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
11 //
12 // Author: Skal (pascal.massimino@gmail.com)
13
14 #include <string.h>
15 #include <math.h>
16
17 #include "./cost.h"
18 #include "./vp8enci.h"
19 #include "../dsp/dsp.h"
20 #include "../webp/format_constants.h" // RIFF constants
21
22 #define SEGMENT_VISU 0
23 #define DEBUG_SEARCH 0 // useful to track search convergence
24
25 //------------------------------------------------------------------------------
26 // multi-pass convergence
27
28 #define HEADER_SIZE_ESTIMATE (RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE + \
29 VP8_FRAME_HEADER_SIZE)
30 #define DQ_LIMIT 0.4 // convergence is considered reached if dq < DQ_LIMIT
31 // we allow 2k of extra head-room in PARTITION0 limit.
32 #define PARTITION0_SIZE_LIMIT ((VP8_MAX_PARTITION0_SIZE - 2048ULL) << 11)
33
34 typedef struct { // struct for organizing convergence in either size or PSNR
35 int is_first;
36 float dq;
37 float q, last_q;
38 double value, last_value; // PSNR or size
39 double target;
40 int do_size_search;
41 } PassStats;
42
InitPassStats(const VP8Encoder * const enc,PassStats * const s)43 static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) {
44 const uint64_t target_size = (uint64_t)enc->config_->target_size;
45 const int do_size_search = (target_size != 0);
46 const float target_PSNR = enc->config_->target_PSNR;
47
48 s->is_first = 1;
49 s->dq = 10.f;
50 s->q = s->last_q = enc->config_->quality;
51 s->target = do_size_search ? (double)target_size
52 : (target_PSNR > 0.) ? target_PSNR
53 : 40.; // default, just in case
54 s->value = s->last_value = 0.;
55 s->do_size_search = do_size_search;
56 return do_size_search;
57 }
58
Clamp(float v,float min,float max)59 static float Clamp(float v, float min, float max) {
60 return (v < min) ? min : (v > max) ? max : v;
61 }
62
ComputeNextQ(PassStats * const s)63 static float ComputeNextQ(PassStats* const s) {
64 float dq;
65 if (s->is_first) {
66 dq = (s->value > s->target) ? -s->dq : s->dq;
67 s->is_first = 0;
68 } else if (s->value != s->last_value) {
69 const double slope = (s->target - s->value) / (s->last_value - s->value);
70 dq = (float)(slope * (s->last_q - s->q));
71 } else {
72 dq = 0.; // we're done?!
73 }
74 // Limit variable to avoid large swings.
75 s->dq = Clamp(dq, -30.f, 30.f);
76 s->last_q = s->q;
77 s->last_value = s->value;
78 s->q = Clamp(s->q + s->dq, 0.f, 100.f);
79 return s->q;
80 }
81
82 //------------------------------------------------------------------------------
83 // Tables for level coding
84
85 const uint8_t VP8Cat3[] = { 173, 148, 140 };
86 const uint8_t VP8Cat4[] = { 176, 155, 140, 135 };
87 const uint8_t VP8Cat5[] = { 180, 157, 141, 134, 130 };
88 const uint8_t VP8Cat6[] =
89 { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
90
91 //------------------------------------------------------------------------------
92 // Reset the statistics about: number of skips, token proba, level cost,...
93
ResetStats(VP8Encoder * const enc)94 static void ResetStats(VP8Encoder* const enc) {
95 VP8EncProba* const proba = &enc->proba_;
96 VP8CalculateLevelCosts(proba);
97 proba->nb_skip_ = 0;
98 }
99
100 //------------------------------------------------------------------------------
101 // Skip decision probability
102
103 #define SKIP_PROBA_THRESHOLD 250 // value below which using skip_proba is OK.
104
CalcSkipProba(uint64_t nb,uint64_t total)105 static int CalcSkipProba(uint64_t nb, uint64_t total) {
106 return (int)(total ? (total - nb) * 255 / total : 255);
107 }
108
109 // Returns the bit-cost for coding the skip probability.
FinalizeSkipProba(VP8Encoder * const enc)110 static int FinalizeSkipProba(VP8Encoder* const enc) {
111 VP8EncProba* const proba = &enc->proba_;
112 const int nb_mbs = enc->mb_w_ * enc->mb_h_;
113 const int nb_events = proba->nb_skip_;
114 int size;
115 proba->skip_proba_ = CalcSkipProba(nb_events, nb_mbs);
116 proba->use_skip_proba_ = (proba->skip_proba_ < SKIP_PROBA_THRESHOLD);
117 size = 256; // 'use_skip_proba' bit
118 if (proba->use_skip_proba_) {
119 size += nb_events * VP8BitCost(1, proba->skip_proba_)
120 + (nb_mbs - nb_events) * VP8BitCost(0, proba->skip_proba_);
121 size += 8 * 256; // cost of signaling the skip_proba_ itself.
122 }
123 return size;
124 }
125
126 // Collect statistics and deduce probabilities for next coding pass.
127 // Return the total bit-cost for coding the probability updates.
CalcTokenProba(int nb,int total)128 static int CalcTokenProba(int nb, int total) {
129 assert(nb <= total);
130 return nb ? (255 - nb * 255 / total) : 255;
131 }
132
133 // Cost of coding 'nb' 1's and 'total-nb' 0's using 'proba' probability.
BranchCost(int nb,int total,int proba)134 static int BranchCost(int nb, int total, int proba) {
135 return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
136 }
137
ResetTokenStats(VP8Encoder * const enc)138 static void ResetTokenStats(VP8Encoder* const enc) {
139 VP8EncProba* const proba = &enc->proba_;
140 memset(proba->stats_, 0, sizeof(proba->stats_));
141 }
142
FinalizeTokenProbas(VP8EncProba * const proba)143 static int FinalizeTokenProbas(VP8EncProba* const proba) {
144 int has_changed = 0;
145 int size = 0;
146 int t, b, c, p;
147 for (t = 0; t < NUM_TYPES; ++t) {
148 for (b = 0; b < NUM_BANDS; ++b) {
149 for (c = 0; c < NUM_CTX; ++c) {
150 for (p = 0; p < NUM_PROBAS; ++p) {
151 const proba_t stats = proba->stats_[t][b][c][p];
152 const int nb = (stats >> 0) & 0xffff;
153 const int total = (stats >> 16) & 0xffff;
154 const int update_proba = VP8CoeffsUpdateProba[t][b][c][p];
155 const int old_p = VP8CoeffsProba0[t][b][c][p];
156 const int new_p = CalcTokenProba(nb, total);
157 const int old_cost = BranchCost(nb, total, old_p)
158 + VP8BitCost(0, update_proba);
159 const int new_cost = BranchCost(nb, total, new_p)
160 + VP8BitCost(1, update_proba)
161 + 8 * 256;
162 const int use_new_p = (old_cost > new_cost);
163 size += VP8BitCost(use_new_p, update_proba);
164 if (use_new_p) { // only use proba that seem meaningful enough.
165 proba->coeffs_[t][b][c][p] = new_p;
166 has_changed |= (new_p != old_p);
167 size += 8 * 256;
168 } else {
169 proba->coeffs_[t][b][c][p] = old_p;
170 }
171 }
172 }
173 }
174 }
175 proba->dirty_ = has_changed;
176 return size;
177 }
178
179 //------------------------------------------------------------------------------
180 // Finalize Segment probability based on the coding tree
181
GetProba(int a,int b)182 static int GetProba(int a, int b) {
183 const int total = a + b;
184 return (total == 0) ? 255 // that's the default probability.
185 : (255 * a + total / 2) / total; // rounded proba
186 }
187
SetSegmentProbas(VP8Encoder * const enc)188 static void SetSegmentProbas(VP8Encoder* const enc) {
189 int p[NUM_MB_SEGMENTS] = { 0 };
190 int n;
191
192 for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
193 const VP8MBInfo* const mb = &enc->mb_info_[n];
194 p[mb->segment_]++;
195 }
196 if (enc->pic_->stats != NULL) {
197 for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
198 enc->pic_->stats->segment_size[n] = p[n];
199 }
200 }
201 if (enc->segment_hdr_.num_segments_ > 1) {
202 uint8_t* const probas = enc->proba_.segments_;
203 probas[0] = GetProba(p[0] + p[1], p[2] + p[3]);
204 probas[1] = GetProba(p[0], p[1]);
205 probas[2] = GetProba(p[2], p[3]);
206
207 enc->segment_hdr_.update_map_ =
208 (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
209 enc->segment_hdr_.size_ =
210 p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
211 p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
212 p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) +
213 p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2]));
214 } else {
215 enc->segment_hdr_.update_map_ = 0;
216 enc->segment_hdr_.size_ = 0;
217 }
218 }
219
220 //------------------------------------------------------------------------------
221 // Coefficient coding
222
PutCoeffs(VP8BitWriter * const bw,int ctx,const VP8Residual * res)223 static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
224 int n = res->first;
225 // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
226 const uint8_t* p = res->prob[n][ctx];
227 if (!VP8PutBit(bw, res->last >= 0, p[0])) {
228 return 0;
229 }
230
231 while (n < 16) {
232 const int c = res->coeffs[n++];
233 const int sign = c < 0;
234 int v = sign ? -c : c;
235 if (!VP8PutBit(bw, v != 0, p[1])) {
236 p = res->prob[VP8EncBands[n]][0];
237 continue;
238 }
239 if (!VP8PutBit(bw, v > 1, p[2])) {
240 p = res->prob[VP8EncBands[n]][1];
241 } else {
242 if (!VP8PutBit(bw, v > 4, p[3])) {
243 if (VP8PutBit(bw, v != 2, p[4]))
244 VP8PutBit(bw, v == 4, p[5]);
245 } else if (!VP8PutBit(bw, v > 10, p[6])) {
246 if (!VP8PutBit(bw, v > 6, p[7])) {
247 VP8PutBit(bw, v == 6, 159);
248 } else {
249 VP8PutBit(bw, v >= 9, 165);
250 VP8PutBit(bw, !(v & 1), 145);
251 }
252 } else {
253 int mask;
254 const uint8_t* tab;
255 if (v < 3 + (8 << 1)) { // VP8Cat3 (3b)
256 VP8PutBit(bw, 0, p[8]);
257 VP8PutBit(bw, 0, p[9]);
258 v -= 3 + (8 << 0);
259 mask = 1 << 2;
260 tab = VP8Cat3;
261 } else if (v < 3 + (8 << 2)) { // VP8Cat4 (4b)
262 VP8PutBit(bw, 0, p[8]);
263 VP8PutBit(bw, 1, p[9]);
264 v -= 3 + (8 << 1);
265 mask = 1 << 3;
266 tab = VP8Cat4;
267 } else if (v < 3 + (8 << 3)) { // VP8Cat5 (5b)
268 VP8PutBit(bw, 1, p[8]);
269 VP8PutBit(bw, 0, p[10]);
270 v -= 3 + (8 << 2);
271 mask = 1 << 4;
272 tab = VP8Cat5;
273 } else { // VP8Cat6 (11b)
274 VP8PutBit(bw, 1, p[8]);
275 VP8PutBit(bw, 1, p[10]);
276 v -= 3 + (8 << 3);
277 mask = 1 << 10;
278 tab = VP8Cat6;
279 }
280 while (mask) {
281 VP8PutBit(bw, !!(v & mask), *tab++);
282 mask >>= 1;
283 }
284 }
285 p = res->prob[VP8EncBands[n]][2];
286 }
287 VP8PutBitUniform(bw, sign);
288 if (n == 16 || !VP8PutBit(bw, n <= res->last, p[0])) {
289 return 1; // EOB
290 }
291 }
292 return 1;
293 }
294
CodeResiduals(VP8BitWriter * const bw,VP8EncIterator * const it,const VP8ModeScore * const rd)295 static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
296 const VP8ModeScore* const rd) {
297 int x, y, ch;
298 VP8Residual res;
299 uint64_t pos1, pos2, pos3;
300 const int i16 = (it->mb_->type_ == 1);
301 const int segment = it->mb_->segment_;
302 VP8Encoder* const enc = it->enc_;
303
304 VP8IteratorNzToBytes(it);
305
306 pos1 = VP8BitWriterPos(bw);
307 if (i16) {
308 VP8InitResidual(0, 1, enc, &res);
309 VP8SetResidualCoeffs(rd->y_dc_levels, &res);
310 it->top_nz_[8] = it->left_nz_[8] =
311 PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res);
312 VP8InitResidual(1, 0, enc, &res);
313 } else {
314 VP8InitResidual(0, 3, enc, &res);
315 }
316
317 // luma-AC
318 for (y = 0; y < 4; ++y) {
319 for (x = 0; x < 4; ++x) {
320 const int ctx = it->top_nz_[x] + it->left_nz_[y];
321 VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
322 it->top_nz_[x] = it->left_nz_[y] = PutCoeffs(bw, ctx, &res);
323 }
324 }
325 pos2 = VP8BitWriterPos(bw);
326
327 // U/V
328 VP8InitResidual(0, 2, enc, &res);
329 for (ch = 0; ch <= 2; ch += 2) {
330 for (y = 0; y < 2; ++y) {
331 for (x = 0; x < 2; ++x) {
332 const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
333 VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
334 it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
335 PutCoeffs(bw, ctx, &res);
336 }
337 }
338 }
339 pos3 = VP8BitWriterPos(bw);
340 it->luma_bits_ = pos2 - pos1;
341 it->uv_bits_ = pos3 - pos2;
342 it->bit_count_[segment][i16] += it->luma_bits_;
343 it->bit_count_[segment][2] += it->uv_bits_;
344 VP8IteratorBytesToNz(it);
345 }
346
347 // Same as CodeResiduals, but doesn't actually write anything.
348 // Instead, it just records the event distribution.
RecordResiduals(VP8EncIterator * const it,const VP8ModeScore * const rd)349 static void RecordResiduals(VP8EncIterator* const it,
350 const VP8ModeScore* const rd) {
351 int x, y, ch;
352 VP8Residual res;
353 VP8Encoder* const enc = it->enc_;
354
355 VP8IteratorNzToBytes(it);
356
357 if (it->mb_->type_ == 1) { // i16x16
358 VP8InitResidual(0, 1, enc, &res);
359 VP8SetResidualCoeffs(rd->y_dc_levels, &res);
360 it->top_nz_[8] = it->left_nz_[8] =
361 VP8RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res);
362 VP8InitResidual(1, 0, enc, &res);
363 } else {
364 VP8InitResidual(0, 3, enc, &res);
365 }
366
367 // luma-AC
368 for (y = 0; y < 4; ++y) {
369 for (x = 0; x < 4; ++x) {
370 const int ctx = it->top_nz_[x] + it->left_nz_[y];
371 VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
372 it->top_nz_[x] = it->left_nz_[y] = VP8RecordCoeffs(ctx, &res);
373 }
374 }
375
376 // U/V
377 VP8InitResidual(0, 2, enc, &res);
378 for (ch = 0; ch <= 2; ch += 2) {
379 for (y = 0; y < 2; ++y) {
380 for (x = 0; x < 2; ++x) {
381 const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
382 VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
383 it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
384 VP8RecordCoeffs(ctx, &res);
385 }
386 }
387 }
388
389 VP8IteratorBytesToNz(it);
390 }
391
392 //------------------------------------------------------------------------------
393 // Token buffer
394
395 #if !defined(DISABLE_TOKEN_BUFFER)
396
RecordTokens(VP8EncIterator * const it,const VP8ModeScore * const rd,VP8TBuffer * const tokens)397 static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
398 VP8TBuffer* const tokens) {
399 int x, y, ch;
400 VP8Residual res;
401 VP8Encoder* const enc = it->enc_;
402
403 VP8IteratorNzToBytes(it);
404 if (it->mb_->type_ == 1) { // i16x16
405 const int ctx = it->top_nz_[8] + it->left_nz_[8];
406 VP8InitResidual(0, 1, enc, &res);
407 VP8SetResidualCoeffs(rd->y_dc_levels, &res);
408 it->top_nz_[8] = it->left_nz_[8] =
409 VP8RecordCoeffTokens(ctx, 1,
410 res.first, res.last, res.coeffs, tokens);
411 VP8RecordCoeffs(ctx, &res);
412 VP8InitResidual(1, 0, enc, &res);
413 } else {
414 VP8InitResidual(0, 3, enc, &res);
415 }
416
417 // luma-AC
418 for (y = 0; y < 4; ++y) {
419 for (x = 0; x < 4; ++x) {
420 const int ctx = it->top_nz_[x] + it->left_nz_[y];
421 VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
422 it->top_nz_[x] = it->left_nz_[y] =
423 VP8RecordCoeffTokens(ctx, res.coeff_type,
424 res.first, res.last, res.coeffs, tokens);
425 VP8RecordCoeffs(ctx, &res);
426 }
427 }
428
429 // U/V
430 VP8InitResidual(0, 2, enc, &res);
431 for (ch = 0; ch <= 2; ch += 2) {
432 for (y = 0; y < 2; ++y) {
433 for (x = 0; x < 2; ++x) {
434 const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
435 VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
436 it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
437 VP8RecordCoeffTokens(ctx, 2,
438 res.first, res.last, res.coeffs, tokens);
439 VP8RecordCoeffs(ctx, &res);
440 }
441 }
442 }
443 VP8IteratorBytesToNz(it);
444 return !tokens->error_;
445 }
446
447 #endif // !DISABLE_TOKEN_BUFFER
448
449 //------------------------------------------------------------------------------
450 // ExtraInfo map / Debug function
451
452 #if SEGMENT_VISU
SetBlock(uint8_t * p,int value,int size)453 static void SetBlock(uint8_t* p, int value, int size) {
454 int y;
455 for (y = 0; y < size; ++y) {
456 memset(p, value, size);
457 p += BPS;
458 }
459 }
460 #endif
461
ResetSSE(VP8Encoder * const enc)462 static void ResetSSE(VP8Encoder* const enc) {
463 enc->sse_[0] = 0;
464 enc->sse_[1] = 0;
465 enc->sse_[2] = 0;
466 // Note: enc->sse_[3] is managed by alpha.c
467 enc->sse_count_ = 0;
468 }
469
StoreSSE(const VP8EncIterator * const it)470 static void StoreSSE(const VP8EncIterator* const it) {
471 VP8Encoder* const enc = it->enc_;
472 const uint8_t* const in = it->yuv_in_;
473 const uint8_t* const out = it->yuv_out_;
474 // Note: not totally accurate at boundary. And doesn't include in-loop filter.
475 enc->sse_[0] += VP8SSE16x16(in + Y_OFF_ENC, out + Y_OFF_ENC);
476 enc->sse_[1] += VP8SSE8x8(in + U_OFF_ENC, out + U_OFF_ENC);
477 enc->sse_[2] += VP8SSE8x8(in + V_OFF_ENC, out + V_OFF_ENC);
478 enc->sse_count_ += 16 * 16;
479 }
480
StoreSideInfo(const VP8EncIterator * const it)481 static void StoreSideInfo(const VP8EncIterator* const it) {
482 VP8Encoder* const enc = it->enc_;
483 const VP8MBInfo* const mb = it->mb_;
484 WebPPicture* const pic = enc->pic_;
485
486 if (pic->stats != NULL) {
487 StoreSSE(it);
488 enc->block_count_[0] += (mb->type_ == 0);
489 enc->block_count_[1] += (mb->type_ == 1);
490 enc->block_count_[2] += (mb->skip_ != 0);
491 }
492
493 if (pic->extra_info != NULL) {
494 uint8_t* const info = &pic->extra_info[it->x_ + it->y_ * enc->mb_w_];
495 switch (pic->extra_info_type) {
496 case 1: *info = mb->type_; break;
497 case 2: *info = mb->segment_; break;
498 case 3: *info = enc->dqm_[mb->segment_].quant_; break;
499 case 4: *info = (mb->type_ == 1) ? it->preds_[0] : 0xff; break;
500 case 5: *info = mb->uv_mode_; break;
501 case 6: {
502 const int b = (int)((it->luma_bits_ + it->uv_bits_ + 7) >> 3);
503 *info = (b > 255) ? 255 : b; break;
504 }
505 case 7: *info = mb->alpha_; break;
506 default: *info = 0; break;
507 }
508 }
509 #if SEGMENT_VISU // visualize segments and prediction modes
510 SetBlock(it->yuv_out_ + Y_OFF_ENC, mb->segment_ * 64, 16);
511 SetBlock(it->yuv_out_ + U_OFF_ENC, it->preds_[0] * 64, 8);
512 SetBlock(it->yuv_out_ + V_OFF_ENC, mb->uv_mode_ * 64, 8);
513 #endif
514 }
515
GetPSNR(uint64_t mse,uint64_t size)516 static double GetPSNR(uint64_t mse, uint64_t size) {
517 return (mse > 0 && size > 0) ? 10. * log10(255. * 255. * size / mse) : 99;
518 }
519
520 //------------------------------------------------------------------------------
521 // StatLoop(): only collect statistics (number of skips, token usage, ...).
522 // This is used for deciding optimal probabilities. It also modifies the
523 // quantizer value if some target (size, PSNR) was specified.
524
SetLoopParams(VP8Encoder * const enc,float q)525 static void SetLoopParams(VP8Encoder* const enc, float q) {
526 // Make sure the quality parameter is inside valid bounds
527 q = Clamp(q, 0.f, 100.f);
528
529 VP8SetSegmentParams(enc, q); // setup segment quantizations and filters
530 SetSegmentProbas(enc); // compute segment probabilities
531
532 ResetStats(enc);
533 ResetSSE(enc);
534 }
535
OneStatPass(VP8Encoder * const enc,VP8RDLevel rd_opt,int nb_mbs,int percent_delta,PassStats * const s)536 static uint64_t OneStatPass(VP8Encoder* const enc, VP8RDLevel rd_opt,
537 int nb_mbs, int percent_delta,
538 PassStats* const s) {
539 VP8EncIterator it;
540 uint64_t size = 0;
541 uint64_t size_p0 = 0;
542 uint64_t distortion = 0;
543 const uint64_t pixel_count = nb_mbs * 384;
544
545 VP8IteratorInit(enc, &it);
546 SetLoopParams(enc, s->q);
547 do {
548 VP8ModeScore info;
549 VP8IteratorImport(&it, NULL);
550 if (VP8Decimate(&it, &info, rd_opt)) {
551 // Just record the number of skips and act like skip_proba is not used.
552 enc->proba_.nb_skip_++;
553 }
554 RecordResiduals(&it, &info);
555 size += info.R + info.H;
556 size_p0 += info.H;
557 distortion += info.D;
558 if (percent_delta && !VP8IteratorProgress(&it, percent_delta))
559 return 0;
560 VP8IteratorSaveBoundary(&it);
561 } while (VP8IteratorNext(&it) && --nb_mbs > 0);
562
563 size_p0 += enc->segment_hdr_.size_;
564 if (s->do_size_search) {
565 size += FinalizeSkipProba(enc);
566 size += FinalizeTokenProbas(&enc->proba_);
567 size = ((size + size_p0 + 1024) >> 11) + HEADER_SIZE_ESTIMATE;
568 s->value = (double)size;
569 } else {
570 s->value = GetPSNR(distortion, pixel_count);
571 }
572 return size_p0;
573 }
574
StatLoop(VP8Encoder * const enc)575 static int StatLoop(VP8Encoder* const enc) {
576 const int method = enc->method_;
577 const int do_search = enc->do_search_;
578 const int fast_probe = ((method == 0 || method == 3) && !do_search);
579 int num_pass_left = enc->config_->pass;
580 const int task_percent = 20;
581 const int percent_per_pass =
582 (task_percent + num_pass_left / 2) / num_pass_left;
583 const int final_percent = enc->percent_ + task_percent;
584 const VP8RDLevel rd_opt =
585 (method >= 3 || do_search) ? RD_OPT_BASIC : RD_OPT_NONE;
586 int nb_mbs = enc->mb_w_ * enc->mb_h_;
587 PassStats stats;
588
589 InitPassStats(enc, &stats);
590 ResetTokenStats(enc);
591
592 // Fast mode: quick analysis pass over few mbs. Better than nothing.
593 if (fast_probe) {
594 if (method == 3) { // we need more stats for method 3 to be reliable.
595 nb_mbs = (nb_mbs > 200) ? nb_mbs >> 1 : 100;
596 } else {
597 nb_mbs = (nb_mbs > 200) ? nb_mbs >> 2 : 50;
598 }
599 }
600
601 while (num_pass_left-- > 0) {
602 const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
603 (num_pass_left == 0) ||
604 (enc->max_i4_header_bits_ == 0);
605 const uint64_t size_p0 =
606 OneStatPass(enc, rd_opt, nb_mbs, percent_per_pass, &stats);
607 if (size_p0 == 0) return 0;
608 #if (DEBUG_SEARCH > 0)
609 printf("#%d value:%.1lf -> %.1lf q:%.2f -> %.2f\n",
610 num_pass_left, stats.last_value, stats.value, stats.last_q, stats.q);
611 #endif
612 if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
613 ++num_pass_left;
614 enc->max_i4_header_bits_ >>= 1; // strengthen header bit limitation...
615 continue; // ...and start over
616 }
617 if (is_last_pass) {
618 break;
619 }
620 // If no target size: just do several pass without changing 'q'
621 if (do_search) {
622 ComputeNextQ(&stats);
623 if (fabs(stats.dq) <= DQ_LIMIT) break;
624 }
625 }
626 if (!do_search || !stats.do_size_search) {
627 // Need to finalize probas now, since it wasn't done during the search.
628 FinalizeSkipProba(enc);
629 FinalizeTokenProbas(&enc->proba_);
630 }
631 VP8CalculateLevelCosts(&enc->proba_); // finalize costs
632 return WebPReportProgress(enc->pic_, final_percent, &enc->percent_);
633 }
634
635 //------------------------------------------------------------------------------
636 // Main loops
637 //
638
639 static const int kAverageBytesPerMB[8] = { 50, 24, 16, 9, 7, 5, 3, 2 };
640
PreLoopInitialize(VP8Encoder * const enc)641 static int PreLoopInitialize(VP8Encoder* const enc) {
642 int p;
643 int ok = 1;
644 const int average_bytes_per_MB = kAverageBytesPerMB[enc->base_quant_ >> 4];
645 const int bytes_per_parts =
646 enc->mb_w_ * enc->mb_h_ * average_bytes_per_MB / enc->num_parts_;
647 // Initialize the bit-writers
648 for (p = 0; ok && p < enc->num_parts_; ++p) {
649 ok = VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
650 }
651 if (!ok) {
652 VP8EncFreeBitWriters(enc); // malloc error occurred
653 WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
654 }
655 return ok;
656 }
657
PostLoopFinalize(VP8EncIterator * const it,int ok)658 static int PostLoopFinalize(VP8EncIterator* const it, int ok) {
659 VP8Encoder* const enc = it->enc_;
660 if (ok) { // Finalize the partitions, check for extra errors.
661 int p;
662 for (p = 0; p < enc->num_parts_; ++p) {
663 VP8BitWriterFinish(enc->parts_ + p);
664 ok &= !enc->parts_[p].error_;
665 }
666 }
667
668 if (ok) { // All good. Finish up.
669 if (enc->pic_->stats != NULL) { // finalize byte counters...
670 int i, s;
671 for (i = 0; i <= 2; ++i) {
672 for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
673 enc->residual_bytes_[i][s] = (int)((it->bit_count_[s][i] + 7) >> 3);
674 }
675 }
676 }
677 VP8AdjustFilterStrength(it); // ...and store filter stats.
678 } else {
679 // Something bad happened -> need to do some memory cleanup.
680 VP8EncFreeBitWriters(enc);
681 }
682 return ok;
683 }
684
685 //------------------------------------------------------------------------------
686 // VP8EncLoop(): does the final bitstream coding.
687
ResetAfterSkip(VP8EncIterator * const it)688 static void ResetAfterSkip(VP8EncIterator* const it) {
689 if (it->mb_->type_ == 1) {
690 *it->nz_ = 0; // reset all predictors
691 it->left_nz_[8] = 0;
692 } else {
693 *it->nz_ &= (1 << 24); // preserve the dc_nz bit
694 }
695 }
696
VP8EncLoop(VP8Encoder * const enc)697 int VP8EncLoop(VP8Encoder* const enc) {
698 VP8EncIterator it;
699 int ok = PreLoopInitialize(enc);
700 if (!ok) return 0;
701
702 StatLoop(enc); // stats-collection loop
703
704 VP8IteratorInit(enc, &it);
705 VP8InitFilter(&it);
706 do {
707 VP8ModeScore info;
708 const int dont_use_skip = !enc->proba_.use_skip_proba_;
709 const VP8RDLevel rd_opt = enc->rd_opt_level_;
710
711 VP8IteratorImport(&it, NULL);
712 // Warning! order is important: first call VP8Decimate() and
713 // *then* decide how to code the skip decision if there's one.
714 if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) {
715 CodeResiduals(it.bw_, &it, &info);
716 } else { // reset predictors after a skip
717 ResetAfterSkip(&it);
718 }
719 StoreSideInfo(&it);
720 VP8StoreFilterStats(&it);
721 VP8IteratorExport(&it);
722 ok = VP8IteratorProgress(&it, 20);
723 VP8IteratorSaveBoundary(&it);
724 } while (ok && VP8IteratorNext(&it));
725
726 return PostLoopFinalize(&it, ok);
727 }
728
729 //------------------------------------------------------------------------------
730 // Single pass using Token Buffer.
731
732 #if !defined(DISABLE_TOKEN_BUFFER)
733
734 #define MIN_COUNT 96 // minimum number of macroblocks before updating stats
735
VP8EncTokenLoop(VP8Encoder * const enc)736 int VP8EncTokenLoop(VP8Encoder* const enc) {
737 // Roughly refresh the proba eight times per pass
738 int max_count = (enc->mb_w_ * enc->mb_h_) >> 3;
739 int num_pass_left = enc->config_->pass;
740 const int do_search = enc->do_search_;
741 VP8EncIterator it;
742 VP8EncProba* const proba = &enc->proba_;
743 const VP8RDLevel rd_opt = enc->rd_opt_level_;
744 const uint64_t pixel_count = enc->mb_w_ * enc->mb_h_ * 384;
745 PassStats stats;
746 int ok;
747
748 InitPassStats(enc, &stats);
749 ok = PreLoopInitialize(enc);
750 if (!ok) return 0;
751
752 if (max_count < MIN_COUNT) max_count = MIN_COUNT;
753
754 assert(enc->num_parts_ == 1);
755 assert(enc->use_tokens_);
756 assert(proba->use_skip_proba_ == 0);
757 assert(rd_opt >= RD_OPT_BASIC); // otherwise, token-buffer won't be useful
758 assert(num_pass_left > 0);
759
760 while (ok && num_pass_left-- > 0) {
761 const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
762 (num_pass_left == 0) ||
763 (enc->max_i4_header_bits_ == 0);
764 uint64_t size_p0 = 0;
765 uint64_t distortion = 0;
766 int cnt = max_count;
767 VP8IteratorInit(enc, &it);
768 SetLoopParams(enc, stats.q);
769 if (is_last_pass) {
770 ResetTokenStats(enc);
771 VP8InitFilter(&it); // don't collect stats until last pass (too costly)
772 }
773 VP8TBufferClear(&enc->tokens_);
774 do {
775 VP8ModeScore info;
776 VP8IteratorImport(&it, NULL);
777 if (--cnt < 0) {
778 FinalizeTokenProbas(proba);
779 VP8CalculateLevelCosts(proba); // refresh cost tables for rd-opt
780 cnt = max_count;
781 }
782 VP8Decimate(&it, &info, rd_opt);
783 ok = RecordTokens(&it, &info, &enc->tokens_);
784 if (!ok) {
785 WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY);
786 break;
787 }
788 size_p0 += info.H;
789 distortion += info.D;
790 if (is_last_pass) {
791 StoreSideInfo(&it);
792 VP8StoreFilterStats(&it);
793 VP8IteratorExport(&it);
794 ok = VP8IteratorProgress(&it, 20);
795 }
796 VP8IteratorSaveBoundary(&it);
797 } while (ok && VP8IteratorNext(&it));
798 if (!ok) break;
799
800 size_p0 += enc->segment_hdr_.size_;
801 if (stats.do_size_search) {
802 uint64_t size = FinalizeTokenProbas(&enc->proba_);
803 size += VP8EstimateTokenSize(&enc->tokens_,
804 (const uint8_t*)proba->coeffs_);
805 size = (size + size_p0 + 1024) >> 11; // -> size in bytes
806 size += HEADER_SIZE_ESTIMATE;
807 stats.value = (double)size;
808 } else { // compute and store PSNR
809 stats.value = GetPSNR(distortion, pixel_count);
810 }
811
812 #if (DEBUG_SEARCH > 0)
813 printf("#%2d metric:%.1lf -> %.1lf last_q=%.2lf q=%.2lf dq=%.2lf\n",
814 num_pass_left, stats.last_value, stats.value,
815 stats.last_q, stats.q, stats.dq);
816 #endif
817 if (size_p0 > PARTITION0_SIZE_LIMIT) {
818 ++num_pass_left;
819 enc->max_i4_header_bits_ >>= 1; // strengthen header bit limitation...
820 continue; // ...and start over
821 }
822 if (is_last_pass) {
823 break; // done
824 }
825 if (do_search) {
826 ComputeNextQ(&stats); // Adjust q
827 }
828 }
829 if (ok) {
830 if (!stats.do_size_search) {
831 FinalizeTokenProbas(&enc->proba_);
832 }
833 ok = VP8EmitTokens(&enc->tokens_, enc->parts_ + 0,
834 (const uint8_t*)proba->coeffs_, 1);
835 }
836 ok = ok && WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
837 return PostLoopFinalize(&it, ok);
838 }
839
840 #else
841
VP8EncTokenLoop(VP8Encoder * const enc)842 int VP8EncTokenLoop(VP8Encoder* const enc) {
843 (void)enc;
844 return 0; // we shouldn't be here.
845 }
846
847 #endif // DISABLE_TOKEN_BUFFER
848
849 //------------------------------------------------------------------------------
850
851