From 4a3fb692430ba602a75a144d6eae8cd6d83d5d0c Mon Sep 17 00:00:00 2001
From: Ted <jhuhong-jheng@kwai.com>
Date: Thu, 6 Feb 2020 23:50:44 +0800
Subject: [PATCH] JVET-Q0493: Palette encoder improvements for lossless coding
Detail
- Use correct precision of distortion based on input bit-depth during the derivation of the palette for lossless coding.
- Remove the unnecessary calculation of distortion during palette coding RD decision on lossless.
---
source/Lib/CommonLib/TypeDef.h | 2 +
source/Lib/EncoderLib/IntraSearch.cpp | 161 ++++++++++++++++++++++++++
source/Lib/EncoderLib/IntraSearch.h | 34 ++++++
3 files changed, 197 insertions(+)
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index 209f4f6a4..fefa0ba8e 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -50,6 +50,8 @@
#include <assert.h>
#include <cassert>
+#define JVET_Q0493_PLT_ENCODER_LOSSLESS 1 // JVET-Q0493: Palette encoder improvements for lossless coding
+
#define JVET_Q0629_REMOVAL_PLT_4X4 1 // JVET-Q0629: Removal of 4x4 blocks in palette mode
#define JVET_Q0291_REDUCE_DUALTREE_PLT_SIZE 1 // JVET-Q0291: reduce palette size of dual tree from 31 to 15 and palette predictor size of dual tree from 63 to 31
diff --git a/source/Lib/EncoderLib/IntraSearch.cpp b/source/Lib/EncoderLib/IntraSearch.cpp
index 617856537..9cdc5d8c7 100644
--- a/source/Lib/EncoderLib/IntraSearch.cpp
+++ b/source/Lib/EncoderLib/IntraSearch.cpp
@@ -1791,6 +1791,9 @@ void IntraSearch::preCalcPLTIndexRD(CodingStructure& cs, Partitioner& partitione
CodingUnit &cu = *cs.getCU(partitioner.chType);
uint32_t height = cu.block(compBegin).height;
uint32_t width = cu.block(compBegin).width;
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ bool lossless = (m_pcEncCfg->getCostMode() == COST_LOSSLESS_CODING);
+#endif
CPelBuf orgBuf[3];
for (int comp = compBegin; comp < (compBegin + numComp); comp++)
@@ -1827,9 +1830,33 @@ void IntraSearch::preCalcPLTIndexRD(CodingStructure& cs, Partitioner& partitione
uint8_t pltIdx = 0;
double minError = MAX_DOUBLE;
uint8_t bestIdx = 0;
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ for (uint8_t z = 0; z < cu.curPLTSize[compBegin]; z++)
+ {
+ m_indexError[z][rasPos] = minError;
+ }
+#endif
while (pltIdx < cu.curPLTSize[compBegin])
{
uint64_t sqrtError = 0;
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (lossless)
+ {
+ for (int comp = compBegin; comp < (discardChroma ? 1 : (compBegin + numComp)); comp++)
+ {
+ sqrtError += int64_t(abs(curPel[comp] - cu.curPLT[comp][pltIdx]));
+ }
+ if (sqrtError == 0)
+ {
+ m_indexError[pltIdx][rasPos] = (double) sqrtError;
+ minError = (double) sqrtError;
+ bestIdx = pltIdx;
+ break;
+ }
+ }
+ else
+ {
+#endif
for (int comp = compBegin; comp < (discardChroma ? 1 : (compBegin + numComp)); comp++)
{
int64_t tmpErr = int64_t(curPel[comp] - cu.curPLT[comp][pltIdx]);
@@ -1848,14 +1875,32 @@ void IntraSearch::preCalcPLTIndexRD(CodingStructure& cs, Partitioner& partitione
minError = (double)sqrtError;
bestIdx = pltIdx;
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
pltIdx++;
}
Pel paPixelValue[3], paRecoValue[3];
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (!lossless)
+ {
+#endif
calcPixelPredRD(cs, partitioner, curPel, paPixelValue, paRecoValue, compBegin, numComp);
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
uint64_t error = 0, rate = 0;
for (int comp = compBegin; comp < (discardChroma ? 1 : (compBegin + numComp)); comp++)
{
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (lossless)
+ {
+ rate += m_escapeNumBins[curPel[comp]];
+ }
+ else
+ {
+#endif
int64_t tmpErr = int64_t(curPel[comp] - paRecoValue[comp]);
if (isChroma((ComponentID)comp))
{
@@ -1866,6 +1911,9 @@ void IntraSearch::preCalcPLTIndexRD(CodingStructure& cs, Partitioner& partitione
error += tmpErr*tmpErr;
}
rate += m_escapeNumBins[paPixelValue[comp]]; // encode quantized escape color
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
}
double rdCost = (double)error + m_pcRdCost->getLambda()*(double)rate;
m_indexError[cu.curPLTSize[compBegin]][rasPos] = rdCost;
@@ -2291,6 +2339,9 @@ void IntraSearch::calcPixelPred(CodingStructure& cs, Partitioner& partitioner, u
{
CodingUnit &cu = *cs.getCU(partitioner.chType);
TransformUnit &tu = *cs.getTU(partitioner.chType);
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ bool lossless = (m_pcEncCfg->getCostMode() == COST_LOSSLESS_CODING);
+#endif
CPelBuf orgBuf[3];
for (int comp = compBegin; comp < (compBegin + numComp); comp++)
@@ -2314,6 +2365,10 @@ void IntraSearch::calcPixelPred(CodingStructure& cs, Partitioner& partitioner, u
int rightShiftOffset[3];
int invquantiserRightShift[3];
int add[3];
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (!lossless)
+ {
+#endif
for (uint32_t ch = compBegin; ch < (compBegin + numComp); ch++)
{
QpParam cQP(tu, ComponentID(ch));
@@ -2326,6 +2381,9 @@ void IntraSearch::calcPixelPred(CodingStructure& cs, Partitioner& partitioner, u
invquantiserRightShift[ch] = IQUANT_SHIFT;
add[ch] = 1 << (invquantiserRightShift[ch] - 1);
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
uint32_t scaleX = getComponentScaleX(COMPONENT_Cb, cs.sps->getChromaFormatIdc());
uint32_t scaleY = getComponentScaleY(COMPONENT_Cb, cs.sps->getChromaFormatIdc());
@@ -2337,19 +2395,43 @@ void IntraSearch::calcPixelPred(CodingStructure& cs, Partitioner& partitioner, u
PLTescapeBuf escapeValue = tu.getescapeValue((ComponentID)ch);
if (compBegin != COMPONENT_Y || ch == 0)
{
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (lossless)
+ {
+ escapeValue.at(xPos, yPos) = orgBuf[ch].at(xPos, yPos);
+ recBuf.at(xPos, yPos) = escapeValue.at(xPos, yPos);
+ }
+ else
+ {
+#endif
escapeValue.at(xPos, yPos) = TCoeff(std::max<int>(0, ((orgBuf[ch].at(xPos, yPos) * quantiserScale[ch] + rightShiftOffset[ch]) >> quantiserRightShift[ch])));
assert(escapeValue.at(xPos, yPos) < (1 << (channelBitDepth + 1)));
recBuf.at(xPos, yPos) = (((escapeValue.at(xPos, yPos)*g_invQuantScales[0][qpRem[ch]]) << qpPer[ch]) + add[ch]) >> invquantiserRightShift[ch];
recBuf.at(xPos, yPos) = Pel(ClipBD<int>(recBuf.at(xPos, yPos), channelBitDepth));//to be checked
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
}
else if (compBegin == COMPONENT_Y && ch > 0 && yPos % (1 << scaleY) == 0 && xPos % (1 << scaleX) == 0)
{
uint32_t yPosC = yPos >> scaleY;
uint32_t xPosC = xPos >> scaleX;
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (lossless)
+ {
+ escapeValue.at(xPosC, yPosC) = orgBuf[ch].at(xPosC, yPosC);
+ recBuf.at(xPosC, yPosC) = escapeValue.at(xPosC, yPosC);
+ }
+ else
+ {
+#endif
escapeValue.at(xPosC, yPosC) = TCoeff(std::max<int>(0, ((orgBuf[ch].at(xPosC, yPosC) * quantiserScale[ch] + rightShiftOffset[ch]) >> quantiserRightShift[ch])));
assert(escapeValue.at(xPosC, yPosC) < (1 << (channelBitDepth + 1)));
recBuf.at(xPosC, yPosC) = (((escapeValue.at(xPosC, yPosC)*g_invQuantScales[0][qpRem[ch]]) << qpPer[ch]) + add[ch]) >> invquantiserRightShift[ch];
recBuf.at(xPosC, yPosC) = Pel(ClipBD<int>(recBuf.at(xPosC, yPosC), channelBitDepth));//to be checked
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
}
}
}
@@ -2358,8 +2440,19 @@ void IntraSearch::derivePLTLossy(CodingStructure& cs, Partitioner& partitioner,
CodingUnit &cu = *cs.getCU(partitioner.chType);
const int channelBitDepth_L = cs.sps->getBitDepth(CHANNEL_TYPE_LUMA);
const int channelBitDepth_C = cs.sps->getBitDepth(CHANNEL_TYPE_CHROMA);
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ bool lossless = (m_pcEncCfg->getCostMode() == COST_LOSSLESS_CODING);
+ int pcmShiftRight_L = (channelBitDepth_L - PLT_ENCBITDEPTH);
+ int pcmShiftRight_C = (channelBitDepth_C - PLT_ENCBITDEPTH);
+ if (lossless)
+ {
+ pcmShiftRight_L = 0;
+ pcmShiftRight_C = 0;
+ }
+#else
const int pcmShiftRight_L = (channelBitDepth_L - PLT_ENCBITDEPTH);
const int pcmShiftRight_C = (channelBitDepth_C - PLT_ENCBITDEPTH);
+#endif
#if JVET_Q0291_REDUCE_DUALTREE_PLT_SIZE
int maxPltSize = cu.isSepTree() ? MAXPLTSIZE_DUALTREE : MAXPLTSIZE;
@@ -2390,6 +2483,12 @@ void IntraSearch::derivePLTLossy(CodingStructure& cs, Partitioner& partitioner,
int errorLimit = g_paletteQuant[qp];
#else
int errorLimit = g_paletteQuant[cu.qp];
+#endif
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (lossless)
+ {
+ errorLimit = 0;
+ }
#endif
uint32_t totalSize = height*width;
SortingElement *pelList = new SortingElement[totalSize];
@@ -2428,12 +2527,38 @@ void IntraSearch::derivePLTLossy(CodingStructure& cs, Partitioner& partitioner,
tmpCompBegin = COMPONENT_Y;
tmpNumComp = 1;
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ int besti = last, bestSAD = (last == -1) ? MAX_UINT : pelList[last].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless);
+ if (lossless)
+ {
+ if (bestSAD)
+ {
+ for (int i = idx - 1; i >= 0; i--)
+ {
+ uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless);
+ if (sad == 0)
+ {
+ bestSAD = sad;
+ besti = i;
+ break;
+ }
+ }
+ }
+ }
+ else
+ {
+#else
int besti = last, bestSAD = (last == -1) ? MAX_UINT : pelList[last].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp);
+#endif
if( bestSAD )
{
for (int i = idx - 1; i >= 0; i--)
{
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless);
+#else
uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp);
+#endif
if (sad < bestSAD)
{
bestSAD = sad;
@@ -2442,7 +2567,12 @@ void IntraSearch::derivePLTLossy(CodingStructure& cs, Partitioner& partitioner,
}
}
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+ if (besti >= 0 && pelList[besti].almostEqualData(element, errorLimit, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp, lossless))
+#else
if (besti >= 0 && pelList[besti].almostEqualData(element, errorLimit, cs.sps->getBitDepths(), tmpCompBegin, tmpNumComp))
+#endif
{
pelList[besti].addElement(element, tmpCompBegin, tmpNumComp);
last = besti;
@@ -2498,12 +2628,38 @@ void IntraSearch::derivePLTLossy(CodingStructure& cs, Partitioner& partitioner,
org[comp] = orgBuf[comp].at(pX, pY);
}
element.setAll(org, compBegin, numComp);
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ int besti = last, bestSAD = (last == -1) ? MAX_UINT : pelList[last].getSAD(element, cs.sps->getBitDepths(), compBegin, numComp, lossless);
+ if (lossless)
+ {
+ if (bestSAD)
+ {
+ for (int i = idx - 1; i >= 0; i--)
+ {
+ uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), compBegin, numComp, lossless);
+ if (sad == 0)
+ {
+ bestSAD = sad;
+ besti = i;
+ break;
+ }
+ }
+ }
+ }
+ else
+ {
+#else
int besti = last, bestSAD = (last == -1) ? MAX_UINT : pelList[last].getSAD(element, cs.sps->getBitDepths(), compBegin, numComp);
+#endif
if (bestSAD)
{
for (int i = idx - 1; i >= 0; i--)
{
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), compBegin, numComp, lossless);
+#else
uint32_t sad = pelList[i].getSAD(element, cs.sps->getBitDepths(), compBegin, numComp);
+#endif
if (sad < bestSAD)
{
bestSAD = sad;
@@ -2512,7 +2668,12 @@ void IntraSearch::derivePLTLossy(CodingStructure& cs, Partitioner& partitioner,
}
}
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+ if (besti >= 0 && pelList[besti].almostEqualData(element, errorLimit, cs.sps->getBitDepths(), compBegin, numComp, lossless))
+#else
if (besti >= 0 && pelList[besti].almostEqualData(element, errorLimit, cs.sps->getBitDepths(), compBegin, numComp))
+#endif
{
pelList[besti].addElement(element, compBegin, numComp);
last = besti;
diff --git a/source/Lib/EncoderLib/IntraSearch.h b/source/Lib/EncoderLib/IntraSearch.h
index da19e57f2..190eff53b 100644
--- a/source/Lib/EncoderLib/IntraSearch.h
+++ b/source/Lib/EncoderLib/IntraSearch.h
@@ -112,11 +112,27 @@ public:
data[ch] = ui[ch];
}
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ bool almostEqualData(SortingElement element, int errorLimit, const BitDepths& bitDepths, ComponentID compBegin, uint32_t numComp, bool lossless)
+#else
bool almostEqualData(SortingElement element, int errorLimit, const BitDepths& bitDepths, ComponentID compBegin, uint32_t numComp)
+#endif
{
bool almostEqual = true;
for (int comp = compBegin; comp < (compBegin + numComp); comp++)
{
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (lossless)
+ {
+ if ((std::abs(data[comp] - element.data[comp])) > errorLimit)
+ {
+ almostEqual = false;
+ break;
+ }
+ }
+ else
+ {
+#endif
uint32_t absError = 0;
if (isChroma((ComponentID) comp))
{
@@ -131,16 +147,34 @@ public:
almostEqual = false;
break;
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
}
return almostEqual;
}
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ uint32_t getSAD(SortingElement element, const BitDepths &bitDepths, ComponentID compBegin, uint32_t numComp, bool lossless)
+#else
uint32_t getSAD(SortingElement element, const BitDepths& bitDepths, ComponentID compBegin, uint32_t numComp)
+#endif
{
uint32_t sumAd = 0;
for (int comp = compBegin; comp < (compBegin + numComp); comp++)
{
ChannelType chType = (comp > 0) ? CHANNEL_TYPE_CHROMA : CHANNEL_TYPE_LUMA;
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ if (lossless)
+ {
+ sumAd += (std::abs(data[comp] - element.data[comp]));
+ }
+ else
+ {
+#endif
sumAd += (std::abs(data[comp] - element.data[comp]) >> (bitDepths.recon[chType] - PLT_ENCBITDEPTH));
+#if JVET_Q0493_PLT_ENCODER_LOSSLESS
+ }
+#endif
}
return sumAd;
}
--
GitLab