPSNR(distortion, pixel_count);
1030 #if (DEBUG_SEARCH > 0)
1031 printf("#%2d metric:%.1lf -> %.1lf last_q=%.2lf q=%.2lf dq=%.2lf\n",
1032 num_pass_left, stats.last_value, stats.value,
1033 stats.last_q, stats.q, stats.dq);
1034 #endif
1035 if (size_p0 > PARTITION0_SIZE_LIMIT) {
1036 ++num_pass_left;
1037 enc->max_i4_header_bits_ >>= 1; // strengthen header bit limitation...
1038 continue; // ...and start over
1040 if (is_last_pass) {
1041 break; // done
1043 if (do_search) {
1044 ComputeNextQ(&stats); // Adjust q
1047 if (ok) {
1048 if (!stats.do_size_search) {
1049 FinalizeTokenProbas(&enc->proba_);
1051 ok = VP8EmitTokens(&enc->tokens_, enc->parts_ + 0,
1052 (const uint8_t*)proba->coeffs_, 1);
1054 ok = ok && WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
1055 return PostLoopFinalize(&it, ok);
1058 #else
1060 int VP8EncTokenLoop(VP8Encoder* const enc) {
1061 (void)enc;
1062 return 0; // we shouldn't be here.
1065 #endif // DISABLE_TOKEN_BUFFER
1067 //------------------------------------------------------------------------------