Imported Upstream version 1.4
[deb_x265.git] / source / encoder / ratecontrol.cpp
CommitLineData
72b9787e
JB
1/*****************************************************************************
2 * Copyright (C) 2013 x265 project
3 *
4 * Authors: Sumalatha Polureddy <sumalatha@multicorewareinc.com>
5 * Aarthi Priya Thirumalai <aarthi@multicorewareinc.com>
6 * Xun Xu, PPLive Corporation <xunxu@pptv.com>
7 *
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * This program 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
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
21 *
22 * This program is also available under a commercial proprietary license.
23 * For more information, contact us at license @ x265.com.
24 *****************************************************************************/
25
26#include "common.h"
27#include "param.h"
28#include "frame.h"
29#include "framedata.h"
30#include "picyuv.h"
31
32#include "encoder.h"
33#include "slicetype.h"
34#include "ratecontrol.h"
35#include "sei.h"
36
37#define BR_SHIFT 6
38#define CPB_SHIFT 4
39
40using namespace x265;
41
42/* Amortize the partial cost of I frames over the next N frames */
43const double RateControl::s_amortizeFraction = 0.85;
44const int RateControl::s_amortizeFrames = 75;
45const int RateControl::s_slidingWindowFrames = 20;
46const char *RateControl::s_defaultStatFileName = "x265_2pass.log";
47
48namespace {
49#define CMP_OPT_FIRST_PASS(opt, param_val)\
50{\
51 bErr = 0;\
52 p = strstr(opts, opt "=");\
53 char* q = strstr(opts, "no-"opt);\
54 if (p && sscanf(p, opt "=%d" , &i) && param_val != i)\
55 bErr = 1;\
56 else if (!param_val && !q && !p)\
57 bErr = 1;\
58 else if (param_val && (q || !strstr(opts, opt)))\
59 bErr = 1;\
60 if (bErr)\
61 {\
62 x265_log(m_param, X265_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i);\
63 return false;\
64 }\
65}
66
67inline int calcScale(uint32_t x)
68{
69 static uint8_t lut[16] = {4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0};
70 int y, z = (((x & 0xffff) - 1) >> 27) & 16;
71 x >>= z;
72 z += y = (((x & 0xff) - 1) >> 28) & 8;
73 x >>= y;
74 z += y = (((x & 0xf) - 1) >> 29) & 4;
75 x >>= y;
76 return z + lut[x&0xf];
77}
78
79inline int calcLength(uint32_t x)
80{
81 static uint8_t lut[16] = {4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0};
82 int y, z = (((x >> 16) - 1) >> 27) & 16;
83 x >>= z ^ 16;
84 z += y = ((x - 0x100) >> 28) & 8;
85 x >>= y ^ 8;
86 z += y = ((x - 0x10) >> 29) & 4;
87 x >>= y ^ 4;
88 return z + lut[x];
89}
90
91inline void reduceFraction(int* n, int* d)
92{
93 int a = *n;
94 int b = *d;
95 int c;
96 if (!a || !b)
97 return;
98 c = a % b;
99 while (c)
100 {
101 a = b;
102 b = c;
103 c = a % b;
104 }
105 *n /= b;
106 *d /= b;
107}
108
109inline char *strcatFilename(const char *input, const char *suffix)
110{
111 char *output = X265_MALLOC(char, strlen(input) + strlen(suffix) + 1);
112 if (!output)
113 {
114 x265_log(NULL, X265_LOG_ERROR, "unable to allocate memory for filename\n");
115 return NULL;
116 }
117 strcpy(output, input);
118 strcat(output, suffix);
119 return output;
120}
121
122inline double qScale2bits(RateControlEntry *rce, double qScale)
123{
124 if (qScale < 0.1)
125 qScale = 0.1;
126 return (rce->coeffBits + .1) * pow(rce->qScale / qScale, 1.1)
127 + rce->mvBits * pow(X265_MAX(rce->qScale, 1) / X265_MAX(qScale, 1), 0.5)
128 + rce->miscBits;
129}
130
131inline void copyRceData(RateControlEntry* rce, RateControlEntry* rce2Pass)
132{
133 rce->coeffBits = rce2Pass->coeffBits;
134 rce->mvBits = rce2Pass->mvBits;
135 rce->miscBits = rce2Pass->miscBits;
136 rce->iCuCount = rce2Pass->iCuCount;
137 rce->pCuCount = rce2Pass->pCuCount;
138 rce->skipCuCount = rce2Pass->skipCuCount;
139 rce->keptAsRef = rce2Pass->keptAsRef;
140 rce->qScale = rce2Pass->qScale;
141 rce->newQScale = rce2Pass->newQScale;
142 rce->expectedBits = rce2Pass->expectedBits;
143 rce->expectedVbv = rce2Pass->expectedVbv;
144 rce->blurredComplexity = rce2Pass->blurredComplexity;
145 rce->sliceType = rce2Pass->sliceType;
146}
147
148} // end anonymous namespace
149/* Compute variance to derive AC energy of each block */
150static inline uint32_t acEnergyVar(Frame *curFrame, uint64_t sum_ssd, int shift, int i)
151{
152 uint32_t sum = (uint32_t)sum_ssd;
153 uint32_t ssd = (uint32_t)(sum_ssd >> 32);
154
155 curFrame->m_lowres.wp_sum[i] += sum;
156 curFrame->m_lowres.wp_ssd[i] += ssd;
157 return ssd - ((uint64_t)sum * sum >> shift);
158}
159
160/* Find the energy of each block in Y/Cb/Cr plane */
161static inline uint32_t acEnergyPlane(Frame *curFrame, pixel* src, intptr_t srcStride, int bChroma, int colorFormat)
162{
163 if ((colorFormat != X265_CSP_I444) && bChroma)
164 {
165 ALIGN_VAR_8(pixel, pix[8 * 8]);
166 primitives.luma_copy_pp[LUMA_8x8](pix, 8, src, srcStride);
167 return acEnergyVar(curFrame, primitives.var[BLOCK_8x8](pix, 8), 6, bChroma);
168 }
169 else
170 return acEnergyVar(curFrame, primitives.var[BLOCK_16x16](src, srcStride), 8, bChroma);
171}
172
173/* Find the total AC energy of each block in all planes */
174uint32_t RateControl::acEnergyCu(Frame* curFrame, uint32_t block_x, uint32_t block_y)
175{
176 intptr_t stride = curFrame->m_origPicYuv->m_stride;
177 intptr_t cStride = curFrame->m_origPicYuv->m_strideC;
178 intptr_t blockOffsetLuma = block_x + (block_y * stride);
179 int colorFormat = m_param->internalCsp;
180 int hShift = CHROMA_H_SHIFT(colorFormat);
181 int vShift = CHROMA_V_SHIFT(colorFormat);
182 intptr_t blockOffsetChroma = (block_x >> hShift) + ((block_y >> vShift) * cStride);
183
184 uint32_t var;
185
186 var = acEnergyPlane(curFrame, curFrame->m_origPicYuv->m_picOrg[0] + blockOffsetLuma, stride, 0, colorFormat);
187 var += acEnergyPlane(curFrame, curFrame->m_origPicYuv->m_picOrg[1] + blockOffsetChroma, cStride, 1, colorFormat);
188 var += acEnergyPlane(curFrame, curFrame->m_origPicYuv->m_picOrg[2] + blockOffsetChroma, cStride, 2, colorFormat);
189 x265_emms();
190 return var;
191}
192
193void RateControl::calcAdaptiveQuantFrame(Frame *curFrame)
194{
195 /* Actual adaptive quantization */
196 int maxCol = curFrame->m_origPicYuv->m_picWidth;
197 int maxRow = curFrame->m_origPicYuv->m_picHeight;
198
199 for (int y = 0; y < 3; y++)
200 {
201 curFrame->m_lowres.wp_ssd[y] = 0;
202 curFrame->m_lowres.wp_sum[y] = 0;
203 }
204
205 /* Calculate Qp offset for each 16x16 block in the frame */
206 int block_xy = 0;
207 int block_x = 0, block_y = 0;
208 double strength = 0.f;
209 if (m_param->rc.aqMode == X265_AQ_NONE || m_param->rc.aqStrength == 0)
210 {
211 /* Need to init it anyways for CU tree */
212 int cuWidth = ((maxCol / 2) + X265_LOWRES_CU_SIZE - 1) >> X265_LOWRES_CU_BITS;
213 int cuHeight = ((maxRow / 2) + X265_LOWRES_CU_SIZE - 1) >> X265_LOWRES_CU_BITS;
214 int cuCount = cuWidth * cuHeight;
215
216 if (m_param->rc.aqMode && m_param->rc.aqStrength == 0)
217 {
218 memset(curFrame->m_lowres.qpCuTreeOffset, 0, cuCount * sizeof(double));
219 memset(curFrame->m_lowres.qpAqOffset, 0, cuCount * sizeof(double));
220 for (int cuxy = 0; cuxy < cuCount; cuxy++)
221 curFrame->m_lowres.invQscaleFactor[cuxy] = 256;
222 }
223
224 /* Need variance data for weighted prediction */
225 if (m_param->bEnableWeightedPred || m_param->bEnableWeightedBiPred)
226 {
227 for (block_y = 0; block_y < maxRow; block_y += 16)
228 for (block_x = 0; block_x < maxCol; block_x += 16)
229 acEnergyCu(curFrame, block_x, block_y);
230 }
231 }
232 else
233 {
234 block_xy = 0;
235 double avg_adj_pow2 = 0, avg_adj = 0, qp_adj = 0;
236 if (m_param->rc.aqMode == X265_AQ_AUTO_VARIANCE)
237 {
238 double bit_depth_correction = pow(1 << (X265_DEPTH - 8), 0.5);
239 for (block_y = 0; block_y < maxRow; block_y += 16)
240 {
241 for (block_x = 0; block_x < maxCol; block_x += 16)
242 {
243 uint32_t energy = acEnergyCu(curFrame, block_x, block_y);
244 qp_adj = pow(energy + 1, 0.1);
245 curFrame->m_lowres.qpCuTreeOffset[block_xy] = qp_adj;
246 avg_adj += qp_adj;
247 avg_adj_pow2 += qp_adj * qp_adj;
248 block_xy++;
249 }
250 }
251
252 avg_adj /= m_ncu;
253 avg_adj_pow2 /= m_ncu;
254 strength = m_param->rc.aqStrength * avg_adj / bit_depth_correction;
255 avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (11.f * bit_depth_correction)) / avg_adj;
256 }
257 else
258 strength = m_param->rc.aqStrength * 1.0397f;
259
260 block_xy = 0;
261 for (block_y = 0; block_y < maxRow; block_y += 16)
262 {
263 for (block_x = 0; block_x < maxCol; block_x += 16)
264 {
265 if (m_param->rc.aqMode == X265_AQ_AUTO_VARIANCE)
266 {
267 qp_adj = curFrame->m_lowres.qpCuTreeOffset[block_xy];
268 qp_adj = strength * (qp_adj - avg_adj);
269 }
270 else
271 {
272 uint32_t energy = acEnergyCu(curFrame, block_x, block_y);
273 qp_adj = strength * (X265_LOG2(X265_MAX(energy, 1)) - (14.427f + 2 * (X265_DEPTH - 8)));
274 }
275 curFrame->m_lowres.qpAqOffset[block_xy] = qp_adj;
276 curFrame->m_lowres.qpCuTreeOffset[block_xy] = qp_adj;
277 curFrame->m_lowres.invQscaleFactor[block_xy] = x265_exp2fix8(qp_adj);
278 block_xy++;
279 }
280 }
281 }
282
283 if (m_param->bEnableWeightedPred || m_param->bEnableWeightedBiPred)
284 {
285 int hShift = CHROMA_H_SHIFT(m_param->internalCsp);
286 int vShift = CHROMA_V_SHIFT(m_param->internalCsp);
287 maxCol = ((maxCol + 8) >> 4) << 4;
288 maxRow = ((maxRow + 8) >> 4) << 4;
289 int width[3] = { maxCol, maxCol >> hShift, maxCol >> hShift };
290 int height[3] = { maxRow, maxRow >> vShift, maxRow >> vShift };
291
292 for (int i = 0; i < 3; i++)
293 {
294 uint64_t sum, ssd;
295 sum = curFrame->m_lowres.wp_sum[i];
296 ssd = curFrame->m_lowres.wp_ssd[i];
297 curFrame->m_lowres.wp_ssd[i] = ssd - (sum * sum + (width[i] * height[i]) / 2) / (width[i] * height[i]);
298 }
299 }
300}
301
302RateControl::RateControl(x265_param *p)
303{
304 m_param = p;
305 int lowresCuWidth = ((m_param->sourceWidth / 2) + X265_LOWRES_CU_SIZE - 1) >> X265_LOWRES_CU_BITS;
306 int lowresCuHeight = ((m_param->sourceHeight / 2) + X265_LOWRES_CU_SIZE - 1) >> X265_LOWRES_CU_BITS;
307 m_ncu = lowresCuWidth * lowresCuHeight;
308
309 if (m_param->rc.cuTree)
310 m_qCompress = 1;
311 else
312 m_qCompress = m_param->rc.qCompress;
313
314 // validate for param->rc, maybe it is need to add a function like x265_parameters_valiate()
315 m_residualFrames = 0;
316 m_partialResidualFrames = 0;
317 m_residualCost = 0;
318 m_partialResidualCost = 0;
319 m_rateFactorMaxIncrement = 0;
320 m_rateFactorMaxDecrement = 0;
321 m_fps = m_param->fpsNum / m_param->fpsDenom;
322 m_startEndOrder.set(0);
323 m_bTerminated = false;
324 m_finalFrameCount = 0;
325 m_numEntries = 0;
326 if (m_param->rc.rateControlMode == X265_RC_CRF)
327 {
328 m_param->rc.qp = (int)m_param->rc.rfConstant;
329 m_param->rc.bitrate = 0;
330
331 double baseCplx = m_ncu * (m_param->bframes ? 120 : 80);
332 double mbtree_offset = m_param->rc.cuTree ? (1.0 - m_param->rc.qCompress) * 13.5 : 0;
333 m_rateFactorConstant = pow(baseCplx, 1 - m_qCompress) /
334 x265_qp2qScale(m_param->rc.rfConstant + mbtree_offset);
335 if (m_param->rc.rfConstantMax)
336 {
337 m_rateFactorMaxIncrement = m_param->rc.rfConstantMax - m_param->rc.rfConstant;
338 if (m_rateFactorMaxIncrement <= 0)
339 {
340 x265_log(m_param, X265_LOG_WARNING, "CRF max must be greater than CRF\n");
341 m_rateFactorMaxIncrement = 0;
342 }
343 }
344 if (m_param->rc.rfConstantMin)
345 m_rateFactorMaxDecrement = m_param->rc.rfConstant - m_param->rc.rfConstantMin;
346 }
347 m_isAbr = m_param->rc.rateControlMode != X265_RC_CQP && !m_param->rc.bStatRead;
348 m_2pass = m_param->rc.rateControlMode == X265_RC_ABR && m_param->rc.bStatRead;
349 m_bitrate = m_param->rc.bitrate * 1000;
350 m_frameDuration = (double)m_param->fpsDenom / m_param->fpsNum;
351 m_qp = m_param->rc.qp;
352 m_lastRceq = 1; /* handles the cmplxrsum when the previous frame cost is zero */
353 m_shortTermCplxSum = 0;
354 m_shortTermCplxCount = 0;
355 m_lastNonBPictType = I_SLICE;
356 m_isAbrReset = false;
357 m_lastAbrResetPoc = -1;
358 m_statFileOut = NULL;
359 m_cutreeStatFileOut = m_cutreeStatFileIn = NULL;
360 m_rce2Pass = NULL;
361
362 // vbv initialization
363 m_param->rc.vbvBufferSize = Clip3(0, 2000000, m_param->rc.vbvBufferSize);
364 m_param->rc.vbvMaxBitrate = Clip3(0, 2000000, m_param->rc.vbvMaxBitrate);
365 m_param->rc.vbvBufferInit = Clip3(0.0, 2000000.0, m_param->rc.vbvBufferInit);
366 m_singleFrameVbv = 0;
367 if (m_param->rc.vbvBufferSize)
368 {
369 if (m_param->rc.rateControlMode == X265_RC_CQP)
370 {
371 x265_log(m_param, X265_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n");
372 m_param->rc.vbvBufferSize = 0;
373 m_param->rc.vbvMaxBitrate = 0;
374 }
375 else if (m_param->rc.vbvMaxBitrate == 0)
376 {
377 if (m_param->rc.rateControlMode == X265_RC_ABR)
378 {
379 x265_log(m_param, X265_LOG_WARNING, "VBV maxrate unspecified, assuming CBR\n");
380 m_param->rc.vbvMaxBitrate = m_param->rc.bitrate;
381 }
382 else
383 {
384 x265_log(m_param, X265_LOG_WARNING, "VBV bufsize set but maxrate unspecified, ignored\n");
385 m_param->rc.vbvBufferSize = 0;
386 }
387 }
388 else if (m_param->rc.vbvMaxBitrate < m_param->rc.bitrate &&
389 m_param->rc.rateControlMode == X265_RC_ABR)
390 {
391 x265_log(m_param, X265_LOG_WARNING, "max bitrate less than average bitrate, assuming CBR\n");
392 m_param->rc.bitrate = m_param->rc.vbvMaxBitrate;
393 }
394 }
395 else if (m_param->rc.vbvMaxBitrate)
396 {
397 x265_log(m_param, X265_LOG_WARNING, "VBV maxrate specified, but no bufsize, ignored\n");
398 m_param->rc.vbvMaxBitrate = 0;
399 }
400 m_isVbv = m_param->rc.vbvMaxBitrate > 0 && m_param->rc.vbvBufferSize > 0;
401 if (m_param->bEmitHRDSEI && !m_isVbv)
402 {
403 x265_log(m_param, X265_LOG_WARNING, "NAL HRD parameters require VBV parameters, ignored\n");
404 m_param->bEmitHRDSEI = 0;
405 }
406
407 m_isCbr = m_param->rc.rateControlMode == X265_RC_ABR && m_isVbv && !m_2pass && m_param->rc.vbvMaxBitrate <= m_param->rc.bitrate;
408 m_leadingBframes = m_param->bframes;
409 m_bframeBits = 0;
410 m_leadingNoBSatd = 0;
411 m_ipOffset = 6.0 * X265_LOG2(m_param->rc.ipFactor);
412 m_pbOffset = 6.0 * X265_LOG2(m_param->rc.pbFactor);
413
414 /* Adjust the first frame in order to stabilize the quality level compared to the rest */
415#define ABR_INIT_QP_MIN (24)
416#define ABR_INIT_QP_MAX (40)
417#define CRF_INIT_QP (int)m_param->rc.rfConstant
418 for (int i = 0; i < 3; i++)
419 m_lastQScaleFor[i] = x265_qp2qScale(m_param->rc.rateControlMode == X265_RC_CRF ? CRF_INIT_QP : ABR_INIT_QP_MIN);
420
421 if (m_param->rc.rateControlMode == X265_RC_CQP)
422 {
423 if (m_qp && !m_param->bLossless)
424 {
425 m_qpConstant[P_SLICE] = m_qp;
426 m_qpConstant[I_SLICE] = Clip3(0, QP_MAX_MAX, (int)(m_qp - m_ipOffset + 0.5));
427 m_qpConstant[B_SLICE] = Clip3(0, QP_MAX_MAX, (int)(m_qp + m_pbOffset + 0.5));
428 }
429 else
430 {
431 m_qpConstant[P_SLICE] = m_qpConstant[I_SLICE] = m_qpConstant[B_SLICE] = m_qp;
432 }
433 }
434
435 /* qstep - value set as encoder specific */
436 m_lstep = pow(2, m_param->rc.qpStep / 6.0);
437
438 for (int i = 0; i < 2; i++)
439 m_cuTreeStats.qpBuffer[i] = NULL;
440}
441
442bool RateControl::init(const SPS *sps)
443{
444 if (m_isVbv)
445 {
446 /* We don't support changing the ABR bitrate right now,
447 * so if the stream starts as CBR, keep it CBR. */
448 if (m_param->rc.vbvBufferSize < (int)(m_param->rc.vbvMaxBitrate / m_fps))
449 {
450 m_param->rc.vbvBufferSize = (int)(m_param->rc.vbvMaxBitrate / m_fps);
451 x265_log(m_param, X265_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n",
452 m_param->rc.vbvBufferSize);
453 }
454 int vbvBufferSize = m_param->rc.vbvBufferSize * 1000;
455 int vbvMaxBitrate = m_param->rc.vbvMaxBitrate * 1000;
456
457 if (m_param->bEmitHRDSEI)
458 {
459 const HRDInfo* hrd = &sps->vuiParameters.hrdParameters;
460 vbvBufferSize = hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT);
461 vbvMaxBitrate = hrd->bitRateValue << (hrd->bitRateScale + BR_SHIFT);
462 }
463 m_bufferRate = vbvMaxBitrate / m_fps;
464 m_vbvMaxRate = vbvMaxBitrate;
465 m_bufferSize = vbvBufferSize;
466 m_singleFrameVbv = m_bufferRate * 1.1 > m_bufferSize;
467
468 if (m_param->rc.vbvBufferInit > 1.)
469 m_param->rc.vbvBufferInit = Clip3(0.0, 1.0, m_param->rc.vbvBufferInit / m_param->rc.vbvBufferSize);
470 m_param->rc.vbvBufferInit = Clip3(0.0, 1.0, X265_MAX(m_param->rc.vbvBufferInit, m_bufferRate / m_bufferSize));
471 m_bufferFillFinal = m_bufferSize * m_param->rc.vbvBufferInit;
472 }
473
474 m_totalBits = 0;
475 m_framesDone = 0;
476 m_residualCost = 0;
477 m_partialResidualCost = 0;
478 for (int i = 0; i < s_slidingWindowFrames; i++)
479 {
480 m_satdCostWindow[i] = 0;
481 m_encodedBitsWindow[i] = 0;
482 }
483 m_sliderPos = 0;
484
485 /* 720p videos seem to be a good cutoff for cplxrSum */
486 double tuneCplxFactor = (m_param->rc.cuTree && m_ncu > 3600) ? 2.5 : 1;
487
488 /* estimated ratio that produces a reasonable QP for the first I-frame */
489 m_cplxrSum = .01 * pow(7.0e5, m_qCompress) * pow(m_ncu, 0.5) * tuneCplxFactor;
490 m_wantedBitsWindow = m_bitrate * m_frameDuration;
491 m_accumPNorm = .01;
492 m_accumPQp = (m_param->rc.rateControlMode == X265_RC_CRF ? CRF_INIT_QP : ABR_INIT_QP_MIN) * m_accumPNorm;
493
494 /* Frame Predictors and Row predictors used in vbv */
495 for (int i = 0; i < 5; i++)
496 {
497 m_pred[i].coeff = 2.0;
498 m_pred[i].count = 1.0;
499 m_pred[i].decay = 0.5;
500 m_pred[i].offset = 0.0;
501 }
502 m_predBfromP = m_pred[0];
503 if (!m_statFileOut && (m_param->rc.bStatWrite || m_param->rc.bStatRead))
504 {
505 /* If the user hasn't defined the stat filename, use the default value */
506 const char *fileName = m_param->rc.statFileName;
507 if (!fileName)
508 fileName = s_defaultStatFileName;
509 /* Load stat file and init 2pass algo */
510 if (m_param->rc.bStatRead)
511 {
512 m_expectedBitsSum = 0;
513 char *p, *statsIn, *statsBuf;
514 /* read 1st pass stats */
515 statsIn = statsBuf = x265_slurp_file(fileName);
516 if (!statsBuf)
517 return false;
518 if (m_param->rc.cuTree)
519 {
520 char *tmpFile = strcatFilename(fileName, ".cutree");
521 if (!tmpFile)
522 return false;
523 m_cutreeStatFileIn = fopen(tmpFile, "rb");
524 X265_FREE(tmpFile);
525 if (!m_cutreeStatFileIn)
526 {
527 x265_log(m_param, X265_LOG_ERROR, "can't open stats file %s\n", tmpFile);
528 return false;
529 }
530 }
531
532 /* check whether 1st pass options were compatible with current options */
533 if (strncmp(statsBuf, "#options:", 9))
534 {
535 x265_log(m_param, X265_LOG_ERROR,"options list in stats file not valid\n");
536 return false;
537 }
538 {
539 int i, j;
540 uint32_t k , l;
541 bool bErr = false;
542 char *opts = statsBuf;
543 statsIn = strchr(statsBuf, '\n');
544 if (!statsIn)
545 {
546 x265_log(m_param, X265_LOG_ERROR, "Malformed stats file\n");
547 return false;
548 }
549 *statsIn = '\0';
550 statsIn++;
551 if (sscanf(opts, "#options: %dx%d", &i, &j) != 2)
552 {
553 x265_log(m_param, X265_LOG_ERROR, "Resolution specified in stats file not valid\n");
554 return false;
555 }
556 if ((p = strstr(opts, " fps=")) == 0 || sscanf(p, " fps=%u/%u", &k, &l) != 2)
557 {
558 x265_log(m_param, X265_LOG_ERROR, "fps specified in stats file not valid\n");
559 return false;
560 }
561 if (k != m_param->fpsNum || l != m_param->fpsDenom)
562 {
563 x265_log(m_param, X265_LOG_ERROR, "fps mismatch with 1st pass (%u/%u vs %u/%u)\n",
564 m_param->fpsNum, m_param->fpsDenom, k, l);
565 return false;
566 }
567 CMP_OPT_FIRST_PASS("bitdepth", m_param->internalBitDepth);
568 CMP_OPT_FIRST_PASS("weightp", m_param->bEnableWeightedPred);
569 CMP_OPT_FIRST_PASS("bframes", m_param->bframes);
570 CMP_OPT_FIRST_PASS("b-pyramid", m_param->bBPyramid);
571 CMP_OPT_FIRST_PASS("open-gop", m_param->bOpenGOP);
572 CMP_OPT_FIRST_PASS("keyint", m_param->keyframeMax);
573 CMP_OPT_FIRST_PASS("scenecut", m_param->scenecutThreshold);
574
575 if ((p = strstr(opts, "b-adapt=")) != 0 && sscanf(p, "b-adapt=%d", &i) && i >= X265_B_ADAPT_NONE && i <= X265_B_ADAPT_TRELLIS)
576 {
577 m_param->bFrameAdaptive = i;
578 }
579 else if (m_param->bframes)
580 {
581 x265_log(m_param, X265_LOG_ERROR, "b-adapt method specified in stats file not valid\n");
582 return false;
583 }
584
585 if ((p = strstr(opts, "rc-lookahead=")) != 0 && sscanf(p, "rc-lookahead=%d", &i))
586 m_param->lookaheadDepth = i;
587 }
588 /* find number of pics */
589 p = statsIn;
590 int numEntries;
591 for (numEntries = -1; p; numEntries++)
592 p = strchr(p + 1, ';');
593 if (!numEntries)
594 {
595 x265_log(m_param, X265_LOG_ERROR, "empty stats file\n");
596 return false;
597 }
598 m_numEntries = numEntries;
599
600 if (m_param->totalFrames < m_numEntries && m_param->totalFrames > 0)
601 {
602 x265_log(m_param, X265_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n",
603 m_param->totalFrames, m_numEntries);
604 }
605 if (m_param->totalFrames > m_numEntries)
606 {
607 x265_log(m_param, X265_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n",
608 m_param->totalFrames, m_numEntries);
609 return false;
610 }
611
612 m_rce2Pass = X265_MALLOC(RateControlEntry, m_numEntries);
613 if (!m_rce2Pass)
614 {
615 x265_log(m_param, X265_LOG_ERROR, "Rce Entries for 2 pass cannot be allocated\n");
616 return false;
617 }
618 /* init all to skipped p frames */
619 for (int i = 0; i < m_numEntries; i++)
620 {
621 RateControlEntry *rce = &m_rce2Pass[i];
622 rce->sliceType = P_SLICE;
623 rce->qScale = rce->newQScale = x265_qp2qScale(20);
624 rce->miscBits = m_ncu + 10;
625 rce->newQp = 0;
626 }
627 /* read stats */
628 p = statsIn;
629 double totalQpAq = 0;
630 for (int i = 0; i < m_numEntries; i++)
631 {
632 RateControlEntry *rce;
633 int frameNumber;
634 char picType;
635 int e;
636 char *next;
637 double qpRc, qpAq;
638 next = strstr(p, ";");
639 if (next)
640 *next++ = 0;
641 e = sscanf(p, " in:%d ", &frameNumber);
642 if (frameNumber < 0 || frameNumber >= m_numEntries)
643 {
644 x265_log(m_param, X265_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frameNumber, i);
645 return false;
646 }
647 rce = &m_rce2Pass[frameNumber];
648 e += sscanf(p, " in:%*d out:%*d type:%c q:%lf q-aq:%lf tex:%d mv:%d misc:%d icu:%lf pcu:%lf scu:%lf",
649 &picType, &qpRc, &qpAq, &rce->coeffBits,
650 &rce->mvBits, &rce->miscBits, &rce->iCuCount, &rce->pCuCount,
651 &rce->skipCuCount);
652 rce->keptAsRef = true;
653 if (picType == 'b' || picType == 'p')
654 rce->keptAsRef = false;
655 if (picType == 'I' || picType == 'i')
656 rce->sliceType = I_SLICE;
657 else if (picType == 'P' || picType == 'p')
658 rce->sliceType = P_SLICE;
659 else if (picType == 'B' || picType == 'b')
660 rce->sliceType = B_SLICE;
661 else
662 e = -1;
663 if (e < 10)
664 {
665 x265_log(m_param, X265_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
666 return false;
667 }
668 rce->qScale = x265_qp2qScale(qpRc);
669 totalQpAq += qpAq;
670 p = next;
671 }
672 X265_FREE(statsBuf);
673
674 if (m_param->rc.rateControlMode == X265_RC_ABR)
675 {
676 if (!initPass2())
677 return false;
678 } /* else we're using constant quant, so no need to run the bitrate allocation */
679 }
680 /* Open output file */
681 /* If input and output files are the same, output to a temp file
682 * and move it to the real name only when it's complete */
683 if (m_param->rc.bStatWrite)
684 {
685 char *p, *statFileTmpname;
686 statFileTmpname = strcatFilename(fileName, ".temp");
687 if (!statFileTmpname)
688 return false;
689 m_statFileOut = fopen(statFileTmpname, "wb");
690 X265_FREE(statFileTmpname);
691 if (!m_statFileOut)
692 {
693 x265_log(m_param, X265_LOG_ERROR, "can't open stats file %s\n", statFileTmpname);
694 return false;
695 }
696 p = x265_param2string(m_param);
697 if (p)
698 fprintf(m_statFileOut, "#options: %s\n", p);
699 X265_FREE(p);
700 if (m_param->rc.cuTree && !m_param->rc.bStatRead)
701 {
702 statFileTmpname = strcatFilename(fileName, ".cutree.temp");
703 if (!statFileTmpname)
704 return false;
705 m_cutreeStatFileOut = fopen(statFileTmpname, "wb");
706 X265_FREE(statFileTmpname);
707 if (!m_cutreeStatFileOut)
708 {
709 x265_log(m_param, X265_LOG_ERROR, "can't open mbtree stats file %s\n", statFileTmpname);
710 return false;
711 }
712 }
713 }
714 if (m_param->rc.cuTree)
715 {
716 m_cuTreeStats.qpBuffer[0] = X265_MALLOC(uint16_t, m_ncu * sizeof(uint16_t));
717 if (m_param->bBPyramid && m_param->rc.bStatRead)
718 m_cuTreeStats.qpBuffer[1] = X265_MALLOC(uint16_t, m_ncu * sizeof(uint16_t));
719 m_cuTreeStats.qpBufPos = -1;
720 }
721 }
722 return true;
723}
724
725void RateControl::initHRD(SPS *sps)
726{
727 int vbvBufferSize = m_param->rc.vbvBufferSize * 1000;
728 int vbvMaxBitrate = m_param->rc.vbvMaxBitrate * 1000;
729
730 // Init HRD
731 HRDInfo* hrd = &sps->vuiParameters.hrdParameters;
732 hrd->cbrFlag = m_isCbr;
733
734 // normalize HRD size and rate to the value / scale notation
735 hrd->bitRateScale = Clip3(0, 15, calcScale(vbvMaxBitrate) - BR_SHIFT);
736 hrd->bitRateValue = (vbvMaxBitrate >> (hrd->bitRateScale + BR_SHIFT));
737
738 hrd->cpbSizeScale = Clip3(0, 15, calcScale(vbvBufferSize) - CPB_SHIFT);
739 hrd->cpbSizeValue = (vbvBufferSize >> (hrd->cpbSizeScale + CPB_SHIFT));
740 int bitRateUnscale = hrd->bitRateValue << (hrd->bitRateScale + BR_SHIFT);
741 int cpbSizeUnscale = hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT);
742
743 // arbitrary
744 #define MAX_DURATION 0.5
745
746 TimingInfo *time = &sps->vuiParameters.timingInfo;
747 int maxCpbOutputDelay = (int)(X265_MIN(m_param->keyframeMax * MAX_DURATION * time->timeScale / time->numUnitsInTick, INT_MAX));
748 int maxDpbOutputDelay = (int)(sps->maxDecPicBuffering * MAX_DURATION * time->timeScale / time->numUnitsInTick);
749 int maxDelay = (int)(90000.0 * cpbSizeUnscale / bitRateUnscale + 0.5);
750
751 hrd->initialCpbRemovalDelayLength = 2 + Clip3(4, 22, 32 - calcLength(maxDelay));
752 hrd->cpbRemovalDelayLength = Clip3(4, 31, 32 - calcLength(maxCpbOutputDelay));
753 hrd->dpbOutputDelayLength = Clip3(4, 31, 32 - calcLength(maxDpbOutputDelay));
754
755 #undef MAX_DURATION
756}
757
758bool RateControl::initPass2()
759{
760 uint64_t allConstBits = 0;
761 uint64_t allAvailableBits = uint64_t(m_param->rc.bitrate * 1000. * m_numEntries * m_frameDuration);
762 double rateFactor, stepMult;
763 double qBlur = m_param->rc.qblur;
764 double cplxBlur = m_param->rc.complexityBlur;
765 const int filterSize = (int)(qBlur * 4) | 1;
766 double expectedBits;
767 double *qScale, *blurredQscale;
768 double baseCplx = m_ncu * (m_param->bframes ? 120 : 80);
769 double clippedDuration = CLIP_DURATION(m_frameDuration) / BASE_FRAME_DURATION;
770
771 /* find total/average complexity & const_bits */
772 for (int i = 0; i < m_numEntries; i++)
773 allConstBits += m_rce2Pass[i].miscBits;
774
775 if (allAvailableBits < allConstBits)
776 {
777 x265_log(m_param, X265_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n",
778 (int)(allConstBits * m_fps / m_numEntries * 1000.));
779 return false;
780 }
781
782 /* Blur complexities, to reduce local fluctuation of QP.
783 * We don't blur the QPs directly, because then one very simple frame
784 * could drag down the QP of a nearby complex frame and give it more
785 * bits than intended. */
786 for (int i = 0; i < m_numEntries; i++)
787 {
788 double weightSum = 0;
789 double cplxSum = 0;
790 double weight = 1.0;
791 double gaussianWeight;
792 /* weighted average of cplx of future frames */
793 for (int j = 1; j < cplxBlur * 2 && j < m_numEntries - i; j++)
794 {
795 RateControlEntry *rcj = &m_rce2Pass[i + j];
796 weight *= 1 - pow(rcj->iCuCount / m_ncu, 2);
797 if (weight < 0.0001)
798 break;
799 gaussianWeight = weight * exp(-j * j / 200.0);
800 weightSum += gaussianWeight;
801 cplxSum += gaussianWeight * (qScale2bits(rcj, 1) - rcj->miscBits) / clippedDuration;
802 }
803 /* weighted average of cplx of past frames */
804 weight = 1.0;
805 for (int j = 0; j <= cplxBlur * 2 && j <= i; j++)
806 {
807 RateControlEntry *rcj = &m_rce2Pass[i - j];
808 gaussianWeight = weight * exp(-j * j / 200.0);
809 weightSum += gaussianWeight;
810 cplxSum += gaussianWeight * (qScale2bits(rcj, 1) - rcj->miscBits) / clippedDuration;
811 weight *= 1 - pow(rcj->iCuCount / m_ncu, 2);
812 if (weight < .0001)
813 break;
814 }
815 m_rce2Pass[i].blurredComplexity = cplxSum / weightSum;
816 }
817
818 CHECKED_MALLOC(qScale, double, m_numEntries);
819 if (filterSize > 1)
820 {
821 CHECKED_MALLOC(blurredQscale, double, m_numEntries);
822 }
823 else
824 blurredQscale = qScale;
825
826 /* Search for a factor which, when multiplied by the RCEQ values from
827 * each frame, adds up to the desired total size.
828 * There is no exact closed-form solution because of VBV constraints and
829 * because qscale2bits is not invertible, but we can start with the simple
830 * approximation of scaling the 1st pass by the ratio of bitrates.
831 * The search range is probably overkill, but speed doesn't matter here. */
832
833 expectedBits = 1;
834 for (int i = 0; i < m_numEntries; i++)
835 {
836 RateControlEntry* rce = &m_rce2Pass[i];
837 double q = getQScale(rce, 1.0);
838 expectedBits += qScale2bits(rce, q);
839 m_lastQScaleFor[rce->sliceType] = q;
840 }
841 stepMult = allAvailableBits / expectedBits;
842
843 rateFactor = 0;
844 for (double step = 1E4 * stepMult; step > 1E-7 * stepMult; step *= 0.5)
845 {
846 expectedBits = 0;
847 rateFactor += step;
848
849 m_lastNonBPictType = -1;
850 m_lastAccumPNorm = 1;
851 m_accumPNorm = 0;
852
853 m_lastQScaleFor[0] = m_lastQScaleFor[1] =
854 m_lastQScaleFor[2] = pow(baseCplx, 1 - m_qCompress) / rateFactor;
855
856 /* find qscale */
857 for (int i = 0; i < m_numEntries; i++)
858 {
859 RateControlEntry *rce = &m_rce2Pass[i];
860 qScale[i] = getQScale(rce, rateFactor);
861 m_lastQScaleFor[rce->sliceType] = qScale[i];
862 }
863
864 /* fixed I/B qscale relative to P */
865 for (int i = m_numEntries - 1; i >= 0; i--)
866 {
867 qScale[i] = getDiffLimitedQScale(&m_rce2Pass[i], qScale[i]);
868 X265_CHECK(qScale[i] >= 0, "qScale became negative\n");
869 }
870
871 /* smooth curve */
872 if (filterSize > 1)
873 {
874 X265_CHECK(filterSize % 2 == 1, "filterSize not an odd number\n");
875 for (int i = 0; i < m_numEntries; i++)
876 {
877 double q = 0.0, sum = 0.0;
878
879 for (int j = 0; j < filterSize; j++)
880 {
881 int idx = i + j - filterSize / 2;
882 double d = idx - i;
883 double coeff = qBlur == 0 ? 1.0 : exp(-d * d / (qBlur * qBlur));
884 if (idx < 0 || idx >= m_numEntries)
885 continue;
886 if (m_rce2Pass[i].sliceType != m_rce2Pass[idx].sliceType)
887 continue;
888 q += qScale[idx] * coeff;
889 sum += coeff;
890 }
891 blurredQscale[i] = q / sum;
892 }
893 }
894
895 /* find expected bits */
896 for (int i = 0; i < m_numEntries; i++)
897 {
898 RateControlEntry *rce = &m_rce2Pass[i];
899 rce->newQScale = clipQscale(NULL, rce, blurredQscale[i]); // check if needed
900 X265_CHECK(rce->newQScale >= 0, "new Qscale is negative\n");
901 expectedBits += qScale2bits(rce, rce->newQScale);
902 }
903
904 if (expectedBits > allAvailableBits)
905 rateFactor -= step;
906 }
907
908 X265_FREE(qScale);
909 if (filterSize > 1)
910 X265_FREE(blurredQscale);
911
912 if (m_isVbv)
913 if (!vbv2Pass(allAvailableBits))
914 return false;
915 expectedBits = countExpectedBits();
916
917 if (fabs(expectedBits / allAvailableBits - 1.0) > 0.01)
918 {
919 double avgq = 0;
920 for (int i = 0; i < m_numEntries; i++)
921 avgq += m_rce2Pass[i].newQScale;
922 avgq = x265_qScale2qp(avgq / m_numEntries);
923
924 if (expectedBits > allAvailableBits || !m_isVbv)
925 x265_log(m_param, X265_LOG_WARNING, "Error: 2pass curve failed to converge\n");
926 x265_log(m_param, X265_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n",
927 (double)m_param->rc.bitrate,
928 expectedBits * m_fps / (m_numEntries * 1000.),
929 avgq);
930 if (expectedBits < allAvailableBits && avgq < QP_MIN + 2)
931 {
932 x265_log(m_param, X265_LOG_WARNING, "try reducing target bitrate\n");
933 }
934 else if (expectedBits > allAvailableBits && avgq > QP_MAX_SPEC - 2)
935 {
936 x265_log(m_param, X265_LOG_WARNING, "try increasing target bitrate\n");
937 }
938 else if (!(m_2pass && m_isVbv))
939 x265_log(m_param, X265_LOG_WARNING, "internal error\n");
940 }
941
942 return true;
943
944fail:
945 x265_log(m_param, X265_LOG_WARNING, "two-pass ABR initialization failed\n");
946 return false;
947}
948
949bool RateControl::vbv2Pass(uint64_t allAvailableBits)
950{
951 /* for each interval of bufferFull .. underflow, uniformly increase the qp of all
952 * frames in the interval until either buffer is full at some intermediate frame or the
953 * last frame in the interval no longer underflows. Recompute intervals and repeat.
954 * Then do the converse to put bits back into overflow areas until target size is met */
955
956 double *fills;
957 double expectedBits = 0;
958 double adjustment;
959 double prevBits = 0;
960 int t0, t1;
961 int iterations = 0 , adjMin, adjMax;
962 CHECKED_MALLOC(fills, double, m_numEntries + 1);
963 fills++;
964
965 /* adjust overall stream size */
966 do
967 {
968 iterations++;
969 prevBits = expectedBits;
970
971 if (expectedBits)
972 { /* not first iteration */
973 adjustment = X265_MAX(X265_MIN(expectedBits / allAvailableBits, 0.999), 0.9);
974 fills[-1] = m_bufferSize * m_param->rc.vbvBufferInit;
975 t0 = 0;
976 /* fix overflows */
977 adjMin = 1;
978 while (adjMin && findUnderflow(fills, &t0, &t1, 1))
979 {
980 adjMin = fixUnderflow(t0, t1, adjustment, MIN_QPSCALE, MAX_MAX_QPSCALE);
981 t0 = t1;
982 }
983 }
984
985 fills[-1] = m_bufferSize * (1. - m_param->rc.vbvBufferInit);
986 t0 = 0;
987 /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */
988 adjMax = 1;
989 while (adjMax && findUnderflow(fills, &t0, &t1, 0))
990 adjMax = fixUnderflow(t0, t1, 1.001, MIN_QPSCALE, MAX_MAX_QPSCALE );
991
992 expectedBits = countExpectedBits();
993 }
994 while ((expectedBits < .995 * allAvailableBits) && ((int64_t)(expectedBits+.5) > (int64_t)(prevBits+.5)));
995
996 if (!adjMax)
997 x265_log(m_param, X265_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n");
998
999 /* store expected vbv filling values for tracking when encoding */
1000 for (int i = 0; i < m_numEntries; i++)
1001 m_rce2Pass[i].expectedVbv = m_bufferSize - fills[i];
1002
1003 X265_FREE(fills - 1);
1004 return true;
1005
1006fail:
1007 x265_log(m_param, X265_LOG_ERROR, "malloc failure in two-pass VBV init\n");
1008 return false;
1009}
1010
1011/* In 2pass, force the same frame types as in the 1st pass */
1012int RateControl::rateControlSliceType(int frameNum)
1013{
1014 if (m_param->rc.bStatRead)
1015 {
1016 if (frameNum >= m_numEntries)
1017 {
1018 /* We could try to initialize everything required for ABR and
1019 * adaptive B-frames, but that would be complicated.
1020 * So just calculate the average QP used so far. */
1021 m_param->rc.qp = (m_accumPQp < 1) ? ABR_INIT_QP_MAX : (int)(m_accumPQp + 0.5);
1022 m_qpConstant[P_SLICE] = Clip3(0, QP_MAX_MAX, m_param->rc.qp);
1023 m_qpConstant[I_SLICE] = Clip3(0, QP_MAX_MAX, (int)(m_param->rc.qp - m_ipOffset + 0.5));
1024 m_qpConstant[B_SLICE] = Clip3(0, QP_MAX_MAX, (int)(m_param->rc.qp + m_pbOffset + 0.5));
1025
1026 x265_log(m_param, X265_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", m_numEntries);
1027 x265_log(m_param, X265_LOG_ERROR, "continuing anyway, at constant QP=%d\n", m_param->rc.qp);
1028 if (m_param->bFrameAdaptive)
1029 x265_log(m_param, X265_LOG_ERROR, "disabling adaptive B-frames\n");
1030
1031 m_isAbr = 0;
1032 m_2pass = 0;
1033 m_param->rc.rateControlMode = X265_RC_CQP;
1034 m_param->rc.bStatRead = 0;
1035 m_param->bFrameAdaptive = 0;
1036 m_param->scenecutThreshold = 0;
1037 m_param->rc.cuTree = 0;
1038 if (m_param->bframes > 1)
1039 m_param->bframes = 1;
1040 return X265_TYPE_AUTO;
1041 }
1042 int frameType = m_rce2Pass[frameNum].sliceType == I_SLICE ? (frameNum > 0 && m_param->bOpenGOP ? X265_TYPE_I : X265_TYPE_IDR)
1043 : m_rce2Pass[frameNum].sliceType == P_SLICE ? X265_TYPE_P
1044 : (m_rce2Pass[frameNum].sliceType == B_SLICE && m_rce2Pass[frameNum].keptAsRef? X265_TYPE_BREF : X265_TYPE_B);
1045 return frameType;
1046 }
1047 else
1048 return X265_TYPE_AUTO;
1049}
1050
1051int RateControl::rateControlStart(Frame* curFrame, RateControlEntry* rce, Encoder* enc)
1052{
1053 int orderValue = m_startEndOrder.get();
1054 int startOrdinal = rce->encodeOrder * 2;
1055
1056 while (orderValue < startOrdinal && !m_bTerminated)
1057 orderValue = m_startEndOrder.waitForChange(orderValue);
1058
1059 if (!curFrame)
1060 {
1061 // faked rateControlStart calls when the encoder is flushing
1062 m_startEndOrder.incr();
1063 return 0;
1064 }
1065
1066 FrameData& curEncData = *curFrame->m_encData;
1067 m_curSlice = curEncData.m_slice;
1068 m_sliceType = m_curSlice->m_sliceType;
1069 rce->sliceType = m_sliceType;
1070 rce->poc = m_curSlice->m_poc;
1071 if (m_param->rc.bStatRead)
1072 {
1073 X265_CHECK(rce->poc >= 0 && rce->poc < m_numEntries, "bad encode ordinal\n");
1074 copyRceData(rce, &m_rce2Pass[rce->poc]);
1075 }
1076 rce->isActive = true;
1077 if (m_sliceType == B_SLICE)
1078 rce->bframes = m_leadingBframes;
1079 else
1080 m_leadingBframes = curFrame->m_lowres.leadingBframes;
1081
1082 rce->bLastMiniGopBFrame = curFrame->m_lowres.bLastMiniGopBFrame;
1083 rce->bufferRate = m_bufferRate;
1084 rce->rowCplxrSum = 0.0;
1085 rce->rowTotalBits = 0;
1086 if (m_isVbv)
1087 {
1088 if (rce->rowPreds[0][0].count == 0)
1089 {
1090 for (int i = 0; i < 3; i++)
1091 {
1092 for (int j = 0; j < 2; j++)
1093 {
1094 rce->rowPreds[i][j].coeff = 0.25;
1095 rce->rowPreds[i][j].count = 1.0;
1096 rce->rowPreds[i][j].decay = 0.5;
1097 rce->rowPreds[i][j].offset = 0.0;
1098 }
1099 }
1100 }
1101 rce->rowPred[0] = &rce->rowPreds[m_sliceType][0];
1102 rce->rowPred[1] = &rce->rowPreds[m_sliceType][1];
1103 m_predictedBits = m_totalBits;
1104 updateVbvPlan(enc);
1105 rce->bufferFill = m_bufferFill;
1106
1107 int mincr = enc->m_vps.ptl.minCrForLevel;
1108 /* Profiles above Main10 don't require maxAU size check, so just set the maximum to a large value. */
1109 if (enc->m_vps.ptl.profileIdc > Profile::MAIN10 || enc->m_vps.ptl.levelIdc == Level::NONE)
1110 rce->frameSizeMaximum = 1e9;
1111 else
1112 {
1113 /* The spec has a special case for the first frame. */
1114 if (rce->encodeOrder == 0)
1115 {
1116 /* 1.5 * (Max( PicSizeInSamplesY, fR * MaxLumaSr) + MaxLumaSr * (AuCpbRemovalTime[ 0 ] -AuNominalRemovalTime[ 0 ])) ? MinCr */
1117 double fr = 1. / 300;
1118 int picSizeInSamplesY = m_param->sourceWidth * m_param->sourceHeight;
1119 rce->frameSizeMaximum = 8 * 1.5 * X265_MAX(picSizeInSamplesY, fr * enc->m_vps.ptl.maxLumaSrForLevel) / mincr;
1120 }
1121 else
1122 {
1123 /* 1.5 * MaxLumaSr * (AuCpbRemovalTime[ n ] - AyCpbRemovalTime[ n - 1 ]) ? MinCr */
1124 rce->frameSizeMaximum = 8 * 1.5 * enc->m_vps.ptl.maxLumaSrForLevel * m_frameDuration / mincr;
1125 }
1126 }
1127 }
1128 if (m_isAbr || m_2pass) // ABR,CRF
1129 {
1130 if (m_isAbr || m_isVbv)
1131 {
1132 m_currentSatd = curFrame->m_lowres.satdCost >> (X265_DEPTH - 8);
1133 /* Update rce for use in rate control VBV later */
1134 rce->lastSatd = m_currentSatd;
1135 }
1136 double q = x265_qScale2qp(rateEstimateQscale(curFrame, rce));
1137 q = Clip3((double)QP_MIN, (double)QP_MAX_MAX, q);
1138 m_qp = int(q + 0.5);
1139 rce->qpaRc = curEncData.m_avgQpRc = curEncData.m_avgQpAq = q;
1140 /* copy value of lastRceq into thread local rce struct *to be used in RateControlEnd() */
1141 rce->qRceq = m_lastRceq;
1142 accumPQpUpdate();
1143 }
1144 else // CQP
1145 {
1146 if (m_sliceType == B_SLICE && IS_REFERENCED(curFrame))
1147 m_qp = (m_qpConstant[B_SLICE] + m_qpConstant[P_SLICE]) / 2;
1148 else
1149 m_qp = m_qpConstant[m_sliceType];
1150 curEncData.m_avgQpAq = curEncData.m_avgQpRc = m_qp;
1151 }
1152 if (m_sliceType != B_SLICE)
1153 {
1154 m_lastNonBPictType = m_sliceType;
1155 m_leadingNoBSatd = m_currentSatd;
1156 }
1157 rce->leadingNoBSatd = m_leadingNoBSatd;
1158 if (curFrame->m_forceqp)
1159 {
1160 m_qp = int32_t(curFrame->m_forceqp + 0.5) - 1;
1161 m_qp = Clip3(QP_MIN, QP_MAX_MAX, m_qp);
1162 rce->qpaRc = curEncData.m_avgQpRc = curEncData.m_avgQpAq = m_qp;
1163 }
1164 // Do not increment m_startEndOrder here. Make rateControlEnd of previous thread
1165 // to wait until rateControlUpdateStats of this frame is called
1166 m_framesDone++;
1167 return m_qp;
1168}
1169
1170void RateControl::accumPQpUpdate()
1171{
1172 m_accumPQp *= .95;
1173 m_accumPNorm *= .95;
1174 m_accumPNorm += 1;
1175 if (m_sliceType == I_SLICE)
1176 m_accumPQp += m_qp + m_ipOffset;
1177 else
1178 m_accumPQp += m_qp;
1179}
1180
1181double RateControl::getDiffLimitedQScale(RateControlEntry *rce, double q)
1182{
1183 // force I/B quants as a function of P quants
1184 const double lastPqScale = m_lastQScaleFor[P_SLICE];
1185 const double lastNonBqScale = m_lastQScaleFor[m_lastNonBPictType];
1186 if (rce->sliceType == I_SLICE)
1187 {
1188 double iq = q;
1189 double pq = x265_qp2qScale(m_accumPQp / m_accumPNorm);
1190 double ipFactor = fabs(m_param->rc.ipFactor);
1191 /* don't apply ipFactor if the following frame is also I */
1192 if (m_accumPNorm <= 0)
1193 q = iq;
1194 else if (m_param->rc.ipFactor < 0)
1195 q = iq / ipFactor;
1196 else if (m_accumPNorm >= 1)
1197 q = pq / ipFactor;
1198 else
1199 q = m_accumPNorm * pq / ipFactor + (1 - m_accumPNorm) * iq;
1200 }
1201 else if (rce->sliceType == B_SLICE)
1202 {
1203 if (m_param->rc.pbFactor > 0)
1204 q = lastNonBqScale;
1205 if (!rce->keptAsRef)
1206 q *= fabs(m_param->rc.pbFactor);
1207 }
1208 else if (rce->sliceType == P_SLICE
1209 && m_lastNonBPictType == P_SLICE
1210 && rce->coeffBits == 0)
1211 {
1212 q = lastPqScale;
1213 }
1214
1215 /* last qscale / qdiff stuff */
1216 if (m_lastNonBPictType == rce->sliceType &&
1217 (rce->sliceType != I_SLICE || m_lastAccumPNorm < 1))
1218 {
1219 double maxQscale = m_lastQScaleFor[rce->sliceType] * m_lstep;
1220 double minQscale = m_lastQScaleFor[rce->sliceType] / m_lstep;
1221 q = Clip3(minQscale, maxQscale, q);
1222 }
1223
1224 m_lastQScaleFor[rce->sliceType] = q;
1225 if (rce->sliceType != B_SLICE)
1226 m_lastNonBPictType = rce->sliceType;
1227 if (rce->sliceType == I_SLICE)
1228 {
1229 m_lastAccumPNorm = m_accumPNorm;
1230 m_accumPNorm = 0;
1231 m_accumPQp = 0;
1232 }
1233 if (rce->sliceType == P_SLICE)
1234 {
1235 double mask = 1 - pow(rce->iCuCount / m_ncu, 2);
1236 m_accumPQp = mask * (x265_qScale2qp(q) + m_accumPQp);
1237 m_accumPNorm = mask * (1 + m_accumPNorm);
1238 }
1239
1240 return q;
1241}
1242
1243double RateControl::countExpectedBits()
1244{
1245 double expectedBits = 0;
1246 for( int i = 0; i < m_numEntries; i++ )
1247 {
1248 RateControlEntry *rce = &m_rce2Pass[i];
1249 rce->expectedBits = (uint64_t)expectedBits;
1250 expectedBits += qScale2bits(rce, rce->newQScale);
1251 }
1252 return expectedBits;
1253}
1254
1255bool RateControl::findUnderflow(double *fills, int *t0, int *t1, int over)
1256{
1257 /* find an interval ending on an overflow or underflow (depending on whether
1258 * we're adding or removing bits), and starting on the earliest frame that
1259 * can influence the buffer fill of that end frame. */
1260 const double bufferMin = .1 * m_bufferSize;
1261 const double bufferMax = .9 * m_bufferSize;
1262 double fill = fills[*t0 - 1];
1263 double parity = over ? 1. : -1.;
1264 int start = -1, end = -1;
1265 for (int i = *t0; i < m_numEntries; i++)
1266 {
1267 fill += (m_frameDuration * m_vbvMaxRate -
1268 qScale2bits(&m_rce2Pass[i], m_rce2Pass[i].newQScale)) * parity;
1269 fill = Clip3(0.0, m_bufferSize, fill);
1270 fills[i] = fill;
1271 if (fill <= bufferMin || i == 0)
1272 {
1273 if (end >= 0)
1274 break;
1275 start = i;
1276 }
1277 else if (fill >= bufferMax && start >= 0)
1278 end = i;
1279 }
1280 *t0 = start;
1281 *t1 = end;
1282 return start >= 0 && end >= 0;
1283}
1284
1285bool RateControl::fixUnderflow(int t0, int t1, double adjustment, double qscaleMin, double qscaleMax)
1286{
1287 double qscaleOrig, qscaleNew;
1288 bool adjusted = false;
1289 if (t0 > 0)
1290 t0++;
1291 for (int i = t0; i <= t1; i++)
1292 {
1293 qscaleOrig = m_rce2Pass[i].newQScale;
1294 qscaleOrig = Clip3(qscaleMin, qscaleMax, qscaleOrig);
1295 qscaleNew = qscaleOrig * adjustment;
1296 qscaleNew = Clip3(qscaleMin, qscaleMax, qscaleNew);
1297 m_rce2Pass[i].newQScale = qscaleNew;
1298 adjusted = adjusted || (qscaleNew != qscaleOrig);
1299 }
1300 return adjusted;
1301}
1302
1303bool RateControl::cuTreeReadFor2Pass(Frame* frame)
1304{
1305 uint8_t sliceTypeActual = (uint8_t)m_rce2Pass[frame->m_poc].sliceType;
1306
1307 if (m_rce2Pass[frame->m_poc].keptAsRef)
1308 {
1309 uint8_t type;
1310 if (m_cuTreeStats.qpBufPos < 0)
1311 {
1312 do
1313 {
1314 m_cuTreeStats.qpBufPos++;
1315
1316 if (!fread(&type, 1, 1, m_cutreeStatFileIn))
1317 goto fail;
1318 if (fread(m_cuTreeStats.qpBuffer[m_cuTreeStats.qpBufPos], sizeof(uint16_t), m_ncu, m_cutreeStatFileIn) != (size_t)m_ncu)
1319 goto fail;
1320
1321 if (type != sliceTypeActual && m_cuTreeStats.qpBufPos == 1)
1322 {
1323 x265_log(m_param, X265_LOG_ERROR, "CU-tree frametype %d doesn't match actual frametype %d.\n", type, sliceTypeActual);
1324 return false;
1325 }
1326 }
1327 while(type != sliceTypeActual);
1328 }
1329 for (int i = 0; i < m_ncu; i++)
1330 {
1331 int16_t qpFix8 = m_cuTreeStats.qpBuffer[m_cuTreeStats.qpBufPos][i];
1332 frame->m_lowres.qpCuTreeOffset[i] = (double)(qpFix8) / 256.0;
1333 frame->m_lowres.invQscaleFactor[i] = x265_exp2fix8(frame->m_lowres.qpCuTreeOffset[i]);
1334 }
1335 m_cuTreeStats.qpBufPos--;
1336 }
1337 else
1338 calcAdaptiveQuantFrame(frame);
1339 return true;
1340
1341fail:
1342 x265_log(m_param, X265_LOG_ERROR, "Incomplete CU-tree stats file.\n");
1343 return false;
1344}
1345
1346double RateControl::rateEstimateQscale(Frame* curFrame, RateControlEntry *rce)
1347{
1348 double q;
1349
1350 if (m_2pass)
1351 {
1352 if (m_sliceType != rce->sliceType)
1353 {
1354 x265_log(m_param, X265_LOG_ERROR, "slice=%c but 2pass stats say %c\n",
1355 g_sliceTypeToChar[m_sliceType], g_sliceTypeToChar[rce->sliceType]);
1356 }
1357 }
1358 else
1359 {
1360 if (m_isAbr)
1361 {
1362 double slidingWindowCplxSum = 0;
1363 int start = m_sliderPos > s_slidingWindowFrames ? m_sliderPos : 0;
1364 for (int cnt = 0; cnt < s_slidingWindowFrames; cnt++, start++)
1365 {
1366 int pos = start % s_slidingWindowFrames;
1367 slidingWindowCplxSum *= 0.5;
1368 if (!m_satdCostWindow[pos])
1369 break;
1370 slidingWindowCplxSum += m_satdCostWindow[pos] / (CLIP_DURATION(m_frameDuration) / BASE_FRAME_DURATION);
1371 }
1372 rce->movingAvgSum = slidingWindowCplxSum;
1373 m_satdCostWindow[m_sliderPos % s_slidingWindowFrames] = rce->lastSatd;
1374 m_sliderPos++;
1375 }
1376 }
1377
1378 if (m_sliceType == B_SLICE)
1379 {
1380 /* B-frames don't have independent rate control, but rather get the
1381 * average QP of the two adjacent P-frames + an offset */
1382 Slice* prevRefSlice = m_curSlice->m_refPicList[0][0]->m_encData->m_slice;
1383 Slice* nextRefSlice = m_curSlice->m_refPicList[1][0]->m_encData->m_slice;
1384 double q0 = m_curSlice->m_refPicList[0][0]->m_encData->m_avgQpRc;
1385 double q1 = m_curSlice->m_refPicList[1][0]->m_encData->m_avgQpRc;
1386 bool i0 = prevRefSlice->m_sliceType == I_SLICE;
1387 bool i1 = nextRefSlice->m_sliceType == I_SLICE;
1388 int dt0 = abs(m_curSlice->m_poc - prevRefSlice->m_poc);
1389 int dt1 = abs(m_curSlice->m_poc - nextRefSlice->m_poc);
1390
1391 // Skip taking a reference frame before the Scenecut if ABR has been reset.
1392 if (m_lastAbrResetPoc >= 0)
1393 {
1394 if (prevRefSlice->m_sliceType == P_SLICE && prevRefSlice->m_poc < m_lastAbrResetPoc)
1395 {
1396 i0 = i1;
1397 dt0 = dt1;
1398 q0 = q1;
1399 }
1400 }
1401 if (prevRefSlice->m_sliceType == B_SLICE && IS_REFERENCED(m_curSlice->m_refPicList[0][0]))
1402 q0 -= m_pbOffset / 2;
1403 if (nextRefSlice->m_sliceType == B_SLICE && IS_REFERENCED(m_curSlice->m_refPicList[1][0]))
1404 q1 -= m_pbOffset / 2;
1405 if (i0 && i1)
1406 q = (q0 + q1) / 2 + m_ipOffset;
1407 else if (i0)
1408 q = q1;
1409 else if (i1)
1410 q = q0;
1411 else
1412 q = (q0 * dt1 + q1 * dt0) / (dt0 + dt1);
1413
1414 if (IS_REFERENCED(curFrame))
1415 q += m_pbOffset / 2;
1416 else
1417 q += m_pbOffset;
1418 rce->qpNoVbv = q;
1419 double qScale = x265_qp2qScale(q);
1420
1421 if (!m_2pass && m_isVbv)
1422 {
1423 if (m_leadingBframes > 5)
1424 {
1425 qScale = clipQscale(curFrame, rce, qScale);
1426 m_lastQScaleFor[m_sliceType] = qScale;
1427 }
1428 rce->frameSizePlanned = predictSize(&m_predBfromP, qScale, (double)m_leadingNoBSatd);
1429 }
1430 else if (m_2pass && m_isVbv)
1431 {
1432 rce->frameSizePlanned = qScale2bits(rce, qScale);
1433 }
1434 /* Limit planned size by MinCR */
1435 if (m_isVbv)
1436 rce->frameSizePlanned = X265_MIN(rce->frameSizePlanned, rce->frameSizeMaximum);
1437 rce->frameSizeEstimated = rce->frameSizePlanned;
1438 rce->newQScale = qScale;
1439 return qScale;
1440 }
1441 else
1442 {
1443 double abrBuffer = 2 * m_param->rc.rateTolerance * m_bitrate;
1444 if (m_2pass)
1445 {
1446 int64_t diff;
1447 if (!m_isVbv)
1448 {
1449 m_predictedBits = m_totalBits;
1450 if (rce->encodeOrder < m_param->frameNumThreads)
1451 m_predictedBits += (int64_t)(rce->encodeOrder * m_bitrate / m_fps);
1452 else
1453 m_predictedBits += (int64_t)(m_param->frameNumThreads * m_bitrate / m_fps);
1454 }
1455 /* Adjust ABR buffer based on distance to the end of the video. */
1456 if (m_numEntries > rce->encodeOrder)
1457 {
1458 uint64_t finalBits = m_rce2Pass[m_numEntries - 1].expectedBits;
1459 double videoPos = (double)rce->expectedBits / finalBits;
1460 double scaleFactor = sqrt((1 - videoPos) * m_numEntries);
1461 abrBuffer *= 0.5 * X265_MAX(scaleFactor, 0.5);
1462 }
1463 diff = m_predictedBits - (int64_t)rce->expectedBits;
1464 q = rce->newQScale;
1465 q /= Clip3(0.5, 2.0, (double)(abrBuffer - diff) / abrBuffer);
1466 if (m_expectedBitsSum > 0)
1467 {
1468 /* Adjust quant based on the difference between
1469 * achieved and expected bitrate so far */
1470 double curTime = (double)rce->encodeOrder / m_numEntries;
1471 double w = Clip3(0.0, 1.0, curTime * 100);
1472 q *= pow((double)m_totalBits / m_expectedBitsSum, w);
1473 }
1474 rce->qpNoVbv = x265_qScale2qp(q);
1475 if (m_isVbv)
1476 {
1477 /* Do not overflow vbv */
1478 double expectedSize = qScale2bits(rce, q);
1479 double expectedVbv = m_bufferFill + m_bufferRate - expectedSize;
1480 double expectedFullness = rce->expectedVbv / m_bufferSize;
1481 double qmax = q * (2 - expectedFullness);
1482 double sizeConstraint = 1 + expectedFullness;
1483 qmax = X265_MAX(qmax, rce->newQScale);
1484 if (expectedFullness < .05)
1485 qmax = MAX_MAX_QPSCALE;
1486 qmax = X265_MIN(qmax, MAX_MAX_QPSCALE);
1487 while (((expectedVbv < rce->expectedVbv/sizeConstraint) && (q < qmax)) ||
1488 ((expectedVbv < 0) && (q < MAX_MAX_QPSCALE)))
1489 {
1490 q *= 1.05;
1491 expectedSize = qScale2bits(rce, q);
1492 expectedVbv = m_bufferFill + m_bufferRate - expectedSize;
1493 }
1494 }
1495 q = Clip3(MIN_QPSCALE, MAX_MAX_QPSCALE, q);
1496 }
1497 else
1498 {
1499 /* 1pass ABR */
1500
1501 /* Calculate the quantizer which would have produced the desired
1502 * average bitrate if it had been applied to all frames so far.
1503 * Then modulate that quant based on the current frame's complexity
1504 * relative to the average complexity so far (using the 2pass RCEQ).
1505 * Then bias the quant up or down if total size so far was far from
1506 * the target.
1507 * Result: Depending on the value of rate_tolerance, there is a
1508 * tradeoff between quality and bitrate precision. But at large
1509 * tolerances, the bit distribution approaches that of 2pass. */
1510
1511 double wantedBits, overflow = 1;
1512
1513 m_shortTermCplxSum *= 0.5;
1514 m_shortTermCplxCount *= 0.5;
1515 m_shortTermCplxSum += m_currentSatd / (CLIP_DURATION(m_frameDuration) / BASE_FRAME_DURATION);
1516 m_shortTermCplxCount++;
1517 /* coeffBits to be used in 2-pass */
1518 rce->coeffBits = (int)m_currentSatd;
1519 rce->blurredComplexity = m_shortTermCplxSum / m_shortTermCplxCount;
1520 rce->mvBits = 0;
1521 rce->sliceType = m_sliceType;
1522
1523 if (m_param->rc.rateControlMode == X265_RC_CRF)
1524 {
1525 q = getQScale(rce, m_rateFactorConstant);
1526 }
1527 else
1528 {
1529 if (!m_param->rc.bStatRead)
1530 checkAndResetABR(rce, false);
1531 q = getQScale(rce, m_wantedBitsWindow / m_cplxrSum);
1532
1533 /* ABR code can potentially be counterproductive in CBR, so just
1534 * don't bother. Don't run it if the frame complexity is zero
1535 * either. */
1536 if (!m_isCbr && m_currentSatd)
1537 {
1538 /* use framesDone instead of POC as poc count is not serial with bframes enabled */
1539 double timeDone = (double)(m_framesDone - m_param->frameNumThreads + 1) * m_frameDuration;
1540 wantedBits = timeDone * m_bitrate;
1541 if (wantedBits > 0 && m_totalBits > 0 && !m_partialResidualFrames)
1542 {
1543 abrBuffer *= X265_MAX(1, sqrt(timeDone));
1544 overflow = Clip3(.5, 2.0, 1.0 + (m_totalBits - wantedBits) / abrBuffer);
1545 q *= overflow;
1546 }
1547 }
1548 }
1549
1550 if (m_sliceType == I_SLICE && m_param->keyframeMax > 1
1551 && m_lastNonBPictType != I_SLICE && !m_isAbrReset)
1552 {
1553 q = x265_qp2qScale(m_accumPQp / m_accumPNorm);
1554 q /= fabs(m_param->rc.ipFactor);
1555 }
1556 else if (m_framesDone > 0)
1557 {
1558 if (m_param->rc.rateControlMode != X265_RC_CRF)
1559 {
1560 double lqmin = 0, lqmax = 0;
1561 lqmin = m_lastQScaleFor[m_sliceType] / m_lstep;
1562 lqmax = m_lastQScaleFor[m_sliceType] * m_lstep;
1563 if (!m_partialResidualFrames)
1564 {
1565 if (overflow > 1.1 && m_framesDone > 3)
1566 lqmax *= m_lstep;
1567 else if (overflow < 0.9)
1568 lqmin /= m_lstep;
1569 }
1570 q = Clip3(lqmin, lqmax, q);
1571 }
1572 }
1573 else if (m_qCompress != 1 && m_param->rc.rateControlMode == X265_RC_CRF)
1574 {
1575 q = x265_qp2qScale(CRF_INIT_QP) / fabs(m_param->rc.ipFactor);
1576 }
1577 else if (m_framesDone == 0 && !m_isVbv)
1578 {
1579 /* for ABR alone, clip the first I frame qp */
1580 double lqmax = x265_qp2qScale(ABR_INIT_QP_MAX) * m_lstep;
1581 q = X265_MIN(lqmax, q);
1582 }
1583 q = Clip3(MIN_QPSCALE, MAX_MAX_QPSCALE, q);
1584 rce->qpNoVbv = x265_qScale2qp(q);
1585 q = clipQscale(curFrame, rce, q);
1586 }
1587 m_lastQScaleFor[m_sliceType] = q;
1588 if ((m_curSlice->m_poc == 0 || m_lastQScaleFor[P_SLICE] < q) && !(m_2pass && !m_isVbv))
1589 m_lastQScaleFor[P_SLICE] = q * fabs(m_param->rc.ipFactor);
1590
1591 if (m_2pass && m_isVbv)
1592 rce->frameSizePlanned = qScale2bits(rce, q);
1593 else
1594 rce->frameSizePlanned = predictSize(&m_pred[m_sliceType], q, (double)m_currentSatd);
1595
1596 /* Always use up the whole VBV in this case. */
1597 if (m_singleFrameVbv)
1598 rce->frameSizePlanned = m_bufferRate;
1599 /* Limit planned size by MinCR */
1600 if (m_isVbv)
1601 rce->frameSizePlanned = X265_MIN(rce->frameSizePlanned, rce->frameSizeMaximum);
1602 rce->frameSizeEstimated = rce->frameSizePlanned;
1603 rce->newQScale = q;
1604 return q;
1605 }
1606}
1607
1608void RateControl::rateControlUpdateStats(RateControlEntry* rce)
1609{
1610 if (!m_param->rc.bStatWrite && !m_param->rc.bStatRead)
1611 {
1612 if (rce->sliceType == I_SLICE)
1613 {
1614 /* previous I still had a residual; roll it into the new loan */
1615 if (m_partialResidualFrames)
1616 rce->rowTotalBits += m_partialResidualCost * m_partialResidualFrames;
1617
1618 m_partialResidualFrames = X265_MIN(s_amortizeFrames, m_param->keyframeMax);
1619 m_partialResidualCost = (int)((rce->rowTotalBits * s_amortizeFraction) /m_partialResidualFrames);
1620 rce->rowTotalBits -= m_partialResidualCost * m_partialResidualFrames;
1621 }
1622 else if (m_partialResidualFrames)
1623 {
1624 rce->rowTotalBits += m_partialResidualCost;
1625 m_partialResidualFrames--;
1626 }
1627 }
1628 if (rce->sliceType != B_SLICE)
1629 rce->rowCplxrSum = rce->rowTotalBits * x265_qp2qScale(rce->qpaRc) / rce->qRceq;
1630 else
1631 rce->rowCplxrSum = rce->rowTotalBits * x265_qp2qScale(rce->qpaRc) / (rce->qRceq * fabs(m_param->rc.pbFactor));
1632
1633 m_cplxrSum += rce->rowCplxrSum;
1634 m_totalBits += rce->rowTotalBits;
1635
1636 /* do not allow the next frame to enter rateControlStart() until this
1637 * frame has updated its mid-frame statistics */
1638 m_startEndOrder.incr();
1639
1640 if (rce->encodeOrder < m_param->frameNumThreads - 1)
1641 m_startEndOrder.incr(); // faked rateControlEnd calls for negative frames
1642}
1643
1644void RateControl::checkAndResetABR(RateControlEntry* rce, bool isFrameDone)
1645{
1646 double abrBuffer = 2 * m_param->rc.rateTolerance * m_bitrate;
1647
1648 // Check if current Slice is a scene cut that follows low detailed/blank frames
1649 if (rce->lastSatd > 4 * rce->movingAvgSum)
1650 {
1651 if (!m_isAbrReset && rce->movingAvgSum > 0)
1652 {
1653 int64_t shrtTermWantedBits = (int64_t) (X265_MIN(m_sliderPos, s_slidingWindowFrames) * m_bitrate * m_frameDuration);
1654 int64_t shrtTermTotalBitsSum = 0;
1655 // Reset ABR if prev frames are blank to prevent further sudden overflows/ high bit rate spikes.
1656 for (int i = 0; i < s_slidingWindowFrames ; i++)
1657 shrtTermTotalBitsSum += m_encodedBitsWindow[i];
1658 double underflow = (shrtTermTotalBitsSum - shrtTermWantedBits) / abrBuffer;
1659 const double epsilon = 0.0001f;
1660 if (underflow < epsilon && !isFrameDone)
1661 {
1662 init(m_curSlice->m_sps);
1663 m_shortTermCplxSum = rce->lastSatd / (CLIP_DURATION(m_frameDuration) / BASE_FRAME_DURATION);
1664 m_shortTermCplxCount = 1;
1665 m_isAbrReset = true;
1666 m_lastAbrResetPoc = rce->poc;
1667 }
1668 }
1669 else
1670 {
1671 // Clear flag to reset ABR and continue as usual.
1672 m_isAbrReset = false;
1673 }
1674 }
1675}
1676
1677void RateControl::hrdFullness(SEIBufferingPeriod *seiBP)
1678{
1679 const VUI* vui = &m_curSlice->m_sps->vuiParameters;
1680 const HRDInfo* hrd = &vui->hrdParameters;
1681 int num = 90000;
1682 int denom = hrd->bitRateValue << (hrd->bitRateScale + BR_SHIFT);
1683 reduceFraction(&num, &denom);
1684 int64_t cpbState = (int64_t)m_bufferFillFinal;
1685 int64_t cpbSize = (int64_t)hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT);
1686
1687 if (cpbState < 0 || cpbState > cpbSize)
1688 {
1689 x265_log(m_param, X265_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n",
1690 cpbState < 0 ? "underflow" : "overflow", (float)cpbState/denom, (float)cpbSize/denom);
1691 }
1692
1693 seiBP->m_initialCpbRemovalDelay = (uint32_t)(num * cpbState + denom) / denom;
1694 seiBP->m_initialCpbRemovalDelayOffset = (uint32_t)(num * cpbSize + denom) / denom - seiBP->m_initialCpbRemovalDelay;
1695}
1696
1697void RateControl::updateVbvPlan(Encoder* enc)
1698{
1699 m_bufferFill = m_bufferFillFinal;
1700 enc->updateVbvPlan(this);
1701}
1702
1703double RateControl::predictSize(Predictor *p, double q, double var)
1704{
1705 return (p->coeff * var + p->offset) / (q * p->count);
1706}
1707
1708double RateControl::clipQscale(Frame* curFrame, RateControlEntry* rce, double q)
1709{
1710 // B-frames are not directly subject to VBV,
1711 // since they are controlled by referenced P-frames' QPs.
1712 double q0 = q;
1713 if (m_isVbv && m_currentSatd > 0 && curFrame)
1714 {
1715 if (m_param->lookaheadDepth || m_param->rc.cuTree ||
1716 m_param->scenecutThreshold ||
1717 (m_param->bFrameAdaptive && m_param->bframes))
1718 {
1719 /* Lookahead VBV: If lookahead is done, raise the quantizer as necessary
1720 * such that no frames in the lookahead overflow and such that the buffer
1721 * is in a reasonable state by the end of the lookahead. */
1722 int loopTerminate = 0;
1723 /* Avoid an infinite loop. */
1724 for (int iterations = 0; iterations < 1000 && loopTerminate != 3; iterations++)
1725 {
1726 double frameQ[3];
1727 double curBits;
1728 if (m_sliceType == B_SLICE)
1729 curBits = predictSize(&m_predBfromP, q, (double)m_currentSatd);
1730 else
1731 curBits = predictSize(&m_pred[m_sliceType], q, (double)m_currentSatd);
1732 double bufferFillCur = m_bufferFill - curBits;
1733 double targetFill;
1734 double totalDuration = 0;
1735 frameQ[P_SLICE] = m_sliceType == I_SLICE ? q * m_param->rc.ipFactor : (m_sliceType == B_SLICE ? q / m_param->rc.pbFactor : q);
1736 frameQ[B_SLICE] = frameQ[P_SLICE] * m_param->rc.pbFactor;
1737 frameQ[I_SLICE] = frameQ[P_SLICE] / m_param->rc.ipFactor;
1738 /* Loop over the planned future frames. */
1739 for (int j = 0; bufferFillCur >= 0; j++)
1740 {
1741 int type = curFrame->m_lowres.plannedType[j];
1742 if (type == X265_TYPE_AUTO)
1743 break;
1744 totalDuration += m_frameDuration;
1745 double wantedFrameSize = m_vbvMaxRate * m_frameDuration;
1746 if (bufferFillCur + wantedFrameSize <= m_bufferSize)
1747 bufferFillCur += wantedFrameSize;
1748 int64_t satd = curFrame->m_lowres.plannedSatd[j] >> (X265_DEPTH - 8);
1749 type = IS_X265_TYPE_I(type) ? I_SLICE : IS_X265_TYPE_B(type) ? B_SLICE : P_SLICE;
1750 curBits = predictSize(&m_pred[type], frameQ[type], (double)satd);
1751 bufferFillCur -= curBits;
1752 }
1753
1754 /* Try to get the buffer at least 50% filled, but don't set an impossible goal. */
1755 targetFill = X265_MIN(m_bufferFill + totalDuration * m_vbvMaxRate * 0.5, m_bufferSize * 0.5);
1756 if (bufferFillCur < targetFill)
1757 {
1758 q *= 1.01;
1759 loopTerminate |= 1;
1760 continue;
1761 }
1762 /* Try to get the buffer no more than 80% filled, but don't set an impossible goal. */
1763 targetFill = Clip3(m_bufferSize * 0.8, m_bufferSize, m_bufferFill - totalDuration * m_vbvMaxRate * 0.5);
1764 if (m_isCbr && bufferFillCur > targetFill)
1765 {
1766 q /= 1.01;
1767 loopTerminate |= 2;
1768 continue;
1769 }
1770 break;
1771 }
1772 q = X265_MAX(q0 / 2, q);
1773 }
1774 else
1775 {
1776 /* Fallback to old purely-reactive algorithm: no lookahead. */
1777 if ((m_sliceType == P_SLICE || m_sliceType == B_SLICE ||
1778 (m_sliceType == I_SLICE && m_lastNonBPictType == I_SLICE)) &&
1779 m_bufferFill / m_bufferSize < 0.5)
1780 {
1781 q /= Clip3(0.5, 1.0, 2.0 * m_bufferFill / m_bufferSize);
1782 }
1783 // Now a hard threshold to make sure the frame fits in VBV.
1784 // This one is mostly for I-frames.
1785 double bits = predictSize(&m_pred[m_sliceType], q, (double)m_currentSatd);
1786
1787 // For small VBVs, allow the frame to use up the entire VBV.
1788 double maxFillFactor;
1789 maxFillFactor = m_bufferSize >= 5 * m_bufferRate ? 2 : 1;
1790 // For single-frame VBVs, request that the frame use up the entire VBV.
1791 double minFillFactor = m_singleFrameVbv ? 1 : 2;
1792
1793 for (int iterations = 0; iterations < 10; iterations++)
1794 {
1795 double qf = 1.0;
1796 if (bits > m_bufferFill / maxFillFactor)
1797 qf = Clip3(0.2, 1.0, m_bufferFill / (maxFillFactor * bits));
1798 q /= qf;
1799 bits *= qf;
1800 if (bits < m_bufferRate / minFillFactor)
1801 q *= bits * minFillFactor / m_bufferRate;
1802 bits = predictSize(&m_pred[m_sliceType], q, (double)m_currentSatd);
1803 }
1804
1805 q = X265_MAX(q0, q);
1806 }
1807
1808 /* Apply MinCR restrictions */
1809 double pbits = predictSize(&m_pred[m_sliceType], q, (double)m_currentSatd);
1810 if (pbits > rce->frameSizeMaximum)
1811 q *= pbits / rce->frameSizeMaximum;
1812
1813 // Check B-frame complexity, and use up any bits that would
1814 // overflow before the next P-frame.
1815 if (m_leadingBframes <= 5 && m_sliceType == P_SLICE && !m_singleFrameVbv)
1816 {
1817 int nb = m_leadingBframes;
1818 double bits = predictSize(&m_pred[m_sliceType], q, (double)m_currentSatd);
1819 double bbits = predictSize(&m_predBfromP, q * m_param->rc.pbFactor, (double)m_currentSatd);
1820 double space;
1821 if (bbits > m_bufferRate)
1822 nb = 0;
1823 double pbbits = nb * bbits;
1824
1825 space = m_bufferFill + (1 + nb) * m_bufferRate - m_bufferSize;
1826 if (pbbits < space)
1827 q *= X265_MAX(pbbits / space, bits / (0.5 * m_bufferSize));
1828
1829 q = X265_MAX(q0 / 2, q);
1830 }
1831
1832 if (!m_isCbr || (m_isAbr && m_currentSatd >= rce->movingAvgSum && q <= q0 / 2))
1833 q = X265_MAX(q0, q);
1834
1835 if (m_rateFactorMaxIncrement)
1836 {
1837 double qpNoVbv = x265_qScale2qp(q0);
1838 double qmax = X265_MIN(MAX_MAX_QPSCALE,x265_qp2qScale(qpNoVbv + m_rateFactorMaxIncrement));
1839 return Clip3(MIN_QPSCALE, qmax, q);
1840 }
1841 }
1842 if (m_2pass)
1843 {
1844 double min = log(MIN_QPSCALE);
1845 double max = log(MAX_MAX_QPSCALE);
1846 q = (log(q) - min) / (max - min) - 0.5;
1847 q = 1.0 / (1.0 + exp(-4 * q));
1848 q = q*(max - min) + min;
1849 return exp(q);
1850 }
1851 return Clip3(MIN_QPSCALE, MAX_MAX_QPSCALE, q);
1852}
1853
1854double RateControl::predictRowsSizeSum(Frame* curFrame, RateControlEntry* rce, double qpVbv, int32_t& encodedBitsSoFar)
1855{
1856 uint32_t rowSatdCostSoFar = 0, totalSatdBits = 0;
1857 encodedBitsSoFar = 0;
1858
1859 double qScale = x265_qp2qScale(qpVbv);
1860 FrameData& curEncData = *curFrame->m_encData;
1861 int picType = curEncData.m_slice->m_sliceType;
1862 Frame* refFrame = curEncData.m_slice->m_refPicList[0][0];
1863
1864 uint32_t maxRows = curEncData.m_slice->m_sps->numCuInHeight;
1865 uint32_t maxCols = curEncData.m_slice->m_sps->numCuInWidth;
1866
1867 for (uint32_t row = 0; row < maxRows; row++)
1868 {
1869 encodedBitsSoFar += curEncData.m_rowStat[row].encodedBits;
1870 rowSatdCostSoFar = curEncData.m_rowStat[row].diagSatd;
1871 uint32_t satdCostForPendingCus = curEncData.m_rowStat[row].satdForVbv - rowSatdCostSoFar;
1872 satdCostForPendingCus >>= X265_DEPTH - 8;
1873 if (satdCostForPendingCus > 0)
1874 {
1875 double pred_s = predictSize(rce->rowPred[0], qScale, satdCostForPendingCus);
1876 uint32_t refRowSatdCost = 0, refRowBits = 0, intraCost = 0;
1877 double refQScale = 0;
1878
1879 if (picType != I_SLICE)
1880 {
1881 FrameData& refEncData = *refFrame->m_encData;
1882 uint32_t endCuAddr = maxCols * (row + 1);
1883 for (uint32_t cuAddr = curEncData.m_rowStat[row].numEncodedCUs + 1; cuAddr < endCuAddr; cuAddr++)
1884 {
1885 refRowSatdCost += refEncData.m_cuStat[cuAddr].vbvCost;
1886 refRowBits += refEncData.m_cuStat[cuAddr].totalBits;
1887 intraCost += curEncData.m_cuStat[cuAddr].intraVbvCost;
1888 }
1889
1890 refRowSatdCost >>= X265_DEPTH - 8;
1891 refQScale = refEncData.m_rowStat[row].diagQpScale;
1892 }
1893
1894 if (picType == I_SLICE || qScale >= refQScale)
1895 {
1896 if (picType == P_SLICE
1897 && !refFrame
1898 && refFrame->m_encData->m_slice->m_sliceType == picType
1899 && refQScale > 0
1900 && refRowSatdCost > 0)
1901 {
1902 if (abs(int32_t(refRowSatdCost - satdCostForPendingCus)) < (int32_t)satdCostForPendingCus / 2)
1903 {
1904 double predTotal = refRowBits * satdCostForPendingCus / refRowSatdCost * refQScale / qScale;
1905 totalSatdBits += int32_t((pred_s + predTotal) * 0.5);
1906 continue;
1907 }
1908 }
1909 totalSatdBits += int32_t(pred_s);
1910 }
1911 else
1912 {
1913 /* Our QP is lower than the reference! */
1914 double pred_intra = predictSize(rce->rowPred[1], qScale, intraCost);
1915 /* Sum: better to overestimate than underestimate by using only one of the two predictors. */
1916 totalSatdBits += int32_t(pred_intra + pred_s);
1917 }
1918 }
1919 }
1920
1921 return totalSatdBits + encodedBitsSoFar;
1922}
1923
1924int RateControl::rowDiagonalVbvRateControl(Frame* curFrame, uint32_t row, RateControlEntry* rce, double& qpVbv)
1925{
1926 FrameData& curEncData = *curFrame->m_encData;
1927 double qScaleVbv = x265_qp2qScale(qpVbv);
1928 uint64_t rowSatdCost = curEncData.m_rowStat[row].diagSatd;
1929 double encodedBits = curEncData.m_rowStat[row].encodedBits;
1930
1931 if (row == 1)
1932 {
1933 rowSatdCost += curEncData.m_rowStat[0].diagSatd;
1934 encodedBits += curEncData.m_rowStat[0].encodedBits;
1935 }
1936 rowSatdCost >>= X265_DEPTH - 8;
1937 updatePredictor(rce->rowPred[0], qScaleVbv, (double)rowSatdCost, encodedBits);
1938 if (curEncData.m_slice->m_sliceType == P_SLICE)
1939 {
1940 Frame* refFrame = curEncData.m_slice->m_refPicList[0][0];
1941 if (qpVbv < refFrame->m_encData->m_rowStat[row].diagQp)
1942 {
1943 uint64_t intraRowSatdCost = curEncData.m_rowStat[row].diagIntraSatd;
1944 if (row == 1)
1945 intraRowSatdCost += curEncData.m_rowStat[0].diagIntraSatd;
1946
1947 updatePredictor(rce->rowPred[1], qScaleVbv, (double)intraRowSatdCost, encodedBits);
1948 }
1949 }
1950
1951 int canReencodeRow = 1;
1952 /* tweak quality based on difference from predicted size */
1953 double prevRowQp = qpVbv;
1954 double qpAbsoluteMax = QP_MAX_MAX;
1955 double qpAbsoluteMin = QP_MIN;
1956 if (m_rateFactorMaxIncrement)
1957 qpAbsoluteMax = X265_MIN(qpAbsoluteMax, rce->qpNoVbv + m_rateFactorMaxIncrement);
1958
1959 if (m_rateFactorMaxDecrement)
1960 qpAbsoluteMin = X265_MAX(qpAbsoluteMin, rce->qpNoVbv - m_rateFactorMaxDecrement);
1961
1962 double qpMax = X265_MIN(prevRowQp + m_param->rc.qpStep, qpAbsoluteMax);
1963 double qpMin = X265_MAX(prevRowQp - m_param->rc.qpStep, qpAbsoluteMin);
1964 double stepSize = 0.5;
1965 double bufferLeftPlanned = rce->bufferFill - rce->frameSizePlanned;
1966
1967 const SPS& sps = *curEncData.m_slice->m_sps;
1968 double maxFrameError = X265_MAX(0.05, 1.0 / sps.numCuInHeight);
1969
1970 if (row < sps.numCuInHeight - 1)
1971 {
1972 /* B-frames shouldn't use lower QP than their reference frames. */
1973 if (rce->sliceType == B_SLICE)
1974 {
1975 Frame* refSlice1 = curEncData.m_slice->m_refPicList[0][0];
1976 Frame* refSlice2 = curEncData.m_slice->m_refPicList[1][0];
1977 qpMin = X265_MAX(qpMin, X265_MAX(refSlice1->m_encData->m_rowStat[row].diagQp, refSlice2->m_encData->m_rowStat[row].diagQp));
1978 qpVbv = X265_MAX(qpVbv, qpMin);
1979 }
1980 /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
1981 double rcTol = bufferLeftPlanned / m_param->frameNumThreads * m_param->rc.rateTolerance;
1982 int32_t encodedBitsSoFar = 0;
1983 double accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar);
1984
1985 /* * Don't increase the row QPs until a sufficent amount of the bits of
1986 * the frame have been processed, in case a flat area at the top of the
1987 * frame was measured inaccurately. */
1988 if (encodedBitsSoFar < 0.05f * rce->frameSizePlanned)
1989 qpMax = qpAbsoluteMax = prevRowQp;
1990
1991 if (rce->sliceType != I_SLICE)
1992 rcTol *= 0.5;
1993
1994 if (!m_isCbr)
1995 qpMin = X265_MAX(qpMin, rce->qpNoVbv);
1996
1997 while (qpVbv < qpMax
1998 && ((accFrameBits > rce->frameSizePlanned + rcTol) ||
1999 (rce->bufferFill - accFrameBits < bufferLeftPlanned * 0.5) ||
2000 (accFrameBits > rce->frameSizePlanned && qpVbv < rce->qpNoVbv)))
2001 {
2002 qpVbv += stepSize;
2003 accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar);
2004 }
2005
2006 while (qpVbv > qpMin
2007 && (qpVbv > curEncData.m_rowStat[0].diagQp || m_singleFrameVbv)
2008 && ((accFrameBits < rce->frameSizePlanned * 0.8f && qpVbv <= prevRowQp)
2009 || accFrameBits < (rce->bufferFill - m_bufferSize + m_bufferRate) * 1.1))
2010 {
2011 qpVbv -= stepSize;
2012 accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar);
2013 }
2014
2015 /* avoid VBV underflow or MinCr violation */
2016 while ((qpVbv < qpAbsoluteMax)
2017 && ((rce->bufferFill - accFrameBits < m_bufferRate * maxFrameError) ||
2018 (rce->frameSizeMaximum - accFrameBits < rce->frameSizeMaximum * maxFrameError)))
2019 {
2020 qpVbv += stepSize;
2021 accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar);
2022 }
2023
2024 rce->frameSizeEstimated = accFrameBits;
2025
2026 /* If the current row was large enough to cause a large QP jump, try re-encoding it. */
2027 if (qpVbv > qpMax && prevRowQp < qpMax && canReencodeRow)
2028 {
2029 /* Bump QP to halfway in between... close enough. */
2030 qpVbv = Clip3(prevRowQp + 1.0f, qpMax, (prevRowQp + qpVbv) * 0.5);
2031 return -1;
2032 }
2033
2034 if (m_param->rc.rfConstantMin)
2035 {
2036 if (qpVbv < qpMin && prevRowQp > qpMin && canReencodeRow)
2037 {
2038 qpVbv = Clip3(qpMin, prevRowQp, (prevRowQp + qpVbv) * 0.5);
2039 return -1;
2040 }
2041 }
2042 }
2043 else
2044 {
2045 int32_t encodedBitsSoFar = 0;
2046 rce->frameSizeEstimated = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar);
2047
2048 /* Last-ditch attempt: if the last row of the frame underflowed the VBV,
2049 * try again. */
2050 if ((rce->frameSizeEstimated > (rce->bufferFill - m_bufferRate * maxFrameError) &&
2051 qpVbv < qpMax && canReencodeRow))
2052 {
2053 qpVbv = qpMax;
2054 return -1;
2055 }
2056 }
2057 return 0;
2058}
2059
2060/* modify the bitrate curve from pass1 for one frame */
2061double RateControl::getQScale(RateControlEntry *rce, double rateFactor)
2062{
2063 double q;
2064
2065 if (m_param->rc.cuTree)
2066 {
2067 // Scale and units are obtained from rateNum and rateDenom for videos with fixed frame rates.
2068 double timescale = (double)m_param->fpsDenom / (2 * m_param->fpsNum);
2069 q = pow(BASE_FRAME_DURATION / CLIP_DURATION(2 * timescale), 1 - m_param->rc.qCompress);
2070 }
2071 else
2072 q = pow(rce->blurredComplexity, 1 - m_param->rc.qCompress);
2073 // avoid NaN's in the Rceq
2074 if (rce->coeffBits + rce->mvBits == 0)
2075 q = m_lastQScaleFor[rce->sliceType];
2076 else
2077 {
2078 m_lastRceq = q;
2079 q /= rateFactor;
2080 }
2081 return q;
2082}
2083
2084void RateControl::updatePredictor(Predictor *p, double q, double var, double bits)
2085{
2086 if (var < 10)
2087 return;
2088 const double range = 1.5;
2089 double old_coeff = p->coeff / p->count;
2090 double new_coeff = bits * q / var;
2091 double new_coeff_clipped = Clip3(old_coeff / range, old_coeff * range, new_coeff);
2092 double new_offset = bits * q - new_coeff_clipped * var;
2093 if (new_offset >= 0)
2094 new_coeff = new_coeff_clipped;
2095 else
2096 new_offset = 0;
2097 p->count *= p->decay;
2098 p->coeff *= p->decay;
2099 p->offset *= p->decay;
2100 p->count++;
2101 p->coeff += new_coeff;
2102 p->offset += new_offset;
2103}
2104
2105void RateControl::updateVbv(int64_t bits, RateControlEntry* rce)
2106{
2107 if (rce->lastSatd >= m_ncu)
2108 updatePredictor(&m_pred[rce->sliceType], x265_qp2qScale(rce->qpaRc), (double)rce->lastSatd, (double)bits);
2109 if (!m_isVbv)
2110 return;
2111
2112 m_bufferFillFinal -= bits;
2113
2114 if (m_bufferFillFinal < 0)
2115 x265_log(m_param, X265_LOG_WARNING, "poc:%d, VBV underflow (%.0f bits)\n", rce->poc, m_bufferFillFinal);
2116
2117 m_bufferFillFinal = X265_MAX(m_bufferFillFinal, 0);
2118 m_bufferFillFinal += m_bufferRate;
2119 m_bufferFillFinal = X265_MIN(m_bufferFillFinal, m_bufferSize);
2120}
2121
2122/* After encoding one frame, update rate control state */
2123int RateControl::rateControlEnd(Frame* curFrame, int64_t bits, RateControlEntry* rce, FrameStats* stats)
2124{
2125 int orderValue = m_startEndOrder.get();
2126 int endOrdinal = (rce->encodeOrder + m_param->frameNumThreads) * 2 - 1;
2127 while (orderValue < endOrdinal && !m_bTerminated)
2128 {
2129 /* no more frames are being encoded, so fake the start event if we would
2130 * have blocked on it. Note that this does not enforce rateControlEnd()
2131 * ordering during flush, but this has no impact on the outputs */
2132 if (m_finalFrameCount && orderValue >= 2 * m_finalFrameCount)
2133 break;
2134 orderValue = m_startEndOrder.waitForChange(orderValue);
2135 }
2136
2137 FrameData& curEncData = *curFrame->m_encData;
2138 int64_t actualBits = bits;
2139 Slice *slice = curEncData.m_slice;
2140 if (m_isAbr)
2141 {
2142 if (m_param->rc.rateControlMode == X265_RC_ABR && !m_param->rc.bStatRead)
2143 checkAndResetABR(rce, true);
2144
2145 if (m_param->rc.rateControlMode == X265_RC_CRF)
2146 {
2147 if (int(curEncData.m_avgQpRc + 0.5) == slice->m_sliceQp)
2148 curEncData.m_rateFactor = m_rateFactorConstant;
2149 else
2150 {
2151 /* If vbv changed the frame QP recalculate the rate-factor */
2152 double baseCplx = m_ncu * (m_param->bframes ? 120 : 80);
2153 double mbtree_offset = m_param->rc.cuTree ? (1.0 - m_param->rc.qCompress) * 13.5 : 0;
2154 curEncData.m_rateFactor = pow(baseCplx, 1 - m_qCompress) /
2155 x265_qp2qScale(int(curEncData.m_avgQpRc + 0.5) + mbtree_offset);
2156 }
2157 }
2158 }
2159
2160 if (m_param->rc.aqMode || m_isVbv)
2161 {
2162 if (m_isVbv)
2163 {
2164 for (uint32_t i = 0; i < slice->m_sps->numCuInHeight; i++)
2165 curEncData.m_avgQpRc += curEncData.m_rowStat[i].sumQpRc;
2166
2167 curEncData.m_avgQpRc /= slice->m_sps->numCUsInFrame;
2168 rce->qpaRc = curEncData.m_avgQpRc;
2169
2170 // copy avg RC qp to m_avgQpAq. To print out the correct qp when aq/cutree is disabled.
2171 curEncData.m_avgQpAq = curEncData.m_avgQpRc;
2172 }
2173
2174 if (m_param->rc.aqMode)
2175 {
2176 for (uint32_t i = 0; i < slice->m_sps->numCuInHeight; i++)
2177 curEncData.m_avgQpAq += curEncData.m_rowStat[i].sumQpAq;
2178
2179 curEncData.m_avgQpAq /= slice->m_sps->numCUsInFrame;
2180 }
2181 }
2182
2183 // Write frame stats into the stats file if 2 pass is enabled.
2184 if (m_param->rc.bStatWrite)
2185 {
2186 char cType = rce->sliceType == I_SLICE ? (rce->poc > 0 && m_param->bOpenGOP ? 'i' : 'I')
2187 : rce->sliceType == P_SLICE ? 'P'
2188 : IS_REFERENCED(curFrame) ? 'B' : 'b';
2189 if (fprintf(m_statFileOut,
2190 "in:%d out:%d type:%c q:%.2f q-aq:%.2f tex:%d mv:%d misc:%d icu:%.2f pcu:%.2f scu:%.2f ;\n",
2191 rce->poc, rce->encodeOrder,
2192 cType, curEncData.m_avgQpRc, curEncData.m_avgQpAq,
2193 stats->coeffBits,
2194 stats->mvBits,
2195 stats->miscBits,
2196 stats->percentIntra * m_ncu,
2197 stats->percentInter * m_ncu,
2198 stats->percentSkip * m_ncu) < 0)
2199 goto writeFailure;
2200 /* Don't re-write the data in multi-pass mode. */
2201 if (m_param->rc.cuTree && IS_REFERENCED(curFrame) && !m_param->rc.bStatRead)
2202 {
2203 uint8_t sliceType = (uint8_t)rce->sliceType;
2204 for (int i = 0; i < m_ncu; i++)
2205 m_cuTreeStats.qpBuffer[0][i] = (uint16_t)(curFrame->m_lowres.qpCuTreeOffset[i] * 256.0);
2206 if (fwrite(&sliceType, 1, 1, m_cutreeStatFileOut) < 1)
2207 goto writeFailure;
2208 if (fwrite(m_cuTreeStats.qpBuffer[0], sizeof(uint16_t), m_ncu, m_cutreeStatFileOut) < (size_t)m_ncu)
2209 goto writeFailure;
2210 }
2211 }
2212 if (m_isAbr && !m_isAbrReset)
2213 {
2214 /* amortize part of each I slice over the next several frames, up to
2215 * keyint-max, to avoid over-compensating for the large I slice cost */
2216 if (!m_param->rc.bStatWrite && !m_param->rc.bStatRead)
2217 {
2218 if (rce->sliceType == I_SLICE)
2219 {
2220 /* previous I still had a residual; roll it into the new loan */
2221 if (m_residualFrames)
2222 bits += m_residualCost * m_residualFrames;
2223 m_residualFrames = X265_MIN(s_amortizeFrames, m_param->keyframeMax);
2224 m_residualCost = (int)((bits * s_amortizeFraction) / m_residualFrames);
2225 bits -= m_residualCost * m_residualFrames;
2226 }
2227 else if (m_residualFrames)
2228 {
2229 bits += m_residualCost;
2230 m_residualFrames--;
2231 }
2232 }
2233 if (rce->sliceType != B_SLICE)
2234 {
2235 /* The factor 1.5 is to tune up the actual bits, otherwise the cplxrSum is scaled too low
2236 * to improve short term compensation for next frame. */
2237 m_cplxrSum += (bits * x265_qp2qScale(rce->qpaRc) / rce->qRceq) - (rce->rowCplxrSum);
2238 }
2239 else
2240 {
2241 /* Depends on the fact that B-frame's QP is an offset from the following P-frame's.
2242 * Not perfectly accurate with B-refs, but good enough. */
2243 m_cplxrSum += (bits * x265_qp2qScale(rce->qpaRc) / (rce->qRceq * fabs(m_param->rc.pbFactor))) - (rce->rowCplxrSum);
2244 }
2245 m_wantedBitsWindow += m_frameDuration * m_bitrate;
2246 m_totalBits += bits - rce->rowTotalBits;
2247 int pos = m_sliderPos - m_param->frameNumThreads;
2248 if (pos >= 0)
2249 m_encodedBitsWindow[pos % s_slidingWindowFrames] = actualBits;
2250 }
2251
2252 if (m_2pass)
2253 {
2254 m_expectedBitsSum += qScale2bits(rce, x265_qp2qScale(rce->newQp));
2255 m_totalBits += bits - rce->rowTotalBits;
2256 }
2257
2258 if (m_isVbv)
2259 {
2260 if (rce->sliceType == B_SLICE)
2261 {
2262 m_bframeBits += actualBits;
2263 if (rce->bLastMiniGopBFrame)
2264 {
2265 if (rce->bframes != 0)
2266 updatePredictor(&m_predBfromP, x265_qp2qScale(rce->qpaRc), (double)rce->leadingNoBSatd, (double)m_bframeBits / rce->bframes);
2267 m_bframeBits = 0;
2268 }
2269 }
2270 updateVbv(actualBits, rce);
2271
2272 if (m_param->bEmitHRDSEI)
2273 {
2274 const VUI *vui = &curEncData.m_slice->m_sps->vuiParameters;
2275 const HRDInfo *hrd = &vui->hrdParameters;
2276 const TimingInfo *time = &vui->timingInfo;
2277 if (!curFrame->m_poc)
2278 {
2279 // first access unit initializes the HRD
2280 rce->hrdTiming->cpbInitialAT = 0;
2281 rce->hrdTiming->cpbRemovalTime = m_nominalRemovalTime = (double)m_bufPeriodSEI.m_initialCpbRemovalDelay / 90000;
2282 }
2283 else
2284 {
2285 rce->hrdTiming->cpbRemovalTime = m_nominalRemovalTime + (double)rce->picTimingSEI->m_auCpbRemovalDelay * time->numUnitsInTick / time->timeScale;
2286 double cpbEarliestAT = rce->hrdTiming->cpbRemovalTime - (double)m_bufPeriodSEI.m_initialCpbRemovalDelay / 90000;
2287 if (!curFrame->m_lowres.bKeyframe)
2288 cpbEarliestAT -= (double)m_bufPeriodSEI.m_initialCpbRemovalDelayOffset / 90000;
2289
2290 rce->hrdTiming->cpbInitialAT = hrd->cbrFlag ? m_prevCpbFinalAT : X265_MAX(m_prevCpbFinalAT, cpbEarliestAT);
2291 }
2292
2293 uint32_t cpbsizeUnscale = hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT);
2294 rce->hrdTiming->cpbFinalAT = m_prevCpbFinalAT = rce->hrdTiming->cpbInitialAT + actualBits / cpbsizeUnscale;
2295 rce->hrdTiming->dpbOutputTime = (double)rce->picTimingSEI->m_picDpbOutputDelay * time->numUnitsInTick / time->timeScale + rce->hrdTiming->cpbRemovalTime;
2296 }
2297 }
2298 // Allow rateControlStart of next frame only when rateControlEnd of previous frame is over
2299 m_startEndOrder.incr();
2300 rce->isActive = false;
2301 return 0;
2302
2303writeFailure:
2304 x265_log(m_param, X265_LOG_ERROR, "RatecontrolEnd: stats file write failure\n");
2305 return 1;
2306}
2307
2308#if defined(_MSC_VER)
2309#pragma warning(disable: 4996) // POSIX function names are just fine, thank you
2310#endif
2311
2312/* called when the encoder is flushing, and thus the final frame count is
2313 * unambiguously known */
2314void RateControl::setFinalFrameCount(int count)
2315{
2316 m_finalFrameCount = count;
2317 /* unblock waiting threads */
2318 m_startEndOrder.set(m_startEndOrder.get());
2319}
2320
2321/* called when the encoder is closing, and no more frames will be output.
2322 * all blocked functions must finish so the frame encoder threads can be
2323 * closed */
2324void RateControl::terminate()
2325{
2326 m_bTerminated = true;
2327 /* unblock waiting threads */
2328 m_startEndOrder.set(m_startEndOrder.get());
2329}
2330
2331void RateControl::destroy()
2332{
2333 const char *fileName = m_param->rc.statFileName;
2334 if (!fileName)
2335 fileName = s_defaultStatFileName;
2336
2337 if (m_statFileOut)
2338 {
2339 fclose(m_statFileOut);
2340 char *tmpFileName = strcatFilename(fileName, ".temp");
2341 int bError = 1;
2342 if (tmpFileName)
2343 {
2344 unlink(fileName);
2345 bError = rename(tmpFileName, fileName);
2346 }
2347 if (bError)
2348 {
2349 x265_log(m_param, X265_LOG_ERROR, "failed to rename output stats file to \"%s\"\n",
2350 fileName);
2351 }
2352 X265_FREE(tmpFileName);
2353 }
2354
2355 if (m_cutreeStatFileOut)
2356 {
2357 fclose(m_cutreeStatFileOut);
2358 char *tmpFileName = strcatFilename(fileName, ".cutree.temp");
2359 char *newFileName = strcatFilename(fileName, ".cutree");
2360 int bError = 1;
2361 if (tmpFileName && newFileName)
2362 {
2363 unlink(newFileName);
2364 bError = rename(tmpFileName, newFileName);
2365 }
2366 if (bError)
2367 {
2368 x265_log(m_param, X265_LOG_ERROR, "failed to rename cutree output stats file to \"%s\"\n",
2369 newFileName);
2370 }
2371 X265_FREE(tmpFileName);
2372 X265_FREE(newFileName);
2373 }
2374
2375 if (m_cutreeStatFileIn)
2376 fclose(m_cutreeStatFileIn);
2377
2378 X265_FREE(m_rce2Pass);
2379 for (int i = 0; i < 2; i++)
2380 X265_FREE(m_cuTreeStats.qpBuffer[i]);
2381}
2382