/src/ffmpeg/libavcodec/ratecontrol.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * Rate control for video encoders |
3 | | * |
4 | | * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at> |
5 | | * |
6 | | * This file is part of FFmpeg. |
7 | | * |
8 | | * FFmpeg is free software; you can redistribute it and/or |
9 | | * modify it under the terms of the GNU Lesser General Public |
10 | | * License as published by the Free Software Foundation; either |
11 | | * version 2.1 of the License, or (at your option) any later version. |
12 | | * |
13 | | * FFmpeg is distributed in the hope that it will be useful, |
14 | | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
15 | | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
16 | | * Lesser General Public License for more details. |
17 | | * |
18 | | * You should have received a copy of the GNU Lesser General Public |
19 | | * License along with FFmpeg; if not, write to the Free Software |
20 | | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
21 | | */ |
22 | | |
23 | | /** |
24 | | * @file |
25 | | * Rate control for video encoders. |
26 | | */ |
27 | | |
28 | | #include "libavutil/attributes.h" |
29 | | #include "libavutil/emms.h" |
30 | | #include "libavutil/internal.h" |
31 | | #include "libavutil/mem.h" |
32 | | |
33 | | #include "avcodec.h" |
34 | | #include "ratecontrol.h" |
35 | | #include "mpegvideoenc.h" |
36 | | #include "libavutil/eval.h" |
37 | | |
38 | | void ff_write_pass1_stats(MpegEncContext *s) |
39 | 0 | { |
40 | 0 | snprintf(s->avctx->stats_out, 256, |
41 | 0 | "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d " |
42 | 0 | "fcode:%d bcode:%d mc-var:%"PRId64" var:%"PRId64" icount:%d hbits:%d;\n", |
43 | 0 | s->cur_pic.ptr->display_picture_number, |
44 | 0 | s->cur_pic.ptr->coded_picture_number, |
45 | 0 | s->pict_type, |
46 | 0 | s->cur_pic.ptr->f->quality, |
47 | 0 | s->i_tex_bits, |
48 | 0 | s->p_tex_bits, |
49 | 0 | s->mv_bits, |
50 | 0 | s->misc_bits, |
51 | 0 | s->f_code, |
52 | 0 | s->b_code, |
53 | 0 | s->mc_mb_var_sum, |
54 | 0 | s->mb_var_sum, |
55 | 0 | s->i_count, |
56 | 0 | s->header_bits); |
57 | 0 | } |
58 | | |
59 | | static AVRational get_fpsQ(AVCodecContext *avctx) |
60 | 195k | { |
61 | 195k | if (avctx->framerate.num > 0 && avctx->framerate.den > 0) |
62 | 183k | return avctx->framerate; |
63 | | |
64 | 12.2k | FF_DISABLE_DEPRECATION_WARNINGS |
65 | 12.2k | #if FF_API_TICKS_PER_FRAME |
66 | 12.2k | return av_div_q((AVRational){1, FFMAX(avctx->ticks_per_frame, 1)}, avctx->time_base); |
67 | | #else |
68 | | return av_inv_q(avctx->time_base); |
69 | | #endif |
70 | 195k | FF_ENABLE_DEPRECATION_WARNINGS |
71 | 195k | } |
72 | | |
73 | | static double get_fps(AVCodecContext *avctx) |
74 | 195k | { |
75 | 195k | return av_q2d(get_fpsQ(avctx)); |
76 | 195k | } |
77 | | |
78 | | static inline double qp2bits(const RateControlEntry *rce, double qp) |
79 | 0 | { |
80 | 0 | if (qp <= 0.0) { |
81 | 0 | av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n"); |
82 | 0 | } |
83 | 0 | return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp; |
84 | 0 | } |
85 | | |
86 | | static double qp2bits_cb(void *rce, double qp) |
87 | 0 | { |
88 | 0 | return qp2bits(rce, qp); |
89 | 0 | } |
90 | | |
91 | | static inline double bits2qp(const RateControlEntry *rce, double bits) |
92 | 93.5k | { |
93 | 93.5k | if (bits < 0.9) { |
94 | 0 | av_log(NULL, AV_LOG_ERROR, "bits<0.9\n"); |
95 | 0 | } |
96 | 93.5k | return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits; |
97 | 93.5k | } |
98 | | |
99 | | static double bits2qp_cb(void *rce, double qp) |
100 | 0 | { |
101 | 0 | return bits2qp(rce, qp); |
102 | 0 | } |
103 | | |
104 | | static double get_diff_limited_q(MpegEncContext *s, const RateControlEntry *rce, double q) |
105 | 93.5k | { |
106 | 93.5k | RateControlContext *rcc = &s->rc_context; |
107 | 93.5k | AVCodecContext *a = s->avctx; |
108 | 93.5k | const int pict_type = rce->new_pict_type; |
109 | 93.5k | const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P]; |
110 | 93.5k | const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type]; |
111 | | |
112 | 93.5k | if (pict_type == AV_PICTURE_TYPE_I && |
113 | 93.5k | (a->i_quant_factor > 0.0 || rcc->last_non_b_pict_type == AV_PICTURE_TYPE_P)) |
114 | 2.77k | q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset; |
115 | 90.8k | else if (pict_type == AV_PICTURE_TYPE_B && |
116 | 90.8k | a->b_quant_factor > 0.0) |
117 | 0 | q = last_non_b_q * a->b_quant_factor + a->b_quant_offset; |
118 | 93.5k | if (q < 1) |
119 | 949 | q = 1; |
120 | | |
121 | | /* last qscale / qdiff stuff */ |
122 | 93.5k | if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) { |
123 | 79.3k | double last_q = rcc->last_qscale_for[pict_type]; |
124 | 79.3k | const int maxdiff = FF_QP2LAMBDA * a->max_qdiff; |
125 | | |
126 | 79.3k | if (q > last_q + maxdiff) |
127 | 25.7k | q = last_q + maxdiff; |
128 | 53.5k | else if (q < last_q - maxdiff) |
129 | 5.50k | q = last_q - maxdiff; |
130 | 79.3k | } |
131 | | |
132 | 93.5k | rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring |
133 | | |
134 | 93.5k | if (pict_type != AV_PICTURE_TYPE_B) |
135 | 93.5k | rcc->last_non_b_pict_type = pict_type; |
136 | | |
137 | 93.5k | return q; |
138 | 93.5k | } |
139 | | |
140 | | /** |
141 | | * Get the qmin & qmax for pict_type. |
142 | | */ |
143 | | static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type) |
144 | 187k | { |
145 | 187k | int qmin = s->lmin; |
146 | 187k | int qmax = s->lmax; |
147 | | |
148 | 187k | av_assert0(qmin <= qmax); |
149 | | |
150 | 187k | switch (pict_type) { |
151 | 0 | case AV_PICTURE_TYPE_B: |
152 | 0 | qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5); |
153 | 0 | qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5); |
154 | 0 | break; |
155 | 66.8k | case AV_PICTURE_TYPE_I: |
156 | 66.8k | qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5); |
157 | 66.8k | qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5); |
158 | 66.8k | break; |
159 | 187k | } |
160 | | |
161 | 187k | qmin = av_clip(qmin, 1, FF_LAMBDA_MAX); |
162 | 187k | qmax = av_clip(qmax, 1, FF_LAMBDA_MAX); |
163 | | |
164 | 187k | if (qmax < qmin) |
165 | 0 | qmax = qmin; |
166 | | |
167 | 187k | *qmin_ret = qmin; |
168 | 187k | *qmax_ret = qmax; |
169 | 187k | } |
170 | | |
171 | | static double modify_qscale(MpegEncContext *s, const RateControlEntry *rce, |
172 | | double q, int frame_num) |
173 | 93.5k | { |
174 | 93.5k | RateControlContext *rcc = &s->rc_context; |
175 | 93.5k | const double buffer_size = s->avctx->rc_buffer_size; |
176 | 93.5k | const double fps = get_fps(s->avctx); |
177 | 93.5k | const double min_rate = s->avctx->rc_min_rate / fps; |
178 | 93.5k | const double max_rate = s->avctx->rc_max_rate / fps; |
179 | 93.5k | const int pict_type = rce->new_pict_type; |
180 | 93.5k | int qmin, qmax; |
181 | | |
182 | 93.5k | get_qminmax(&qmin, &qmax, s, pict_type); |
183 | | |
184 | | /* modulation */ |
185 | 93.5k | if (s->rc_qmod_freq && |
186 | 93.5k | frame_num % s->rc_qmod_freq == 0 && |
187 | 93.5k | pict_type == AV_PICTURE_TYPE_P) |
188 | 0 | q *= s->rc_qmod_amp; |
189 | | |
190 | | /* buffer overflow/underflow protection */ |
191 | 93.5k | if (buffer_size) { |
192 | 0 | double expected_size = rcc->buffer_index; |
193 | 0 | double q_limit; |
194 | |
|
195 | 0 | if (min_rate) { |
196 | 0 | double d = 2 * (buffer_size - expected_size) / buffer_size; |
197 | 0 | if (d > 1.0) |
198 | 0 | d = 1.0; |
199 | 0 | else if (d < 0.0001) |
200 | 0 | d = 0.0001; |
201 | 0 | q *= pow(d, 1.0 / s->rc_buffer_aggressivity); |
202 | |
|
203 | 0 | q_limit = bits2qp(rce, |
204 | 0 | FFMAX((min_rate - buffer_size + rcc->buffer_index) * |
205 | 0 | s->avctx->rc_min_vbv_overflow_use, 1)); |
206 | |
|
207 | 0 | if (q > q_limit) { |
208 | 0 | if (s->avctx->debug & FF_DEBUG_RC) |
209 | 0 | av_log(s->avctx, AV_LOG_DEBUG, |
210 | 0 | "limiting QP %f -> %f\n", q, q_limit); |
211 | 0 | q = q_limit; |
212 | 0 | } |
213 | 0 | } |
214 | |
|
215 | 0 | if (max_rate) { |
216 | 0 | double d = 2 * expected_size / buffer_size; |
217 | 0 | if (d > 1.0) |
218 | 0 | d = 1.0; |
219 | 0 | else if (d < 0.0001) |
220 | 0 | d = 0.0001; |
221 | 0 | q /= pow(d, 1.0 / s->rc_buffer_aggressivity); |
222 | |
|
223 | 0 | q_limit = bits2qp(rce, |
224 | 0 | FFMAX(rcc->buffer_index * |
225 | 0 | s->avctx->rc_max_available_vbv_use, |
226 | 0 | 1)); |
227 | 0 | if (q < q_limit) { |
228 | 0 | if (s->avctx->debug & FF_DEBUG_RC) |
229 | 0 | av_log(s->avctx, AV_LOG_DEBUG, |
230 | 0 | "limiting QP %f -> %f\n", q, q_limit); |
231 | 0 | q = q_limit; |
232 | 0 | } |
233 | 0 | } |
234 | 0 | } |
235 | 93.5k | ff_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n", |
236 | 93.5k | q, max_rate, min_rate, buffer_size, rcc->buffer_index, |
237 | 93.5k | s->rc_buffer_aggressivity); |
238 | 93.5k | if (s->rc_qsquish == 0.0 || qmin == qmax) { |
239 | 93.5k | if (q < qmin) |
240 | 43.5k | q = qmin; |
241 | 49.9k | else if (q > qmax) |
242 | 19.9k | q = qmax; |
243 | 93.5k | } else { |
244 | 0 | double min2 = log(qmin); |
245 | 0 | double max2 = log(qmax); |
246 | |
|
247 | 0 | q = log(q); |
248 | 0 | q = (q - min2) / (max2 - min2) - 0.5; |
249 | 0 | q *= -4.0; |
250 | 0 | q = 1.0 / (1.0 + exp(q)); |
251 | 0 | q = q * (max2 - min2) + min2; |
252 | |
|
253 | 0 | q = exp(q); |
254 | 0 | } |
255 | | |
256 | 93.5k | return q; |
257 | 93.5k | } |
258 | | |
259 | | /** |
260 | | * Modify the bitrate curve from pass1 for one frame. |
261 | | */ |
262 | | static double get_qscale(MpegEncContext *s, RateControlEntry *rce, |
263 | | double rate_factor, int frame_num) |
264 | 93.5k | { |
265 | 93.5k | RateControlContext *rcc = &s->rc_context; |
266 | 93.5k | AVCodecContext *a = s->avctx; |
267 | 93.5k | const int pict_type = rce->new_pict_type; |
268 | 93.5k | const double mb_num = s->mb_num; |
269 | 93.5k | double q, bits; |
270 | 93.5k | int i; |
271 | | |
272 | 93.5k | double const_values[] = { |
273 | 93.5k | M_PI, |
274 | 93.5k | M_E, |
275 | 93.5k | rce->i_tex_bits * rce->qscale, |
276 | 93.5k | rce->p_tex_bits * rce->qscale, |
277 | 93.5k | (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale, |
278 | 93.5k | rce->mv_bits / mb_num, |
279 | 93.5k | rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code, |
280 | 93.5k | rce->i_count / mb_num, |
281 | 93.5k | rce->mc_mb_var_sum / mb_num, |
282 | 93.5k | rce->mb_var_sum / mb_num, |
283 | 93.5k | rce->pict_type == AV_PICTURE_TYPE_I, |
284 | 93.5k | rce->pict_type == AV_PICTURE_TYPE_P, |
285 | 93.5k | rce->pict_type == AV_PICTURE_TYPE_B, |
286 | 93.5k | rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type], |
287 | 93.5k | a->qcompress, |
288 | 93.5k | rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I], |
289 | 93.5k | rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P], |
290 | 93.5k | rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P], |
291 | 93.5k | rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B], |
292 | 93.5k | (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type], |
293 | 93.5k | 0 |
294 | 93.5k | }; |
295 | | |
296 | 93.5k | bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce); |
297 | 93.5k | if (isnan(bits)) { |
298 | 0 | av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->rc_eq); |
299 | 0 | return -1; |
300 | 0 | } |
301 | | |
302 | 93.5k | rcc->pass1_rc_eq_output_sum += bits; |
303 | 93.5k | bits *= rate_factor; |
304 | 93.5k | if (bits < 0.0) |
305 | 28.3k | bits = 0.0; |
306 | 93.5k | bits += 1.0; // avoid 1/0 issues |
307 | | |
308 | | /* user override */ |
309 | 93.5k | for (i = 0; i < s->avctx->rc_override_count; i++) { |
310 | 0 | RcOverride *rco = s->avctx->rc_override; |
311 | 0 | if (rco[i].start_frame > frame_num) |
312 | 0 | continue; |
313 | 0 | if (rco[i].end_frame < frame_num) |
314 | 0 | continue; |
315 | | |
316 | 0 | if (rco[i].qscale) |
317 | 0 | bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it? |
318 | 0 | else |
319 | 0 | bits *= rco[i].quality_factor; |
320 | 0 | } |
321 | | |
322 | 93.5k | q = bits2qp(rce, bits); |
323 | | |
324 | | /* I/B difference */ |
325 | 93.5k | if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0) |
326 | 33.4k | q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset; |
327 | 60.1k | else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0) |
328 | 0 | q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset; |
329 | 93.5k | if (q < 1) |
330 | 24.6k | q = 1; |
331 | | |
332 | 93.5k | return q; |
333 | 93.5k | } |
334 | | |
335 | | static int init_pass2(MpegEncContext *s) |
336 | 0 | { |
337 | 0 | RateControlContext *rcc = &s->rc_context; |
338 | 0 | AVCodecContext *a = s->avctx; |
339 | 0 | int i, toobig; |
340 | 0 | AVRational fps = get_fpsQ(s->avctx); |
341 | 0 | double complexity[5] = { 0 }; // approximate bits at quant=1 |
342 | 0 | uint64_t const_bits[5] = { 0 }; // quantizer independent bits |
343 | 0 | uint64_t all_const_bits; |
344 | 0 | uint64_t all_available_bits = av_rescale_q(s->bit_rate, |
345 | 0 | (AVRational){rcc->num_entries,1}, |
346 | 0 | fps); |
347 | 0 | double rate_factor = 0; |
348 | 0 | double step; |
349 | 0 | const int filter_size = (int)(a->qblur * 4) | 1; |
350 | 0 | double expected_bits = 0; // init to silence gcc warning |
351 | 0 | double *qscale, *blurred_qscale, qscale_sum; |
352 | | |
353 | | /* find complexity & const_bits & decide the pict_types */ |
354 | 0 | for (i = 0; i < rcc->num_entries; i++) { |
355 | 0 | RateControlEntry *rce = &rcc->entry[i]; |
356 | |
|
357 | 0 | rce->new_pict_type = rce->pict_type; |
358 | 0 | rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale; |
359 | 0 | rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale; |
360 | 0 | rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits; |
361 | 0 | rcc->frame_count[rce->pict_type]++; |
362 | |
|
363 | 0 | complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) * |
364 | 0 | (double)rce->qscale; |
365 | 0 | const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits; |
366 | 0 | } |
367 | |
|
368 | 0 | all_const_bits = const_bits[AV_PICTURE_TYPE_I] + |
369 | 0 | const_bits[AV_PICTURE_TYPE_P] + |
370 | 0 | const_bits[AV_PICTURE_TYPE_B]; |
371 | |
|
372 | 0 | if (all_available_bits < all_const_bits) { |
373 | 0 | av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n"); |
374 | 0 | return -1; |
375 | 0 | } |
376 | | |
377 | 0 | qscale = av_malloc_array(rcc->num_entries, sizeof(double)); |
378 | 0 | blurred_qscale = av_malloc_array(rcc->num_entries, sizeof(double)); |
379 | 0 | if (!qscale || !blurred_qscale) { |
380 | 0 | av_free(qscale); |
381 | 0 | av_free(blurred_qscale); |
382 | 0 | return AVERROR(ENOMEM); |
383 | 0 | } |
384 | 0 | toobig = 0; |
385 | |
|
386 | 0 | for (step = 256 * 256; step > 0.0000001; step *= 0.5) { |
387 | 0 | expected_bits = 0; |
388 | 0 | rate_factor += step; |
389 | |
|
390 | 0 | rcc->buffer_index = s->avctx->rc_buffer_size / 2; |
391 | | |
392 | | /* find qscale */ |
393 | 0 | for (i = 0; i < rcc->num_entries; i++) { |
394 | 0 | const RateControlEntry *rce = &rcc->entry[i]; |
395 | |
|
396 | 0 | qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i); |
397 | 0 | rcc->last_qscale_for[rce->pict_type] = qscale[i]; |
398 | 0 | } |
399 | 0 | av_assert0(filter_size % 2 == 1); |
400 | | |
401 | | /* fixed I/B QP relative to P mode */ |
402 | 0 | for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) { |
403 | 0 | const RateControlEntry *rce = &rcc->entry[i]; |
404 | |
|
405 | 0 | qscale[i] = get_diff_limited_q(s, rce, qscale[i]); |
406 | 0 | } |
407 | |
|
408 | 0 | for (i = rcc->num_entries - 1; i >= 0; i--) { |
409 | 0 | const RateControlEntry *rce = &rcc->entry[i]; |
410 | |
|
411 | 0 | qscale[i] = get_diff_limited_q(s, rce, qscale[i]); |
412 | 0 | } |
413 | | |
414 | | /* smooth curve */ |
415 | 0 | for (i = 0; i < rcc->num_entries; i++) { |
416 | 0 | const RateControlEntry *rce = &rcc->entry[i]; |
417 | 0 | const int pict_type = rce->new_pict_type; |
418 | 0 | int j; |
419 | 0 | double q = 0.0, sum = 0.0; |
420 | |
|
421 | 0 | for (j = 0; j < filter_size; j++) { |
422 | 0 | int index = i + j - filter_size / 2; |
423 | 0 | double d = index - i; |
424 | 0 | double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur)); |
425 | |
|
426 | 0 | if (index < 0 || index >= rcc->num_entries) |
427 | 0 | continue; |
428 | 0 | if (pict_type != rcc->entry[index].new_pict_type) |
429 | 0 | continue; |
430 | 0 | q += qscale[index] * coeff; |
431 | 0 | sum += coeff; |
432 | 0 | } |
433 | 0 | blurred_qscale[i] = q / sum; |
434 | 0 | } |
435 | | |
436 | | /* find expected bits */ |
437 | 0 | for (i = 0; i < rcc->num_entries; i++) { |
438 | 0 | RateControlEntry *rce = &rcc->entry[i]; |
439 | 0 | double bits; |
440 | |
|
441 | 0 | rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i); |
442 | |
|
443 | 0 | bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits; |
444 | 0 | bits += 8 * ff_vbv_update(s, bits); |
445 | |
|
446 | 0 | rce->expected_bits = expected_bits; |
447 | 0 | expected_bits += bits; |
448 | 0 | } |
449 | |
|
450 | 0 | ff_dlog(s->avctx, |
451 | 0 | "expected_bits: %f all_available_bits: %d rate_factor: %f\n", |
452 | 0 | expected_bits, (int)all_available_bits, rate_factor); |
453 | 0 | if (expected_bits > all_available_bits) { |
454 | 0 | rate_factor -= step; |
455 | 0 | ++toobig; |
456 | 0 | } |
457 | 0 | } |
458 | 0 | av_free(qscale); |
459 | 0 | av_free(blurred_qscale); |
460 | | |
461 | | /* check bitrate calculations and print info */ |
462 | 0 | qscale_sum = 0.0; |
463 | 0 | for (i = 0; i < rcc->num_entries; i++) { |
464 | 0 | ff_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n", |
465 | 0 | i, |
466 | 0 | rcc->entry[i].new_qscale, |
467 | 0 | rcc->entry[i].new_qscale / FF_QP2LAMBDA); |
468 | 0 | qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, |
469 | 0 | s->avctx->qmin, s->avctx->qmax); |
470 | 0 | } |
471 | 0 | av_assert0(toobig <= 40); |
472 | 0 | av_log(s->avctx, AV_LOG_DEBUG, |
473 | 0 | "[lavc rc] requested bitrate: %"PRId64" bps expected bitrate: %"PRId64" bps\n", |
474 | 0 | s->bit_rate, |
475 | 0 | (int64_t)(expected_bits / ((double)all_available_bits / s->bit_rate))); |
476 | 0 | av_log(s->avctx, AV_LOG_DEBUG, |
477 | 0 | "[lavc rc] estimated target average qp: %.3f\n", |
478 | 0 | (float)qscale_sum / rcc->num_entries); |
479 | 0 | if (toobig == 0) { |
480 | 0 | av_log(s->avctx, AV_LOG_INFO, |
481 | 0 | "[lavc rc] Using all of requested bitrate is not " |
482 | 0 | "necessary for this video with these parameters.\n"); |
483 | 0 | } else if (toobig == 40) { |
484 | 0 | av_log(s->avctx, AV_LOG_ERROR, |
485 | 0 | "[lavc rc] Error: bitrate too low for this video " |
486 | 0 | "with these parameters.\n"); |
487 | 0 | return -1; |
488 | 0 | } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) { |
489 | 0 | av_log(s->avctx, AV_LOG_ERROR, |
490 | 0 | "[lavc rc] Error: 2pass curve failed to converge\n"); |
491 | 0 | return -1; |
492 | 0 | } |
493 | | |
494 | 0 | return 0; |
495 | 0 | } |
496 | | |
497 | | av_cold int ff_rate_control_init(MpegEncContext *s) |
498 | 11.9k | { |
499 | 11.9k | RateControlContext *rcc = &s->rc_context; |
500 | 11.9k | int i, res; |
501 | 11.9k | static const char * const const_names[] = { |
502 | 11.9k | "PI", |
503 | 11.9k | "E", |
504 | 11.9k | "iTex", |
505 | 11.9k | "pTex", |
506 | 11.9k | "tex", |
507 | 11.9k | "mv", |
508 | 11.9k | "fCode", |
509 | 11.9k | "iCount", |
510 | 11.9k | "mcVar", |
511 | 11.9k | "var", |
512 | 11.9k | "isI", |
513 | 11.9k | "isP", |
514 | 11.9k | "isB", |
515 | 11.9k | "avgQP", |
516 | 11.9k | "qComp", |
517 | 11.9k | "avgIITex", |
518 | 11.9k | "avgPITex", |
519 | 11.9k | "avgPPTex", |
520 | 11.9k | "avgBPTex", |
521 | 11.9k | "avgTex", |
522 | 11.9k | NULL |
523 | 11.9k | }; |
524 | 11.9k | static double (* const func1[])(void *, double) = { |
525 | 11.9k | bits2qp_cb, |
526 | 11.9k | qp2bits_cb, |
527 | 11.9k | NULL |
528 | 11.9k | }; |
529 | 11.9k | static const char * const func1_names[] = { |
530 | 11.9k | "bits2qp", |
531 | 11.9k | "qp2bits", |
532 | 11.9k | NULL |
533 | 11.9k | }; |
534 | 11.9k | emms_c(); |
535 | | |
536 | 11.9k | if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) { |
537 | 0 | if (s->avctx->rc_max_rate) { |
538 | 0 | s->avctx->rc_max_available_vbv_use = av_clipf(s->avctx->rc_max_rate/(s->avctx->rc_buffer_size*get_fps(s->avctx)), 1.0/3, 1.0); |
539 | 0 | } else |
540 | 0 | s->avctx->rc_max_available_vbv_use = 1.0; |
541 | 0 | } |
542 | | |
543 | 11.9k | res = av_expr_parse(&rcc->rc_eq_eval, |
544 | 11.9k | s->rc_eq ? s->rc_eq : "tex^qComp", |
545 | 11.9k | const_names, func1_names, func1, |
546 | 11.9k | NULL, NULL, 0, s->avctx); |
547 | 11.9k | if (res < 0) { |
548 | 0 | av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->rc_eq); |
549 | 0 | return res; |
550 | 0 | } |
551 | | |
552 | 71.7k | for (i = 0; i < 5; i++) { |
553 | 59.7k | rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0; |
554 | 59.7k | rcc->pred[i].count = 1.0; |
555 | 59.7k | rcc->pred[i].decay = 0.4; |
556 | | |
557 | 59.7k | rcc->i_cplx_sum [i] = |
558 | 59.7k | rcc->p_cplx_sum [i] = |
559 | 59.7k | rcc->mv_bits_sum[i] = |
560 | 59.7k | rcc->qscale_sum [i] = |
561 | 59.7k | rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such |
562 | | |
563 | 59.7k | rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5; |
564 | 59.7k | } |
565 | 11.9k | rcc->buffer_index = s->avctx->rc_initial_buffer_occupancy; |
566 | 11.9k | if (!rcc->buffer_index) |
567 | 11.9k | rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4; |
568 | | |
569 | 11.9k | if (s->avctx->flags & AV_CODEC_FLAG_PASS2) { |
570 | 0 | int i; |
571 | 0 | char *p; |
572 | | |
573 | | /* find number of pics */ |
574 | 0 | p = s->avctx->stats_in; |
575 | 0 | for (i = -1; p; i++) |
576 | 0 | p = strchr(p + 1, ';'); |
577 | 0 | i += s->max_b_frames; |
578 | 0 | if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry)) |
579 | 0 | return -1; |
580 | 0 | rcc->entry = av_mallocz(i * sizeof(RateControlEntry)); |
581 | 0 | if (!rcc->entry) |
582 | 0 | return AVERROR(ENOMEM); |
583 | 0 | rcc->num_entries = i; |
584 | | |
585 | | /* init all to skipped P-frames |
586 | | * (with B-frames we might have a not encoded frame at the end FIXME) */ |
587 | 0 | for (i = 0; i < rcc->num_entries; i++) { |
588 | 0 | RateControlEntry *rce = &rcc->entry[i]; |
589 | |
|
590 | 0 | rce->pict_type = rce->new_pict_type = AV_PICTURE_TYPE_P; |
591 | 0 | rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2; |
592 | 0 | rce->misc_bits = s->mb_num + 10; |
593 | 0 | rce->mb_var_sum = s->mb_num * 100; |
594 | 0 | } |
595 | | |
596 | | /* read stats */ |
597 | 0 | p = s->avctx->stats_in; |
598 | 0 | for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) { |
599 | 0 | RateControlEntry *rce; |
600 | 0 | int picture_number; |
601 | 0 | int e; |
602 | 0 | char *next; |
603 | |
|
604 | 0 | next = strchr(p, ';'); |
605 | 0 | if (next) { |
606 | 0 | (*next) = 0; // sscanf is unbelievably slow on looong strings // FIXME copy / do not write |
607 | 0 | next++; |
608 | 0 | } |
609 | 0 | e = sscanf(p, " in:%d ", &picture_number); |
610 | |
|
611 | 0 | av_assert0(picture_number >= 0); |
612 | 0 | av_assert0(picture_number < rcc->num_entries); |
613 | 0 | rce = &rcc->entry[picture_number]; |
614 | |
|
615 | 0 | e += sscanf(p, " in:%*d out:%*d type:%d q:%f itex:%d ptex:%d " |
616 | 0 | "mv:%d misc:%d " |
617 | 0 | "fcode:%d bcode:%d " |
618 | 0 | "mc-var:%"SCNd64" var:%"SCNd64" " |
619 | 0 | "icount:%d hbits:%d", |
620 | 0 | &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, |
621 | 0 | &rce->mv_bits, &rce->misc_bits, |
622 | 0 | &rce->f_code, &rce->b_code, |
623 | 0 | &rce->mc_mb_var_sum, &rce->mb_var_sum, |
624 | 0 | &rce->i_count, &rce->header_bits); |
625 | 0 | if (e != 13) { |
626 | 0 | av_log(s->avctx, AV_LOG_ERROR, |
627 | 0 | "statistics are damaged at line %d, parser out=%d\n", |
628 | 0 | i, e); |
629 | 0 | return -1; |
630 | 0 | } |
631 | | |
632 | 0 | p = next; |
633 | 0 | } |
634 | | |
635 | 0 | res = init_pass2(s); |
636 | 0 | if (res < 0) |
637 | 0 | return res; |
638 | 0 | } |
639 | | |
640 | 11.9k | if (!(s->avctx->flags & AV_CODEC_FLAG_PASS2)) { |
641 | 11.9k | rcc->short_term_qsum = 0.001; |
642 | 11.9k | rcc->short_term_qcount = 0.001; |
643 | | |
644 | 11.9k | rcc->pass1_rc_eq_output_sum = 0.001; |
645 | 11.9k | rcc->pass1_wanted_bits = 0.001; |
646 | | |
647 | 11.9k | if (s->avctx->qblur > 1.0) { |
648 | 0 | av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n"); |
649 | 0 | return -1; |
650 | 0 | } |
651 | | /* init stuff with the user specified complexity */ |
652 | 11.9k | if (s->rc_initial_cplx) { |
653 | 0 | for (i = 0; i < 60 * 30; i++) { |
654 | 0 | double bits = s->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num; |
655 | 0 | RateControlEntry rce; |
656 | |
|
657 | 0 | if (i % ((s->gop_size + 3) / 4) == 0) |
658 | 0 | rce.pict_type = AV_PICTURE_TYPE_I; |
659 | 0 | else if (i % (s->max_b_frames + 1)) |
660 | 0 | rce.pict_type = AV_PICTURE_TYPE_B; |
661 | 0 | else |
662 | 0 | rce.pict_type = AV_PICTURE_TYPE_P; |
663 | |
|
664 | 0 | rce.new_pict_type = rce.pict_type; |
665 | 0 | rce.mc_mb_var_sum = bits * s->mb_num / 100000; |
666 | 0 | rce.mb_var_sum = s->mb_num; |
667 | |
|
668 | 0 | rce.qscale = FF_QP2LAMBDA * 2; |
669 | 0 | rce.f_code = 2; |
670 | 0 | rce.b_code = 1; |
671 | 0 | rce.misc_bits = 1; |
672 | |
|
673 | 0 | if (s->pict_type == AV_PICTURE_TYPE_I) { |
674 | 0 | rce.i_count = s->mb_num; |
675 | 0 | rce.i_tex_bits = bits; |
676 | 0 | rce.p_tex_bits = 0; |
677 | 0 | rce.mv_bits = 0; |
678 | 0 | } else { |
679 | 0 | rce.i_count = 0; // FIXME we do know this approx |
680 | 0 | rce.i_tex_bits = 0; |
681 | 0 | rce.p_tex_bits = bits * 0.9; |
682 | 0 | rce.mv_bits = bits * 0.1; |
683 | 0 | } |
684 | 0 | rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale; |
685 | 0 | rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale; |
686 | 0 | rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits; |
687 | 0 | rcc->frame_count[rce.pict_type]++; |
688 | |
|
689 | 0 | get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i); |
690 | | |
691 | | // FIXME misbehaves a little for variable fps |
692 | 0 | rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx); |
693 | 0 | } |
694 | 0 | } |
695 | 11.9k | } |
696 | | |
697 | 11.9k | return 0; |
698 | 11.9k | } |
699 | | |
700 | | av_cold void ff_rate_control_uninit(RateControlContext *rcc) |
701 | 13.4k | { |
702 | 13.4k | emms_c(); |
703 | | |
704 | 13.4k | av_expr_free(rcc->rc_eq_eval); |
705 | 13.4k | rcc->rc_eq_eval = NULL; |
706 | 13.4k | av_freep(&rcc->entry); |
707 | 13.4k | } |
708 | | |
709 | | int ff_vbv_update(MpegEncContext *s, int frame_size) |
710 | 8.40k | { |
711 | 8.40k | RateControlContext *rcc = &s->rc_context; |
712 | 8.40k | const double fps = get_fps(s->avctx); |
713 | 8.40k | const int buffer_size = s->avctx->rc_buffer_size; |
714 | 8.40k | const double min_rate = s->avctx->rc_min_rate / fps; |
715 | 8.40k | const double max_rate = s->avctx->rc_max_rate / fps; |
716 | | |
717 | 8.40k | ff_dlog(s, "%d %f %d %f %f\n", |
718 | 8.40k | buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate); |
719 | | |
720 | 8.40k | if (buffer_size) { |
721 | 0 | int left; |
722 | |
|
723 | 0 | rcc->buffer_index -= frame_size; |
724 | 0 | if (rcc->buffer_index < 0) { |
725 | 0 | av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n"); |
726 | 0 | if (frame_size > max_rate && s->qscale == s->avctx->qmax) { |
727 | 0 | av_log(s->avctx, AV_LOG_ERROR, "max bitrate possibly too small or try trellis with large lmax or increase qmax\n"); |
728 | 0 | } |
729 | 0 | rcc->buffer_index = 0; |
730 | 0 | } |
731 | |
|
732 | 0 | left = buffer_size - rcc->buffer_index - 1; |
733 | 0 | rcc->buffer_index += av_clip(left, min_rate, max_rate); |
734 | |
|
735 | 0 | if (rcc->buffer_index > buffer_size) { |
736 | 0 | int stuffing = ceil((rcc->buffer_index - buffer_size) / 8); |
737 | |
|
738 | 0 | if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4) |
739 | 0 | stuffing = 4; |
740 | 0 | rcc->buffer_index -= 8 * stuffing; |
741 | |
|
742 | 0 | if (s->avctx->debug & FF_DEBUG_RC) |
743 | 0 | av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing); |
744 | |
|
745 | 0 | return stuffing; |
746 | 0 | } |
747 | 0 | } |
748 | 8.40k | return 0; |
749 | 8.40k | } |
750 | | |
751 | | static double predict_size(Predictor *p, double q, double var) |
752 | 93.5k | { |
753 | 93.5k | return p->coeff * var / (q * p->count); |
754 | 93.5k | } |
755 | | |
756 | | static void update_predictor(Predictor *p, double q, double var, double size) |
757 | 35.5k | { |
758 | 35.5k | double new_coeff = size * q / (var + 1); |
759 | 35.5k | if (var < 10) |
760 | 24.3k | return; |
761 | | |
762 | 11.2k | p->count *= p->decay; |
763 | 11.2k | p->coeff *= p->decay; |
764 | 11.2k | p->count++; |
765 | 11.2k | p->coeff += new_coeff; |
766 | 11.2k | } |
767 | | |
768 | | static void adaptive_quantization(MpegEncContext *s, double q) |
769 | 0 | { |
770 | 0 | int i; |
771 | 0 | const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0); |
772 | 0 | const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0); |
773 | 0 | const float temp_cplx_masking = s->avctx->temporal_cplx_masking; |
774 | 0 | const float spatial_cplx_masking = s->avctx->spatial_cplx_masking; |
775 | 0 | const float p_masking = s->avctx->p_masking; |
776 | 0 | const float border_masking = s->border_masking; |
777 | 0 | float bits_sum = 0.0; |
778 | 0 | float cplx_sum = 0.0; |
779 | 0 | float *cplx_tab = s->cplx_tab; |
780 | 0 | float *bits_tab = s->bits_tab; |
781 | 0 | const int qmin = s->avctx->mb_lmin; |
782 | 0 | const int qmax = s->avctx->mb_lmax; |
783 | 0 | const int mb_width = s->mb_width; |
784 | 0 | const int mb_height = s->mb_height; |
785 | |
|
786 | 0 | for (i = 0; i < s->mb_num; i++) { |
787 | 0 | const int mb_xy = s->mb_index2xy[i]; |
788 | 0 | float temp_cplx = sqrt(s->mc_mb_var[mb_xy]); // FIXME merge in pow() |
789 | 0 | float spat_cplx = sqrt(s->mb_var[mb_xy]); |
790 | 0 | const int lumi = s->mb_mean[mb_xy]; |
791 | 0 | float bits, cplx, factor; |
792 | 0 | int mb_x = mb_xy % s->mb_stride; |
793 | 0 | int mb_y = mb_xy / s->mb_stride; |
794 | 0 | int mb_distance; |
795 | 0 | float mb_factor = 0.0; |
796 | 0 | if (spat_cplx < 4) |
797 | 0 | spat_cplx = 4; // FIXME fine-tune |
798 | 0 | if (temp_cplx < 4) |
799 | 0 | temp_cplx = 4; // FIXME fine-tune |
800 | |
|
801 | 0 | if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode |
802 | 0 | cplx = spat_cplx; |
803 | 0 | factor = 1.0 + p_masking; |
804 | 0 | } else { |
805 | 0 | cplx = temp_cplx; |
806 | 0 | factor = pow(temp_cplx, -temp_cplx_masking); |
807 | 0 | } |
808 | 0 | factor *= pow(spat_cplx, -spatial_cplx_masking); |
809 | |
|
810 | 0 | if (lumi > 127) |
811 | 0 | factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking); |
812 | 0 | else |
813 | 0 | factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking); |
814 | |
|
815 | 0 | if (mb_x < mb_width / 5) { |
816 | 0 | mb_distance = mb_width / 5 - mb_x; |
817 | 0 | mb_factor = (float)mb_distance / (float)(mb_width / 5); |
818 | 0 | } else if (mb_x > 4 * mb_width / 5) { |
819 | 0 | mb_distance = mb_x - 4 * mb_width / 5; |
820 | 0 | mb_factor = (float)mb_distance / (float)(mb_width / 5); |
821 | 0 | } |
822 | 0 | if (mb_y < mb_height / 5) { |
823 | 0 | mb_distance = mb_height / 5 - mb_y; |
824 | 0 | mb_factor = FFMAX(mb_factor, |
825 | 0 | (float)mb_distance / (float)(mb_height / 5)); |
826 | 0 | } else if (mb_y > 4 * mb_height / 5) { |
827 | 0 | mb_distance = mb_y - 4 * mb_height / 5; |
828 | 0 | mb_factor = FFMAX(mb_factor, |
829 | 0 | (float)mb_distance / (float)(mb_height / 5)); |
830 | 0 | } |
831 | |
|
832 | 0 | factor *= 1.0 - border_masking * mb_factor; |
833 | |
|
834 | 0 | if (factor < 0.00001) |
835 | 0 | factor = 0.00001; |
836 | |
|
837 | 0 | bits = cplx * factor; |
838 | 0 | cplx_sum += cplx; |
839 | 0 | bits_sum += bits; |
840 | 0 | cplx_tab[i] = cplx; |
841 | 0 | bits_tab[i] = bits; |
842 | 0 | } |
843 | | |
844 | | /* handle qmin/qmax clipping */ |
845 | 0 | if (s->mpv_flags & FF_MPV_FLAG_NAQ) { |
846 | 0 | float factor = bits_sum / cplx_sum; |
847 | 0 | for (i = 0; i < s->mb_num; i++) { |
848 | 0 | float newq = q * cplx_tab[i] / bits_tab[i]; |
849 | 0 | newq *= factor; |
850 | |
|
851 | 0 | if (newq > qmax) { |
852 | 0 | bits_sum -= bits_tab[i]; |
853 | 0 | cplx_sum -= cplx_tab[i] * q / qmax; |
854 | 0 | } else if (newq < qmin) { |
855 | 0 | bits_sum -= bits_tab[i]; |
856 | 0 | cplx_sum -= cplx_tab[i] * q / qmin; |
857 | 0 | } |
858 | 0 | } |
859 | 0 | if (bits_sum < 0.001) |
860 | 0 | bits_sum = 0.001; |
861 | 0 | if (cplx_sum < 0.001) |
862 | 0 | cplx_sum = 0.001; |
863 | 0 | } |
864 | |
|
865 | 0 | for (i = 0; i < s->mb_num; i++) { |
866 | 0 | const int mb_xy = s->mb_index2xy[i]; |
867 | 0 | float newq = q * cplx_tab[i] / bits_tab[i]; |
868 | 0 | int intq; |
869 | |
|
870 | 0 | if (s->mpv_flags & FF_MPV_FLAG_NAQ) { |
871 | 0 | newq *= bits_sum / cplx_sum; |
872 | 0 | } |
873 | |
|
874 | 0 | intq = (int)(newq + 0.5); |
875 | |
|
876 | 0 | if (intq > qmax) |
877 | 0 | intq = qmax; |
878 | 0 | else if (intq < qmin) |
879 | 0 | intq = qmin; |
880 | 0 | s->lambda_table[mb_xy] = intq; |
881 | 0 | } |
882 | 0 | } |
883 | | |
884 | | void ff_get_2pass_fcode(MpegEncContext *s) |
885 | 0 | { |
886 | 0 | const RateControlContext *rcc = &s->rc_context; |
887 | 0 | const RateControlEntry *rce = &rcc->entry[s->picture_number]; |
888 | |
|
889 | 0 | s->f_code = rce->f_code; |
890 | 0 | s->b_code = rce->b_code; |
891 | 0 | } |
892 | | |
893 | | // FIXME rd or at least approx for dquant |
894 | | |
895 | | float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run) |
896 | 93.5k | { |
897 | 93.5k | float q; |
898 | 93.5k | int qmin, qmax; |
899 | 93.5k | float br_compensation; |
900 | 93.5k | double diff; |
901 | 93.5k | double short_term_q; |
902 | 93.5k | double fps; |
903 | 93.5k | int picture_number = s->picture_number; |
904 | 93.5k | int64_t wanted_bits; |
905 | 93.5k | RateControlContext *rcc = &s->rc_context; |
906 | 93.5k | AVCodecContext *a = s->avctx; |
907 | 93.5k | RateControlEntry local_rce, *rce; |
908 | 93.5k | double bits; |
909 | 93.5k | double rate_factor; |
910 | 93.5k | int64_t var; |
911 | 93.5k | const int pict_type = s->pict_type; |
912 | 93.5k | emms_c(); |
913 | | |
914 | 93.5k | get_qminmax(&qmin, &qmax, s, pict_type); |
915 | | |
916 | 93.5k | fps = get_fps(s->avctx); |
917 | | /* update predictors */ |
918 | 93.5k | if (picture_number > 2 && !dry_run) { |
919 | 35.5k | const int64_t last_var = |
920 | 35.5k | s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum |
921 | 35.5k | : rcc->last_mc_mb_var_sum; |
922 | 35.5k | av_assert1(s->frame_bits >= s->stuffing_bits); |
923 | 35.5k | update_predictor(&rcc->pred[s->last_pict_type], |
924 | 35.5k | rcc->last_qscale, |
925 | 35.5k | sqrt(last_var), |
926 | 35.5k | s->frame_bits - s->stuffing_bits); |
927 | 35.5k | } |
928 | | |
929 | 93.5k | if (s->avctx->flags & AV_CODEC_FLAG_PASS2) { |
930 | 0 | av_assert0(picture_number >= 0); |
931 | 0 | if (picture_number >= rcc->num_entries) { |
932 | 0 | av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n"); |
933 | 0 | return -1; |
934 | 0 | } |
935 | 0 | rce = &rcc->entry[picture_number]; |
936 | 0 | wanted_bits = rce->expected_bits; |
937 | 93.5k | } else { |
938 | 93.5k | const MPVPicture *dts_pic; |
939 | 93.5k | double wanted_bits_double; |
940 | 93.5k | rce = &local_rce; |
941 | | |
942 | | /* FIXME add a dts field to AVFrame and ensure it is set and use it |
943 | | * here instead of reordering but the reordering is simpler for now |
944 | | * until H.264 B-pyramid must be handled. */ |
945 | 93.5k | if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay) |
946 | 6.78k | dts_pic = s->cur_pic.ptr; |
947 | 86.7k | else |
948 | 86.7k | dts_pic = s->last_pic.ptr; |
949 | | |
950 | 93.5k | if (!dts_pic || dts_pic->f->pts == AV_NOPTS_VALUE) |
951 | 49.6k | wanted_bits_double = s->bit_rate * (double)picture_number / fps; |
952 | 43.9k | else |
953 | 43.9k | wanted_bits_double = s->bit_rate * (double)dts_pic->f->pts / fps; |
954 | 93.5k | if (wanted_bits_double > INT64_MAX) { |
955 | 9.27k | av_log(s, AV_LOG_WARNING, "Bits exceed 64bit range\n"); |
956 | 9.27k | wanted_bits = INT64_MAX; |
957 | 9.27k | } else |
958 | 84.3k | wanted_bits = (int64_t)wanted_bits_double; |
959 | 93.5k | } |
960 | | |
961 | 93.5k | diff = s->total_bits - wanted_bits; |
962 | 93.5k | br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance; |
963 | 93.5k | if (br_compensation <= 0.0) |
964 | 9.27k | br_compensation = 0.001; |
965 | | |
966 | 93.5k | var = pict_type == AV_PICTURE_TYPE_I ? s->mb_var_sum : s->mc_mb_var_sum; |
967 | | |
968 | 93.5k | short_term_q = 0; /* avoid warning */ |
969 | 93.5k | if (s->avctx->flags & AV_CODEC_FLAG_PASS2) { |
970 | 0 | if (pict_type != AV_PICTURE_TYPE_I) |
971 | 0 | av_assert0(pict_type == rce->new_pict_type); |
972 | | |
973 | 0 | q = rce->new_qscale / br_compensation; |
974 | 0 | ff_dlog(s, "%f %f %f last:%d var:%"PRId64" type:%d//\n", q, rce->new_qscale, |
975 | 0 | br_compensation, s->frame_bits, var, pict_type); |
976 | 93.5k | } else { |
977 | 93.5k | rce->pict_type = |
978 | 93.5k | rce->new_pict_type = pict_type; |
979 | 93.5k | rce->mc_mb_var_sum = s->mc_mb_var_sum; |
980 | 93.5k | rce->mb_var_sum = s->mb_var_sum; |
981 | 93.5k | rce->qscale = FF_QP2LAMBDA * 2; |
982 | 93.5k | rce->f_code = s->f_code; |
983 | 93.5k | rce->b_code = s->b_code; |
984 | 93.5k | rce->misc_bits = 1; |
985 | | |
986 | 93.5k | bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var)); |
987 | 93.5k | if (pict_type == AV_PICTURE_TYPE_I) { |
988 | 33.4k | rce->i_count = s->mb_num; |
989 | 33.4k | rce->i_tex_bits = bits; |
990 | 33.4k | rce->p_tex_bits = 0; |
991 | 33.4k | rce->mv_bits = 0; |
992 | 60.1k | } else { |
993 | 60.1k | rce->i_count = 0; // FIXME we do know this approx |
994 | 60.1k | rce->i_tex_bits = 0; |
995 | 60.1k | rce->p_tex_bits = bits * 0.9; |
996 | 60.1k | rce->mv_bits = bits * 0.1; |
997 | 60.1k | } |
998 | 93.5k | rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale; |
999 | 93.5k | rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale; |
1000 | 93.5k | rcc->mv_bits_sum[pict_type] += rce->mv_bits; |
1001 | 93.5k | rcc->frame_count[pict_type]++; |
1002 | | |
1003 | 93.5k | rate_factor = rcc->pass1_wanted_bits / |
1004 | 93.5k | rcc->pass1_rc_eq_output_sum * br_compensation; |
1005 | | |
1006 | 93.5k | q = get_qscale(s, rce, rate_factor, picture_number); |
1007 | 93.5k | if (q < 0) |
1008 | 0 | return -1; |
1009 | | |
1010 | 93.5k | av_assert0(q > 0.0); |
1011 | 93.5k | q = get_diff_limited_q(s, rce, q); |
1012 | 93.5k | av_assert0(q > 0.0); |
1013 | | |
1014 | | // FIXME type dependent blur like in 2-pass |
1015 | 93.5k | if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) { |
1016 | 62.0k | rcc->short_term_qsum *= a->qblur; |
1017 | 62.0k | rcc->short_term_qcount *= a->qblur; |
1018 | | |
1019 | 62.0k | rcc->short_term_qsum += q; |
1020 | 62.0k | rcc->short_term_qcount++; |
1021 | 62.0k | q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount; |
1022 | 62.0k | } |
1023 | 93.5k | av_assert0(q > 0.0); |
1024 | | |
1025 | 93.5k | q = modify_qscale(s, rce, q, picture_number); |
1026 | | |
1027 | 93.5k | rcc->pass1_wanted_bits += s->bit_rate / fps; |
1028 | | |
1029 | 93.5k | av_assert0(q > 0.0); |
1030 | 93.5k | } |
1031 | | |
1032 | 93.5k | if (s->avctx->debug & FF_DEBUG_RC) { |
1033 | 0 | av_log(s->avctx, AV_LOG_DEBUG, |
1034 | 0 | "%c qp:%d<%2.1f<%d %d want:%"PRId64" total:%"PRId64" comp:%f st_q:%2.2f " |
1035 | 0 | "size:%d var:%"PRId64"/%"PRId64" br:%"PRId64" fps:%d\n", |
1036 | 0 | av_get_picture_type_char(pict_type), |
1037 | 0 | qmin, q, qmax, picture_number, |
1038 | 0 | wanted_bits / 1000, s->total_bits / 1000, |
1039 | 0 | br_compensation, short_term_q, s->frame_bits, |
1040 | 0 | s->mb_var_sum, s->mc_mb_var_sum, |
1041 | 0 | s->bit_rate / 1000, (int)fps); |
1042 | 0 | } |
1043 | | |
1044 | 93.5k | if (q < qmin) |
1045 | 0 | q = qmin; |
1046 | 93.5k | else if (q > qmax) |
1047 | 0 | q = qmax; |
1048 | | |
1049 | 93.5k | if (s->adaptive_quant) |
1050 | 0 | adaptive_quantization(s, q); |
1051 | 93.5k | else |
1052 | 93.5k | q = (int)(q + 0.5); |
1053 | | |
1054 | 93.5k | if (!dry_run) { |
1055 | 50.9k | rcc->last_qscale = q; |
1056 | 50.9k | rcc->last_mc_mb_var_sum = s->mc_mb_var_sum; |
1057 | 50.9k | rcc->last_mb_var_sum = s->mb_var_sum; |
1058 | 50.9k | } |
1059 | 93.5k | return q; |
1060 | 93.5k | } |