|
1 /* |
|
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. |
|
3 * |
|
4 * Use of this source code is governed by a BSD-style license |
|
5 * that can be found in the LICENSE file in the root of the source |
|
6 * tree. An additional intellectual property rights grant can be found |
|
7 * in the file PATENTS. All contributing project authors may |
|
8 * be found in the AUTHORS file in the root of the source tree. |
|
9 */ |
|
10 |
|
11 |
|
12 #include <stdlib.h> |
|
13 #include <stdio.h> |
|
14 #include <string.h> |
|
15 #include <limits.h> |
|
16 #include <assert.h> |
|
17 #include <math.h> |
|
18 |
|
19 #include "vp9/common/vp9_alloccommon.h" |
|
20 #include "vp9/common/vp9_common.h" |
|
21 #include "vp9/encoder/vp9_ratectrl.h" |
|
22 #include "vp9/common/vp9_entropymode.h" |
|
23 #include "vpx_mem/vpx_mem.h" |
|
24 #include "vp9/common/vp9_systemdependent.h" |
|
25 #include "vp9/encoder/vp9_encodemv.h" |
|
26 #include "vp9/common/vp9_quant_common.h" |
|
27 #include "vp9/common/vp9_seg_common.h" |
|
28 |
|
29 #define MIN_BPB_FACTOR 0.005 |
|
30 #define MAX_BPB_FACTOR 50 |
|
31 |
|
32 // Bits Per MB at different Q (Multiplied by 512) |
|
33 #define BPER_MB_NORMBITS 9 |
|
34 |
|
35 static const unsigned int prior_key_frame_weight[KEY_FRAME_CONTEXT] = |
|
36 { 1, 2, 3, 4, 5 }; |
|
37 |
|
38 // These functions use formulaic calculations to make playing with the |
|
39 // quantizer tables easier. If necessary they can be replaced by lookup |
|
40 // tables if and when things settle down in the experimental bitstream |
|
41 double vp9_convert_qindex_to_q(int qindex) { |
|
42 // Convert the index to a real Q value (scaled down to match old Q values) |
|
43 return vp9_ac_quant(qindex, 0) / 4.0; |
|
44 } |
|
45 |
|
46 int vp9_gfboost_qadjust(int qindex) { |
|
47 const double q = vp9_convert_qindex_to_q(qindex); |
|
48 return (int)((0.00000828 * q * q * q) + |
|
49 (-0.0055 * q * q) + |
|
50 (1.32 * q) + 79.3); |
|
51 } |
|
52 |
|
53 static int kfboost_qadjust(int qindex) { |
|
54 const double q = vp9_convert_qindex_to_q(qindex); |
|
55 return (int)((0.00000973 * q * q * q) + |
|
56 (-0.00613 * q * q) + |
|
57 (1.316 * q) + 121.2); |
|
58 } |
|
59 |
|
60 int vp9_bits_per_mb(FRAME_TYPE frame_type, int qindex, |
|
61 double correction_factor) { |
|
62 const double q = vp9_convert_qindex_to_q(qindex); |
|
63 int enumerator = frame_type == KEY_FRAME ? 3300000 : 2250000; |
|
64 |
|
65 // q based adjustment to baseline enumerator |
|
66 enumerator += (int)(enumerator * q) >> 12; |
|
67 return (int)(0.5 + (enumerator * correction_factor / q)); |
|
68 } |
|
69 |
|
70 void vp9_save_coding_context(VP9_COMP *cpi) { |
|
71 CODING_CONTEXT *const cc = &cpi->coding_context; |
|
72 VP9_COMMON *cm = &cpi->common; |
|
73 |
|
74 // Stores a snapshot of key state variables which can subsequently be |
|
75 // restored with a call to vp9_restore_coding_context. These functions are |
|
76 // intended for use in a re-code loop in vp9_compress_frame where the |
|
77 // quantizer value is adjusted between loop iterations. |
|
78 vp9_copy(cc->nmvjointcost, cpi->mb.nmvjointcost); |
|
79 vp9_copy(cc->nmvcosts, cpi->mb.nmvcosts); |
|
80 vp9_copy(cc->nmvcosts_hp, cpi->mb.nmvcosts_hp); |
|
81 |
|
82 vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs); |
|
83 |
|
84 vpx_memcpy(cpi->coding_context.last_frame_seg_map_copy, |
|
85 cm->last_frame_seg_map, (cm->mi_rows * cm->mi_cols)); |
|
86 |
|
87 vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas); |
|
88 vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas); |
|
89 |
|
90 cc->fc = cm->fc; |
|
91 } |
|
92 |
|
93 void vp9_restore_coding_context(VP9_COMP *cpi) { |
|
94 CODING_CONTEXT *const cc = &cpi->coding_context; |
|
95 VP9_COMMON *cm = &cpi->common; |
|
96 |
|
97 // Restore key state variables to the snapshot state stored in the |
|
98 // previous call to vp9_save_coding_context. |
|
99 vp9_copy(cpi->mb.nmvjointcost, cc->nmvjointcost); |
|
100 vp9_copy(cpi->mb.nmvcosts, cc->nmvcosts); |
|
101 vp9_copy(cpi->mb.nmvcosts_hp, cc->nmvcosts_hp); |
|
102 |
|
103 vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs); |
|
104 |
|
105 vpx_memcpy(cm->last_frame_seg_map, |
|
106 cpi->coding_context.last_frame_seg_map_copy, |
|
107 (cm->mi_rows * cm->mi_cols)); |
|
108 |
|
109 vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas); |
|
110 vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas); |
|
111 |
|
112 cm->fc = cc->fc; |
|
113 } |
|
114 |
|
115 void vp9_setup_key_frame(VP9_COMP *cpi) { |
|
116 VP9_COMMON *cm = &cpi->common; |
|
117 |
|
118 vp9_setup_past_independence(cm); |
|
119 |
|
120 // interval before next GF |
|
121 cpi->frames_till_gf_update_due = cpi->baseline_gf_interval; |
|
122 /* All buffers are implicitly updated on key frames. */ |
|
123 cpi->refresh_golden_frame = 1; |
|
124 cpi->refresh_alt_ref_frame = 1; |
|
125 } |
|
126 |
|
127 void vp9_setup_inter_frame(VP9_COMP *cpi) { |
|
128 VP9_COMMON *cm = &cpi->common; |
|
129 if (cm->error_resilient_mode || cm->intra_only) |
|
130 vp9_setup_past_independence(cm); |
|
131 |
|
132 assert(cm->frame_context_idx < NUM_FRAME_CONTEXTS); |
|
133 cm->fc = cm->frame_contexts[cm->frame_context_idx]; |
|
134 } |
|
135 |
|
136 static int estimate_bits_at_q(int frame_kind, int q, int mbs, |
|
137 double correction_factor) { |
|
138 const int bpm = (int)(vp9_bits_per_mb(frame_kind, q, correction_factor)); |
|
139 |
|
140 // Attempt to retain reasonable accuracy without overflow. The cutoff is |
|
141 // chosen such that the maximum product of Bpm and MBs fits 31 bits. The |
|
142 // largest Bpm takes 20 bits. |
|
143 return (mbs > (1 << 11)) ? (bpm >> BPER_MB_NORMBITS) * mbs |
|
144 : (bpm * mbs) >> BPER_MB_NORMBITS; |
|
145 } |
|
146 |
|
147 |
|
148 static void calc_iframe_target_size(VP9_COMP *cpi) { |
|
149 // boost defaults to half second |
|
150 int target; |
|
151 |
|
152 // Clear down mmx registers to allow floating point in what follows |
|
153 vp9_clear_system_state(); // __asm emms; |
|
154 |
|
155 // New Two pass RC |
|
156 target = cpi->per_frame_bandwidth; |
|
157 |
|
158 if (cpi->oxcf.rc_max_intra_bitrate_pct) { |
|
159 int max_rate = cpi->per_frame_bandwidth |
|
160 * cpi->oxcf.rc_max_intra_bitrate_pct / 100; |
|
161 |
|
162 if (target > max_rate) |
|
163 target = max_rate; |
|
164 } |
|
165 |
|
166 cpi->this_frame_target = target; |
|
167 } |
|
168 |
|
169 |
|
170 // Do the best we can to define the parameters for the next GF based |
|
171 // on what information we have available. |
|
172 // |
|
173 // In this experimental code only two pass is supported |
|
174 // so we just use the interval determined in the two pass code. |
|
175 static void calc_gf_params(VP9_COMP *cpi) { |
|
176 // Set the gf interval |
|
177 cpi->frames_till_gf_update_due = cpi->baseline_gf_interval; |
|
178 } |
|
179 |
|
180 |
|
181 static void calc_pframe_target_size(VP9_COMP *cpi) { |
|
182 const int min_frame_target = MAX(cpi->min_frame_bandwidth, |
|
183 cpi->av_per_frame_bandwidth >> 5); |
|
184 if (cpi->refresh_alt_ref_frame) { |
|
185 // Special alt reference frame case |
|
186 // Per frame bit target for the alt ref frame |
|
187 cpi->per_frame_bandwidth = cpi->twopass.gf_bits; |
|
188 cpi->this_frame_target = cpi->per_frame_bandwidth; |
|
189 } else { |
|
190 // Normal frames (gf,and inter) |
|
191 cpi->this_frame_target = cpi->per_frame_bandwidth; |
|
192 } |
|
193 |
|
194 // Check that the total sum of adjustments is not above the maximum allowed. |
|
195 // That is, having allowed for the KF and GF penalties, we have not pushed |
|
196 // the current inter-frame target too low. If the adjustment we apply here is |
|
197 // not capable of recovering all the extra bits we have spent in the KF or GF, |
|
198 // then the remainder will have to be recovered over a longer time span via |
|
199 // other buffer / rate control mechanisms. |
|
200 if (cpi->this_frame_target < min_frame_target) |
|
201 cpi->this_frame_target = min_frame_target; |
|
202 |
|
203 if (!cpi->refresh_alt_ref_frame) |
|
204 // Note the baseline target data rate for this inter frame. |
|
205 cpi->inter_frame_target = cpi->this_frame_target; |
|
206 |
|
207 // Adjust target frame size for Golden Frames: |
|
208 if (cpi->frames_till_gf_update_due == 0) { |
|
209 const int q = (cpi->oxcf.fixed_q < 0) ? cpi->last_q[INTER_FRAME] |
|
210 : cpi->oxcf.fixed_q; |
|
211 |
|
212 cpi->refresh_golden_frame = 1; |
|
213 |
|
214 calc_gf_params(cpi); |
|
215 |
|
216 // If we are using alternate ref instead of gf then do not apply the boost |
|
217 // It will instead be applied to the altref update |
|
218 // Jims modified boost |
|
219 if (!cpi->source_alt_ref_active) { |
|
220 if (cpi->oxcf.fixed_q < 0) { |
|
221 // The spend on the GF is defined in the two pass code |
|
222 // for two pass encodes |
|
223 cpi->this_frame_target = cpi->per_frame_bandwidth; |
|
224 } else { |
|
225 cpi->this_frame_target = |
|
226 (estimate_bits_at_q(1, q, cpi->common.MBs, 1.0) |
|
227 * cpi->last_boost) / 100; |
|
228 } |
|
229 } else { |
|
230 // If there is an active ARF at this location use the minimum |
|
231 // bits on this frame even if it is a constructed arf. |
|
232 // The active maximum quantizer insures that an appropriate |
|
233 // number of bits will be spent if needed for constructed ARFs. |
|
234 cpi->this_frame_target = 0; |
|
235 } |
|
236 } |
|
237 } |
|
238 |
|
239 |
|
240 void vp9_update_rate_correction_factors(VP9_COMP *cpi, int damp_var) { |
|
241 const int q = cpi->common.base_qindex; |
|
242 int correction_factor = 100; |
|
243 double rate_correction_factor; |
|
244 double adjustment_limit; |
|
245 |
|
246 int projected_size_based_on_q = 0; |
|
247 |
|
248 // Clear down mmx registers to allow floating point in what follows |
|
249 vp9_clear_system_state(); // __asm emms; |
|
250 |
|
251 if (cpi->common.frame_type == KEY_FRAME) { |
|
252 rate_correction_factor = cpi->key_frame_rate_correction_factor; |
|
253 } else { |
|
254 if (cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) |
|
255 rate_correction_factor = cpi->gf_rate_correction_factor; |
|
256 else |
|
257 rate_correction_factor = cpi->rate_correction_factor; |
|
258 } |
|
259 |
|
260 // Work out how big we would have expected the frame to be at this Q given |
|
261 // the current correction factor. |
|
262 // Stay in double to avoid int overflow when values are large |
|
263 projected_size_based_on_q = estimate_bits_at_q(cpi->common.frame_type, q, |
|
264 cpi->common.MBs, |
|
265 rate_correction_factor); |
|
266 |
|
267 // Work out a size correction factor. |
|
268 if (projected_size_based_on_q > 0) |
|
269 correction_factor = |
|
270 (100 * cpi->projected_frame_size) / projected_size_based_on_q; |
|
271 |
|
272 // More heavily damped adjustment used if we have been oscillating either side |
|
273 // of target. |
|
274 switch (damp_var) { |
|
275 case 0: |
|
276 adjustment_limit = 0.75; |
|
277 break; |
|
278 case 1: |
|
279 adjustment_limit = 0.375; |
|
280 break; |
|
281 case 2: |
|
282 default: |
|
283 adjustment_limit = 0.25; |
|
284 break; |
|
285 } |
|
286 |
|
287 // if ( (correction_factor > 102) && (Q < cpi->active_worst_quality) ) |
|
288 if (correction_factor > 102) { |
|
289 // We are not already at the worst allowable quality |
|
290 correction_factor = |
|
291 (int)(100 + ((correction_factor - 100) * adjustment_limit)); |
|
292 rate_correction_factor = |
|
293 ((rate_correction_factor * correction_factor) / 100); |
|
294 |
|
295 // Keep rate_correction_factor within limits |
|
296 if (rate_correction_factor > MAX_BPB_FACTOR) |
|
297 rate_correction_factor = MAX_BPB_FACTOR; |
|
298 } else if (correction_factor < 99) { |
|
299 // We are not already at the best allowable quality |
|
300 correction_factor = |
|
301 (int)(100 - ((100 - correction_factor) * adjustment_limit)); |
|
302 rate_correction_factor = |
|
303 ((rate_correction_factor * correction_factor) / 100); |
|
304 |
|
305 // Keep rate_correction_factor within limits |
|
306 if (rate_correction_factor < MIN_BPB_FACTOR) |
|
307 rate_correction_factor = MIN_BPB_FACTOR; |
|
308 } |
|
309 |
|
310 if (cpi->common.frame_type == KEY_FRAME) { |
|
311 cpi->key_frame_rate_correction_factor = rate_correction_factor; |
|
312 } else { |
|
313 if (cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) |
|
314 cpi->gf_rate_correction_factor = rate_correction_factor; |
|
315 else |
|
316 cpi->rate_correction_factor = rate_correction_factor; |
|
317 } |
|
318 } |
|
319 |
|
320 |
|
321 int vp9_regulate_q(VP9_COMP *cpi, int target_bits_per_frame) { |
|
322 int q = cpi->active_worst_quality; |
|
323 |
|
324 int i; |
|
325 int last_error = INT_MAX; |
|
326 int target_bits_per_mb; |
|
327 int bits_per_mb_at_this_q; |
|
328 double correction_factor; |
|
329 |
|
330 // Select the appropriate correction factor based upon type of frame. |
|
331 if (cpi->common.frame_type == KEY_FRAME) { |
|
332 correction_factor = cpi->key_frame_rate_correction_factor; |
|
333 } else { |
|
334 if (cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) |
|
335 correction_factor = cpi->gf_rate_correction_factor; |
|
336 else |
|
337 correction_factor = cpi->rate_correction_factor; |
|
338 } |
|
339 |
|
340 // Calculate required scaling factor based on target frame size and size of |
|
341 // frame produced using previous Q. |
|
342 if (target_bits_per_frame >= (INT_MAX >> BPER_MB_NORMBITS)) |
|
343 target_bits_per_mb = |
|
344 (target_bits_per_frame / cpi->common.MBs) |
|
345 << BPER_MB_NORMBITS; // Case where we would overflow int |
|
346 else |
|
347 target_bits_per_mb = |
|
348 (target_bits_per_frame << BPER_MB_NORMBITS) / cpi->common.MBs; |
|
349 |
|
350 i = cpi->active_best_quality; |
|
351 |
|
352 do { |
|
353 bits_per_mb_at_this_q = (int)vp9_bits_per_mb(cpi->common.frame_type, i, |
|
354 correction_factor); |
|
355 |
|
356 if (bits_per_mb_at_this_q <= target_bits_per_mb) { |
|
357 if ((target_bits_per_mb - bits_per_mb_at_this_q) <= last_error) |
|
358 q = i; |
|
359 else |
|
360 q = i - 1; |
|
361 |
|
362 break; |
|
363 } else { |
|
364 last_error = bits_per_mb_at_this_q - target_bits_per_mb; |
|
365 } |
|
366 } while (++i <= cpi->active_worst_quality); |
|
367 |
|
368 return q; |
|
369 } |
|
370 |
|
371 |
|
372 static int estimate_keyframe_frequency(VP9_COMP *cpi) { |
|
373 int i; |
|
374 |
|
375 // Average key frame frequency |
|
376 int av_key_frame_frequency = 0; |
|
377 |
|
378 /* First key frame at start of sequence is a special case. We have no |
|
379 * frequency data. |
|
380 */ |
|
381 if (cpi->key_frame_count == 1) { |
|
382 /* Assume a default of 1 kf every 2 seconds, or the max kf interval, |
|
383 * whichever is smaller. |
|
384 */ |
|
385 int key_freq = cpi->oxcf.key_freq > 0 ? cpi->oxcf.key_freq : 1; |
|
386 av_key_frame_frequency = (int)cpi->output_framerate * 2; |
|
387 |
|
388 if (cpi->oxcf.auto_key && av_key_frame_frequency > key_freq) |
|
389 av_key_frame_frequency = cpi->oxcf.key_freq; |
|
390 |
|
391 cpi->prior_key_frame_distance[KEY_FRAME_CONTEXT - 1] |
|
392 = av_key_frame_frequency; |
|
393 } else { |
|
394 unsigned int total_weight = 0; |
|
395 int last_kf_interval = |
|
396 (cpi->frames_since_key > 0) ? cpi->frames_since_key : 1; |
|
397 |
|
398 /* reset keyframe context and calculate weighted average of last |
|
399 * KEY_FRAME_CONTEXT keyframes |
|
400 */ |
|
401 for (i = 0; i < KEY_FRAME_CONTEXT; i++) { |
|
402 if (i < KEY_FRAME_CONTEXT - 1) |
|
403 cpi->prior_key_frame_distance[i] |
|
404 = cpi->prior_key_frame_distance[i + 1]; |
|
405 else |
|
406 cpi->prior_key_frame_distance[i] = last_kf_interval; |
|
407 |
|
408 av_key_frame_frequency += prior_key_frame_weight[i] |
|
409 * cpi->prior_key_frame_distance[i]; |
|
410 total_weight += prior_key_frame_weight[i]; |
|
411 } |
|
412 |
|
413 av_key_frame_frequency /= total_weight; |
|
414 } |
|
415 return av_key_frame_frequency; |
|
416 } |
|
417 |
|
418 |
|
419 void vp9_adjust_key_frame_context(VP9_COMP *cpi) { |
|
420 // Clear down mmx registers to allow floating point in what follows |
|
421 vp9_clear_system_state(); |
|
422 |
|
423 cpi->frames_since_key = 0; |
|
424 cpi->key_frame_count++; |
|
425 } |
|
426 |
|
427 |
|
428 void vp9_compute_frame_size_bounds(VP9_COMP *cpi, int *frame_under_shoot_limit, |
|
429 int *frame_over_shoot_limit) { |
|
430 // Set-up bounds on acceptable frame size: |
|
431 if (cpi->oxcf.fixed_q >= 0) { |
|
432 // Fixed Q scenario: frame size never outranges target (there is no target!) |
|
433 *frame_under_shoot_limit = 0; |
|
434 *frame_over_shoot_limit = INT_MAX; |
|
435 } else { |
|
436 if (cpi->common.frame_type == KEY_FRAME) { |
|
437 *frame_over_shoot_limit = cpi->this_frame_target * 9 / 8; |
|
438 *frame_under_shoot_limit = cpi->this_frame_target * 7 / 8; |
|
439 } else { |
|
440 if (cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) { |
|
441 *frame_over_shoot_limit = cpi->this_frame_target * 9 / 8; |
|
442 *frame_under_shoot_limit = cpi->this_frame_target * 7 / 8; |
|
443 } else { |
|
444 // Stron overshoot limit for constrained quality |
|
445 if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) { |
|
446 *frame_over_shoot_limit = cpi->this_frame_target * 11 / 8; |
|
447 *frame_under_shoot_limit = cpi->this_frame_target * 2 / 8; |
|
448 } else { |
|
449 *frame_over_shoot_limit = cpi->this_frame_target * 11 / 8; |
|
450 *frame_under_shoot_limit = cpi->this_frame_target * 5 / 8; |
|
451 } |
|
452 } |
|
453 } |
|
454 |
|
455 // For very small rate targets where the fractional adjustment |
|
456 // (eg * 7/8) may be tiny make sure there is at least a minimum |
|
457 // range. |
|
458 *frame_over_shoot_limit += 200; |
|
459 *frame_under_shoot_limit -= 200; |
|
460 if (*frame_under_shoot_limit < 0) |
|
461 *frame_under_shoot_limit = 0; |
|
462 } |
|
463 } |
|
464 |
|
465 |
|
466 // return of 0 means drop frame |
|
467 int vp9_pick_frame_size(VP9_COMP *cpi) { |
|
468 VP9_COMMON *cm = &cpi->common; |
|
469 |
|
470 if (cm->frame_type == KEY_FRAME) |
|
471 calc_iframe_target_size(cpi); |
|
472 else |
|
473 calc_pframe_target_size(cpi); |
|
474 |
|
475 return 1; |
|
476 } |