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