本文详细介绍了HEVC编码器x265中的率失真优化量化(RDOQ)过程，包括变换量化、逆变换量化、RDOQ量化等核心模块的实现细节和技术原理。
注：问号以及未注释部分 会在x265-1.8版本内更新
/*****************************************************************************
* Copyright (C) 2015 x265 project
*
* Authors: Steve Borho <steve@borho.org>
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
*
* This program is also available under a commercial proprietary license.
* For more information, contact us at license @ x265.com.
*****************************************************************************/
#include "common.h"
#include "primitives.h"
#include "quant.h"
#include "framedata.h"
#include "entropy.h"
#include "yuv.h"
#include "cudata.h"
#include "contexts.h"
using namespace x265;
#define SIGN(x,y) ((x^(y >> 31))-(y >> 31))
namespace {
struct coeffGroupRDStats
{
int nnzBeforePos0; /* indicates coeff other than pos 0 are coded */
int64_t codedLevelAndDist; /* distortion and level cost of coded coefficients */
int64_t uncodedDist; /* uncoded distortion cost of coded coefficients */
int64_t sigCost; /* cost of signaling significant coeff bitmap */
int64_t sigCost0; /* cost of signaling sig coeff bit of coeff 0 */
};
inline int fastMin(int x, int y)
{
return y + ((x - y) & ((x - y) >> (sizeof(int) * CHAR_BIT - 1))); // min(x, y)
}
inline int getICRate(uint32_t absLevel, int32_t diffLevel, const int* greaterOneBits, const int* levelAbsBits, const uint32_t absGoRice, const uint32_t maxVlc, uint32_t c1c2Idx)
{
X265_CHECK(c1c2Idx <= 3, "c1c2Idx check failure\n");
X265_CHECK(absGoRice <= 4, "absGoRice check failure\n");
if (!absLevel)
{
X265_CHECK(diffLevel < 0, "diffLevel check failure\n");
return 0;
}
int rate = 0;
if (diffLevel < 0)
{
X265_CHECK(absLevel <= 2, "absLevel check failure\n");
rate += greaterOneBits[(absLevel == 2)];
if (absLevel == 2)
rate += levelAbsBits[0];
}
else
{
uint32_t symbol = diffLevel;
bool expGolomb = (symbol > maxVlc);
if (expGolomb)
{
absLevel = symbol - maxVlc;
// NOTE: mapping to x86 hardware instruction BSR
unsigned long size;
CLZ(size, absLevel);
int egs = size * 2 + 1;
rate += egs << 15;
// NOTE: in here, expGolomb=true means (symbol >= maxVlc + 1)
X265_CHECK(fastMin(symbol, (maxVlc + 1)) == (int)maxVlc + 1, "min check failure\n");
symbol = maxVlc + 1;
}
uint32_t prefLen = (symbol >> absGoRice) + 1;
uint32_t numBins = fastMin(prefLen + absGoRice, 8 /* g_goRicePrefixLen[absGoRice] + absGoRice */);
rate += numBins << 15;
if (c1c2Idx & 1)
rate += greaterOneBits[1];
if (c1c2Idx == 3)
rate += levelAbsBits[1];
}
return rate;
}
#if CHECKED_BUILD || _DEBUG
inline int getICRateNegDiff(uint32_t absLevel, const int* greaterOneBits, const int* levelAbsBits)
{
X265_CHECK(absLevel <= 2, "absLevel check failure\n");
int rate;
if (absLevel == 0)
rate = 0;
else if (absLevel == 2)
rate = greaterOneBits[1] + levelAbsBits[0];
else
rate = greaterOneBits[0];
return rate;
}
#endif
inline int getICRateLessVlc(uint32_t absLevel, int32_t diffLevel, const uint32_t absGoRice)
{
X265_CHECK(absGoRice <= 4, "absGoRice check failure\n");
if (!absLevel)
{
X265_CHECK(diffLevel < 0, "diffLevel check failure\n");
return 0;
}
int rate;
uint32_t symbol = diffLevel;
uint32_t prefLen = (symbol >> absGoRice) + 1;
uint32_t numBins = fastMin(prefLen + absGoRice, 8 /* g_goRicePrefixLen[absGoRice] + absGoRice */);
rate = numBins << 15;
return rate;
}
/* Calculates the cost for specific absolute transform level */
inline uint32_t getICRateCost(uint32_t absLevel, int32_t diffLevel, const int* greaterOneBits, const int* levelAbsBits, uint32_t absGoRice, uint32_t c1c2Idx)
{
X265_CHECK(absLevel, "absLevel should not be zero\n");
if (diffLevel < 0)
{
X265_CHECK((absLevel == 1) || (absLevel == 2), "absLevel range check failure\n");
uint32_t rate = greaterOneBits[(absLevel == 2)];
if (absLevel == 2)
rate += levelAbsBits[0];
return rate;
}
else
{
uint32_t rate;
uint32_t symbol = diffLevel;
if ((symbol >> absGoRice) < COEF_REMAIN_BIN_REDUCTION)
{
uint32_t length = symbol >> absGoRice;
rate = (length + 1 + absGoRice) << 15;
}
else
{
uint32_t length = 0;
symbol = (symbol >> absGoRice) - COEF_REMAIN_BIN_REDUCTION;
if (symbol)
{
unsigned long idx;
CLZ(idx, symbol + 1);
length = idx;
}
rate = (COEF_REMAIN_BIN_REDUCTION + length + absGoRice + 1 + length) << 15;
}
if (c1c2Idx & 1)
rate += greaterOneBits[1];
if (c1c2Idx == 3)
rate += levelAbsBits[1];
return rate;
}
}
}
Quant::Quant()
{
m_resiDctCoeff = NULL;
m_fencDctCoeff = NULL;
m_fencShortBuf = NULL;
m_frameNr = NULL;
m_nr = NULL;
}
bool Quant::init(int rdoqLevel, double psyScale, const ScalingList& scalingList, Entropy& entropy)
{
m_entropyCoder = &entropy; // 初始化熵编码器
m_rdoqLevel = rdoqLevel; // 初始化rdoq级别
m_psyRdoqScale = (int32_t)(psyScale * 256.0); // 针对RDOQ的心理视觉优化
X265_CHECK((psyScale * 256.0) < (double)MAX_INT, "psyScale value too large\n");
m_scalingList = &scalingList;
m_resiDctCoeff = X265_MALLOC(int16_t, MAX_TR_SIZE * MAX_TR_SIZE * 2);
m_fencDctCoeff = m_resiDctCoeff + (MAX_TR_SIZE * MAX_TR_SIZE);
m_fencShortBuf = X265_MALLOC(int16_t, MAX_TR_SIZE * MAX_TR_SIZE);
m_tqBypass = false;
return m_resiDctCoeff && m_fencShortBuf;
}
bool Quant::allocNoiseReduction(const x265_param& param)
{
m_frameNr = X265_MALLOC(NoiseReduction, param.frameNumThreads);
if (m_frameNr)
memset(m_frameNr, 0, sizeof(NoiseReduction) * param.frameNumThreads);
else
return false;
return true;
}
Quant::~Quant()
{
X265_FREE(m_frameNr);
X265_FREE(m_resiDctCoeff);
X265_FREE(m_fencShortBuf);
}
void Quant::setQPforQuant(const CUData& ctu, int qp)
{
m_tqBypass = !!ctu.m_tqBypass[0];
if (m_tqBypass)
return;
m_nr = m_frameNr ? &m_frameNr[ctu.m_encData->m_frameEncoderID] : NULL;
m_qpParam[TEXT_LUMA].setQpParam(qp + QP_BD_OFFSET);
setChromaQP(qp + ctu.m_slice->m_pps->chromaQpOffset[0], TEXT_CHROMA_U, ctu.m_chromaFormat);
setChromaQP(qp + ctu.m_slice->m_pps->chromaQpOffset[1], TEXT_CHROMA_V, ctu.m_chromaFormat);
}
void Quant::setChromaQP(int qpin, TextType ttype, int chFmt)
{
int qp = x265_clip3(-QP_BD_OFFSET, 57, qpin);
if (qp >= 30)
{
if (chFmt == X265_CSP_I420)
qp = g_chromaScale[qp];
else
qp = X265_MIN(qp, QP_MAX_SPEC);
}
m_qpParam[ttype].setQpParam(qp + QP_BD_OFFSET);
}
/* To minimize the distortion only. No rate is considered */
uint32_t Quant::signBitHidingHDQ(int16_t* coeff, int32_t* deltaU, uint32_t numSig, const TUEntropyCodingParameters &codeParams)
{
const uint32_t log2TrSizeCG = codeParams.log2TrSizeCG;
const uint16_t* scan = codeParams.scan;
bool lastCG = true;
for (int cg = (1 << (log2TrSizeCG * 2)) - 1; cg >= 0; cg--)
{
int cgStartPos = cg << LOG2_SCAN_SET_SIZE;
int n;
for (n = SCAN_SET_SIZE - 1; n >= 0; --n)
if (coeff[scan[n + cgStartPos]])
break;
if (n < 0)
continue;
int lastNZPosInCG = n;
for (n = 0;; n++)
if (coeff[scan[n + cgStartPos]])
break;
int firstNZPosInCG = n;
if (lastNZPosInCG - firstNZPosInCG >= SBH_THRESHOLD)
{
uint32_t signbit = coeff[scan[cgStartPos + firstNZPosInCG]] > 0 ? 0 : 1;
uint32_t absSum = 0;
for (n = firstNZPosInCG; n <= lastNZPosInCG; n++)
absSum += coeff[scan[n + cgStartPos]];
if (signbit != (absSum & 0x1)) // compare signbit with sum_parity
{
int minCostInc = MAX_INT, minPos = -1, curCost = MAX_INT;
int16_t finalChange = 0, curChange = 0;
for (n = (lastCG ? lastNZPosInCG : SCAN_SET_SIZE - 1); n >= 0; --n)
{
uint32_t blkPos = scan[n + cgStartPos];
if (coeff[blkPos])
{
if (deltaU[blkPos] > 0)
{
curCost = -deltaU[blkPos];
curChange = 1;
}
else
{
if (n == firstNZPosInCG && abs(coeff[blkPos]) == 1)
curCost = MAX_INT;
else
{
curCost = deltaU[blkPos];
curChange = -1;
}
}
}
else
{
if (n < firstNZPosInCG)
{
uint32_t thisSignBit = m_resiDctCoeff[blkPos] >= 0 ? 0 : 1;
if (thisSignBit != signbit)
curCost = MAX_INT;
else
{
curCost = -deltaU[blkPos];
curChange = 1;
}
}
else
{
curCost = -deltaU[blkPos];
curChange = 1;
}
}
if (curCost < minCostInc)
{
minCostInc = curCost;
finalChange = curChange;
minPos = blkPos;
}
}
/* do not allow change to violate coeff clamp */
if (coeff[minPos] == 32767 || coeff[minPos] == -32768)
finalChange = -1;
if (!coeff[minPos])
numSig++;
else if (finalChange == -1 && abs(coeff[minPos]) == 1)
numSig--;
if (m_resiDctCoeff[minPos] >= 0)
coeff[minPos] += finalChange;
else
coeff[minPos] -= finalChange;
}
}
lastCG = false;
}
return numSig;
}
/** 函数功能 ： 对残差块进行变换、量化
* \参数 cu ：CUData对象
* \参数 fenc ：原始图像
* \参数 fencStride ：原始图像块的步长
* \参数 residual ：残差数据
* \参数 resiStride ：残差数据的步长
* \参数 coeff ：存储残差经过变换、量化后的系数
* \参数 log2TrSize ：TU尺寸
* \参数 ttype ：数据分量类型（亮度/色度）
* \参数 absPartIdx ：CU地址
* \参数 useTransformSkip ：是否使用变换跳过模式
* \返回 ：量化后非零系数的个数
**/
uint32_t Quant::transformNxN(const CUData& cu, const pixel* fenc, uint32_t fencStride, const int16_t* residual, uint32_t resiStride,
coeff_t* coeff, uint32_t log2TrSize, TextType ttype, uint32_t absPartIdx, bool useTransformSkip)
{
const uint32_t sizeIdx = log2TrSize - 2;
if (m_tqBypass) // 如果使用 变换/量化的bypass模式，即跳过变换/量化，则直接将残差块拷贝到变换系数块
{
X265_CHECK(log2TrSize >= 2 && log2TrSize <= 5, "Block size mistake!\n");
return primitives.cu[sizeIdx].copy_cnt(coeff, residual, resiStride); // 拷贝残差块到变换系数块coeff，返回非零系数的个数
}
bool isLuma = ttype == TEXT_LUMA;
bool usePsy = m_psyRdoqScale && isLuma && !useTransformSkip;
int transformShift = MAX_TR_DYNAMIC_RANGE - X265_DEPTH - log2TrSize; // Represents scaling through forward transform
X265_CHECK((cu.m_slice->m_sps->quadtreeTULog2MaxSize >= log2TrSize), "transform size too large\n");
if (useTransformSkip) // 如果应用"跳过变换"模式，则只需将残差进行相应的移位，无需进行其他操作
{
#if X265_DEPTH <= 10
X265_CHECK(transformShift >= 0, "invalid transformShift\n");
primitives.cu[sizeIdx].cpy2Dto1D_shl(m_resiDctCoeff, residual, resiStride, transformShift); // 将残差数据进行左移操作
#else
if (transformShift >= 0)
primitives.cu[sizeIdx].cpy2Dto1D_shl(m_resiDctCoeff, residual, resiStride, transformShift);
else
primitives.cu[sizeIdx].cpy2Dto1D_shr(m_resiDctCoeff, residual, resiStride, -transformShift);
#endif
}
else // 进行常规变换
{
bool isIntra = cu.isIntra(absPartIdx);
if (!sizeIdx && isLuma && isIntra) // 如果变换块是4x4(sizeIdx=0)，且是亮度块、intra预测模式，则使用4x4的dst变换
primitives.dst4x4(residual, m_resiDctCoeff, resiStride);
else // 否则使用dct变换
primitives.cu[sizeIdx].dct(residual, m_resiDctCoeff, resiStride); // 对残差做DCT变换，得到的结果存储在m_resiDctCoeff中
/* NOTE: if RDOQ is disabled globally, psy-rdoq is also disabled, so
* there is no risk of performing this DCT unnecessarily */
if (usePsy)
{
int trSize = 1 << log2TrSize;
/* perform DCT on source pixels for psy-rdoq */
primitives.cu[sizeIdx].copy_ps(m_fencShortBuf, trSize, fenc, fencStride);
primitives.cu[sizeIdx].dct(m_fencShortBuf, m_fencDctCoeff, trSize);
}
if (m_nr)
{
/* denoise is not applied to intra residual, so DST can be ignored */
int cat = sizeIdx + 4 * !isLuma + 8 * !isIntra;
int numCoeff = 1 << (log2TrSize * 2);
primitives.denoiseDct(m_resiDctCoeff, m_nr->residualSum[cat], m_nr->offsetDenoise[cat], numCoeff);
m_nr->count[cat]++;
}
}
if (m_rdoqLevel) // 如果RDOQ的级别大于0，才进行RDOQ量化，否则使用常规（均匀）量化
return rdoQuant(cu, coeff, log2TrSize, ttype, absPartIdx, usePsy);
else // 常规量化（均匀量化或非均匀量化）
{
int deltaU[32 * 32]; // 用于存储量化误差矩阵，在常规量化中，deltaU只用于进行符号位隐藏的操作
int scalingListType = (cu.isIntra(absPartIdx) ? 0 : 3) + ttype; // 根据预测模式和亮度/色度分量得到 前向量化表的类型，用于选择不同的前向量化表
int rem = m_qpParam[ttype].rem; // 得到Qp的余数部分，实际上 rem = Qp%6
int per = m_qpParam[ttype].per; // 得到Qp的倍数部分，实际上 per = Qp/6
const int32_t* quantCoeff = m_scalingList->m_quantCoef[log2TrSize - 2][scalingListType][rem]; // 根据TU的尺寸、前向量化类型和Qp余数部分，选择对应的量化表
int qbits = QUANT_SHIFT + per + transformShift; // 量化右移的位数，由3部分组成：1.量化带来的位数增加 2.Qp/6部分带来的位数增加 3.前向变换所带来的位数增加
int add = (cu.m_slice->m_sliceType == I_SLICE ? 171 : 85) << (qbits - 9); // 量化后右移可能会带来低位上的损失，这里对右移可能带来的损失进行补偿，HEVC规定I_SLICE补偿1/3，其他类型SLICE补偿1/6
// 结合量化公式，I_SLICE中的add实际相当于： add >> qbits = (171 << (qbits-9))>>qbits = 171>>9 = 171/512 = 1/3
// 非I_SLICE中add实际相当于： add >> qbits = (85 << (qbits-9))>>qbits = 85>>9 = 85/512 = 1/6
int numCoeff = 1 << (log2TrSize * 2); // 当前变换块中包含得系数个数
uint32_t numSig = primitives.quant(m_resiDctCoeff, quantCoeff, deltaU, coeff, qbits, add, numCoeff); // 进行常规量化，参看C版本函数 quant_c，返回值为量化后非零系数的个数
if (numSig >= 2 && cu.m_slice->m_pps->bSignHideEnabled) // 假如非零系数的个数大于等于2，并且使能符号位隐藏，则进行符号位隐藏的操作
{
TUEntropyCodingParameters codeParams;
cu.getTUEntropyCodingParameters(codeParams, absPartIdx, log2TrSize, isLuma);
return signBitHidingHDQ(coeff, deltaU, numSig, codeParams);
}
else
return numSig;
}
}
void Quant::invtransformNxN(int16_t* residual, uint32_t resiStride, const coeff_t* coeff,
uint32_t log2TrSize, TextType ttype, bool bIntra, bool useTransformSkip, uint32_t numSig)
{
const uint32_t sizeIdx = log2TrSize - 2;
if (m_tqBypass)
{
primitives.cu[sizeIdx].cpy1Dto2D_shl(residual, coeff, resiStride, 0);
return;
}
// Values need to pass as input parameter in dequant
int rem = m_qpParam[ttype].rem;
int per = m_qpParam[ttype].per;
int transformShift = MAX_TR_DYNAMIC_RANGE - X265_DEPTH - log2TrSize;
int shift = QUANT_IQUANT_SHIFT - QUANT_SHIFT - transformShift;
int numCoeff = 1 << (log2TrSize * 2);
if (m_scalingList->m_bEnabled)
{
int scalingListType = (bIntra ? 0 : 3) + ttype;
const int32_t* dequantCoef = m_scalingList->m_dequantCoef[sizeIdx][scalingListType][rem];
primitives.dequant_scaling(coeff, dequantCoef, m_resiDctCoeff, numCoeff, per, shift);
}
else
{
int scale = m_scalingList->s_invQuantScales[rem] << per;
primitives.dequant_normal(coeff, m_resiDctCoeff, numCoeff, scale, shift);
}
if (useTransformSkip)
{
#if X265_DEPTH <= 10
X265_CHECK(transformShift > 0, "invalid transformShift\n");
primitives.cu[sizeIdx].cpy1Dto2D_shr(residual, m_resiDctCoeff, resiStride, transformShift);
#else
if (transformShift > 0)
primitives.cu[sizeIdx].cpy1Dto2D_shr(residual, m_resiDctCoeff, resiStride, transformShift);
else
primitives.cu[sizeIdx].cpy1Dto2D_shl(residual, m_resiDctCoeff, resiStride, -transformShift);
#endif
}
else
{
int useDST = !sizeIdx && ttype == TEXT_LUMA && bIntra;
X265_CHECK((int)numSig == primitives.cu[log2TrSize - 2].count_nonzero(coeff), "numSig differ\n");
// DC only
if (numSig == 1 && coeff[0] != 0 && !useDST)
{
const int shift_1st = 7 - 6;
const int add_1st = 1 << (shift_1st - 1);
const int shift_2nd = 12 - (X265_DEPTH - 8) - 3;
const int add_2nd = 1 << (shift_2nd - 1);
int dc_val = (((m_resiDctCoeff[0] * (64 >> 6) + add_1st) >> shift_1st) * (64 >> 3) + add_2nd) >> shift_2nd;
primitives.cu[sizeIdx].blockfill_s(residual, resiStride, (int16_t)dc_val);
return;
}
if (useDST)
primitives.idst4x4(m_resiDctCoeff, residual, resiStride);
else
primitives.cu[sizeIdx].idct(m_resiDctCoeff, residual, resiStride);
}
}
/* Rate distortion optimized quantization for entropy coding engines using
* probability models like CABAC */
/** 函数功能 ： 使用RDO（率失真优化）技术对变换后的系数进行量化
** RDOQ主要可以分成三步：
** step1. 对每个系数单独做RDO优化，找到率失真意义上的最优量化值
** step2. 对每一个系数组（Coefficient Group，下面都缩写为CG）进行优化，试图将整个CG都设置为0
** step3. 找到最优的最后一个非零系数的位置，尝试从最后一个非零位置开始将量化后的系数设置为0
** 调用范围 ：只在Quant::transformNxN函数中被调用
* \参数 cu ：CUData对象
* \参数 dstCoeff ：进行RDOQ量化后的系数（变换后的系数存储在m_resiDctCoeff中）
* \参数 log2TrSize ：TU尺寸
* \参数 ttype ：数据分量类型（亮度/色度）
* \参数 absPartIdx ：CU地址
* \参数 usePsy ：是否使用心理视觉量化
* \返回 ：非零系数的个数
**/
uint32_t Quant::rdoQuant(const CUData& cu, int16_t* dstCoeff, uint32_t log2TrSize, TextType ttype, uint32_t absPartIdx, bool usePsy)
{
int transformShift = MAX_TR_DYNAMIC_RANGE - X265_DEPTH - log2TrSize; // 前变换的需要的右移位数，需要在量化中完成 /* Represents scaling through forward transform */
int scalingListType = (cu.isIntra(absPartIdx) ? 0 : 3) + ttype; // 根据预测类型(Intra/Inter)和当前分量(Y/U/V)判断list类型
const uint32_t usePsyMask = usePsy ? -1 : 0; // 是否使用心理视觉量化
X265_CHECK(scalingListType < 6, "scaling list type out of range\n"); //scalingListType 最大为6
int rem = m_qpParam[ttype].rem; // 得到Qp的余数部分，=Qp%6
int per = m_qpParam[ttype].per; // 得到Qp的整数部分，=Qp/6
int qbits = QUANT_SHIFT + per + transformShift; // 常规量化中需要右移的位数 /* Right shift of non-RDOQ quantizer level = (coeff*Q + offset)>>q_bits */
int add = (1 << (qbits - 1)); // 常规量化中右移前需要补偿的加数
const int32_t* qCoef = m_scalingList->m_quantCoef[log2TrSize - 2][scalingListType][rem]; // 得到常规量化使用的量化乘数
int numCoeff = 1 << (log2TrSize * 2); // 当前TU中系数的个数
uint32_t numSig = primitives.nquant(m_resiDctCoeff, qCoef, dstCoeff, qbits, add, numCoeff); // 对变换系数进行常规量化，参考C语言版本的函数 nquant_c
X265_CHECK((int)numSig == primitives.cu[log2TrSize - 2].count_nonzero(dstCoeff), "numSig differ\n"); // 再次统计量化后的非零系数个数，并判断与常规量化的结果是否一致
if (!numSig) // 如果常规量化后系数为全零，则跳过RDOQ过程（这是RDOQ提前终止的快速算法）
return 0;
uint32_t trSize = 1 << log2TrSize; // 得到TU大小
int64_t lambda2 = m_qpParam[ttype].lambda2; // 得到RDO中的lambda
const int64_t psyScale = ((int64_t)m_psyRdoqScale * m_qpParam[ttype].lambda); // 得到心理视觉量化系数
/* unquant constants for measuring distortion. Scaling list quant coefficients have a (1 << 4)
* scale applied that must be removed during unquant. Note that in real dequant there is clipping
* at several stages. We skip the clipping for simplicity when measuring RD cost */
const int32_t* unquantScale = m_scalingList->m_dequantCoef[log2TrSize - 2][scalingListType][rem]; // 得到反量化乘数
int unquantShift = QUANT_IQUANT_SHIFT - QUANT_SHIFT - transformShift + (m_scalingList->m_bEnabled ? 4 : 0); // 得到反量化右移位数
int unquantRound = (unquantShift > per) ? 1 << (unquantShift - per - 1) : 0; // 反量化右移时补偿加数
int scaleBits = SCALE_BITS - 2 * transformShift; //
#define UNQUANT(lvl) (((lvl) * (unquantScale[blkPos] << per) + unquantRound) >> unquantShift)
#define SIGCOST(bits) ((lambda2 * (bits)) >> 8)
#define RDCOST(d, bits) ((((int64_t)d * d) << scaleBits) + SIGCOST(bits))
#define PSYVALUE(rec) ((psyScale * (rec)) >> (2 * transformShift + 1))
// 以下是系数级别的变量，即每个系数都占用一个存储单元
int64_t costCoeff[32 * 32]; // 每一个系数花费 /* d*d + lambda * bits */
int64_t costUncoded[32 * 32]; // 每一个系数被量化为0的花费（Z型顺序存储） /* d*d + lambda * 0 */
int64_t costSig[32 * 32]; // 每一个系数的是否为0标记(sig_coeff_flag)的花费 /* lambda * bits */
int rateIncUp[32 * 32]; /* signal overhead of increasing level */
int rateIncDown[32 * 32]; /* signal overhead of decreasing level */
int sigRateDelta[32 * 32]; // 将系数量化为0和量化为非0，系数标记（sig_coeff_flag）的花费差异 /* signal difference between zero and non-zero */
// 以下是系数组（CG）级别的变量
int64_t costCoeffGroupSig[MLS_GRP_NUM]; // 每一个系数组CG的花费 /* lambda * bits of group coding cost */
uint64_t sigCoeffGroupFlag64 = 0; // CG的非零标记，每一位代表一个CG，某一位为1代表对应的CG不是全0，反之代表对应的CG为全0
const uint32_t cgSize = (1 << MLS_CG_SIZE); // 一个系数组CG中的系数个数 /* 4x4 num coef = 16 */
bool bIsLuma = ttype == TEXT_LUMA; // 当前是否是亮度分量
/* total rate distortion cost of transform block, as CBF=0 */
int64_t totalUncodedCost = 0; // 当前块都被量化为0时的cost
/* Total rate distortion cost of this transform block, counting te distortion of uncoded blocks,
* the distortion and signal cost of coded blocks, and the coding cost of significant
* coefficient and coefficient group bitmaps */
int64_t totalRdCost = 0;
TUEntropyCodingParameters codeParams;
cu.getTUEntropyCodingParameters(codeParams, absPartIdx, log2TrSize, bIsLuma); // 得到熵编码参数
const uint32_t cgNum = 1 << (codeParams.log2TrSizeCG * 2); // TU中的系数组CG的个数
const uint32_t cgStride = (trSize >> MLS_CG_LOG2_SIZE); // TU中CG的步长
uint8_t coeffNum[MLS_GRP_NUM]; // 每个CG中的非零系数的个数 // value range[0, 16]
uint16_t coeffSign[MLS_GRP_NUM]; // 每个CG中的非零系数的符号 // bit mask map for non-zero coeff sign
uint16_t coeffFlag[MLS_GRP_NUM]; // 每个CG中的非零系数的标记 // bit mask map for non-zero coeff
#if CHECKED_BUILD || _DEBUG
// clean output buffer, the asm version of scanPosLast Never output anything after latest non-zero coeff group
memset(coeffNum, 0, sizeof(coeffNum));
memset(coeffSign, 0, sizeof(coeffNum)); // 这里size应该是写错了，应该改为 sizeof(coeffSign)，下面也是一样
memset(coeffFlag, 0, sizeof(coeffNum));
#endif
// 统计每个CG中的非零系数的符号、非零系数的标志（是否是非零系数）、非零系数个数，以及最后一个非零系数的扫描位置（lastScanPos）
const int lastScanPos = primitives.scanPosLast(codeParams.scan, dstCoeff, coeffSign, coeffFlag, coeffNum, numSig, g_scan4x4[codeParams.scanType], trSize); // 参看 scanPosLast_c
const int cgLastScanPos = (lastScanPos >> LOG2_SCAN_SET_SIZE); // 得到最后一个非零系数的扫描位置（lastScanPos）所对应的系数组CG的位置
// 但是如果 lastScanPos%(2^LOG2_SCAN_SET_SIZE) != 0，即如果最后一个非零系数位置并不能被16整除，这里得到的实际上倒数第二个非零CG的位置
/* TODO: update bit estimates if dirty */
EstBitsSbac& estBitsSbac = m_entropyCoder->m_estBitsSbac;
uint32_t scanPos = 0;
uint32_t c1 = 1;
// process trail all zero Coeff Group
/* coefficients after lastNZ have no distortion signal cost */
const int zeroCG = cgNum - 1 - cgLastScanPos; // 得到全零CG的个数
memset(&costCoeff[(cgLastScanPos + 1) << MLS_CG_SIZE], 0, zeroCG * MLS_CG_BLK_SIZE * sizeof(int64_t)); // 将全零CG中的每个系数花费都设置为0
memset(&costSig[(cgLastScanPos + 1) << MLS_CG_SIZE], 0, zeroCG * MLS_CG_BLK_SIZE * sizeof(int64_t)); // 将全零CG中的每个系数的标记花费设置为0
/* sum zero coeff (uncodec) cost */
// TODO: does we need these cost?
if (usePsyMask) // 如果使用心理视觉量化
{
for (int cgScanPos = cgLastScanPos + 1; cgScanPos < (int)cgNum ; cgScanPos++)
{
X265_CHECK(coeffNum[cgScanPos] == 0, "count of coeff failure\n");
uint32_t scanPosBase = (cgScanPos << MLS_CG_SIZE);
uint32_t blkPos = codeParams.scan[scanPosBase];
// TODO: we can't SIMD optimize because PSYVALUE need 64-bits multiplication, convert to Double can work faster by FMA
for (int y = 0; y < MLS_CG_SIZE; y++)
{
for (int x = 0; x < MLS_CG_SIZE; x++)
{
int signCoef = m_resiDctCoeff[blkPos + x]; // 得到变换系数 /* pre-quantization DCT coeff */
int predictedCoef = m_fencDctCoeff[blkPos + x] - signCoef; /* predicted DCT = source DCT - residual DCT*/
costUncoded[blkPos + x] = ((int64_t)signCoef * signCoef) << scaleBits; // 计算将变换系数量化为0的cost（distortion 部分）
/* when no residual coefficient is coded, predicted coef == recon coef */
costUncoded[blkPos + x] -= PSYVALUE(predictedCoef);
totalUncodedCost += costUncoded[blkPos + x]; // 累加量化为0的cost
totalRdCost += costUncoded[blkPos + x]; // 累加量化为0的cost
}
blkPos += trSize;
}
}
}
else // 不使用心理视觉量化，对量化为全零的CG计算每个系数的cost
{
// non-psy path
for (int cgScanPos = cgLastScanPos + 1; cgScanPos < (int)cgNum ; cgScanPos++) // 遍历每一个全零的CG
{
X265_CHECK(coeffNum[cgScanPos] == 0, "count of coeff failure\n"); // 再次确认是否该CG中非零系数的个数是0
uint32_t scanPosBase = (cgScanPos << MLS_CG_SIZE); // 得到每个CG的首地址（这里的CG顺序是Z型扫描，而不是按照选择的scan模式扫描）
uint32_t blkPos = codeParams.scan[scanPosBase]; // 找到一个CG首地址对应的扫描位置
for (int y = 0; y < MLS_CG_SIZE; y++) // 每个CG内部仍然按照Z型扫描存储 costUncoded，但CG位置是通过选择的扫描顺序得到的
{
for (int x = 0; x < MLS_CG_SIZE; x++)
{
int signCoef = m_resiDctCoeff[blkPos + x]; // 得到变换系数 /* pre-quantization DCT coeff */
costUncoded[blkPos + x] = ((int64_t)signCoef * signCoef) << scaleBits; // 计算将变换系数量化为0的cost（distortion 部分）
totalUncodedCost += costUncoded[blkPos + x]; // 累加量化为0的cost
totalRdCost += costUncoded[blkPos + x]; // 累加量化为0的cost
}
blkPos += trSize;
}
}
}
static const uint8_t table_cnt[5][SCAN_SET_SIZE] =
{
// patternSigCtx = 0
{
2, 1, 1, 0,
1, 1, 0, 0,
1, 0, 0, 0,
0, 0, 0, 0,
},
// patternSigCtx = 1
{
2, 2, 2, 2,
1, 1, 1, 1,
0, 0, 0, 0,
0, 0, 0, 0,
},
// patternSigCtx = 2
{
2, 1, 0, 0,
2, 1, 0, 0,
2, 1, 0, 0,
2, 1, 0, 0,
},
// patternSigCtx = 3
{
2, 2, 2, 2,
2, 2, 2, 2,
2, 2, 2, 2,
2, 2, 2, 2,
},
// 4x4
{
0, 1, 4, 5,
2, 3, 4, 5,
6, 6, 8, 8,
7, 7, 8, 8
}
};
/* iterate over coding groups in reverse scan order */
// step1. 对每个系数单独做RDO优化，找到率失真意义上的最优量化值
for (int cgScanPos = cgLastScanPos; cgScanPos >= 0; cgScanPos--) // 从最后一非零CG开始遍历每个CG，从最后一个到第一个
{
uint32_t ctxSet = (cgScanPos && bIsLuma) ? 2 : 0;
const uint32_t cgBlkPos = codeParams.scanCG[cgScanPos]; // 得到CG的扫描位置
const uint32_t cgPosY = cgBlkPos >> codeParams.log2TrSizeCG; // CG扫描位置的Y坐标
const uint32_t cgPosX = cgBlkPos - (cgPosY << codeParams.log2TrSizeCG); // CG扫描位置的X坐标
const uint64_t cgBlkPosMask = ((uint64_t)1 << cgBlkPos); // 当前CG的位置，使用cgBlkPosMask中1的位置来表示
const int patternSigCtx = calcPatternSigCtx(sigCoeffGroupFlag64, cgPosX, cgPosY, cgBlkPos, cgStride);
const int ctxSigOffset = codeParams.firstSignificanceMapContext + (cgScanPos && bIsLuma ? 3 : 0);
if (c1 == 0)
ctxSet++;
c1 = 1;
if (cgScanPos && (coeffNum[cgScanPos] == 0)) // 如果当前CG不是第一个CG，并且CG系数为全零，即在最后一个非全零CG可能还会存在全零的CG
{ // 如果发现全零的CG，处理方法与上文中，最后一个非零CG之后的CG的处理方法相同；但是对这些系数又计算了量化为0和非零时，系数标记的花费，这是为了之后进行step2、step3的优化
// TODO: does we need zero-coeff cost?
const uint32_t scanPosBase = (cgScanPos << MLS_CG_SIZE);
uint32_t blkPos = codeParams.scan[scanPosBase];
if (usePsyMask) // 如果使用心理视觉量化
{
// TODO: we can't SIMD optimize because PSYVALUE need 64-bits multiplication, convert to Double can work faster by FMA
for (int y = 0; y < MLS_CG_SIZE; y++)
{
for (int x = 0; x < MLS_CG_SIZE; x++)
{
int signCoef = m_resiDctCoeff[blkPos + x]; /* pre-quantization DCT coeff */
int predictedCoef = m_fencDctCoeff[blkPos + x] - signCoef; /* predicted DCT = source DCT - residual DCT*/
costUncoded[blkPos + x] = ((int64_t)signCoef * signCoef) << scaleBits;
/* when no residual coefficient is coded, predicted coef == recon coef */
costUncoded[blkPos + x] -= PSYVALUE(predictedCoef);
totalUncodedCost += costUncoded[blkPos + x];
totalRdCost += costUncoded[blkPos + x];
const uint32_t scanPosOffset = y * MLS_CG_SIZE + x;
const uint32_t ctxSig = table_cnt[patternSigCtx][g_scan4x4[codeParams.scanType][scanPosOffset]] + ctxSigOffset;
X265_CHECK(trSize > 4, "trSize check failure\n");
X265_CHECK(ctxSig == getSigCtxInc(patternSigCtx, log2TrSize, trSize, codeParams.scan[scanPosBase + scanPosOffset], bIsLuma, codeParams.firstSignificanceMapContext), "sigCtx check failure\n");
costSig[scanPosBase + scanPosOffset] = SIGCOST(estBitsSbac.significantBits[0][ctxSig]);
costCoeff[scanPosBase + scanPosOffset] = costUncoded[blkPos + x];
sigRateDelta[blkPos + x] = estBitsSbac.significantBits[1][ctxSig] - estBitsSbac.significantBits[0][ctxSig];
}
blkPos += trSize;
}
}
else // 如果不使用心理视觉量化
{
// non-psy path
for (int y = 0; y < MLS_CG_SIZE; y++) // 对整个CG进行遍历
{
for (int x = 0; x < MLS_CG_SIZE; x++)
{
int signCoef = m_resiDctCoeff[blkPos + x]; // 得到变换系数 /* pre-quantization DCT coeff */
costUncoded[blkPos + x] = ((int64_t)signCoef * signCoef) << scaleBits; // 计算将变换系数量化为0的cost（distortion 部分）
totalUncodedCost += costUncoded[blkPos + x]; // 累加量化为0的cost
totalRdCost += costUncoded[blkPos + x]; // 累加量化为0的cost
const uint32_t scanPosOffset = y * MLS_CG_SIZE + x;
const uint32_t ctxSig = table_cnt[patternSigCtx][g_scan4x4[codeParams.scanType][scanPosOffset]] + ctxSigOffset;
X265_CHECK(trSize > 4, "trSize check failure\n");
X265_CHECK(ctxSig == getSigCtxInc(patternSigCtx, log2TrSize, trSize, codeParams.scan[scanPosBase + scanPosOffset], bIsLuma, codeParams.firstSignificanceMapContext), "sigCtx check failure\n");
costSig[scanPosBase + scanPosOffset] = SIGCOST(estBitsSbac.significantBits[0][ctxSig]); // 计算系数标记的cost，实际上全零的CG没有这部分的花费
costCoeff[scanPosBase + scanPosOffset] = costUncoded[blkPos + x]; // 对于全零的CG，每个系数的花费就是量化为0的distortion部分
sigRateDelta[blkPos + x] = estBitsSbac.significantBits[1][ctxSig] - estBitsSbac.significantBits[0][ctxSig]; // 计算系数被量化为0和非0，系数标记的bit花费差异
}
blkPos += trSize;
}
}
/* there were no coded coefficients in this coefficient group */
{
uint32_t ctxSig = getSigCoeffGroupCtxInc(sigCoeffGroupFlag64, cgPosX, cgPosY, cgBlkPos, cgStride);
costCoeffGroupSig[cgScanPos] = SIGCOST(estBitsSbac.significantCoeffGroupBits[ctxSig][0]); // 得到CG的非零标记的cost
totalRdCost += costCoeffGroupSig[cgScanPos]; /* add cost of 0 bit in significant CG bitmap */
}
continue;
}
coeffGroupRDStats cgRdStats;
memset(&cgRdStats, 0, sizeof(coeffGroupRDStats));
uint32_t subFlagMask = coeffFlag[cgScanPos]; // 得到一个CG内系数的标记
int c2 = 0;
uint32_t goRiceParam = 0;
uint32_t c1Idx = 0;
uint32_t c2Idx = 0;
/* iterate over coefficients in each group in reverse scan order */
for (int scanPosinCG = cgSize - 1; scanPosinCG >= 0; scanPosinCG--) // 扫描CG内部的每一个系数
{
scanPos = (cgScanPos << MLS_CG_SIZE) + scanPosinCG; // 找到每个系数的Z型扫描顺序
uint32_t blkPos = codeParams.scan[scanPos]; // 找到对应的扫描位置
uint32_t maxAbsLevel = abs(dstCoeff[blkPos]); // 得到常规量化后的值 /* abs(quantized coeff) */
int signCoef = m_resiDctCoeff[blkPos]; // 得到DCT系数 /* pre-quantization DCT coeff */
int predictedCoef = m_fencDctCoeff[blkPos] - signCoef; /* predicted DCT = source DCT - residual DCT*/
/* RDOQ measures distortion as the squared difference between the unquantized coded level
* and the original DCT coefficient. The result is shifted scaleBits to account for the
* FIX15 nature of the CABAC cost tables minus the forward transform scale */
/* cost of not coding this coefficient (all distortion, no signal bits) */
costUncoded[blkPos] = ((int64_t)signCoef * signCoef) << scaleBits; // 得到量化为0的cost
X265_CHECK((!!scanPos ^ !!blkPos) == 0, "failed on (blkPos=0 && scanPos!=0)\n"); // ^表示按位异或，若参加运算的两个二进制位值相同则为0，否则为1。这里是保证scanPos和blkPos同时为0或者同时不为0
if (usePsyMask & scanPos)
/* when no residual coefficient is coded, predicted coef == recon coef */
costUncoded[blkPos] -= PSYVALUE(predictedCoef);
totalUncodedCost += costUncoded[blkPos]; // 累加量化为0的cost
// coefficient level estimation
const int* greaterOneBits = estBitsSbac.greaterOneBits[4 * ctxSet + c1];
const uint32_t ctxSig = (blkPos == 0) ? 0 : table_cnt[(trSize == 4) ? 4 : patternSigCtx][g_scan4x4[codeParams.scanType][scanPosinCG]] + ctxSigOffset;
X265_CHECK(ctxSig == getSigCtxInc(patternSigCtx, log2TrSize, trSize, blkPos, bIsLuma, codeParams.firstSignificanceMapContext), "sigCtx check failure\n");
// before find lastest non-zero coeff
if (scanPos > (uint32_t)lastScanPos) // 如果当前扫描位置在最后一个非零系数之后，则将这些系数的cost都设置为0
{ // 这种情况出现是由于最后一个非零系数可能出现在最后一个非零CG中的任何位置，所以对最后一个CG扫描就会出现这种情况
/* coefficients after lastNZ have no distortion signal cost */
costCoeff[scanPos] = 0;
costSig[scanPos] = 0;
/* No non-zero coefficient yet found, but this does not mean
* there is no uncoded-cost for this coefficient. Pre-
* quantization the coefficient may have been non-zero */
totalRdCost += costUncoded[blkPos];
}
else if (!(subFlagMask & 1)) // 如果当前位置位于最后一个非零系数之前，并且该系数为0，计算量化为0的cost
{
// fast zero coeff path
/* set default costs to uncoded costs */
costSig[scanPos] = SIGCOST(estBitsSbac.significantBits[0][ctxSig]); // 计算系数标记sig_coeff_flag为0的cost
costCoeff[scanPos] = costUncoded[blkPos] + costSig[scanPos]; // 得到一个系数的总花费， = 量化为0的distortion + 系数标记为0的cost
sigRateDelta[blkPos] = estBitsSbac.significantBits[1][ctxSig] - estBitsSbac.significantBits[0][ctxSig]; // 将系数量化为0和量化为非0，系数标记（sig_coeff_flag）的花费差异
totalRdCost += costCoeff[scanPos]; // 累加系数总花费
rateIncUp[blkPos] = greaterOneBits[0]; // 得到比当前量化系数大1所需要花费的bit ???
subFlagMask >>= 1; // 系数非零标记右移移位，下次取到CG中前一个系数的非零标记
}
else // 如果当前位置位于最后一个非零系数之前，并且该系数为1，估计每一个系数的最优量化值
{
subFlagMask >>= 1; // 系数非零标记右移移位，下次取到CG中前一个系数的非零标记
const uint32_t c1c2Idx = ((c1Idx - 8) >> (sizeof(int) * CHAR_BIT - 1)) + (((-(int)c2Idx) >> (sizeof(int) * CHAR_BIT - 1)) + 1) * 2;
const uint32_t baseLevel = ((uint32_t)0xD9 >> (c1c2Idx * 2)) & 3; // {1, 2, 1, 3}
X265_CHECK(!!((int)c1Idx < C1FLAG_NUMBER) == (int)((c1Idx - 8) >> (sizeof(int) * CHAR_BIT - 1)), "scan validation 1\n");
X265_CHECK(!!(c2Idx == 0) == ((-(int)c2Idx) >> (sizeof(int) * CHAR_BIT - 1)) + 1, "scan validation 2\n");
X265_CHECK((int)baseLevel == ((c1Idx < C1FLAG_NUMBER) ? (2 + (c2Idx == 0)) : 1), "scan validation 3\n");
// coefficient level estimation
const int* levelAbsBits = estBitsSbac.levelAbsBits[ctxSet + c2];
uint32_t level = 0;
uint32_t sigCoefBits = 0;
costCoeff[scanPos] = MAX_INT64;
if ((int)scanPos == lastScanPos) // 如果当前位置是最后一个非零系数的位置，则将量化为0和非0的系数标记（sig_coeff_flag）的花费差异设为0
sigRateDelta[blkPos] = 0;
else
{
if (maxAbsLevel < 3) // 如果该系数的常规量化值很小，则尝试将它量化为0
{
/* set default costs to uncoded costs */
costSig[scanPos] = SIGCOST(estBitsSbac.significantBits[0][ctxSig]); // 计算系数标记sig_coeff_flag为0的cost
costCoeff[scanPos] = costUncoded[blkPos] + costSig[scanPos]; // 得到一个系数的总花费， = 量化为0的distortion + 系数标记为0的cost
}
sigRateDelta[blkPos] = estBitsSbac.significantBits[1][ctxSig] - estBitsSbac.significantBits[0][ctxSig]; // 将系数量化为0和量化为非0，系数标记（sig_coeff_flag）的花费差异
sigCoefBits = estBitsSbac.significantBits[1][ctxSig]; // 将系数量化为非0，系数标记（sig_coeff_flag）的花费
}
// NOTE: X265_MAX(maxAbsLevel - 1, 1) ==> (X>=2 -> X-1), (X<2 -> 1) | (0 < X < 2 ==> X=1)
if (maxAbsLevel == 1) // 如果常规量化值为1，则与量化为0相比较
{
uint32_t levelBits = (c1c2Idx & 1) ? greaterOneBits[0] + IEP_RATE : ((1 + goRiceParam) << 15) + IEP_RATE; // 计算量化为1时，系数幅值的bit消耗
X265_CHECK(levelBits == getICRateCost(1, 1 - baseLevel, greaterOneBits, levelAbsBits, goRiceParam, c1c2Idx) + IEP_RATE, "levelBits mistake\n");
int unquantAbsLevel = UNQUANT(1); // 得到1的反量化值
int d = abs(signCoef) - unquantAbsLevel; // 得到distortion
int64_t curCost = RDCOST(d, sigCoefBits + levelBits); // 计算量化为1的RDcost
/* Psy RDOQ: bias in favor of higher AC coefficients in the reconstructed frame */
if (usePsyMask & scanPos)
{
int reconCoef = abs(unquantAbsLevel + SIGN(predictedCoef, signCoef));
curCost -= PSYVALUE(reconCoef);
}
if (curCost < costCoeff[scanPos]) // 将量化为1的cost与量化为0相比较，如果量化为1更好，则将其量化为1
{
level = 1; // 更新量化值为1
costCoeff[scanPos] = curCost; // 更新量化当前系数的cost
costSig[scanPos] = SIGCOST(sigCoefBits); // 更新sig_coeff_flag的cost
}
}
else if (maxAbsLevel) // 如果常规量化值大于1
{
// 计算量化为当前常规量化值所花费的比特数，并计算量化为常规量化值减1所花费的比特数
uint32_t levelBits0 = getICRateCost(maxAbsLevel, maxAbsLevel - baseLevel, greaterOneBits, levelAbsBits, goRiceParam, c1c2Idx) + IEP_RATE;
uint32_t levelBits1 = getICRateCost(maxAbsLevel - 1, maxAbsLevel - 1 - baseLevel, greaterOneBits, levelAbsBits, goRiceParam, c1c2Idx) + IEP_RATE;
// 计算量化为当前常规量化值所产生的cost
int unquantAbsLevel0 = UNQUANT(maxAbsLevel); // 得到反量化值
int d0 = abs(signCoef) - unquantAbsLevel0; // 得到distortion
int64_t curCost0 = RDCOST(d0, sigCoefBits + levelBits0); // 计算得到RDcost
// 计算量化为当前常规量化值减1所产生的cost
int unquantAbsLevel1 = UNQUANT(maxAbsLevel - 1); // 得到反量化值
int d1 = abs(signCoef) - unquantAbsLevel1; // 得到distortion
int64_t curCost1 = RDCOST(d1, sigCoefBits + levelBits1); // 计算得到RDcost
/* Psy RDOQ: bias in favor of higher AC coefficients in the reconstructed frame */
if (usePsyMask & scanPos)
{
int reconCoef;
reconCoef = abs(unquantAbsLevel0 + SIGN(predictedCoef, signCoef));
curCost0 -= PSYVALUE(reconCoef);
reconCoef = abs(unquantAbsLevel1 + SIGN(predictedCoef, signCoef));
curCost1 -= PSYVALUE(reconCoef);
}
if (curCost0 < costCoeff[scanPos]) // 如果量化为当前常规量化值的cost更小，则更新量化值和cost
{
level = maxAbsLevel; // 更新量化值为1
costCoeff[scanPos] = curCost0; // 更新量化当前系数的cost
costSig[scanPos] = SIGCOST(sigCoefBits); // 更新sig_coeff_flag的cost
}
if (curCost1 < costCoeff[scanPos]) // 如果量化为当前常规量化值减1的cost更小，则更新量化值和cost
{
level = maxAbsLevel - 1; // 更新量化值为1
costCoeff[scanPos] = curCost1; // 更新量化当前系数的cost
costSig[scanPos] = SIGCOST(sigCoefBits); // 更新sig_coeff_flag的cost
}
}
dstCoeff[blkPos] = (int16_t)level; // 在对当前量化系数进行完RDO优化后，更新目标量化值
totalRdCost += costCoeff[scanPos]; // 累加总的RDcost
/* record costs for sign-hiding performed at the end */
if ((cu.m_slice->m_pps->bSignHideEnabled ? ~0 : 0) & level)
{
const int32_t diff0 = level - 1 - baseLevel;
const int32_t diff2 = level + 1 - baseLevel;
const int32_t maxVlc = g_goRiceRange[goRiceParam];
int rate0, rate1, rate2;
if (diff0 < -2) // prob (92.9, 86.5, 74.5)%
{
// NOTE: Min: L - 1 - {1,2,1,3} < -2 ==> L < {0,1,0,2}
// additional L > 0, so I got (L > 0 && L < 2) ==> L = 1
X265_CHECK(level == 1, "absLevel check failure\n");
const int rateEqual2 = greaterOneBits[1] + levelAbsBits[0];;
const int rateNotEqual2 = greaterOneBits[0];
rate0 = 0;
rate2 = rateEqual2;
rate1 = rateNotEqual2;
X265_CHECK(rate1 == getICRateNegDiff(level + 0, greaterOneBits, levelAbsBits), "rate1 check failure!\n");
X265_CHECK(rate2 == getICRateNegDiff(level + 1, greaterOneBits, levelAbsBits), "rate1 check failure!\n");
X265_CHECK(rate0 == getICRateNegDiff(level - 1, greaterOneBits, levelAbsBits), "rate1 check failure!\n");
}
else if (diff0 >= 0 && diff2 <= maxVlc) // prob except from above path (98.6, 97.9, 96.9)%
{
// NOTE: no c1c2 correct rate since all of rate include this factor
rate1 = getICRateLessVlc(level + 0, diff0 + 1, goRiceParam);
rate2 = getICRateLessVlc(level + 1, diff0 + 2, goRiceParam);
rate0 = getICRateLessVlc(level - 1, diff0 + 0, goRiceParam);
}
else
{
rate1 = getICRate(level + 0, diff0 + 1, greaterOneBits, levelAbsBits, goRiceParam, maxVlc, c1c2Idx);
rate2 = getICRate(level + 1, diff0 + 2, greaterOneBits, levelAbsBits, goRiceParam, maxVlc, c1c2Idx);
rate0 = getICRate(level - 1, diff0 + 0, greaterOneBits, levelAbsBits, goRiceParam, maxVlc, c1c2Idx);
}
rateIncUp[blkPos] = rate2 - rate1;
rateIncDown[blkPos] = rate0 - rate1;
}
else
{
rateIncUp[blkPos] = greaterOneBits[0];
rateIncDown[blkPos] = 0;
}
/* Update CABAC estimation state */
if (level >= baseLevel && goRiceParam < 4 && level > (3U << goRiceParam))
goRiceParam++;
c1Idx -= (-(int32_t)level) >> 31;
/* update bin model */
if (level > 1)
{
c1 = 0;
c2 += (uint32_t)(c2 - 2) >> 31;
c2Idx++;
}
else if ((c1 < 3) && (c1 > 0) && level)
c1++;
if (dstCoeff[blkPos]) // 如果当前量化系数为非零
{
sigCoeffGroupFlag64 |= cgBlkPosMask; // 与当前CG位置的Mask相或，标志该CG不是全0
cgRdStats.codedLevelAndDist += costCoeff[scanPos] - costSig[scanPos]; //
cgRdStats.uncodedDist += costUncoded[blkPos];
cgRdStats.nnzBeforePos0 += scanPosinCG;
}
}
cgRdStats.sigCost += costSig[scanPos];
} /* end for (scanPosinCG) */
X265_CHECK((cgScanPos << MLS_CG_SIZE) == (int)scanPos, "scanPos mistake\n");
cgRdStats.sigCost0 = costSig[scanPos];
costCoeffGroupSig[cgScanPos] = 0;
/* nothing to do at this case */
X265_CHECK(cgLastScanPos >= 0, "cgLastScanPos check failure\n");
if (!cgScanPos || cgScanPos == cgLastScanPos)
{
/* coeff group 0 is implied to be present, no signal cost */
/* coeff group with last NZ is implied to be present, handled below */
}
else if (sigCoeffGroupFlag64 & cgBlkPosMask)
{
if (!cgRdStats.nnzBeforePos0)
{
/* if only coeff 0 in this CG is coded, its significant coeff bit is implied */
totalRdCost -= cgRdStats.sigCost0;
cgRdStats.sigCost -= cgRdStats.sigCost0;
}
/* there are coded coefficients in this group, but now we include the signaling cost
* of the significant coefficient group flag and evaluate whether the RD cost of the
* coded group is more than the RD cost of the uncoded group */
uint32_t sigCtx = getSigCoeffGroupCtxInc(sigCoeffGroupFlag64, cgPosX, cgPosY, cgBlkPos, cgStride);
int64_t costZeroCG = totalRdCost + SIGCOST(estBitsSbac.significantCoeffGroupBits[sigCtx][0]);
costZeroCG += cgRdStats.uncodedDist; /* add distortion for resetting non-zero levels to zero levels */
costZeroCG -= cgRdStats.codedLevelAndDist; /* remove distortion and level cost of coded coefficients */
costZeroCG -= cgRdStats.sigCost; /* remove signaling cost of significant coeff bitmap */
costCoeffGroupSig[cgScanPos] = SIGCOST(estBitsSbac.significantCoeffGroupBits[sigCtx][1]);
totalRdCost += costCoeffGroupSig[cgScanPos]; /* add the cost of 1 bit in significant CG bitmap */
if (costZeroCG < totalRdCost && m_rdoqLevel > 1)
{
sigCoeffGroupFlag64 &= ~cgBlkPosMask;
totalRdCost = costZeroCG;
costCoeffGroupSig[cgScanPos] = SIGCOST(estBitsSbac.significantCoeffGroupBits[sigCtx][0]);
/* reset all coeffs to 0. UNCODE THIS COEFF GROUP! */
const uint32_t blkPos = codeParams.scan[cgScanPos * cgSize];
memset(&dstCoeff[blkPos + 0 * trSize], 0, 4 * sizeof(*dstCoeff));
memset(&dstCoeff[blkPos + 1 * trSize], 0, 4 * sizeof(*dstCoeff));
memset(&dstCoeff[blkPos + 2 * trSize], 0, 4 * sizeof(*dstCoeff));
memset(&dstCoeff[blkPos + 3 * trSize], 0, 4 * sizeof(*dstCoeff));
}
}
else
{
/* there were no coded coefficients in this coefficient group */
uint32_t ctxSig = getSigCoeffGroupCtxInc(sigCoeffGroupFlag64, cgPosX, cgPosY, cgBlkPos, cgStride);
costCoeffGroupSig[cgScanPos] = SIGCOST(estBitsSbac.significantCoeffGroupBits[ctxSig][0]);
totalRdCost += costCoeffGroupSig[cgScanPos]; /* add cost of 0 bit in significant CG bitmap */
totalRdCost -= cgRdStats.sigCost; /* remove cost of significant coefficient bitmap */
}
} /* end for (cgScanPos) */
X265_CHECK(lastScanPos >= 0, "numSig non zero, but no coded CG\n");
/* calculate RD cost of uncoded block CBF=0, and add cost of CBF=1 to total */
int64_t bestCost;
if (!cu.isIntra(absPartIdx) && bIsLuma && !cu.m_tuDepth[absPartIdx])
{
bestCost = totalUncodedCost + SIGCOST(estBitsSbac.blockRootCbpBits[0]);
totalRdCost += SIGCOST(estBitsSbac.blockRootCbpBits[1]);
}
else
{
int ctx = ctxCbf[ttype][cu.m_tuDepth[absPartIdx]];
bestCost = totalUncodedCost + SIGCOST(estBitsSbac.blockCbpBits[ctx][0]);
totalRdCost += SIGCOST(estBitsSbac.blockCbpBits[ctx][1]);
}
/* This loop starts with the last non-zero found in the first loop and then refines this last
* non-zero by measuring the true RD cost of the last NZ at this position, and then the RD costs
* at all previous coefficients until a coefficient greater than 1 is encountered or we run out
* of coefficients to evaluate. This will factor in the cost of coding empty groups and empty
* coeff prior to the last NZ. The base best cost is the RD cost of CBF=0 */
int bestLastIdx = 0;
bool foundLast = false;
for (int cgScanPos = cgLastScanPos; cgScanPos >= 0 && !foundLast; cgScanPos--)
{
if (!cgScanPos || cgScanPos == cgLastScanPos)
{
/* the presence of these coefficient groups are inferred, they have no bit in
* sigCoeffGroupFlag64 and no saved costCoeffGroupSig[] cost */
}
else if (sigCoeffGroupFlag64 & (1ULL << codeParams.scanCG[cgScanPos]))
{
/* remove cost of significant coeff group flag, the group's presence would be inferred
* from lastNZ if it were present in this group */
totalRdCost -= costCoeffGroupSig[cgScanPos];
}
else
{
/* remove cost of signaling this empty group as not present */
totalRdCost -= costCoeffGroupSig[cgScanPos];
continue;
}
for (int scanPosinCG = cgSize - 1; scanPosinCG >= 0; scanPosinCG--)
{
scanPos = cgScanPos * cgSize + scanPosinCG;
if ((int)scanPos > lastScanPos)
continue;
/* if the coefficient was coded, measure the RD cost of it as the last non-zero and then
* continue as if it were uncoded. If the coefficient was already uncoded, remove the
* cost of signaling it as not-significant */
uint32_t blkPos = codeParams.scan[scanPos];
if (dstCoeff[blkPos]) // 如果目标量化系数不是0，则试图将该系数设置为0
{
// Calculates the cost of signaling the last significant coefficient in the block
uint32_t pos[2] = { (blkPos & (trSize - 1)), (blkPos >> log2TrSize) }; // 得到该系数所在位置的X/Y坐标，X = blkPos&(trSize-1); Y = blkPos>>log2TrSize
if (codeParams.scanType == SCAN_VER) // 如果当前的扫描类型是竖直扫描，则调换X和Y坐标
std::swap(pos[0], pos[1]);
uint32_t bitsLastNZ = 0;
for (int i = 0; i < 2; i++) // 估计该位置的X/Y坐标所需要的bit花费，X=pos[0]，Y=pos[1]
{
int temp = g_lastCoeffTable[pos[i]]; // 得到该系数位置的前缀和后缀
int prefixOnes = temp & 15;
int suffixLen = temp >> 4;
bitsLastNZ += m_entropyCoder->m_estBitsSbac.lastBits[i][prefixOnes]; // 估计对前缀熵编码所消耗的bits
bitsLastNZ += IEP_RATE * suffixLen; // 加上后缀所消耗的bits
}
int64_t costAsLast = totalRdCost - costSig[scanPos] + SIGCOST(bitsLastNZ);
if (costAsLast < bestCost)
{
bestLastIdx = scanPos + 1;
bestCost = costAsLast;
}
if (dstCoeff[blkPos] > 1 || m_rdoqLevel == 1)
{
foundLast = true;
break;
}
totalRdCost -= costCoeff[scanPos];
totalRdCost += costUncoded[blkPos];
}
else
totalRdCost -= costSig[scanPos];
}
}
/* recount non-zero coefficients and re-apply sign of DCT coef */
numSig = 0;
for (int pos = 0; pos < bestLastIdx; pos++)
{
int blkPos = codeParams.scan[pos];
int level = dstCoeff[blkPos];
numSig += (level != 0);
uint32_t mask = (int32_t)m_resiDctCoeff[blkPos] >> 31;
dstCoeff[blkPos] = (int16_t)((level ^ mask) - mask);
}
// Average 49.62 pixels
/* clean uncoded coefficients */
for (int pos = bestLastIdx; pos <= fastMin(lastScanPos, (bestLastIdx | (SCAN_SET_SIZE - 1))); pos++)
{
dstCoeff[codeParams.scan[pos]] = 0;
}
for (int pos = (bestLastIdx & ~(SCAN_SET_SIZE - 1)) + SCAN_SET_SIZE; pos <= lastScanPos; pos += SCAN_SET_SIZE)
{
const uint32_t blkPos = codeParams.scan[pos];
memset(&dstCoeff[blkPos + 0 * trSize], 0, 4 * sizeof(*dstCoeff));
memset(&dstCoeff[blkPos + 1 * trSize], 0, 4 * sizeof(*dstCoeff));
memset(&dstCoeff[blkPos + 2 * trSize], 0, 4 * sizeof(*dstCoeff));
memset(&dstCoeff[blkPos + 3 * trSize], 0, 4 * sizeof(*dstCoeff));
}
/* rate-distortion based sign-hiding */
if (cu.m_slice->m_pps->bSignHideEnabled && numSig >= 2)
{
const int realLastScanPos = (bestLastIdx - 1) >> LOG2_SCAN_SET_SIZE;
int lastCG = true;
for (int subSet = realLastScanPos; subSet >= 0; subSet--)
{
int subPos = subSet << LOG2_SCAN_SET_SIZE;
int n;
if (!(sigCoeffGroupFlag64 & (1ULL << codeParams.scanCG[subSet])))
continue;
/* measure distance between first and last non-zero coef in this
* coding group */
const uint32_t posFirstLast = primitives.findPosFirstLast(&dstCoeff[codeParams.scan[subPos]], trSize, g_scan4x4[codeParams.scanType]);
int firstNZPosInCG = (uint16_t)posFirstLast;
int lastNZPosInCG = posFirstLast >> 16;
if (lastNZPosInCG - firstNZPosInCG >= SBH_THRESHOLD)
{
uint32_t signbit = (dstCoeff[codeParams.scan[subPos + firstNZPosInCG]] > 0 ? 0 : 1);
int absSum = 0;
for (n = firstNZPosInCG; n <= lastNZPosInCG; n++)
absSum += dstCoeff[codeParams.scan[n + subPos]];
if (signbit != (absSum & 1U))
{
/* We must find a coeff to toggle up or down so the sign bit of the first non-zero coeff
* is properly implied. Note dstCoeff[] are signed by this point but curChange and
* finalChange imply absolute levels (+1 is away from zero, -1 is towards zero) */
int64_t minCostInc = MAX_INT64, curCost = MAX_INT64;
int minPos = -1;
int16_t finalChange = 0, curChange = 0;
for (n = (lastCG ? lastNZPosInCG : SCAN_SET_SIZE - 1); n >= 0; --n)
{
uint32_t blkPos = codeParams.scan[n + subPos];
int signCoef = m_resiDctCoeff[blkPos]; /* pre-quantization DCT coeff */
int absLevel = abs(dstCoeff[blkPos]);
int d = abs(signCoef) - UNQUANT(absLevel);
int64_t origDist = (((int64_t)d * d)) << scaleBits;
#define DELTARDCOST(d, deltabits) ((((int64_t)d * d) << scaleBits) - origDist + ((lambda2 * (int64_t)(deltabits)) >> 8))
if (dstCoeff[blkPos])
{
d = abs(signCoef) - UNQUANT(absLevel + 1);
int64_t costUp = DELTARDCOST(d, rateIncUp[blkPos]);
/* if decrementing would make the coeff 0, we can include the
* significant coeff flag cost savings */
d = abs(signCoef) - UNQUANT(absLevel - 1);
bool isOne = abs(dstCoeff[blkPos]) == 1;
int downBits = rateIncDown[blkPos] - (isOne ? (IEP_RATE + sigRateDelta[blkPos]) : 0);
int64_t costDown = DELTARDCOST(d, downBits);
if (lastCG && lastNZPosInCG == n && isOne)
costDown -= 4 * IEP_RATE;
if (costUp < costDown)
{
curCost = costUp;
curChange = 1;
}
else
{
curChange = -1;
if (n == firstNZPosInCG && isOne)
curCost = MAX_INT64;
else
curCost = costDown;
}
}
else if (n < firstNZPosInCG && signbit != (signCoef >= 0 ? 0 : 1U))
{
/* don't try to make a new coded coeff before the first coeff if its
* sign would be different than the first coeff, the inferred sign would
* still be wrong and we'd have to do this again. */
curCost = MAX_INT64;
}
else
{
/* evaluate changing an uncoded coeff 0 to a coded coeff +/-1 */
d = abs(signCoef) - UNQUANT(1);
curCost = DELTARDCOST(d, rateIncUp[blkPos] + IEP_RATE + sigRateDelta[blkPos]);
curChange = 1;
}
if (curCost < minCostInc)
{
minCostInc = curCost;
finalChange = curChange;
minPos = blkPos;
}
}
if (dstCoeff[minPos] == 32767 || dstCoeff[minPos] == -32768)
/* don't allow sign hiding to violate the SPEC range */
finalChange = -1;
if (dstCoeff[minPos] == 0)
numSig++;
else if (finalChange == -1 && abs(dstCoeff[minPos]) == 1)
numSig--;
if (m_resiDctCoeff[minPos] >= 0)
dstCoeff[minPos] += finalChange;
else
dstCoeff[minPos] -= finalChange;
}
}
lastCG = false;
}
}
return numSig;
}
/* Context derivation process of coeff_abs_significant_flag */
uint32_t Quant::getSigCtxInc(uint32_t patternSigCtx, uint32_t log2TrSize, uint32_t trSize, uint32_t blkPos, bool bIsLuma,
uint32_t firstSignificanceMapContext)
{
static const uint8_t ctxIndMap[16] =
{
0, 1, 4, 5,
2, 3, 4, 5,
6, 6, 8, 8,
7, 7, 8, 8
};
if (!blkPos) // special case for the DC context variable
return 0;
if (log2TrSize == 2) // 4x4
return ctxIndMap[blkPos];
const uint32_t posY = blkPos >> log2TrSize;
const uint32_t posX = blkPos & (trSize - 1);
X265_CHECK((blkPos - (posY << log2TrSize)) == posX, "block pos check failed\n");
int posXinSubset = blkPos & 3;
X265_CHECK((posX & 3) == (blkPos & 3), "pos alignment fail\n");
int posYinSubset = posY & 3;
// NOTE: [patternSigCtx][posXinSubset][posYinSubset]
static const uint8_t table_cnt[4][4][4] =
{
// patternSigCtx = 0
{
{ 2, 1, 1, 0 },
{ 1, 1, 0, 0 },
{ 1, 0, 0, 0 },
{ 0, 0, 0, 0 },
},
// patternSigCtx = 1
{
{ 2, 1, 0, 0 },
{ 2, 1, 0, 0 },
{ 2, 1, 0, 0 },
{ 2, 1, 0, 0 },
},
// patternSigCtx = 2
{
{ 2, 2, 2, 2 },
{ 1, 1, 1, 1 },
{ 0, 0, 0, 0 },
{ 0, 0, 0, 0 },
},
// patternSigCtx = 3
{
{ 2, 2, 2, 2 },
{ 2, 2, 2, 2 },
{ 2, 2, 2, 2 },
{ 2, 2, 2, 2 },
}
};
int cnt = table_cnt[patternSigCtx][posXinSubset][posYinSubset];
int offset = firstSignificanceMapContext;
offset += cnt;
return (bIsLuma && (posX | posY) >= 4) ? 3 + offset : offset;
}
扫一扫
1万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
