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 "src/enc/cost_enc.h" |
18 | #include "src/enc/vp8i_enc.h" |
19 | #include "src/dsp/dsp.h" |
20 | #include "src/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 (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 | static float Clamp(float v, float min, float max) { |
35 | return (v < min) ? min : (v > max) ? max : v; |
36 | } |
37 | |
38 | typedef struct { // struct for organizing convergence in either size or PSNR |
39 | int is_first; |
40 | float dq; |
41 | float q, last_q; |
42 | float qmin, qmax; |
43 | double value, last_value; // PSNR or size |
44 | double target; |
45 | int do_size_search; |
46 | } PassStats; |
47 | |
48 | static int InitPassStats(const VP8Encoder* const enc, PassStats* const s) { |
49 | const uint64_t target_size = (uint64_t)enc->config_->target_size; |
50 | const int do_size_search = (target_size != 0); |
51 | const float target_PSNR = enc->config_->target_PSNR; |
52 | |
53 | s->is_first = 1; |
54 | s->dq = 10.f; |
55 | s->qmin = 1.f * enc->config_->qmin; |
56 | s->qmax = 1.f * enc->config_->qmax; |
57 | s->q = s->last_q = Clamp(enc->config_->quality, s->qmin, s->qmax); |
58 | s->target = do_size_search ? (double)target_size |
59 | : (target_PSNR > 0.) ? target_PSNR |
60 | : 40.; // default, just in case |
61 | s->value = s->last_value = 0.; |
62 | s->do_size_search = do_size_search; |
63 | return do_size_search; |
64 | } |
65 | |
66 | static float ComputeNextQ(PassStats* const s) { |
67 | float dq; |
68 | if (s->is_first) { |
69 | dq = (s->value > s->target) ? -s->dq : s->dq; |
70 | s->is_first = 0; |
71 | } else if (s->value != s->last_value) { |
72 | const double slope = (s->target - s->value) / (s->last_value - s->value); |
73 | dq = (float)(slope * (s->last_q - s->q)); |
74 | } else { |
75 | dq = 0.; // we're done?! |
76 | } |
77 | // Limit variable to avoid large swings. |
78 | s->dq = Clamp(dq, -30.f, 30.f); |
79 | s->last_q = s->q; |
80 | s->last_value = s->value; |
81 | s->q = Clamp(s->q + s->dq, s->qmin, s->qmax); |
82 | return s->q; |
83 | } |
84 | |
85 | //------------------------------------------------------------------------------ |
86 | // Tables for level coding |
87 | |
88 | const uint8_t VP8Cat3[] = { 173, 148, 140 }; |
89 | const uint8_t VP8Cat4[] = { 176, 155, 140, 135 }; |
90 | const uint8_t VP8Cat5[] = { 180, 157, 141, 134, 130 }; |
91 | const uint8_t VP8Cat6[] = |
92 | { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129 }; |
93 | |
94 | //------------------------------------------------------------------------------ |
95 | // Reset the statistics about: number of skips, token proba, level cost,... |
96 | |
97 | static void ResetStats(VP8Encoder* const enc) { |
98 | VP8EncProba* const proba = &enc->proba_; |
99 | VP8CalculateLevelCosts(proba); |
100 | proba->nb_skip_ = 0; |
101 | } |
102 | |
103 | //------------------------------------------------------------------------------ |
104 | // Skip decision probability |
105 | |
106 | #define SKIP_PROBA_THRESHOLD 250 // value below which using skip_proba is OK. |
107 | |
108 | static int CalcSkipProba(uint64_t nb, uint64_t total) { |
109 | return (int)(total ? (total - nb) * 255 / total : 255); |
110 | } |
111 | |
112 | // Returns the bit-cost for coding the skip probability. |
113 | static int FinalizeSkipProba(VP8Encoder* const enc) { |
114 | VP8EncProba* const proba = &enc->proba_; |
115 | const int nb_mbs = enc->mb_w_ * enc->mb_h_; |
116 | const int nb_events = proba->nb_skip_; |
117 | int size; |
118 | proba->skip_proba_ = CalcSkipProba(nb_events, nb_mbs); |
119 | proba->use_skip_proba_ = (proba->skip_proba_ < SKIP_PROBA_THRESHOLD); |
120 | size = 256; // 'use_skip_proba' bit |
121 | if (proba->use_skip_proba_) { |
122 | size += nb_events * VP8BitCost(1, proba->skip_proba_) |
123 | + (nb_mbs - nb_events) * VP8BitCost(0, proba->skip_proba_); |
124 | size += 8 * 256; // cost of signaling the skip_proba_ itself. |
125 | } |
126 | return size; |
127 | } |
128 | |
129 | // Collect statistics and deduce probabilities for next coding pass. |
130 | // Return the total bit-cost for coding the probability updates. |
131 | static int CalcTokenProba(int nb, int total) { |
132 | assert(nb <= total); |
133 | return nb ? (255 - nb * 255 / total) : 255; |
134 | } |
135 | |
136 | // Cost of coding 'nb' 1's and 'total-nb' 0's using 'proba' probability. |
137 | static int BranchCost(int nb, int total, int proba) { |
138 | return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba); |
139 | } |
140 | |
141 | static void ResetTokenStats(VP8Encoder* const enc) { |
142 | VP8EncProba* const proba = &enc->proba_; |
143 | memset(proba->stats_, 0, sizeof(proba->stats_)); |
144 | } |
145 | |
146 | static int FinalizeTokenProbas(VP8EncProba* const proba) { |
147 | int has_changed = 0; |
148 | int size = 0; |
149 | int t, b, c, p; |
150 | for (t = 0; t < NUM_TYPES; ++t) { |
151 | for (b = 0; b < NUM_BANDS; ++b) { |
152 | for (c = 0; c < NUM_CTX; ++c) { |
153 | for (p = 0; p < NUM_PROBAS; ++p) { |
154 | const proba_t stats = proba->stats_[t][b][c][p]; |
155 | const int nb = (stats >> 0) & 0xffff; |
156 | const int total = (stats >> 16) & 0xffff; |
157 | const int update_proba = VP8CoeffsUpdateProba[t][b][c][p]; |
158 | const int old_p = VP8CoeffsProba0[t][b][c][p]; |
159 | const int new_p = CalcTokenProba(nb, total); |
160 | const int old_cost = BranchCost(nb, total, old_p) |
161 | + VP8BitCost(0, update_proba); |
162 | const int new_cost = BranchCost(nb, total, new_p) |
163 | + VP8BitCost(1, update_proba) |
164 | + 8 * 256; |
165 | const int use_new_p = (old_cost > new_cost); |
166 | size += VP8BitCost(use_new_p, update_proba); |
167 | if (use_new_p) { // only use proba that seem meaningful enough. |
168 | proba->coeffs_[t][b][c][p] = new_p; |
169 | has_changed |= (new_p != old_p); |
170 | size += 8 * 256; |
171 | } else { |
172 | proba->coeffs_[t][b][c][p] = old_p; |
173 | } |
174 | } |
175 | } |
176 | } |
177 | } |
178 | proba->dirty_ = has_changed; |
179 | return size; |
180 | } |
181 | |
182 | //------------------------------------------------------------------------------ |
183 | // Finalize Segment probability based on the coding tree |
184 | |
185 | static int GetProba(int a, int b) { |
186 | const int total = a + b; |
187 | return (total == 0) ? 255 // that's the default probability. |
188 | : (255 * a + total / 2) / total; // rounded proba |
189 | } |
190 | |
191 | static void ResetSegments(VP8Encoder* const enc) { |
192 | int n; |
193 | for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) { |
194 | enc->mb_info_[n].segment_ = 0; |
195 | } |
196 | } |
197 | |
198 | static void SetSegmentProbas(VP8Encoder* const enc) { |
199 | int p[NUM_MB_SEGMENTS] = { 0 }; |
200 | int n; |
201 | |
202 | for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) { |
203 | const VP8MBInfo* const mb = &enc->mb_info_[n]; |
204 | ++p[mb->segment_]; |
205 | } |
206 | #if !defined(WEBP_DISABLE_STATS) |
207 | if (enc->pic_->stats != NULL) { |
208 | for (n = 0; n < NUM_MB_SEGMENTS; ++n) { |
209 | enc->pic_->stats->segment_size[n] = p[n]; |
210 | } |
211 | } |
212 | #endif |
213 | if (enc->segment_hdr_.num_segments_ > 1) { |
214 | uint8_t* const probas = enc->proba_.segments_; |
215 | probas[0] = GetProba(p[0] + p[1], p[2] + p[3]); |
216 | probas[1] = GetProba(p[0], p[1]); |
217 | probas[2] = GetProba(p[2], p[3]); |
218 | |
219 | enc->segment_hdr_.update_map_ = |
220 | (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255); |
221 | if (!enc->segment_hdr_.update_map_) ResetSegments(enc); |
222 | enc->segment_hdr_.size_ = |
223 | p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) + |
224 | p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) + |
225 | p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) + |
226 | p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2])); |
227 | } else { |
228 | enc->segment_hdr_.update_map_ = 0; |
229 | enc->segment_hdr_.size_ = 0; |
230 | } |
231 | } |
232 | |
233 | //------------------------------------------------------------------------------ |
234 | // Coefficient coding |
235 | |
236 | static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) { |
237 | int n = res->first; |
238 | // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1 |
239 | const uint8_t* p = res->prob[n][ctx]; |
240 | if (!VP8PutBit(bw, res->last >= 0, p[0])) { |
241 | return 0; |
242 | } |
243 | |
244 | while (n < 16) { |
245 | const int c = res->coeffs[n++]; |
246 | const int sign = c < 0; |
247 | int v = sign ? -c : c; |
248 | if (!VP8PutBit(bw, v != 0, p[1])) { |
249 | p = res->prob[VP8EncBands[n]][0]; |
250 | continue; |
251 | } |
252 | if (!VP8PutBit(bw, v > 1, p[2])) { |
253 | p = res->prob[VP8EncBands[n]][1]; |
254 | } else { |
255 | if (!VP8PutBit(bw, v > 4, p[3])) { |
256 | if (VP8PutBit(bw, v != 2, p[4])) { |
257 | VP8PutBit(bw, v == 4, p[5]); |
258 | } |
259 | } else if (!VP8PutBit(bw, v > 10, p[6])) { |
260 | if (!VP8PutBit(bw, v > 6, p[7])) { |
261 | VP8PutBit(bw, v == 6, 159); |
262 | } else { |
263 | VP8PutBit(bw, v >= 9, 165); |
264 | VP8PutBit(bw, !(v & 1), 145); |
265 | } |
266 | } else { |
267 | int mask; |
268 | const uint8_t* tab; |
269 | if (v < 3 + (8 << 1)) { // VP8Cat3 (3b) |
270 | VP8PutBit(bw, 0, p[8]); |
271 | VP8PutBit(bw, 0, p[9]); |
272 | v -= 3 + (8 << 0); |
273 | mask = 1 << 2; |
274 | tab = VP8Cat3; |
275 | } else if (v < 3 + (8 << 2)) { // VP8Cat4 (4b) |
276 | VP8PutBit(bw, 0, p[8]); |
277 | VP8PutBit(bw, 1, p[9]); |
278 | v -= 3 + (8 << 1); |
279 | mask = 1 << 3; |
280 | tab = VP8Cat4; |
281 | } else if (v < 3 + (8 << 3)) { // VP8Cat5 (5b) |
282 | VP8PutBit(bw, 1, p[8]); |
283 | VP8PutBit(bw, 0, p[10]); |
284 | v -= 3 + (8 << 2); |
285 | mask = 1 << 4; |
286 | tab = VP8Cat5; |
287 | } else { // VP8Cat6 (11b) |
288 | VP8PutBit(bw, 1, p[8]); |
289 | VP8PutBit(bw, 1, p[10]); |
290 | v -= 3 + (8 << 3); |
291 | mask = 1 << 10; |
292 | tab = VP8Cat6; |
293 | } |
294 | while (mask) { |
295 | VP8PutBit(bw, !!(v & mask), *tab++); |
296 | mask >>= 1; |
297 | } |
298 | } |
299 | p = res->prob[VP8EncBands[n]][2]; |
300 | } |
301 | VP8PutBitUniform(bw, sign); |
302 | if (n == 16 || !VP8PutBit(bw, n <= res->last, p[0])) { |
303 | return 1; // EOB |
304 | } |
305 | } |
306 | return 1; |
307 | } |
308 | |
309 | static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it, |
310 | const VP8ModeScore* const rd) { |
311 | int x, y, ch; |
312 | VP8Residual res; |
313 | uint64_t pos1, pos2, pos3; |
314 | const int i16 = (it->mb_->type_ == 1); |
315 | const int segment = it->mb_->segment_; |
316 | VP8Encoder* const enc = it->enc_; |
317 | |
318 | VP8IteratorNzToBytes(it); |
319 | |
320 | pos1 = VP8BitWriterPos(bw); |
321 | if (i16) { |
322 | VP8InitResidual(0, 1, enc, &res); |
323 | VP8SetResidualCoeffs(rd->y_dc_levels, &res); |
324 | it->top_nz_[8] = it->left_nz_[8] = |
325 | PutCoeffs(bw, it->top_nz_[8] + it->left_nz_[8], &res); |
326 | VP8InitResidual(1, 0, enc, &res); |
327 | } else { |
328 | VP8InitResidual(0, 3, enc, &res); |
329 | } |
330 | |
331 | // luma-AC |
332 | for (y = 0; y < 4; ++y) { |
333 | for (x = 0; x < 4; ++x) { |
334 | const int ctx = it->top_nz_[x] + it->left_nz_[y]; |
335 | VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res); |
336 | it->top_nz_[x] = it->left_nz_[y] = PutCoeffs(bw, ctx, &res); |
337 | } |
338 | } |
339 | pos2 = VP8BitWriterPos(bw); |
340 | |
341 | // U/V |
342 | VP8InitResidual(0, 2, enc, &res); |
343 | for (ch = 0; ch <= 2; ch += 2) { |
344 | for (y = 0; y < 2; ++y) { |
345 | for (x = 0; x < 2; ++x) { |
346 | const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y]; |
347 | VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res); |
348 | it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = |
349 | PutCoeffs(bw, ctx, &res); |
350 | } |
351 | } |
352 | } |
353 | pos3 = VP8BitWriterPos(bw); |
354 | it->luma_bits_ = pos2 - pos1; |
355 | it->uv_bits_ = pos3 - pos2; |
356 | it->bit_count_[segment][i16] += it->luma_bits_; |
357 | it->bit_count_[segment][2] += it->uv_bits_; |
358 | VP8IteratorBytesToNz(it); |
359 | } |
360 | |
361 | // Same as CodeResiduals, but doesn't actually write anything. |
362 | // Instead, it just records the event distribution. |
363 | static void RecordResiduals(VP8EncIterator* const it, |
364 | const VP8ModeScore* const rd) { |
365 | int x, y, ch; |
366 | VP8Residual res; |
367 | VP8Encoder* const enc = it->enc_; |
368 | |
369 | VP8IteratorNzToBytes(it); |
370 | |
371 | if (it->mb_->type_ == 1) { // i16x16 |
372 | VP8InitResidual(0, 1, enc, &res); |
373 | VP8SetResidualCoeffs(rd->y_dc_levels, &res); |
374 | it->top_nz_[8] = it->left_nz_[8] = |
375 | VP8RecordCoeffs(it->top_nz_[8] + it->left_nz_[8], &res); |
376 | VP8InitResidual(1, 0, enc, &res); |
377 | } else { |
378 | VP8InitResidual(0, 3, enc, &res); |
379 | } |
380 | |
381 | // luma-AC |
382 | for (y = 0; y < 4; ++y) { |
383 | for (x = 0; x < 4; ++x) { |
384 | const int ctx = it->top_nz_[x] + it->left_nz_[y]; |
385 | VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res); |
386 | it->top_nz_[x] = it->left_nz_[y] = VP8RecordCoeffs(ctx, &res); |
387 | } |
388 | } |
389 | |
390 | // U/V |
391 | VP8InitResidual(0, 2, enc, &res); |
392 | for (ch = 0; ch <= 2; ch += 2) { |
393 | for (y = 0; y < 2; ++y) { |
394 | for (x = 0; x < 2; ++x) { |
395 | const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y]; |
396 | VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res); |
397 | it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = |
398 | VP8RecordCoeffs(ctx, &res); |
399 | } |
400 | } |
401 | } |
402 | |
403 | VP8IteratorBytesToNz(it); |
404 | } |
405 | |
406 | //------------------------------------------------------------------------------ |
407 | // Token buffer |
408 | |
409 | #if !defined(DISABLE_TOKEN_BUFFER) |
410 | |
411 | static int RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd, |
412 | VP8TBuffer* const tokens) { |
413 | int x, y, ch; |
414 | VP8Residual res; |
415 | VP8Encoder* const enc = it->enc_; |
416 | |
417 | VP8IteratorNzToBytes(it); |
418 | if (it->mb_->type_ == 1) { // i16x16 |
419 | const int ctx = it->top_nz_[8] + it->left_nz_[8]; |
420 | VP8InitResidual(0, 1, enc, &res); |
421 | VP8SetResidualCoeffs(rd->y_dc_levels, &res); |
422 | it->top_nz_[8] = it->left_nz_[8] = |
423 | VP8RecordCoeffTokens(ctx, &res, tokens); |
424 | VP8InitResidual(1, 0, enc, &res); |
425 | } else { |
426 | VP8InitResidual(0, 3, enc, &res); |
427 | } |
428 | |
429 | // luma-AC |
430 | for (y = 0; y < 4; ++y) { |
431 | for (x = 0; x < 4; ++x) { |
432 | const int ctx = it->top_nz_[x] + it->left_nz_[y]; |
433 | VP8SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res); |
434 | it->top_nz_[x] = it->left_nz_[y] = |
435 | VP8RecordCoeffTokens(ctx, &res, tokens); |
436 | } |
437 | } |
438 | |
439 | // U/V |
440 | VP8InitResidual(0, 2, enc, &res); |
441 | for (ch = 0; ch <= 2; ch += 2) { |
442 | for (y = 0; y < 2; ++y) { |
443 | for (x = 0; x < 2; ++x) { |
444 | const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y]; |
445 | VP8SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res); |
446 | it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] = |
447 | VP8RecordCoeffTokens(ctx, &res, tokens); |
448 | } |
449 | } |
450 | } |
451 | VP8IteratorBytesToNz(it); |
452 | return !tokens->error_; |
453 | } |
454 | |
455 | #endif // !DISABLE_TOKEN_BUFFER |
456 | |
457 | //------------------------------------------------------------------------------ |
458 | // ExtraInfo map / Debug function |
459 | |
460 | #if !defined(WEBP_DISABLE_STATS) |
461 | |
462 | #if SEGMENT_VISU |
463 | static void SetBlock(uint8_t* p, int value, int size) { |
464 | int y; |
465 | for (y = 0; y < size; ++y) { |
466 | memset(p, value, size); |
467 | p += BPS; |
468 | } |
469 | } |
470 | #endif |
471 | |
472 | static void ResetSSE(VP8Encoder* const enc) { |
473 | enc->sse_[0] = 0; |
474 | enc->sse_[1] = 0; |
475 | enc->sse_[2] = 0; |
476 | // Note: enc->sse_[3] is managed by alpha.c |
477 | enc->sse_count_ = 0; |
478 | } |
479 | |
480 | static void StoreSSE(const VP8EncIterator* const it) { |
481 | VP8Encoder* const enc = it->enc_; |
482 | const uint8_t* const in = it->yuv_in_; |
483 | const uint8_t* const out = it->yuv_out_; |
484 | // Note: not totally accurate at boundary. And doesn't include in-loop filter. |
485 | enc->sse_[0] += VP8SSE16x16(in + Y_OFF_ENC, out + Y_OFF_ENC); |
486 | enc->sse_[1] += VP8SSE8x8(in + U_OFF_ENC, out + U_OFF_ENC); |
487 | enc->sse_[2] += VP8SSE8x8(in + V_OFF_ENC, out + V_OFF_ENC); |
488 | enc->sse_count_ += 16 * 16; |
489 | } |
490 | |
491 | static void StoreSideInfo(const VP8EncIterator* const it) { |
492 | VP8Encoder* const enc = it->enc_; |
493 | const VP8MBInfo* const mb = it->mb_; |
494 | WebPPicture* const pic = enc->pic_; |
495 | |
496 | if (pic->stats != NULL) { |
497 | StoreSSE(it); |
498 | enc->block_count_[0] += (mb->type_ == 0); |
499 | enc->block_count_[1] += (mb->type_ == 1); |
500 | enc->block_count_[2] += (mb->skip_ != 0); |
501 | } |
502 | |
503 | if (pic->extra_info != NULL) { |
504 | uint8_t* const info = &pic->extra_info[it->x_ + it->y_ * enc->mb_w_]; |
505 | switch (pic->extra_info_type) { |
506 | case 1: *info = mb->type_; break; |
507 | case 2: *info = mb->segment_; break; |
508 | case 3: *info = enc->dqm_[mb->segment_].quant_; break; |
509 | case 4: *info = (mb->type_ == 1) ? it->preds_[0] : 0xff; break; |
510 | case 5: *info = mb->uv_mode_; break; |
511 | case 6: { |
512 | const int b = (int)((it->luma_bits_ + it->uv_bits_ + 7) >> 3); |
513 | *info = (b > 255) ? 255 : b; break; |
514 | } |
515 | case 7: *info = mb->alpha_; break; |
516 | default: *info = 0; break; |
517 | } |
518 | } |
519 | #if SEGMENT_VISU // visualize segments and prediction modes |
520 | SetBlock(it->yuv_out_ + Y_OFF_ENC, mb->segment_ * 64, 16); |
521 | SetBlock(it->yuv_out_ + U_OFF_ENC, it->preds_[0] * 64, 8); |
522 | SetBlock(it->yuv_out_ + V_OFF_ENC, mb->uv_mode_ * 64, 8); |
523 | #endif |
524 | } |
525 | |
526 | static void ResetSideInfo(const VP8EncIterator* const it) { |
527 | VP8Encoder* const enc = it->enc_; |
528 | WebPPicture* const pic = enc->pic_; |
529 | if (pic->stats != NULL) { |
530 | memset(enc->block_count_, 0, sizeof(enc->block_count_)); |
531 | } |
532 | ResetSSE(enc); |
533 | } |
534 | #else // defined(WEBP_DISABLE_STATS) |
535 | static void ResetSSE(VP8Encoder* const enc) { |
536 | (void)enc; |
537 | } |
538 | static void StoreSideInfo(const VP8EncIterator* const it) { |
539 | VP8Encoder* const enc = it->enc_; |
540 | WebPPicture* const pic = enc->pic_; |
541 | if (pic->extra_info != NULL) { |
542 | if (it->x_ == 0 && it->y_ == 0) { // only do it once, at start |
543 | memset(pic->extra_info, 0, |
544 | enc->mb_w_ * enc->mb_h_ * sizeof(*pic->extra_info)); |
545 | } |
546 | } |
547 | } |
548 | |
549 | static void ResetSideInfo(const VP8EncIterator* const it) { |
550 | (void)it; |
551 | } |
552 | #endif // !defined(WEBP_DISABLE_STATS) |
553 | |
554 | static double GetPSNR(uint64_t mse, uint64_t size) { |
555 | return (mse > 0 && size > 0) ? 10. * log10(255. * 255. * size / mse) : 99; |
556 | } |
557 | |
558 | //------------------------------------------------------------------------------ |
559 | // StatLoop(): only collect statistics (number of skips, token usage, ...). |
560 | // This is used for deciding optimal probabilities. It also modifies the |
561 | // quantizer value if some target (size, PSNR) was specified. |
562 | |
563 | static void SetLoopParams(VP8Encoder* const enc, float q) { |
564 | // Make sure the quality parameter is inside valid bounds |
565 | q = Clamp(q, 0.f, 100.f); |
566 | |
567 | VP8SetSegmentParams(enc, q); // setup segment quantizations and filters |
568 | SetSegmentProbas(enc); // compute segment probabilities |
569 | |
570 | ResetStats(enc); |
571 | ResetSSE(enc); |
572 | } |
573 | |
574 | static uint64_t OneStatPass(VP8Encoder* const enc, VP8RDLevel rd_opt, |
575 | int nb_mbs, int percent_delta, |
576 | PassStats* const s) { |
577 | VP8EncIterator it; |
578 | uint64_t size = 0; |
579 | uint64_t size_p0 = 0; |
580 | uint64_t distortion = 0; |
581 | const uint64_t pixel_count = nb_mbs * 384; |
582 | |
583 | VP8IteratorInit(enc, &it); |
584 | SetLoopParams(enc, s->q); |
585 | do { |
586 | VP8ModeScore info; |
587 | VP8IteratorImport(&it, NULL); |
588 | if (VP8Decimate(&it, &info, rd_opt)) { |
589 | // Just record the number of skips and act like skip_proba is not used. |
590 | ++enc->proba_.nb_skip_; |
591 | } |
592 | RecordResiduals(&it, &info); |
593 | size += info.R + info.H; |
594 | size_p0 += info.H; |
595 | distortion += info.D; |
596 | if (percent_delta && !VP8IteratorProgress(&it, percent_delta)) { |
597 | return 0; |
598 | } |
599 | VP8IteratorSaveBoundary(&it); |
600 | } while (VP8IteratorNext(&it) && --nb_mbs > 0); |
601 | |
602 | size_p0 += enc->segment_hdr_.size_; |
603 | if (s->do_size_search) { |
604 | size += FinalizeSkipProba(enc); |
605 | size += FinalizeTokenProbas(&enc->proba_); |
606 | size = ((size + size_p0 + 1024) >> 11) + HEADER_SIZE_ESTIMATE; |
607 | s->value = (double)size; |
608 | } else { |
609 | s->value = GetPSNR(distortion, pixel_count); |
610 | } |
611 | return size_p0; |
612 | } |
613 | |
614 | static int StatLoop(VP8Encoder* const enc) { |
615 | const int method = enc->method_; |
616 | const int do_search = enc->do_search_; |
617 | const int fast_probe = ((method == 0 || method == 3) && !do_search); |
618 | int num_pass_left = enc->config_->pass; |
619 | const int task_percent = 20; |
620 | const int percent_per_pass = |
621 | (task_percent + num_pass_left / 2) / num_pass_left; |
622 | const int final_percent = enc->percent_ + task_percent; |
623 | const VP8RDLevel rd_opt = |
624 | (method >= 3 || do_search) ? RD_OPT_BASIC : RD_OPT_NONE; |
625 | int nb_mbs = enc->mb_w_ * enc->mb_h_; |
626 | PassStats stats; |
627 | |
628 | InitPassStats(enc, &stats); |
629 | ResetTokenStats(enc); |
630 | |
631 | // Fast mode: quick analysis pass over few mbs. Better than nothing. |
632 | if (fast_probe) { |
633 | if (method == 3) { // we need more stats for method 3 to be reliable. |
634 | nb_mbs = (nb_mbs > 200) ? nb_mbs >> 1 : 100; |
635 | } else { |
636 | nb_mbs = (nb_mbs > 200) ? nb_mbs >> 2 : 50; |
637 | } |
638 | } |
639 | |
640 | while (num_pass_left-- > 0) { |
641 | const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) || |
642 | (num_pass_left == 0) || |
643 | (enc->max_i4_header_bits_ == 0); |
644 | const uint64_t size_p0 = |
645 | OneStatPass(enc, rd_opt, nb_mbs, percent_per_pass, &stats); |
646 | if (size_p0 == 0) return 0; |
647 | #if (DEBUG_SEARCH > 0) |
648 | printf("#%d value:%.1lf -> %.1lf q:%.2f -> %.2f\n" , |
649 | num_pass_left, stats.last_value, stats.value, stats.last_q, stats.q); |
650 | #endif |
651 | if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) { |
652 | ++num_pass_left; |
653 | enc->max_i4_header_bits_ >>= 1; // strengthen header bit limitation... |
654 | continue; // ...and start over |
655 | } |
656 | if (is_last_pass) { |
657 | break; |
658 | } |
659 | // If no target size: just do several pass without changing 'q' |
660 | if (do_search) { |
661 | ComputeNextQ(&stats); |
662 | if (fabs(stats.dq) <= DQ_LIMIT) break; |
663 | } |
664 | } |
665 | if (!do_search || !stats.do_size_search) { |
666 | // Need to finalize probas now, since it wasn't done during the search. |
667 | FinalizeSkipProba(enc); |
668 | FinalizeTokenProbas(&enc->proba_); |
669 | } |
670 | VP8CalculateLevelCosts(&enc->proba_); // finalize costs |
671 | return WebPReportProgress(enc->pic_, final_percent, &enc->percent_); |
672 | } |
673 | |
674 | //------------------------------------------------------------------------------ |
675 | // Main loops |
676 | // |
677 | |
678 | static const uint8_t kAverageBytesPerMB[8] = { 50, 24, 16, 9, 7, 5, 3, 2 }; |
679 | |
680 | static int PreLoopInitialize(VP8Encoder* const enc) { |
681 | int p; |
682 | int ok = 1; |
683 | const int average_bytes_per_MB = kAverageBytesPerMB[enc->base_quant_ >> 4]; |
684 | const int bytes_per_parts = |
685 | enc->mb_w_ * enc->mb_h_ * average_bytes_per_MB / enc->num_parts_; |
686 | // Initialize the bit-writers |
687 | for (p = 0; ok && p < enc->num_parts_; ++p) { |
688 | ok = VP8BitWriterInit(enc->parts_ + p, bytes_per_parts); |
689 | } |
690 | if (!ok) { |
691 | VP8EncFreeBitWriters(enc); // malloc error occurred |
692 | return WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY); |
693 | } |
694 | return ok; |
695 | } |
696 | |
697 | static int PostLoopFinalize(VP8EncIterator* const it, int ok) { |
698 | VP8Encoder* const enc = it->enc_; |
699 | if (ok) { // Finalize the partitions, check for extra errors. |
700 | int p; |
701 | for (p = 0; p < enc->num_parts_; ++p) { |
702 | VP8BitWriterFinish(enc->parts_ + p); |
703 | ok &= !enc->parts_[p].error_; |
704 | } |
705 | } |
706 | |
707 | if (ok) { // All good. Finish up. |
708 | #if !defined(WEBP_DISABLE_STATS) |
709 | if (enc->pic_->stats != NULL) { // finalize byte counters... |
710 | int i, s; |
711 | for (i = 0; i <= 2; ++i) { |
712 | for (s = 0; s < NUM_MB_SEGMENTS; ++s) { |
713 | enc->residual_bytes_[i][s] = (int)((it->bit_count_[s][i] + 7) >> 3); |
714 | } |
715 | } |
716 | } |
717 | #endif |
718 | VP8AdjustFilterStrength(it); // ...and store filter stats. |
719 | } else { |
720 | // Something bad happened -> need to do some memory cleanup. |
721 | VP8EncFreeBitWriters(enc); |
722 | return WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY); |
723 | } |
724 | return ok; |
725 | } |
726 | |
727 | //------------------------------------------------------------------------------ |
728 | // VP8EncLoop(): does the final bitstream coding. |
729 | |
730 | static void ResetAfterSkip(VP8EncIterator* const it) { |
731 | if (it->mb_->type_ == 1) { |
732 | *it->nz_ = 0; // reset all predictors |
733 | it->left_nz_[8] = 0; |
734 | } else { |
735 | *it->nz_ &= (1 << 24); // preserve the dc_nz bit |
736 | } |
737 | } |
738 | |
739 | int VP8EncLoop(VP8Encoder* const enc) { |
740 | VP8EncIterator it; |
741 | int ok = PreLoopInitialize(enc); |
742 | if (!ok) return 0; |
743 | |
744 | StatLoop(enc); // stats-collection loop |
745 | |
746 | VP8IteratorInit(enc, &it); |
747 | VP8InitFilter(&it); |
748 | do { |
749 | VP8ModeScore info; |
750 | const int dont_use_skip = !enc->proba_.use_skip_proba_; |
751 | const VP8RDLevel rd_opt = enc->rd_opt_level_; |
752 | |
753 | VP8IteratorImport(&it, NULL); |
754 | // Warning! order is important: first call VP8Decimate() and |
755 | // *then* decide how to code the skip decision if there's one. |
756 | if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) { |
757 | CodeResiduals(it.bw_, &it, &info); |
758 | if (it.bw_->error_) { |
759 | // enc->pic_->error_code is set in PostLoopFinalize(). |
760 | ok = 0; |
761 | break; |
762 | } |
763 | } else { // reset predictors after a skip |
764 | ResetAfterSkip(&it); |
765 | } |
766 | StoreSideInfo(&it); |
767 | VP8StoreFilterStats(&it); |
768 | VP8IteratorExport(&it); |
769 | ok = VP8IteratorProgress(&it, 20); |
770 | VP8IteratorSaveBoundary(&it); |
771 | } while (ok && VP8IteratorNext(&it)); |
772 | |
773 | return PostLoopFinalize(&it, ok); |
774 | } |
775 | |
776 | //------------------------------------------------------------------------------ |
777 | // Single pass using Token Buffer. |
778 | |
779 | #if !defined(DISABLE_TOKEN_BUFFER) |
780 | |
781 | #define MIN_COUNT 96 // minimum number of macroblocks before updating stats |
782 | |
783 | int VP8EncTokenLoop(VP8Encoder* const enc) { |
784 | // Roughly refresh the proba eight times per pass |
785 | int max_count = (enc->mb_w_ * enc->mb_h_) >> 3; |
786 | int num_pass_left = enc->config_->pass; |
787 | int remaining_progress = 40; // percents |
788 | const int do_search = enc->do_search_; |
789 | VP8EncIterator it; |
790 | VP8EncProba* const proba = &enc->proba_; |
791 | const VP8RDLevel rd_opt = enc->rd_opt_level_; |
792 | const uint64_t pixel_count = enc->mb_w_ * enc->mb_h_ * 384; |
793 | PassStats stats; |
794 | int ok; |
795 | |
796 | InitPassStats(enc, &stats); |
797 | ok = PreLoopInitialize(enc); |
798 | if (!ok) return 0; |
799 | |
800 | if (max_count < MIN_COUNT) max_count = MIN_COUNT; |
801 | |
802 | assert(enc->num_parts_ == 1); |
803 | assert(enc->use_tokens_); |
804 | assert(proba->use_skip_proba_ == 0); |
805 | assert(rd_opt >= RD_OPT_BASIC); // otherwise, token-buffer won't be useful |
806 | assert(num_pass_left > 0); |
807 | |
808 | while (ok && num_pass_left-- > 0) { |
809 | const int is_last_pass = (fabs(stats.dq) <= DQ_LIMIT) || |
810 | (num_pass_left == 0) || |
811 | (enc->max_i4_header_bits_ == 0); |
812 | uint64_t size_p0 = 0; |
813 | uint64_t distortion = 0; |
814 | int cnt = max_count; |
815 | // The final number of passes is not trivial to know in advance. |
816 | const int pass_progress = remaining_progress / (2 + num_pass_left); |
817 | remaining_progress -= pass_progress; |
818 | VP8IteratorInit(enc, &it); |
819 | SetLoopParams(enc, stats.q); |
820 | if (is_last_pass) { |
821 | ResetTokenStats(enc); |
822 | VP8InitFilter(&it); // don't collect stats until last pass (too costly) |
823 | } |
824 | VP8TBufferClear(&enc->tokens_); |
825 | do { |
826 | VP8ModeScore info; |
827 | VP8IteratorImport(&it, NULL); |
828 | if (--cnt < 0) { |
829 | FinalizeTokenProbas(proba); |
830 | VP8CalculateLevelCosts(proba); // refresh cost tables for rd-opt |
831 | cnt = max_count; |
832 | } |
833 | VP8Decimate(&it, &info, rd_opt); |
834 | ok = RecordTokens(&it, &info, &enc->tokens_); |
835 | if (!ok) { |
836 | WebPEncodingSetError(enc->pic_, VP8_ENC_ERROR_OUT_OF_MEMORY); |
837 | break; |
838 | } |
839 | size_p0 += info.H; |
840 | distortion += info.D; |
841 | if (is_last_pass) { |
842 | StoreSideInfo(&it); |
843 | VP8StoreFilterStats(&it); |
844 | VP8IteratorExport(&it); |
845 | ok = VP8IteratorProgress(&it, pass_progress); |
846 | } |
847 | VP8IteratorSaveBoundary(&it); |
848 | } while (ok && VP8IteratorNext(&it)); |
849 | if (!ok) break; |
850 | |
851 | size_p0 += enc->segment_hdr_.size_; |
852 | if (stats.do_size_search) { |
853 | uint64_t size = FinalizeTokenProbas(&enc->proba_); |
854 | size += VP8EstimateTokenSize(&enc->tokens_, |
855 | (const uint8_t*)proba->coeffs_); |
856 | size = (size + size_p0 + 1024) >> 11; // -> size in bytes |
857 | size += HEADER_SIZE_ESTIMATE; |
858 | stats.value = (double)size; |
859 | } else { // compute and store PSNR |
860 | stats.value = GetPSNR(distortion, pixel_count); |
861 | } |
862 | |
863 | #if (DEBUG_SEARCH > 0) |
864 | printf("#%2d metric:%.1lf -> %.1lf last_q=%.2lf q=%.2lf dq=%.2lf " |
865 | " range:[%.1f, %.1f]\n" , |
866 | num_pass_left, stats.last_value, stats.value, |
867 | stats.last_q, stats.q, stats.dq, stats.qmin, stats.qmax); |
868 | #endif |
869 | if (enc->max_i4_header_bits_ > 0 && size_p0 > PARTITION0_SIZE_LIMIT) { |
870 | ++num_pass_left; |
871 | enc->max_i4_header_bits_ >>= 1; // strengthen header bit limitation... |
872 | if (is_last_pass) { |
873 | ResetSideInfo(&it); |
874 | } |
875 | continue; // ...and start over |
876 | } |
877 | if (is_last_pass) { |
878 | break; // done |
879 | } |
880 | if (do_search) { |
881 | ComputeNextQ(&stats); // Adjust q |
882 | } |
883 | } |
884 | if (ok) { |
885 | if (!stats.do_size_search) { |
886 | FinalizeTokenProbas(&enc->proba_); |
887 | } |
888 | ok = VP8EmitTokens(&enc->tokens_, enc->parts_ + 0, |
889 | (const uint8_t*)proba->coeffs_, 1); |
890 | } |
891 | ok = ok && WebPReportProgress(enc->pic_, enc->percent_ + remaining_progress, |
892 | &enc->percent_); |
893 | return PostLoopFinalize(&it, ok); |
894 | } |
895 | |
896 | #else |
897 | |
898 | int VP8EncTokenLoop(VP8Encoder* const enc) { |
899 | (void)enc; |
900 | return 0; // we shouldn't be here. |
901 | } |
902 | |
903 | #endif // DISABLE_TOKEN_BUFFER |
904 | |
905 | //------------------------------------------------------------------------------ |
906 | |