| 1 | /***************************************************************************** |
| 2 | * Copyright (C) 2013 x265 project |
| 3 | * |
| 4 | * Authors: Steve Borho <steve@borho.org> |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License as published by |
| 8 | * the Free Software Foundation; either version 2 of the License, or |
| 9 | * (at your option) any later version. |
| 10 | * |
| 11 | * This program is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 14 | * GNU General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU General Public License |
| 17 | * along with this program; if not, write to the Free Software |
| 18 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA. |
| 19 | * |
| 20 | * This program is also available under a commercial proprietary license. |
| 21 | * For more information, contact us at license @ x265.com. |
| 22 | *****************************************************************************/ |
| 23 | |
| 24 | #ifndef X265_SEARCH_H |
| 25 | #define X265_SEARCH_H |
| 26 | |
| 27 | #include "common.h" |
| 28 | #include "predict.h" |
| 29 | #include "quant.h" |
| 30 | #include "bitcost.h" |
| 31 | #include "yuv.h" |
| 32 | #include "threadpool.h" |
| 33 | |
| 34 | #include "rdcost.h" |
| 35 | #include "entropy.h" |
| 36 | #include "motion.h" |
| 37 | |
| 38 | namespace x265 { |
| 39 | // private namespace |
| 40 | |
| 41 | class Entropy; |
| 42 | struct ThreadLocalData; |
| 43 | |
| 44 | /* All the CABAC contexts that Analysis needs to keep track of at each depth |
| 45 | * and temp buffers for residual, coeff, and recon for use during residual |
| 46 | * quad-tree depth recursion */ |
| 47 | struct RQTData |
| 48 | { |
| 49 | Entropy cur; /* starting context for current CU */ |
| 50 | |
| 51 | /* these are indexed by qtLayer (log2size - 2) so nominally 0=4x4, 1=8x8, 2=16x16, 3=32x32 |
| 52 | * the coeffRQT and reconQtYuv are allocated to the max CU size at every depth. The parts |
| 53 | * which are reconstructed at each depth are valid. At the end, the transform depth table |
| 54 | * is walked and the coeff and recon at the final split depths are collected */ |
| 55 | Entropy rqtRoot; /* residual quad-tree start context */ |
| 56 | Entropy rqtTemp; /* residual quad-tree temp context */ |
| 57 | Entropy rqtTest; /* residual quad-tree test context */ |
| 58 | coeff_t* coeffRQT[3]; /* coeff storage for entire CTU for each RQT layer */ |
| 59 | Yuv reconQtYuv; /* recon storage for entire CTU for each RQT layer (intra) */ |
| 60 | ShortYuv resiQtYuv; /* residual storage for entire CTU for each RQT layer (inter) */ |
| 61 | |
| 62 | /* per-depth temp buffers for inter prediction */ |
| 63 | ShortYuv tmpResiYuv; |
| 64 | Yuv tmpPredYuv; |
| 65 | Yuv bidirPredYuv[2]; |
| 66 | }; |
| 67 | |
| 68 | struct MotionData |
| 69 | { |
| 70 | MV mv; |
| 71 | MV mvp; |
| 72 | int mvpIdx; |
| 73 | int ref; |
| 74 | uint32_t cost; |
| 75 | int bits; |
| 76 | bool costZero; |
| 77 | }; |
| 78 | |
| 79 | struct Mode |
| 80 | { |
| 81 | CUData cu; |
| 82 | const Yuv* fencYuv; |
| 83 | Yuv predYuv; |
| 84 | Yuv reconYuv; |
| 85 | Entropy contexts; |
| 86 | |
| 87 | enum { MAX_INTER_PARTS = 2 }; |
| 88 | |
| 89 | MotionData bestME[MAX_INTER_PARTS][2]; |
| 90 | MV amvpCand[2][MAX_NUM_REF][AMVP_NUM_CANDS]; |
| 91 | |
| 92 | uint64_t rdCost; // sum of partition (psy) RD costs (sse(fenc, recon) + lambda2 * bits) |
| 93 | uint64_t sa8dCost; // sum of partition sa8d distortion costs (sa8d(fenc, pred) + lambda * bits) |
| 94 | uint32_t sa8dBits; // signal bits used in sa8dCost calculation |
| 95 | uint32_t psyEnergy; // sum of partition psycho-visual energy difference |
| 96 | uint32_t distortion; // sum of partition SSE distortion |
| 97 | uint32_t totalBits; // sum of partition bits (mv + coeff) |
| 98 | uint32_t mvBits; // Mv bits + Ref + block type (or intra mode) |
| 99 | uint32_t coeffBits; // Texture bits (DCT Coeffs) |
| 100 | |
| 101 | void initCosts() |
| 102 | { |
| 103 | rdCost = 0; |
| 104 | sa8dCost = 0; |
| 105 | sa8dBits = 0; |
| 106 | psyEnergy = 0; |
| 107 | distortion = 0; |
| 108 | totalBits = 0; |
| 109 | mvBits = 0; |
| 110 | coeffBits = 0; |
| 111 | } |
| 112 | |
| 113 | void addSubCosts(const Mode& subMode) |
| 114 | { |
| 115 | rdCost += subMode.rdCost; |
| 116 | sa8dCost += subMode.sa8dCost; |
| 117 | sa8dBits += subMode.sa8dBits; |
| 118 | psyEnergy += subMode.psyEnergy; |
| 119 | distortion += subMode.distortion; |
| 120 | totalBits += subMode.totalBits; |
| 121 | mvBits += subMode.mvBits; |
| 122 | coeffBits += subMode.coeffBits; |
| 123 | } |
| 124 | }; |
| 125 | |
| 126 | inline int getTUBits(int idx, int numIdx) |
| 127 | { |
| 128 | return idx + (idx < numIdx - 1); |
| 129 | } |
| 130 | |
| 131 | class Search : public JobProvider, public Predict |
| 132 | { |
| 133 | public: |
| 134 | |
| 135 | static const pixel zeroPixel[MAX_CU_SIZE]; |
| 136 | static const int16_t zeroShort[MAX_CU_SIZE]; |
| 137 | |
| 138 | MotionEstimate m_me; |
| 139 | Quant m_quant; |
| 140 | RDCost m_rdCost; |
| 141 | const x265_param* m_param; |
| 142 | Frame* m_frame; |
| 143 | const Slice* m_slice; |
| 144 | |
| 145 | Entropy m_entropyCoder; |
| 146 | RQTData m_rqt[NUM_FULL_DEPTH]; |
| 147 | |
| 148 | uint8_t* m_qtTempCbf[3]; |
| 149 | uint8_t* m_qtTempTransformSkipFlag[3]; |
| 150 | |
| 151 | bool m_bFrameParallel; |
| 152 | bool m_bEnableRDOQ; |
| 153 | uint32_t m_numLayers; |
| 154 | uint32_t m_refLagPixels; |
| 155 | |
| 156 | Search(); |
| 157 | ~Search(); |
| 158 | |
| 159 | bool initSearch(const x265_param& param, ScalingList& scalingList); |
| 160 | void setQP(const Slice& slice, int qp); |
| 161 | |
| 162 | // mark temp RD entropy contexts as uninitialized; useful for finding loads without stores |
| 163 | void invalidateContexts(int fromDepth); |
| 164 | |
| 165 | // full RD search of intra modes. if sharedModes is not NULL, it directly uses them |
| 166 | void checkIntra(Mode& intraMode, const CUGeom& cuGeom, PartSize partSize, uint8_t* sharedModes); |
| 167 | |
| 168 | // select best intra mode using only sa8d costs, cannot measure NxN intra |
| 169 | void checkIntraInInter(Mode& intraMode, const CUGeom& cuGeom); |
| 170 | // encode luma mode selected by checkIntraInInter, then pick and encode a chroma mode |
| 171 | void encodeIntraInInter(Mode& intraMode, const CUGeom& cuGeom); |
| 172 | |
| 173 | // estimation inter prediction (non-skip) |
| 174 | bool predInterSearch(Mode& interMode, const CUGeom& cuGeom, bool bMergeOnly, bool bChroma); |
| 175 | |
| 176 | // encode residual and compute rd-cost for inter mode |
| 177 | void encodeResAndCalcRdInterCU(Mode& interMode, const CUGeom& cuGeom); |
| 178 | void encodeResAndCalcRdSkipCU(Mode& interMode); |
| 179 | |
| 180 | // encode residual without rd-cost |
| 181 | void residualTransformQuantInter(Mode& mode, const CUGeom& cuGeom, uint32_t absPartIdx, uint32_t depth, const uint32_t depthRange[2]); |
| 182 | void residualTransformQuantIntra(Mode& mode, const CUGeom& cuGeom, uint32_t tuDepth, uint32_t absPartIdx, const uint32_t depthRange[2]); |
| 183 | void residualQTIntraChroma(Mode& mode, const CUGeom& cuGeom, uint32_t tuDepth, uint32_t absPartIdx); |
| 184 | |
| 185 | // pick be chroma mode from available using just sa8d costs |
| 186 | void getBestIntraModeChroma(Mode& intraMode, const CUGeom& cuGeom); |
| 187 | |
| 188 | protected: |
| 189 | |
| 190 | /* motion estimation distribution */ |
| 191 | ThreadLocalData* m_tld; |
| 192 | Mode* m_curInterMode; |
| 193 | const CUGeom* m_curGeom; |
| 194 | int m_curPart; |
| 195 | uint32_t m_listSelBits[3]; |
| 196 | int m_totalNumME; |
| 197 | volatile int m_numAcquiredME; |
| 198 | volatile int m_numCompletedME; |
| 199 | Event m_meCompletionEvent; |
| 200 | Lock m_meLock; |
| 201 | bool m_bJobsQueued; |
| 202 | void singleMotionEstimation(Search& master, Mode& interMode, const CUGeom& cuGeom, int part, int list, int ref); |
| 203 | |
| 204 | void saveResidualQTData(CUData& cu, ShortYuv& resiYuv, uint32_t absPartIdx, uint32_t depth); |
| 205 | |
| 206 | // RDO search of luma intra modes; result is fully encoded luma. luma distortion is returned |
| 207 | uint32_t estIntraPredQT(Mode &intraMode, const CUGeom& cuGeom, const uint32_t depthRange[2], uint8_t* sharedModes); |
| 208 | |
| 209 | // RDO select best chroma mode from luma; result is fully encode chroma. chroma distortion is returned |
| 210 | uint32_t estIntraPredChromaQT(Mode &intraMode, const CUGeom& cuGeom); |
| 211 | |
| 212 | void codeSubdivCbfQTChroma(const CUData& cu, uint32_t tuDepth, uint32_t absPartIdx); |
| 213 | void codeInterSubdivCbfQT(CUData& cu, uint32_t absPartIdx, const uint32_t depth, const uint32_t depthRange[2]); |
| 214 | void codeCoeffQTChroma(const CUData& cu, uint32_t tuDepth, uint32_t absPartIdx, TextType ttype); |
| 215 | |
| 216 | struct Cost |
| 217 | { |
| 218 | uint64_t rdcost; |
| 219 | uint32_t bits; |
| 220 | uint32_t distortion; |
| 221 | uint32_t energy; |
| 222 | Cost() { rdcost = 0; bits = 0; distortion = 0; energy = 0; } |
| 223 | }; |
| 224 | |
| 225 | uint64_t estimateNullCbfCost(uint32_t &dist, uint32_t &psyEnergy, uint32_t tuDepth, TextType compId); |
| 226 | void estimateResidualQT(Mode& mode, const CUGeom& cuGeom, uint32_t absPartIdx, uint32_t depth, ShortYuv& resiYuv, Cost& costs, const uint32_t depthRange[2]); |
| 227 | |
| 228 | // estimate bit cost of residual QT |
| 229 | void encodeResidualQT(CUData& cu, uint32_t absPartIdx, uint32_t depth, TextType ttype, const uint32_t depthRange[2]); |
| 230 | |
| 231 | // generate prediction, generate residual and recon. if bAllowSplit, find optimal RQT splits |
| 232 | void codeIntraLumaQT(Mode& mode, const CUGeom& cuGeom, uint32_t tuDepth, uint32_t absPartIdx, bool bAllowSplit, Cost& costs, const uint32_t depthRange[2]); |
| 233 | void codeIntraLumaTSkip(Mode& mode, const CUGeom& cuGeom, uint32_t tuDepth, uint32_t absPartIdx, Cost& costs); |
| 234 | void extractIntraResultQT(CUData& cu, Yuv& reconYuv, uint32_t tuDepth, uint32_t absPartIdx); |
| 235 | |
| 236 | // generate chroma prediction, generate residual and recon |
| 237 | uint32_t codeIntraChromaQt(Mode& mode, const CUGeom& cuGeom, uint32_t tuDepth, uint32_t absPartIdx, uint32_t& psyEnergy); |
| 238 | uint32_t codeIntraChromaTSkip(Mode& mode, const CUGeom& cuGeom, uint32_t tuDepth, uint32_t tuDepthC, uint32_t absPartIdx, uint32_t& psyEnergy); |
| 239 | void extractIntraResultChromaQT(CUData& cu, Yuv& reconYuv, uint32_t absPartIdx, uint32_t tuDepth); |
| 240 | |
| 241 | // reshuffle CBF flags after coding a pair of 4:2:2 chroma blocks |
| 242 | void offsetSubTUCBFs(CUData& cu, TextType ttype, uint32_t tuDepth, uint32_t absPartIdx); |
| 243 | |
| 244 | struct MergeData |
| 245 | { |
| 246 | /* merge candidate data, cached between calls to mergeEstimation */ |
| 247 | MVField mvFieldNeighbours[MRG_MAX_NUM_CANDS][2]; |
| 248 | uint8_t interDirNeighbours[MRG_MAX_NUM_CANDS]; |
| 249 | uint32_t maxNumMergeCand; |
| 250 | |
| 251 | /* data updated for each partition */ |
| 252 | uint32_t absPartIdx; |
| 253 | int width; |
| 254 | int height; |
| 255 | |
| 256 | /* outputs */ |
| 257 | MVField mvField[2]; |
| 258 | uint32_t interDir; |
| 259 | uint32_t index; |
| 260 | uint32_t bits; |
| 261 | }; |
| 262 | |
| 263 | /* inter/ME helper functions */ |
| 264 | void checkBestMVP(MV* amvpCand, MV cMv, MV& mvPred, int& mvpIdx, uint32_t& outBits, uint32_t& outCost) const; |
| 265 | void setSearchRange(const CUData& cu, MV mvp, int merange, MV& mvmin, MV& mvmax) const; |
| 266 | uint32_t mergeEstimation(CUData& cu, const CUGeom& cuGeom, int partIdx, MergeData& m); |
| 267 | static void getBlkBits(PartSize cuMode, bool bPSlice, int partIdx, uint32_t lastMode, uint32_t blockBit[3]); |
| 268 | |
| 269 | /* intra helper functions */ |
| 270 | enum { MAX_RD_INTRA_MODES = 16 }; |
| 271 | static void updateCandList(uint32_t mode, uint64_t cost, int maxCandCount, uint32_t* candModeList, uint64_t* candCostList); |
| 272 | |
| 273 | // get most probable luma modes for CU part, and bit cost of all non mpm modes |
| 274 | uint32_t getIntraRemModeBits(CUData & cu, uint32_t absPartIdx, uint32_t preds[3], uint64_t& mpms) const; |
| 275 | |
| 276 | void updateModeCost(Mode& m) const { m.rdCost = m_rdCost.m_psyRd ? m_rdCost.calcPsyRdCost(m.distortion, m.totalBits, m.psyEnergy) : m_rdCost.calcRdCost(m.distortion, m.totalBits); } |
| 277 | }; |
| 278 | } |
| 279 | |
| 280 | #endif // ifndef X265_SEARCH_H |