Book a Demo!
CoCalc Logo Icon
StoreFeaturesDocsShareSupportNewsAboutPoliciesSign UpSign In
godotengine
GitHub Repository: godotengine/godot
Path: blob/master/thirdparty/libwebp/src/enc/frame_enc.c
21653 views
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 ([email protected])
13
14
#include <assert.h>
15
#include <math.h>
16
#include <string.h>
17
18
#include "src/dec/common_dec.h"
19
#include "src/webp/types.h"
20
#include "src/dsp/dsp.h"
21
#include "src/enc/cost_enc.h"
22
#include "src/enc/vp8i_enc.h"
23
#include "src/utils/bit_writer_utils.h"
24
#include "src/webp/encode.h"
25
#include "src/webp/format_constants.h" // RIFF constants
26
27
#define SEGMENT_VISU 0
28
#define DEBUG_SEARCH 0 // useful to track search convergence
29
30
//------------------------------------------------------------------------------
31
// multi-pass convergence
32
33
#define HEADER_SIZE_ESTIMATE (RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE + \
34
VP8_FRAME_HEADER_SIZE)
35
#define DQ_LIMIT 0.4 // convergence is considered reached if dq < DQ_LIMIT
36
// we allow 2k of extra head-room in PARTITION0 limit.
37
#define PARTITION0_SIZE_LIMIT ((VP8_MAX_PARTITION0_SIZE - 2048ULL) << 11)
38
39
static float Clamp(float v, float min, float max) {
40
return (v < min) ? min : (v > max) ? max : v;
41
}
42
43
typedef struct { // struct for organizing convergence in either size or PSNR
44
int is_first;
45
float dq;
46
float q, last_q;
47
float qmin, qmax;
48
double value, last_value; // PSNR or size
49
double target;
50
int do_size_search;
51
} PassStats;
52
53
static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) {
54
const uint64_t target_size = (uint64_t)enc->config->target_size;
55
const int do_size_search = (target_size != 0);
56
const float target_PSNR = enc->config->target_PSNR;
57
58
s->is_first = 1;
59
s->dq = 10.f;
60
s->qmin = 1.f * enc->config->qmin;
61
s->qmax = 1.f * enc->config->qmax;
62
s->q = s->last_q = Clamp(enc->config->quality, s->qmin, s->qmax);
63
s->target = do_size_search ? (double)target_size
64
: (target_PSNR > 0.) ? target_PSNR
65
: 40.; // default, just in case
66
s->value = s->last_value = 0.;
67
s->do_size_search = do_size_search;
68
return do_size_search;
69
}
70
71
static float ComputeNextQ(PassStats* const s) {
72
float dq;
73
if (s->is_first) {
74
dq = (s->value > s->target) ? -s->dq : s->dq;
75
s->is_first = 0;
76
} else if (s->value != s->last_value) {
77
const double slope = (s->target - s->value) / (s->last_value - s->value);
78
dq = (float)(slope * (s->last_q - s->q));
79
} else {
80
dq = 0.; // we're done?!
81
}
82
// Limit variable to avoid large swings.
83
s->dq = Clamp(dq, -30.f, 30.f);
84
s->last_q = s->q;
85
s->last_value = s->value;
86
s->q = Clamp(s->q + s->dq, s->qmin, s->qmax);
87
return s->q;
88
}
89
90
//------------------------------------------------------------------------------
91
// Tables for level coding
92
93
const uint8_t VP8Cat3[] = { 173, 148, 140 };
94
const uint8_t VP8Cat4[] = { 176, 155, 140, 135 };
95
const uint8_t VP8Cat5[] = { 180, 157, 141, 134, 130 };
96
const uint8_t VP8Cat6[] =
97
{ 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
98
99
//------------------------------------------------------------------------------
100
// Reset the statistics about: number of skips, token proba, level cost,...
101
102
static void ResetStats(VP8Encoder* const enc) {
103
VP8EncProba* const proba = &enc->proba;
104
VP8CalculateLevelCosts(proba);
105
proba->nb_skip = 0;
106
}
107
108
//------------------------------------------------------------------------------
109
// Skip decision probability
110
111
#define SKIP_PROBA_THRESHOLD 250 // value below which using skip_proba is OK.
112
113
static int CalcSkipProba(uint64_t nb, uint64_t total) {
114
return (int)(total ? (total - nb) * 255 / total : 255);
115
}
116
117
// Returns the bit-cost for coding the skip probability.
118
static int FinalizeSkipProba(VP8Encoder* const enc) {
119
VP8EncProba* const proba = &enc->proba;
120
const int nb_mbs = enc->mb_w * enc->mb_h;
121
const int nb_events = proba->nb_skip;
122
int size;
123
proba->skip_proba = CalcSkipProba(nb_events, nb_mbs);
124
proba->use_skip_proba = (proba->skip_proba < SKIP_PROBA_THRESHOLD);
125
size = 256; // 'use_skip_proba' bit
126
if (proba->use_skip_proba) {
127
size += nb_events * VP8BitCost(1, proba->skip_proba)
128
+ (nb_mbs - nb_events) * VP8BitCost(0, proba->skip_proba);
129
size += 8 * 256; // cost of signaling the 'skip_proba' itself.
130
}
131
return size;
132
}
133
134
// Collect statistics and deduce probabilities for next coding pass.
135
// Return the total bit-cost for coding the probability updates.
136
static int CalcTokenProba(int nb, int total) {
137
assert(nb <= total);
138
return nb ? (255 - nb * 255 / total) : 255;
139
}
140
141
// Cost of coding 'nb' 1's and 'total-nb' 0's using 'proba' probability.
142
static int BranchCost(int nb, int total, int proba) {
143
return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
144
}
145
146
static void ResetTokenStats(VP8Encoder* const enc) {
147
VP8EncProba* const proba = &enc->proba;
148
memset(proba->stats, 0, sizeof(proba->stats));
149
}
150
151
static int FinalizeTokenProbas(VP8EncProba* const proba) {
152
int has_changed = 0;
153
int size = 0;
154
int t, b, c, p;
155
for (t = 0; t < NUM_TYPES; ++t) {
156
for (b = 0; b < NUM_BANDS; ++b) {
157
for (c = 0; c < NUM_CTX; ++c) {
158
for (p = 0; p < NUM_PROBAS; ++p) {
159
const proba_t stats = proba->stats[t][b][c][p];
160
const int nb = (stats >> 0) & 0xffff;
161
const int total = (stats >> 16) & 0xffff;
162
const int update_proba = VP8CoeffsUpdateProba[t][b][c][p];
163
const int old_p = VP8CoeffsProba0[t][b][c][p];
164
const int new_p = CalcTokenProba(nb, total);
165
const int old_cost = BranchCost(nb, total, old_p)
166
+ VP8BitCost(0, update_proba);
167
const int new_cost = BranchCost(nb, total, new_p)
168
+ VP8BitCost(1, update_proba)
169
+ 8 * 256;
170
const int use_new_p = (old_cost > new_cost);
171
size += VP8BitCost(use_new_p, update_proba);
172
if (use_new_p) { // only use proba that seem meaningful enough.
173
proba->coeffs[t][b][c][p] = new_p;
174
has_changed |= (new_p != old_p);
175
size += 8 * 256;
176
} else {
177
proba->coeffs[t][b][c][p] = old_p;
178
}
179
}
180
}
181
}
182
}
183
proba->dirty = has_changed;
184
return size;
185
}
186
187
//------------------------------------------------------------------------------
188
// Finalize Segment probability based on the coding tree
189
190
static int GetProba(int a, int b) {
191
const int total = a + b;
192
return (total == 0) ? 255 // that's the default probability.
193
: (255 * a + total / 2) / total; // rounded proba
194
}
195
196
static void ResetSegments(VP8Encoder* const enc) {
197
int n;
198
for (n = 0; n < enc->mb_w * enc->mb_h; ++n) {
199
enc->mb_info[n].segment = 0;
200
}
201
}
202
203
static void SetSegmentProbas(VP8Encoder* const enc) {
204
int p[NUM_MB_SEGMENTS] = { 0 };
205
int n;
206
207
for (n = 0; n < enc->mb_w * enc->mb_h; ++n) {
208
const VP8MBInfo* const mb = &enc->mb_info[n];
209
++p[mb->segment];
210
}
211
#if !defined(WEBP_DISABLE_STATS)
212
if (enc->pic->stats != NULL) {
213
for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
214
enc->pic->stats->segment_size[n] = p[n];
215
}
216
}
217
#endif
218
if (enc->segment_hdr.num_segments > 1) {
219
uint8_t* const probas = enc->proba.segments;
220
probas[0] = GetProba(p[0] + p[1], p[2] + p[3]);
221
probas[1] = GetProba(p[0], p[1]);
222
probas[2] = GetProba(p[2], p[3]);
223
224
enc->segment_hdr.update_map =
225
(probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
226
if (!enc->segment_hdr.update_map) ResetSegments(enc);
227
enc->segment_hdr.size =
228
p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
229
p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
230
p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) +
231
p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2]));
232
} else {
233
enc->segment_hdr.update_map = 0;
234
enc->segment_hdr.size = 0;
235
}
236
}
237
238
//------------------------------------------------------------------------------
239
// Coefficient coding
240
241
static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
242
int n = res->first;
243
// should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
244
const uint8_t* p = res->prob[n][ctx];
245
if (!VP8PutBit(bw, res->last >= 0, p[0])) {
246
return 0;
247
}
248
249
while (n < 16) {
250
const int c = res->coeffs[n++];
251
const int sign = c < 0;
252
int v = sign ? -c : c;
253
if (!VP8PutBit(bw, v != 0, p[1])) {
254
p = res->prob[VP8EncBands[n]][0];
255
continue;
256
}
257
if (!VP8PutBit(bw, v > 1, p[2])) {
258
p = res->prob[VP8EncBands[n]][1];
259
} else {
260
if (!VP8PutBit(bw, v > 4, p[3])) {
261
if (VP8PutBit(bw, v != 2, p[4])) {
262
VP8PutBit(bw, v == 4, p[5]);
263
}
264
} else if (!VP8PutBit(bw, v > 10, p[6])) {
265
if (!VP8PutBit(bw, v > 6, p[7])) {
266
VP8PutBit(bw, v == 6, 159);
267
} else {
268
VP8PutBit(bw, v >= 9, 165);
269
VP8PutBit(bw, !(v & 1), 145);
270
}
271
} else {
272
int mask;
273
const uint8_t* tab;
274
if (v < 3 + (8 << 1)) { // VP8Cat3 (3b)
275
VP8PutBit(bw, 0, p[8]);
276
VP8PutBit(bw, 0, p[9]);
277
v -= 3 + (8 << 0);
278
mask = 1 << 2;
279
tab = VP8Cat3;
280
} else if (v < 3 + (8 << 2)) { // VP8Cat4 (4b)
281
VP8PutBit(bw, 0, p[8]);
282
VP8PutBit(bw, 1, p[9]);
283
v -= 3 + (8 << 1);
284
mask = 1 << 3;
285
tab = VP8Cat4;
286
} else if (v < 3 + (8 << 3)) { // VP8Cat5 (5b)
287
VP8PutBit(bw, 1, p[8]);
288
VP8PutBit(bw, 0, p[10]);
289
v -= 3 + (8 << 2);
290
mask = 1 << 4;
291
tab = VP8Cat5;
292
} else { // VP8Cat6 (11b)
293
VP8PutBit(bw, 1, p[8]);
294
VP8PutBit(bw, 1, p[10]);
295
v -= 3 + (8 << 3);
296
mask = 1 << 10;
297
tab = VP8Cat6;
298
}
299
while (mask) {
300
VP8PutBit(bw, !!(v & mask), *tab++);
301
mask >>= 1;
302
}
303
}
304
p = res->prob[VP8EncBands[n]][2];
305
}
306
VP8PutBitUniform(bw, sign);
307
if (n == 16 || !VP8PutBit(bw, n <= res->last, p[0])) {
308
return 1; // EOB
309
}
310
}
311
return 1;
312
}
313
314
static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
315
const VP8ModeScore* const rd) {
316
int x, y, ch;
317
VP8Residual res;
318
uint64_t pos1, pos2, pos3;
319
const int i16 = (it->mb->type == 1);
320
const int segment = it->mb->segment;
321
VP8Encoder* const enc = it->enc;
322
323
VP8IteratorNzToBytes(it);
324
325
pos1 = VP8BitWriterPos(bw);
326
if (i16) {
327
VP8InitResidual(0, 1, enc, &res);
328
VP8SetResidualCoeffs(rd->y_dc_levels, &res);
329
it->top_nz[8] = it->left_nz[8] =
330
PutCoeffs(bw, it->top_nz[8] + it->left_nz[8], &res);
331
VP8InitResidual(1, 0, enc, &res);
332
} else {
333
VP8InitResidual(0, 3, enc, &res);
334
}
335
336
// luma-AC
337
for (y = 0; y < 4; ++y) {
338
for (x = 0; x < 4; ++x) {
339
const int ctx = it->top_nz[x] + it->left_nz[y];
340
VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
341
it->top_nz[x] = it->left_nz[y] = PutCoeffs(bw, ctx, &res);
342
}
343
}
344
pos2 = VP8BitWriterPos(bw);
345
346
// U/V
347
VP8InitResidual(0, 2, enc, &res);
348
for (ch = 0; ch <= 2; ch += 2) {
349
for (y = 0; y < 2; ++y) {
350
for (x = 0; x < 2; ++x) {
351
const int ctx = it->top_nz[4 + ch + x] + it->left_nz[4 + ch + y];
352
VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
353
it->top_nz[4 + ch + x] = it->left_nz[4 + ch + y] =
354
PutCoeffs(bw, ctx, &res);
355
}
356
}
357
}
358
pos3 = VP8BitWriterPos(bw);
359
it->luma_bits = pos2 - pos1;
360
it->uv_bits = pos3 - pos2;
361
it->bit_count[segment][i16] += it->luma_bits;
362
it->bit_count[segment][2] += it->uv_bits;
363
VP8IteratorBytesToNz(it);
364
}
365
366
// Same as CodeResiduals, but doesn't actually write anything.
367
// Instead, it just records the event distribution.
368
static void RecordResiduals(VP8EncIterator* const it,
369
const VP8ModeScore* const rd) {
370
int x, y, ch;
371
VP8Residual res;
372
VP8Encoder* const enc = it->enc;
373
374
VP8IteratorNzToBytes(it);
375
376
if (it->mb->type == 1) { // i16x16
377
VP8InitResidual(0, 1, enc, &res);
378
VP8SetResidualCoeffs(rd->y_dc_levels, &res);
379
it->top_nz[8] = it->left_nz[8] =
380
VP8RecordCoeffs(it->top_nz[8] + it->left_nz[8], &res);
381
VP8InitResidual(1, 0, enc, &res);
382
} else {
383
VP8InitResidual(0, 3, enc, &res);
384
}
385
386
// luma-AC
387
for (y = 0; y < 4; ++y) {
388
for (x = 0; x < 4; ++x) {
389
const int ctx = it->top_nz[x] + it->left_nz[y];
390
VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
391
it->top_nz[x] = it->left_nz[y] = VP8RecordCoeffs(ctx, &res);
392
}
393
}
394
395
// U/V
396
VP8InitResidual(0, 2, enc, &res);
397
for (ch = 0; ch <= 2; ch += 2) {
398
for (y = 0; y < 2; ++y) {
399
for (x = 0; x < 2; ++x) {
400
const int ctx = it->top_nz[4 + ch + x] + it->left_nz[4 + ch + y];
401
VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
402
it->top_nz[4 + ch + x] = it->left_nz[4 + ch + y] =
403
VP8RecordCoeffs(ctx, &res);
404
}
405
}
406
}
407
408
VP8IteratorBytesToNz(it);
409
}
410
411
//------------------------------------------------------------------------------
412
// Token buffer
413
414
#if !defined(DISABLE_TOKEN_BUFFER)
415
416
static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
417
VP8TBuffer* const tokens) {
418
int x, y, ch;
419
VP8Residual res;
420
VP8Encoder* const enc = it->enc;
421
422
VP8IteratorNzToBytes(it);
423
if (it->mb->type == 1) { // i16x16
424
const int ctx = it->top_nz[8] + it->left_nz[8];
425
VP8InitResidual(0, 1, enc, &res);
426
VP8SetResidualCoeffs(rd->y_dc_levels, &res);
427
it->top_nz[8] = it->left_nz[8] =
428
VP8RecordCoeffTokens(ctx, &res, tokens);
429
VP8InitResidual(1, 0, enc, &res);
430
} else {
431
VP8InitResidual(0, 3, enc, &res);
432
}
433
434
// luma-AC
435
for (y = 0; y < 4; ++y) {
436
for (x = 0; x < 4; ++x) {
437
const int ctx = it->top_nz[x] + it->left_nz[y];
438
VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
439
it->top_nz[x] = it->left_nz[y] =
440
VP8RecordCoeffTokens(ctx, &res, tokens);
441
}
442
}
443
444
// U/V
445
VP8InitResidual(0, 2, enc, &res);
446
for (ch = 0; ch <= 2; ch += 2) {
447
for (y = 0; y < 2; ++y) {
448
for (x = 0; x < 2; ++x) {
449
const int ctx = it->top_nz[4 + ch + x] + it->left_nz[4 + ch + y];
450
VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
451
it->top_nz[4 + ch + x] = it->left_nz[4 + ch + y] =
452
VP8RecordCoeffTokens(ctx, &res, tokens);
453
}
454
}
455
}
456
VP8IteratorBytesToNz(it);
457
return !tokens->error;
458
}
459
460
#endif // !DISABLE_TOKEN_BUFFER
461
462
//------------------------------------------------------------------------------
463
// ExtraInfo map / Debug function
464
465
#if !defined(WEBP_DISABLE_STATS)
466
467
#if SEGMENT_VISU
468
static void SetBlock(uint8_t* p, int value, int size) {
469
int y;
470
for (y = 0; y < size; ++y) {
471
memset(p, value, size);
472
p += BPS;
473
}
474
}
475
#endif
476
477
static void ResetSSE(VP8Encoder* const enc) {
478
enc->sse[0] = 0;
479
enc->sse[1] = 0;
480
enc->sse[2] = 0;
481
// Note: enc->sse[3] is managed by alpha.c
482
enc->sse_count = 0;
483
}
484
485
static void StoreSSE(const VP8EncIterator* const it) {
486
VP8Encoder* const enc = it->enc;
487
const uint8_t* const in = it->yuv_in;
488
const uint8_t* const out = it->yuv_out;
489
// Note: not totally accurate at boundary. And doesn't include in-loop filter.
490
enc->sse[0] += VP8SSE16x16(in + Y_OFF_ENC, out + Y_OFF_ENC);
491
enc->sse[1] += VP8SSE8x8(in + U_OFF_ENC, out + U_OFF_ENC);
492
enc->sse[2] += VP8SSE8x8(in + V_OFF_ENC, out + V_OFF_ENC);
493
enc->sse_count += 16 * 16;
494
}
495
496
static void StoreSideInfo(const VP8EncIterator* const it) {
497
VP8Encoder* const enc = it->enc;
498
const VP8MBInfo* const mb = it->mb;
499
WebPPicture* const pic = enc->pic;
500
501
if (pic->stats != NULL) {
502
StoreSSE(it);
503
enc->block_count[0] += (mb->type == 0);
504
enc->block_count[1] += (mb->type == 1);
505
enc->block_count[2] += (mb->skip != 0);
506
}
507
508
if (pic->extra_info != NULL) {
509
uint8_t* const info = &pic->extra_info[it->x + it->y * enc->mb_w];
510
switch (pic->extra_info_type) {
511
case 1: *info = mb->type; break;
512
case 2: *info = mb->segment; break;
513
case 3: *info = enc->dqm[mb->segment].quant; break;
514
case 4: *info = (mb->type == 1) ? it->preds[0] : 0xff; break;
515
case 5: *info = mb->uv_mode; break;
516
case 6: {
517
const int b = (int)((it->luma_bits + it->uv_bits + 7) >> 3);
518
*info = (b > 255) ? 255 : b; break;
519
}
520
case 7: *info = mb->alpha; break;
521
default: *info = 0; break;
522
}
523
}
524
#if SEGMENT_VISU // visualize segments and prediction modes
525
SetBlock(it->yuv_out + Y_OFF_ENC, mb->segment * 64, 16);
526
SetBlock(it->yuv_out + U_OFF_ENC, it->preds[0] * 64, 8);
527
SetBlock(it->yuv_out + V_OFF_ENC, mb->uv_mode * 64, 8);
528
#endif
529
}
530
531
static void ResetSideInfo(const VP8EncIterator* const it) {
532
VP8Encoder* const enc = it->enc;
533
WebPPicture* const pic = enc->pic;
534
if (pic->stats != NULL) {
535
memset(enc->block_count, 0, sizeof(enc->block_count));
536
}
537
ResetSSE(enc);
538
}
539
#else // defined(WEBP_DISABLE_STATS)
540
static void ResetSSE(VP8Encoder* const enc) {
541
(void)enc;
542
}
543
static void StoreSideInfo(const VP8EncIterator* const it) {
544
VP8Encoder* const enc = it->enc;
545
WebPPicture* const pic = enc->pic;
546
if (pic->extra_info != NULL) {
547
if (it->x == 0 && it->y == 0) { // only do it once, at start
548
memset(pic->extra_info, 0,
549
enc->mb_w * enc->mb_h * sizeof(*pic->extra_info));
550
}
551
}
552
}
553
554
static void ResetSideInfo(const VP8EncIterator* const it) {
555
(void)it;
556
}
557
#endif // !defined(WEBP_DISABLE_STATS)
558
559
static double GetPSNR(uint64_t mse, uint64_t size) {
560
return (mse > 0 && size > 0) ? 10. * log10(255. * 255. * size / mse) : 99;
561
}
562
563
//------------------------------------------------------------------------------
564
// StatLoop(): only collect statistics (number of skips, token usage, ...).
565
// This is used for deciding optimal probabilities. It also modifies the
566
// quantizer value if some target (size, PSNR) was specified.
567
568
static void SetLoopParams(VP8Encoder* const enc, float q) {
569
// Make sure the quality parameter is inside valid bounds
570
q = Clamp(q, 0.f, 100.f);
571
572
VP8SetSegmentParams(enc, q); // setup segment quantizations and filters
573
SetSegmentProbas(enc); // compute segment probabilities
574
575
ResetStats(enc);
576
ResetSSE(enc);
577
}
578
579
static uint64_t OneStatPass(VP8Encoder* const enc, VP8RDLevel rd_opt,
580
int nb_mbs, int percent_delta,
581
PassStats* const s) {
582
VP8EncIterator it;
583
uint64_t size = 0;
584
uint64_t size_p0 = 0;
585
uint64_t distortion = 0;
586
const uint64_t pixel_count = (uint64_t)nb_mbs * 384;
587
588
VP8IteratorInit(enc, &it);
589
SetLoopParams(enc, s->q);
590
do {
591
VP8ModeScore info;
592
VP8IteratorImport(&it, NULL);
593
if (VP8Decimate(&it, &info, rd_opt)) {
594
// Just record the number of skips and act like skip_proba is not used.
595
++enc->proba.nb_skip;
596
}
597
RecordResiduals(&it, &info);
598
size += info.R + info.H;
599
size_p0 += info.H;
600
distortion += info.D;
601
if (percent_delta && !VP8IteratorProgress(&it, percent_delta)) {
602
return 0;
603
}
604
VP8IteratorSaveBoundary(&it);
605
} while (VP8IteratorNext(&it) && --nb_mbs > 0);
606
607
size_p0 += enc->segment_hdr.size;
608
if (s->do_size_search) {
609
size += FinalizeSkipProba(enc);
610
size += FinalizeTokenProbas(&enc->proba);
611
size = ((size + size_p0 + 1024) >> 11) + HEADER_SIZE_ESTIMATE;
612
s->value = (double)size;
613
} else {
614
s->value = GetPSNR(distortion, pixel_count);
615
}
616
return size_p0;
617
}
618
619
static int StatLoop(VP8Encoder* const enc) {
620
const int method = enc->method;
621
const int do_search = enc->do_search;
622
const int fast_probe = ((method == 0 || method == 3) && !do_search);
623
int num_pass_left = enc->config->pass;
624
const int task_percent = 20;
625
const int percent_per_pass =
626
(task_percent + num_pass_left / 2) / num_pass_left;
627
const int final_percent = enc->percent + task_percent;
628
const VP8RDLevel rd_opt =
629
(method >= 3 || do_search) ? RD_OPT_BASIC : RD_OPT_NONE;
630
int nb_mbs = enc->mb_w * enc->mb_h;
631
PassStats stats;
632
633
InitPassStats(enc, &stats);
634
ResetTokenStats(enc);
635
636
// Fast mode: quick analysis pass over few mbs. Better than nothing.
637
if (fast_probe) {
638
if (method == 3) { // we need more stats for method 3 to be reliable.
639
nb_mbs = (nb_mbs > 200) ? nb_mbs >> 1 : 100;
640
} else {
641
nb_mbs = (nb_mbs > 200) ? nb_mbs >> 2 : 50;
642
}
643
}
644
645
while (num_pass_left-- > 0) {
646
const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
647
(num_pass_left == 0) ||
648
(enc->max_i4_header_bits == 0);
649
const uint64_t size_p0 =
650
OneStatPass(enc, rd_opt, nb_mbs, percent_per_pass, &stats);
651
if (size_p0 == 0) return 0;
652
#if (DEBUG_SEARCH > 0)
653
printf("#%d value:%.1lf -> %.1lf q:%.2f -> %.2f\n",
654
num_pass_left, stats.last_value, stats.value, stats.last_q, stats.q);
655
#endif
656
if (enc->max_i4_header_bits > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
657
++num_pass_left;
658
enc->max_i4_header_bits >>= 1; // strengthen header bit limitation...
659
continue; // ...and start over
660
}
661
if (is_last_pass) {
662
break;
663
}
664
// If no target size: just do several pass without changing 'q'
665
if (do_search) {
666
ComputeNextQ(&stats);
667
if (fabs(stats.dq) <= DQ_LIMIT) break;
668
}
669
}
670
if (!do_search || !stats.do_size_search) {
671
// Need to finalize probas now, since it wasn't done during the search.
672
FinalizeSkipProba(enc);
673
FinalizeTokenProbas(&enc->proba);
674
}
675
VP8CalculateLevelCosts(&enc->proba); // finalize costs
676
return WebPReportProgress(enc->pic, final_percent, &enc->percent);
677
}
678
679
//------------------------------------------------------------------------------
680
// Main loops
681
//
682
683
static const uint8_t kAverageBytesPerMB[8] = { 50, 24, 16, 9, 7, 5, 3, 2 };
684
685
static int PreLoopInitialize(VP8Encoder* const enc) {
686
int p;
687
int ok = 1;
688
const int average_bytes_per_MB = kAverageBytesPerMB[enc->base_quant >> 4];
689
const int bytes_per_parts =
690
enc->mb_w * enc->mb_h * average_bytes_per_MB / enc->num_parts;
691
// Initialize the bit-writers
692
for (p = 0; ok && p < enc->num_parts; ++p) {
693
ok = VP8BitWriterInit(enc->parts + p, bytes_per_parts);
694
}
695
if (!ok) {
696
VP8EncFreeBitWriters(enc); // malloc error occurred
697
return WebPEncodingSetError(enc->pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
698
}
699
return ok;
700
}
701
702
static int PostLoopFinalize(VP8EncIterator* const it, int ok) {
703
VP8Encoder* const enc = it->enc;
704
if (ok) { // Finalize the partitions, check for extra errors.
705
int p;
706
for (p = 0; p < enc->num_parts; ++p) {
707
VP8BitWriterFinish(enc->parts + p);
708
ok &= !enc->parts[p].error;
709
}
710
}
711
712
if (ok) { // All good. Finish up.
713
#if !defined(WEBP_DISABLE_STATS)
714
if (enc->pic->stats != NULL) { // finalize byte counters...
715
int i, s;
716
for (i = 0; i <= 2; ++i) {
717
for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
718
enc->residual_bytes[i][s] = (int)((it->bit_count[s][i] + 7) >> 3);
719
}
720
}
721
}
722
#endif
723
VP8AdjustFilterStrength(it); // ...and store filter stats.
724
} else {
725
// Something bad happened -> need to do some memory cleanup.
726
VP8EncFreeBitWriters(enc);
727
return WebPEncodingSetError(enc->pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
728
}
729
return ok;
730
}
731
732
//------------------------------------------------------------------------------
733
// VP8EncLoop(): does the final bitstream coding.
734
735
static void ResetAfterSkip(VP8EncIterator* const it) {
736
if (it->mb->type == 1) {
737
*it->nz = 0; // reset all predictors
738
it->left_nz[8] = 0;
739
} else {
740
*it->nz &= (1 << 24); // preserve the dc_nz bit
741
}
742
}
743
744
int VP8EncLoop(VP8Encoder* const enc) {
745
VP8EncIterator it;
746
int ok = PreLoopInitialize(enc);
747
if (!ok) return 0;
748
749
StatLoop(enc); // stats-collection loop
750
751
VP8IteratorInit(enc, &it);
752
VP8InitFilter(&it);
753
do {
754
VP8ModeScore info;
755
const int dont_use_skip = !enc->proba.use_skip_proba;
756
const VP8RDLevel rd_opt = enc->rd_opt_level;
757
758
VP8IteratorImport(&it, NULL);
759
// Warning! order is important: first call VP8Decimate() and
760
// *then* decide how to code the skip decision if there's one.
761
if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) {
762
CodeResiduals(it.bw, &it, &info);
763
if (it.bw->error) {
764
// enc->pic->error_code is set in PostLoopFinalize().
765
ok = 0;
766
break;
767
}
768
} else { // reset predictors after a skip
769
ResetAfterSkip(&it);
770
}
771
StoreSideInfo(&it);
772
VP8StoreFilterStats(&it);
773
VP8IteratorExport(&it);
774
ok = VP8IteratorProgress(&it, 20);
775
VP8IteratorSaveBoundary(&it);
776
} while (ok && VP8IteratorNext(&it));
777
778
return PostLoopFinalize(&it, ok);
779
}
780
781
//------------------------------------------------------------------------------
782
// Single pass using Token Buffer.
783
784
#if !defined(DISABLE_TOKEN_BUFFER)
785
786
#define MIN_COUNT 96 // minimum number of macroblocks before updating stats
787
788
int VP8EncTokenLoop(VP8Encoder* const enc) {
789
// Roughly refresh the proba eight times per pass
790
int max_count = (enc->mb_w * enc->mb_h) >> 3;
791
int num_pass_left = enc->config->pass;
792
int remaining_progress = 40; // percents
793
const int do_search = enc->do_search;
794
VP8EncIterator it;
795
VP8EncProba* const proba = &enc->proba;
796
const VP8RDLevel rd_opt = enc->rd_opt_level;
797
const uint64_t pixel_count = (uint64_t)enc->mb_w * enc->mb_h * 384;
798
PassStats stats;
799
int ok;
800
801
InitPassStats(enc, &stats);
802
ok = PreLoopInitialize(enc);
803
if (!ok) return 0;
804
805
if (max_count < MIN_COUNT) max_count = MIN_COUNT;
806
807
assert(enc->num_parts == 1);
808
assert(enc->use_tokens);
809
assert(proba->use_skip_proba == 0);
810
assert(rd_opt >= RD_OPT_BASIC); // otherwise, token-buffer won't be useful
811
assert(num_pass_left > 0);
812
813
while (ok && num_pass_left-- > 0) {
814
const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) ||
815
(num_pass_left == 0) ||
816
(enc->max_i4_header_bits == 0);
817
uint64_t size_p0 = 0;
818
uint64_t distortion = 0;
819
int cnt = max_count;
820
// The final number of passes is not trivial to know in advance.
821
const int pass_progress = remaining_progress / (2 + num_pass_left);
822
remaining_progress -= pass_progress;
823
VP8IteratorInit(enc, &it);
824
SetLoopParams(enc, stats.q);
825
if (is_last_pass) {
826
ResetTokenStats(enc);
827
VP8InitFilter(&it); // don't collect stats until last pass (too costly)
828
}
829
VP8TBufferClear(&enc->tokens);
830
do {
831
VP8ModeScore info;
832
VP8IteratorImport(&it, NULL);
833
if (--cnt < 0) {
834
FinalizeTokenProbas(proba);
835
VP8CalculateLevelCosts(proba); // refresh cost tables for rd-opt
836
cnt = max_count;
837
}
838
VP8Decimate(&it, &info, rd_opt);
839
ok = RecordTokens(&it, &info, &enc->tokens);
840
if (!ok) {
841
WebPEncodingSetError(enc->pic, VP8_ENC_ERROR_OUT_OF_MEMORY);
842
break;
843
}
844
size_p0 += info.H;
845
distortion += info.D;
846
if (is_last_pass) {
847
StoreSideInfo(&it);
848
VP8StoreFilterStats(&it);
849
VP8IteratorExport(&it);
850
ok = VP8IteratorProgress(&it, pass_progress);
851
}
852
VP8IteratorSaveBoundary(&it);
853
} while (ok && VP8IteratorNext(&it));
854
if (!ok) break;
855
856
size_p0 += enc->segment_hdr.size;
857
if (stats.do_size_search) {
858
uint64_t size = FinalizeTokenProbas(&enc->proba);
859
size += VP8EstimateTokenSize(&enc->tokens,
860
(const uint8_t*)proba->coeffs);
861
size = (size + size_p0 + 1024) >> 11; // -> size in bytes
862
size += HEADER_SIZE_ESTIMATE;
863
stats.value = (double)size;
864
} else { // compute and store PSNR
865
stats.value = GetPSNR(distortion, pixel_count);
866
}
867
868
#if (DEBUG_SEARCH > 0)
869
printf("#%2d metric:%.1lf -> %.1lf last_q=%.2lf q=%.2lf dq=%.2lf "
870
" range:[%.1f, %.1f]\n",
871
num_pass_left, stats.last_value, stats.value,
872
stats.last_q, stats.q, stats.dq, stats.qmin, stats.qmax);
873
#endif
874
if (enc->max_i4_header_bits > 0 && size_p0 > PARTITION0_SIZE_LIMIT) {
875
++num_pass_left;
876
enc->max_i4_header_bits >>= 1; // strengthen header bit limitation...
877
if (is_last_pass) {
878
ResetSideInfo(&it);
879
}
880
continue; // ...and start over
881
}
882
if (is_last_pass) {
883
break; // done
884
}
885
if (do_search) {
886
ComputeNextQ(&stats); // Adjust q
887
}
888
}
889
if (ok) {
890
if (!stats.do_size_search) {
891
FinalizeTokenProbas(&enc->proba);
892
}
893
ok = VP8EmitTokens(&enc->tokens, enc->parts + 0,
894
(const uint8_t*)proba->coeffs, 1);
895
}
896
ok = ok && WebPReportProgress(enc->pic, enc->percent + remaining_progress,
897
&enc->percent);
898
return PostLoopFinalize(&it, ok);
899
}
900
901
#else
902
903
int VP8EncTokenLoop(VP8Encoder* const enc) {
904
(void)enc;
905
return 0; // we shouldn't be here.
906
}
907
908
#endif // DISABLE_TOKEN_BUFFER
909
910
//------------------------------------------------------------------------------
911
912