/src/x265/source/encoder/ratecontrol.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | /***************************************************************************** |
2 | | * Copyright (C) 2013-2020 MulticoreWare, Inc |
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 | | #if _MSC_VER |
27 | | #pragma warning(disable: 4127) // conditional expression is constant, yes I know |
28 | | #endif |
29 | | |
30 | | #include "common.h" |
31 | | #include "param.h" |
32 | | #include "frame.h" |
33 | | #include "framedata.h" |
34 | | #include "picyuv.h" |
35 | | |
36 | | #include "encoder.h" |
37 | | #include "slicetype.h" |
38 | | #include "ratecontrol.h" |
39 | | #include "sei.h" |
40 | | |
41 | 0 | #define BR_SHIFT 6 |
42 | 0 | #define CPB_SHIFT 4 |
43 | | |
44 | | using namespace X265_NS; |
45 | | |
46 | | /* Amortize the partial cost of I frames over the next N frames */ |
47 | | |
48 | | const int RateControl::s_slidingWindowFrames = 20; |
49 | | const char *RateControl::s_defaultStatFileName = "x265_2pass.log"; |
50 | | |
51 | | namespace { |
52 | 0 | #define CMP_OPT_FIRST_PASS(opt, param_val)\ |
53 | 0 | {\ |
54 | 0 | bErr = 0;\ |
55 | 0 | p = strstr(opts, opt "=");\ |
56 | 0 | char* q = strstr(opts, "no-" opt " ");\ |
57 | 0 | if (p && sscanf(p, opt "=%d" , &i) && param_val != i)\ |
58 | 0 | bErr = 1;\ |
59 | 0 | else if (!param_val && !q && !p)\ |
60 | 0 | bErr = 1;\ |
61 | 0 | else if (param_val && (q || !strstr(opts, opt)))\ |
62 | 0 | bErr = 1;\ |
63 | 0 | if (bErr)\ |
64 | 0 | {\ |
65 | 0 | x265_log(m_param, X265_LOG_ERROR, "different " opt " setting than first pass (%d vs %d)\n", param_val, i);\ |
66 | 0 | return false;\ |
67 | 0 | }\ |
68 | 0 | } |
69 | | |
70 | | inline int calcScale(uint32_t x) |
71 | 0 | { |
72 | 0 | static uint8_t lut[16] = {4, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0}; |
73 | 0 | int y, z = (((x & 0xffff) - 1) >> 27) & 16; |
74 | 0 | x >>= z; |
75 | 0 | z += y = (((x & 0xff) - 1) >> 28) & 8; |
76 | 0 | x >>= y; |
77 | 0 | z += y = (((x & 0xf) - 1) >> 29) & 4; |
78 | 0 | x >>= y; |
79 | 0 | return z + lut[x&0xf]; |
80 | 0 | } |
81 | | |
82 | | inline int calcLength(uint32_t x) |
83 | 0 | { |
84 | 0 | static uint8_t lut[16] = {4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0}; |
85 | 0 | int y, z = (((x >> 16) - 1) >> 27) & 16; |
86 | 0 | x >>= z ^ 16; |
87 | 0 | z += y = ((x - 0x100) >> 28) & 8; |
88 | 0 | x >>= y ^ 8; |
89 | 0 | z += y = ((x - 0x10) >> 29) & 4; |
90 | 0 | x >>= y ^ 4; |
91 | 0 | return z + lut[x]; |
92 | 0 | } |
93 | | |
94 | | inline char *strcatFilename(const char *input, const char *suffix) |
95 | 0 | { |
96 | 0 | char *output = X265_MALLOC(char, strlen(input) + strlen(suffix) + 1); |
97 | 0 | if (!output) |
98 | 0 | { |
99 | 0 | x265_log(NULL, X265_LOG_ERROR, "unable to allocate memory for filename\n"); |
100 | 0 | return NULL; |
101 | 0 | } |
102 | 0 | strcpy(output, input); |
103 | 0 | strcat(output, suffix); |
104 | 0 | return output; |
105 | 0 | } |
106 | | |
107 | | inline double qScale2bits(RateControlEntry *rce, double qScale) |
108 | 0 | { |
109 | 0 | if (qScale < 0.1) |
110 | 0 | qScale = 0.1; |
111 | 0 | return (rce->coeffBits + .1) * pow(rce->qScale / qScale, 1.1) |
112 | 0 | + rce->mvBits * pow(X265_MAX(rce->qScale, 1) / X265_MAX(qScale, 1), 0.5) |
113 | 0 | + rce->miscBits; |
114 | 0 | } |
115 | | |
116 | | inline void copyRceData(RateControlEntry* rce, RateControlEntry* rce2Pass) |
117 | 0 | { |
118 | 0 | rce->coeffBits = rce2Pass->coeffBits; |
119 | 0 | rce->mvBits = rce2Pass->mvBits; |
120 | 0 | rce->miscBits = rce2Pass->miscBits; |
121 | 0 | rce->iCuCount = rce2Pass->iCuCount; |
122 | 0 | rce->pCuCount = rce2Pass->pCuCount; |
123 | 0 | rce->skipCuCount = rce2Pass->skipCuCount; |
124 | 0 | rce->keptAsRef = rce2Pass->keptAsRef; |
125 | 0 | rce->qScale = rce2Pass->qScale; |
126 | 0 | rce->newQScale = rce2Pass->newQScale; |
127 | 0 | rce->expectedBits = rce2Pass->expectedBits; |
128 | 0 | rce->expectedVbv = rce2Pass->expectedVbv; |
129 | 0 | rce->blurredComplexity = rce2Pass->blurredComplexity; |
130 | 0 | rce->sliceType = rce2Pass->sliceType; |
131 | 0 | rce->qpNoVbv = rce2Pass->qpNoVbv; |
132 | 0 | rce->newQp = rce2Pass->newQp; |
133 | 0 | rce->qRceq = rce2Pass->qRceq; |
134 | 0 | } |
135 | | |
136 | | } // end anonymous namespace |
137 | | /* Returns the zone for the current frame */ |
138 | | x265_zone* RateControl::getZone() |
139 | 0 | { |
140 | 0 | for (int i = m_param->rc.zoneCount - 1; i >= 0; i--) |
141 | 0 | { |
142 | 0 | x265_zone *z = &m_param->rc.zones[i]; |
143 | 0 | if (m_framesDone + 1 >= z->startFrame && m_framesDone < z->endFrame) |
144 | 0 | return z; |
145 | 0 | } |
146 | 0 | return NULL; |
147 | 0 | } |
148 | | |
149 | | RateControl::RateControl(x265_param& p, Encoder *top) |
150 | 0 | { |
151 | 0 | m_param = &p; |
152 | 0 | m_top = top; |
153 | 0 | int lowresCuWidth = ((m_param->sourceWidth / 2) + X265_LOWRES_CU_SIZE - 1) >> X265_LOWRES_CU_BITS; |
154 | 0 | int lowresCuHeight = ((m_param->sourceHeight / 2) + X265_LOWRES_CU_SIZE - 1) >> X265_LOWRES_CU_BITS; |
155 | 0 | m_ncu = lowresCuWidth * lowresCuHeight; |
156 | |
|
157 | 0 | m_qCompress = (m_param->rc.cuTree && !m_param->rc.hevcAq) ? 1 : m_param->rc.qCompress; |
158 | | |
159 | | // validate for param->rc, maybe it is need to add a function like x265_parameters_valiate() |
160 | 0 | m_zoneBufferIdx = 0; |
161 | 0 | m_residualFrames = 0; |
162 | 0 | m_partialResidualFrames = 0; |
163 | 0 | m_residualCost = 0; |
164 | 0 | m_partialResidualCost = 0; |
165 | 0 | m_rateFactorMaxIncrement = 0; |
166 | 0 | m_rateFactorMaxDecrement = 0; |
167 | 0 | m_fps = (double)m_param->fpsNum / m_param->fpsDenom; |
168 | 0 | m_startEndOrder.set(0); |
169 | 0 | m_bTerminated = false; |
170 | 0 | m_finalFrameCount = 0; |
171 | 0 | m_numEntries = 0; |
172 | 0 | m_isSceneTransition = false; |
173 | 0 | m_lastPredictorReset = 0; |
174 | 0 | m_avgPFrameQp = 0; |
175 | 0 | m_isFirstMiniGop = false; |
176 | 0 | m_lastScenecut = -1; |
177 | 0 | m_lastScenecutAwareIFrame = -1; |
178 | 0 | if (m_param->rc.rateControlMode == X265_RC_CRF) |
179 | 0 | { |
180 | 0 | m_param->rc.qp = (int)m_param->rc.rfConstant; |
181 | 0 | m_param->rc.bitrate = 0; |
182 | |
|
183 | 0 | double baseCplx = m_ncu * (m_param->bframes ? 120 : 80); |
184 | 0 | double mbtree_offset = m_param->rc.cuTree ? (1.0 - m_param->rc.qCompress) * 13.5 : 0; |
185 | 0 | m_rateFactorConstant = pow(baseCplx, 1 - m_qCompress) / |
186 | 0 | x265_qp2qScale(m_param->rc.rfConstant + mbtree_offset); |
187 | 0 | if (m_param->rc.rfConstantMax) |
188 | 0 | { |
189 | 0 | m_rateFactorMaxIncrement = m_param->rc.rfConstantMax - m_param->rc.rfConstant; |
190 | 0 | if (m_rateFactorMaxIncrement <= 0) |
191 | 0 | { |
192 | 0 | x265_log(m_param, X265_LOG_WARNING, "CRF max must be greater than CRF\n"); |
193 | 0 | m_rateFactorMaxIncrement = 0; |
194 | 0 | } |
195 | 0 | } |
196 | 0 | if (m_param->rc.rfConstantMin) |
197 | 0 | m_rateFactorMaxDecrement = m_param->rc.rfConstant - m_param->rc.rfConstantMin; |
198 | 0 | } |
199 | 0 | m_isAbr = m_param->rc.rateControlMode != X265_RC_CQP && !m_param->rc.bStatRead; |
200 | 0 | m_2pass = m_param->rc.rateControlMode != X265_RC_CQP && m_param->rc.bStatRead; |
201 | 0 | m_bitrate = m_param->rc.bitrate * 1000; |
202 | 0 | m_frameDuration = (double)m_param->fpsDenom / m_param->fpsNum; |
203 | 0 | m_qp = m_param->rc.qp; |
204 | 0 | m_lastRceq = 1; /* handles the cmplxrsum when the previous frame cost is zero */ |
205 | 0 | m_shortTermCplxSum = 0; |
206 | 0 | m_shortTermCplxCount = 0; |
207 | 0 | m_lastNonBPictType = I_SLICE; |
208 | 0 | m_isAbrReset = false; |
209 | 0 | m_lastAbrResetPoc = -1; |
210 | 0 | m_statFileOut = NULL; |
211 | 0 | m_cutreeStatFileOut = m_cutreeStatFileIn = NULL; |
212 | 0 | m_rce2Pass = NULL; |
213 | 0 | m_encOrder = NULL; |
214 | 0 | m_lastBsliceSatdCost = 0; |
215 | 0 | m_movingAvgSum = 0.0; |
216 | 0 | m_isNextGop = false; |
217 | 0 | m_relativeComplexity = NULL; |
218 | | |
219 | | // vbv initialization |
220 | 0 | m_param->rc.vbvBufferSize = x265_clip3(0, 2000000, m_param->rc.vbvBufferSize); |
221 | 0 | m_param->rc.vbvMaxBitrate = x265_clip3(0, 2000000, m_param->rc.vbvMaxBitrate); |
222 | 0 | m_param->rc.vbvBufferInit = x265_clip3(0.0, 2000000.0, m_param->rc.vbvBufferInit); |
223 | 0 | m_param->vbvBufferEnd = x265_clip3(0.0, 2000000.0, m_param->vbvBufferEnd); |
224 | 0 | m_initVbv = false; |
225 | 0 | m_singleFrameVbv = 0; |
226 | 0 | m_rateTolerance = 1.0; |
227 | |
|
228 | 0 | if (m_param->rc.vbvBufferSize) |
229 | 0 | { |
230 | 0 | if (m_param->rc.rateControlMode == X265_RC_CQP) |
231 | 0 | { |
232 | 0 | x265_log(m_param, X265_LOG_WARNING, "VBV is incompatible with constant QP, ignored.\n"); |
233 | 0 | m_param->rc.vbvBufferSize = 0; |
234 | 0 | m_param->rc.vbvMaxBitrate = 0; |
235 | 0 | } |
236 | 0 | else if (m_param->rc.vbvMaxBitrate == 0) |
237 | 0 | { |
238 | 0 | if (m_param->rc.rateControlMode == X265_RC_ABR) |
239 | 0 | { |
240 | 0 | x265_log(m_param, X265_LOG_WARNING, "VBV maxrate unspecified, assuming CBR\n"); |
241 | 0 | m_param->rc.vbvMaxBitrate = m_param->rc.bitrate; |
242 | 0 | } |
243 | 0 | else |
244 | 0 | { |
245 | 0 | x265_log(m_param, X265_LOG_WARNING, "VBV bufsize set but maxrate unspecified, ignored\n"); |
246 | 0 | m_param->rc.vbvBufferSize = 0; |
247 | 0 | } |
248 | 0 | } |
249 | 0 | else if (m_param->rc.vbvMaxBitrate < m_param->rc.bitrate && |
250 | 0 | m_param->rc.rateControlMode == X265_RC_ABR) |
251 | 0 | { |
252 | 0 | x265_log(m_param, X265_LOG_WARNING, "max bitrate less than average bitrate, assuming CBR\n"); |
253 | 0 | m_param->rc.bitrate = m_param->rc.vbvMaxBitrate; |
254 | 0 | } |
255 | 0 | } |
256 | 0 | else if (m_param->rc.vbvMaxBitrate) |
257 | 0 | { |
258 | 0 | x265_log(m_param, X265_LOG_WARNING, "VBV maxrate specified, but no bufsize, ignored\n"); |
259 | 0 | m_param->rc.vbvMaxBitrate = 0; |
260 | 0 | } |
261 | 0 | m_isVbv = m_param->rc.vbvMaxBitrate > 0 && m_param->rc.vbvBufferSize > 0; |
262 | 0 | if (m_param->vbvBufferEnd && !m_isVbv) |
263 | 0 | { |
264 | 0 | x265_log(m_param, X265_LOG_WARNING, "vbv-end requires VBV parameters, ignored\n"); |
265 | 0 | m_param->vbvBufferEnd = 0; |
266 | 0 | } |
267 | 0 | if (m_param->bEmitHRDSEI && !m_isVbv) |
268 | 0 | { |
269 | 0 | x265_log(m_param, X265_LOG_WARNING, "NAL HRD parameters require VBV parameters, ignored\n"); |
270 | 0 | m_param->bEmitHRDSEI = 0; |
271 | 0 | } |
272 | 0 | m_isCbr = m_param->rc.rateControlMode == X265_RC_ABR && m_isVbv && m_param->rc.vbvMaxBitrate <= m_param->rc.bitrate; |
273 | 0 | if (m_param->rc.bStrictCbr && !m_isCbr) |
274 | 0 | { |
275 | 0 | x265_log(m_param, X265_LOG_WARNING, "strict CBR set without CBR mode, ignored\n"); |
276 | 0 | m_param->rc.bStrictCbr = 0; |
277 | 0 | } |
278 | 0 | if(m_param->rc.bStrictCbr) |
279 | 0 | m_rateTolerance = 0.7; |
280 | |
|
281 | 0 | m_bframeBits = 0; |
282 | 0 | m_leadingNoBSatd = 0; |
283 | 0 | m_ipOffset = 6.0 * X265_LOG2(m_param->rc.ipFactor); |
284 | 0 | m_pbOffset = 6.0 * X265_LOG2(m_param->rc.pbFactor); |
285 | |
|
286 | 0 | for (int i = 0; i < QP_MAX_MAX; i++) |
287 | 0 | m_qpToEncodedBits[i] = 0; |
288 | | |
289 | | /* Adjust the first frame in order to stabilize the quality level compared to the rest */ |
290 | 0 | #define ABR_INIT_QP_MIN (24) |
291 | 0 | #define ABR_INIT_QP_MAX (37) |
292 | 0 | #define ABR_INIT_QP_GRAIN_MAX (33) |
293 | 0 | #define ABR_SCENECUT_INIT_QP_MIN (12) |
294 | 0 | #define CRF_INIT_QP (int)m_param->rc.rfConstant |
295 | 0 | for (int i = 0; i < 3; i++) |
296 | 0 | { |
297 | 0 | m_lastQScaleFor[i] = x265_qp2qScale(m_param->rc.rateControlMode == X265_RC_CRF ? CRF_INIT_QP : ABR_INIT_QP_MIN); |
298 | 0 | m_lmin[i] = x265_qp2qScale(m_param->rc.qpMin); |
299 | 0 | m_lmax[i] = x265_qp2qScale(m_param->rc.qpMax); |
300 | 0 | } |
301 | |
|
302 | 0 | if (m_param->rc.rateControlMode == X265_RC_CQP) |
303 | 0 | { |
304 | 0 | if (m_qp && !m_param->bLossless) |
305 | 0 | { |
306 | 0 | m_qpConstant[P_SLICE] = m_qp; |
307 | 0 | m_qpConstant[I_SLICE] = x265_clip3(QP_MIN, QP_MAX_MAX, (int)(m_qp - m_ipOffset + 0.5)); |
308 | 0 | m_qpConstant[B_SLICE] = x265_clip3(QP_MIN, QP_MAX_MAX, (int)(m_qp + m_pbOffset + 0.5)); |
309 | 0 | } |
310 | 0 | else |
311 | 0 | { |
312 | 0 | m_qpConstant[P_SLICE] = m_qpConstant[I_SLICE] = m_qpConstant[B_SLICE] = m_qp; |
313 | 0 | } |
314 | 0 | } |
315 | | |
316 | | /* qpstep - value set as encoder specific */ |
317 | 0 | m_lstep = pow(2, m_param->rc.qpStep / 6.0); |
318 | |
|
319 | 0 | for (int i = 0; i < 2; i++) |
320 | 0 | m_cuTreeStats.qpBuffer[i] = NULL; |
321 | 0 | } |
322 | | |
323 | | bool RateControl::init(const SPS& sps) |
324 | 0 | { |
325 | 0 | if (m_isVbv && !m_initVbv) |
326 | 0 | { |
327 | | /* We don't support changing the ABR bitrate right now, |
328 | | * so if the stream starts as CBR, keep it CBR. */ |
329 | 0 | if (m_param->rc.vbvBufferSize < (int)(m_param->rc.vbvMaxBitrate / m_fps)) |
330 | 0 | { |
331 | 0 | m_param->rc.vbvBufferSize = (int)(m_param->rc.vbvMaxBitrate / m_fps); |
332 | 0 | x265_log(m_param, X265_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n", |
333 | 0 | m_param->rc.vbvBufferSize); |
334 | 0 | } |
335 | 0 | int vbvBufferSize = m_param->rc.vbvBufferSize * 1000; |
336 | 0 | int vbvMaxBitrate = m_param->rc.vbvMaxBitrate * 1000; |
337 | |
|
338 | 0 | if (m_param->bEmitHRDSEI && !m_param->decoderVbvMaxRate) |
339 | 0 | { |
340 | 0 | const HRDInfo* hrd = &sps.vuiParameters.hrdParameters; |
341 | 0 | vbvBufferSize = hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT); |
342 | 0 | vbvMaxBitrate = hrd->bitRateValue << (hrd->bitRateScale + BR_SHIFT); |
343 | 0 | } |
344 | 0 | m_bufferRate = vbvMaxBitrate / m_fps; |
345 | 0 | m_vbvMaxRate = vbvMaxBitrate; |
346 | 0 | m_bufferSize = vbvBufferSize; |
347 | 0 | m_singleFrameVbv = m_bufferRate * 1.1 > m_bufferSize; |
348 | |
|
349 | 0 | if (m_param->rc.vbvBufferInit > 1.) |
350 | 0 | m_param->rc.vbvBufferInit = x265_clip3(0.0, 1.0, m_param->rc.vbvBufferInit / m_param->rc.vbvBufferSize); |
351 | 0 | if (m_param->vbvBufferEnd > 1.) |
352 | 0 | m_param->vbvBufferEnd = x265_clip3(0.0, 1.0, m_param->vbvBufferEnd / m_param->rc.vbvBufferSize); |
353 | 0 | if (m_param->vbvEndFrameAdjust > 1.) |
354 | 0 | m_param->vbvEndFrameAdjust = x265_clip3(0.0, 1.0, m_param->vbvEndFrameAdjust); |
355 | 0 | m_param->rc.vbvBufferInit = x265_clip3(0.0, 1.0, X265_MAX(m_param->rc.vbvBufferInit, m_bufferRate / m_bufferSize)); |
356 | 0 | m_bufferFillFinal = m_bufferSize * m_param->rc.vbvBufferInit; |
357 | 0 | m_bufferFillActual = m_bufferFillFinal; |
358 | 0 | m_bufferExcess = 0; |
359 | 0 | m_initVbv = true; |
360 | 0 | } |
361 | |
|
362 | 0 | if (!m_param->bResetZoneConfig && (m_relativeComplexity == NULL)) |
363 | 0 | { |
364 | 0 | m_relativeComplexity = X265_MALLOC(double, m_param->reconfigWindowSize); |
365 | 0 | if (m_relativeComplexity == NULL) |
366 | 0 | { |
367 | 0 | x265_log(m_param, X265_LOG_ERROR, "Failed to allocate memory for m_relativeComplexity\n"); |
368 | 0 | return false; |
369 | 0 | } |
370 | 0 | } |
371 | | |
372 | 0 | m_totalBits = 0; |
373 | 0 | m_encodedBits = 0; |
374 | 0 | m_framesDone = 0; |
375 | 0 | m_residualCost = 0; |
376 | 0 | m_partialResidualCost = 0; |
377 | 0 | m_amortizeFraction = 0.85; |
378 | 0 | m_amortizeFrames = 75; |
379 | 0 | if (m_param->totalFrames && m_param->totalFrames <= 2 * m_fps && m_param->rc.bStrictCbr) /* Strict CBR segment encode */ |
380 | 0 | { |
381 | 0 | m_amortizeFraction = 0.85; |
382 | 0 | m_amortizeFrames = m_param->totalFrames / 2; |
383 | 0 | } |
384 | |
|
385 | 0 | for (int i = 0; i < s_slidingWindowFrames; i++) |
386 | 0 | { |
387 | 0 | m_satdCostWindow[i] = 0; |
388 | 0 | m_encodedBitsWindow[i] = 0; |
389 | 0 | } |
390 | 0 | m_sliderPos = 0; |
391 | 0 | m_isPatternPresent = false; |
392 | 0 | m_numBframesInPattern = 0; |
393 | |
|
394 | 0 | m_isGrainEnabled = false; |
395 | 0 | if(m_param->rc.bEnableGrain) // tune for grainy content OR equal p-b frame sizes |
396 | 0 | m_isGrainEnabled = true; |
397 | 0 | for (int i = 0; i < 3; i++) |
398 | 0 | m_lastQScaleFor[i] = x265_qp2qScale(m_param->rc.rateControlMode == X265_RC_CRF ? CRF_INIT_QP : ABR_INIT_QP_MIN); |
399 | 0 | m_avgPFrameQp = 0 ; |
400 | | |
401 | | /* 720p videos seem to be a good cutoff for cplxrSum */ |
402 | 0 | double tuneCplxFactor = (m_ncu > 3600 && m_param->rc.cuTree && !m_param->rc.hevcAq) ? 2.5 : m_param->rc.hevcAq ? 1.5 : m_isGrainEnabled ? 1.9 : 1.0; |
403 | | |
404 | | /* estimated ratio that produces a reasonable QP for the first I-frame */ |
405 | 0 | m_cplxrSum = .01 * pow(7.0e5, m_qCompress) * pow(m_ncu, 0.5) * tuneCplxFactor; |
406 | 0 | m_wantedBitsWindow = m_bitrate * m_frameDuration; |
407 | 0 | m_accumPNorm = .01; |
408 | 0 | m_accumPQp = (m_param->rc.rateControlMode == X265_RC_CRF ? CRF_INIT_QP : ABR_INIT_QP_MIN) * m_accumPNorm; |
409 | | |
410 | | |
411 | | /* Frame Predictors used in vbv */ |
412 | 0 | initFramePredictors(); |
413 | 0 | if (!m_statFileOut && (m_param->rc.bStatWrite || m_param->rc.bStatRead)) |
414 | 0 | { |
415 | | /* If the user hasn't defined the stat filename, use the default value */ |
416 | 0 | const char *fileName = m_param->rc.statFileName; |
417 | 0 | if (!fileName) |
418 | 0 | fileName = s_defaultStatFileName; |
419 | | /* Load stat file and init 2pass algo */ |
420 | 0 | if (m_param->rc.bStatRead) |
421 | 0 | { |
422 | 0 | m_expectedBitsSum = 0; |
423 | 0 | char *p, *statsIn, *statsBuf; |
424 | | /* read 1st pass stats */ |
425 | 0 | statsIn = statsBuf = x265_slurp_file(fileName); |
426 | 0 | if (!statsBuf) |
427 | 0 | return false; |
428 | 0 | if (m_param->rc.cuTree) |
429 | 0 | { |
430 | 0 | char *tmpFile = strcatFilename(fileName, ".cutree"); |
431 | 0 | if (!tmpFile) |
432 | 0 | return false; |
433 | 0 | m_cutreeStatFileIn = x265_fopen(tmpFile, "rb"); |
434 | 0 | X265_FREE(tmpFile); |
435 | 0 | if (!m_cutreeStatFileIn) |
436 | 0 | { |
437 | 0 | x265_log_file(m_param, X265_LOG_ERROR, "can't open stats file %s.cutree\n", fileName); |
438 | 0 | return false; |
439 | 0 | } |
440 | 0 | } |
441 | | |
442 | | /* check whether 1st pass options were compatible with current options */ |
443 | 0 | if (strncmp(statsBuf, "#options:", 9)) |
444 | 0 | { |
445 | 0 | x265_log(m_param, X265_LOG_ERROR,"options list in stats file not valid\n"); |
446 | 0 | return false; |
447 | 0 | } |
448 | 0 | { |
449 | 0 | int i, j, m; |
450 | 0 | uint32_t k , l; |
451 | 0 | bool bErr = false; |
452 | 0 | char *opts = statsBuf; |
453 | 0 | statsIn = strchr(statsBuf, '\n'); |
454 | 0 | if (!statsIn) |
455 | 0 | { |
456 | 0 | x265_log(m_param, X265_LOG_ERROR, "Malformed stats file\n"); |
457 | 0 | return false; |
458 | 0 | } |
459 | 0 | *statsIn = '\0'; |
460 | 0 | statsIn++; |
461 | 0 | if ((p = strstr(opts, " input-res=")) == 0 || sscanf(p, " input-res=%dx%d", &i, &j) != 2) |
462 | 0 | { |
463 | 0 | x265_log(m_param, X265_LOG_ERROR, "Resolution specified in stats file not valid\n"); |
464 | 0 | return false; |
465 | 0 | } |
466 | 0 | if ((p = strstr(opts, " fps=")) == 0 || sscanf(p, " fps=%u/%u", &k, &l) != 2) |
467 | 0 | { |
468 | 0 | x265_log(m_param, X265_LOG_ERROR, "fps specified in stats file not valid\n"); |
469 | 0 | return false; |
470 | 0 | } |
471 | 0 | if (((p = strstr(opts, " vbv-maxrate=")) == 0 || sscanf(p, " vbv-maxrate=%d", &m) != 1) && m_param->rc.rateControlMode == X265_RC_CRF) |
472 | 0 | { |
473 | 0 | x265_log(m_param, X265_LOG_ERROR, "Constant rate-factor is incompatible with 2pass without vbv-maxrate in the previous pass\n"); |
474 | 0 | return false; |
475 | 0 | } |
476 | 0 | if (k != m_param->fpsNum || l != m_param->fpsDenom) |
477 | 0 | { |
478 | 0 | x265_log(m_param, X265_LOG_ERROR, "fps mismatch with 1st pass (%u/%u vs %u/%u)\n", |
479 | 0 | m_param->fpsNum, m_param->fpsDenom, k, l); |
480 | 0 | return false; |
481 | 0 | } |
482 | 0 | if (m_param->analysisMultiPassRefine) |
483 | 0 | { |
484 | 0 | p = strstr(opts, "ref="); |
485 | 0 | sscanf(p, "ref=%d", &i); |
486 | 0 | if (i > m_param->maxNumReferences) |
487 | 0 | { |
488 | 0 | x265_log(m_param, X265_LOG_ERROR, "maxNumReferences cannot be less than 1st pass (%d vs %d)\n", |
489 | 0 | i, m_param->maxNumReferences); |
490 | 0 | return false; |
491 | 0 | } |
492 | 0 | } |
493 | 0 | if (m_param->analysisMultiPassRefine || m_param->analysisMultiPassDistortion) |
494 | 0 | { |
495 | 0 | p = strstr(opts, "ctu="); |
496 | 0 | sscanf(p, "ctu=%u", &k); |
497 | 0 | if (k != m_param->maxCUSize) |
498 | 0 | { |
499 | 0 | x265_log(m_param, X265_LOG_ERROR, "maxCUSize mismatch with 1st pass (%u vs %u)\n", |
500 | 0 | k, m_param->maxCUSize); |
501 | 0 | return false; |
502 | 0 | } |
503 | 0 | } |
504 | 0 | CMP_OPT_FIRST_PASS("bitdepth", m_param->internalBitDepth); |
505 | 0 | CMP_OPT_FIRST_PASS("weightp", m_param->bEnableWeightedPred); |
506 | 0 | CMP_OPT_FIRST_PASS("bframes", m_param->bframes); |
507 | 0 | CMP_OPT_FIRST_PASS("b-pyramid", m_param->bBPyramid); |
508 | 0 | CMP_OPT_FIRST_PASS("open-gop", m_param->bOpenGOP); |
509 | 0 | CMP_OPT_FIRST_PASS(" keyint", m_param->keyframeMax); |
510 | 0 | CMP_OPT_FIRST_PASS("scenecut", m_param->scenecutThreshold); |
511 | 0 | CMP_OPT_FIRST_PASS("intra-refresh", m_param->bIntraRefresh); |
512 | 0 | CMP_OPT_FIRST_PASS("frame-dup", m_param->bEnableFrameDuplication); |
513 | 0 | if (m_param->bMultiPassOptRPS) |
514 | 0 | { |
515 | 0 | CMP_OPT_FIRST_PASS("multi-pass-opt-rps", m_param->bMultiPassOptRPS); |
516 | 0 | CMP_OPT_FIRST_PASS("repeat-headers", m_param->bRepeatHeaders); |
517 | 0 | CMP_OPT_FIRST_PASS("min-keyint", m_param->keyframeMin); |
518 | 0 | } |
519 | | |
520 | 0 | if ((p = strstr(opts, "b-adapt=")) != 0 && sscanf(p, "b-adapt=%d", &i) && i >= X265_B_ADAPT_NONE && i <= X265_B_ADAPT_TRELLIS) |
521 | 0 | { |
522 | 0 | m_param->bFrameAdaptive = i; |
523 | 0 | } |
524 | 0 | else if (m_param->bframes) |
525 | 0 | { |
526 | 0 | x265_log(m_param, X265_LOG_ERROR, "b-adapt method specified in stats file not valid\n"); |
527 | 0 | return false; |
528 | 0 | } |
529 | | |
530 | 0 | if ((p = strstr(opts, "rc-lookahead=")) != 0 && sscanf(p, "rc-lookahead=%d", &i)) |
531 | 0 | m_param->lookaheadDepth = i; |
532 | 0 | } |
533 | | /* find number of pics */ |
534 | 0 | p = statsIn; |
535 | 0 | int numEntries; |
536 | 0 | for (numEntries = -1; p; numEntries++) |
537 | 0 | p = strchr(p + 1, ';'); |
538 | 0 | if (!numEntries) |
539 | 0 | { |
540 | 0 | x265_log(m_param, X265_LOG_ERROR, "empty stats file\n"); |
541 | 0 | return false; |
542 | 0 | } |
543 | 0 | m_numEntries = numEntries; |
544 | |
|
545 | 0 | if (m_param->totalFrames < m_numEntries && m_param->totalFrames > 0) |
546 | 0 | { |
547 | 0 | x265_log(m_param, X265_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n", |
548 | 0 | m_param->totalFrames, m_numEntries); |
549 | 0 | } |
550 | 0 | if (m_param->totalFrames > m_numEntries && !m_param->bEnableFrameDuplication) |
551 | 0 | { |
552 | 0 | x265_log(m_param, X265_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n", |
553 | 0 | m_param->totalFrames, m_numEntries); |
554 | 0 | return false; |
555 | 0 | } |
556 | | |
557 | 0 | m_rce2Pass = X265_MALLOC(RateControlEntry, m_numEntries); |
558 | 0 | if (!m_rce2Pass) |
559 | 0 | { |
560 | 0 | x265_log(m_param, X265_LOG_ERROR, "Rce Entries for 2 pass cannot be allocated\n"); |
561 | 0 | return false; |
562 | 0 | } |
563 | 0 | m_encOrder = X265_MALLOC(int, m_numEntries); |
564 | 0 | if (!m_encOrder) |
565 | 0 | { |
566 | 0 | x265_log(m_param, X265_LOG_ERROR, "Encode order for 2 pass cannot be allocated\n"); |
567 | 0 | return false; |
568 | 0 | } |
569 | | /* init all to skipped p frames */ |
570 | 0 | for (int i = 0; i < m_numEntries; i++) |
571 | 0 | { |
572 | 0 | RateControlEntry *rce = &m_rce2Pass[i]; |
573 | 0 | rce->sliceType = P_SLICE; |
574 | 0 | rce->qScale = rce->newQScale = x265_qp2qScale(20); |
575 | 0 | rce->miscBits = m_ncu + 10; |
576 | 0 | rce->newQp = 0; |
577 | 0 | } |
578 | | /* read stats */ |
579 | 0 | p = statsIn; |
580 | 0 | double totalQpAq = 0; |
581 | 0 | for (int i = 0; i < m_numEntries; i++) |
582 | 0 | { |
583 | 0 | RateControlEntry *rce; |
584 | 0 | int frameNumber; |
585 | 0 | int encodeOrder; |
586 | 0 | char picType; |
587 | 0 | int e; |
588 | 0 | char *next; |
589 | 0 | double qpRc, qpAq, qNoVbv, qRceq; |
590 | 0 | next = strstr(p, ";"); |
591 | 0 | if (next) |
592 | 0 | *next++ = 0; |
593 | 0 | e = sscanf(p, " in:%d out:%d", &frameNumber, &encodeOrder); |
594 | 0 | if (frameNumber < 0 || frameNumber >= m_numEntries) |
595 | 0 | { |
596 | 0 | x265_log(m_param, X265_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frameNumber, i); |
597 | 0 | return false; |
598 | 0 | } |
599 | 0 | rce = &m_rce2Pass[encodeOrder]; |
600 | 0 | m_encOrder[frameNumber] = encodeOrder; |
601 | 0 | if (!m_param->bMultiPassOptRPS) |
602 | 0 | { |
603 | 0 | e += sscanf(p, " in:%*d out:%*d type:%c q:%lf q-aq:%lf q-noVbv:%lf q-Rceq:%lf tex:%d mv:%d misc:%d icu:%lf pcu:%lf scu:%lf", |
604 | 0 | &picType, &qpRc, &qpAq, &qNoVbv, &qRceq, &rce->coeffBits, |
605 | 0 | &rce->mvBits, &rce->miscBits, &rce->iCuCount, &rce->pCuCount, |
606 | 0 | &rce->skipCuCount); |
607 | 0 | } |
608 | 0 | else |
609 | 0 | { |
610 | 0 | char deltaPOC[128]; |
611 | 0 | char bUsed[40]; |
612 | 0 | memset(deltaPOC, 0, sizeof(deltaPOC)); |
613 | 0 | memset(bUsed, 0, sizeof(bUsed)); |
614 | 0 | e += sscanf(p, " in:%*d out:%*d type:%c q:%lf q-aq:%lf q-noVbv:%lf q-Rceq:%lf tex:%d mv:%d misc:%d icu:%lf pcu:%lf scu:%lf nump:%d numnegp:%d numposp:%d deltapoc:%s bused:%s", |
615 | 0 | &picType, &qpRc, &qpAq, &qNoVbv, &qRceq, &rce->coeffBits, |
616 | 0 | &rce->mvBits, &rce->miscBits, &rce->iCuCount, &rce->pCuCount, |
617 | 0 | &rce->skipCuCount, &rce->rpsData.numberOfPictures, &rce->rpsData.numberOfNegativePictures, &rce->rpsData.numberOfPositivePictures, deltaPOC, bUsed); |
618 | 0 | splitdeltaPOC(deltaPOC, rce); |
619 | 0 | splitbUsed(bUsed, rce); |
620 | 0 | rce->rpsIdx = -1; |
621 | 0 | } |
622 | 0 | rce->keptAsRef = true; |
623 | 0 | rce->isIdr = false; |
624 | 0 | if (picType == 'b' || picType == 'p') |
625 | 0 | rce->keptAsRef = false; |
626 | 0 | if (picType == 'I') |
627 | 0 | rce->isIdr = true; |
628 | 0 | if (picType == 'I' || picType == 'i') |
629 | 0 | rce->sliceType = I_SLICE; |
630 | 0 | else if (picType == 'P' || picType == 'p') |
631 | 0 | rce->sliceType = P_SLICE; |
632 | 0 | else if (picType == 'B' || picType == 'b') |
633 | 0 | rce->sliceType = B_SLICE; |
634 | 0 | else |
635 | 0 | e = -1; |
636 | 0 | if (e < 10) |
637 | 0 | { |
638 | 0 | x265_log(m_param, X265_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e); |
639 | 0 | return false; |
640 | 0 | } |
641 | 0 | rce->qScale = rce->newQScale = x265_qp2qScale(qpRc); |
642 | 0 | totalQpAq += qpAq; |
643 | 0 | rce->qpNoVbv = qNoVbv; |
644 | 0 | rce->qpaRc = qpRc; |
645 | 0 | rce->qpAq = qpAq; |
646 | 0 | rce->qRceq = qRceq; |
647 | 0 | p = next; |
648 | 0 | } |
649 | 0 | X265_FREE(statsBuf); |
650 | 0 | if (m_param->rc.rateControlMode != X265_RC_CQP) |
651 | 0 | { |
652 | 0 | m_start = 0; |
653 | 0 | m_isQpModified = true; |
654 | 0 | if (!initPass2()) |
655 | 0 | return false; |
656 | 0 | } /* else we're using constant quant, so no need to run the bitrate allocation */ |
657 | 0 | } |
658 | | /* Open output file */ |
659 | | /* If input and output files are the same, output to a temp file |
660 | | * and move it to the real name only when it's complete */ |
661 | 0 | if (m_param->rc.bStatWrite) |
662 | 0 | { |
663 | 0 | char *p, *statFileTmpname; |
664 | 0 | statFileTmpname = strcatFilename(fileName, ".temp"); |
665 | 0 | if (!statFileTmpname) |
666 | 0 | return false; |
667 | 0 | m_statFileOut = x265_fopen(statFileTmpname, "wb"); |
668 | 0 | X265_FREE(statFileTmpname); |
669 | 0 | if (!m_statFileOut) |
670 | 0 | { |
671 | 0 | x265_log_file(m_param, X265_LOG_ERROR, "can't open stats file %s.temp\n", fileName); |
672 | 0 | return false; |
673 | 0 | } |
674 | 0 | p = x265_param2string(m_param, sps.conformanceWindow.rightOffset, sps.conformanceWindow.bottomOffset); |
675 | 0 | if (p) |
676 | 0 | fprintf(m_statFileOut, "#options: %s\n", p); |
677 | 0 | X265_FREE(p); |
678 | 0 | if (m_param->rc.cuTree && !m_param->rc.bStatRead) |
679 | 0 | { |
680 | 0 | statFileTmpname = strcatFilename(fileName, ".cutree.temp"); |
681 | 0 | if (!statFileTmpname) |
682 | 0 | return false; |
683 | 0 | m_cutreeStatFileOut = x265_fopen(statFileTmpname, "wb"); |
684 | 0 | X265_FREE(statFileTmpname); |
685 | 0 | if (!m_cutreeStatFileOut) |
686 | 0 | { |
687 | 0 | x265_log_file(m_param, X265_LOG_ERROR, "can't open mbtree stats file %s.cutree.temp\n", fileName); |
688 | 0 | return false; |
689 | 0 | } |
690 | 0 | } |
691 | 0 | } |
692 | 0 | if (m_param->rc.cuTree) |
693 | 0 | { |
694 | 0 | if (m_param->rc.qgSize == 8) |
695 | 0 | { |
696 | 0 | m_cuTreeStats.qpBuffer[0] = X265_MALLOC(uint16_t, m_ncu * 4 * sizeof(uint16_t)); |
697 | 0 | if (m_param->bBPyramid && m_param->rc.bStatRead) |
698 | 0 | m_cuTreeStats.qpBuffer[1] = X265_MALLOC(uint16_t, m_ncu * 4 * sizeof(uint16_t)); |
699 | 0 | } |
700 | 0 | else |
701 | 0 | { |
702 | 0 | m_cuTreeStats.qpBuffer[0] = X265_MALLOC(uint16_t, m_ncu * sizeof(uint16_t)); |
703 | 0 | if (m_param->bBPyramid && m_param->rc.bStatRead) |
704 | 0 | m_cuTreeStats.qpBuffer[1] = X265_MALLOC(uint16_t, m_ncu * sizeof(uint16_t)); |
705 | 0 | } |
706 | 0 | m_cuTreeStats.qpBufPos = -1; |
707 | 0 | } |
708 | 0 | } |
709 | 0 | return true; |
710 | 0 | } |
711 | | |
712 | | void RateControl::reconfigureRC() |
713 | 0 | { |
714 | 0 | if (m_isVbv) |
715 | 0 | { |
716 | 0 | m_param->rc.vbvBufferSize = x265_clip3(0, 2000000, m_param->rc.vbvBufferSize); |
717 | 0 | m_param->rc.vbvMaxBitrate = x265_clip3(0, 2000000, m_param->rc.vbvMaxBitrate); |
718 | 0 | if (m_param->reconfigWindowSize) |
719 | 0 | m_param->rc.vbvMaxBitrate = (int)(m_param->rc.vbvMaxBitrate * (double)(m_fps / m_param->reconfigWindowSize)); |
720 | 0 | if (m_param->rc.vbvMaxBitrate < m_param->rc.bitrate && |
721 | 0 | m_param->rc.rateControlMode == X265_RC_ABR) |
722 | 0 | { |
723 | 0 | x265_log(m_param, X265_LOG_WARNING, "max bitrate less than average bitrate, assuming CBR\n"); |
724 | 0 | m_param->rc.bitrate = m_param->rc.vbvMaxBitrate; |
725 | 0 | } |
726 | |
|
727 | 0 | if (m_param->rc.vbvBufferSize < (int)(m_param->rc.vbvMaxBitrate / m_fps)) |
728 | 0 | { |
729 | 0 | m_param->rc.vbvBufferSize = (int)(m_param->rc.vbvMaxBitrate / m_fps); |
730 | 0 | x265_log(m_param, X265_LOG_WARNING, "VBV buffer size cannot be smaller than one frame, using %d kbit\n", |
731 | 0 | m_param->rc.vbvBufferSize); |
732 | 0 | } |
733 | 0 | int vbvBufferSize = m_param->rc.vbvBufferSize * 1000; |
734 | 0 | int vbvMaxBitrate = m_param->rc.vbvMaxBitrate * 1000; |
735 | 0 | m_bufferRate = vbvMaxBitrate / m_fps; |
736 | 0 | m_vbvMaxRate = vbvMaxBitrate; |
737 | 0 | m_bufferSize = vbvBufferSize; |
738 | 0 | m_singleFrameVbv = m_bufferRate * 1.1 > m_bufferSize; |
739 | 0 | } |
740 | 0 | if (m_param->rc.rateControlMode == X265_RC_CRF) |
741 | 0 | { |
742 | 0 | #define CRF_INIT_QP (int)m_param->rc.rfConstant |
743 | 0 | m_param->rc.bitrate = 0; |
744 | 0 | double baseCplx = m_ncu * (m_param->bframes ? 120 : 80); |
745 | 0 | double mbtree_offset = m_param->rc.cuTree ? (1.0 - m_param->rc.qCompress) * 13.5 : 0; |
746 | 0 | m_rateFactorConstant = pow(baseCplx, 1 - m_qCompress) / |
747 | 0 | x265_qp2qScale(m_param->rc.rfConstant + mbtree_offset); |
748 | 0 | if (m_param->rc.rfConstantMax) |
749 | 0 | { |
750 | 0 | m_rateFactorMaxIncrement = m_param->rc.rfConstantMax - m_param->rc.rfConstant; |
751 | 0 | if (m_rateFactorMaxIncrement <= 0) |
752 | 0 | { |
753 | 0 | x265_log(m_param, X265_LOG_WARNING, "CRF max must be greater than CRF\n"); |
754 | 0 | m_rateFactorMaxIncrement = 0; |
755 | 0 | } |
756 | 0 | } |
757 | 0 | if (m_param->rc.rfConstantMin) |
758 | 0 | m_rateFactorMaxDecrement = m_param->rc.rfConstant - m_param->rc.rfConstantMin; |
759 | 0 | } |
760 | 0 | if (m_param->rc.rateControlMode == X265_RC_CQP) |
761 | 0 | { |
762 | 0 | m_qp = m_param->rc.qp; |
763 | 0 | if (m_qp && !m_param->bLossless) |
764 | 0 | { |
765 | 0 | m_qpConstant[P_SLICE] = m_qp; |
766 | 0 | m_qpConstant[I_SLICE] = x265_clip3(QP_MIN, QP_MAX_MAX, (int)(m_qp - m_ipOffset + 0.5)); |
767 | 0 | m_qpConstant[B_SLICE] = x265_clip3(QP_MIN, QP_MAX_MAX, (int)(m_qp + m_pbOffset + 0.5)); |
768 | 0 | } |
769 | 0 | else |
770 | 0 | { |
771 | 0 | m_qpConstant[P_SLICE] = m_qpConstant[I_SLICE] = m_qpConstant[B_SLICE] = m_qp; |
772 | 0 | } |
773 | 0 | } |
774 | 0 | m_bitrate = (double)m_param->rc.bitrate * 1000; |
775 | 0 | } |
776 | | |
777 | | void RateControl::initHRD(SPS& sps) |
778 | 0 | { |
779 | 0 | int vbvBufferSize = m_param->rc.vbvBufferSize * 1000; |
780 | 0 | int vbvMaxBitrate = m_param->rc.vbvMaxBitrate * 1000; |
781 | | |
782 | | // Init HRD |
783 | 0 | HRDInfo* hrd = &sps.vuiParameters.hrdParameters; |
784 | 0 | hrd->cbrFlag = m_isCbr; |
785 | 0 | if (m_param->reconfigWindowSize) { |
786 | 0 | hrd->cbrFlag = 0; |
787 | 0 | vbvMaxBitrate = m_param->decoderVbvMaxRate * 1000; |
788 | 0 | } |
789 | | |
790 | | // normalize HRD size and rate to the value / scale notation |
791 | 0 | hrd->bitRateScale = x265_clip3(0, 15, calcScale(vbvMaxBitrate) - BR_SHIFT); |
792 | 0 | hrd->bitRateValue = (vbvMaxBitrate >> (hrd->bitRateScale + BR_SHIFT)); |
793 | |
|
794 | 0 | hrd->cpbSizeScale = x265_clip3(0, 15, calcScale(vbvBufferSize) - CPB_SHIFT); |
795 | 0 | hrd->cpbSizeValue = (vbvBufferSize >> (hrd->cpbSizeScale + CPB_SHIFT)); |
796 | 0 | int bitRateUnscale = hrd->bitRateValue << (hrd->bitRateScale + BR_SHIFT); |
797 | 0 | int cpbSizeUnscale = hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT); |
798 | | |
799 | | // arbitrary |
800 | 0 | #define MAX_DURATION 0.5 |
801 | |
|
802 | 0 | TimingInfo *time = &sps.vuiParameters.timingInfo; |
803 | 0 | int maxCpbOutputDelay = (int)(X265_MIN(m_param->keyframeMax * MAX_DURATION * time->timeScale / time->numUnitsInTick, INT_MAX)); |
804 | 0 | int maxDpbOutputDelay = (int)(sps.maxDecPicBuffering * MAX_DURATION * time->timeScale / time->numUnitsInTick); |
805 | 0 | int maxDelay = (int)(90000.0 * cpbSizeUnscale / bitRateUnscale + 0.5); |
806 | |
|
807 | 0 | hrd->initialCpbRemovalDelayLength = 2 + x265_clip3(4, 22, 32 - calcLength(maxDelay)); |
808 | 0 | hrd->cpbRemovalDelayLength = x265_clip3(4, 31, 32 - calcLength(maxCpbOutputDelay)); |
809 | 0 | hrd->dpbOutputDelayLength = x265_clip3(4, 31, 32 - calcLength(maxDpbOutputDelay)); |
810 | |
|
811 | 0 | #undef MAX_DURATION |
812 | 0 | } |
813 | | |
814 | | bool RateControl::analyseABR2Pass(uint64_t allAvailableBits) |
815 | 0 | { |
816 | 0 | double rateFactor, stepMult; |
817 | 0 | double qBlur = m_param->rc.qblur; |
818 | 0 | double cplxBlur = m_param->rc.complexityBlur; |
819 | 0 | const int filterSize = (int)(qBlur * 4) | 1; |
820 | 0 | double expectedBits; |
821 | 0 | double *qScale, *blurredQscale; |
822 | 0 | double baseCplx = m_ncu * (m_param->bframes ? 120 : 80); |
823 | 0 | double clippedDuration = CLIP_DURATION(m_frameDuration) / BASE_FRAME_DURATION; |
824 | | /* Blur complexities, to reduce local fluctuation of QP. |
825 | | * We don't blur the QPs directly, because then one very simple frame |
826 | | * could drag down the QP of a nearby complex frame and give it more |
827 | | * bits than intended. */ |
828 | 0 | for (int i = 0; i < m_numEntries; i++) |
829 | 0 | { |
830 | 0 | double weightSum = 0; |
831 | 0 | double cplxSum = 0; |
832 | 0 | double weight = 1.0; |
833 | 0 | double gaussianWeight; |
834 | | /* weighted average of cplx of future frames */ |
835 | 0 | for (int j = 1; j < cplxBlur * 2 && j < m_numEntries - i; j++) |
836 | 0 | { |
837 | 0 | int index = i+j; |
838 | 0 | RateControlEntry *rcj = &m_rce2Pass[index]; |
839 | 0 | weight *= 1 - pow(rcj->iCuCount / m_ncu, 2); |
840 | 0 | if (weight < 0.0001) |
841 | 0 | break; |
842 | 0 | gaussianWeight = weight * exp(-j * j / 200.0); |
843 | 0 | weightSum += gaussianWeight; |
844 | 0 | cplxSum += gaussianWeight * (qScale2bits(rcj, 1) - rcj->miscBits) / clippedDuration; |
845 | 0 | } |
846 | | /* weighted average of cplx of past frames */ |
847 | 0 | weight = 1.0; |
848 | 0 | for (int j = 0; j <= cplxBlur * 2 && j <= i; j++) |
849 | 0 | { |
850 | 0 | int index = i-j; |
851 | 0 | RateControlEntry *rcj = &m_rce2Pass[index]; |
852 | 0 | gaussianWeight = weight * exp(-j * j / 200.0); |
853 | 0 | weightSum += gaussianWeight; |
854 | 0 | cplxSum += gaussianWeight * (qScale2bits(rcj, 1) - rcj->miscBits) / clippedDuration; |
855 | 0 | weight *= 1 - pow(rcj->iCuCount / m_ncu, 2); |
856 | 0 | if (weight < .0001) |
857 | 0 | break; |
858 | 0 | } |
859 | 0 | m_rce2Pass[i].blurredComplexity= cplxSum / weightSum; |
860 | 0 | } |
861 | 0 | CHECKED_MALLOC(qScale, double, m_numEntries); |
862 | 0 | if (filterSize > 1) |
863 | 0 | { |
864 | 0 | CHECKED_MALLOC(blurredQscale, double, m_numEntries); |
865 | 0 | } |
866 | 0 | else |
867 | 0 | blurredQscale = qScale; |
868 | | |
869 | | /* Search for a factor which, when multiplied by the RCEQ values from |
870 | | * each frame, adds up to the desired total size. |
871 | | * There is no exact closed-form solution because of VBV constraints and |
872 | | * because qscale2bits is not invertible, but we can start with the simple |
873 | | * approximation of scaling the 1st pass by the ratio of bitrates. |
874 | | * The search range is probably overkill, but speed doesn't matter here. */ |
875 | 0 | expectedBits = 1; |
876 | 0 | for (int i = 0; i < m_numEntries; i++) |
877 | 0 | { |
878 | 0 | RateControlEntry* rce = &m_rce2Pass[i]; |
879 | 0 | double q = getQScale(rce, 1.0); |
880 | 0 | expectedBits += qScale2bits(rce, q); |
881 | 0 | m_lastQScaleFor[rce->sliceType] = q; |
882 | 0 | } |
883 | 0 | stepMult = allAvailableBits / expectedBits; |
884 | |
|
885 | 0 | rateFactor = 0; |
886 | 0 | for (double step = 1E4 * stepMult; step > 1E-7 * stepMult; step *= 0.5) |
887 | 0 | { |
888 | 0 | expectedBits = 0; |
889 | 0 | rateFactor += step; |
890 | |
|
891 | 0 | m_lastNonBPictType = -1; |
892 | 0 | m_lastAccumPNorm = 1; |
893 | 0 | m_accumPNorm = 0; |
894 | |
|
895 | 0 | m_lastQScaleFor[0] = m_lastQScaleFor[1] = |
896 | 0 | m_lastQScaleFor[2] = pow(baseCplx, 1 - m_qCompress) / rateFactor; |
897 | | |
898 | | /* find qscale */ |
899 | 0 | for (int i = 0; i < m_numEntries; i++) |
900 | 0 | { |
901 | 0 | RateControlEntry *rce = &m_rce2Pass[i]; |
902 | 0 | qScale[i] = getQScale(rce, rateFactor); |
903 | 0 | m_lastQScaleFor[rce->sliceType] = qScale[i]; |
904 | 0 | } |
905 | | |
906 | | /* fixed I/B qscale relative to P */ |
907 | 0 | for (int i = 0; i < m_numEntries; i++) |
908 | 0 | { |
909 | 0 | qScale[i] = getDiffLimitedQScale(&m_rce2Pass[i], qScale[i]); |
910 | 0 | X265_CHECK(qScale[i] >= 0, "qScale became negative\n"); |
911 | 0 | } |
912 | | |
913 | | /* smooth curve */ |
914 | 0 | if (filterSize > 1) |
915 | 0 | { |
916 | 0 | X265_CHECK(filterSize % 2 == 1, "filterSize not an odd number\n"); |
917 | 0 | for (int i = 0; i < m_numEntries; i++) |
918 | 0 | { |
919 | 0 | double q = 0.0, sum = 0.0; |
920 | 0 | for (int j = 0; j < filterSize; j++) |
921 | 0 | { |
922 | 0 | int idx = i + j - filterSize / 2; |
923 | 0 | double d = idx - i; |
924 | 0 | double coeff = qBlur == 0 ? 1.0 : exp(-d * d / (qBlur * qBlur)); |
925 | 0 | if (idx < 0 || idx >= m_numEntries) |
926 | 0 | continue; |
927 | 0 | if (m_rce2Pass[i].sliceType != m_rce2Pass[idx].sliceType) |
928 | 0 | continue; |
929 | 0 | q += qScale[idx] * coeff; |
930 | 0 | sum += coeff; |
931 | 0 | } |
932 | 0 | blurredQscale[i] = q / sum; |
933 | 0 | } |
934 | 0 | } |
935 | | |
936 | | /* find expected bits */ |
937 | 0 | for (int i = 0; i < m_numEntries; i++) |
938 | 0 | { |
939 | 0 | RateControlEntry *rce = &m_rce2Pass[i]; |
940 | 0 | rce->newQScale = clipQscale(NULL, rce, blurredQscale[i]); // check if needed |
941 | 0 | X265_CHECK(rce->newQScale >= 0, "new Qscale is negative\n"); |
942 | 0 | expectedBits += qScale2bits(rce, rce->newQScale); |
943 | 0 | } |
944 | |
|
945 | 0 | if (expectedBits > allAvailableBits) |
946 | 0 | rateFactor -= step; |
947 | 0 | } |
948 | |
|
949 | 0 | X265_FREE(qScale); |
950 | 0 | if (filterSize > 1) |
951 | 0 | X265_FREE(blurredQscale); |
952 | 0 | if (m_isVbv) |
953 | 0 | if (!vbv2Pass(allAvailableBits, m_numEntries - 1, 0)) |
954 | 0 | return false; |
955 | 0 | expectedBits = countExpectedBits(0, m_numEntries - 1); |
956 | 0 | if (fabs(expectedBits / allAvailableBits - 1.0) > 0.01) |
957 | 0 | { |
958 | 0 | double avgq = 0; |
959 | 0 | for (int i = 0; i < m_numEntries; i++) |
960 | 0 | avgq += m_rce2Pass[i].newQScale; |
961 | 0 | avgq = x265_qScale2qp(avgq / m_numEntries); |
962 | |
|
963 | 0 | if (expectedBits > allAvailableBits || !m_isVbv) |
964 | 0 | x265_log(m_param, X265_LOG_WARNING, "Error: 2pass curve failed to converge\n"); |
965 | 0 | x265_log(m_param, X265_LOG_WARNING, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n", |
966 | 0 | (double)m_param->rc.bitrate, |
967 | 0 | expectedBits * m_fps / (m_numEntries * 1000.), |
968 | 0 | avgq); |
969 | 0 | if (expectedBits < allAvailableBits && avgq < m_param->rc.qpMin + 2) |
970 | 0 | { |
971 | 0 | if (m_param->rc.qpMin > 0) |
972 | 0 | x265_log(m_param, X265_LOG_WARNING, "try reducing target bitrate or reducing qp_min (currently %d)\n", m_param->rc.qpMin); |
973 | 0 | else |
974 | 0 | x265_log(m_param, X265_LOG_WARNING, "try reducing target bitrate\n"); |
975 | 0 | } |
976 | 0 | else if (expectedBits > allAvailableBits && avgq > m_param->rc.qpMax - 2) |
977 | 0 | { |
978 | 0 | if (m_param->rc.qpMax < QP_MAX_MAX) |
979 | 0 | x265_log(m_param, X265_LOG_WARNING, "try increasing target bitrate or increasing qp_max (currently %d)\n", m_param->rc.qpMax); |
980 | 0 | else |
981 | 0 | x265_log(m_param, X265_LOG_WARNING, "try increasing target bitrate\n"); |
982 | 0 | } |
983 | 0 | else if (!(m_2pass && m_isVbv)) |
984 | 0 | x265_log(m_param, X265_LOG_WARNING, "internal error\n"); |
985 | 0 | } |
986 | |
|
987 | 0 | return true; |
988 | | |
989 | 0 | fail: |
990 | 0 | x265_log(m_param, X265_LOG_WARNING, "two-pass ABR initialization failed\n"); |
991 | 0 | return false; |
992 | 0 | } |
993 | | |
994 | | bool RateControl::initPass2() |
995 | 0 | { |
996 | 0 | uint64_t allConstBits = 0, allCodedBits = 0; |
997 | 0 | uint64_t allAvailableBits = uint64_t(m_param->rc.bitrate * 1000. * m_numEntries * m_frameDuration); |
998 | 0 | int startIndex, framesCount, endIndex; |
999 | 0 | int fps = X265_MIN(m_param->keyframeMax, (int)(m_fps + 0.5)); |
1000 | 0 | startIndex = endIndex = framesCount = 0; |
1001 | 0 | int diffQp = 0; |
1002 | 0 | double targetBits = 0; |
1003 | 0 | double expectedBits = 0; |
1004 | 0 | for (startIndex = m_start, endIndex = m_start; endIndex < m_numEntries; endIndex++) |
1005 | 0 | { |
1006 | 0 | allConstBits += m_rce2Pass[endIndex].miscBits; |
1007 | 0 | allCodedBits += m_rce2Pass[endIndex].coeffBits + m_rce2Pass[endIndex].mvBits; |
1008 | 0 | if (m_param->rc.rateControlMode == X265_RC_CRF) |
1009 | 0 | { |
1010 | 0 | framesCount = endIndex - startIndex + 1; |
1011 | 0 | diffQp += int (m_rce2Pass[endIndex].qpaRc - m_rce2Pass[endIndex].qpNoVbv); |
1012 | 0 | if (framesCount > fps) |
1013 | 0 | diffQp -= int (m_rce2Pass[endIndex - fps].qpaRc - m_rce2Pass[endIndex - fps].qpNoVbv); |
1014 | 0 | if (framesCount >= fps) |
1015 | 0 | { |
1016 | 0 | if (diffQp >= 1) |
1017 | 0 | { |
1018 | 0 | if (!m_isQpModified && endIndex > fps) |
1019 | 0 | { |
1020 | 0 | double factor = 2; |
1021 | 0 | double step = 0; |
1022 | 0 | if (endIndex + fps >= m_numEntries) |
1023 | 0 | { |
1024 | 0 | m_start = endIndex - (endIndex % fps); |
1025 | 0 | return true; |
1026 | 0 | } |
1027 | 0 | for (int start = endIndex + 1; start <= endIndex + fps && start < m_numEntries; start++) |
1028 | 0 | { |
1029 | 0 | RateControlEntry *rce = &m_rce2Pass[start]; |
1030 | 0 | targetBits += qScale2bits(rce, x265_qp2qScale(rce->qpNoVbv)); |
1031 | 0 | expectedBits += qScale2bits(rce, rce->qScale); |
1032 | 0 | } |
1033 | 0 | if (expectedBits < 0.95 * targetBits) |
1034 | 0 | { |
1035 | 0 | m_isQpModified = true; |
1036 | 0 | m_isGopReEncoded = true; |
1037 | 0 | while (endIndex + fps < m_numEntries) |
1038 | 0 | { |
1039 | 0 | step = pow(2, factor / 6.0); |
1040 | 0 | expectedBits = 0; |
1041 | 0 | for (int start = endIndex + 1; start <= endIndex + fps; start++) |
1042 | 0 | { |
1043 | 0 | RateControlEntry *rce = &m_rce2Pass[start]; |
1044 | 0 | rce->newQScale = rce->qScale / step; |
1045 | 0 | X265_CHECK(rce->newQScale >= 0, "new Qscale is negative\n"); |
1046 | 0 | expectedBits += qScale2bits(rce, rce->newQScale); |
1047 | 0 | rce->newQp = x265_qScale2qp(rce->newQScale); |
1048 | 0 | } |
1049 | 0 | if (expectedBits >= targetBits && step > 1) |
1050 | 0 | factor *= 0.90; |
1051 | 0 | else |
1052 | 0 | break; |
1053 | 0 | } |
1054 | |
|
1055 | 0 | if (m_isVbv && endIndex + fps < m_numEntries) |
1056 | 0 | if (!vbv2Pass((uint64_t)targetBits, endIndex + fps, endIndex + 1)) |
1057 | 0 | return false; |
1058 | | |
1059 | 0 | targetBits = 0; |
1060 | 0 | expectedBits = 0; |
1061 | |
|
1062 | 0 | for (int start = endIndex - fps + 1; start <= endIndex; start++) |
1063 | 0 | { |
1064 | 0 | RateControlEntry *rce = &m_rce2Pass[start]; |
1065 | 0 | targetBits += qScale2bits(rce, x265_qp2qScale(rce->qpNoVbv)); |
1066 | 0 | } |
1067 | 0 | while (1) |
1068 | 0 | { |
1069 | 0 | step = pow(2, factor / 6.0); |
1070 | 0 | expectedBits = 0; |
1071 | 0 | for (int start = endIndex - fps + 1; start <= endIndex; start++) |
1072 | 0 | { |
1073 | 0 | RateControlEntry *rce = &m_rce2Pass[start]; |
1074 | 0 | rce->newQScale = rce->qScale * step; |
1075 | 0 | X265_CHECK(rce->newQScale >= 0, "new Qscale is negative\n"); |
1076 | 0 | expectedBits += qScale2bits(rce, rce->newQScale); |
1077 | 0 | rce->newQp = x265_qScale2qp(rce->newQScale); |
1078 | 0 | } |
1079 | 0 | if (expectedBits > targetBits && step > 1) |
1080 | 0 | factor *= 1.1; |
1081 | 0 | else |
1082 | 0 | break; |
1083 | 0 | } |
1084 | 0 | if (m_isVbv) |
1085 | 0 | if (!vbv2Pass((uint64_t)targetBits, endIndex, endIndex - fps + 1)) |
1086 | 0 | return false; |
1087 | 0 | diffQp = 0; |
1088 | 0 | m_reencode = endIndex - fps + 1; |
1089 | 0 | endIndex = endIndex + fps; |
1090 | 0 | startIndex = endIndex + 1; |
1091 | 0 | m_start = startIndex; |
1092 | 0 | targetBits = expectedBits = 0; |
1093 | 0 | } |
1094 | 0 | else |
1095 | 0 | targetBits = expectedBits = 0; |
1096 | 0 | } |
1097 | 0 | } |
1098 | 0 | else |
1099 | 0 | m_isQpModified = false; |
1100 | 0 | } |
1101 | 0 | } |
1102 | 0 | } |
1103 | | |
1104 | 0 | if (m_param->rc.rateControlMode == X265_RC_ABR) |
1105 | 0 | { |
1106 | 0 | if (allAvailableBits < allConstBits) |
1107 | 0 | { |
1108 | 0 | x265_log(m_param, X265_LOG_ERROR, "requested bitrate is too low. estimated minimum is %d kbps\n", |
1109 | 0 | (int)(allConstBits * m_fps / framesCount * 1000.)); |
1110 | 0 | return false; |
1111 | 0 | } |
1112 | 0 | if (!analyseABR2Pass(allAvailableBits)) |
1113 | 0 | return false; |
1114 | 0 | } |
1115 | | |
1116 | 0 | m_start = X265_MAX(m_start, endIndex - fps); |
1117 | |
|
1118 | 0 | return true; |
1119 | 0 | } |
1120 | | |
1121 | | bool RateControl::vbv2Pass(uint64_t allAvailableBits, int endPos, int startPos) |
1122 | 0 | { |
1123 | | /* for each interval of bufferFull .. underflow, uniformly increase the qp of all |
1124 | | * frames in the interval until either buffer is full at some intermediate frame or the |
1125 | | * last frame in the interval no longer underflows. Recompute intervals and repeat. |
1126 | | * Then do the converse to put bits back into overflow areas until target size is met */ |
1127 | |
|
1128 | 0 | double *fills; |
1129 | 0 | double expectedBits = 0; |
1130 | 0 | double adjustment; |
1131 | 0 | double prevBits = 0; |
1132 | 0 | int t0, t1; |
1133 | 0 | double qScaleMin = x265_qp2qScale(m_param->rc.qpMin); |
1134 | 0 | double qScaleMax = x265_qp2qScale(m_param->rc.qpMax); |
1135 | 0 | int iterations = 0 , adjMin, adjMax; |
1136 | 0 | CHECKED_MALLOC(fills, double, m_numEntries + 1); |
1137 | 0 | fills++; |
1138 | | |
1139 | | /* adjust overall stream size */ |
1140 | 0 | do |
1141 | 0 | { |
1142 | 0 | iterations++; |
1143 | 0 | prevBits = expectedBits; |
1144 | |
|
1145 | 0 | if (expectedBits) |
1146 | 0 | { /* not first iteration */ |
1147 | 0 | adjustment = X265_MAX(X265_MIN(expectedBits / allAvailableBits, 0.999), 0.9); |
1148 | 0 | fills[-1] = m_bufferSize * m_param->rc.vbvBufferInit; |
1149 | 0 | t0 = startPos; |
1150 | | /* fix overflows */ |
1151 | 0 | adjMin = 1; |
1152 | 0 | while (adjMin && findUnderflow(fills, &t0, &t1, 1, endPos)) |
1153 | 0 | { |
1154 | 0 | adjMin = fixUnderflow(t0, t1, adjustment, qScaleMin, qScaleMax); |
1155 | 0 | t0 = t1; |
1156 | 0 | } |
1157 | 0 | } |
1158 | |
|
1159 | 0 | fills[-1] = m_bufferSize * (1. - m_param->rc.vbvBufferInit); |
1160 | 0 | t0 = 0; |
1161 | | /* fix underflows -- should be done after overflow, as we'd better undersize target than underflowing VBV */ |
1162 | 0 | adjMax = 1; |
1163 | 0 | while (adjMax && findUnderflow(fills, &t0, &t1, 0, endPos)) |
1164 | 0 | adjMax = fixUnderflow(t0, t1, 1.001, qScaleMin, qScaleMax); |
1165 | 0 | expectedBits = countExpectedBits(startPos, endPos); |
1166 | 0 | } |
1167 | 0 | while ((expectedBits < .995 * allAvailableBits) && ((int64_t)(expectedBits+.5) > (int64_t)(prevBits+.5)) && !(m_param->rc.rateControlMode == X265_RC_CRF)); |
1168 | 0 | if (!adjMax) |
1169 | 0 | x265_log(m_param, X265_LOG_WARNING, "vbv-maxrate issue, qpmax or vbv-maxrate too low\n"); |
1170 | | /* store expected vbv filling values for tracking when encoding */ |
1171 | 0 | for (int i = startPos; i <= endPos; i++) |
1172 | 0 | m_rce2Pass[i].expectedVbv = m_bufferSize - fills[i]; |
1173 | 0 | X265_FREE(fills - 1); |
1174 | 0 | return true; |
1175 | | |
1176 | 0 | fail: |
1177 | 0 | x265_log(m_param, X265_LOG_ERROR, "malloc failure in two-pass VBV init\n"); |
1178 | 0 | return false; |
1179 | 0 | } |
1180 | | |
1181 | | /* In 2pass, force the same frame types as in the 1st pass */ |
1182 | | int RateControl::rateControlSliceType(int frameNum) |
1183 | 0 | { |
1184 | 0 | if (m_param->rc.bStatRead) |
1185 | 0 | { |
1186 | 0 | if (frameNum >= m_numEntries) |
1187 | 0 | { |
1188 | | /* We could try to initialize everything required for ABR and |
1189 | | * adaptive B-frames, but that would be complicated. |
1190 | | * So just calculate the average QP used so far. */ |
1191 | 0 | m_param->rc.qp = (m_accumPQp < 1) ? ABR_INIT_QP_MAX : (int)(m_accumPQp + 0.5); |
1192 | 0 | m_qpConstant[P_SLICE] = x265_clip3(QP_MIN, QP_MAX_MAX, m_param->rc.qp); |
1193 | 0 | m_qpConstant[I_SLICE] = x265_clip3(QP_MIN, QP_MAX_MAX, (int)(m_param->rc.qp - m_ipOffset + 0.5)); |
1194 | 0 | m_qpConstant[B_SLICE] = x265_clip3(QP_MIN, QP_MAX_MAX, (int)(m_param->rc.qp + m_pbOffset + 0.5)); |
1195 | |
|
1196 | 0 | x265_log(m_param, X265_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", m_numEntries); |
1197 | 0 | x265_log(m_param, X265_LOG_ERROR, "continuing anyway, at constant QP=%d\n", m_param->rc.qp); |
1198 | 0 | if (m_param->bFrameAdaptive) |
1199 | 0 | x265_log(m_param, X265_LOG_ERROR, "disabling adaptive B-frames\n"); |
1200 | |
|
1201 | 0 | m_isAbr = 0; |
1202 | 0 | m_2pass = 0; |
1203 | 0 | m_param->rc.rateControlMode = X265_RC_CQP; |
1204 | 0 | m_param->rc.bStatRead = 0; |
1205 | 0 | m_param->bFrameAdaptive = 0; |
1206 | 0 | m_param->scenecutThreshold = 0; |
1207 | 0 | m_param->bHistBasedSceneCut = 0; |
1208 | 0 | m_param->rc.cuTree = 0; |
1209 | 0 | if (m_param->bframes > 1) |
1210 | 0 | m_param->bframes = 1; |
1211 | 0 | return X265_TYPE_AUTO; |
1212 | 0 | } |
1213 | 0 | int index = m_encOrder[frameNum]; |
1214 | 0 | int frameType = m_rce2Pass[index].sliceType == I_SLICE ? (m_rce2Pass[index].isIdr ? X265_TYPE_IDR : X265_TYPE_I) |
1215 | 0 | : m_rce2Pass[index].sliceType == P_SLICE ? X265_TYPE_P |
1216 | 0 | : (m_rce2Pass[index].sliceType == B_SLICE && m_rce2Pass[index].keptAsRef ? X265_TYPE_BREF : X265_TYPE_B); |
1217 | 0 | return frameType; |
1218 | 0 | } |
1219 | 0 | else |
1220 | 0 | return X265_TYPE_AUTO; |
1221 | 0 | } |
1222 | | |
1223 | | void RateControl::initFramePredictors() |
1224 | 0 | { |
1225 | | /* Frame Predictors used in vbv */ |
1226 | 0 | for (int i = 0; i < 4; i++) |
1227 | 0 | { |
1228 | 0 | m_pred[i].coeffMin = 1.0 / 4; |
1229 | 0 | m_pred[i].coeff = 1.0; |
1230 | 0 | m_pred[i].count = 1.0; |
1231 | 0 | m_pred[i].decay = 0.5; |
1232 | 0 | m_pred[i].offset = 0.0; |
1233 | 0 | } |
1234 | 0 | m_pred[0].coeff = m_pred[3].coeff = 0.75; |
1235 | 0 | m_pred[0].coeffMin = m_pred[3].coeffMin = 0.75 / 4; |
1236 | 0 | if (m_isGrainEnabled) // when tuned for grain |
1237 | 0 | { |
1238 | 0 | m_pred[1].coeffMin = 0.75 / 4; |
1239 | 0 | m_pred[1].coeff = 0.75; |
1240 | 0 | m_pred[0].coeff = m_pred[3].coeff = 0.75; |
1241 | 0 | m_pred[0].coeffMin = m_pred[3].coeffMin = 0.75 / 4; |
1242 | 0 | } |
1243 | 0 | } |
1244 | | |
1245 | | int RateControl::rateControlStart(Frame* curFrame, RateControlEntry* rce, Encoder* enc) |
1246 | 0 | { |
1247 | 0 | int orderValue = m_startEndOrder.get(); |
1248 | 0 | int startOrdinal = rce->encodeOrder * 2; |
1249 | |
|
1250 | 0 | while (orderValue < startOrdinal && !m_bTerminated) |
1251 | 0 | orderValue = m_startEndOrder.waitForChange(orderValue); |
1252 | |
|
1253 | 0 | if (!curFrame) |
1254 | 0 | { |
1255 | | // faked rateControlStart calls when the encoder is flushing |
1256 | 0 | m_startEndOrder.incr(); |
1257 | 0 | return 0; |
1258 | 0 | } |
1259 | | |
1260 | 0 | FrameData& curEncData = *curFrame->m_encData; |
1261 | 0 | m_curSlice = curEncData.m_slice; |
1262 | 0 | m_sliceType = m_curSlice->m_sliceType; |
1263 | 0 | rce->sliceType = m_sliceType; |
1264 | 0 | if (!m_2pass) |
1265 | 0 | rce->keptAsRef = IS_REFERENCED(curFrame); |
1266 | 0 | m_predType = getPredictorType(curFrame->m_lowres.sliceType, m_sliceType); |
1267 | 0 | rce->poc = m_curSlice->m_poc; |
1268 | |
|
1269 | 0 | if (!m_param->bResetZoneConfig && (rce->encodeOrder % m_param->reconfigWindowSize == 0)) |
1270 | 0 | { |
1271 | 0 | int index = m_zoneBufferIdx % m_param->rc.zonefileCount; |
1272 | 0 | int read = m_top->zoneReadCount[index].get(); |
1273 | 0 | int write = m_top->zoneWriteCount[index].get(); |
1274 | 0 | if (write <= read) |
1275 | 0 | write = m_top->zoneWriteCount[index].waitForChange(write); |
1276 | 0 | m_zoneBufferIdx++; |
1277 | |
|
1278 | 0 | for (int i = 0; i < m_param->rc.zonefileCount; i++) |
1279 | 0 | { |
1280 | 0 | if (m_param->rc.zones[i].startFrame == rce->encodeOrder) |
1281 | 0 | { |
1282 | 0 | m_param->rc.bitrate = m_param->rc.zones[i].zoneParam->rc.bitrate; |
1283 | 0 | m_param->rc.vbvMaxBitrate = m_param->rc.zones[i].zoneParam->rc.vbvMaxBitrate; |
1284 | 0 | memcpy(m_relativeComplexity, m_param->rc.zones[i].relativeComplexity, sizeof(double) * m_param->reconfigWindowSize); |
1285 | 0 | reconfigureRC(); |
1286 | 0 | m_isCbr = 1; /* Always vbvmaxrate == bitrate here*/ |
1287 | 0 | m_top->zoneReadCount[i].incr(); |
1288 | 0 | } |
1289 | 0 | } |
1290 | 0 | } |
1291 | | |
1292 | | |
1293 | 0 | if (m_param->bResetZoneConfig) |
1294 | 0 | { |
1295 | | /* change ratecontrol stats for next zone if specified */ |
1296 | 0 | for (int i = 0; i < m_param->rc.zonefileCount; i++) |
1297 | 0 | { |
1298 | 0 | if (m_param->rc.zones[i].startFrame == curFrame->m_encodeOrder) |
1299 | 0 | { |
1300 | 0 | m_param = m_param->rc.zones[i].zoneParam; |
1301 | 0 | reconfigureRC(); |
1302 | 0 | init(*m_curSlice->m_sps); |
1303 | 0 | } |
1304 | 0 | } |
1305 | 0 | } |
1306 | |
|
1307 | 0 | if (m_param->rc.bStatRead) |
1308 | 0 | { |
1309 | 0 | X265_CHECK(rce->poc >= 0 && rce->poc < m_numEntries, "bad encode ordinal\n"); |
1310 | 0 | int index = m_encOrder[rce->poc]; |
1311 | 0 | copyRceData(rce, &m_rce2Pass[index]); |
1312 | 0 | } |
1313 | 0 | rce->isActive = true; |
1314 | 0 | rce->scenecut = false; |
1315 | 0 | rce->isFadeEnd = curFrame->m_lowres.bIsFadeEnd; |
1316 | 0 | bool isRefFrameScenecut = m_sliceType!= I_SLICE && m_curSlice->m_refFrameList[0][0]->m_lowres.bScenecut; |
1317 | 0 | m_isFirstMiniGop = m_sliceType == I_SLICE ? true : m_isFirstMiniGop; |
1318 | 0 | if (curFrame->m_lowres.bScenecut) |
1319 | 0 | { |
1320 | 0 | m_isSceneTransition = true; |
1321 | 0 | rce->scenecut = true; |
1322 | 0 | m_lastPredictorReset = rce->encodeOrder; |
1323 | |
|
1324 | 0 | initFramePredictors(); |
1325 | 0 | } |
1326 | 0 | else if (m_sliceType != B_SLICE && !isRefFrameScenecut) |
1327 | 0 | m_isSceneTransition = false; |
1328 | |
|
1329 | 0 | if (rce->encodeOrder < m_lastPredictorReset + m_param->frameNumThreads) |
1330 | 0 | { |
1331 | 0 | rce->rowPreds[0][0].count = 0; |
1332 | 0 | } |
1333 | |
|
1334 | 0 | rce->bLastMiniGopBFrame = curFrame->m_lowres.bLastMiniGopBFrame; |
1335 | 0 | rce->bufferRate = m_bufferRate; |
1336 | 0 | rce->rowCplxrSum = 0.0; |
1337 | 0 | rce->rowTotalBits = 0; |
1338 | 0 | if (m_isVbv) |
1339 | 0 | { |
1340 | 0 | if (rce->rowPreds[0][0].count == 0) |
1341 | 0 | { |
1342 | 0 | for (int i = 0; i < 3; i++) |
1343 | 0 | { |
1344 | 0 | for (int j = 0; j < 2; j++) |
1345 | 0 | { |
1346 | 0 | rce->rowPreds[i][j].coeffMin = 0.25 / 4; |
1347 | 0 | rce->rowPreds[i][j].coeff = 0.25; |
1348 | 0 | rce->rowPreds[i][j].count = 1.0; |
1349 | 0 | rce->rowPreds[i][j].decay = 0.5; |
1350 | 0 | rce->rowPreds[i][j].offset = 0.0; |
1351 | 0 | } |
1352 | 0 | } |
1353 | 0 | } |
1354 | 0 | rce->rowPred[0] = &rce->rowPreds[m_sliceType][0]; |
1355 | 0 | rce->rowPred[1] = &rce->rowPreds[m_sliceType][1]; |
1356 | 0 | m_predictedBits = m_totalBits; |
1357 | 0 | updateVbvPlan(enc); |
1358 | 0 | rce->bufferFill = m_bufferFill; |
1359 | 0 | rce->vbvEndAdj = false; |
1360 | 0 | if (m_param->vbvBufferEnd && rce->encodeOrder >= m_param->vbvEndFrameAdjust * m_param->totalFrames) |
1361 | 0 | { |
1362 | 0 | rce->vbvEndAdj = true; |
1363 | 0 | rce->targetFill = 0; |
1364 | 0 | } |
1365 | |
|
1366 | 0 | int mincr = enc->m_vps.ptl.minCrForLevel; |
1367 | | /* Profiles above Main10 don't require maxAU size check, so just set the maximum to a large value. */ |
1368 | 0 | if (enc->m_vps.ptl.profileIdc > Profile::MAIN10 || enc->m_vps.ptl.levelIdc == Level::NONE) |
1369 | 0 | rce->frameSizeMaximum = 1e9; |
1370 | 0 | else |
1371 | 0 | { |
1372 | | /* The spec has a special case for the first frame. */ |
1373 | 0 | if (curFrame->m_lowres.bKeyframe) |
1374 | 0 | { |
1375 | | /* 1.5 * (Max( PicSizeInSamplesY, fR * MaxLumaSr) + MaxLumaSr * (AuCpbRemovalTime[ 0 ] -AuNominalRemovalTime[ 0 ])) ? MinCr */ |
1376 | 0 | double fr = 1. / 300; |
1377 | 0 | int picSizeInSamplesY = m_param->sourceWidth * m_param->sourceHeight; |
1378 | 0 | rce->frameSizeMaximum = 8 * 1.5 * X265_MAX(picSizeInSamplesY, fr * enc->m_vps.ptl.maxLumaSrForLevel) / mincr; |
1379 | 0 | } |
1380 | 0 | else |
1381 | 0 | { |
1382 | | /* 1.5 * MaxLumaSr * (AuCpbRemovalTime[ n ] - AuCpbRemovalTime[ n - 1 ]) / MinCr */ |
1383 | 0 | rce->frameSizeMaximum = 8 * 1.5 * enc->m_vps.ptl.maxLumaSrForLevel * m_frameDuration / mincr; |
1384 | 0 | } |
1385 | 0 | rce->frameSizeMaximum *= m_param->maxAUSizeFactor; |
1386 | 0 | } |
1387 | 0 | } |
1388 | 0 | if (!m_isAbr && m_2pass && m_param->rc.rateControlMode == X265_RC_CRF) |
1389 | 0 | { |
1390 | 0 | rce->qpPrev = x265_qScale2qp(rce->qScale); |
1391 | 0 | rce->qScale = rce->newQScale; |
1392 | 0 | rce->qpaRc = curEncData.m_avgQpRc = curEncData.m_avgQpAq = x265_qScale2qp(rce->newQScale); |
1393 | 0 | m_qp = int(rce->qpaRc + 0.5); |
1394 | 0 | rce->frameSizePlanned = qScale2bits(rce, rce->qScale); |
1395 | 0 | m_framesDone++; |
1396 | 0 | return m_qp; |
1397 | 0 | } |
1398 | | |
1399 | 0 | if (m_isAbr || m_2pass) // ABR,CRF |
1400 | 0 | { |
1401 | 0 | if (m_isAbr || m_isVbv) |
1402 | 0 | { |
1403 | 0 | m_currentSatd = curFrame->m_lowres.satdCost >> (X265_DEPTH - 8); |
1404 | | /* Update rce for use in rate control VBV later */ |
1405 | 0 | rce->lastSatd = m_currentSatd; |
1406 | 0 | X265_CHECK(rce->lastSatd, "satdcost cannot be zero\n"); |
1407 | | /* Detect a pattern for B frames with same SATDcost to identify a series of static frames |
1408 | | * and the P frame at the end of the series marks a possible case for ABR reset logic */ |
1409 | 0 | if (m_param->bframes) |
1410 | 0 | { |
1411 | 0 | if (m_sliceType != B_SLICE && m_numBframesInPattern > m_param->bframes) |
1412 | 0 | { |
1413 | 0 | m_isPatternPresent = true; |
1414 | 0 | } |
1415 | 0 | else if (m_sliceType == B_SLICE && !IS_REFERENCED(curFrame)) |
1416 | 0 | { |
1417 | 0 | if (m_currentSatd != m_lastBsliceSatdCost && !rce->bLastMiniGopBFrame) |
1418 | 0 | { |
1419 | 0 | m_isPatternPresent = false; |
1420 | 0 | m_lastBsliceSatdCost = m_currentSatd; |
1421 | 0 | m_numBframesInPattern = 0; |
1422 | 0 | } |
1423 | 0 | else if (m_currentSatd == m_lastBsliceSatdCost) |
1424 | 0 | m_numBframesInPattern++; |
1425 | 0 | } |
1426 | 0 | } |
1427 | 0 | if (rce->isFadeEnd) |
1428 | 0 | m_isPatternPresent = true; |
1429 | 0 | } |
1430 | | /* For a scenecut that occurs within the mini-gop, enable scene transition |
1431 | | * switch until the next mini-gop to ensure a min qp for all the frames within |
1432 | | * the scene-transition mini-gop */ |
1433 | |
|
1434 | 0 | double q = x265_qScale2qp(rateEstimateQscale(curFrame, rce)); |
1435 | 0 | q = x265_clip3((double)m_param->rc.qpMin, (double)m_param->rc.qpMax, q); |
1436 | 0 | m_qp = int(q + 0.5); |
1437 | 0 | q = m_isGrainEnabled ? m_qp : q; |
1438 | 0 | rce->qpaRc = curEncData.m_avgQpRc = curEncData.m_avgQpAq = q; |
1439 | | /* copy value of lastRceq into thread local rce struct *to be used in RateControlEnd() */ |
1440 | 0 | rce->qRceq = m_lastRceq; |
1441 | 0 | accumPQpUpdate(); |
1442 | 0 | curFrame->m_rcData->cumulativePQp = m_accumPQp; |
1443 | 0 | curFrame->m_rcData->cumulativePNorm = m_accumPNorm; |
1444 | 0 | for (int i = 0; i < 3; i++) |
1445 | 0 | curFrame->m_rcData->lastQScaleFor[i] = m_lastQScaleFor[i]; |
1446 | 0 | curFrame->m_rcData->shortTermCplxSum = m_shortTermCplxSum; |
1447 | 0 | curFrame->m_rcData->shortTermCplxCount = m_shortTermCplxCount; |
1448 | 0 | } |
1449 | 0 | else // CQP |
1450 | 0 | { |
1451 | 0 | if (m_sliceType == B_SLICE && IS_REFERENCED(curFrame)) |
1452 | 0 | m_qp = (m_qpConstant[B_SLICE] + m_qpConstant[P_SLICE]) / 2; |
1453 | 0 | else |
1454 | 0 | m_qp = m_qpConstant[m_sliceType]; |
1455 | 0 | curEncData.m_avgQpAq = curEncData.m_avgQpRc = m_qp; |
1456 | | |
1457 | 0 | x265_zone* zone = getZone(); |
1458 | 0 | if (zone) |
1459 | 0 | { |
1460 | 0 | if (zone->bForceQp) |
1461 | 0 | m_qp += zone->qp - m_qpConstant[P_SLICE]; |
1462 | 0 | else |
1463 | 0 | m_qp -= (int)(6.0 * X265_LOG2(zone->bitrateFactor)); |
1464 | 0 | } |
1465 | 0 | } |
1466 | 0 | if (m_sliceType != B_SLICE) |
1467 | 0 | { |
1468 | 0 | m_lastNonBPictType = m_sliceType; |
1469 | 0 | m_leadingNoBSatd = m_currentSatd; |
1470 | 0 | } |
1471 | 0 | rce->leadingNoBSatd = m_leadingNoBSatd; |
1472 | 0 | if (curFrame->m_forceqp) |
1473 | 0 | { |
1474 | 0 | m_qp = (int32_t)(curFrame->m_forceqp + 0.5) - 1; |
1475 | 0 | m_qp = x265_clip3(m_param->rc.qpMin, m_param->rc.qpMax, m_qp); |
1476 | 0 | rce->qpaRc = curEncData.m_avgQpRc = curEncData.m_avgQpAq = m_qp; |
1477 | 0 | if (m_isAbr || m_2pass) |
1478 | 0 | { |
1479 | 0 | rce->qpNoVbv = rce->qpaRc; |
1480 | 0 | m_lastQScaleFor[m_sliceType] = x265_qp2qScale(rce->qpaRc); |
1481 | 0 | if (rce->poc == 0) |
1482 | 0 | m_lastQScaleFor[P_SLICE] = m_lastQScaleFor[m_sliceType] * fabs(m_param->rc.ipFactor); |
1483 | 0 | rce->frameSizePlanned = predictSize(&m_pred[m_predType], m_qp, (double)m_currentSatd); |
1484 | 0 | } |
1485 | 0 | } |
1486 | 0 | m_framesDone++; |
1487 | |
|
1488 | 0 | return m_qp; |
1489 | 0 | } |
1490 | | |
1491 | | void RateControl::accumPQpUpdate() |
1492 | 0 | { |
1493 | 0 | m_accumPQp *= .95; |
1494 | 0 | m_accumPNorm *= .95; |
1495 | 0 | m_accumPNorm += 1; |
1496 | 0 | if (m_sliceType == I_SLICE) |
1497 | 0 | m_accumPQp += m_qp + m_ipOffset; |
1498 | 0 | else |
1499 | 0 | m_accumPQp += m_qp; |
1500 | 0 | } |
1501 | | |
1502 | | int RateControl::getPredictorType(int lowresSliceType, int sliceType) |
1503 | 0 | { |
1504 | | /* Use a different predictor for B Ref and B frames for vbv frame size predictions */ |
1505 | 0 | if (lowresSliceType == X265_TYPE_BREF) |
1506 | 0 | return 3; |
1507 | 0 | return sliceType; |
1508 | 0 | } |
1509 | | |
1510 | | double RateControl::getDiffLimitedQScale(RateControlEntry *rce, double q) |
1511 | 0 | { |
1512 | | // force I/B quants as a function of P quants |
1513 | 0 | const double lastPqScale = m_lastQScaleFor[P_SLICE]; |
1514 | 0 | const double lastNonBqScale = m_lastQScaleFor[m_lastNonBPictType]; |
1515 | 0 | if (rce->sliceType == I_SLICE) |
1516 | 0 | { |
1517 | 0 | double iq = q; |
1518 | 0 | double pq = x265_qp2qScale(m_accumPQp / m_accumPNorm); |
1519 | 0 | double ipFactor = fabs(m_param->rc.ipFactor); |
1520 | | /* don't apply ipFactor if the following frame is also I */ |
1521 | 0 | if (m_accumPNorm <= 0) |
1522 | 0 | q = iq; |
1523 | 0 | else if (m_param->rc.ipFactor < 0) |
1524 | 0 | q = iq / ipFactor; |
1525 | 0 | else if (m_accumPNorm >= 1) |
1526 | 0 | q = pq / ipFactor; |
1527 | 0 | else |
1528 | 0 | q = m_accumPNorm * pq / ipFactor + (1 - m_accumPNorm) * iq; |
1529 | 0 | } |
1530 | 0 | else if (rce->sliceType == B_SLICE) |
1531 | 0 | { |
1532 | 0 | if (m_param->rc.pbFactor > 0) |
1533 | 0 | q = lastNonBqScale; |
1534 | 0 | if (!rce->keptAsRef) |
1535 | 0 | q *= fabs(m_param->rc.pbFactor); |
1536 | 0 | } |
1537 | 0 | else if (rce->sliceType == P_SLICE |
1538 | 0 | && m_lastNonBPictType == P_SLICE |
1539 | 0 | && rce->coeffBits == 0) |
1540 | 0 | { |
1541 | 0 | q = lastPqScale; |
1542 | 0 | } |
1543 | | |
1544 | | /* last qscale / qdiff stuff */ |
1545 | 0 | if (m_lastNonBPictType == rce->sliceType && |
1546 | 0 | (rce->sliceType != I_SLICE || m_lastAccumPNorm < 1)) |
1547 | 0 | { |
1548 | 0 | double maxQscale = m_lastQScaleFor[rce->sliceType] * m_lstep; |
1549 | 0 | double minQscale = m_lastQScaleFor[rce->sliceType] / m_lstep; |
1550 | 0 | q = x265_clip3(minQscale, maxQscale, q); |
1551 | 0 | } |
1552 | |
|
1553 | 0 | m_lastQScaleFor[rce->sliceType] = q; |
1554 | 0 | if (rce->sliceType != B_SLICE) |
1555 | 0 | m_lastNonBPictType = rce->sliceType; |
1556 | 0 | if (rce->sliceType == I_SLICE) |
1557 | 0 | { |
1558 | 0 | m_lastAccumPNorm = m_accumPNorm; |
1559 | 0 | m_accumPNorm = 0; |
1560 | 0 | m_accumPQp = 0; |
1561 | 0 | } |
1562 | 0 | if (rce->sliceType == P_SLICE) |
1563 | 0 | { |
1564 | 0 | double mask = 1 - pow(rce->iCuCount / m_ncu, 2); |
1565 | 0 | m_accumPQp = mask * (x265_qScale2qp(q) + m_accumPQp); |
1566 | 0 | m_accumPNorm = mask * (1 + m_accumPNorm); |
1567 | 0 | } |
1568 | |
|
1569 | 0 | return q; |
1570 | 0 | } |
1571 | | |
1572 | | double RateControl::countExpectedBits(int startPos, int endPos) |
1573 | 0 | { |
1574 | 0 | double expectedBits = 0; |
1575 | 0 | for (int i = startPos; i <= endPos; i++) |
1576 | 0 | { |
1577 | 0 | RateControlEntry *rce = &m_rce2Pass[i]; |
1578 | 0 | rce->expectedBits = (uint64_t)expectedBits; |
1579 | 0 | expectedBits += qScale2bits(rce, rce->newQScale); |
1580 | 0 | } |
1581 | 0 | return expectedBits; |
1582 | 0 | } |
1583 | | |
1584 | | bool RateControl::findUnderflow(double *fills, int *t0, int *t1, int over, int endPos) |
1585 | 0 | { |
1586 | | /* find an interval ending on an overflow or underflow (depending on whether |
1587 | | * we're adding or removing bits), and starting on the earliest frame that |
1588 | | * can influence the buffer fill of that end frame. */ |
1589 | 0 | const double bufferMin = .1 * m_bufferSize; |
1590 | 0 | const double bufferMax = .9 * m_bufferSize; |
1591 | 0 | double fill = fills[*t0 - 1]; |
1592 | 0 | double parity = over ? 1. : -1.; |
1593 | 0 | int start = -1, end = -1; |
1594 | 0 | for (int i = *t0; i <= endPos; i++) |
1595 | 0 | { |
1596 | 0 | fill += (m_frameDuration * m_vbvMaxRate - |
1597 | 0 | qScale2bits(&m_rce2Pass[i], m_rce2Pass[i].newQScale)) * parity; |
1598 | 0 | fill = x265_clip3(0.0, m_bufferSize, fill); |
1599 | 0 | fills[i] = fill; |
1600 | 0 | if (fill <= bufferMin || i == 0) |
1601 | 0 | { |
1602 | 0 | if (end >= 0) |
1603 | 0 | break; |
1604 | 0 | start = i; |
1605 | 0 | } |
1606 | 0 | else if (fill >= bufferMax && start >= 0) |
1607 | 0 | end = i; |
1608 | 0 | } |
1609 | 0 | *t0 = start; |
1610 | 0 | *t1 = end; |
1611 | 0 | return start >= 0 && end >= 0; |
1612 | 0 | } |
1613 | | |
1614 | | bool RateControl::fixUnderflow(int t0, int t1, double adjustment, double qscaleMin, double qscaleMax) |
1615 | 0 | { |
1616 | 0 | double qscaleOrig, qscaleNew; |
1617 | 0 | bool adjusted = false; |
1618 | 0 | if (t0 > 0) |
1619 | 0 | t0++; |
1620 | 0 | for (int i = t0; i <= t1; i++) |
1621 | 0 | { |
1622 | 0 | qscaleOrig = m_rce2Pass[i].newQScale; |
1623 | 0 | qscaleOrig = x265_clip3(qscaleMin, qscaleMax, qscaleOrig); |
1624 | 0 | qscaleNew = qscaleOrig * adjustment; |
1625 | 0 | qscaleNew = x265_clip3(qscaleMin, qscaleMax, qscaleNew); |
1626 | 0 | m_rce2Pass[i].newQScale = qscaleNew; |
1627 | 0 | adjusted = adjusted || (qscaleNew != qscaleOrig); |
1628 | 0 | } |
1629 | 0 | return adjusted; |
1630 | 0 | } |
1631 | | |
1632 | | bool RateControl::cuTreeReadFor2Pass(Frame* frame) |
1633 | 0 | { |
1634 | 0 | int index = m_encOrder[frame->m_poc]; |
1635 | 0 | uint8_t sliceTypeActual = (uint8_t)m_rce2Pass[index].sliceType; |
1636 | 0 | int ncu; |
1637 | 0 | if (m_param->rc.qgSize == 8) |
1638 | 0 | ncu = m_ncu * 4; |
1639 | 0 | else |
1640 | 0 | ncu = m_ncu; |
1641 | 0 | if (m_rce2Pass[index].keptAsRef) |
1642 | 0 | { |
1643 | | /* TODO: We don't need pre-lookahead to measure AQ offsets, but there is currently |
1644 | | * no way to signal this */ |
1645 | 0 | uint8_t type; |
1646 | 0 | if (m_cuTreeStats.qpBufPos < 0) |
1647 | 0 | { |
1648 | 0 | do |
1649 | 0 | { |
1650 | 0 | m_cuTreeStats.qpBufPos++; |
1651 | |
|
1652 | 0 | if (!fread(&type, 1, 1, m_cutreeStatFileIn)) |
1653 | 0 | goto fail; |
1654 | 0 | if (fread(m_cuTreeStats.qpBuffer[m_cuTreeStats.qpBufPos], sizeof(uint16_t), ncu, m_cutreeStatFileIn) != (size_t)ncu) |
1655 | 0 | goto fail; |
1656 | | |
1657 | 0 | if (type != sliceTypeActual && m_cuTreeStats.qpBufPos == 1) |
1658 | 0 | { |
1659 | 0 | x265_log(m_param, X265_LOG_ERROR, "CU-tree frametype %d doesn't match actual frametype %d.\n", type, sliceTypeActual); |
1660 | 0 | return false; |
1661 | 0 | } |
1662 | 0 | } |
1663 | 0 | while(type != sliceTypeActual); |
1664 | 0 | } |
1665 | 0 | primitives.fix8Unpack(frame->m_lowres.qpCuTreeOffset, m_cuTreeStats.qpBuffer[m_cuTreeStats.qpBufPos], ncu); |
1666 | 0 | for (int i = 0; i < ncu; i++) |
1667 | 0 | frame->m_lowres.invQscaleFactor[i] = x265_exp2fix8(frame->m_lowres.qpCuTreeOffset[i]); |
1668 | 0 | m_cuTreeStats.qpBufPos--; |
1669 | 0 | } |
1670 | 0 | return true; |
1671 | | |
1672 | 0 | fail: |
1673 | 0 | x265_log(m_param, X265_LOG_ERROR, "Incomplete CU-tree stats file.\n"); |
1674 | 0 | return false; |
1675 | 0 | } |
1676 | | |
1677 | | double RateControl::tuneAbrQScaleFromFeedback(double qScale) |
1678 | 0 | { |
1679 | 0 | double abrBuffer = 2 * m_rateTolerance * m_bitrate; |
1680 | | /* use framesDone instead of POC as poc count is not serial with bframes enabled */ |
1681 | 0 | double overflow = 1.0; |
1682 | 0 | double timeDone = (double)(m_framesDone - m_param->frameNumThreads + 1) * m_frameDuration; |
1683 | 0 | double wantedBits = timeDone * m_bitrate; |
1684 | 0 | int64_t encodedBits = m_totalBits; |
1685 | 0 | if (m_param->totalFrames && m_param->totalFrames <= 2 * m_fps) |
1686 | 0 | { |
1687 | 0 | abrBuffer = m_param->totalFrames * (m_bitrate / m_fps); |
1688 | 0 | encodedBits = m_encodedBits; |
1689 | 0 | } |
1690 | |
|
1691 | 0 | if (wantedBits > 0 && encodedBits > 0 && (!m_partialResidualFrames || |
1692 | 0 | m_param->rc.bStrictCbr || m_isGrainEnabled)) |
1693 | 0 | { |
1694 | 0 | abrBuffer *= X265_MAX(1, sqrt(timeDone)); |
1695 | 0 | overflow = x265_clip3(.5, 2.0, 1.0 + (encodedBits - wantedBits) / abrBuffer); |
1696 | 0 | qScale *= overflow; |
1697 | 0 | } |
1698 | 0 | return qScale; |
1699 | 0 | } |
1700 | | |
1701 | | double RateControl::tuneQScaleForZone(RateControlEntry *rce, double qScale) |
1702 | 0 | { |
1703 | 0 | rce->frameSizePlanned = predictSize(&m_pred[m_predType], qScale, (double)m_currentSatd); |
1704 | 0 | int loop = 0; |
1705 | |
|
1706 | 0 | double availableBits = (double)m_param->rc.bitrate * 1000 * m_relativeComplexity[rce->encodeOrder % m_param->reconfigWindowSize]; |
1707 | | |
1708 | | // Tune qScale to adhere to the available frame bits. |
1709 | 0 | for (int i = 0; i < 1000 && loop != 3; i++) |
1710 | 0 | { |
1711 | 0 | if (rce->frameSizePlanned < availableBits) |
1712 | 0 | { |
1713 | 0 | qScale = qScale / 1.01; |
1714 | 0 | loop = loop | 1; |
1715 | 0 | } |
1716 | 0 | else if (rce->frameSizePlanned > availableBits) |
1717 | 0 | { |
1718 | 0 | qScale = qScale * 1.01; |
1719 | 0 | loop = loop | 2; |
1720 | 0 | } |
1721 | 0 | rce->frameSizePlanned = predictSize(&m_pred[m_predType], qScale, (double)m_currentSatd); |
1722 | 0 | } |
1723 | 0 | return qScale; |
1724 | 0 | } |
1725 | | |
1726 | | double RateControl::tuneQScaleForGrain(double rcOverflow) |
1727 | 0 | { |
1728 | 0 | double qpstep = rcOverflow > 1.1 ? rcOverflow : m_lstep; |
1729 | 0 | double qScaleAvg = x265_qp2qScale(m_avgPFrameQp); |
1730 | 0 | double q = m_lastQScaleFor[P_SLICE]; |
1731 | 0 | int curQp = int (x265_qScale2qp(m_lastQScaleFor[P_SLICE]) + 0.5); |
1732 | 0 | double curBitrate = m_qpToEncodedBits[curQp] * int(m_fps + 0.5); |
1733 | 0 | int newQp = rcOverflow > 1.1 ? curQp + 2 : rcOverflow > 1 ? curQp + 1 : curQp - 1 ; |
1734 | 0 | double projectedBitrate = int(m_fps + 0.5) * m_qpToEncodedBits[newQp]; |
1735 | 0 | if (curBitrate > 0 && projectedBitrate > 0) |
1736 | 0 | q = abs(projectedBitrate - m_bitrate) < abs (curBitrate - m_bitrate) ? x265_qp2qScale(newQp) : m_lastQScaleFor[P_SLICE]; |
1737 | 0 | else |
1738 | 0 | q = rcOverflow > 1 ? qScaleAvg * qpstep : rcOverflow < 1 ? qScaleAvg / qpstep : m_lastQScaleFor[P_SLICE]; |
1739 | 0 | return q; |
1740 | 0 | } |
1741 | | |
1742 | | double RateControl::rateEstimateQscale(Frame* curFrame, RateControlEntry *rce) |
1743 | 0 | { |
1744 | 0 | double q; |
1745 | |
|
1746 | 0 | if (m_2pass) |
1747 | 0 | { |
1748 | 0 | if (m_sliceType != rce->sliceType) |
1749 | 0 | { |
1750 | 0 | x265_log(m_param, X265_LOG_ERROR, "slice=%c but 2pass stats say %c\n", |
1751 | 0 | g_sliceTypeToChar[m_sliceType], g_sliceTypeToChar[rce->sliceType]); |
1752 | 0 | } |
1753 | 0 | } |
1754 | 0 | else |
1755 | 0 | { |
1756 | 0 | if (m_isAbr) |
1757 | 0 | { |
1758 | 0 | int pos = m_sliderPos % s_slidingWindowFrames; |
1759 | 0 | int addPos = (pos + s_slidingWindowFrames - 1) % s_slidingWindowFrames; |
1760 | 0 | if (m_sliderPos > s_slidingWindowFrames) |
1761 | 0 | { |
1762 | 0 | const static double base = pow(0.5, s_slidingWindowFrames - 1); |
1763 | 0 | m_movingAvgSum -= m_lastRemovedSatdCost * base; |
1764 | 0 | m_movingAvgSum *= 0.5; |
1765 | 0 | m_movingAvgSum += m_satdCostWindow[addPos]; |
1766 | 0 | } |
1767 | 0 | else if (m_sliderPos == s_slidingWindowFrames) |
1768 | 0 | { |
1769 | 0 | m_movingAvgSum += m_satdCostWindow[addPos]; |
1770 | 0 | } |
1771 | 0 | else if (m_sliderPos > 0) |
1772 | 0 | { |
1773 | 0 | m_movingAvgSum += m_satdCostWindow[addPos]; |
1774 | 0 | m_movingAvgSum *= 0.5; |
1775 | 0 | } |
1776 | |
|
1777 | 0 | rce->movingAvgSum = m_movingAvgSum; |
1778 | 0 | m_lastRemovedSatdCost = m_satdCostWindow[pos]; |
1779 | 0 | m_satdCostWindow[pos] = rce->lastSatd; |
1780 | 0 | m_sliderPos++; |
1781 | 0 | } |
1782 | 0 | } |
1783 | |
|
1784 | 0 | if (m_sliceType == B_SLICE) |
1785 | 0 | { |
1786 | | /* B-frames don't have independent rate control, but rather get the |
1787 | | * average QP of the two adjacent P-frames + an offset */ |
1788 | 0 | Slice* prevRefSlice = m_curSlice->m_refFrameList[0][0]->m_encData->m_slice; |
1789 | 0 | Slice* nextRefSlice = m_curSlice->m_refFrameList[1][0]->m_encData->m_slice; |
1790 | 0 | double q0 = m_curSlice->m_refFrameList[0][0]->m_encData->m_avgQpRc; |
1791 | 0 | double q1 = m_curSlice->m_refFrameList[1][0]->m_encData->m_avgQpRc; |
1792 | 0 | bool i0 = prevRefSlice->m_sliceType == I_SLICE; |
1793 | 0 | bool i1 = nextRefSlice->m_sliceType == I_SLICE; |
1794 | 0 | int dt0 = abs(m_curSlice->m_poc - prevRefSlice->m_poc); |
1795 | 0 | int dt1 = abs(m_curSlice->m_poc - nextRefSlice->m_poc); |
1796 | | |
1797 | | // Skip taking a reference frame before the Scenecut if ABR has been reset. |
1798 | 0 | if (m_lastAbrResetPoc >= 0) |
1799 | 0 | { |
1800 | 0 | if (prevRefSlice->m_sliceType == P_SLICE && prevRefSlice->m_poc < m_lastAbrResetPoc) |
1801 | 0 | { |
1802 | 0 | i0 = i1; |
1803 | 0 | dt0 = dt1; |
1804 | 0 | q0 = q1; |
1805 | 0 | } |
1806 | 0 | } |
1807 | |
|
1808 | 0 | if (prevRefSlice->m_sliceType == B_SLICE && IS_REFERENCED(m_curSlice->m_refFrameList[0][0])) |
1809 | 0 | q0 -= m_pbOffset / 2; |
1810 | 0 | if (nextRefSlice->m_sliceType == B_SLICE && IS_REFERENCED(m_curSlice->m_refFrameList[1][0])) |
1811 | 0 | q1 -= m_pbOffset / 2; |
1812 | 0 | if (i0 && i1) |
1813 | 0 | q = (q0 + q1) / 2 + m_ipOffset; |
1814 | 0 | else if (i0) |
1815 | 0 | q = q1; |
1816 | 0 | else if (i1) |
1817 | 0 | q = q0; |
1818 | 0 | else if(m_isGrainEnabled && !m_2pass) |
1819 | 0 | q = q1; |
1820 | 0 | else |
1821 | 0 | q = (q0 * dt1 + q1 * dt0) / (dt0 + dt1); |
1822 | |
|
1823 | 0 | if (IS_REFERENCED(curFrame)) |
1824 | 0 | q += m_pbOffset / 2; |
1825 | 0 | else |
1826 | 0 | q += m_pbOffset; |
1827 | | |
1828 | | /* Set a min qp at scenechanges and transitions */ |
1829 | 0 | if (m_isSceneTransition) |
1830 | 0 | { |
1831 | 0 | q = X265_MAX(ABR_SCENECUT_INIT_QP_MIN, q); |
1832 | 0 | double minScenecutQscale =x265_qp2qScale(ABR_SCENECUT_INIT_QP_MIN); |
1833 | 0 | m_lastQScaleFor[P_SLICE] = X265_MAX(minScenecutQscale, m_lastQScaleFor[P_SLICE]); |
1834 | 0 | } |
1835 | 0 | double qScale = x265_qp2qScale(q); |
1836 | 0 | rce->qpNoVbv = q; |
1837 | 0 | double lmin = 0, lmax = 0; |
1838 | 0 | if (m_isGrainEnabled && m_isFirstMiniGop) |
1839 | 0 | { |
1840 | 0 | lmin = m_lastQScaleFor[P_SLICE] / m_lstep; |
1841 | 0 | lmax = m_lastQScaleFor[P_SLICE] * m_lstep; |
1842 | 0 | double tunedQscale = tuneAbrQScaleFromFeedback(qScale); |
1843 | 0 | double overflow = tunedQscale / qScale; |
1844 | 0 | if (!m_isAbrReset) |
1845 | 0 | qScale = x265_clip3(lmin, lmax, qScale); |
1846 | 0 | m_avgPFrameQp = m_avgPFrameQp == 0 ? rce->qpNoVbv : m_avgPFrameQp; |
1847 | 0 | if (overflow != 1) |
1848 | 0 | { |
1849 | 0 | qScale = tuneQScaleForGrain(overflow); |
1850 | 0 | q = x265_qScale2qp(qScale); |
1851 | 0 | } |
1852 | 0 | rce->qpNoVbv = q; |
1853 | 0 | } |
1854 | | /* Scenecut Aware QP offsets*/ |
1855 | 0 | if (m_param->bEnableSceneCutAwareQp) |
1856 | 0 | { |
1857 | 0 | double lqmin = m_lmin[m_sliceType]; |
1858 | 0 | double lqmax = m_lmax[m_sliceType]; |
1859 | 0 | qScale = scenecutAwareQp(curFrame, qScale); |
1860 | 0 | qScale = x265_clip3(lqmin, lqmax, qScale); |
1861 | 0 | q = x265_qScale2qp(qScale); |
1862 | 0 | rce->qpNoVbv = q; |
1863 | 0 | } |
1864 | 0 | if (m_isVbv) |
1865 | 0 | { |
1866 | 0 | lmin = m_lastQScaleFor[P_SLICE] / m_lstep; |
1867 | 0 | lmax = m_lastQScaleFor[P_SLICE] * m_lstep; |
1868 | |
|
1869 | 0 | if (m_isCbr && !m_isGrainEnabled) |
1870 | 0 | { |
1871 | 0 | qScale = tuneAbrQScaleFromFeedback(qScale); |
1872 | 0 | if (!m_isAbrReset) |
1873 | 0 | qScale = x265_clip3(lmin, lmax, qScale); |
1874 | 0 | q = x265_qScale2qp(qScale); |
1875 | 0 | } |
1876 | |
|
1877 | 0 | if (!m_param->bResetZoneConfig) |
1878 | 0 | { |
1879 | 0 | double lqmin = m_lmin[m_sliceType]; |
1880 | 0 | double lqmax = m_lmax[m_sliceType]; |
1881 | 0 | qScale = tuneQScaleForZone(rce, qScale); |
1882 | 0 | qScale = x265_clip3(lqmin, lqmax, qScale); |
1883 | 0 | } |
1884 | |
|
1885 | 0 | if (!m_2pass) |
1886 | 0 | { |
1887 | | /* clip qp to permissible range after vbv-lookahead estimation to avoid possible |
1888 | | * mispredictions by initial frame size predictors */ |
1889 | 0 | qScale = clipQscale(curFrame, rce, qScale); |
1890 | |
|
1891 | 0 | if (m_pred[m_predType].count == 1) |
1892 | 0 | qScale = x265_clip3(lmin, lmax, qScale); |
1893 | 0 | m_lastQScaleFor[m_sliceType] = qScale; |
1894 | 0 | } |
1895 | 0 | } |
1896 | |
|
1897 | 0 | if (m_2pass) |
1898 | 0 | rce->frameSizePlanned = qScale2bits(rce, qScale); |
1899 | 0 | else |
1900 | 0 | rce->frameSizePlanned = predictSize(&m_pred[m_predType], qScale, (double)m_currentSatd); |
1901 | | |
1902 | | /* Limit planned size by MinCR */ |
1903 | 0 | if (m_isVbv) |
1904 | 0 | rce->frameSizePlanned = X265_MIN(rce->frameSizePlanned, rce->frameSizeMaximum); |
1905 | 0 | rce->frameSizeEstimated = rce->frameSizePlanned; |
1906 | |
|
1907 | 0 | rce->newQScale = qScale; |
1908 | 0 | if(rce->bLastMiniGopBFrame) |
1909 | 0 | { |
1910 | 0 | if (m_isFirstMiniGop && m_isGrainEnabled) |
1911 | 0 | { |
1912 | 0 | m_avgPFrameQp = (m_avgPFrameQp + rce->qpNoVbv) / 2; |
1913 | 0 | m_lastQScaleFor[P_SLICE] = x265_qp2qScale(m_avgPFrameQp); |
1914 | 0 | } |
1915 | 0 | m_isFirstMiniGop = false; |
1916 | 0 | } |
1917 | 0 | return qScale; |
1918 | 0 | } |
1919 | 0 | else |
1920 | 0 | { |
1921 | 0 | double abrBuffer = 2 * m_rateTolerance * m_bitrate; |
1922 | 0 | if (m_2pass) |
1923 | 0 | { |
1924 | 0 | double lmin = m_lmin[m_sliceType]; |
1925 | 0 | double lmax = m_lmax[m_sliceType]; |
1926 | 0 | int64_t diff; |
1927 | 0 | if (!m_isVbv) |
1928 | 0 | { |
1929 | 0 | m_predictedBits = m_totalBits; |
1930 | 0 | if (rce->encodeOrder < m_param->frameNumThreads) |
1931 | 0 | m_predictedBits += (int64_t)(rce->encodeOrder * m_bitrate / m_fps); |
1932 | 0 | else |
1933 | 0 | m_predictedBits += (int64_t)(m_param->frameNumThreads * m_bitrate / m_fps); |
1934 | 0 | } |
1935 | | /* Adjust ABR buffer based on distance to the end of the video. */ |
1936 | 0 | if (m_numEntries > rce->encodeOrder) |
1937 | 0 | { |
1938 | 0 | uint64_t finalBits = m_rce2Pass[m_numEntries - 1].expectedBits; |
1939 | 0 | double videoPos = (double)rce->expectedBits / finalBits; |
1940 | 0 | double scaleFactor = sqrt((1 - videoPos) * m_numEntries); |
1941 | 0 | abrBuffer *= 0.5 * X265_MAX(scaleFactor, 0.5); |
1942 | 0 | } |
1943 | 0 | diff = m_predictedBits - (int64_t)rce->expectedBits; |
1944 | 0 | q = rce->newQScale; |
1945 | 0 | x265_zone* zone = getZone(); |
1946 | 0 | if (zone) |
1947 | 0 | { |
1948 | 0 | if (zone->bForceQp) |
1949 | 0 | q = x265_qp2qScale(zone->qp); |
1950 | 0 | else |
1951 | 0 | q /= zone->bitrateFactor; |
1952 | 0 | } |
1953 | 0 | q /= x265_clip3(0.5, 2.0, (double)(abrBuffer - diff) / abrBuffer); |
1954 | 0 | if (m_expectedBitsSum > 0) |
1955 | 0 | { |
1956 | | /* Adjust quant based on the difference between |
1957 | | * achieved and expected bitrate so far */ |
1958 | 0 | double curTime = (double)rce->encodeOrder / m_numEntries; |
1959 | 0 | double w = x265_clip3(0.0, 1.0, curTime); |
1960 | 0 | q *= pow((double)m_totalBits / m_expectedBitsSum, w); |
1961 | 0 | } |
1962 | 0 | if (m_framesDone == 0 && m_param->rc.rateControlMode == X265_RC_ABR && m_isGrainEnabled) |
1963 | 0 | q = X265_MIN(x265_qp2qScale(ABR_INIT_QP_GRAIN_MAX), q); |
1964 | 0 | rce->qpNoVbv = x265_qScale2qp(q); |
1965 | 0 | if ((m_sliceType == I_SLICE && m_param->keyframeMax > 1 |
1966 | 0 | && m_lastNonBPictType != I_SLICE && !m_isAbrReset) || (m_isNextGop && !m_framesDone)) |
1967 | 0 | m_avgPFrameQp = 0; |
1968 | 0 | if (m_sliceType == P_SLICE) |
1969 | 0 | { |
1970 | 0 | m_avgPFrameQp = m_avgPFrameQp == 0 ? rce->qpNoVbv : m_avgPFrameQp; |
1971 | 0 | m_avgPFrameQp = (m_avgPFrameQp + rce->qpNoVbv) / 2; |
1972 | 0 | } |
1973 | |
|
1974 | 0 | if (m_isVbv) |
1975 | 0 | { |
1976 | | /* Do not overflow vbv */ |
1977 | 0 | double expectedSize = qScale2bits(rce, q); |
1978 | 0 | double expectedVbv = m_bufferFill + m_bufferRate - expectedSize; |
1979 | 0 | double expectedFullness = rce->expectedVbv / m_bufferSize; |
1980 | 0 | double qmax = q * (2 - expectedFullness); |
1981 | 0 | double sizeConstraint = 1 + expectedFullness; |
1982 | 0 | qmax = X265_MAX(qmax, rce->newQScale); |
1983 | 0 | if (expectedFullness < .05) |
1984 | 0 | qmax = lmax; |
1985 | 0 | qmax = X265_MIN(qmax, lmax); |
1986 | 0 | while (((expectedVbv < rce->expectedVbv/sizeConstraint) && (q < qmax)) || |
1987 | 0 | ((expectedVbv < 0) && (q < lmax))) |
1988 | 0 | { |
1989 | 0 | q *= 1.05; |
1990 | 0 | expectedSize = qScale2bits(rce, q); |
1991 | 0 | expectedVbv = m_bufferFill + m_bufferRate - expectedSize; |
1992 | 0 | } |
1993 | 0 | } |
1994 | 0 | q = x265_clip3(lmin, lmax, q); |
1995 | 0 | } |
1996 | 0 | else |
1997 | 0 | { |
1998 | | /* 1pass ABR */ |
1999 | | |
2000 | | /* Calculate the quantizer which would have produced the desired |
2001 | | * average bitrate if it had been applied to all frames so far. |
2002 | | * Then modulate that quant based on the current frame's complexity |
2003 | | * relative to the average complexity so far (using the 2pass RCEQ). |
2004 | | * Then bias the quant up or down if total size so far was far from |
2005 | | * the target. |
2006 | | * Result: Depending on the value of rate_tolerance, there is a |
2007 | | * trade-off between quality and bitrate precision. But at large |
2008 | | * tolerances, the bit distribution approaches that of 2pass. */ |
2009 | |
|
2010 | 0 | double overflow = 1; |
2011 | 0 | double lqmin = m_lmin[m_sliceType]; |
2012 | 0 | double lqmax = m_lmax[m_sliceType]; |
2013 | 0 | m_shortTermCplxSum *= 0.5; |
2014 | 0 | m_shortTermCplxCount *= 0.5; |
2015 | 0 | m_shortTermCplxSum += m_currentSatd / (CLIP_DURATION(m_frameDuration) / BASE_FRAME_DURATION); |
2016 | 0 | m_shortTermCplxCount++; |
2017 | | /* coeffBits to be used in 2-pass */ |
2018 | 0 | rce->coeffBits = (int)m_currentSatd; |
2019 | 0 | rce->blurredComplexity = m_shortTermCplxSum / m_shortTermCplxCount; |
2020 | 0 | rce->mvBits = 0; |
2021 | 0 | rce->sliceType = m_sliceType; |
2022 | |
|
2023 | 0 | if (m_param->rc.rateControlMode == X265_RC_CRF) |
2024 | 0 | { |
2025 | 0 | q = getQScale(rce, m_rateFactorConstant); |
2026 | 0 | x265_zone* zone = getZone(); |
2027 | 0 | if (zone) |
2028 | 0 | { |
2029 | 0 | if (zone->bForceQp) |
2030 | 0 | q = x265_qp2qScale(zone->qp); |
2031 | 0 | else |
2032 | 0 | q /= zone->bitrateFactor; |
2033 | 0 | } |
2034 | 0 | } |
2035 | 0 | else |
2036 | 0 | { |
2037 | 0 | if (!m_param->rc.bStatRead) |
2038 | 0 | checkAndResetABR(rce, false); |
2039 | 0 | double initialQScale = getQScale(rce, m_wantedBitsWindow / m_cplxrSum); |
2040 | 0 | x265_zone* zone = getZone(); |
2041 | 0 | if (zone) |
2042 | 0 | { |
2043 | 0 | if (zone->bForceQp) |
2044 | 0 | initialQScale = x265_qp2qScale(zone->qp); |
2045 | 0 | else |
2046 | 0 | initialQScale /= zone->bitrateFactor; |
2047 | 0 | } |
2048 | 0 | double tunedQScale = tuneAbrQScaleFromFeedback(initialQScale); |
2049 | 0 | overflow = tunedQScale / initialQScale; |
2050 | 0 | q = !m_partialResidualFrames? tunedQScale : initialQScale; |
2051 | 0 | bool isEncodeEnd = (m_param->totalFrames && |
2052 | 0 | m_framesDone > 0.75 * m_param->totalFrames) ? 1 : 0; |
2053 | 0 | bool isEncodeBeg = m_framesDone < (int)(m_fps + 0.5); |
2054 | 0 | if (m_isGrainEnabled) |
2055 | 0 | { |
2056 | 0 | if(m_sliceType!= I_SLICE && m_framesDone && !isEncodeEnd && |
2057 | 0 | ((overflow < 1.05 && overflow > 0.95) || isEncodeBeg)) |
2058 | 0 | { |
2059 | 0 | q = tuneQScaleForGrain(overflow); |
2060 | 0 | } |
2061 | 0 | } |
2062 | 0 | } |
2063 | 0 | if ((m_sliceType == I_SLICE && m_param->keyframeMax > 1 |
2064 | 0 | && m_lastNonBPictType != I_SLICE && !m_isAbrReset) || (m_isNextGop && !m_framesDone)) |
2065 | 0 | { |
2066 | 0 | if (!m_param->rc.bStrictCbr) |
2067 | 0 | q = x265_qp2qScale(m_accumPQp / m_accumPNorm); |
2068 | 0 | q /= fabs(m_param->rc.ipFactor); |
2069 | 0 | m_avgPFrameQp = 0; |
2070 | 0 | } |
2071 | 0 | else if (m_framesDone > 0) |
2072 | 0 | { |
2073 | 0 | if (m_param->rc.rateControlMode != X265_RC_CRF) |
2074 | 0 | { |
2075 | 0 | lqmin = m_lastQScaleFor[m_sliceType] / m_lstep; |
2076 | 0 | lqmax = m_lastQScaleFor[m_sliceType] * m_lstep; |
2077 | 0 | if (!m_partialResidualFrames || m_isGrainEnabled) |
2078 | 0 | { |
2079 | 0 | if (overflow > 1.1 && m_framesDone > 3) |
2080 | 0 | lqmax *= m_lstep; |
2081 | 0 | else if (overflow < 0.9) |
2082 | 0 | lqmin /= m_lstep; |
2083 | 0 | } |
2084 | 0 | q = x265_clip3(lqmin, lqmax, q); |
2085 | 0 | } |
2086 | 0 | } |
2087 | 0 | else if (m_qCompress != 1 && m_param->rc.rateControlMode == X265_RC_CRF) |
2088 | 0 | { |
2089 | 0 | q = x265_qp2qScale(CRF_INIT_QP) / fabs(m_param->rc.ipFactor); |
2090 | 0 | } |
2091 | 0 | else if (m_framesDone == 0 && !m_isVbv && m_param->rc.rateControlMode == X265_RC_ABR) |
2092 | 0 | { |
2093 | | /* for ABR alone, clip the first I frame qp */ |
2094 | 0 | lqmax = (m_isGrainEnabled && m_lstep) ? x265_qp2qScale(ABR_INIT_QP_GRAIN_MAX) : |
2095 | 0 | x265_qp2qScale(ABR_INIT_QP_MAX); |
2096 | 0 | q = X265_MIN(lqmax, q); |
2097 | 0 | } |
2098 | 0 | q = x265_clip3(lqmin, lqmax, q); |
2099 | | /* Set a min qp at scenechanges and transitions */ |
2100 | 0 | if (m_isSceneTransition) |
2101 | 0 | { |
2102 | 0 | double minScenecutQscale = x265_qp2qScale(ABR_SCENECUT_INIT_QP_MIN); |
2103 | 0 | q = X265_MAX(minScenecutQscale, q); |
2104 | 0 | m_lastQScaleFor[P_SLICE] = X265_MAX(minScenecutQscale, m_lastQScaleFor[P_SLICE]); |
2105 | 0 | } |
2106 | 0 | rce->qpNoVbv = x265_qScale2qp(q); |
2107 | 0 | if (m_sliceType == P_SLICE) |
2108 | 0 | { |
2109 | 0 | m_avgPFrameQp = m_avgPFrameQp == 0 ? rce->qpNoVbv : m_avgPFrameQp; |
2110 | 0 | m_avgPFrameQp = (m_avgPFrameQp + rce->qpNoVbv) / 2; |
2111 | 0 | } |
2112 | |
|
2113 | 0 | if (!m_param->bResetZoneConfig) |
2114 | 0 | { |
2115 | 0 | q = tuneQScaleForZone(rce, q); |
2116 | 0 | q = x265_clip3(lqmin, lqmax, q); |
2117 | 0 | } |
2118 | | /* Scenecut Aware QP offsets*/ |
2119 | 0 | if (m_param->bEnableSceneCutAwareQp) |
2120 | 0 | { |
2121 | 0 | double qmin = m_lmin[m_sliceType]; |
2122 | 0 | double qmax = m_lmax[m_sliceType]; |
2123 | 0 | q = scenecutAwareQp(curFrame, q); |
2124 | 0 | q = x265_clip3(qmin, qmax, q); |
2125 | 0 | rce->qpNoVbv = x265_qScale2qp(q); |
2126 | 0 | } |
2127 | 0 | q = clipQscale(curFrame, rce, q); |
2128 | | |
2129 | |
|
2130 | 0 | if (m_2pass) |
2131 | 0 | rce->frameSizePlanned = qScale2bits(rce, q); |
2132 | 0 | else |
2133 | 0 | rce->frameSizePlanned = predictSize(&m_pred[m_predType], q, (double)m_currentSatd); |
2134 | | |
2135 | | /* clip qp to permissible range after vbv-lookahead estimation to avoid possible |
2136 | | * mispredictions by initial frame size predictors, after each scenecut */ |
2137 | 0 | bool isFrameAfterScenecut = m_sliceType!= I_SLICE && m_curSlice->m_refFrameList[0][0]->m_lowres.bScenecut; |
2138 | 0 | if (!m_2pass && m_isVbv && isFrameAfterScenecut) |
2139 | 0 | q = x265_clip3(lqmin, lqmax, q); |
2140 | 0 | } |
2141 | 0 | m_lastQScaleFor[m_sliceType] = q; |
2142 | 0 | if ((m_curSlice->m_poc == 0 || m_lastQScaleFor[P_SLICE] < q) && !(m_2pass && !m_isVbv)) |
2143 | 0 | m_lastQScaleFor[P_SLICE] = q * fabs(m_param->rc.ipFactor); |
2144 | |
|
2145 | 0 | if (m_2pass) |
2146 | 0 | rce->frameSizePlanned = qScale2bits(rce, q); |
2147 | 0 | else |
2148 | 0 | rce->frameSizePlanned = predictSize(&m_pred[m_predType], q, (double)m_currentSatd); |
2149 | | |
2150 | | /* Always use up the whole VBV in this case. */ |
2151 | 0 | if (m_singleFrameVbv) |
2152 | 0 | rce->frameSizePlanned = m_bufferRate; |
2153 | | /* Limit planned size by MinCR */ |
2154 | 0 | if (m_isVbv) |
2155 | 0 | rce->frameSizePlanned = X265_MIN(rce->frameSizePlanned, rce->frameSizeMaximum); |
2156 | 0 | rce->frameSizeEstimated = rce->frameSizePlanned; |
2157 | 0 | rce->newQScale = q; |
2158 | 0 | return q; |
2159 | 0 | } |
2160 | 0 | } |
2161 | | |
2162 | | void RateControl::rateControlUpdateStats(RateControlEntry* rce) |
2163 | 0 | { |
2164 | 0 | if (!m_param->rc.bStatWrite && !m_param->rc.bStatRead) |
2165 | 0 | { |
2166 | 0 | if (rce->sliceType == I_SLICE) |
2167 | 0 | { |
2168 | | /* previous I still had a residual; roll it into the new loan */ |
2169 | 0 | if (m_partialResidualFrames) |
2170 | 0 | rce->rowTotalBits += m_partialResidualCost * m_partialResidualFrames; |
2171 | 0 | if ((m_param->totalFrames != 0) && (m_amortizeFrames > (m_param->totalFrames - m_framesDone))) |
2172 | 0 | { |
2173 | 0 | m_amortizeFrames = 0; |
2174 | 0 | m_amortizeFraction = 0; |
2175 | 0 | } |
2176 | 0 | else |
2177 | 0 | { |
2178 | 0 | double depreciateRate = 1.1; |
2179 | 0 | m_amortizeFrames = (int)(m_amortizeFrames / depreciateRate); |
2180 | 0 | m_amortizeFraction /= depreciateRate; |
2181 | 0 | m_amortizeFrames = X265_MAX(m_amortizeFrames, MIN_AMORTIZE_FRAME); |
2182 | 0 | m_amortizeFraction = X265_MAX(m_amortizeFraction, MIN_AMORTIZE_FRACTION); |
2183 | 0 | } |
2184 | 0 | rce->amortizeFrames = m_amortizeFrames; |
2185 | 0 | rce->amortizeFraction = m_amortizeFraction; |
2186 | 0 | m_partialResidualFrames = X265_MIN((int)rce->amortizeFrames, m_param->keyframeMax); |
2187 | 0 | m_partialResidualCost = (int)((rce->rowTotalBits * rce->amortizeFraction) / m_partialResidualFrames); |
2188 | 0 | rce->rowTotalBits -= m_partialResidualCost * m_partialResidualFrames; |
2189 | 0 | } |
2190 | 0 | else if (m_partialResidualFrames) |
2191 | 0 | { |
2192 | 0 | rce->rowTotalBits += m_partialResidualCost; |
2193 | 0 | m_partialResidualFrames--; |
2194 | 0 | } |
2195 | 0 | } |
2196 | 0 | if (rce->sliceType != B_SLICE) |
2197 | 0 | rce->rowCplxrSum = rce->rowTotalBits * x265_qp2qScale(rce->qpaRc) / rce->qRceq; |
2198 | 0 | else |
2199 | 0 | rce->rowCplxrSum = rce->rowTotalBits * x265_qp2qScale(rce->qpaRc) / (rce->qRceq * fabs(m_param->rc.pbFactor)); |
2200 | |
|
2201 | 0 | m_cplxrSum += rce->rowCplxrSum; |
2202 | 0 | m_totalBits += rce->rowTotalBits; |
2203 | | |
2204 | | /* do not allow the next frame to enter rateControlStart() until this |
2205 | | * frame has updated its mid-frame statistics */ |
2206 | 0 | if (m_param->rc.rateControlMode == X265_RC_ABR || m_isVbv) |
2207 | 0 | { |
2208 | 0 | m_startEndOrder.incr(); |
2209 | |
|
2210 | 0 | if (rce->encodeOrder < m_param->frameNumThreads - 1) |
2211 | 0 | m_startEndOrder.incr(); // faked rateControlEnd calls for negative frames |
2212 | 0 | } |
2213 | 0 | } |
2214 | | |
2215 | | void RateControl::checkAndResetABR(RateControlEntry* rce, bool isFrameDone) |
2216 | 0 | { |
2217 | 0 | double abrBuffer = 2 * m_rateTolerance * m_bitrate; |
2218 | | |
2219 | | // Check if current Slice is a scene cut that follows low detailed/blank frames |
2220 | 0 | if (rce->lastSatd > 4 * rce->movingAvgSum || rce->scenecut || rce->isFadeEnd) |
2221 | 0 | { |
2222 | 0 | if (!m_isAbrReset && rce->movingAvgSum > 0 |
2223 | 0 | && (m_isPatternPresent || !m_param->bframes)) |
2224 | 0 | { |
2225 | 0 | int pos = X265_MAX(m_sliderPos - m_param->frameNumThreads, 0); |
2226 | 0 | int64_t shrtTermWantedBits = (int64_t) (X265_MIN(pos, s_slidingWindowFrames) * m_bitrate * m_frameDuration); |
2227 | 0 | int64_t shrtTermTotalBitsSum = 0; |
2228 | | // Reset ABR if prev frames are blank to prevent further sudden overflows/ high bit rate spikes. |
2229 | 0 | for (int i = 0; i < s_slidingWindowFrames ; i++) |
2230 | 0 | shrtTermTotalBitsSum += m_encodedBitsWindow[i]; |
2231 | 0 | double underflow = (shrtTermTotalBitsSum - shrtTermWantedBits) / abrBuffer; |
2232 | 0 | const double epsilon = 0.0001f; |
2233 | 0 | if ((underflow < epsilon || rce->isFadeEnd) && !isFrameDone) |
2234 | 0 | { |
2235 | 0 | init(*m_curSlice->m_sps); |
2236 | | // Reduce tune complexity factor for scenes that follow blank frames |
2237 | 0 | double tuneCplxFactor = (m_ncu > 3600 && m_param->rc.cuTree && !m_param->rc.hevcAq) ? 2.5 : m_param->rc.hevcAq ? 1.5 : m_isGrainEnabled ? 1.9 : 1.0; |
2238 | 0 | m_cplxrSum /= tuneCplxFactor; |
2239 | 0 | m_shortTermCplxSum = rce->lastSatd / (CLIP_DURATION(m_frameDuration) / BASE_FRAME_DURATION); |
2240 | 0 | m_shortTermCplxCount = 1; |
2241 | 0 | m_isAbrReset = true; |
2242 | 0 | m_lastAbrResetPoc = rce->poc; |
2243 | 0 | } |
2244 | 0 | } |
2245 | 0 | else if (m_isAbrReset && isFrameDone) |
2246 | 0 | { |
2247 | | // Clear flag to reset ABR and continue as usual. |
2248 | 0 | m_isAbrReset = false; |
2249 | 0 | } |
2250 | 0 | } |
2251 | 0 | } |
2252 | | |
2253 | | void RateControl::hrdFullness(SEIBufferingPeriod *seiBP) |
2254 | 0 | { |
2255 | 0 | const VUI* vui = &m_curSlice->m_sps->vuiParameters; |
2256 | 0 | const HRDInfo* hrd = &vui->hrdParameters; |
2257 | 0 | int num = 90000; |
2258 | 0 | int denom = hrd->bitRateValue << (hrd->bitRateScale + BR_SHIFT); |
2259 | 0 | int64_t cpbState = (int64_t)m_bufferFillFinal; |
2260 | 0 | int64_t cpbSize = (int64_t)hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT); |
2261 | |
|
2262 | 0 | if (cpbState < 0 || cpbState > cpbSize) |
2263 | 0 | { |
2264 | 0 | x265_log(m_param, X265_LOG_WARNING, "CPB %s: %.0lf bits in a %.0lf-bit buffer\n", |
2265 | 0 | cpbState < 0 ? "underflow" : "overflow", (float)cpbState, (float)cpbSize); |
2266 | 0 | } |
2267 | |
|
2268 | 0 | seiBP->m_initialCpbRemovalDelay = (uint32_t)(num * cpbState / denom); |
2269 | 0 | seiBP->m_initialCpbRemovalDelayOffset = (uint32_t)(num * cpbSize / denom - seiBP->m_initialCpbRemovalDelay); |
2270 | 0 | } |
2271 | | |
2272 | | void RateControl::updateVbvPlan(Encoder* enc) |
2273 | 0 | { |
2274 | 0 | m_bufferFill = m_bufferFillFinal; |
2275 | 0 | enc->updateVbvPlan(this); |
2276 | 0 | } |
2277 | | |
2278 | | double RateControl::predictSize(Predictor *p, double q, double var) |
2279 | 0 | { |
2280 | 0 | return (p->coeff * var + p->offset) / (q * p->count); |
2281 | 0 | } |
2282 | | |
2283 | | double RateControl::clipQscale(Frame* curFrame, RateControlEntry* rce, double q) |
2284 | 0 | { |
2285 | | // B-frames are not directly subject to VBV, |
2286 | | // since they are controlled by referenced P-frames' QPs. |
2287 | 0 | double lmin = m_lmin[rce->sliceType]; |
2288 | 0 | double lmax = m_lmax[rce->sliceType]; |
2289 | 0 | double q0 = q; |
2290 | 0 | if (m_isVbv && m_currentSatd > 0 && curFrame) |
2291 | 0 | { |
2292 | 0 | if (m_param->lookaheadDepth || m_param->rc.cuTree || |
2293 | 0 | (m_param->scenecutThreshold || m_param->bHistBasedSceneCut) || |
2294 | 0 | (m_param->bFrameAdaptive && m_param->bframes)) |
2295 | 0 | { |
2296 | | /* Lookahead VBV: If lookahead is done, raise the quantizer as necessary |
2297 | | * such that no frames in the lookahead overflow and such that the buffer |
2298 | | * is in a reasonable state by the end of the lookahead. */ |
2299 | 0 | int loopTerminate = 0; |
2300 | | /* Avoid an infinite loop. */ |
2301 | 0 | for (int iterations = 0; iterations < 1000 && loopTerminate != 3; iterations++) |
2302 | 0 | { |
2303 | 0 | double frameQ[3]; |
2304 | 0 | double curBits; |
2305 | 0 | curBits = predictSize(&m_pred[m_predType], q, (double)m_currentSatd); |
2306 | 0 | double bufferFillCur = m_bufferFill - curBits; |
2307 | 0 | double targetFill; |
2308 | 0 | double totalDuration = m_frameDuration; |
2309 | 0 | frameQ[P_SLICE] = m_sliceType == I_SLICE ? q * m_param->rc.ipFactor : (m_sliceType == B_SLICE ? q / m_param->rc.pbFactor : q); |
2310 | 0 | frameQ[B_SLICE] = frameQ[P_SLICE] * m_param->rc.pbFactor; |
2311 | 0 | frameQ[I_SLICE] = frameQ[P_SLICE] / m_param->rc.ipFactor; |
2312 | | /* Loop over the planned future frames. */ |
2313 | 0 | bool iter = true; |
2314 | 0 | for (int j = 0; bufferFillCur >= 0 && iter ; j++) |
2315 | 0 | { |
2316 | 0 | int type = curFrame->m_lowres.plannedType[j]; |
2317 | 0 | if (type == X265_TYPE_AUTO || totalDuration >= 1.0) |
2318 | 0 | break; |
2319 | 0 | totalDuration += m_frameDuration; |
2320 | 0 | double wantedFrameSize = m_vbvMaxRate * m_frameDuration; |
2321 | 0 | if (bufferFillCur + wantedFrameSize <= m_bufferSize) |
2322 | 0 | bufferFillCur += wantedFrameSize; |
2323 | 0 | int64_t satd = curFrame->m_lowres.plannedSatd[j] >> (X265_DEPTH - 8); |
2324 | 0 | type = IS_X265_TYPE_I(type) ? I_SLICE : IS_X265_TYPE_B(type) ? B_SLICE : P_SLICE; |
2325 | 0 | int predType = getPredictorType(curFrame->m_lowres.plannedType[j], type); |
2326 | 0 | curBits = predictSize(&m_pred[predType], frameQ[type], (double)satd); |
2327 | 0 | bufferFillCur -= curBits; |
2328 | 0 | if (!m_param->bResetZoneConfig && ((uint64_t)j == (m_param->reconfigWindowSize - 1))) |
2329 | 0 | iter = false; |
2330 | 0 | } |
2331 | 0 | if (rce->vbvEndAdj) |
2332 | 0 | { |
2333 | 0 | bool loopBreak = false; |
2334 | 0 | double bufferDiff = m_param->vbvBufferEnd - (m_bufferFill / m_bufferSize); |
2335 | 0 | rce->targetFill = m_bufferFill + m_bufferSize * (bufferDiff / (m_param->totalFrames - rce->encodeOrder)); |
2336 | 0 | if (bufferFillCur < rce->targetFill) |
2337 | 0 | { |
2338 | 0 | q *= 1.01; |
2339 | 0 | loopTerminate |= 1; |
2340 | 0 | loopBreak = true; |
2341 | 0 | } |
2342 | 0 | if (bufferFillCur > m_param->vbvBufferEnd * m_bufferSize) |
2343 | 0 | { |
2344 | 0 | q /= 1.01; |
2345 | 0 | loopTerminate |= 2; |
2346 | 0 | loopBreak = true; |
2347 | 0 | } |
2348 | 0 | if (!loopBreak) |
2349 | 0 | break; |
2350 | 0 | } |
2351 | 0 | else |
2352 | 0 | { |
2353 | | /* Try to get the buffer at least 50% filled, but don't set an impossible goal. */ |
2354 | 0 | double finalDur = 1; |
2355 | 0 | if (m_param->rc.bStrictCbr) |
2356 | 0 | { |
2357 | 0 | finalDur = x265_clip3(0.4, 1.0, totalDuration); |
2358 | 0 | } |
2359 | 0 | targetFill = X265_MIN(m_bufferFill + totalDuration * m_vbvMaxRate * 0.5, m_bufferSize * (1 - 0.5 * finalDur)); |
2360 | 0 | if (bufferFillCur < targetFill) |
2361 | 0 | { |
2362 | 0 | q *= 1.01; |
2363 | 0 | loopTerminate |= 1; |
2364 | 0 | continue; |
2365 | 0 | } |
2366 | | /* Try to get the buffer not more than 80% filled, but don't set an impossible goal. */ |
2367 | 0 | targetFill = x265_clip3(m_bufferSize * (1 - 0.2 * finalDur), m_bufferSize, m_bufferFill - totalDuration * m_vbvMaxRate * 0.5); |
2368 | 0 | if (m_isCbr && bufferFillCur > targetFill && !m_isSceneTransition) |
2369 | 0 | { |
2370 | 0 | q /= 1.01; |
2371 | 0 | loopTerminate |= 2; |
2372 | 0 | continue; |
2373 | 0 | } |
2374 | 0 | break; |
2375 | 0 | } |
2376 | 0 | } |
2377 | 0 | q = X265_MAX(q0 / 2, q); |
2378 | 0 | } |
2379 | 0 | else |
2380 | 0 | { |
2381 | | /* Fallback to old purely-reactive algorithm: no lookahead. */ |
2382 | 0 | if ((m_sliceType == P_SLICE || m_sliceType == B_SLICE || |
2383 | 0 | (m_sliceType == I_SLICE && m_lastNonBPictType == I_SLICE)) && |
2384 | 0 | m_bufferFill / m_bufferSize < 0.5) |
2385 | 0 | { |
2386 | 0 | q /= x265_clip3(0.5, 1.0, 2.0 * m_bufferFill / m_bufferSize); |
2387 | 0 | } |
2388 | | // Now a hard threshold to make sure the frame fits in VBV. |
2389 | | // This one is mostly for I-frames. |
2390 | 0 | double bits = predictSize(&m_pred[m_predType], q, (double)m_currentSatd); |
2391 | | |
2392 | | // For small VBVs, allow the frame to use up the entire VBV. |
2393 | 0 | double maxFillFactor; |
2394 | 0 | maxFillFactor = m_bufferSize >= 5 * m_bufferRate ? 2 : 1; |
2395 | | // For single-frame VBVs, request that the frame use up the entire VBV. |
2396 | 0 | double minFillFactor = m_singleFrameVbv ? 1 : 2; |
2397 | |
|
2398 | 0 | for (int iterations = 0; iterations < 10; iterations++) |
2399 | 0 | { |
2400 | 0 | double qf = 1.0; |
2401 | 0 | if (bits > m_bufferFill / maxFillFactor) |
2402 | 0 | qf = x265_clip3(0.2, 1.0, m_bufferFill / (maxFillFactor * bits)); |
2403 | 0 | q /= qf; |
2404 | 0 | bits *= qf; |
2405 | 0 | if (bits < m_bufferRate / minFillFactor) |
2406 | 0 | q *= bits * minFillFactor / m_bufferRate; |
2407 | 0 | bits = predictSize(&m_pred[m_predType], q, (double)m_currentSatd); |
2408 | 0 | } |
2409 | |
|
2410 | 0 | q = X265_MAX(q0, q); |
2411 | 0 | } |
2412 | | |
2413 | | /* Apply MinCR restrictions */ |
2414 | 0 | double pbits = predictSize(&m_pred[m_predType], q, (double)m_currentSatd); |
2415 | 0 | if (pbits > rce->frameSizeMaximum) |
2416 | 0 | q *= pbits / rce->frameSizeMaximum; |
2417 | | /* To detect frames that are more complex in SATD costs compared to prev window, yet |
2418 | | * lookahead vbv reduces its qscale by half its value. Be on safer side and avoid drastic |
2419 | | * qscale reductions for frames high in complexity */ |
2420 | 0 | bool mispredCheck = rce->movingAvgSum && m_currentSatd >= rce->movingAvgSum && q <= q0 / 2; |
2421 | 0 | if (!m_isCbr || (m_isAbr && mispredCheck)) |
2422 | 0 | q = X265_MAX(q0, q); |
2423 | |
|
2424 | 0 | if (m_rateFactorMaxIncrement) |
2425 | 0 | { |
2426 | 0 | double qpNoVbv = x265_qScale2qp(q0); |
2427 | 0 | double qmax = X265_MIN(lmax,x265_qp2qScale(qpNoVbv + m_rateFactorMaxIncrement)); |
2428 | 0 | return x265_clip3(lmin, qmax, q); |
2429 | 0 | } |
2430 | 0 | } |
2431 | 0 | if (m_2pass) |
2432 | 0 | { |
2433 | 0 | double min = log(lmin); |
2434 | 0 | double max = log(lmax); |
2435 | 0 | q = (log(q) - min) / (max - min) - 0.5; |
2436 | 0 | q = 1.0 / (1.0 + exp(-4 * q)); |
2437 | 0 | q = q*(max - min) + min; |
2438 | 0 | return exp(q); |
2439 | 0 | } |
2440 | 0 | return x265_clip3(lmin, lmax, q); |
2441 | 0 | } |
2442 | | |
2443 | | double RateControl::predictRowsSizeSum(Frame* curFrame, RateControlEntry* rce, double qpVbv, int32_t& encodedBitsSoFar) |
2444 | 0 | { |
2445 | 0 | uint32_t rowSatdCostSoFar = 0, totalSatdBits = 0; |
2446 | 0 | encodedBitsSoFar = 0; |
2447 | |
|
2448 | 0 | double qScale = x265_qp2qScale(qpVbv); |
2449 | 0 | FrameData& curEncData = *curFrame->m_encData; |
2450 | 0 | int picType = curEncData.m_slice->m_sliceType; |
2451 | 0 | Frame* refFrame = curEncData.m_slice->m_refFrameList[0][0]; |
2452 | |
|
2453 | 0 | uint32_t maxRows = curEncData.m_slice->m_sps->numCuInHeight; |
2454 | 0 | uint32_t maxCols = curEncData.m_slice->m_sps->numCuInWidth; |
2455 | |
|
2456 | 0 | for (uint32_t row = 0; row < maxRows; row++) |
2457 | 0 | { |
2458 | 0 | encodedBitsSoFar += curEncData.m_rowStat[row].encodedBits; |
2459 | 0 | rowSatdCostSoFar = curEncData.m_rowStat[row].rowSatd; |
2460 | 0 | uint32_t satdCostForPendingCus = curEncData.m_rowStat[row].satdForVbv - rowSatdCostSoFar; |
2461 | 0 | satdCostForPendingCus >>= X265_DEPTH - 8; |
2462 | 0 | if (satdCostForPendingCus > 0) |
2463 | 0 | { |
2464 | 0 | double pred_s = predictSize(rce->rowPred[0], qScale, satdCostForPendingCus); |
2465 | 0 | uint32_t refRowSatdCost = 0, refRowBits = 0, intraCostForPendingCus = 0; |
2466 | 0 | double refQScale = 0; |
2467 | |
|
2468 | 0 | if (picType != I_SLICE && !m_param->rc.bEnableConstVbv) |
2469 | 0 | { |
2470 | 0 | FrameData& refEncData = *refFrame->m_encData; |
2471 | 0 | uint32_t endCuAddr = maxCols * (row + 1); |
2472 | 0 | uint32_t startCuAddr = curEncData.m_rowStat[row].numEncodedCUs; |
2473 | 0 | if (startCuAddr) |
2474 | 0 | { |
2475 | 0 | for (uint32_t cuAddr = startCuAddr + 1 ; cuAddr < endCuAddr; cuAddr++) |
2476 | 0 | { |
2477 | 0 | refRowSatdCost += refEncData.m_cuStat[cuAddr].vbvCost; |
2478 | 0 | refRowBits += refEncData.m_cuStat[cuAddr].totalBits; |
2479 | 0 | } |
2480 | 0 | } |
2481 | 0 | else |
2482 | 0 | { |
2483 | 0 | refRowBits = refEncData.m_rowStat[row].encodedBits; |
2484 | 0 | refRowSatdCost = refEncData.m_rowStat[row].satdForVbv; |
2485 | 0 | } |
2486 | |
|
2487 | 0 | refRowSatdCost >>= X265_DEPTH - 8; |
2488 | 0 | refQScale = refEncData.m_rowStat[row].rowQpScale; |
2489 | 0 | } |
2490 | |
|
2491 | 0 | if (picType == I_SLICE || qScale >= refQScale) |
2492 | 0 | { |
2493 | 0 | if (picType == P_SLICE |
2494 | 0 | && refFrame |
2495 | 0 | && refFrame->m_encData->m_slice->m_sliceType == picType |
2496 | 0 | && refQScale > 0 |
2497 | 0 | && refRowBits > 0 |
2498 | 0 | && !m_param->rc.bEnableConstVbv) |
2499 | 0 | { |
2500 | 0 | if (abs((int32_t)(refRowSatdCost - satdCostForPendingCus)) < (int32_t)satdCostForPendingCus / 2) |
2501 | 0 | { |
2502 | 0 | double predTotal = refRowBits * satdCostForPendingCus / refRowSatdCost * refQScale / qScale; |
2503 | 0 | totalSatdBits += (int32_t)((pred_s + predTotal) * 0.5); |
2504 | 0 | continue; |
2505 | 0 | } |
2506 | 0 | } |
2507 | 0 | totalSatdBits += (int32_t)pred_s; |
2508 | 0 | } |
2509 | 0 | else if (picType == P_SLICE) |
2510 | 0 | { |
2511 | 0 | intraCostForPendingCus = curEncData.m_rowStat[row].intraSatdForVbv - curEncData.m_rowStat[row].rowIntraSatd; |
2512 | 0 | intraCostForPendingCus >>= X265_DEPTH - 8; |
2513 | | /* Our QP is lower than the reference! */ |
2514 | 0 | double pred_intra = predictSize(rce->rowPred[1], qScale, intraCostForPendingCus); |
2515 | | /* Sum: better to overestimate than underestimate by using only one of the two predictors. */ |
2516 | 0 | totalSatdBits += (int32_t)(pred_intra + pred_s); |
2517 | 0 | } |
2518 | 0 | else |
2519 | 0 | totalSatdBits += (int32_t)pred_s; |
2520 | 0 | } |
2521 | 0 | } |
2522 | |
|
2523 | 0 | return totalSatdBits + encodedBitsSoFar; |
2524 | 0 | } |
2525 | | |
2526 | | int RateControl::rowVbvRateControl(Frame* curFrame, uint32_t row, RateControlEntry* rce, double& qpVbv, uint32_t* m_sliceBaseRow, uint32_t sliceId) |
2527 | 0 | { |
2528 | 0 | FrameData& curEncData = *curFrame->m_encData; |
2529 | 0 | double qScaleVbv = x265_qp2qScale(qpVbv); |
2530 | 0 | uint64_t rowSatdCost = curEncData.m_rowStat[row].rowSatd; |
2531 | 0 | double encodedBits = curEncData.m_rowStat[row].encodedBits; |
2532 | 0 | uint32_t rowInSlice = row - m_sliceBaseRow[sliceId]; |
2533 | |
|
2534 | 0 | if (m_param->bEnableWavefront && rowInSlice == 1) |
2535 | 0 | { |
2536 | 0 | rowSatdCost += curEncData.m_rowStat[row - 1].rowSatd; |
2537 | 0 | encodedBits += curEncData.m_rowStat[row - 1].encodedBits; |
2538 | 0 | } |
2539 | 0 | rowSatdCost >>= X265_DEPTH - 8; |
2540 | 0 | updatePredictor(rce->rowPred[0], qScaleVbv, (double)rowSatdCost, encodedBits); |
2541 | 0 | if (curEncData.m_slice->m_sliceType != I_SLICE && !m_param->rc.bEnableConstVbv) |
2542 | 0 | { |
2543 | 0 | Frame* refFrame = curEncData.m_slice->m_refFrameList[0][0]; |
2544 | 0 | if (qpVbv < refFrame->m_encData->m_rowStat[row].rowQp) |
2545 | 0 | { |
2546 | 0 | uint64_t intraRowSatdCost = curEncData.m_rowStat[row].rowIntraSatd; |
2547 | 0 | if (m_param->bEnableWavefront && rowInSlice == 1) |
2548 | 0 | intraRowSatdCost += curEncData.m_rowStat[row - 1].rowIntraSatd; |
2549 | 0 | intraRowSatdCost >>= X265_DEPTH - 8; |
2550 | 0 | updatePredictor(rce->rowPred[1], qScaleVbv, (double)intraRowSatdCost, encodedBits); |
2551 | 0 | } |
2552 | 0 | } |
2553 | |
|
2554 | 0 | int canReencodeRow = 1; |
2555 | | /* tweak quality based on difference from predicted size */ |
2556 | 0 | double prevRowQp = qpVbv; |
2557 | 0 | double qpAbsoluteMax = m_param->rc.qpMax; |
2558 | 0 | double qpAbsoluteMin = m_param->rc.qpMin; |
2559 | 0 | if (m_rateFactorMaxIncrement) |
2560 | 0 | qpAbsoluteMax = X265_MIN(qpAbsoluteMax, rce->qpNoVbv + m_rateFactorMaxIncrement); |
2561 | |
|
2562 | 0 | if (m_rateFactorMaxDecrement) |
2563 | 0 | qpAbsoluteMin = X265_MAX(qpAbsoluteMin, rce->qpNoVbv - m_rateFactorMaxDecrement); |
2564 | |
|
2565 | 0 | double qpMax = X265_MIN(prevRowQp + m_param->rc.qpStep, qpAbsoluteMax); |
2566 | 0 | double qpMin = X265_MAX(prevRowQp - m_param->rc.qpStep, qpAbsoluteMin); |
2567 | 0 | double stepSize = 0.5; |
2568 | 0 | double bufferLeftPlanned = rce->bufferFill - rce->frameSizePlanned; |
2569 | |
|
2570 | 0 | const SPS& sps = *curEncData.m_slice->m_sps; |
2571 | 0 | double maxFrameError = X265_MAX(0.05, 1.0 / sps.numCuInHeight); |
2572 | |
|
2573 | 0 | if (row < m_sliceBaseRow[sliceId + 1] - 1) |
2574 | 0 | { |
2575 | | /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */ |
2576 | 0 | double rcTol = bufferLeftPlanned / m_param->frameNumThreads * m_rateTolerance; |
2577 | 0 | int32_t encodedBitsSoFar = 0; |
2578 | 0 | double accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar); |
2579 | 0 | double vbvEndBias = 0.95; |
2580 | | |
2581 | | /* * Don't increase the row QPs until a sufficent amount of the bits of |
2582 | | * the frame have been processed, in case a flat area at the top of the |
2583 | | * frame was measured inaccurately. */ |
2584 | 0 | if (encodedBitsSoFar < 0.05f * rce->frameSizePlanned) |
2585 | 0 | qpMax = qpAbsoluteMax = prevRowQp; |
2586 | |
|
2587 | 0 | if (rce->sliceType != I_SLICE || (m_param->rc.bStrictCbr && rce->poc > 0)) |
2588 | 0 | rcTol *= 0.5; |
2589 | |
|
2590 | 0 | if (!m_isCbr) |
2591 | 0 | qpMin = X265_MAX(qpMin, rce->qpNoVbv); |
2592 | |
|
2593 | 0 | double totalBitsNeeded = m_wantedBitsWindow; |
2594 | 0 | if (m_param->totalFrames) |
2595 | 0 | totalBitsNeeded = (m_param->totalFrames * m_bitrate) / m_fps; |
2596 | 0 | double abrOvershoot = (accFrameBits + m_totalBits - m_wantedBitsWindow) / totalBitsNeeded; |
2597 | |
|
2598 | 0 | while (qpVbv < qpMax |
2599 | 0 | && (((accFrameBits > rce->frameSizePlanned + rcTol) || |
2600 | 0 | (rce->bufferFill - accFrameBits < bufferLeftPlanned * 0.5) || |
2601 | 0 | (accFrameBits > rce->frameSizePlanned && qpVbv < rce->qpNoVbv) || |
2602 | 0 | (rce->vbvEndAdj && ((rce->bufferFill - accFrameBits) < (rce->targetFill * vbvEndBias)))) |
2603 | 0 | && (!m_param->rc.bStrictCbr ? 1 : abrOvershoot > 0.1))) |
2604 | 0 | { |
2605 | 0 | qpVbv += stepSize; |
2606 | 0 | accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar); |
2607 | 0 | abrOvershoot = (accFrameBits + m_totalBits - m_wantedBitsWindow) / totalBitsNeeded; |
2608 | 0 | } |
2609 | |
|
2610 | 0 | while (qpVbv > qpMin |
2611 | 0 | && (qpVbv > curEncData.m_rowStat[0].rowQp || m_singleFrameVbv) |
2612 | 0 | && (((accFrameBits < rce->frameSizePlanned * 0.8f && qpVbv <= prevRowQp) |
2613 | 0 | || accFrameBits < (rce->bufferFill - m_bufferSize + m_bufferRate) * 1.1 |
2614 | 0 | || (rce->vbvEndAdj && ((rce->bufferFill - accFrameBits) > (rce->targetFill * vbvEndBias)))) |
2615 | 0 | && (!m_param->rc.bStrictCbr ? 1 : abrOvershoot < 0))) |
2616 | 0 | { |
2617 | 0 | qpVbv -= stepSize; |
2618 | 0 | accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar); |
2619 | 0 | abrOvershoot = (accFrameBits + m_totalBits - m_wantedBitsWindow) / totalBitsNeeded; |
2620 | 0 | } |
2621 | |
|
2622 | 0 | if (m_param->rc.bStrictCbr && m_param->totalFrames) |
2623 | 0 | { |
2624 | 0 | double timeDone = (double)(m_framesDone) / m_param->totalFrames; |
2625 | 0 | while (qpVbv < qpMax && (qpVbv < rce->qpNoVbv + (m_param->rc.qpStep * timeDone)) && |
2626 | 0 | (timeDone > 0.75 && abrOvershoot > 0)) |
2627 | 0 | { |
2628 | 0 | qpVbv += stepSize; |
2629 | 0 | accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar); |
2630 | 0 | abrOvershoot = (accFrameBits + m_totalBits - m_wantedBitsWindow) / totalBitsNeeded; |
2631 | 0 | } |
2632 | 0 | if (qpVbv > curEncData.m_rowStat[0].rowQp && |
2633 | 0 | abrOvershoot < -0.1 && timeDone > 0.5 && accFrameBits < rce->frameSizePlanned - rcTol) |
2634 | 0 | { |
2635 | 0 | qpVbv -= stepSize; |
2636 | 0 | accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar); |
2637 | 0 | } |
2638 | 0 | } |
2639 | | |
2640 | | /* avoid VBV underflow or MinCr violation */ |
2641 | 0 | while ((qpVbv < qpAbsoluteMax) |
2642 | 0 | && ((rce->bufferFill - accFrameBits < m_bufferRate * maxFrameError) || |
2643 | 0 | (rce->frameSizeMaximum - accFrameBits < rce->frameSizeMaximum * maxFrameError))) |
2644 | 0 | { |
2645 | 0 | qpVbv += stepSize; |
2646 | 0 | accFrameBits = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar); |
2647 | 0 | } |
2648 | |
|
2649 | 0 | rce->frameSizeEstimated = accFrameBits; |
2650 | | |
2651 | | /* If the current row was large enough to cause a large QP jump, try re-encoding it. */ |
2652 | 0 | if (qpVbv > qpMax && prevRowQp < qpMax && canReencodeRow) |
2653 | 0 | { |
2654 | | /* Bump QP to halfway in between... close enough. */ |
2655 | 0 | qpVbv = x265_clip3(prevRowQp + 1.0f, qpMax, (prevRowQp + qpVbv) * 0.5); |
2656 | 0 | return -1; |
2657 | 0 | } |
2658 | | |
2659 | 0 | if (m_param->rc.rfConstantMin) |
2660 | 0 | { |
2661 | 0 | if (qpVbv < qpMin && prevRowQp > qpMin && canReencodeRow) |
2662 | 0 | { |
2663 | 0 | qpVbv = x265_clip3(qpMin, prevRowQp, (prevRowQp + qpVbv) * 0.5); |
2664 | 0 | return -1; |
2665 | 0 | } |
2666 | 0 | } |
2667 | 0 | } |
2668 | 0 | else |
2669 | 0 | { |
2670 | 0 | int32_t encodedBitsSoFar = 0; |
2671 | 0 | rce->frameSizeEstimated = predictRowsSizeSum(curFrame, rce, qpVbv, encodedBitsSoFar); |
2672 | | |
2673 | | /* Last-ditch attempt: if the last row of the frame underflowed the VBV, |
2674 | | * try again. */ |
2675 | 0 | if ((rce->frameSizeEstimated > (rce->bufferFill - m_bufferRate * maxFrameError) && |
2676 | 0 | qpVbv < qpMax && canReencodeRow)) |
2677 | 0 | { |
2678 | 0 | qpVbv = qpMax; |
2679 | 0 | return -1; |
2680 | 0 | } |
2681 | 0 | } |
2682 | 0 | return 0; |
2683 | 0 | } |
2684 | | |
2685 | | /* modify the bitrate curve from pass1 for one frame */ |
2686 | | double RateControl::getQScale(RateControlEntry *rce, double rateFactor) |
2687 | 0 | { |
2688 | 0 | double q; |
2689 | |
|
2690 | 0 | if (m_param->rc.cuTree && !m_param->rc.hevcAq) |
2691 | 0 | { |
2692 | | // Scale and units are obtained from rateNum and rateDenom for videos with fixed frame rates. |
2693 | 0 | double timescale = (double)m_param->fpsDenom / (2 * m_param->fpsNum); |
2694 | 0 | q = pow(BASE_FRAME_DURATION / CLIP_DURATION(2 * timescale), 1 - m_param->rc.qCompress); |
2695 | 0 | } |
2696 | 0 | else |
2697 | 0 | q = pow(rce->blurredComplexity, 1 - m_param->rc.qCompress); |
2698 | | |
2699 | | // avoid NaN's in the Rceq |
2700 | 0 | if (rce->coeffBits + rce->mvBits == 0) |
2701 | 0 | q = m_lastQScaleFor[rce->sliceType]; |
2702 | 0 | else |
2703 | 0 | { |
2704 | 0 | m_lastRceq = q; |
2705 | 0 | q /= rateFactor; |
2706 | 0 | } |
2707 | |
|
2708 | 0 | return q; |
2709 | 0 | } |
2710 | | |
2711 | | void RateControl::updatePredictor(Predictor *p, double q, double var, double bits) |
2712 | 0 | { |
2713 | 0 | if (var < 10) |
2714 | 0 | return; |
2715 | 0 | const double range = 2; |
2716 | 0 | double old_coeff = p->coeff / p->count; |
2717 | 0 | double old_offset = p->offset / p->count; |
2718 | 0 | double new_coeff = X265_MAX((bits * q - old_offset) / var, p->coeffMin ); |
2719 | 0 | double new_coeff_clipped = x265_clip3(old_coeff / range, old_coeff * range, new_coeff); |
2720 | 0 | double new_offset = bits * q - new_coeff_clipped * var; |
2721 | 0 | if (new_offset >= 0) |
2722 | 0 | new_coeff = new_coeff_clipped; |
2723 | 0 | else |
2724 | 0 | new_offset = 0; |
2725 | 0 | p->count *= p->decay; |
2726 | 0 | p->coeff *= p->decay; |
2727 | 0 | p->offset *= p->decay; |
2728 | 0 | p->count++; |
2729 | 0 | p->coeff += new_coeff; |
2730 | 0 | p->offset += new_offset; |
2731 | 0 | } |
2732 | | |
2733 | | int RateControl::updateVbv(int64_t bits, RateControlEntry* rce) |
2734 | 0 | { |
2735 | 0 | int predType = rce->sliceType; |
2736 | 0 | int filler = 0; |
2737 | 0 | double bufferBits; |
2738 | 0 | predType = rce->sliceType == B_SLICE && rce->keptAsRef ? 3 : predType; |
2739 | 0 | if (rce->lastSatd >= m_ncu && rce->encodeOrder >= m_lastPredictorReset) |
2740 | 0 | updatePredictor(&m_pred[predType], x265_qp2qScale(rce->qpaRc), (double)rce->lastSatd, (double)bits); |
2741 | 0 | if (!m_isVbv) |
2742 | 0 | return 0; |
2743 | | |
2744 | 0 | m_bufferFillFinal -= bits; |
2745 | |
|
2746 | 0 | if (m_bufferFillFinal < 0) |
2747 | 0 | x265_log(m_param, X265_LOG_WARNING, "poc:%d, VBV underflow (%.0f bits)\n", rce->poc, m_bufferFillFinal); |
2748 | |
|
2749 | 0 | m_bufferFillFinal = X265_MAX(m_bufferFillFinal, 0); |
2750 | 0 | m_bufferFillFinal += rce->bufferRate; |
2751 | 0 | if (m_param->csvLogLevel >= 2) |
2752 | 0 | m_unclippedBufferFillFinal = m_bufferFillFinal; |
2753 | |
|
2754 | 0 | if (m_param->rc.bStrictCbr) |
2755 | 0 | { |
2756 | 0 | if (m_bufferFillFinal > m_bufferSize) |
2757 | 0 | { |
2758 | 0 | filler = (int)(m_bufferFillFinal - m_bufferSize); |
2759 | 0 | filler += FILLER_OVERHEAD * 8; |
2760 | 0 | } |
2761 | 0 | m_bufferFillFinal -= filler; |
2762 | 0 | bufferBits = X265_MIN(bits + filler + m_bufferExcess, rce->bufferRate); |
2763 | 0 | m_bufferExcess = X265_MAX(m_bufferExcess - bufferBits + bits + filler, 0); |
2764 | 0 | m_bufferFillActual += bufferBits - bits - filler; |
2765 | 0 | } |
2766 | 0 | else |
2767 | 0 | { |
2768 | 0 | m_bufferFillFinal = X265_MIN(m_bufferFillFinal, m_bufferSize); |
2769 | 0 | bufferBits = X265_MIN(bits + m_bufferExcess, rce->bufferRate); |
2770 | 0 | m_bufferExcess = X265_MAX(m_bufferExcess - bufferBits + bits, 0); |
2771 | 0 | m_bufferFillActual += bufferBits - bits; |
2772 | 0 | m_bufferFillActual = X265_MIN(m_bufferFillActual, m_bufferSize); |
2773 | 0 | } |
2774 | 0 | return filler; |
2775 | 0 | } |
2776 | | |
2777 | | /* After encoding one frame, update rate control state */ |
2778 | | int RateControl::rateControlEnd(Frame* curFrame, int64_t bits, RateControlEntry* rce, int *filler) |
2779 | 0 | { |
2780 | 0 | int orderValue = m_startEndOrder.get(); |
2781 | 0 | int endOrdinal = (rce->encodeOrder + m_param->frameNumThreads) * 2 - 1; |
2782 | 0 | while (orderValue < endOrdinal && !m_bTerminated) |
2783 | 0 | { |
2784 | | /* no more frames are being encoded, so fake the start event if we would |
2785 | | * have blocked on it. Note that this does not enforce rateControlEnd() |
2786 | | * ordering during flush, but this has no impact on the outputs */ |
2787 | 0 | if (m_finalFrameCount && orderValue >= 2 * m_finalFrameCount) |
2788 | 0 | break; |
2789 | 0 | orderValue = m_startEndOrder.waitForChange(orderValue); |
2790 | 0 | } |
2791 | |
|
2792 | 0 | FrameData& curEncData = *curFrame->m_encData; |
2793 | 0 | int64_t actualBits = bits; |
2794 | 0 | Slice *slice = curEncData.m_slice; |
2795 | 0 | bool bEnableDistOffset = m_param->analysisMultiPassDistortion && m_param->rc.bStatRead; |
2796 | |
|
2797 | 0 | if (m_param->rc.aqMode || m_isVbv || m_param->bAQMotion || bEnableDistOffset) |
2798 | 0 | { |
2799 | 0 | if (m_isVbv && !(m_2pass && m_param->rc.rateControlMode == X265_RC_CRF)) |
2800 | 0 | { |
2801 | 0 | double avgQpRc = 0; |
2802 | | /* determine avg QP decided by VBV rate control */ |
2803 | 0 | for (uint32_t i = 0; i < slice->m_sps->numCuInHeight; i++) |
2804 | 0 | avgQpRc += curEncData.m_rowStat[i].sumQpRc; |
2805 | |
|
2806 | 0 | avgQpRc /= slice->m_sps->numCUsInFrame; |
2807 | 0 | curEncData.m_avgQpRc = x265_clip3((double)m_param->rc.qpMin, (double)m_param->rc.qpMax, avgQpRc); |
2808 | 0 | rce->qpaRc = curEncData.m_avgQpRc; |
2809 | 0 | } |
2810 | |
|
2811 | 0 | if (m_param->rc.aqMode || m_param->bAQMotion || bEnableDistOffset) |
2812 | 0 | { |
2813 | 0 | double avgQpAq = 0; |
2814 | | /* determine actual avg encoded QP, after AQ/cutree/distortion adjustments */ |
2815 | 0 | for (uint32_t i = 0; i < slice->m_sps->numCuInHeight; i++) |
2816 | 0 | avgQpAq += curEncData.m_rowStat[i].sumQpAq; |
2817 | |
|
2818 | 0 | avgQpAq /= (slice->m_sps->numCUsInFrame * m_param->num4x4Partitions); |
2819 | 0 | curEncData.m_avgQpAq = avgQpAq; |
2820 | 0 | } |
2821 | 0 | else |
2822 | 0 | curEncData.m_avgQpAq = curEncData.m_avgQpRc; |
2823 | 0 | } |
2824 | |
|
2825 | 0 | if (m_isAbr) |
2826 | 0 | { |
2827 | 0 | if (m_param->rc.rateControlMode == X265_RC_ABR && !m_param->rc.bStatRead) |
2828 | 0 | checkAndResetABR(rce, true); |
2829 | 0 | } |
2830 | 0 | if (m_param->rc.rateControlMode == X265_RC_CRF) |
2831 | 0 | { |
2832 | 0 | double crfVal, qpRef = curEncData.m_avgQpRc; |
2833 | 0 | bool is2passCrfChange = false; |
2834 | 0 | if (m_2pass) |
2835 | 0 | { |
2836 | 0 | if (fabs(curEncData.m_avgQpRc - rce->qpPrev) > 0.1) |
2837 | 0 | { |
2838 | 0 | qpRef = rce->qpPrev; |
2839 | 0 | is2passCrfChange = true; |
2840 | 0 | } |
2841 | 0 | } |
2842 | 0 | if (is2passCrfChange || fabs(qpRef - rce->qpNoVbv) > 0.5) |
2843 | 0 | { |
2844 | 0 | double crfFactor = rce->qRceq /x265_qp2qScale(qpRef); |
2845 | 0 | double baseCplx = m_ncu * (m_param->bframes ? 120 : 80); |
2846 | 0 | double mbtree_offset = m_param->rc.cuTree ? (1.0 - m_param->rc.qCompress) * 13.5 : 0; |
2847 | 0 | crfVal = x265_qScale2qp(pow(baseCplx, 1 - m_qCompress) / crfFactor) - mbtree_offset; |
2848 | 0 | } |
2849 | 0 | else |
2850 | 0 | crfVal = rce->sliceType == I_SLICE ? m_param->rc.rfConstant - m_ipOffset : |
2851 | 0 | (rce->sliceType == B_SLICE ? m_param->rc.rfConstant + m_pbOffset : m_param->rc.rfConstant); |
2852 | |
|
2853 | 0 | curEncData.m_rateFactor = crfVal; |
2854 | 0 | } |
2855 | |
|
2856 | 0 | if (m_isAbr && !m_isAbrReset) |
2857 | 0 | { |
2858 | | /* amortize part of each I slice over the next several frames, up to |
2859 | | * keyint-max, to avoid over-compensating for the large I slice cost */ |
2860 | 0 | if (!m_param->rc.bStatWrite && !m_param->rc.bStatRead) |
2861 | 0 | { |
2862 | 0 | if (rce->sliceType == I_SLICE) |
2863 | 0 | { |
2864 | | /* previous I still had a residual; roll it into the new loan */ |
2865 | 0 | if (m_residualFrames) |
2866 | 0 | bits += m_residualCost * m_residualFrames; |
2867 | 0 | m_residualFrames = X265_MIN((int)rce->amortizeFrames, m_param->keyframeMax); |
2868 | 0 | m_residualCost = (int)((bits * rce->amortizeFraction) / m_residualFrames); |
2869 | 0 | bits -= m_residualCost * m_residualFrames; |
2870 | 0 | } |
2871 | 0 | else if (m_residualFrames) |
2872 | 0 | { |
2873 | 0 | bits += m_residualCost; |
2874 | 0 | m_residualFrames--; |
2875 | 0 | } |
2876 | 0 | } |
2877 | 0 | if (rce->sliceType != B_SLICE) |
2878 | 0 | { |
2879 | | /* The factor 1.5 is to tune up the actual bits, otherwise the cplxrSum is scaled too low |
2880 | | * to improve short term compensation for next frame. */ |
2881 | 0 | m_cplxrSum += (bits * x265_qp2qScale(rce->qpaRc) / rce->qRceq) - (rce->rowCplxrSum); |
2882 | 0 | } |
2883 | 0 | else |
2884 | 0 | { |
2885 | | /* Depends on the fact that B-frame's QP is an offset from the following P-frame's. |
2886 | | * Not perfectly accurate with B-refs, but good enough. */ |
2887 | 0 | m_cplxrSum += (bits * x265_qp2qScale(rce->qpaRc) / (rce->qRceq * fabs(m_param->rc.pbFactor))) - (rce->rowCplxrSum); |
2888 | 0 | } |
2889 | 0 | m_wantedBitsWindow += m_frameDuration * m_bitrate; |
2890 | 0 | m_totalBits += bits - rce->rowTotalBits; |
2891 | 0 | m_encodedBits += actualBits; |
2892 | 0 | int pos = m_sliderPos - m_param->frameNumThreads; |
2893 | 0 | if (pos >= 0) |
2894 | 0 | m_encodedBitsWindow[pos % s_slidingWindowFrames] = actualBits; |
2895 | 0 | if(rce->sliceType != I_SLICE) |
2896 | 0 | { |
2897 | 0 | int qp = int (rce->qpaRc + 0.5); |
2898 | 0 | m_qpToEncodedBits[qp] = m_qpToEncodedBits[qp] == 0 ? actualBits : (m_qpToEncodedBits[qp] + actualBits) * 0.5; |
2899 | 0 | } |
2900 | 0 | curFrame->m_rcData->wantedBitsWindow = m_wantedBitsWindow; |
2901 | 0 | curFrame->m_rcData->cplxrSum = m_cplxrSum; |
2902 | 0 | curFrame->m_rcData->totalBits = m_totalBits; |
2903 | 0 | curFrame->m_rcData->encodedBits = m_encodedBits; |
2904 | 0 | } |
2905 | |
|
2906 | 0 | if (m_2pass) |
2907 | 0 | { |
2908 | 0 | m_expectedBitsSum += qScale2bits(rce, x265_qp2qScale(rce->newQp)); |
2909 | 0 | m_totalBits += bits - rce->rowTotalBits; |
2910 | 0 | } |
2911 | |
|
2912 | 0 | if (m_isVbv) |
2913 | 0 | { |
2914 | 0 | *filler = updateVbv(actualBits, rce); |
2915 | |
|
2916 | 0 | curFrame->m_rcData->bufferFillFinal = m_bufferFillFinal; |
2917 | 0 | for (int i = 0; i < 4; i++) |
2918 | 0 | { |
2919 | 0 | curFrame->m_rcData->coeff[i] = m_pred[i].coeff; |
2920 | 0 | curFrame->m_rcData->count[i] = m_pred[i].count; |
2921 | 0 | curFrame->m_rcData->offset[i] = m_pred[i].offset; |
2922 | 0 | } |
2923 | 0 | if (m_param->bEmitHRDSEI) |
2924 | 0 | { |
2925 | 0 | const VUI *vui = &curEncData.m_slice->m_sps->vuiParameters; |
2926 | 0 | const HRDInfo *hrd = &vui->hrdParameters; |
2927 | 0 | const TimingInfo *time = &vui->timingInfo; |
2928 | 0 | if (!curFrame->m_poc) |
2929 | 0 | { |
2930 | | // first access unit initializes the HRD |
2931 | 0 | rce->hrdTiming->cpbInitialAT = 0; |
2932 | 0 | rce->hrdTiming->cpbRemovalTime = m_nominalRemovalTime = (double)m_bufPeriodSEI.m_initialCpbRemovalDelay / 90000; |
2933 | 0 | } |
2934 | 0 | else |
2935 | 0 | { |
2936 | 0 | rce->hrdTiming->cpbRemovalTime = m_nominalRemovalTime + (double)rce->picTimingSEI->m_auCpbRemovalDelay * time->numUnitsInTick / time->timeScale; |
2937 | 0 | double cpbEarliestAT = rce->hrdTiming->cpbRemovalTime - (double)m_bufPeriodSEI.m_initialCpbRemovalDelay / 90000; |
2938 | 0 | if (!curFrame->m_lowres.bKeyframe) |
2939 | 0 | cpbEarliestAT -= (double)m_bufPeriodSEI.m_initialCpbRemovalDelayOffset / 90000; |
2940 | |
|
2941 | 0 | rce->hrdTiming->cpbInitialAT = hrd->cbrFlag ? m_prevCpbFinalAT : X265_MAX(m_prevCpbFinalAT, cpbEarliestAT); |
2942 | 0 | } |
2943 | 0 | int filler_bits = *filler ? (*filler - START_CODE_OVERHEAD * 8) : 0; |
2944 | 0 | uint32_t cpbsizeUnscale = hrd->cpbSizeValue << (hrd->cpbSizeScale + CPB_SHIFT); |
2945 | 0 | rce->hrdTiming->cpbFinalAT = m_prevCpbFinalAT = rce->hrdTiming->cpbInitialAT + (actualBits + filler_bits)/ cpbsizeUnscale; |
2946 | 0 | rce->hrdTiming->dpbOutputTime = (double)rce->picTimingSEI->m_picDpbOutputDelay * time->numUnitsInTick / time->timeScale + rce->hrdTiming->cpbRemovalTime; |
2947 | 0 | } |
2948 | 0 | } |
2949 | 0 | rce->isActive = false; |
2950 | | // Allow rateControlStart of next frame only when rateControlEnd of previous frame is over |
2951 | 0 | m_startEndOrder.incr(); |
2952 | 0 | return 0; |
2953 | 0 | } |
2954 | | |
2955 | | /* called to write out the rate control frame stats info in multipass encodes */ |
2956 | | int RateControl::writeRateControlFrameStats(Frame* curFrame, RateControlEntry* rce) |
2957 | 0 | { |
2958 | 0 | FrameData& curEncData = *curFrame->m_encData; |
2959 | 0 | int ncu = (m_param->rc.qgSize == 8) ? m_ncu * 4 : m_ncu; |
2960 | 0 | char cType = rce->sliceType == I_SLICE ? (curFrame->m_lowres.sliceType == X265_TYPE_IDR ? 'I' : 'i') |
2961 | 0 | : rce->sliceType == P_SLICE ? 'P' |
2962 | 0 | : IS_REFERENCED(curFrame) ? 'B' : 'b'; |
2963 | | |
2964 | 0 | if (!curEncData.m_param->bMultiPassOptRPS) |
2965 | 0 | { |
2966 | 0 | if (fprintf(m_statFileOut, |
2967 | 0 | "in:%d out:%d type:%c q:%.2f q-aq:%.2f q-noVbv:%.2f q-Rceq:%.2f tex:%d mv:%d misc:%d icu:%.2f pcu:%.2f scu:%.2f ;\n", |
2968 | 0 | rce->poc, rce->encodeOrder, |
2969 | 0 | cType, curEncData.m_avgQpRc, curEncData.m_avgQpAq, |
2970 | 0 | rce->qpNoVbv, rce->qRceq, |
2971 | 0 | curFrame->m_encData->m_frameStats.coeffBits, |
2972 | 0 | curFrame->m_encData->m_frameStats.mvBits, |
2973 | 0 | curFrame->m_encData->m_frameStats.miscBits, |
2974 | 0 | curFrame->m_encData->m_frameStats.percent8x8Intra * m_ncu, |
2975 | 0 | curFrame->m_encData->m_frameStats.percent8x8Inter * m_ncu, |
2976 | 0 | curFrame->m_encData->m_frameStats.percent8x8Skip * m_ncu) < 0) |
2977 | 0 | goto writeFailure; |
2978 | 0 | } |
2979 | 0 | else |
2980 | 0 | { |
2981 | 0 | RPS* rpsWriter = &curFrame->m_encData->m_slice->m_rps; |
2982 | 0 | int i, num = rpsWriter->numberOfPictures; |
2983 | 0 | char deltaPOC[128]; |
2984 | 0 | char bUsed[40]; |
2985 | 0 | memset(deltaPOC, 0, sizeof(deltaPOC)); |
2986 | 0 | memset(bUsed, 0, sizeof(bUsed)); |
2987 | 0 | sprintf(deltaPOC, "deltapoc:~"); |
2988 | 0 | sprintf(bUsed, "bused:~"); |
2989 | |
|
2990 | 0 | for (i = 0; i < num; i++) |
2991 | 0 | { |
2992 | 0 | sprintf(deltaPOC, "%s%d~", deltaPOC, rpsWriter->deltaPOC[i]); |
2993 | 0 | sprintf(bUsed, "%s%d~", bUsed, rpsWriter->bUsed[i]); |
2994 | 0 | } |
2995 | |
|
2996 | 0 | if (fprintf(m_statFileOut, |
2997 | 0 | "in:%d out:%d type:%c q:%.2f q-aq:%.2f q-noVbv:%.2f q-Rceq:%.2f tex:%d mv:%d misc:%d icu:%.2f pcu:%.2f scu:%.2f nump:%d numnegp:%d numposp:%d %s %s ;\n", |
2998 | 0 | rce->poc, rce->encodeOrder, |
2999 | 0 | cType, curEncData.m_avgQpRc, curEncData.m_avgQpAq, |
3000 | 0 | rce->qpNoVbv, rce->qRceq, |
3001 | 0 | curFrame->m_encData->m_frameStats.coeffBits, |
3002 | 0 | curFrame->m_encData->m_frameStats.mvBits, |
3003 | 0 | curFrame->m_encData->m_frameStats.miscBits, |
3004 | 0 | curFrame->m_encData->m_frameStats.percent8x8Intra * m_ncu, |
3005 | 0 | curFrame->m_encData->m_frameStats.percent8x8Inter * m_ncu, |
3006 | 0 | curFrame->m_encData->m_frameStats.percent8x8Skip * m_ncu, |
3007 | 0 | rpsWriter->numberOfPictures, |
3008 | 0 | rpsWriter->numberOfNegativePictures, |
3009 | 0 | rpsWriter->numberOfPositivePictures, |
3010 | 0 | deltaPOC, bUsed) < 0) |
3011 | 0 | goto writeFailure; |
3012 | 0 | } |
3013 | | /* Don't re-write the data in multi-pass mode. */ |
3014 | 0 | if (m_param->rc.cuTree && IS_REFERENCED(curFrame) && !m_param->rc.bStatRead) |
3015 | 0 | { |
3016 | 0 | uint8_t sliceType = (uint8_t)rce->sliceType; |
3017 | 0 | primitives.fix8Pack(m_cuTreeStats.qpBuffer[0], curFrame->m_lowres.qpCuTreeOffset, ncu); |
3018 | 0 | if (fwrite(&sliceType, 1, 1, m_cutreeStatFileOut) < 1) |
3019 | 0 | goto writeFailure; |
3020 | 0 | if (fwrite(m_cuTreeStats.qpBuffer[0], sizeof(uint16_t), ncu, m_cutreeStatFileOut) < (size_t)ncu) |
3021 | 0 | goto writeFailure; |
3022 | 0 | } |
3023 | 0 | return 0; |
3024 | | |
3025 | 0 | writeFailure: |
3026 | 0 | x265_log(m_param, X265_LOG_ERROR, "RatecontrolEnd: stats file write failure\n"); |
3027 | 0 | return 1; |
3028 | 0 | } |
3029 | | #if defined(_MSC_VER) |
3030 | | #pragma warning(disable: 4996) // POSIX function names are just fine, thank you |
3031 | | #endif |
3032 | | |
3033 | | /* called when the encoder is flushing, and thus the final frame count is |
3034 | | * unambiguously known */ |
3035 | | void RateControl::setFinalFrameCount(int count) |
3036 | 0 | { |
3037 | 0 | m_finalFrameCount = count; |
3038 | | /* unblock waiting threads */ |
3039 | 0 | m_startEndOrder.poke(); |
3040 | 0 | } |
3041 | | |
3042 | | /* called when the encoder is closing, and no more frames will be output. |
3043 | | * all blocked functions must finish so the frame encoder threads can be |
3044 | | * closed */ |
3045 | | void RateControl::terminate() |
3046 | 0 | { |
3047 | 0 | m_bTerminated = true; |
3048 | | /* unblock waiting threads */ |
3049 | 0 | m_startEndOrder.poke(); |
3050 | 0 | } |
3051 | | |
3052 | | void RateControl::destroy() |
3053 | 0 | { |
3054 | 0 | const char *fileName = m_param->rc.statFileName; |
3055 | 0 | if (!fileName) |
3056 | 0 | fileName = s_defaultStatFileName; |
3057 | |
|
3058 | 0 | if (m_statFileOut) |
3059 | 0 | { |
3060 | 0 | fclose(m_statFileOut); |
3061 | 0 | char *tmpFileName = strcatFilename(fileName, ".temp"); |
3062 | 0 | int bError = 1; |
3063 | 0 | if (tmpFileName) |
3064 | 0 | { |
3065 | 0 | x265_unlink(fileName); |
3066 | 0 | bError = x265_rename(tmpFileName, fileName); |
3067 | 0 | } |
3068 | 0 | if (bError) |
3069 | 0 | { |
3070 | 0 | x265_log_file(m_param, X265_LOG_ERROR, "failed to rename output stats file to \"%s\"\n", fileName); |
3071 | 0 | } |
3072 | 0 | X265_FREE(tmpFileName); |
3073 | 0 | } |
3074 | |
|
3075 | 0 | if (m_cutreeStatFileOut) |
3076 | 0 | { |
3077 | 0 | fclose(m_cutreeStatFileOut); |
3078 | 0 | char *tmpFileName = strcatFilename(fileName, ".cutree.temp"); |
3079 | 0 | char *newFileName = strcatFilename(fileName, ".cutree"); |
3080 | 0 | int bError = 1; |
3081 | 0 | if (tmpFileName && newFileName) |
3082 | 0 | { |
3083 | 0 | x265_unlink(newFileName); |
3084 | 0 | bError = x265_rename(tmpFileName, newFileName); |
3085 | 0 | } |
3086 | 0 | if (bError) |
3087 | 0 | { |
3088 | 0 | x265_log_file(m_param, X265_LOG_ERROR, "failed to rename cutree output stats file to \"%s\"\n", newFileName); |
3089 | 0 | } |
3090 | 0 | X265_FREE(tmpFileName); |
3091 | 0 | X265_FREE(newFileName); |
3092 | 0 | } |
3093 | |
|
3094 | 0 | if (m_cutreeStatFileIn) |
3095 | 0 | fclose(m_cutreeStatFileIn); |
3096 | |
|
3097 | 0 | X265_FREE(m_rce2Pass); |
3098 | 0 | X265_FREE(m_encOrder); |
3099 | 0 | for (int i = 0; i < 2; i++) |
3100 | 0 | X265_FREE(m_cuTreeStats.qpBuffer[i]); |
3101 | | |
3102 | 0 | if (m_relativeComplexity) |
3103 | 0 | X265_FREE(m_relativeComplexity); |
3104 | |
|
3105 | 0 | } |
3106 | | |
3107 | | void RateControl::splitdeltaPOC(char deltapoc[], RateControlEntry *rce) |
3108 | 0 | { |
3109 | 0 | int idx = 0, length = 0; |
3110 | 0 | char tmpStr[128]; |
3111 | 0 | char* src = deltapoc; |
3112 | 0 | char* buf = strstr(src, "~"); |
3113 | 0 | while (buf) |
3114 | 0 | { |
3115 | 0 | memset(tmpStr, 0, sizeof(tmpStr)); |
3116 | 0 | length = (int)(buf - src); |
3117 | 0 | if (length != 0) |
3118 | 0 | { |
3119 | 0 | strncpy(tmpStr, src, length); |
3120 | 0 | rce->rpsData.deltaPOC[idx] = atoi(tmpStr); |
3121 | 0 | idx++; |
3122 | 0 | if (idx == rce->rpsData.numberOfPictures) |
3123 | 0 | break; |
3124 | 0 | } |
3125 | 0 | src += (length + 1); |
3126 | 0 | buf = strstr(src, "~"); |
3127 | 0 | } |
3128 | 0 | } |
3129 | | |
3130 | | void RateControl::splitbUsed(char bused[], RateControlEntry *rce) |
3131 | 0 | { |
3132 | 0 | int idx = 0, length = 0; |
3133 | 0 | char tmpStr[128]; |
3134 | 0 | char* src = bused; |
3135 | 0 | char* buf = strstr(src, "~"); |
3136 | 0 | while (buf) |
3137 | 0 | { |
3138 | 0 | memset(tmpStr, 0, sizeof(tmpStr)); |
3139 | 0 | length = (int)(buf - src); |
3140 | 0 | if (length != 0) |
3141 | 0 | { |
3142 | 0 | strncpy(tmpStr, src, length); |
3143 | 0 | rce->rpsData.bUsed[idx] = atoi(tmpStr) > 0; |
3144 | 0 | idx++; |
3145 | 0 | if (idx == rce->rpsData.numberOfPictures) |
3146 | 0 | break; |
3147 | 0 | } |
3148 | 0 | src += (length + 1); |
3149 | 0 | buf = strstr(src, "~"); |
3150 | 0 | } |
3151 | 0 | } |
3152 | | |
3153 | | double RateControl::scenecutAwareQp(Frame* curFrame, double q) |
3154 | 0 | { |
3155 | 0 | uint32_t maxWindowSize = uint32_t((m_param->scenecutWindow / 1000.0) * (m_param->fpsNum / m_param->fpsDenom) + 0.5); |
3156 | 0 | uint32_t windowSize = maxWindowSize / 3; |
3157 | 0 | int lastScenecut = m_top->m_rateControl->m_lastScenecut; |
3158 | 0 | int lastIFrame = m_top->m_rateControl->m_lastScenecutAwareIFrame; |
3159 | 0 | double maxQpDelta = x265_qp2qScale(double(m_param->maxQpDelta)); |
3160 | 0 | double iSliceDelta = x265_qp2qScale(double(I_SLICE_DELTA)); |
3161 | 0 | double sliceTypeDelta = SLICE_TYPE_DELTA * maxQpDelta; |
3162 | 0 | double window2Delta = WINDOW2_DELTA * maxQpDelta; |
3163 | 0 | double window3Delta = WINDOW3_DELTA * maxQpDelta; |
3164 | |
|
3165 | 0 | bool isFrameInsideWindow = curFrame->m_poc > lastScenecut && curFrame->m_poc <= (lastScenecut + int(maxWindowSize)); |
3166 | |
|
3167 | 0 | if (isFrameInsideWindow && IS_X265_TYPE_I(curFrame->m_lowres.sliceType)) |
3168 | 0 | { |
3169 | 0 | m_top->m_rateControl->m_lastScenecutAwareIFrame = curFrame->m_poc; |
3170 | 0 | } |
3171 | 0 | else if (isFrameInsideWindow && (curFrame->m_lowres.sliceType == X265_TYPE_P)) |
3172 | 0 | { |
3173 | 0 | if (!(lastIFrame > lastScenecut && lastIFrame <= (lastScenecut + int(maxWindowSize)) |
3174 | 0 | && curFrame->m_poc > lastIFrame)) |
3175 | 0 | { |
3176 | 0 | q += maxQpDelta - sliceTypeDelta; |
3177 | 0 | if (((curFrame->m_poc) > (lastScenecut + int(windowSize))) && ((curFrame->m_poc) <= (lastScenecut + 2 * int(windowSize)))) |
3178 | 0 | q -= window2Delta; |
3179 | 0 | else if (curFrame->m_poc > lastScenecut + 2 * int(windowSize)) |
3180 | 0 | q -= window3Delta; |
3181 | 0 | } |
3182 | 0 | } |
3183 | 0 | else if (isFrameInsideWindow && IS_X265_TYPE_B(curFrame->m_lowres.sliceType)) |
3184 | 0 | { |
3185 | 0 | if (!(lastIFrame > lastScenecut && lastIFrame <= (lastScenecut + int(maxWindowSize)) |
3186 | 0 | && curFrame->m_poc > lastIFrame)) |
3187 | 0 | { |
3188 | 0 | q += maxQpDelta; |
3189 | 0 | if (curFrame->m_lowres.sliceType == X265_TYPE_B) |
3190 | 0 | q += sliceTypeDelta; |
3191 | 0 | if (((curFrame->m_poc) > (lastScenecut + int(windowSize))) && ((curFrame->m_poc) <= (lastScenecut + 2 * int(windowSize)))) |
3192 | 0 | q -= window2Delta; |
3193 | 0 | else if (curFrame->m_poc > lastScenecut + 2 * int(windowSize)) |
3194 | 0 | q -= window3Delta; |
3195 | 0 | } |
3196 | 0 | } |
3197 | 0 | if (IS_X265_TYPE_I(curFrame->m_lowres.sliceType) && curFrame->m_lowres.bScenecut) |
3198 | 0 | q = q - iSliceDelta; |
3199 | 0 | return q; |
3200 | 0 | } |