source/encoder/ratecontrol.cpp

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