diff --git a/source/App/EncoderApp/EncApp.cpp b/source/App/EncoderApp/EncApp.cpp
index a2881df31554040e244d372d2ca78c24a3912e29..6713cb839e84e1f97994f5ff32603b8f670e9489 100644
--- a/source/App/EncoderApp/EncApp.cpp
+++ b/source/App/EncoderApp/EncApp.cpp
@@ -1256,6 +1256,9 @@ void EncApp::xInitLibCfg()
#endif
#if JVET_AH0057_CCALF_COEFF_PRECISION
m_cEncLib.setUseCCALFPrecision ( m_ccalfPrecision );
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_cEncLib.setAlfLumaFixedFilterAdjust ( m_alfLumaFixedFilterAdjust );
#endif
m_cEncLib.setTestSAODisableAtPictureLevel ( m_bTestSAODisableAtPictureLevel );
m_cEncLib.setSaoEncodingRate ( m_saoEncodingRate );
diff --git a/source/App/EncoderApp/EncAppCfg.cpp b/source/App/EncoderApp/EncAppCfg.cpp
index 737460da0d59fcf4437eb1a11f688ece2af21650..49aa88e1dbec1b60d4ee66748bfa6f026006bc3d 100644
--- a/source/App/EncoderApp/EncAppCfg.cpp
+++ b/source/App/EncoderApp/EncAppCfg.cpp
@@ -1567,6 +1567,9 @@ bool EncAppCfg::parseCfg( int argc, char* argv[] )
#endif
#if JVET_AH0057_CCALF_COEFF_PRECISION
("CCALFPrecision", m_ccalfPrecision, true, "Cross-component Alf with variable precision coefficients")
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ ("AlfLumaFixedFilterAdjust", m_alfLumaFixedFilterAdjust, true, "Alf Luma Fixed Filter Adjustment" )
#endif
("TestSAODisableAtPictureLevel", m_bTestSAODisableAtPictureLevel, false, "Enables the testing of disabling SAO at the picture level after having analysed all blocks")
("SaoEncodingRate", m_saoEncodingRate, 0.75, "When >0 SAO early picture termination is enabled for luma and chroma")
diff --git a/source/App/EncoderApp/EncAppCfg.h b/source/App/EncoderApp/EncAppCfg.h
index 9a478c16a0901ac17536b60ee3dbf14f38c0769d..2b1b81859dbbe74e8f9f24fbd01cc41e9871808d 100644
--- a/source/App/EncoderApp/EncAppCfg.h
+++ b/source/App/EncoderApp/EncAppCfg.h
@@ -756,6 +756,9 @@ protected:
#endif
#if JVET_AH0057_CCALF_COEFF_PRECISION
bool m_ccalfPrecision;
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ bool m_alfLumaFixedFilterAdjust;
#endif
bool m_bTestSAODisableAtPictureLevel;
double m_saoEncodingRate; ///< When >0 SAO early picture termination is enabled for luma and chroma
diff --git a/source/Lib/CommonLib/AdaptiveLoopFilter.cpp b/source/Lib/CommonLib/AdaptiveLoopFilter.cpp
index 4a8a3c36089e9cc4441f4dead5b38e76c28f1669..ceaa4d3d184ca1d214f7bc3878b571862737a2b8 100644
--- a/source/Lib/CommonLib/AdaptiveLoopFilter.cpp
+++ b/source/Lib/CommonLib/AdaptiveLoopFilter.cpp
@@ -64,6 +64,12 @@ AdaptiveLoopFilter::AdaptiveLoopFilter()
{
m_classifier[i] = nullptr;
}
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ for( int i = 0; i < 1; i++ )
+ {
+ m_classifierCodingInfo[i] = nullptr;
+ }
#endif
for (size_t i = 0; i < NUM_DIRECTIONS; i++)
{
@@ -163,6 +169,10 @@ AdaptiveLoopFilter::AdaptiveLoopFilter()
#else
m_deriveClassificationBlk = deriveClassificationBlk;
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_textureClassMapping = textureClassMapping;
+ m_calcAlfLumaCodingInfoBlk = calcAlfLumaCodingInfoBlk;
+#endif
#if ENABLE_SIMD_OPT_ALF
#ifdef TARGET_SIMD_X86
@@ -671,6 +681,9 @@ void AdaptiveLoopFilter::ALFProcess(CodingStructure& cs)
#endif
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
memset(m_ctuPadFlag, 0, sizeof(uint8_t) * m_numCTUsInPic);
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf tmpYuvCodingInfo = m_tempBufCodingInfo.getBuf( cs.area );
#endif
const PreCalcValues& pcv = *cs.pcv;
@@ -769,6 +782,23 @@ void AdaptiveLoopFilter::ALFProcess(CodingStructure& cs)
#endif
buf = buf.subBuf( UnitArea( cs.area.chromaFormat, Area( clipL ? 0 : MAX_ALF_PADDING_SIZE, clipT ? 0 : MAX_ALF_PADDING_SIZE, w, h ) ) );
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf bufCodingInfo = m_tempBufCodingInfo2.subBuf( UnitArea( CHROMA_400, Area( 0, 0, wBuf, hBuf ) ) );
+ bufCodingInfo.copyFrom( tmpYuvCodingInfo.subBuf( UnitArea( CHROMA_400, Area( xStart - ( clipL ? 0 : MAX_ALF_PADDING_SIZE ), yStart - ( clipT ? 0 : MAX_ALF_PADDING_SIZE ), wBuf, hBuf ) ) ) );
+ // pad top-left unavailable samples for raster slice
+ if( xStart == xPos && yStart == yPos && ( rasterSliceAlfPad & 1 ) )
+ {
+ bufCodingInfo.padBorderPel( MAX_ALF_PADDING_SIZE, 1 );
+ }
+
+ // pad bottom-right unavailable samples for raster slice
+ if( xEnd == xPos + width && yEnd == yPos + height && ( rasterSliceAlfPad & 2 ) )
+ {
+ bufCodingInfo.padBorderPel( MAX_ALF_PADDING_SIZE, 2 );
+ }
+ mirroredPaddingForAlf(cs, bufCodingInfo, MAX_ALF_PADDING_SIZE, true, false);
+ bufCodingInfo = bufCodingInfo.subBuf( UnitArea( CHROMA_400, Area( clipL ? 0 : MAX_ALF_PADDING_SIZE, clipT ? 0 : MAX_ALF_PADDING_SIZE, w, h ) ) );
+#endif
#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
#if JVET_AF0197_LUMA_RESIDUAL_TAP_IN_CCALF
PelUnitBuf bufResi = m_tempBufResi2.subBuf(UnitArea(CHROMA_400, Area(0, 0, wBuf, hBuf)));
@@ -817,6 +847,9 @@ void AdaptiveLoopFilter::ALFProcess(CodingStructure& cs)
bufDb.extendBorderPel(NUM_DB_PAD);
bufDb = bufDb.subBuf(UnitArea(CHROMA_400, Area(clipL ? 0 : NUM_DB_PAD, clipT ? 0 : NUM_DB_PAD, w, h)));
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ calcAlfLumaCodingInfoBlk(cs, m_classifierCodingInfo[0], blkDst, blkSrc, buf.get(COMPONENT_Y), 2, 2, m_inputBitDepth[CHANNEL_TYPE_LUMA], bufResi.get(COMPONENT_Y), m_laplacian[0], bufCodingInfo.get(COMPONENT_Y) );
+#endif
#if JVET_X0071_ALF_BAND_CLASSIFIER
deriveClassification( m_classifier, buf.get(COMPONENT_Y),
#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
@@ -1213,6 +1246,9 @@ void AdaptiveLoopFilter::ALFProcess(CodingStructure& cs)
{
Area blk( xPos, yPos, width, height );
short filterSetIndex = alfCtuFilterIndex[ctuIdx];
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ calcAlfLumaCodingInfoBlk(cs, m_classifierCodingInfo[0], blk, blk, recYuv.get(COMPONENT_Y), 2, 2, m_inputBitDepth[CHANNEL_TYPE_LUMA], tmpYuvResi.get(COMPONENT_Y), m_laplacian[0], tmpYuvCodingInfo.get(COMPONENT_Y) );
+#endif
#if JVET_X0071_ALF_BAND_CLASSIFIER
deriveClassification( m_classifier, tmpYuv.get(COMPONENT_Y),
#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
@@ -1232,13 +1268,21 @@ void AdaptiveLoopFilter::ALFProcess(CodingStructure& cs)
if( filterSetIndex != 0 )
{
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
- deriveFixedFilterResultsCtuBoundary( m_classifier, m_fixFilterResult[COMPONENT_Y], tmpYuv.get( COMPONENT_Y ), tmpYuvBeforeDb.get( COMPONENT_Y ), blk, m_inputBitDepth[CHANNEL_TYPE_LUMA], cs, m_clpRngs.comp[COMPONENT_Y], m_alfClippingValues[CHANNEL_TYPE_LUMA], cs.slice->getSliceQp(), fixedFilterSetIdx, m_mappingDir, m_laplacian, m_ctuEnableFlag[COMPONENT_Y], m_ctuEnableOnlineLumaFlag, ctuIdx, 0 );
+ deriveFixedFilterResultsCtuBoundary( m_classifier, m_fixFilterResult[COMPONENT_Y], tmpYuv.get( COMPONENT_Y ), tmpYuvBeforeDb.get( COMPONENT_Y ), blk, m_inputBitDepth[CHANNEL_TYPE_LUMA], cs, m_clpRngs.comp[COMPONENT_Y], m_alfClippingValues[CHANNEL_TYPE_LUMA], cs.slice->getSliceQp(), fixedFilterSetIdx, m_mappingDir, m_laplacian, m_ctuEnableFlag[COMPONENT_Y], m_ctuEnableOnlineLumaFlag, ctuIdx, 0
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , tmpYuvCodingInfo.get(COMPONENT_Y), tmpYuvResi.get( COMPONENT_Y )
+#endif
+ );
#else
deriveFixedFilterResultsCtuBoundary( m_classifier, m_fixFilterResult, tmpYuv.get( COMPONENT_Y ), tmpYuvBeforeDb.get( COMPONENT_Y ), blk, m_inputBitDepth[CHANNEL_TYPE_LUMA], cs, m_clpRngs.comp[COMPONENT_Y], m_alfClippingValues[CHANNEL_TYPE_LUMA], cs.slice->getSliceQp(), fixedFilterSetIdx, m_mappingDir, m_laplacian, m_ctuEnableFlag[COMPONENT_Y], m_ctuEnableOnlineLumaFlag, ctuIdx, 0 );
#endif
deriveFixedFilterResults( m_classifier, tmpYuv.get( COMPONENT_Y ), m_tempBufBeforeDb.get( COMPONENT_Y ), blk, blk, cs, 1, fixedFilterSetIdx );
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
- deriveFixedFilterResultsCtuBoundary( m_classifier, m_fixFilterResult[COMPONENT_Y], tmpYuv.get( COMPONENT_Y ), tmpYuvBeforeDb.get( COMPONENT_Y ), blk, m_inputBitDepth[CHANNEL_TYPE_LUMA], cs, m_clpRngs.comp[COMPONENT_Y], m_alfClippingValues[CHANNEL_TYPE_LUMA], cs.slice->getSliceQp(), fixedFilterSetIdx, m_mappingDir, m_laplacian, m_ctuEnableFlag[COMPONENT_Y], m_ctuEnableOnlineLumaFlag, ctuIdx, 1 );
+ deriveFixedFilterResultsCtuBoundary( m_classifier, m_fixFilterResult[COMPONENT_Y], tmpYuv.get( COMPONENT_Y ), tmpYuvBeforeDb.get( COMPONENT_Y ), blk, m_inputBitDepth[CHANNEL_TYPE_LUMA], cs, m_clpRngs.comp[COMPONENT_Y], m_alfClippingValues[CHANNEL_TYPE_LUMA], cs.slice->getSliceQp(), fixedFilterSetIdx, m_mappingDir, m_laplacian, m_ctuEnableFlag[COMPONENT_Y], m_ctuEnableOnlineLumaFlag, ctuIdx, 1
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , tmpYuvCodingInfo.get(COMPONENT_Y), tmpYuvResi.get( COMPONENT_Y )
+#endif
+ );
#else
deriveFixedFilterResultsCtuBoundary( m_classifier, m_fixFilterResult, tmpYuv.get( COMPONENT_Y ), tmpYuvBeforeDb.get( COMPONENT_Y ), blk, m_inputBitDepth[CHANNEL_TYPE_LUMA], cs, m_clpRngs.comp[COMPONENT_Y], m_alfClippingValues[CHANNEL_TYPE_LUMA], cs.slice->getSliceQp(), fixedFilterSetIdx, m_mappingDir, m_laplacian, m_ctuEnableFlag[COMPONENT_Y], m_ctuEnableOnlineLumaFlag, ctuIdx, 1 );
#endif
@@ -1842,6 +1886,12 @@ void AdaptiveLoopFilter::create(const int picWidth, const int picHeight, const C
m_tempBufSAO2.destroy();
m_tempBufSAO2.create(format, Area(0, 0, maxCUWidth + (MAX_ALF_PADDING_SIZE << 1), maxCUHeight + (MAX_ALF_PADDING_SIZE << 1)), maxCUWidth, MAX_ALF_PADDING_SIZE, 0, false);
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_tempBufCodingInfo.destroy();
+ m_tempBufCodingInfo.create(CHROMA_400, Area(0, 0, picWidth, picHeight), maxCUWidth, MAX_FILTER_LENGTH_FIXED, 0, false);
+ m_tempBufCodingInfo2.destroy();
+ m_tempBufCodingInfo2.create(CHROMA_400, Area( 0, 0, maxCUWidth + (MAX_ALF_PADDING_SIZE << 1), maxCUHeight + (MAX_ALF_PADDING_SIZE << 1) ), maxCUWidth, MAX_ALF_PADDING_SIZE, 0, false );
+#endif
#if ALF_IMPROVEMENT
int numFixedFilters = EXT_LENGTH << 1;
@@ -2142,6 +2192,21 @@ void AdaptiveLoopFilter::create(const int picWidth, const int picHeight, const C
}
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ for( int classifier = 0; classifier < 1; classifier++ )
+ {
+ if( m_classifierCodingInfo[classifier] == nullptr )
+ {
+ m_classifierCodingInfo[classifier] = new AlfClassifier*[picHeight];
+ m_classifierCodingInfo[classifier][0] = new AlfClassifier[picWidth * picHeight];
+
+ for( int i = 1; i < picHeight; i++ )
+ {
+ m_classifierCodingInfo[classifier][i] = m_classifierCodingInfo[classifier][0] + i * picWidth;
+ }
+ }
+ }
+#endif
#if !ALF_IMPROVEMENT
for (int filterSetIndex = 0; filterSetIndex < NUM_FIXED_FILTER_SETS; filterSetIndex++)
{
@@ -2199,6 +2264,17 @@ void AdaptiveLoopFilter::destroy()
m_classifier = nullptr;
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ for (int classifier = 0; classifier < 1; classifier++)
+ {
+ if (m_classifierCodingInfo[classifier] )
+ {
+ delete[] m_classifierCodingInfo[classifier][0];
+ delete[] m_classifierCodingInfo[classifier];
+ m_classifierCodingInfo[classifier] = nullptr;
+ }
+ }
+#endif
#if ALF_IMPROVEMENT
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
@@ -2391,6 +2467,10 @@ void AdaptiveLoopFilter::destroy()
#if JVET_AI0166_CCALF_CHROMA_SAO_INPUT
m_tempBufSAO.destroy();
m_tempBufSAO2.destroy();
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_tempBufCodingInfo.destroy();
+ m_tempBufCodingInfo2.destroy();
#endif
m_filterShapes[CHANNEL_TYPE_LUMA].clear();
m_filterShapes[CHANNEL_TYPE_CHROMA].clear();
@@ -3127,7 +3207,11 @@ void AdaptiveLoopFilter::deriveFixFilterResultsBlkChroma(AlfClassifier ***classi
{
if (fixedFiltSetInd == targetFixedFilterSetInd || targetFixedFilterSetInd == -1)
{
- alfFixedFilterBlk(classifier[ALF_NUM_CLASSIFIER + 1], src, blk, blkDst, srcBeforeDb, fixedFilterResults, m_picWidth, fixedFiltInd, fixedFiltSetInd, 0, clpRng, clippingValues, false);
+ alfFixedFilterBlk(classifier[ALF_NUM_CLASSIFIER + 1], src, blk, blkDst, srcBeforeDb, fixedFilterResults, m_picWidth, fixedFiltInd, fixedFiltSetInd, 0, clpRng, clippingValues, false
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
}
fixedFiltInd++;
}
@@ -3485,54 +3569,89 @@ void AdaptiveLoopFilter::calcClass0Var( AlfClassifier **classifier, const Area &
{
for( int jj = posXDst + j; jj < posXDst + j + subBlkSize; jj++ )
{
+ classifier[ii][jj] = (actDirInd << 2) + transposeIdx;
+ }
+ }
#else
for( int ii = curBlk.y + i; ii < curBlk.y + i + subBlkSize; ii++ )
{
for( int jj = curBlk.x + j; jj < curBlk.x + j + subBlkSize; jj++ )
{
-#endif
classifier[ii][jj] = (actDirInd << 2) + transposeIdx;
}
}
+#endif
}
}
}
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
-void AdaptiveLoopFilter::alfFixedFilterBlkNonSimd( AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma )
+void AdaptiveLoopFilter::alfFixedFilterBlkNonSimd( AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , CodingStructure &cs
+#endif
+ )
{
if( isLuma )
{
if( dirWindSize == 0 )
{
- fixedFilterBlk<ALF_FIXED_FILTER_9_DB_9>( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues );
+ fixedFilterBlk<ALF_FIXED_FILTER_9_DB_9>( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , isLuma, cs, m_classifierCodingInfo[0]
+#endif
+ );
}
else
{
- fixedFilterBlk<ALF_FIXED_FILTER_13_DB_9>( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter13Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues );
+ fixedFilterBlk<ALF_FIXED_FILTER_13_DB_9>( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter13Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , isLuma, cs, m_classifierCodingInfo[0]
+#endif
+ );
}
}
else
{
- fixedFilterBlk<ALF_FIXED_FILTER_9_DB_9>(classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues);
+ fixedFilterBlk<ALF_FIXED_FILTER_9_DB_9>(classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , isLuma, cs, m_classifierCodingInfo[0]
+#endif
+ );
}
}
-void AdaptiveLoopFilter::alfFixedFilterBlk( AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma )
+void AdaptiveLoopFilter::alfFixedFilterBlk( AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , CodingStructure &cs
+#endif
+ )
{
if( isLuma )
{
if( dirWindSize == 0 )
{
- m_fixFilter9x9Db9Blk( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues );
+ m_fixFilter9x9Db9Blk( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , isLuma, cs, m_classifierCodingInfo[0]
+#endif
+ );
}
else
{
- m_fixFilter13x13Db9Blk( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter13Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues );
+ m_fixFilter13x13Db9Blk( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter13Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , isLuma, cs, m_classifierCodingInfo[0]
+#endif
+ );
}
}
else
{
- m_fixFilter9x9Db9Blk( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues );
+ m_fixFilter9x9Db9Blk( classifier, src, curBlk, blkDst, srcBeforeDb, fixedFilterResults, picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltQpInd], fixedFiltQpInd, dirWindSize, clpRng, clippingValues
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , isLuma, cs, m_classifierCodingInfo[0]
+#endif
+ );
}
}
#else
@@ -3562,7 +3681,11 @@ void AdaptiveLoopFilter::alfFixedFilterBlk( AlfClassifier **classifier, const CP
template<AlfFixedFilterType filtType>
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
-void AdaptiveLoopFilter::fixedFilterBlk( AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4] )
+void AdaptiveLoopFilter::fixedFilterBlk( AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ )
#else
void AdaptiveLoopFilter::fixedFilterBlk( AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4] )
#endif
@@ -3604,6 +3727,15 @@ void AdaptiveLoopFilter::fixedFilterBlk( AlfClassifier **classifier, const CPelB
const Pel *pImgYBeforeDbPad0 = srcLumaBeforeDb.buf + posY * srcBeforeDbStride + posX;
#endif
const int srcBeforeDbStride2 = srcBeforeDbStride * clsSizeY;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+ const bool useCodingInfo = true;
+
+ const bool useBounCondition = applyCodingInfo && !( !isSpsAdjust && isIntraSlice ) && useCodingInfo;
+ const bool useResiCondition = applyCodingInfo && (isSpsAdjust || !isSpsAdjust) && !isIntraSlice && useCodingInfo;
+ const int offsetClipValue = 1 << ( clpRng.bd - 1 );
+#endif
int fixedFiltIndF0 = -1;
int numCoeff;
@@ -3635,7 +3767,25 @@ void AdaptiveLoopFilter::fixedFilterBlk( AlfClassifier **classifier, const CPelB
int classIdx = classifier[posY + i][posX + j] >> 2;
int transposeIdx = classifier[posY + i][posX + j] & 0x3;
#endif
- int filterIdx = classIndFixed[classIdx];
+ int filterIdx = classIndFixed[classIdx];
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ int classIdxBs = 0;
+ if( useBounCondition )
+ {
+ //0: BS0 Resi0, 1:BS0 Resi1, 2:BS1 Resi0 3:BS1 Resi1
+ // A >> 1 = BS
+ // A - ( A >> 1) * 2
+ classIdxBs = classifierCodingInfo[posYDst + i][posXDst + j] >> 1;
+ }
+ int classIdxResi = 0;
+ if( useResiCondition )
+ {
+ //0: BS0 Resi0, 1:BS0 Resi1, 2:BS1 Resi0 3:BS1 Resi1
+ // A >> 1 = BS
+ // A - ( A >> 1) * 2
+ classIdxResi = classifierCodingInfo[posYDst + i][posXDst + j] - ((classifierCodingInfo[posYDst + i][posXDst + j] >> 1 ) * 2);
+ }
+#endif
const short* coeff;
const short* clipp;
@@ -3889,6 +4039,26 @@ void AdaptiveLoopFilter::fixedFilterBlk( AlfClassifier **classifier, const CPelB
CHECK( 1, "not supported" );
}
sum = ( sum + offset ) >> shift;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ sum = Clip3(-offsetClipValue, +offsetClipValue, sum);
+
+ int sign = sum < 0 ? -1 : +1;
+
+ int boundaryStrengthFactor = isIntraSlice ? 4 : 3;
+ sum = classIdxBs ? sign * ((abs(sum) * (16 + boundaryStrengthFactor) + 8 ) >> 4) : sum;
+ }
+
+ if( useResiCondition )
+ {
+ sum = Clip3(-offsetClipValue, +offsetClipValue, sum);
+
+ int sign = sum < 0 ? -1 : +1;
+ int resiStrengthFactor = isIntraSlice ? 0 >> ( !isSpsAdjust ? 1 : 0) : 3 >> (!isSpsAdjust ? 1 : 0);
+ sum = classIdxResi ? sign * ((abs(sum) * (16 + resiStrengthFactor) + 8 ) >> 4) : sum;
+ }
+#endif
sum += curr;
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
@@ -3918,6 +4088,9 @@ void AdaptiveLoopFilter::fixedFilterBlk( AlfClassifier **classifier, const CPelB
void AdaptiveLoopFilter::calcClassNew( AlfClassifier **classifier, const Area &blkDst, const Area &curBlk, const CPelBuf& srcLuma, int subBlkSize, AlfClassifier **classifier0, int classifierIdx, int bitDepth
#if JVET_AD0222_ALF_RESI_CLASS
, const CPelBuf& srcLumaResi, uint32_t **buffer
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , AlfClassifier ** classifierCodingInfo
#endif
)
{
@@ -3939,7 +4112,21 @@ void AdaptiveLoopFilter::calcClassNew( AlfClassifier **classifier, const Area &b
const Pel *pY0 = src0 + xOffset;
const Pel *pY1 = src1 + xOffset;
int sum = pY0[0] + pY0[1] + pY1[0] + pY1[1];
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ int boundShift = classifierCodingInfo[curBlk.y + i][curBlk.x + j ] >> 1;
+
+ int classIdx = 0;
+ if( boundShift == 0 )
+ {
+ classIdx = ( sum * 12 ) >> (bitDepth + 2);
+ }
+ else
+ {
+ classIdx = 12 + (( sum * 12 ) >> (bitDepth + 2));
+ }
+#else
int classIdx = (sum * ALF_NUM_CLASSES_CLASSIFIER[classifierIdx]) >> (bitDepth + 2);
+#endif
for (int ii = curBlk.y + i; ii < curBlk.y + i + subBlkSize; ii++)
{
for (int jj = curBlk.x + j; jj < curBlk.x + j + subBlkSize; jj++)
@@ -3996,11 +4183,34 @@ void AdaptiveLoopFilter::calcClassNew( AlfClassifier **classifier, const Area &b
int j2 = j >> 1;
int sum = buffer[i2][j2] + buffer[i2][j2 + 2] + buffer[i2 + 2][j2] + buffer[i2 + 2][j2 + 2];
int shiftOffset = ALF_RESI_SHIFT_OFFSET;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ int boundShift = classifierCodingInfo[curBlk.y + i][curBlk.x + j] >> 1;
+
+ int classIdx = sum >> (bitDepth - shiftOffset);
+ // Merge Neighbor Class, Then Clip
+ classIdx >>= 1;
+ if( boundShift == 0)
+ {
+ if( classIdx > 12 - 1)
+ {
+ classIdx = 12 - 1;
+ }
+ }
+ else
+ {
+ if( classIdx > 12 - 1)
+ {
+ classIdx = 12 - 1;
+ }
+ classIdx += 12;
+ }
+#else
int classIdx = sum >> (bitDepth - shiftOffset);
if (classIdx > 24)
{
classIdx = 24;
}
+#endif
for (int ii = curBlk.y + i; ii < curBlk.y + i + subBlkSize; ii++)
{
for (int jj = curBlk.x + j; jj < curBlk.x + j + subBlkSize; jj++)
@@ -4223,15 +4433,18 @@ void AdaptiveLoopFilter::calcClass(AlfClassifier **classifier, const Area &blkDs
{
for (int jj = posXDst + j; jj < posXDst + j + subBlkSize; jj++)
{
+ classifier[ii][jj] = (actDirInd << 2) + transposeIdx;
+ }
+ }
#else
for (int ii = curBlk.y + i; ii < curBlk.y + i + subBlkSize; ii++)
{
for (int jj = curBlk.x + j; jj < curBlk.x + j + subBlkSize; jj++)
{
-#endif
classifier[ii][jj] = (actDirInd << 2) + transposeIdx;
}
}
+#endif
}
}
}
@@ -4411,7 +4624,11 @@ void AdaptiveLoopFilter::fixedFilteringResi(AlfClassifier **classifier, const CP
#endif
Pel ***fixedFilterResiResults, int picWidth, const int fixedFiltInd,
const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd,
- int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4])
+ int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ )
{
const int shift = m_NUM_BITS_FIXED_FILTER - 1;
const int offset = 1 << (shift - 1);
@@ -4432,6 +4649,14 @@ void AdaptiveLoopFilter::fixedFilteringResi(AlfClassifier **classifier, const CP
const int clsSizeY = 2;
const int clsSizeX = 2;
const int srcStride2 = srcStride * clsSizeY;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+ const bool useCodingInfo = isSpsAdjust ? true : false;
+ const bool useBounCondition = applyCodingInfo && !( !isSpsAdjust && isIntraSlice ) && useCodingInfo;
+ const bool useResiCondition = applyCodingInfo && (isSpsAdjust || !isSpsAdjust) && !isIntraSlice && useCodingInfo;
+ const int offsetClipValue = 1 << ( clpRng.bd - 1 );
+#endif
for (int i = 0; i < curBlk.height; i += clsSizeY)
{
@@ -4445,6 +4670,25 @@ void AdaptiveLoopFilter::fixedFilteringResi(AlfClassifier **classifier, const CP
int transposeIdx = classifier[posY + i][posX + j] & 0x3;
#endif
int filterIdx = classIndFixed[classIdx];
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ int classIdxBs = 0;
+ if( useBounCondition )
+ {
+ //0: BS0 Resi0, 1:BS0 Resi1, 2:BS1 Resi0 3:BS1 Resi1
+ // A >> 1 = BS
+ // A - ( A >> 1) * 2
+ classIdxBs = classifierCodingInfo[posYDst + i][posXDst + j] >> 1;
+ }
+
+ int classIdxResi = 0;
+ if( useResiCondition )
+ {
+ //0: BS0 Resi0, 1:BS0 Resi1, 2:BS1 Resi0 3:BS1 Resi1
+ // A >> 1 = BS
+ // A - ( A >> 1) * 2
+ classIdxResi = classifierCodingInfo[posYDst + i][posXDst + j] - ((classifierCodingInfo[posYDst + i][posXDst + j] >> 1 ) * 2);
+ }
+#endif
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
std::array<short, FIX_FILTER_NUM_COEFF_9_DB_9> filterCoeff;
std::array<short, FIX_FILTER_NUM_COEFF_9_DB_9> filterClipp;
@@ -4656,6 +4900,25 @@ void AdaptiveLoopFilter::fixedFilteringResi(AlfClassifier **classifier, const CP
#endif
sum = (sum + offset) >> shift;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ sum = Clip3(-offsetClipValue, +offsetClipValue, sum);
+
+ int sign = sum < 0 ? -1 : +1;
+
+ int boundaryStrengthFactor = isIntraSlice ? 4 : 3;
+ sum = classIdxBs ? sign * ((abs(sum) * (16 + boundaryStrengthFactor) + 8 ) >> 4) : sum;
+ }
+ if( useResiCondition )
+ {
+ sum = Clip3(-offsetClipValue, +offsetClipValue, sum);
+
+ int sign = sum < 0 ? -1 : +1;
+ int resiStrengthFactor = isIntraSlice ? 0 >> (!isSpsAdjust ? 1 : 0) : 3 >> (!isSpsAdjust ? 1 : 0);
+ sum = classIdxResi ? sign * ((abs(sum) * (16 + resiStrengthFactor) + 8 ) >> 4) : sum;
+ }
+#endif
sum += curr;
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
@@ -4901,7 +5164,11 @@ void AdaptiveLoopFilter::deriveClassificationAndFixFilterResultsBlk( AlfClassifi
#if JVET_X0071_ALF_BAND_CLASSIFIER
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
- alfFixedFilterBlk( classifier[storeIdx], srcLuma, blkNew, blkDstNew, srcLumaBeforeDb, fixedFilterResults, m_picWidth, fixedFiltInd, fixedFiltSetInd, dirWindSize, clpRng, clippingValues, true );
+ alfFixedFilterBlk( classifier[storeIdx], srcLuma, blkNew, blkDstNew, srcLumaBeforeDb, fixedFilterResults, m_picWidth, fixedFiltInd, fixedFiltSetInd, dirWindSize, clpRng, clippingValues, true
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
#else
alfFixedFilterBlk( classifier[storeIdx], srcLuma, blkNew, blkDstNew, srcLumaBeforeDb, fixedFilterResults, m_picWidth, fixedFiltInd, fixedFiltSetInd, dirWindSize, clpRng, clippingValues );
#endif
@@ -4969,7 +5236,11 @@ void AdaptiveLoopFilter::deriveClassificationAndFixFilterResultsBlk( AlfClassifi
if (bResiFixed && dirWindSize == 0)
{
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
- m_filterResi9x9Blk(classifier[0], srcResiLuma, blk, blkDst, fixedFilterResiResults, m_picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltSetInd], fixedFiltSetInd, dirWindSize, clpRng, clippingValues);
+ m_filterResi9x9Blk(classifier[0], srcResiLuma, blk, blkDst, fixedFilterResiResults, m_picWidth, fixedFiltInd, m_classIdnFixedFilter9Db9[fixedFiltSetInd], fixedFiltSetInd, dirWindSize, clpRng, clippingValues
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , true, cs, m_classifierCodingInfo[0]
+#endif
+ );
#else
m_filterResi13x13Blk(classifier[0], srcResiLuma,
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
@@ -5005,6 +5276,9 @@ void AdaptiveLoopFilter::deriveClassificationAndFixFilterResultsBlk( AlfClassifi
#else
m_calcClass1(classifier[0], blkDst, Area(blkDst.pos().x, blkDst.pos().y, blkDst.width, blkDst.height), 5, 0, 5, 5,
bits, 2, mappingDir, laplacian);
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_textureClassMapping(classifier[0], blkDst, 0, 2, m_classifierCodingInfo[0] );
#endif
}
for( int curClassifierIdx = 1; curClassifierIdx < ALF_NUM_CLASSIFIER; curClassifierIdx++ )
@@ -5017,7 +5291,11 @@ void AdaptiveLoopFilter::deriveClassificationAndFixFilterResultsBlk( AlfClassifi
{
continue;
}
- m_calcClass2( classifier[curClassifierIdx], blk, Area( blkDst.pos().x, blkDst.pos().y, blkDst.width, blkDst.height ), srcLuma, 2, classifier[0], curClassifierIdx, bits, srcResiLuma, laplacian[0] );
+ m_calcClass2( classifier[curClassifierIdx], blk, Area( blkDst.pos().x, blkDst.pos().y, blkDst.width, blkDst.height ), srcLuma, 2, classifier[0], curClassifierIdx, bits, srcResiLuma, laplacian[0]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , m_classifierCodingInfo[0]
+#endif
+ );
#else
m_calcClass2( classifier[curClassifierIdx], blk, Area( blkDst.pos().x, blkDst.pos().y, blkDst.width, blkDst.height ), srcLuma, 2, classifier[0], curClassifierIdx, bits );
#endif
@@ -6855,7 +7133,11 @@ void AdaptiveLoopFilter::deriveFixedFilterResultsBlk(AlfClassifier*** classifier
if (fixedFiltSetInd == targetFixedFilterSetInd || targetFixedFilterSetInd == -1)
{
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
- alfFixedFilterBlk(winIdx == 0 ? classifier[0] : classifier[ALF_NUM_CLASSIFIER], srcLuma, blk, blkDst, srcLumaBeforeDb, m_fixFilterResult[COMPONENT_Y], m_picWidth, fixedFiltInd, fixedFiltSetInd, winIdx, clpRng, clippingValues, true);
+ alfFixedFilterBlk(winIdx == 0 ? classifier[0] : classifier[ALF_NUM_CLASSIFIER], srcLuma, blk, blkDst, srcLumaBeforeDb, m_fixFilterResult[COMPONENT_Y], m_picWidth, fixedFiltInd, fixedFiltSetInd, winIdx, clpRng, clippingValues, true
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
#else
alfFixedFilterBlk(winIdx == 0 ? classifier[0] : classifier[ALF_NUM_CLASSIFIER], srcLuma, blk, blkDst, srcLumaBeforeDb, m_fixFilterResult, m_picWidth, fixedFiltInd, fixedFiltSetInd, winIdx, clpRng, clippingValues);
#endif
@@ -6905,14 +7187,22 @@ void AdaptiveLoopFilter::deriveFixedFilterResultsPerBlkChroma(AlfClassifier ***c
m_deriveVariance(src, blk, blk, laplacian);
m_deriveClassificationLaplacian(src, blk, blk, laplacian, ALF_CLASSIFIER_FL_CHROMA);
m_calcClass0(classifier[ALF_NUM_CLASSIFIER + 1], blk, blk, ALF_CLASSIFIER_FL_CHROMA + 10, 1, NUM_DIR_FIX, NUM_ACT_FIX, bits, 2, mappingDir, laplacian);
- alfFixedFilterBlk(classifier[ALF_NUM_CLASSIFIER + 1], src, blk, blk, srcBeforeDb, fixedFilterResults, m_picWidth, fixedFilterSetIdx, targetFixedFilterSetInd, 0, clpRng, clippingValues, false);
+ alfFixedFilterBlk(classifier[ALF_NUM_CLASSIFIER + 1], src, blk, blk, srcBeforeDb, fixedFilterResults, m_picWidth, fixedFilterSetIdx, targetFixedFilterSetInd, 0, clpRng, clippingValues, false
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
}
else
{
deriveVariance(src, blk, blk, laplacian);
deriveClassificationLaplacian(src, blk, blk, laplacian, ALF_CLASSIFIER_FL_CHROMA);
calcClass0Var(classifier[ALF_NUM_CLASSIFIER + 1], blk, blk, ALF_CLASSIFIER_FL_CHROMA + 10, 1, NUM_DIR_FIX, NUM_ACT_FIX, bits, 2, mappingDir, laplacian);
- alfFixedFilterBlkNonSimd(classifier[ALF_NUM_CLASSIFIER + 1], src, blk, blk, srcBeforeDb, fixedFilterResults, m_picWidth, fixedFilterSetIdx, targetFixedFilterSetInd, 0, clpRng, clippingValues, false);
+ alfFixedFilterBlkNonSimd(classifier[ALF_NUM_CLASSIFIER + 1], src, blk, blk, srcBeforeDb, fixedFilterResults, m_picWidth, fixedFilterSetIdx, targetFixedFilterSetInd, 0, clpRng, clippingValues, false
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
}
}
@@ -7000,7 +7290,11 @@ void AdaptiveLoopFilter::deriveFixedFilterResultsCtuBoundaryChroma(AlfClassifier
}
#endif
-void AdaptiveLoopFilter::deriveFixedFilterResultsCtuBoundary(AlfClassifier ***classifier, Pel ***fixedFilterResults, const CPelBuf &srcLuma, const CPelBuf &srcLumaBeforeDb, const Area &blkDst, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS], uint8_t* ctuEnableFlagLuma, uint8_t* ctuEnableOnlineLuma, int ctuIdx, int classifierIdx)
+void AdaptiveLoopFilter::deriveFixedFilterResultsCtuBoundary(AlfClassifier ***classifier, Pel ***fixedFilterResults, const CPelBuf &srcLuma, const CPelBuf &srcLumaBeforeDb, const Area &blkDst, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS], uint8_t* ctuEnableFlagLuma, uint8_t* ctuEnableOnlineLuma, int ctuIdx, int classifierIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , const CPelBuf& srcCodingInfo, const CPelBuf& srcResi
+#endif
+ )
#else
void AdaptiveLoopFilter::deriveFixedFilterResultsCtuBoundary(AlfClassifier **classifier, Pel ***fixedFilterResults, const CPelBuf &srcLuma, const Area &blkDst, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS], uint8_t* ctuEnableFlagLuma, uint8_t* ctuEnableOnlineLuma, int ctuIdx )
#endif
@@ -7080,6 +7374,13 @@ void AdaptiveLoopFilter::deriveFixedFilterResultsCtuBoundary(AlfClassifier **cla
{
if(isBoundaryValid[boundaryIdx] && !isNeighborAvai[boundaryIdx])
{
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ //classifierIdx = 1 can reuse coding info results
+ if( classifierIdx == 0 )
+ {
+ calcAlfLumaCodingInfoBlk(cs, m_classifierCodingInfo[0], blkCur[boundaryIdx], blkCur[boundaryIdx], srcLuma, 2, 2, m_inputBitDepth[CHANNEL_TYPE_LUMA], srcResi, m_laplacian[0], srcCodingInfo);
+ }
+#endif
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
if (boundaryIdx == 1)
{
@@ -7429,11 +7730,19 @@ void AdaptiveLoopFilter::deriveFixedFilterResultsPerBlk( AlfClassifier **classif
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
if( useSimd)
{
- alfFixedFilterBlk( classifier[ALF_NUM_CLASSIFIER], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true );
+ alfFixedFilterBlk( classifier[ALF_NUM_CLASSIFIER], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
}
else
{
- alfFixedFilterBlkNonSimd( classifier[ALF_NUM_CLASSIFIER], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true );
+ alfFixedFilterBlkNonSimd( classifier[ALF_NUM_CLASSIFIER], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
}
#else
if( useSimd)
@@ -7456,7 +7765,11 @@ void AdaptiveLoopFilter::deriveFixedFilterResultsPerBlk( AlfClassifier **classif
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
m_calcClass0( classifier[0], blkCur, blkCur, usedWindowIdx[0], 1, NUM_DIR_FIX, NUM_ACT_FIX, bits, 2, mappingDir, laplacian);
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
- alfFixedFilterBlk( classifier[0], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true );
+ alfFixedFilterBlk( classifier[0], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
#else
alfFixedFilterBlk(classifier[0], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues );
#endif
@@ -7471,7 +7784,11 @@ void AdaptiveLoopFilter::deriveFixedFilterResultsPerBlk( AlfClassifier **classif
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
calcClass0Var( classifier[0], blkCur, blkCur, usedWindowIdx[0], 1, NUM_DIR_FIX, NUM_ACT_FIX, bits, 2, mappingDir, laplacian );
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
- alfFixedFilterBlkNonSimd( classifier[0], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true );
+ alfFixedFilterBlkNonSimd( classifier[0], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues, true
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , cs
+#endif
+ );
#else
alfFixedFilterBlkNonSimd( classifier[0], srcLuma, blkCur, blkCur, srcLumaBeforeDb, fixedFilterResults, m_picWidth, classifierIdx * NUM_FIXED_FILTER_SETS + fixedFilterSetIdx, targetFixedFilterSetInd, classifierIdx, clpRng, clippingValues );
#endif
@@ -7779,15 +8096,27 @@ void AdaptiveLoopFilter::deriveGaussResultsBlk(Pel ***gaussPic, const CPelBuf &s
bool useSimd = blkDst.size().width % 8 == 0 ? true : false;
if( useSimd )
{
- m_gaussFiltering(cs, gaussPic, srcLuma, blkDst, blk, clpRng, clippingValues, filterSetIdx, storeIdx);
+ m_gaussFiltering(cs, gaussPic, srcLuma, blkDst, blk, clpRng, clippingValues, filterSetIdx, storeIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , true, m_classifierCodingInfo[0]
+#endif
+ );
}
else
{
- gaussFiltering(cs, gaussPic, srcLuma, blkDst, blk, clpRng, clippingValues, filterSetIdx, storeIdx);
+ gaussFiltering(cs, gaussPic, srcLuma, blkDst, blk, clpRng, clippingValues, filterSetIdx, storeIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , true, m_classifierCodingInfo[0]
+#endif
+ );
}
}
-void AdaptiveLoopFilter::gaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx )
+void AdaptiveLoopFilter::gaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, AlfClassifier** classifierCodingInfo
+#endif
+ )
{
int strideSrc = srcLuma.stride;
int xPosSrc = blk.pos().x;
@@ -7798,6 +8127,13 @@ void AdaptiveLoopFilter::gaussFiltering(CodingStructure &cs, Pel ***gaussPic, co
int shift = 10;
const int numCoeff = 12;
int diffTH = 32;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+ const bool useBounCondition = applyCodingInfo && !(!isSpsAdjust && isIntraSlice);
+ const bool useResiCondition = applyCodingInfo && (isSpsAdjust || !isSpsAdjust) && !isIntraSlice && false;
+ const int offsetClipValue = 1 << ( clpRng.bd - 1 );
+#endif
int gaussTable[NUM_GAUSS_FILTERED_SOURCE][25] =
{
@@ -7827,6 +8163,24 @@ void AdaptiveLoopFilter::gaussFiltering(CodingStructure &cs, Pel ***gaussPic, co
{
for (int j = 0; j < width; j++)
{
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ int classIdxBs = 0;
+ if( useBounCondition )
+ {
+ //0: BS0 Resi0, 1:BS0 Resi1, 2:BS1 Resi0 3:BS1 Resi1
+ // A >> 1 = BS
+ // A - ( A >> 1) * 2
+ classIdxBs = classifierCodingInfo[blkDst.y + i][blkDst.x + j] >> 1;
+ }
+ int classIdxResi = 0;
+ if( useResiCondition )
+ {
+ //0: BS0 Resi0, 1:BS0 Resi1, 2:BS1 Resi0 3:BS1 Resi1
+ // A >> 1 = BS
+ // A - ( A >> 1) * 2
+ classIdxResi = classifierCodingInfo[blkDst.y + i][blkDst.x + j] - ((classifierCodingInfo[blkDst.y + i][blkDst.x + j] >> 1 ) * 2);
+ }
+#endif
int dstPosY = blkDst.y + i + padSize;
int dstPosX = blkDst.x + j + padSize;
@@ -7867,6 +8221,27 @@ void AdaptiveLoopFilter::gaussFiltering(CodingStructure &cs, Pel ***gaussPic, co
sum += 1 << (shift - 1);
sum >>= shift;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ sum = Clip3(-offsetClipValue, +offsetClipValue, sum);
+
+ int sign = sum < 0 ? -1 : +1;
+
+ int boundaryStrengthFactor = isIntraSlice ? 4 + 2 : 3 + 2;
+
+ sum = classIdxBs ? sign * ((abs(sum) * (16 + boundaryStrengthFactor) + 8 ) >> 4) : sum;
+ }
+
+ if( useResiCondition )
+ {
+ sum = Clip3(-offsetClipValue, +offsetClipValue, sum);
+
+ int sign = sum < 0 ? -1 : +1;
+ int resiStrengthFactor = isIntraSlice ? 0 >> (!isSpsAdjust ? 1 : 0) : 3 >> (!isSpsAdjust ? 1 : 0);
+ sum = classIdxResi ? sign * ((abs(sum) * (16 + resiStrengthFactor) + 8 ) >> 4) : sum;
+ }
+#endif
int diff = Clip3<int>(-diffTH, +diffTH, sum);
sum = curr + diff;
@@ -7876,3 +8251,143 @@ void AdaptiveLoopFilter::gaussFiltering(CodingStructure &cs, Pel ***gaussPic, co
}//height
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+void AdaptiveLoopFilter::textureClassMapping(AlfClassifier **classifier, const Area& blk, int classifierIdx, int subBlkSize, AlfClassifier **classifierCodingInfo )
+{
+ CHECK(classifierIdx != 0, "Wrong Classifier Index for DBF-BS Mapping");
+
+ int bsMappingTable[2][25] =
+ {
+ //A: (0)(12)(3)(4)
+ //D: (0)(12)(34)
+ //4A x 3D
+ {
+ 0, 1, 1, 2, 2,
+ 3, 4, 4, 5, 5,
+ 3, 4, 4, 5, 5,
+ 6, 7, 7, 8, 8,
+ 9, 10, 10, 11, 11,
+ },
+ //4A x 3D
+ {
+ 0, 1, 1, 2, 2,
+ 3, 4, 4, 5, 5,
+ 3, 4, 4, 5, 5,
+ 6, 7, 7, 8, 8,
+ 9, 10, 10, 11, 11,
+ },
+ };
+
+ int width = blk.width;
+ int height = blk.height;
+ int posY = blk.pos().y;
+ int posX = blk.pos().x;
+
+ int bsUnit = 0;
+ int transposeIdx = 0;
+ int classIdx = 0;
+ int classIdxMap = 0;
+ int bsClassOffet = 12;
+
+ for(int y = 0; y < height; y += subBlkSize)
+ {
+ for(int x = 0; x < width; x += subBlkSize)
+ {
+ bsUnit = classifierCodingInfo[posY + y][posX + x] >> 1;
+
+ transposeIdx = classifier[posY + y][posX + x] & 0x3;
+ classIdx = classifier[posY + y][posX + x] >> 2;
+
+ classIdxMap = bsMappingTable[bsUnit][classIdx] + bsUnit * bsClassOffet;
+
+ classIdxMap = (classIdxMap << 2) + transposeIdx;
+
+ for(int ii = 0; ii < subBlkSize; ii++)
+ {
+ for(int jj = 0; jj < subBlkSize; jj++)
+ {
+ int curPosY = posY + y + ii;
+ int curPosX = posX + x + jj;
+ classifier[curPosY][curPosX] = classIdxMap;
+ }
+ }
+ //Unit 2x2
+ }
+ }
+}
+
+void AdaptiveLoopFilter::calcAlfLumaCodingInfoBlk( CodingStructure& cs, AlfClassifier** classifier, const Area &blkDst, const Area &blkSrc, const CPelBuf& srcLuma, int subBlkSize, int classifierIdx, int bitDepth, const CPelBuf& srcLumaResi, uint32_t **buffer, const CPelBuf& srcCodingInfo )
+{
+
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+ const bool calcResi = !isIntraSlice && (isSpsAdjust || !isSpsAdjust);
+
+ const Pel *srcResiPtr = srcLumaResi.buf;
+ int srcResiStride = srcLumaResi.stride;
+ int yOffset = blkSrc.pos().y * srcResiStride;
+ const Pel *srcResi0 = &srcResiPtr[yOffset];
+ int srcResiStride2 = srcResiStride * 2;
+
+ const Pel *srcResiUp = srcResi0 - 1 * srcResiStride + blkSrc.pos().x - 1;
+
+ const Pel *srcResiDn = srcResiUp + srcResiStride;
+
+ if( calcResi )
+ {
+ //2x2 sum
+ for (int i = 0; i < blkSrc.height + 1 * 2; i += 2)
+ {
+ for (int j = 0; j < blkSrc.width + 1 * 2; j += 2)
+ {
+ buffer[i >> 1][j >> 1] = abs(srcResiUp[j]) + abs(srcResiUp[j + 1]) + abs(srcResiDn[j]) + abs(srcResiDn[j + 1]);
+ }
+ srcResiUp += srcResiStride2;
+ srcResiDn += srcResiStride2;
+ }
+ }
+
+ const int srcCodingInfoStride = srcCodingInfo.stride;
+ const Pel* srcCodingInfoPtr = srcCodingInfo.buf + blkSrc.pos().y * srcCodingInfoStride + blkSrc.pos().x;
+ const Pel* srcCodingInfoPtr0 = srcCodingInfoPtr;
+ const Pel* srcCodingInfoPtr1 = srcCodingInfoPtr;
+
+ for (int i = 0; i < blkDst.height; i += subBlkSize)
+ {
+ for (int j = 0; j < blkDst.width; j += subBlkSize)
+ {
+ int classIdxResi = 0;
+ if( calcResi )
+ {
+ int i2 = i >> 1;
+ int j2 = j >> 1;
+
+ int sum = buffer[i2][j2] + buffer[i2][j2 + 1] + buffer[i2 + 1][j2] + buffer[i2 + 1][j2 + 1];
+ int shiftOffset = 6;
+
+ int avgResi = sum >> (bitDepth - shiftOffset);
+ int highResiTh = +4;
+ classIdxResi = avgResi > highResiTh ? 1 : 0;
+ }
+
+ int classIdxCodingInfo = 0;
+ srcCodingInfoPtr0 = srcCodingInfoPtr + i * srcCodingInfoStride + j;
+ srcCodingInfoPtr1 = srcCodingInfoPtr0 + srcCodingInfoStride;
+ classIdxCodingInfo = ( srcCodingInfoPtr0[+0] + srcCodingInfoPtr0[+1] + srcCodingInfoPtr1[+0] + srcCodingInfoPtr1[+1] ) > 0 ? 1 : 0;
+
+ for (int ii = blkDst.y + i; ii < blkDst.y + i + subBlkSize; ii++)
+ {
+ for (int jj = blkDst.x + j; jj < blkDst.x + j + subBlkSize; jj++)
+ {
+ //0: BS0 Resi0, 1:BS0 Resi1, 2:BS1 Resi0 3:BS1 Resi1
+ // A >> 1 = BS
+ // A - ( A >> 1) * 2
+ classifier[ii][jj] = classIdxCodingInfo * 2 + classIdxResi;
+ }
+ }
+
+ }
+ }
+
+}
+#endif
diff --git a/source/Lib/CommonLib/AdaptiveLoopFilter.h b/source/Lib/CommonLib/AdaptiveLoopFilter.h
index 9808eb42a09f654bd07fcb88dbe22f97414cf661..d2dacb490aee9c719ad26b12472efe41bcd319eb 100644
--- a/source/Lib/CommonLib/AdaptiveLoopFilter.h
+++ b/source/Lib/CommonLib/AdaptiveLoopFilter.h
@@ -113,6 +113,9 @@ public:
#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
void copyResiData(CodingStructure &cs) { m_tempBufResi.bufs[COMPONENT_Y].copyFrom(cs.getResiBuf().bufs[COMPONENT_Y]); }
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf callCodingInfoBuf( CodingStructure &cs ) { return m_tempBufCodingInfo; }
+#endif
static constexpr int AlfNumClippingValues[MAX_NUM_CHANNEL_TYPE] = { 4, 4 };
static constexpr int MaxAlfNumClippingValues = 4;
@@ -160,7 +163,11 @@ public:
#endif
Pel ***fixedFilterResiResults, int picWidth, const int fixedFiltInd,
const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize,
- const ClpRng &clpRng, const Pel clippingValues[4]);
+ const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ );
#endif
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
@@ -174,7 +181,11 @@ public:
void deriveFixedFilterResults( AlfClassifier*** classifier, const CPelBuf& srcLuma, const CPelBuf& srcLumaBeforeDb, const Area& blkDst, const Area& blk, CodingStructure &cs, int winIdx, int fixedFilterSetIdx );
static void calcClass0Var( AlfClassifier **classifier, const Area &blkDst, const Area &cu, int dirWindSize, int classDir, int noDir, int noAct, int bitDepth, int subBlkSize, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS] );
static void deriveVariance( const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, uint32_t ***laplacian );
- void deriveFixedFilterResultsCtuBoundary( AlfClassifier ***classifier, Pel ***fixedFilterResults, const CPelBuf &srcLuma, const CPelBuf &srcLumaBeforeDb, const Area &blkDst, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS], uint8_t* ctuEnableFlagLuma, uint8_t* ctuEnableOnlineLuma, int ctuIdx, int classifierIdx );
+ void deriveFixedFilterResultsCtuBoundary( AlfClassifier ***classifier, Pel ***fixedFilterResults, const CPelBuf &srcLuma, const CPelBuf &srcLumaBeforeDb, const Area &blkDst, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS], uint8_t* ctuEnableFlagLuma, uint8_t* ctuEnableOnlineLuma, int ctuIdx, int classifierIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , const CPelBuf& srcCodingInfo, const CPelBuf& srcResi
+#endif
+ );
void deriveFixedFilterResultsPerBlk( AlfClassifier ***classifier, Pel ***fixedFilterResults, const CPelBuf &srcLuma, const CPelBuf &srcLumaBeforeDb, const Area &blkCur, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS], const int classifierIdx );
void(*m_deriveVariance)(const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, uint32_t ***variance);
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
@@ -182,17 +193,37 @@ public:
void deriveFixedFilterResultsPerBlkChroma(AlfClassifier ***classifier, Pel ***fixedFilterResults, const CPelBuf &src, const CPelBuf &srcBeforeDb, const Area &blk, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS]);
void deriveFixFilterResultsBlkChroma( AlfClassifier ***classifier, Pel ***fixedFilterResults, const CPelBuf &src, const CPelBuf &srcBeforeDb, const Area &blkDst, const Area &blk, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS] );
void deriveFixedFilterChroma(AlfClassifier*** classifier, const PelUnitBuf& src, const PelUnitBuf& srcBeforeDb, const Area& blkDst, const Area& blk, CodingStructure &cs, const int classifierIdx, ComponentID compID);
- void alfFixedFilterBlkNonSimd(AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma);
- void alfFixedFilterBlk(AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma);
+ void alfFixedFilterBlkNonSimd(AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , CodingStructure &cs
+#endif
+ );
+ void alfFixedFilterBlk(AlfClassifier **classifier, const CPelBuf &src, const Area &curBlk, const Area &blkDst, const CPelBuf &srcBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4], bool isLuma
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , CodingStructure &cs
+#endif
+ );
template<AlfFixedFilterType filtType>
#else
void alfFixedFilterBlkNonSimd(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]);
void alfFixedFilterBlk(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]);
template<AlfFixedFilterType filtType>
#endif
- static void fixedFilterBlk(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]);
- void(*m_fixFilter13x13Db9Blk)(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]);
- void(*m_fixFilter9x9Db9Blk)(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]);
+ static void fixedFilterBlk(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ );
+ void(*m_fixFilter13x13Db9Blk)(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ );
+ void(*m_fixFilter9x9Db9Blk)(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area &curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ );
#else
void paddingFixedFilterResultsCtu(Pel*** fixedFilterResultsPic, Pel*** fixedFilterResultsCtu, const int fixedFilterSetIdx, const Area &blk);
void deriveFixedFilterResultsCtuBoundary(AlfClassifier **classifier, Pel ***fixedFilterResults, const CPelBuf &srcLuma, const Area &blkDst, const int bits, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int qp, int fixedFilterSetIdx, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS], uint8_t* ctuEnableFlagLuma, uint8_t* ctuEnableOnlineLuma, int ctuIdx);
@@ -206,8 +237,23 @@ public:
void deriveGaussResultsBlk( Pel*** gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, CodingStructure& cs, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, const int storeIdx);
void deriveGaussResults(const CPelBuf& srcLumaDb, const Area& blkDst, const Area& blk, CodingStructure &cs, const int filterSetIdx, const int storeIdx );
- static void gaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx );
- void(*m_gaussFiltering) (CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx );
+ static void gaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, AlfClassifier** classifierCodingInfo
+#endif
+ );
+ void(*m_gaussFiltering) (CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, AlfClassifier** classifierCodingInfo
+#endif
+ );
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ static void textureClassMapping(AlfClassifier **classifier, const Area& blk, int classifierIdx, int subBlkSize, AlfClassifier **classifierCodingInfo);
+ void( *m_textureClassMapping ) (AlfClassifier **classifier, const Area& blk, int classifierIdx, int subBlkSize, AlfClassifier **classifierCodingInfo);
+
+ static void calcAlfLumaCodingInfoBlk( CodingStructure& cs, AlfClassifier** classifier, const Area &blkDst, const Area &blkSrc, const CPelBuf& srcLuma, int subBlkSize, int classifierIdx, int bitDepth, const CPelBuf& srcLumaResi, uint32_t **buffer, const CPelBuf& srcCodingInfo );
+ void( *m_calcAlfLumaCodingInfoBlk )( CodingStructure& cs, AlfClassifier** classifier, const Area &blkDst, const Area &blkSrc, const CPelBuf& srcLuma, int subBlkSize, int classifierIdx, int bitDepth, const CPelBuf& srcLumaResi, uint32_t **buffer, const CPelBuf& srcCodingInfo );
#endif
int assignAct(int avg_varPrec, int shift, int noAct);
@@ -230,6 +276,9 @@ public:
static void calcClassNew( AlfClassifier **classifier, const Area &blkDst, const Area &cu, const CPelBuf& srcLuma, int subBlkSize, AlfClassifier **classifier0, int classifierIdx, int bitDepth
#if JVET_AD0222_ALF_RESI_CLASS
, const CPelBuf& srcResiLuma, uint32_t **buffer
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , AlfClassifier **classifierCodingInfo
#endif
);
#else
@@ -292,7 +341,11 @@ public:
void(*m_calcClass1)(AlfClassifier **classifier, const Area &blkDst, const Area &cu, int dirWindSize, int classDir, int noDir, int noAct, int bitDepth, int subBlkSize, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS]);
#if JVET_X0071_ALF_BAND_CLASSIFIER
#if JVET_AD0222_ALF_RESI_CLASS
- void(*m_calcClass2)(AlfClassifier **classifier, const Area &blkDst, const Area &cu, const CPelBuf& srcLuma, int subBlkSize, AlfClassifier **classifier0, int classifierIdx, int bitDepth, const CPelBuf& srcLumaResi, uint32_t **buffer);
+ void(*m_calcClass2)(AlfClassifier **classifier, const Area &blkDst, const Area &cu, const CPelBuf& srcLuma, int subBlkSize, AlfClassifier **classifier0, int classifierIdx, int bitDepth, const CPelBuf& srcLumaResi, uint32_t **buffer
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , AlfClassifier **classifierCodingInfo
+#endif
+ );
#else
void(*m_calcClass2)(AlfClassifier **classifier, const Area &blkDst, const Area &cu, const CPelBuf& srcLuma, int subBlkSize, AlfClassifier **classifier0, int classifierIdx, int bitDepth);
#endif
@@ -318,7 +371,11 @@ public:
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
const Area &blkDst,
#endif
- Pel ***fixedFilterResiResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]);
+ Pel ***fixedFilterResiResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ );
#else
void (*m_filterResi13x13Blk)(AlfClassifier **classifier, const CPelBuf &srcResiLuma, const Area &curBlk,
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
@@ -783,6 +840,9 @@ protected:
#else
AlfClassifier** m_classifier;
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier** m_classifierCodingInfo[1];
+#endif
#if ALF_IMPROVEMENT
int m_numLumaAltAps[ALF_CTB_MAX_NUM_APS];
short m_coeffApsLuma[ALF_CTB_MAX_NUM_APS][MAX_NUM_ALF_ALTERNATIVES_LUMA][MAX_NUM_ALF_LUMA_COEFF * MAX_NUM_ALF_CLASSES];
@@ -861,6 +921,10 @@ protected:
#if JVET_AI0166_CCALF_CHROMA_SAO_INPUT
PelStorage m_tempBufSAO;
PelStorage m_tempBufSAO2;
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelStorage m_tempBufCodingInfo;
+ PelStorage m_tempBufCodingInfo2;
#endif
int m_inputBitDepth[MAX_NUM_CHANNEL_TYPE];
int m_picWidth;
diff --git a/source/Lib/CommonLib/CommonDef.h b/source/Lib/CommonLib/CommonDef.h
index a5a6c32fc81feba3cbfcafb07e19305a1a398b66..2eb2493d741c3cb8b6f8aaec6e689be212e07ca1 100644
--- a/source/Lib/CommonLib/CommonDef.h
+++ b/source/Lib/CommonLib/CommonDef.h
@@ -417,7 +417,11 @@ static const int MAX_CCSAO_BAND_NUM_U_BAND_BITS = 4;
#endif
static const int MAX_NUM_ALF_ALTERNATIVES_CHROMA = 8;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+static const int MAX_NUM_ALF_CLASSES = 24;
+#else
static const int MAX_NUM_ALF_CLASSES = 25;
+#endif
#if ALF_IMPROVEMENT
static const int MAX_NUM_ALF_ALTERNATIVES_LUMA = 4;
static const int EXT_LENGTH = 2;
@@ -544,8 +548,13 @@ static const int ALF_RESI_SHIFT_OFFSET = 4;
static const int NUM_RESI_ABS_PAD = 8;
static const int ALF_PADDING_SIZE_PRED = 3;
static const int ALF_NUM_CLASSIFIER = 3;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+static const int ALF_CLASSES_RESI = 24;
+static const int ALF_CLASSES_NEW = 24;
+#else
static const int ALF_CLASSES_RESI = 25;
static const int ALF_CLASSES_NEW = 25;
+#endif
static const int ALF_NUM_CLASSES_CLASSIFIER[ALF_NUM_CLASSIFIER] = { MAX_NUM_ALF_CLASSES, ALF_CLASSES_NEW, ALF_CLASSES_RESI };
#else
static const int ALF_NUM_CLASSIFIER = 2;
diff --git a/source/Lib/CommonLib/LoopFilter.cpp b/source/Lib/CommonLib/LoopFilter.cpp
index a5c2be2e0d7590bf3a54564b77804397f0f3000e..cb24c5113b75d1420bdcbaf86f0ce8d46e16d8f9 100644
--- a/source/Lib/CommonLib/LoopFilter.cpp
+++ b/source/Lib/CommonLib/LoopFilter.cpp
@@ -151,6 +151,9 @@ void LoopFilter::destroy()
\param pcPic picture class (Pic) pointer
*/
void LoopFilter::loopFilterPic( CodingStructure& cs
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , PelUnitBuf& alfCodingInfo, bool storeInfo
+#endif
)
{
const PreCalcValues& pcv = *cs.pcv;
@@ -193,7 +196,11 @@ void LoopFilter::loopFilterPic( CodingStructure& cs
continue;
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ xDeblockCU( currCU, EDGE_VER, storeInfo, alfCodingInfo );
+#else
xDeblockCU( currCU, EDGE_VER );
+#endif
}
@@ -215,7 +222,11 @@ void LoopFilter::loopFilterPic( CodingStructure& cs
continue;
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ xDeblockCU( currCU, EDGE_VER, storeInfo, alfCodingInfo );
+#else
xDeblockCU( currCU, EDGE_VER );
+#endif
}
}
}
@@ -245,7 +256,11 @@ void LoopFilter::loopFilterPic( CodingStructure& cs
continue;
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ xDeblockCU( currCU, EDGE_HOR, storeInfo, alfCodingInfo );
+#else
xDeblockCU( currCU, EDGE_HOR );
+#endif
}
#if JVET_AI0136_ADAPTIVE_DUAL_TREE
@@ -266,7 +281,11 @@ void LoopFilter::loopFilterPic( CodingStructure& cs
continue;
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ xDeblockCU( currCU, EDGE_HOR, storeInfo, alfCodingInfo );
+#else
xDeblockCU( currCU, EDGE_HOR );
+#endif
}
}
}
@@ -299,7 +318,11 @@ void LoopFilter::resetFilterLengths()
\param cu the CU to be deblocked
\param edgeDir the direction of the edge in block boundary (horizontal/vertical), which is added newly
*/
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+void LoopFilter::xDeblockCU( CodingUnit& cu, const DeblockEdgeDir edgeDir, bool storeInfo, PelUnitBuf& alfCodingInfo )
+#else
void LoopFilter::xDeblockCU( CodingUnit& cu, const DeblockEdgeDir edgeDir )
+#endif
{
const PreCalcValues& pcv = *cu.cs->pcv;
const Area area = cu.Y().valid() ? cu.Y() : Area( recalcPosition( cu.chromaFormat, cu.chType, CHANNEL_TYPE_LUMA, cu.blocks[cu.chType].pos() ), recalcSize( cu.chromaFormat, cu.chType, CHANNEL_TYPE_LUMA, cu.blocks[cu.chType].size() ) );
@@ -425,6 +448,46 @@ void LoopFilter::xDeblockCU( CodingUnit& cu, const DeblockEdgeDir edgeDir )
}
const unsigned uiPelsInPart = pcv.minCUWidth;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ auto storeBoundaryInfo = [this, cu, uiPelsInPart, edgeDir]( char bs, Position pos, ComponentID comp, PelUnitBuf& bsBuf) -> void
+ {
+ Size sz;
+ Position posPQ;
+
+ int edgeLongside = uiPelsInPart;
+ int edgeShortside = 2;
+
+ int scaleX = getComponentScaleX( comp, cu.chromaFormat );
+ int scaleY = getComponentScaleY( comp, cu.chromaFormat );
+
+ if(edgeDir == EDGE_HOR)
+ {
+ sz.width = edgeLongside;
+ sz.height= edgeShortside << scaleY;
+ posPQ = Position( pos.x, pos.y - ( 1 << scaleY) );
+ }
+ else
+ {
+ sz.height= edgeLongside;
+ sz.width = edgeShortside << scaleX;
+ posPQ = Position( pos.x - ( 1 << scaleX) , pos.y);
+ }
+
+ auto bsComp = BsGet(bs, comp);
+ Pel toFill = bsComp > 0 ? 1 : 0;
+
+ int bsStride = bsBuf.get(comp).stride;
+ Pel* bsPtr = bsBuf.get(comp).buf + (posPQ.y >> scaleY) * bsStride + (posPQ.x >> scaleX);
+ for(int y = 0; y < sz.height >> scaleY; y++)
+ {
+ for(int x = 0; x < sz.width >> scaleX; x++)
+ {
+ bsPtr[x] = std::max( bsPtr[x], toFill );
+ }
+ bsPtr += bsStride;
+ }
+ };
+#endif
for( int y = 0; y < area.height; y += uiPelsInPart )
{
@@ -442,6 +505,12 @@ void LoopFilter::xDeblockCU( CodingUnit& cu, const DeblockEdgeDir edgeDir )
#endif
{
bS |= xGetBoundaryStrengthSingle( cu, edgeDir, localPos, CHANNEL_TYPE_LUMA );
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( cu.blocks[COMPONENT_Y].valid() && storeInfo )
+ {
+ storeBoundaryInfo(bS, Position(area.x + x, area.y + y), COMPONENT_Y, alfCodingInfo);
+ }
+#endif
}
#if !INTRA_RM_SMALL_BLOCK_SIZE_CONSTRAINTS
if(cu.treeType != TREE_L && cu.chromaFormat != CHROMA_400 && cu.blocks[COMPONENT_Cb].valid())
diff --git a/source/Lib/CommonLib/LoopFilter.h b/source/Lib/CommonLib/LoopFilter.h
index 0059914d42b3a59d22aafab0fbf6747f0e20e3d6..db56b9cea17a8343cf73863054b3f59ff6be63f1 100644
--- a/source/Lib/CommonLib/LoopFilter.h
+++ b/source/Lib/CommonLib/LoopFilter.h
@@ -129,7 +129,11 @@ public:
~LoopFilter();
/// CU-level deblocking function
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ void xDeblockCU(CodingUnit& cu, const DeblockEdgeDir edgeDir, bool storeInfo, PelUnitBuf& alfCodingInfo);
+#else
void xDeblockCU(CodingUnit& cu, const DeblockEdgeDir edgeDir);
+#endif
void initEncPicYuvBuffer(ChromaFormat chromaFormat, const Size &size, const unsigned maxCUSize);
PelStorage& getDbEncPicYuvBuffer() { return m_encPicYuvBuffer; }
void setEnc(bool b) { m_enc = b; }
@@ -139,6 +143,9 @@ public:
/// picture-level deblocking filter
void loopFilterPic ( CodingStructure& cs
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , PelUnitBuf& alfCodingInfo, bool storeInfo
+#endif
);
static int getBeta ( const int qp )
diff --git a/source/Lib/CommonLib/Slice.h b/source/Lib/CommonLib/Slice.h
index 4e132f78c0cf89d346a997752c0b1f695e89ea10..f523b5be70dd2b47289ac351bde04b2c0728083c 100644
--- a/source/Lib/CommonLib/Slice.h
+++ b/source/Lib/CommonLib/Slice.h
@@ -1775,6 +1775,9 @@ private:
bool m_alfEnabledFlag;
bool m_ccalfEnabledFlag;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ bool m_alfLumaFixedFilterAdjust;
+#endif
bool m_wrapAroundEnabledFlag;
unsigned m_IBCFlag;
#if JVET_AD0208_IBC_ADAPT_FOR_CAM_CAPTURED_CONTENTS
@@ -2239,6 +2242,10 @@ public:
void setALFEnabledFlag( bool b ) { m_alfEnabledFlag = b; }
bool getCCALFEnabledFlag() const { return m_ccalfEnabledFlag; }
void setCCALFEnabledFlag( bool b ) { m_ccalfEnabledFlag = b; }
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ bool getAlfLumaFixedFilterAdjust() const { return m_alfLumaFixedFilterAdjust; }
+ void setAlfLumaFixedFilterAdjust( bool b ) { m_alfLumaFixedFilterAdjust = b; }
+#endif
void setJointCbCrEnabledFlag(bool bVal) { m_JointCbCrEnabledFlag = bVal; }
bool getJointCbCrEnabledFlag() const { return m_JointCbCrEnabledFlag; }
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index c2c5e192c06ced35c0536cd3819d17e6a412e109..6a8daf3b5f335ec492032221c60387793009695f 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -482,6 +482,7 @@
#define JVET_AI0084_ALF_RESIDUALS_SCALING 1 // JVET_AI0084: non-fixed ALF residuals scaling
#define JVET_AI0058_ALF_RELAXED_RDO_LUMA 1 // JVET-AI0058: Relaxed ALF Luma RDO
#define JVET_AI0166_CCALF_CHROMA_SAO_INPUT 1 // JVET-AI0166: CCALF with Chroma inputs
+#define JVET_AJ0188_CODING_INFO_CLASSIFICATION 1 // JVET-AJ0188: Coding Information based Classification for ALF
// SIMD optimizations
#if IF_12TAP
diff --git a/source/Lib/CommonLib/x86/AdaptiveLoopFilterX86.h b/source/Lib/CommonLib/x86/AdaptiveLoopFilterX86.h
index fd0efa2db34a06577a2e59f6067a90a99ca4216f..bd42c0c21426d14f05823d270766a21357508e72 100644
--- a/source/Lib/CommonLib/x86/AdaptiveLoopFilterX86.h
+++ b/source/Lib/CommonLib/x86/AdaptiveLoopFilterX86.h
@@ -1193,14 +1193,302 @@ static void simdFilter9x9Blk(AlfClassifier **classifier, const PelUnitBuf &recDs
const Pel *src = srcBuffer.buf + blk.y * srcStride + blk.x;
Pel * dst = dstBuffer.buf + blkDst.y * dstStride + blkDst.x;
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
const __m128i mmOffset = _mm_set1_epi32(round);
const __m128i mmMin = _mm_set1_epi16(clpRng.min);
const __m128i mmMax = _mm_set1_epi16(clpRng.max);
+#endif
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
const int padSize = ALF_PADDING_SIZE_FIXED_RESULTS;
#endif
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool use256BitSimd = vext >= AVX2 && blkDst.width % 16 == 0 ? true : false;
+
+ if( use256BitSimd )
+ {
+ const __m256i mmOffset = _mm256_set1_epi32(round);
+ const __m256i mmMin = _mm256_set1_epi16(clpRng.min);
+ const __m256i mmMax = _mm256_set1_epi16(clpRng.max);
+
+ for (size_t i = 0; i < height; i += stepY)
+ {
+ const AlfClassifier *pClass = isChroma(compId) ? nullptr : classifier[blkDst.y + i] + blkDst.x;
+ for (size_t j = 0; j < width; j += stepX * 2)
+ {
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ __m256i params[2][2][13];
+#else
+ __m256i params[2][2][10];
+#endif
+ for (int k = 0; k < 2; k++)
+ {
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ __m128i rawCoeffTmp[2][2][4], rawClipTmp[2][2][4], s0Tmp[2], s1Tmp[2], s2Tmp[2], s3Tmp[2];
+ __m256i rawCoeff[2][4], rawClip[2][4];
+#else
+ __m128i rawCoeffTmp[2][2][3], rawClipTmp[2][2][3], s0Tmp[2], s1Tmp[2], s2Tmp[2], s3Tmp[2];
+ __m256i rawCoeff[2][3], rawClip[2][3];
+#endif
+
+ for (int l = 0; l < 2; l++)
+ {
+ const int transposeIdx0 = pClass ? (pClass[j + 4 * k + 2 * l + 0] & 0x3) : 0;
+ const int classIdx0 = pClass ? (pClass[j + 4 * k + 2 * l + 0] >> 2) : 0;
+
+ rawCoeffTmp[0][l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF));
+ rawCoeffTmp[0][l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 8));
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ rawCoeffTmp[0][l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawCoeffTmp[0][l][3] = _mm_loadl_epi64((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 24));
+#else
+ rawCoeffTmp[0][l][2] = _mm_loadl_epi64((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+#endif
+ rawClipTmp[0][l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF));
+ rawClipTmp[0][l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 8));
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ rawClipTmp[0][l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawClipTmp[0][l][3] = _mm_loadl_epi64((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 24));
+#else
+ rawClipTmp[0][l][2] = _mm_loadl_epi64((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+#endif
+
+ for (int m = 0; m < shuffleTime9[transposeIdx0]; m++)
+ {
+ int op0 = shuffleOp9[transposeIdx0][m][0];
+ int op1 = shuffleOp9[transposeIdx0][m][1];
+
+ s0Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab9[transposeIdx0][m][0]);
+ s1Tmp[0] = _mm_xor_si128(s0Tmp[0], _mm_set1_epi8((char) 0x80));
+ s2Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab9[transposeIdx0][m][1]);
+ s3Tmp[0] = _mm_xor_si128(s2Tmp[0], _mm_set1_epi8((char) 0x80));
+
+ __m128i rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoeffTmp[0][l][op0], s0Tmp[0]), _mm_shuffle_epi8(rawCoeffTmp[0][l][op1], s1Tmp[0]));
+ __m128i rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoeffTmp[0][l][op0], s2Tmp[0]), _mm_shuffle_epi8(rawCoeffTmp[0][l][op1], s3Tmp[0]));
+ rawCoeffTmp[0][l][op0] = rawTmp0;
+ rawCoeffTmp[0][l][op1] = rawTmp1;
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[0][l][op0], s0Tmp[0]), _mm_shuffle_epi8(rawClipTmp[0][l][op1], s1Tmp[0]));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[0][l][op0], s2Tmp[0]), _mm_shuffle_epi8(rawClipTmp[0][l][op1], s3Tmp[0]));
+ rawClipTmp[0][l][op0] = rawTmp0;
+ rawClipTmp[0][l][op1] = rawTmp1;
+ }
+
+ const int transposeIdx1 = pClass ? (pClass[j + 4 * k + 2 * l + 8] & 0x3) : 0;
+ const int classIdx1 = pClass ? (pClass[j + 4 * k + 2 * l + 8] >> 2) : 0;
+
+ rawCoeffTmp[1][l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF));
+ rawCoeffTmp[1][l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 8));
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ rawCoeffTmp[1][l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawCoeffTmp[1][l][3] = _mm_loadl_epi64((const __m128i *) (filterSet + classIdx1* MAX_NUM_ALF_LUMA_COEFF + 24));
+#else
+ rawCoeffTmp[1][l][2] = _mm_loadl_epi64((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+#endif
+ rawClipTmp[1][l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF));
+ rawClipTmp[1][l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 8));
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ rawClipTmp[1][l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawClipTmp[1][l][3] = _mm_loadl_epi64((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 24));
+#else
+ rawClipTmp[1][l][2] = _mm_loadl_epi64((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+#endif
+
+ for (int m = 0; m < shuffleTime9[transposeIdx1]; m++)
+ {
+ int op0 = shuffleOp9[transposeIdx1][m][0];
+ int op1 = shuffleOp9[transposeIdx1][m][1];
+
+ s0Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab9[transposeIdx1][m][0]);
+ s1Tmp[1] = _mm_xor_si128(s0Tmp[1], _mm_set1_epi8((char) 0x80));
+ s2Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab9[transposeIdx1][m][1]);
+ s3Tmp[1] = _mm_xor_si128(s2Tmp[1], _mm_set1_epi8((char) 0x80));
+
+ __m128i rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoeffTmp[1][l][op0], s0Tmp[1]), _mm_shuffle_epi8(rawCoeffTmp[1][l][op1], s1Tmp[1]));
+ __m128i rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoeffTmp[1][l][op0], s2Tmp[1]), _mm_shuffle_epi8(rawCoeffTmp[1][l][op1], s3Tmp[1]));
+ rawCoeffTmp[1][l][op0] = rawTmp0;
+ rawCoeffTmp[1][l][op1] = rawTmp1;
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[1][l][op0], s0Tmp[1]), _mm_shuffle_epi8(rawClipTmp[1][l][op1], s1Tmp[1]));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[1][l][op0], s2Tmp[1]), _mm_shuffle_epi8(rawClipTmp[1][l][op1], s3Tmp[1]));
+ rawClipTmp[1][l][op0] = rawTmp0;
+ rawClipTmp[1][l][op1] = rawTmp1;
+ }
+
+ rawCoeff[l][0] = _mm256_castsi128_si256( rawCoeffTmp[0][l][0]);
+ rawCoeff[l][0] = _mm256_insertf128_si256(rawCoeff[l][0], rawCoeffTmp[1][l][0], 1);
+ rawCoeff[l][1] = _mm256_castsi128_si256(rawCoeffTmp[0][l][1]);
+ rawCoeff[l][1] = _mm256_insertf128_si256(rawCoeff[l][1], rawCoeffTmp[1][l][1], 1);
+ rawCoeff[l][2] = _mm256_castsi128_si256(rawCoeffTmp[0][l][2]);
+ rawCoeff[l][2] = _mm256_insertf128_si256(rawCoeff[l][2], rawCoeffTmp[1][l][2], 1);
+ rawCoeff[l][3] = _mm256_castsi128_si256(rawCoeffTmp[0][l][3]);
+ rawCoeff[l][3] = _mm256_insertf128_si256(rawCoeff[l][3], rawCoeffTmp[1][l][3], 1);
+
+ rawClip[l][0] = _mm256_castsi128_si256(rawClipTmp[0][l][0]);
+ rawClip[l][0] = _mm256_insertf128_si256(rawClip[l][0], rawClipTmp[1][l][0], 1);
+ rawClip[l][1] = _mm256_castsi128_si256(rawClipTmp[0][l][1]);
+ rawClip[l][1] = _mm256_insertf128_si256(rawClip[l][1], rawClipTmp[1][l][1], 1);
+ rawClip[l][2] = _mm256_castsi128_si256(rawClipTmp[0][l][2]);
+ rawClip[l][2] = _mm256_insertf128_si256(rawClip[l][2], rawClipTmp[1][l][2], 1);
+ rawClip[l][3] = _mm256_castsi128_si256(rawClipTmp[0][l][3]);
+ rawClip[l][3] = _mm256_insertf128_si256(rawClip[l][3], rawClipTmp[1][l][3], 1);
+ }
+
+ params[k][0][0] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][0], 0x00), _mm256_shuffle_epi32(rawCoeff[1][0], 0x00));
+ params[k][0][1] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][0], 0x55), _mm256_shuffle_epi32(rawCoeff[1][0], 0x55));
+ params[k][0][2] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][0], 0xaa), _mm256_shuffle_epi32(rawCoeff[1][0], 0xaa));
+ params[k][0][3] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][0], 0xff), _mm256_shuffle_epi32(rawCoeff[1][0], 0xff));
+ params[k][0][4] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][1], 0x00), _mm256_shuffle_epi32(rawCoeff[1][1], 0x00));
+ params[k][0][5] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][1], 0x55), _mm256_shuffle_epi32(rawCoeff[1][1], 0x55));
+ params[k][0][6] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][1], 0xaa), _mm256_shuffle_epi32(rawCoeff[1][1], 0xaa));
+ params[k][0][7] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][1], 0xff), _mm256_shuffle_epi32(rawCoeff[1][1], 0xff));
+ params[k][0][8] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][2], 0x00), _mm256_shuffle_epi32(rawCoeff[1][2], 0x00));
+ params[k][0][9] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][2], 0x55), _mm256_shuffle_epi32(rawCoeff[1][2], 0x55));
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ params[k][0][10] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][2], 0xaa), _mm256_shuffle_epi32(rawCoeff[1][2], 0xaa));
+ params[k][0][11] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][2], 0xff), _mm256_shuffle_epi32(rawCoeff[1][2], 0xff));
+ params[k][0][12] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoeff[0][3], 0x00), _mm256_shuffle_epi32(rawCoeff[1][3], 0x00));
+#endif
+
+ params[k][1][0] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][0], 0x00), _mm256_shuffle_epi32(rawClip[1][0], 0x00));
+ params[k][1][1] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][0], 0x55), _mm256_shuffle_epi32(rawClip[1][0], 0x55));
+ params[k][1][2] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][0], 0xaa), _mm256_shuffle_epi32(rawClip[1][0], 0xaa));
+ params[k][1][3] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][0], 0xff), _mm256_shuffle_epi32(rawClip[1][0], 0xff));
+ params[k][1][4] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][1], 0x00), _mm256_shuffle_epi32(rawClip[1][1], 0x00));
+ params[k][1][5] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][1], 0x55), _mm256_shuffle_epi32(rawClip[1][1], 0x55));
+ params[k][1][6] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][1], 0xaa), _mm256_shuffle_epi32(rawClip[1][1], 0xaa));
+ params[k][1][7] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][1], 0xff), _mm256_shuffle_epi32(rawClip[1][1], 0xff));
+ params[k][1][8] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][2], 0x00), _mm256_shuffle_epi32(rawClip[1][2], 0x00));
+ params[k][1][9] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][2], 0x55), _mm256_shuffle_epi32(rawClip[1][2], 0x55));
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ params[k][1][10] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][2], 0xaa), _mm256_shuffle_epi32(rawClip[1][2], 0xaa));
+ params[k][1][11] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][2], 0xff), _mm256_shuffle_epi32(rawClip[1][2], 0xff));
+ params[k][1][12] = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][3], 0x00), _mm256_shuffle_epi32(rawClip[1][3], 0x00));
+#endif
+ }
+
+ for (size_t ii = 0; ii < stepY; ii++)
+ {
+ const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
+
+ pImg0 = src + j + ii * srcStride;
+ pImg1 = pImg0 + srcStride;
+ pImg2 = pImg0 - srcStride;
+ pImg3 = pImg1 + srcStride;
+ pImg4 = pImg2 - srcStride;
+ pImg5 = pImg3 + srcStride;
+ pImg6 = pImg4 - srcStride;
+ pImg7 = pImg5 + srcStride;
+ pImg8 = pImg6 - srcStride;
+
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ const Pel *pImg0FixedBased = fixedFilterResults[fixedFilterSetIdx][blkDst.y + i + ii + padSize + 0] + blkDst.x + j + padSize;
+ const Pel *pImg1FixedBased = fixedFilterResults[fixedFilterSetIdx][blkDst.y + i + ii + padSize + 1] + blkDst.x + j + padSize;
+ const Pel *pImg2FixedBased = fixedFilterResults[fixedFilterSetIdx][blkDst.y + i + ii + padSize - 1] + blkDst.x + j + padSize;
+ const Pel *pImg3FixedBased = fixedFilterResults[fixedFilterSetIdx][blkDst.y + i + ii + padSize + 2] + blkDst.x + j + padSize;
+ const Pel *pImg4FixedBased = fixedFilterResults[fixedFilterSetIdx][blkDst.y + i + ii + padSize - 2] + blkDst.x + j + padSize;
+#endif
+ __m256i cur = _mm256_loadu_si256((const __m256i *) pImg0);
+ __m256i accumA = mmOffset;
+ __m256i accumB = mmOffset;
+
+ auto process2coeffs = [&](const int i, const Pel *ptr0, const Pel *ptr1, const Pel *ptr2, const Pel *ptr3)
+ {
+ const __m256i val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr0), cur);
+ const __m256i val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr2), cur);
+ const __m256i val01 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr1), cur);
+ const __m256i val11 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr3), cur);
+
+ __m256i val01A = _mm256_unpacklo_epi16(val00, val10);
+ __m256i val01B = _mm256_unpackhi_epi16(val00, val10);
+ __m256i val01C = _mm256_unpacklo_epi16(val01, val11);
+ __m256i val01D = _mm256_unpackhi_epi16(val01, val11);
+
+ __m256i limit01A = params[0][1][i];
+ __m256i limit01B = params[1][1][i];
+
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ val01C = _mm256_min_epi16(val01C, limit01A);
+ val01D = _mm256_min_epi16(val01D, limit01B);
+
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+ val01C = _mm256_max_epi16(val01C, limit01A);
+ val01D = _mm256_max_epi16(val01D, limit01B);
+
+ val01A = _mm256_add_epi16(val01A, val01C);
+ val01B = _mm256_add_epi16(val01B, val01D);
+
+ const __m256i coeff01A = params[0][0][i];
+ const __m256i coeff01B = params[1][0][i];
+
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+ };
+
+ process2coeffs(0, pImg7 + 0, pImg8 + 0, pImg5 + 1, pImg6 - 1);
+ process2coeffs(1, pImg5 + 0, pImg6 + 0, pImg5 - 1, pImg6 + 1);
+ process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
+ process2coeffs(3, pImg3 + 0, pImg4 + 0, pImg3 - 1, pImg4 + 1);
+ process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 3, pImg2 - 3);
+ process2coeffs(5, pImg1 + 2, pImg2 - 2, pImg1 + 1, pImg2 - 1);
+ process2coeffs(6, pImg1 + 0, pImg2 + 0, pImg1 - 1, pImg2 + 1);
+ process2coeffs(7, pImg1 - 2, pImg2 + 2, pImg1 - 3, pImg2 + 3);
+ process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ process2coeffs(10, pImg3FixedBased + 0, pImg4FixedBased + 0, pImg1FixedBased + 0, pImg2FixedBased + 0);
+ process2coeffs(11, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+
+ __m256i val00 = _mm256_sub_epi16( _mm256_loadu_si256((const __m256i *) (fixedFilterResults[fixedFilterSetIdx][blkDst.y + i + ii + padSize] + blkDst.x + j + padSize)), cur);
+ __m256i val10 = _mm256_setzero_si256();
+ __m256i val01A = _mm256_unpacklo_epi16(val00, val10);
+ __m256i val01B = _mm256_unpackhi_epi16(val00, val10);
+ __m256i limit01A = params[0][1][12];
+ __m256i limit01B = params[1][1][12];
+
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+
+ __m256i coeff01A = params[0][0][12];
+ __m256i coeff01B = params[1][0][12];
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+#endif
+
+ accumA = _mm256_srai_epi32(accumA, shift);
+ accumB = _mm256_srai_epi32(accumB, shift);
+
+ accumA = _mm256_packs_epi32(accumA, accumB);
+ accumA = _mm256_add_epi16(accumA, cur);
+ accumA = _mm256_min_epi16(mmMax, _mm256_max_epi16(accumA, mmMin));
+
+ _mm256_storeu_si256((__m256i *) (dst + ii * dstStride + j), accumA);
+
+ } // for (size_t ii = 0; ii < stepY; ii++)
+ } // for (size_t j = 0; j < width; j += stepX)
+ src += srcStride * stepY;
+ dst += dstStride * stepY;
+ }
+ }
+ else
+ {
+
+ const __m128i mmOffset = _mm_set1_epi32(round);
+ const __m128i mmMin = _mm_set1_epi16(clpRng.min);
+ const __m128i mmMax = _mm_set1_epi16(clpRng.max);
+#endif
+
for (size_t i = 0; i < height; i += stepY)
{
const AlfClassifier *pClass = isChroma(compId) ? nullptr : classifier[blkDst.y + i] + blkDst.x;
@@ -1413,6 +1701,9 @@ static void simdFilter9x9Blk(AlfClassifier **classifier, const PelUnitBuf &recDs
src += srcStride * stepY;
dst += dstStride * stepY;
}
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ }//use 256 Bit Simd
+#endif
}
#endif
@@ -3473,7 +3764,9 @@ static void simdFilter13x13BlkExtDbResiDirect(
const Pel currBase = 512;
int round = 1 << (shift - 1);
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
__m128i curBase = _mm_set_epi16( currBase, currBase, currBase, currBase, currBase, currBase, currBase, currBase );
+#endif
#else
#if JVET_AG0158_ALF_LUMA_COEFF_PRECISION
int shift = coeffBits;
@@ -3491,9 +3784,11 @@ static void simdFilter13x13BlkExtDbResiDirect(
constexpr size_t stepX = 8;
size_t stepY = 1;
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
const __m128i mmOffset = _mm_set1_epi32(round);
const __m128i mmMin = _mm_set1_epi16(clpRng.min);
const __m128i mmMax = _mm_set1_epi16(clpRng.max);
+#endif
static_assert(sizeof(*filterSet) == 2, "ALF coeffs must be 16-bit wide");
static_assert(sizeof(*fClipSet) == 2, "ALF clip values must be 16-bit wide");
@@ -3509,600 +3804,1798 @@ static void simdFilter13x13BlkExtDbResiDirect(
const int padSizeGauss = ALF_PADDING_SIZE_GAUSS_RESULTS;
#endif
- for (size_t i = 0; i < height; i += stepY)
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool use256BitSimd = vext >= AVX2 && blkDst.width % 16 == 0 ? true : false;
+
+ if( use256BitSimd )
{
- const AlfClassifier *pClass = classifier[blkDst.y + i] + blkDst.x;
- for (size_t j = 0; j < width; j += stepX)
+ const __m256i mmOffset = _mm256_set1_epi32(round);
+ const __m256i mmMin = _mm256_set1_epi16(clpRng.min);
+ const __m256i mmMax = _mm256_set1_epi16(clpRng.max);
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ const __m256i curBase = _mm256_set1_epi16(currBase);
+#endif
+
+ for (size_t i = 0; i < height; i += stepY)
{
+ const AlfClassifier *pClass = classifier[blkDst.y + i] + blkDst.x;
+ for (size_t j = 0; j < width; j += stepX * 2)
+ {
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- __m128i params[2][2][20];
+ __m256i params[2][2][20];
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- __m128i params[2][2][17];
+ __m256i params[2][2][17];
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- __m128i params[2][2][19];
+ __m256i params[2][2][19];
#else
- __m128i params[2][2][16];
+ __m256i params[2][2][16];
#endif
#else
- __m128i params[2][2][13];
+ __m256i params[2][2][13];
#endif
- for (int k = 0; k < 2; k++)
- {
+ for (int k = 0; k < 2; k++)
+ {
#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- __m128i rawCoef[4][5], rawClip[4][5], s0, s1, s2, s3, rawTmp0, rawTmp1;
+ __m256i rawCoef[4][5], rawClip[4][5], s0, s1;
+ __m128i rawCoefTmp[2][4][5], rawClipTmp[2][4][5], s0Tmp[2], s1Tmp[2], s2Tmp[2], s3Tmp[2];
#else
- __m128i rawCoef[4][4], rawClip[4][4], s0, s1, s2, s3, rawTmp0, rawTmp1;
+ __m256i rawCoef[4][4], rawClip[4][4], s0, s1;
+ __m128i rawCoefTmp[2][4][4], rawClipTmp[2][4][4], s0Tmp[2], s1Tmp[2], s2Tmp[2], s3Tmp[2];
#endif
- for (int l = 0; l < 4; l++)
- {
- const int transposeIdx = pClass[j + 4 * k + l] & 0x3;
- const int classIdx = pClass[j + 4 * k + l] >> 2;
+ for (int l = 0; l < 4; l++)
+ {
+ const int transposeIdx0 = pClass[j + 4 * k + l + 0] & 0x3;
+ const int classIdx0 = pClass[j + 4 * k + l + 0] >> 2;
- rawCoef[l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF));
- rawCoef[l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 8));
- rawCoef[l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 16));
- rawCoef[l][3] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 24));
+ rawCoefTmp[0][l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF));
+ rawCoefTmp[0][l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawCoefTmp[0][l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawCoefTmp[0][l][3] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 24));
#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- rawCoef[l][4] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 32));
+ rawCoefTmp[0][l][4] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 32));
#endif
- rawClip[l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF));
- rawClip[l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 8));
- rawClip[l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 16));
- rawClip[l][3] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 24));
+ rawClipTmp[0][l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF));
+ rawClipTmp[0][l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawClipTmp[0][l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawClipTmp[0][l][3] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 24));
#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- rawClip[l][4] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 32));
+ rawClipTmp[0][l][4] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 32));
#endif
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- for (int m = 0; m < shuffleTime13FixedBasedLongLength[transposeIdx]; m++)
+ for (int m = 0; m < shuffleTime13FixedBasedLongLength[transposeIdx0]; m++)
#else
- for (int m = 0; m < shuffleTime13LongLength[transposeIdx]; m++)
+ for (int m = 0; m < shuffleTime13LongLength[transposeIdx0]; m++)
#endif
- {
+ {
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- int op0 = shuffleOp13FixedBasedLongLength[transposeIdx][m][0];
- int op1 = shuffleOp13FixedBasedLongLength[transposeIdx][m][1];
+ int op0 = shuffleOp13FixedBasedLongLength[transposeIdx0][m][0];
+ int op1 = shuffleOp13FixedBasedLongLength[transposeIdx0][m][1];
#else
- int op0 = shuffleOp13LongLength[transposeIdx][m][0];
- int op1 = shuffleOp13LongLength[transposeIdx][m][1];
+ int op0 = shuffleOp13LongLength[transposeIdx0][m][0];
+ int op1 = shuffleOp13LongLength[transposeIdx0][m][1];
#endif
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- s0 = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx][m][0]);
- s1 = _mm_xor_si128(s0, _mm_set1_epi8((char) 0x80));
- s2 = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx][m][1]);
- s3 = _mm_xor_si128(s2, _mm_set1_epi8((char) 0x80));
+ s0Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx0][m][0]);
+ s1Tmp[0] = _mm_xor_si128(s0Tmp[0], _mm_set1_epi8((char) 0x80));
+ s2Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx0][m][1]);
+ s3Tmp[0] = _mm_xor_si128(s2Tmp[0], _mm_set1_epi8((char) 0x80));
+#else
+ s0Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx0][m][0]);
+ s1Tmp[0] = _mm_xor_si128(s0Tmp[0], _mm_set1_epi8((char) 0x80));
+ s2Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx0][m][1]);
+ s3Tmp[0] = _mm_xor_si128(s2Tmp[0], _mm_set1_epi8((char) 0x80));
+#endif
+
+ __m128i rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[0][l][op0], s0Tmp[0]), _mm_shuffle_epi8(rawCoefTmp[0][l][op1], s1Tmp[0]));
+ __m128i rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[0][l][op0], s2Tmp[0]), _mm_shuffle_epi8(rawCoefTmp[0][l][op1], s3Tmp[0]));
+ rawCoefTmp[0][l][op0] = rawTmp0;
+ rawCoefTmp[0][l][op1] = rawTmp1;
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[0][l][op0], s0Tmp[0]), _mm_shuffle_epi8(rawClipTmp[0][l][op1], s1Tmp[0]));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[0][l][op0], s2Tmp[0]), _mm_shuffle_epi8(rawClipTmp[0][l][op1], s3Tmp[0]));
+ rawClipTmp[0][l][op0] = rawTmp0;
+ rawClipTmp[0][l][op1] = rawTmp1;
+ }
+
+ const int transposeIdx1 = pClass[j + 4 * k + l + 8] & 0x3;
+ const int classIdx1 = pClass[j + 4 * k + l + 8] >> 2;
+
+ rawCoefTmp[1][l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF));
+ rawCoefTmp[1][l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawCoefTmp[1][l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawCoefTmp[1][l][3] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 24));
+#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ rawCoefTmp[1][l][4] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 32));
+#endif
+ rawClipTmp[1][l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF));
+ rawClipTmp[1][l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawClipTmp[1][l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawClipTmp[1][l][3] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 24));
+#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ rawClipTmp[1][l][4] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 32));
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ for (int m = 0; m < shuffleTime13FixedBasedLongLength[transposeIdx1]; m++)
#else
- s0 = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx][m][0]);
- s1 = _mm_xor_si128(s0, _mm_set1_epi8((char) 0x80));
- s2 = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx][m][1]);
- s3 = _mm_xor_si128(s2, _mm_set1_epi8((char) 0x80));
+ for (int m = 0; m < shuffleTime13LongLength[transposeIdx1]; m++)
#endif
+ {
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ int op0 = shuffleOp13FixedBasedLongLength[transposeIdx1][m][0];
+ int op1 = shuffleOp13FixedBasedLongLength[transposeIdx1][m][1];
+#else
+ int op0 = shuffleOp13LongLength[transposeIdx1][m][0];
+ int op1 = shuffleOp13LongLength[transposeIdx1][m][1];
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ s0Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx1][m][0]);
+ s1Tmp[1] = _mm_xor_si128(s0Tmp[1], _mm_set1_epi8((char) 0x80));
+ s2Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx1][m][1]);
+ s3Tmp[1] = _mm_xor_si128(s2Tmp[1], _mm_set1_epi8((char) 0x80));
+#else
+ s0Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx1][m][0]);
+ s1Tmp[1] = _mm_xor_si128(s0Tmp[1], _mm_set1_epi8((char) 0x80));
+ s2Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx1][m][1]);
+ s3Tmp[1] = _mm_xor_si128(s2Tmp[1], _mm_set1_epi8((char) 0x80));
+#endif
+
+ __m128i rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[1][l][op0], s0Tmp[1]), _mm_shuffle_epi8(rawCoefTmp[1][l][op1], s1Tmp[1]));
+ __m128i rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[1][l][op0], s2Tmp[1]), _mm_shuffle_epi8(rawCoefTmp[1][l][op1], s3Tmp[1]));
+ rawCoefTmp[1][l][op0] = rawTmp0;
+ rawCoefTmp[1][l][op1] = rawTmp1;
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[1][l][op0], s0Tmp[1]), _mm_shuffle_epi8(rawClipTmp[1][l][op1], s1Tmp[1]));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[1][l][op0], s2Tmp[1]), _mm_shuffle_epi8(rawClipTmp[1][l][op1], s3Tmp[1]));
+ rawClipTmp[1][l][op0] = rawTmp0;
+ rawClipTmp[1][l][op1] = rawTmp1;
+ }
- rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoef[l][op0], s0), _mm_shuffle_epi8(rawCoef[l][op1], s1));
- rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoef[l][op0], s2), _mm_shuffle_epi8(rawCoef[l][op1], s3));
- rawCoef[l][op0] = rawTmp0;
- rawCoef[l][op1] = rawTmp1;
-
- rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClip[l][op0], s0), _mm_shuffle_epi8(rawClip[l][op1], s1));
- rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClip[l][op0], s2), _mm_shuffle_epi8(rawClip[l][op1], s3));
- rawClip[l][op0] = rawTmp0;
- rawClip[l][op1] = rawTmp1;
- }
- } // for l
+ rawCoef[l][0] = _mm256_castsi128_si256(rawCoefTmp[0][l][0]);
+ rawCoef[l][0] = _mm256_insertf128_si256(rawCoef[l][0], rawCoefTmp[1][l][0], 1);
+ rawCoef[l][1] = _mm256_castsi128_si256(rawCoefTmp[0][l][1]);
+ rawCoef[l][1] = _mm256_insertf128_si256(rawCoef[l][1], rawCoefTmp[1][l][1], 1);
+ rawCoef[l][2] = _mm256_castsi128_si256(rawCoefTmp[0][l][2]);
+ rawCoef[l][2] = _mm256_insertf128_si256(rawCoef[l][2], rawCoefTmp[1][l][2], 1);
+ rawCoef[l][3] = _mm256_castsi128_si256(rawCoefTmp[0][l][3]);
+ rawCoef[l][3] = _mm256_insertf128_si256(rawCoef[l][3], rawCoefTmp[1][l][3], 1);
+ rawCoef[l][4] = _mm256_castsi128_si256(rawCoefTmp[0][l][4]);
+ rawCoef[l][4] = _mm256_insertf128_si256(rawCoef[l][4], rawCoefTmp[1][l][4], 1);
+
+ rawClip[l][0] = _mm256_castsi128_si256(rawClipTmp[0][l][0]);
+ rawClip[l][0] = _mm256_insertf128_si256(rawClip[l][0], rawClipTmp[1][l][0], 1);
+ rawClip[l][1] = _mm256_castsi128_si256(rawClipTmp[0][l][1]);
+ rawClip[l][1] = _mm256_insertf128_si256(rawClip[l][1], rawClipTmp[1][l][1], 1);
+ rawClip[l][2] = _mm256_castsi128_si256(rawClipTmp[0][l][2]);
+ rawClip[l][2] = _mm256_insertf128_si256(rawClip[l][2], rawClipTmp[1][l][2], 1);
+ rawClip[l][3] = _mm256_castsi128_si256(rawClipTmp[0][l][3]);
+ rawClip[l][3] = _mm256_insertf128_si256(rawClip[l][3], rawClipTmp[1][l][3], 1);
+ rawClip[l][4] = _mm256_castsi128_si256(rawClipTmp[0][l][4]);
+ rawClip[l][4] = _mm256_insertf128_si256(rawClip[l][4], rawClipTmp[1][l][4], 1);
+ } // for l
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- int limR, lim0, lim1, lim2, lim3;
+ int limR, lim0, lim1, lim2, lim3;
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- limR = 5, lim0 = 5, lim1 = 5, lim2 = 5, lim3 = 5;
+ limR = 5, lim0 = 5, lim1 = 5, lim2 = 5, lim3 = 5;
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- limR = 5, lim0 = 5, lim1 = 4, lim2 = 4, lim3 = 4;
+ limR = 5, lim0 = 5, lim1 = 4, lim2 = 4, lim3 = 4;
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- limR = 5, lim0 = 5, lim1 = 5, lim2 = 5, lim3 = 4;
+ limR = 5, lim0 = 5, lim1 = 5, lim2 = 5, lim3 = 4;
#else
- limR = 4, lim0 = 4, lim1 = 4, lim2 = 4, lim3 = 4;
+ limR = 4, lim0 = 4, lim1 = 4, lim2 = 4, lim3 = 4;
#endif
- for (unsigned char l = 0; l < limR; l++)
+ for (unsigned char l = 0; l < limR; l++)
#else
- for (unsigned char l = 0; l < 4; l++)
+ for (unsigned char l = 0; l < 4; l++)
#endif
- {
- int m = l << 2;
-#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- if (l < lim0)
{
+ int m = l << 2;
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ if (l < lim0)
+ {
#endif
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0x00), _mm_shuffle_epi32(rawCoef[1][l], 0x00));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0x00), _mm_shuffle_epi32(rawCoef[3][l], 0x00));
- params[k][0][0 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0x00), _mm_shuffle_epi32(rawClip[1][l], 0x00));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0x00), _mm_shuffle_epi32(rawClip[3][l], 0x00));
- params[k][1][0 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0x00), _mm256_shuffle_epi32(rawCoef[1][l], 0x00));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0x00), _mm256_shuffle_epi32(rawCoef[3][l], 0x00));
+ params[k][0][0 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0x00), _mm256_shuffle_epi32(rawClip[1][l], 0x00));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0x00), _mm256_shuffle_epi32(rawClip[3][l], 0x00));
+ params[k][1][0 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- }
- if (l < lim1)
- {
+ }
+ if (l < lim1)
+ {
#endif
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0x55), _mm_shuffle_epi32(rawCoef[1][l], 0x55));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0x55), _mm_shuffle_epi32(rawCoef[3][l], 0x55));
- params[k][0][1 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0x55), _mm_shuffle_epi32(rawClip[1][l], 0x55));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0x55), _mm_shuffle_epi32(rawClip[3][l], 0x55));
- params[k][1][1 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0x55), _mm256_shuffle_epi32(rawCoef[1][l], 0x55));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0x55), _mm256_shuffle_epi32(rawCoef[3][l], 0x55));
+ params[k][0][1 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0x55), _mm256_shuffle_epi32(rawClip[1][l], 0x55));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0x55), _mm256_shuffle_epi32(rawClip[3][l], 0x55));
+ params[k][1][1 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- }
- if (l < lim2)
- {
+ }
+ if (l < lim2)
+ {
#endif
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0xaa), _mm_shuffle_epi32(rawCoef[1][l], 0xaa));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0xaa), _mm_shuffle_epi32(rawCoef[3][l], 0xaa));
- params[k][0][2 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0xaa), _mm_shuffle_epi32(rawClip[1][l], 0xaa));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0xaa), _mm_shuffle_epi32(rawClip[3][l], 0xaa));
- params[k][1][2 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0xaa), _mm256_shuffle_epi32(rawCoef[1][l], 0xaa));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0xaa), _mm256_shuffle_epi32(rawCoef[3][l], 0xaa));
+ params[k][0][2 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0xaa), _mm256_shuffle_epi32(rawClip[1][l], 0xaa));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0xaa), _mm256_shuffle_epi32(rawClip[3][l], 0xaa));
+ params[k][1][2 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- }
- if (l < lim3)
- {
+ }
+ if (l < lim3)
+ {
#endif
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0xff), _mm_shuffle_epi32(rawCoef[1][l], 0xff));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0xff), _mm_shuffle_epi32(rawCoef[3][l], 0xff));
- params[k][0][3 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
- s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0xff), _mm_shuffle_epi32(rawClip[1][l], 0xff));
- s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0xff), _mm_shuffle_epi32(rawClip[3][l], 0xff));
- params[k][1][3 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0xff), _mm256_shuffle_epi32(rawCoef[1][l], 0xff));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0xff), _mm256_shuffle_epi32(rawCoef[3][l], 0xff));
+ params[k][0][3 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0xff), _mm256_shuffle_epi32(rawClip[1][l], 0xff));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0xff), _mm256_shuffle_epi32(rawClip[3][l], 0xff));
+ params[k][1][3 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- }
+ }
#endif
- } // for l
- } // for k
+ } // for l
+ } // for k
- const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
+ const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
#if !JVET_AD0222_ALF_LONG_FIXFILTER
- const Pel *pImg9, *pImg10, *pImg11, *pImg12;
+ const Pel *pImg9, *pImg10, *pImg11, *pImg12;
#endif
- const Pel *pImgP0;
+ const Pel *pImgP0;
- pImg0 = src + j;
- pImg1 = pImg0 + srcStride;
- pImg2 = pImg0 - srcStride;
- pImg3 = pImg1 + srcStride;
- pImg4 = pImg2 - srcStride;
- pImg5 = pImg3 + srcStride;
- pImg6 = pImg4 - srcStride;
- pImg7 = pImg5 + srcStride;
- pImg8 = pImg6 - srcStride;
-#if !JVET_AD0222_ALF_LONG_FIXFILTER
- pImg9 = pImg7 + srcStride;
- pImg10 = pImg8 - srcStride;
- pImg11 = pImg9 + srcStride;
- pImg12 = pImg10 - srcStride;
-#endif
-#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ pImg0 = src + j;
+ pImg1 = pImg0 + srcStride;
+ pImg2 = pImg0 - srcStride;
+ pImg3 = pImg1 + srcStride;
+ pImg4 = pImg2 - srcStride;
+ pImg5 = pImg3 + srcStride;
+ pImg6 = pImg4 - srcStride;
+ pImg7 = pImg5 + srcStride;
+ pImg8 = pImg6 - srcStride;
+#if !JVET_AD0222_ALF_LONG_FIXFILTER
+ pImg9 = pImg7 + srcStride;
+ pImg10 = pImg8 - srcStride;
+ pImg11 = pImg9 + srcStride;
+ pImg12 = pImg10 - srcStride;
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
- int filterSetIdx = 2 + fixedFilterSetIdx;
+ int filterSetIdx = 2 + fixedFilterSetIdx;
#else
- int filterSetIdx = 0 + fixedFilterSetIdx;
+ int filterSetIdx = 0 + fixedFilterSetIdx;
#endif
- const Pel *pImg0FixedBased, *pImg1FixedBased, *pImg2FixedBased, *pImg3FixedBased, *pImg4FixedBased;
+ const Pel *pImg0FixedBased, *pImg1FixedBased, *pImg2FixedBased, *pImg3FixedBased, *pImg4FixedBased;
#if JVET_AD0222_ALF_LONG_FIXFILTER
- const Pel *pImg5FixedBased, *pImg6FixedBased, *pImg7FixedBased, *pImg8FixedBased, *pImg9FixedBased, *pImg10FixedBased, *pImg11FixedBased, *pImg12FixedBased;
+ const Pel *pImg5FixedBased, *pImg6FixedBased, *pImg7FixedBased, *pImg8FixedBased, *pImg9FixedBased,
+ *pImg10FixedBased, *pImg11FixedBased, *pImg12FixedBased;
#endif
- if (isFixedFilterPaddedPerCtu)
- {
- pImg0FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 0] + j + padSize;
- pImg1FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 1] + j + padSize;
- pImg2FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 1] + j + padSize;
- pImg3FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 2] + j + padSize;
- pImg4FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 2] + j + padSize;
+ if (isFixedFilterPaddedPerCtu)
+ {
+ pImg0FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 0] + j + padSize;
+ pImg1FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 1] + j + padSize;
+ pImg2FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 1] + j + padSize;
+ pImg3FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 2] + j + padSize;
+ pImg4FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 2] + j + padSize;
#if JVET_AD0222_ALF_LONG_FIXFILTER
- pImg5FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 3] + j + padSize;
- pImg6FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 3] + j + padSize;
- pImg7FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 4] + j + padSize;
- pImg8FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 4] + j + padSize;
- pImg9FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 5] + j + padSize;
- pImg10FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 5] + j + padSize;
- pImg11FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 6] + j + padSize;
- pImg12FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 6] + j + padSize;
+ pImg5FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 3] + j + padSize;
+ pImg6FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 3] + j + padSize;
+ pImg7FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 4] + j + padSize;
+ pImg8FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 4] + j + padSize;
+ pImg9FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 5] + j + padSize;
+ pImg10FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 5] + j + padSize;
+ pImg11FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 6] + j + padSize;
+ pImg12FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 6] + j + padSize;
#endif
- }
- else
- {
- pImg0FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 0] + blkDst.x + j + padSize;
- pImg1FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 1] + blkDst.x + j + padSize;
- pImg2FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 1] + blkDst.x + j + padSize;
- pImg3FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 2] + blkDst.x + j + padSize;
- pImg4FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 2] + blkDst.x + j + padSize;
+ }
+ else
+ {
+ pImg0FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 0] + blkDst.x + j + padSize;
+ pImg1FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 1] + blkDst.x + j + padSize;
+ pImg2FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 1] + blkDst.x + j + padSize;
+ pImg3FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 2] + blkDst.x + j + padSize;
+ pImg4FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 2] + blkDst.x + j + padSize;
#if JVET_AD0222_ALF_LONG_FIXFILTER
- pImg5FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 3] + blkDst.x + j + padSize;
- pImg6FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 3] + blkDst.x + j + padSize;
- pImg7FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 4] + blkDst.x + j + padSize;
- pImg8FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 4] + blkDst.x + j + padSize;
- pImg9FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 5] + blkDst.x + j + padSize;
- pImg10FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 5] + blkDst.x + j + padSize;
- pImg11FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 6] + blkDst.x + j + padSize;
- pImg12FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 6] + blkDst.x + j + padSize;
+ pImg5FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 3] + blkDst.x + j + padSize;
+ pImg6FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 3] + blkDst.x + j + padSize;
+ pImg7FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 4] + blkDst.x + j + padSize;
+ pImg8FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 4] + blkDst.x + j + padSize;
+ pImg9FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 5] + blkDst.x + j + padSize;
+ pImg10FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 5] + blkDst.x + j + padSize;
+ pImg11FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 6] + blkDst.x + j + padSize;
+ pImg12FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 6] + blkDst.x + j + padSize;
#endif
- }
+ }
#endif
#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- const Pel *pImg0Gauss[NUM_GAUSS_FILTERED_SOURCE];
- const Pel *pImg1Gauss[NUM_GAUSS_FILTERED_SOURCE], *pImg2Gauss[NUM_GAUSS_FILTERED_SOURCE];
- const Pel *pImg3Gauss[NUM_GAUSS_FILTERED_SOURCE], *pImg4Gauss[NUM_GAUSS_FILTERED_SOURCE];
+ const Pel *pImg0Gauss[NUM_GAUSS_FILTERED_SOURCE];
+ const Pel *pImg1Gauss[NUM_GAUSS_FILTERED_SOURCE], *pImg2Gauss[NUM_GAUSS_FILTERED_SOURCE];
+ const Pel *pImg3Gauss[NUM_GAUSS_FILTERED_SOURCE], *pImg4Gauss[NUM_GAUSS_FILTERED_SOURCE];
- for( int gaussIdx = 0; gaussIdx < NUM_GAUSS_FILTERED_SOURCE; gaussIdx++ )
- {
- if( isFixedFilterPaddedPerCtu )
- {
- pImg0Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 0] + j + padSizeGauss;
- pImg1Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 1] + j + padSizeGauss;
- pImg2Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss - 1] + j + padSizeGauss;
- pImg3Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 2] + j + padSizeGauss;
- pImg4Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss - 2] + j + padSizeGauss;
- }
- else
+ for (int gaussIdx = 0; gaussIdx < NUM_GAUSS_FILTERED_SOURCE; gaussIdx++)
{
- pImg0Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 0] + blkDst.x + j + padSizeGauss;
- pImg1Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 1] + blkDst.x + j + padSizeGauss;
- pImg2Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss - 1] + blkDst.x + j + padSizeGauss;
- pImg3Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 2] + blkDst.x + j + padSizeGauss;
- pImg4Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss - 2] + blkDst.x + j + padSizeGauss;
+ if (isFixedFilterPaddedPerCtu)
+ {
+ pImg0Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 0] + j + padSizeGauss;
+ pImg1Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 1] + j + padSizeGauss;
+ pImg2Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss - 1] + j + padSizeGauss;
+ pImg3Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 2] + j + padSizeGauss;
+ pImg4Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss - 2] + j + padSizeGauss;
+ }
+ else
+ {
+ pImg0Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 0] + blkDst.x + j + padSizeGauss;
+ pImg1Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 1] + blkDst.x + j + padSizeGauss;
+ pImg2Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss - 1] + blkDst.x + j + padSizeGauss;
+ pImg3Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 2] + blkDst.x + j + padSizeGauss;
+ pImg4Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss - 2] + blkDst.x + j + padSizeGauss;
+ }
}
- }
#endif
- __m128i cur = _mm_loadu_si128((const __m128i *) pImg0);
- __m128i accumA = mmOffset;
- __m128i accumB = mmOffset;
+ __m256i cur = _mm256_loadu_si256((const __m256i *) pImg0);
+ __m256i accumA = mmOffset;
+ __m256i accumB = mmOffset;
- auto process2coeffs = [&](const int i, const Pel *ptr0, const Pel *ptr1, const Pel *ptr2, const Pel *ptr3)
- {
- const __m128i val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr0), cur);
- const __m128i val10 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr2), cur);
- const __m128i val01 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr1), cur);
- const __m128i val11 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr3), cur);
+ auto process2coeffs = [&](const int i, const Pel *ptr0, const Pel *ptr1, const Pel *ptr2, const Pel *ptr3)
+ {
+ const __m256i val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr0), cur);
+ const __m256i val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr2), cur);
+ const __m256i val01 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr1), cur);
+ const __m256i val11 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr3), cur);
- __m128i val01A = _mm_unpacklo_epi16(val00, val10);
- __m128i val01B = _mm_unpackhi_epi16(val00, val10);
- __m128i val01C = _mm_unpacklo_epi16(val01, val11);
- __m128i val01D = _mm_unpackhi_epi16(val01, val11);
+ __m256i val01A = _mm256_unpacklo_epi16(val00, val10);
+ __m256i val01B = _mm256_unpackhi_epi16(val00, val10);
+ __m256i val01C = _mm256_unpacklo_epi16(val01, val11);
+ __m256i val01D = _mm256_unpackhi_epi16(val01, val11);
- __m128i limit01A = params[0][1][i];
- __m128i limit01B = params[1][1][i];
+ __m256i limit01A = params[0][1][i];
+ __m256i limit01B = params[1][1][i];
- val01A = _mm_min_epi16(val01A, limit01A);
- val01B = _mm_min_epi16(val01B, limit01B);
- val01C = _mm_min_epi16(val01C, limit01A);
- val01D = _mm_min_epi16(val01D, limit01B);
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ val01C = _mm256_min_epi16(val01C, limit01A);
+ val01D = _mm256_min_epi16(val01D, limit01B);
- limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
- limit01B = _mm_sub_epi16(_mm_setzero_si128(), limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
- val01A = _mm_max_epi16(val01A, limit01A);
- val01B = _mm_max_epi16(val01B, limit01B);
- val01C = _mm_max_epi16(val01C, limit01A);
- val01D = _mm_max_epi16(val01D, limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+ val01C = _mm256_max_epi16(val01C, limit01A);
+ val01D = _mm256_max_epi16(val01D, limit01B);
- val01A = _mm_add_epi16(val01A, val01C);
- val01B = _mm_add_epi16(val01B, val01D);
+ val01A = _mm256_add_epi16(val01A, val01C);
+ val01B = _mm256_add_epi16(val01B, val01D);
- const __m128i coeff01A = params[0][0][i];
- const __m128i coeff01B = params[1][0][i];
+ const __m256i coeff01A = params[0][0][i];
+ const __m256i coeff01B = params[1][0][i];
- accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
- accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
- };
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+ };
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
- process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
- process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
- process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
- process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
- process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
- process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
- process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
- process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
- process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
- process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
- process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
- process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
- process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
- process2coeffs(14, pImg3Gauss[0] - 0, pImg4Gauss[0] + 0, pImg1Gauss[0] - 0, pImg2Gauss[0] + 0);
- process2coeffs(15, pImg0Gauss[0] - 2, pImg0Gauss[0] + 2, pImg0Gauss[0] - 1, pImg0Gauss[0] + 1);
+ process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
+ process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
+ process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
+ process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
+ process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
+ process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
+ process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
+ process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
+ process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
+ process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
+ process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+ process2coeffs(14, pImg3Gauss[0] - 0, pImg4Gauss[0] + 0, pImg1Gauss[0] - 0, pImg2Gauss[0] + 0);
+ process2coeffs(15, pImg0Gauss[0] - 2, pImg0Gauss[0] + 2, pImg0Gauss[0] - 1, pImg0Gauss[0] + 1);
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
- process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
- process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
- process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
- process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
- process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
- process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
- process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
- process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
- process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
- process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
- process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
- process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
- process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+ process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
+ process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
+ process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
+ process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
+ process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
+ process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
+ process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
+ process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
+ process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
+ process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
+ process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
- process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
- process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
- process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
- process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
- process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
- process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
- process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
- process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
- process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
- process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
- process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
- process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
- process2coeffs(13, pImg3Gauss[0] - 0, pImg4Gauss[0] + 0, pImg1Gauss[0] - 0, pImg2Gauss[0] + 0);
- process2coeffs(14, pImg0Gauss[0] - 2, pImg0Gauss[0] + 2, pImg0Gauss[0] - 1, pImg0Gauss[0] + 1);
-#else
- process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
- process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
- process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
- process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
- process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
- process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
- process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
- process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
- process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
- process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
+ process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
+ process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
+ process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
+ process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
+ process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
+ process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
+ process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
+ process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
+ process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+ process2coeffs(13, pImg3Gauss[0] - 0, pImg4Gauss[0] + 0, pImg1Gauss[0] - 0, pImg2Gauss[0] + 0);
+ process2coeffs(14, pImg0Gauss[0] - 2, pImg0Gauss[0] + 2, pImg0Gauss[0] - 1, pImg0Gauss[0] + 1);
+#else
+ process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
+ process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
+ process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
+ process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
+ process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
+ process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
+ process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
+ process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
- process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
- process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+ process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
+ process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
+ process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
#endif
#endif
- pImg0 = srcBeforeDb + j;
- pImg1 = pImg0 + srcBeforeDbStride;
- pImg2 = pImg0 - srcBeforeDbStride;
+ pImg0 = srcBeforeDb + j;
+ pImg1 = pImg0 + srcBeforeDbStride;
+ pImg2 = pImg0 - srcBeforeDbStride;
- pImgP0 = srcResi + j;
+ pImgP0 = srcResi + j;
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- process2coeffs(16, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+ process2coeffs(16, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- process2coeffs(14, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+ process2coeffs(14, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- process2coeffs(15, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+ process2coeffs(15, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
#else
- process2coeffs(13, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+ process2coeffs(13, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
#endif
#else
- process2coeffs(10, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+ process2coeffs(10, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
#endif
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- __m128i val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *)(fixedFilterResults[0 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)),
- cur);
- __m128i val10 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *)(fixedFilterResults[2 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)),
- cur);
+ __m256i val00 = _mm256_sub_epi16( _mm256_loadu_si256((const __m256i *) (fixedFilterResults[0 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)), cur);
+ __m256i val10 = _mm256_sub_epi16( _mm256_loadu_si256((const __m256i *) (fixedFilterResults[2 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)), cur);
#else
- __m128i val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) pImg0), cur);
- __m128i val10 = _mm_sub_epi16(
- _mm_loadu_si128((const __m128i *) (fixedFilterResults[fixedFilterSetIdx][blkDst.y + i] + blkDst.x + j)), cur);
+ __m256i val00 = _mm265_sub_epi16(_mm256_loadu_si256((const __m256i *) pImg0), cur);
+ __m256i val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) (fixedFilterResults[fixedFilterSetIdx][blkDst.y + i] + blkDst.x + j)), cur);
#endif
- __m128i val01A = _mm_unpacklo_epi16(val00, val10);
- __m128i val01B = _mm_unpackhi_epi16(val00, val10);
+ __m256i val01A = _mm256_unpacklo_epi16(val00, val10);
+ __m256i val01B = _mm256_unpackhi_epi16(val00, val10);
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- __m128i limit01A = params[0][1][17];
- __m128i limit01B = params[1][1][17];
+ __m256i limit01A = params[0][1][17];
+ __m256i limit01B = params[1][1][17];
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- __m128i limit01A = params[0][1][15];
- __m128i limit01B = params[1][1][15];
+ __m256i limit01A = params[0][1][15];
+ __m256i limit01B = params[1][1][15];
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- __m128i limit01A = params[0][1][16];
- __m128i limit01B = params[1][1][16];
+ __m256i limit01A = params[0][1][16];
+ __m256i limit01B = params[1][1][16];
#else
- __m128i limit01A = params[0][1][14];
- __m128i limit01B = params[1][1][14];
+ __m256i limit01A = params[0][1][14];
+ __m256i limit01B = params[1][1][14];
#endif
#else
- __m128i limit01A = params[0][1][11];
- __m128i limit01B = params[1][1][11];
+ __m256i limit01A = params[0][1][11];
+ __m256i limit01B = params[1][1][11];
#endif
- val01A = _mm_min_epi16(val01A, limit01A);
- val01B = _mm_min_epi16(val01B, limit01B);
- limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
- limit01B = _mm_sub_epi16(_mm_setzero_si128(), limit01B);
- val01A = _mm_max_epi16(val01A, limit01A);
- val01B = _mm_max_epi16(val01B, limit01B);
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- __m128i coeff01A = params[0][0][17];
- __m128i coeff01B = params[1][0][17];
+ __m256i coeff01A = params[0][0][17];
+ __m256i coeff01B = params[1][0][17];
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- __m128i coeff01A = params[0][0][15];
- __m128i coeff01B = params[1][0][15];
+ __m256i coeff01A = params[0][0][15];
+ __m256i coeff01B = params[1][0][15];
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- __m128i coeff01A = params[0][0][16];
- __m128i coeff01B = params[1][0][16];
+ __m256i coeff01A = params[0][0][16];
+ __m256i coeff01B = params[1][0][16];
#else
- __m128i coeff01A = params[0][0][14];
- __m128i coeff01B = params[1][0][14];
+ __m256i coeff01A = params[0][0][14];
+ __m256i coeff01B = params[1][0][14];
#endif
#else
- __m128i coeff01A = params[0][0][11];
- __m128i coeff01B = params[1][0][11];
+ __m256i coeff01A = params[0][0][11];
+ __m256i coeff01B = params[1][0][11];
#endif
- accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
- accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
- // start prediction fixed filter
- __m128i zero = _mm_setzero_si128();
- val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) (pImg0)), cur);
- val10 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) pImgP0), zero);
- val01A = _mm_unpacklo_epi16(val00, val10);
- val01B = _mm_unpackhi_epi16(val00, val10);
+ // start prediction fixed filter
+ __m256i zero = _mm256_setzero_si256();
+ val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) (pImg0)), cur);
+ val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) pImgP0), zero);
+ val01A = _mm256_unpacklo_epi16(val00, val10);
+ val01B = _mm256_unpackhi_epi16(val00, val10);
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- limit01A = params[0][1][18];
- limit01B = params[1][1][18];
+ limit01A = params[0][1][18];
+ limit01B = params[1][1][18];
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- limit01A = params[0][1][16];
- limit01B = params[1][1][16];
+ limit01A = params[0][1][16];
+ limit01B = params[1][1][16];
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- limit01A = params[0][1][17];
- limit01B = params[1][1][17];
+ limit01A = params[0][1][17];
+ limit01B = params[1][1][17];
#else
- limit01A = params[0][1][15];
- limit01B = params[1][1][15];
+ limit01A = params[0][1][15];
+ limit01B = params[1][1][15];
#endif
- val01A = _mm_min_epi16(val01A, limit01A);
- val01B = _mm_min_epi16(val01B, limit01B);
- limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ coeff01A = params[0][0][18];
+ coeff01B = params[1][0][18];
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ coeff01A = params[0][0][16];
+ coeff01B = params[1][0][16];
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ coeff01A = params[0][0][17];
+ coeff01B = params[1][0][17];
+#else
+ coeff01A = params[0][0][15];
+ coeff01B = params[1][0][15];
+#endif
+
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+ // end prediction fixed filter
+#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) (pImg0Gauss[0])), cur);
+ val10 = _mm256_setzero_si256();
+ val01A = _mm256_unpacklo_epi16(val00, val10);
+ val01B = _mm256_unpackhi_epi16(val00, val10);
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ limit01A = params[0][1][19];
+ limit01B = params[1][1][19];
+#else
+ limit01A = params[0][1][18];
+ limit01B = params[1][1][18];
+#endif
+
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ coeff01A = params[0][0][19];
+ coeff01B = params[1][0][19];
+#else
+ coeff01A = params[0][0][18];
+ coeff01B = params[1][0][18];
+#endif
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+#endif
+ accumA = _mm256_srai_epi32(accumA, shift);
+ accumB = _mm256_srai_epi32(accumB, shift);
+
+ accumA = _mm256_packs_epi32(accumA, accumB);
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ if( bScalingCorr )
+ {
+ accumA = _mm256_add_epi16(accumA, curBase);
+ }
+ else
+#endif
+ accumA = _mm256_add_epi16(accumA, cur);
+ accumA = _mm256_min_epi16(mmMax, _mm256_max_epi16(accumA, mmMin));
+
+ _mm256_storeu_si256((__m256i *) (dst + j), accumA);
+ } // for j
+ src += srcStride * stepY;
+ dst += dstStride * stepY;
+ srcBeforeDb += srcBeforeDbStride * stepY;
+ srcResi += srcResiStride * stepY;
+ } // for i
+ }
+ else
+ {
+
+ const __m128i mmOffset = _mm_set1_epi32(round);
+ const __m128i mmMin = _mm_set1_epi16(clpRng.min);
+ const __m128i mmMax = _mm_set1_epi16(clpRng.max);
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ const __m128i curBase = _mm_set1_epi16(currBase);
+#endif
+#endif//Use AVX2 SIMD
+ for (size_t i = 0; i < height; i += stepY)
+ {
+ const AlfClassifier *pClass = classifier[blkDst.y + i] + blkDst.x;
+ for (size_t j = 0; j < width; j += stepX)
+ {
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ __m128i params[2][2][20];
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ __m128i params[2][2][17];
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ __m128i params[2][2][19];
+#else
+ __m128i params[2][2][16];
+#endif
+#else
+ __m128i params[2][2][13];
+#endif
+ for (int k = 0; k < 2; k++)
+ {
+#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ __m128i rawCoef[4][5], rawClip[4][5], s0, s1, s2, s3, rawTmp0, rawTmp1;
+#else
+ __m128i rawCoef[4][4], rawClip[4][4], s0, s1, s2, s3, rawTmp0, rawTmp1;
+#endif
+ for (int l = 0; l < 4; l++)
+ {
+ const int transposeIdx = pClass[j + 4 * k + l] & 0x3;
+ const int classIdx = pClass[j + 4 * k + l] >> 2;
+
+ rawCoef[l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF));
+ rawCoef[l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawCoef[l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawCoef[l][3] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 24));
+#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ rawCoef[l][4] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 32));
+#endif
+ rawClip[l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF));
+ rawClip[l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawClip[l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawClip[l][3] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 24));
+#if JVET_AD0222_ALF_LONG_FIXFILTER || JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ rawClip[l][4] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx * MAX_NUM_ALF_LUMA_COEFF + 32));
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ for (int m = 0; m < shuffleTime13FixedBasedLongLength[transposeIdx]; m++)
+#else
+ for (int m = 0; m < shuffleTime13LongLength[transposeIdx]; m++)
+#endif
+ {
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ int op0 = shuffleOp13FixedBasedLongLength[transposeIdx][m][0];
+ int op1 = shuffleOp13FixedBasedLongLength[transposeIdx][m][1];
+#else
+ int op0 = shuffleOp13LongLength[transposeIdx][m][0];
+ int op1 = shuffleOp13LongLength[transposeIdx][m][1];
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ s0 = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx][m][0]);
+ s1 = _mm_xor_si128(s0, _mm_set1_epi8((char) 0x80));
+ s2 = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx][m][1]);
+ s3 = _mm_xor_si128(s2, _mm_set1_epi8((char) 0x80));
+#else
+ s0 = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx][m][0]);
+ s1 = _mm_xor_si128(s0, _mm_set1_epi8((char) 0x80));
+ s2 = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx][m][1]);
+ s3 = _mm_xor_si128(s2, _mm_set1_epi8((char) 0x80));
+#endif
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoef[l][op0], s0), _mm_shuffle_epi8(rawCoef[l][op1], s1));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoef[l][op0], s2), _mm_shuffle_epi8(rawCoef[l][op1], s3));
+ rawCoef[l][op0] = rawTmp0;
+ rawCoef[l][op1] = rawTmp1;
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClip[l][op0], s0), _mm_shuffle_epi8(rawClip[l][op1], s1));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClip[l][op0], s2), _mm_shuffle_epi8(rawClip[l][op1], s3));
+ rawClip[l][op0] = rawTmp0;
+ rawClip[l][op1] = rawTmp1;
+ }
+ } // for l
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ int limR, lim0, lim1, lim2, lim3;
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ limR = 5, lim0 = 5, lim1 = 5, lim2 = 5, lim3 = 5;
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ limR = 5, lim0 = 5, lim1 = 4, lim2 = 4, lim3 = 4;
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ limR = 5, lim0 = 5, lim1 = 5, lim2 = 5, lim3 = 4;
+#else
+ limR = 4, lim0 = 4, lim1 = 4, lim2 = 4, lim3 = 4;
+#endif
+ for (unsigned char l = 0; l < limR; l++)
+#else
+ for (unsigned char l = 0; l < 4; l++)
+#endif
+ {
+ int m = l << 2;
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ if (l < lim0)
+ {
+#endif
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0x00), _mm_shuffle_epi32(rawCoef[1][l], 0x00));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0x00), _mm_shuffle_epi32(rawCoef[3][l], 0x00));
+ params[k][0][0 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0x00), _mm_shuffle_epi32(rawClip[1][l], 0x00));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0x00), _mm_shuffle_epi32(rawClip[3][l], 0x00));
+ params[k][1][0 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+ if (l < lim1)
+ {
+#endif
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0x55), _mm_shuffle_epi32(rawCoef[1][l], 0x55));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0x55), _mm_shuffle_epi32(rawCoef[3][l], 0x55));
+ params[k][0][1 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0x55), _mm_shuffle_epi32(rawClip[1][l], 0x55));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0x55), _mm_shuffle_epi32(rawClip[3][l], 0x55));
+ params[k][1][1 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+ if (l < lim2)
+ {
+#endif
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0xaa), _mm_shuffle_epi32(rawCoef[1][l], 0xaa));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0xaa), _mm_shuffle_epi32(rawCoef[3][l], 0xaa));
+ params[k][0][2 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0xaa), _mm_shuffle_epi32(rawClip[1][l], 0xaa));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0xaa), _mm_shuffle_epi32(rawClip[3][l], 0xaa));
+ params[k][1][2 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+ if (l < lim3)
+ {
+#endif
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[0][l], 0xff), _mm_shuffle_epi32(rawCoef[1][l], 0xff));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawCoef[2][l], 0xff), _mm_shuffle_epi32(rawCoef[3][l], 0xff));
+ params[k][0][3 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[0][l], 0xff), _mm_shuffle_epi32(rawClip[1][l], 0xff));
+ s1 = _mm_unpacklo_epi64(_mm_shuffle_epi32(rawClip[2][l], 0xff), _mm_shuffle_epi32(rawClip[3][l], 0xff));
+ params[k][1][3 + m] = _mm_blend_epi16(_mm_shuffle_epi32(s0, 0x88), _mm_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+#endif
+ } // for l
+ } // for k
+
+ const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
+#if !JVET_AD0222_ALF_LONG_FIXFILTER
+ const Pel *pImg9, *pImg10, *pImg11, *pImg12;
+#endif
+ const Pel *pImgP0;
+
+ pImg0 = src + j;
+ pImg1 = pImg0 + srcStride;
+ pImg2 = pImg0 - srcStride;
+ pImg3 = pImg1 + srcStride;
+ pImg4 = pImg2 - srcStride;
+ pImg5 = pImg3 + srcStride;
+ pImg6 = pImg4 - srcStride;
+ pImg7 = pImg5 + srcStride;
+ pImg8 = pImg6 - srcStride;
+#if !JVET_AD0222_ALF_LONG_FIXFILTER
+ pImg9 = pImg7 + srcStride;
+ pImg10 = pImg8 - srcStride;
+ pImg11 = pImg9 + srcStride;
+ pImg12 = pImg10 - srcStride;
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
+ int filterSetIdx = 2 + fixedFilterSetIdx;
+#else
+ int filterSetIdx = 0 + fixedFilterSetIdx;
+#endif
+ const Pel *pImg0FixedBased, *pImg1FixedBased, *pImg2FixedBased, *pImg3FixedBased, *pImg4FixedBased;
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ const Pel *pImg5FixedBased, *pImg6FixedBased, *pImg7FixedBased, *pImg8FixedBased, *pImg9FixedBased, *pImg10FixedBased, *pImg11FixedBased, *pImg12FixedBased;
+#endif
+ if (isFixedFilterPaddedPerCtu)
+ {
+ pImg0FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 0] + j + padSize;
+ pImg1FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 1] + j + padSize;
+ pImg2FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 1] + j + padSize;
+ pImg3FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 2] + j + padSize;
+ pImg4FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 2] + j + padSize;
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ pImg5FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 3] + j + padSize;
+ pImg6FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 3] + j + padSize;
+ pImg7FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 4] + j + padSize;
+ pImg8FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 4] + j + padSize;
+ pImg9FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 5] + j + padSize;
+ pImg10FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 5] + j + padSize;
+ pImg11FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 6] + j + padSize;
+ pImg12FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 6] + j + padSize;
+#endif
+ }
+ else
+ {
+ pImg0FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 0] + blkDst.x + j + padSize;
+ pImg1FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 1] + blkDst.x + j + padSize;
+ pImg2FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 1] + blkDst.x + j + padSize;
+ pImg3FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 2] + blkDst.x + j + padSize;
+ pImg4FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 2] + blkDst.x + j + padSize;
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ pImg5FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 3] + blkDst.x + j + padSize;
+ pImg6FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 3] + blkDst.x + j + padSize;
+ pImg7FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 4] + blkDst.x + j + padSize;
+ pImg8FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 4] + blkDst.x + j + padSize;
+ pImg9FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 5] + blkDst.x + j + padSize;
+ pImg10FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 5] + blkDst.x + j + padSize;
+ pImg11FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 6] + blkDst.x + j + padSize;
+ pImg12FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 6] + blkDst.x + j + padSize;
+#endif
+ }
+#endif
+#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ const Pel *pImg0Gauss[NUM_GAUSS_FILTERED_SOURCE];
+ const Pel *pImg1Gauss[NUM_GAUSS_FILTERED_SOURCE], *pImg2Gauss[NUM_GAUSS_FILTERED_SOURCE];
+ const Pel *pImg3Gauss[NUM_GAUSS_FILTERED_SOURCE], *pImg4Gauss[NUM_GAUSS_FILTERED_SOURCE];
+
+ for( int gaussIdx = 0; gaussIdx < NUM_GAUSS_FILTERED_SOURCE; gaussIdx++ )
+ {
+ if( isFixedFilterPaddedPerCtu )
+ {
+ pImg0Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 0] + j + padSizeGauss;
+ pImg1Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 1] + j + padSizeGauss;
+ pImg2Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss - 1] + j + padSizeGauss;
+ pImg3Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 2] + j + padSizeGauss;
+ pImg4Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss - 2] + j + padSizeGauss;
+ }
+ else
+ {
+ pImg0Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 0] + blkDst.x + j + padSizeGauss;
+ pImg1Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 1] + blkDst.x + j + padSizeGauss;
+ pImg2Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss - 1] + blkDst.x + j + padSizeGauss;
+ pImg3Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 2] + blkDst.x + j + padSizeGauss;
+ pImg4Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss - 2] + blkDst.x + j + padSizeGauss;
+ }
+ }
+#endif
+ __m128i cur = _mm_loadu_si128((const __m128i *) pImg0);
+ __m128i accumA = mmOffset;
+ __m128i accumB = mmOffset;
+
+ auto process2coeffs = [&](const int i, const Pel *ptr0, const Pel *ptr1, const Pel *ptr2, const Pel *ptr3)
+ {
+ const __m128i val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr0), cur);
+ const __m128i val10 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr2), cur);
+ const __m128i val01 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr1), cur);
+ const __m128i val11 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) ptr3), cur);
+
+ __m128i val01A = _mm_unpacklo_epi16(val00, val10);
+ __m128i val01B = _mm_unpackhi_epi16(val00, val10);
+ __m128i val01C = _mm_unpacklo_epi16(val01, val11);
+ __m128i val01D = _mm_unpackhi_epi16(val01, val11);
+
+ __m128i limit01A = params[0][1][i];
+ __m128i limit01B = params[1][1][i];
+
+ val01A = _mm_min_epi16(val01A, limit01A);
+ val01B = _mm_min_epi16(val01B, limit01B);
+ val01C = _mm_min_epi16(val01C, limit01A);
+ val01D = _mm_min_epi16(val01D, limit01B);
+
+ limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
+ limit01B = _mm_sub_epi16(_mm_setzero_si128(), limit01B);
+
+ val01A = _mm_max_epi16(val01A, limit01A);
+ val01B = _mm_max_epi16(val01B, limit01B);
+ val01C = _mm_max_epi16(val01C, limit01A);
+ val01D = _mm_max_epi16(val01D, limit01B);
+
+ val01A = _mm_add_epi16(val01A, val01C);
+ val01B = _mm_add_epi16(val01B, val01D);
+
+ const __m128i coeff01A = params[0][0][i];
+ const __m128i coeff01B = params[1][0][i];
+
+ accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
+ accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
+ };
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
+ process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
+ process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
+ process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
+ process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
+ process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
+ process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
+ process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
+ process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
+ process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
+ process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+ process2coeffs(14, pImg3Gauss[0] - 0, pImg4Gauss[0] + 0, pImg1Gauss[0] - 0, pImg2Gauss[0] + 0);
+ process2coeffs(15, pImg0Gauss[0] - 2, pImg0Gauss[0] + 2, pImg0Gauss[0] - 1, pImg0Gauss[0] + 1);
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
+ process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
+ process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
+ process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
+ process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
+ process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
+ process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
+ process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
+ process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
+ process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
+ process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
+ process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
+ process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
+ process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
+ process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
+ process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
+ process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
+ process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
+ process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
+ process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+ process2coeffs(13, pImg3Gauss[0] - 0, pImg4Gauss[0] + 0, pImg1Gauss[0] - 0, pImg2Gauss[0] + 0);
+ process2coeffs(14, pImg0Gauss[0] - 2, pImg0Gauss[0] + 2, pImg0Gauss[0] - 1, pImg0Gauss[0] + 1);
+#else
+ process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
+ process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
+ process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
+ process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
+ process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
+ process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
+ process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
+ process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
+ process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
+ process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
+#endif
+#endif
+ pImg0 = srcBeforeDb + j;
+ pImg1 = pImg0 + srcBeforeDbStride;
+ pImg2 = pImg0 - srcBeforeDbStride;
+
+ pImgP0 = srcResi + j;
+
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ process2coeffs(16, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ process2coeffs(14, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ process2coeffs(15, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+#else
+ process2coeffs(13, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+#endif
+#else
+ process2coeffs(10, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+#endif
+
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ __m128i val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *)(fixedFilterResults[0 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)),
+ cur);
+ __m128i val10 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *)(fixedFilterResults[2 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)),
+ cur);
+#else
+ __m128i val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) pImg0), cur);
+ __m128i val10 = _mm_sub_epi16(
+ _mm_loadu_si128((const __m128i *) (fixedFilterResults[fixedFilterSetIdx][blkDst.y + i] + blkDst.x + j)), cur);
+#endif
+ __m128i val01A = _mm_unpacklo_epi16(val00, val10);
+ __m128i val01B = _mm_unpackhi_epi16(val00, val10);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ __m128i limit01A = params[0][1][17];
+ __m128i limit01B = params[1][1][17];
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ __m128i limit01A = params[0][1][15];
+ __m128i limit01B = params[1][1][15];
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ __m128i limit01A = params[0][1][16];
+ __m128i limit01B = params[1][1][16];
+#else
+ __m128i limit01A = params[0][1][14];
+ __m128i limit01B = params[1][1][14];
+#endif
+#else
+ __m128i limit01A = params[0][1][11];
+ __m128i limit01B = params[1][1][11];
+#endif
+ val01A = _mm_min_epi16(val01A, limit01A);
+ val01B = _mm_min_epi16(val01B, limit01B);
+ limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
+ limit01B = _mm_sub_epi16(_mm_setzero_si128(), limit01B);
+ val01A = _mm_max_epi16(val01A, limit01A);
+ val01B = _mm_max_epi16(val01B, limit01B);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ __m128i coeff01A = params[0][0][17];
+ __m128i coeff01B = params[1][0][17];
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ __m128i coeff01A = params[0][0][15];
+ __m128i coeff01B = params[1][0][15];
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ __m128i coeff01A = params[0][0][16];
+ __m128i coeff01B = params[1][0][16];
+#else
+ __m128i coeff01A = params[0][0][14];
+ __m128i coeff01B = params[1][0][14];
+#endif
+#else
+ __m128i coeff01A = params[0][0][11];
+ __m128i coeff01B = params[1][0][11];
+#endif
+ accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
+ accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
+
+ // start prediction fixed filter
+ __m128i zero = _mm_setzero_si128();
+ val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) (pImg0)), cur);
+ val10 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) pImgP0), zero);
+ val01A = _mm_unpacklo_epi16(val00, val10);
+ val01B = _mm_unpackhi_epi16(val00, val10);
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ limit01A = params[0][1][18];
+ limit01B = params[1][1][18];
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ limit01A = params[0][1][16];
+ limit01B = params[1][1][16];
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ limit01A = params[0][1][17];
+ limit01B = params[1][1][17];
+#else
+ limit01A = params[0][1][15];
+ limit01B = params[1][1][15];
+#endif
+
+ val01A = _mm_min_epi16(val01A, limit01A);
+ val01B = _mm_min_epi16(val01B, limit01B);
+ limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
+ limit01B = _mm_sub_epi16(_mm_setzero_si128(), limit01B);
+ val01A = _mm_max_epi16(val01A, limit01A);
+ val01B = _mm_max_epi16(val01B, limit01B);
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ coeff01A = params[0][0][18];
+ coeff01B = params[1][0][18];
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ coeff01A = params[0][0][16];
+ coeff01B = params[1][0][16];
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ coeff01A = params[0][0][17];
+ coeff01B = params[1][0][17];
+#else
+ coeff01A = params[0][0][15];
+ coeff01B = params[1][0][15];
+#endif
+
+ accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
+ accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
+ // end prediction fixed filter
+#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) (pImg0Gauss[0])), cur);
+ val10 = _mm_setzero_si128();
+ val01A = _mm_unpacklo_epi16(val00, val10);
+ val01B = _mm_unpackhi_epi16(val00, val10);
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ limit01A = params[0][1][19];
+ limit01B = params[1][1][19];
+#else
+ limit01A = params[0][1][18];
+ limit01B = params[1][1][18];
+#endif
+
+ val01A = _mm_min_epi16(val01A, limit01A);
+ val01B = _mm_min_epi16(val01B, limit01B);
+ limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
limit01B = _mm_sub_epi16(_mm_setzero_si128(), limit01B);
val01A = _mm_max_epi16(val01A, limit01A);
val01B = _mm_max_epi16(val01B, limit01B);
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ coeff01A = params[0][0][19];
+ coeff01B = params[1][0][19];
+#else
+ coeff01A = params[0][0][18];
+ coeff01B = params[1][0][18];
+#endif
+ accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
+ accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
+#endif
+ accumA = _mm_srai_epi32(accumA, shift);
+ accumB = _mm_srai_epi32(accumB, shift);
+
+ accumA = _mm_packs_epi32(accumA, accumB);
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ if ( bScalingCorr )
+ {
+ accumA = _mm_add_epi16(accumA, curBase);
+ }
+ else
+#endif
+ accumA = _mm_add_epi16(accumA, cur);
+ accumA = _mm_min_epi16(mmMax, _mm_max_epi16(accumA, mmMin));
+
+ _mm_storeu_si128((__m128i *) (dst + j), accumA);
+ } // for j
+ src += srcStride * stepY;
+ dst += dstStride * stepY;
+ srcBeforeDb += srcBeforeDbStride * stepY;
+ srcResi += srcResiStride * stepY;
+ } // for i
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ }//Use 256 Bit Simd
+ #endif
+}
+
+template<X86_VEXT vext>
+static void simdFilter13x13BlkExtDbResi(
+ AlfClassifier * *classifier, const PelUnitBuf &recDst, const PelUnitBuf &recBeforeDb, const PelUnitBuf &resi,
+ const CPelUnitBuf &recSrc, const Area &blkDst, const Area &blk, const ComponentID compId, const short *filterSet,
+#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+ const Pel *fClipSet
+#else
+ const short *fClipSet
+#endif
+ ,const ClpRng &clpRng, CodingStructure &cs, Pel ***fixedFilterResults, Pel ***fixedFilterResiResults, int fixedFilterSetIdx
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ ,Pel ***fixedFilterResultsPerCtu, bool isFixedFilterPaddedPerCtu
+#endif
+#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ , Pel ***gaussPic, Pel ***gaussCtu
+#endif
+#if JVET_AG0158_ALF_LUMA_COEFF_PRECISION
+ , char coeffBits
+#endif
+)
+{
+ const CPelBuf srcBuffer = recSrc.get(compId);
+ PelBuf dstBuffer = recDst.get(compId);
+ const CPelBuf scrBufferBeforeDb = recBeforeDb.get(compId);
+ const CPelBuf scrBufferResi = resi.get(compId);
+
+ const size_t srcStride = srcBuffer.stride;
+ const size_t dstStride = dstBuffer.stride;
+ const size_t srcBeforeDbStride = scrBufferBeforeDb.stride;
+ const size_t srcResiStride = scrBufferResi.stride;
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ int adjustShift = coeffBits - 1;
+ const bool bScalingCorr = isLuma(compId) && fixedFilterSetIdx < 0;
+ if ( bScalingCorr )
+ {
+ fixedFilterSetIdx = -fixedFilterSetIdx - 1;
+ adjustShift -= shiftPrecis; // add more precision
+ }
+ const int shift = adjustShift;
+ const Pel currBase = 512;
+ int round = 1 << (shift - 1);
+
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
+ __m128i curBase = _mm_set_epi16( currBase, currBase, currBase, currBase, currBase, currBase, currBase, currBase );
+#endif
+#else
+#if JVET_AG0158_ALF_LUMA_COEFF_PRECISION
+ int shift = coeffBits;
+ shift -= 1;
+ int round = 1 << (shift - 1);
+#else
+ constexpr int shift = AdaptiveLoopFilter::m_NUM_BITS - 1;
+ constexpr int round = 1 << (shift - 1);
+#endif
+#endif
+
+ const size_t width = blk.width;
+ const size_t height = blk.height;
+
+ constexpr size_t stepX = 8;
+ size_t stepY = 1;
+
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
+ const __m128i mmOffset = _mm_set1_epi32(round);
+ const __m128i mmMin = _mm_set1_epi16(clpRng.min);
+ const __m128i mmMax = _mm_set1_epi16(clpRng.max);
+#endif
+
+ static_assert(sizeof(*filterSet) == 2, "ALF coeffs must be 16-bit wide");
+ static_assert(sizeof(*fClipSet) == 2, "ALF clip values must be 16-bit wide");
+
+ const Pel *src = srcBuffer.buf + blk.y * srcStride + blk.x;
+ Pel *dst = dstBuffer.buf + blkDst.y * dstStride + blkDst.x;
+ const Pel *srcBeforeDb = scrBufferBeforeDb.buf + blk.y * srcBeforeDbStride + blk.x;
+ const Pel *srcResi = scrBufferResi.buf + blk.y * srcResiStride + blk.x;
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ const int padSize = ALF_PADDING_SIZE_FIXED_RESULTS;
+#endif
+#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ const int padSizeGauss = ALF_PADDING_SIZE_GAUSS_RESULTS;
+#endif
+
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool use256BitSimd = vext >= AVX2 && blkDst.width % 16 == 0 ? true : false;
+
+ if( use256BitSimd )
+ {
+ const __m256i mmOffset = _mm256_set1_epi32(round);
+ const __m256i mmMin = _mm256_set1_epi16(clpRng.min);
+ const __m256i mmMax = _mm256_set1_epi16(clpRng.max);
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ const __m256i curBase = _mm256_set1_epi16(currBase);
+#endif
+ for (size_t i = 0; i < height; i += stepY)
+ {
+ const AlfClassifier *pClass = classifier[blkDst.y + i] + blkDst.x;
+ for (size_t j = 0; j < width; j += stepX * 2)
+ {
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ __m256i params[2][2][18];
+#else
+ __m256i params[2][2][17];
+#endif
+#else
+ __m256i params[2][2][13];
+#endif
+ for (int k = 0; k < 2; k++)
+ {
+ __m256i rawCoef[4][5], rawClip[4][5], s0, s1;
+ __m128i rawCoefTmp[2][4][5], rawClipTmp[2][4][5], s0Tmp[2], s1Tmp[2], s2Tmp[2], s3Tmp[2];
+ for (int l = 0; l < 4; l++)
+ {
+ const int transposeIdx0 = pClass[j + 4 * k + l + 0] & 0x3;
+ const int classIdx0 = pClass[j + 4 * k + l + 0] >> 2;
+
+ rawCoefTmp[0][l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF));
+ rawCoefTmp[0][l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawCoefTmp[0][l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawCoefTmp[0][l][3] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 24));
+ rawCoefTmp[0][l][4] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 32));
+
+ rawClipTmp[0][l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF));
+ rawClipTmp[0][l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawClipTmp[0][l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawClipTmp[0][l][3] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 24));
+ rawClipTmp[0][l][4] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx0 * MAX_NUM_ALF_LUMA_COEFF + 32));
+
+ const int transposeIdx1 = pClass[j + 4 * k + l + 8] & 0x3;
+ const int classIdx1 = pClass[j + 4 * k + l + 8] >> 2;
+
+ rawCoefTmp[1][l][0] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF));
+ rawCoefTmp[1][l][1] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawCoefTmp[1][l][2] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawCoefTmp[1][l][3] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 24));
+ rawCoefTmp[1][l][4] = _mm_loadu_si128((const __m128i *) (filterSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 32));
+
+ rawClipTmp[1][l][0] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF));
+ rawClipTmp[1][l][1] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 8));
+ rawClipTmp[1][l][2] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 16));
+ rawClipTmp[1][l][3] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 24));
+ rawClipTmp[1][l][4] = _mm_loadu_si128((const __m128i *) (fClipSet + classIdx1 * MAX_NUM_ALF_LUMA_COEFF + 32));
+
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ for (int m = 0; m < shuffleTime13FixedBasedLongLength[transposeIdx0]; m++)
+#else
+ for (int m = 0; m < shuffleTime13LongLength[transposeIdx0]; m++)
+#endif
+ {
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ int op0 = shuffleOp13FixedBasedLongLength[transposeIdx0][m][0];
+ int op1 = shuffleOp13FixedBasedLongLength[transposeIdx0][m][1];
+#else
+ int op0 = shuffleOp13LongLength[transposeIdx0][m][0];
+ int op1 = shuffleOp13LongLength[transposeIdx0][m][1];
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ s0Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx0][m][0]);
+ s1Tmp[0] = _mm_xor_si128(s0Tmp[0], _mm_set1_epi8((char) 0x80));
+ s2Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx0][m][1]);
+ s3Tmp[0] = _mm_xor_si128(s2Tmp[0], _mm_set1_epi8((char) 0x80));
+#else
+ s0Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx0][m][0]);
+ s1Tmp[0] = _mm_xor_si128(s0Tmp[0], _mm_set1_epi8((char) 0x80));
+ s2Tmp[0] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx0][m][1]);
+ s3Tmp[0] = _mm_xor_si128(s2Tmp[0], _mm_set1_epi8((char) 0x80));
+#endif
+
+ __m128i rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[0][l][op0], s0Tmp[0]), _mm_shuffle_epi8(rawCoefTmp[0][l][op1], s1Tmp[0]));
+ __m128i rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[0][l][op0], s2Tmp[0]), _mm_shuffle_epi8(rawCoefTmp[0][l][op1], s3Tmp[0]));
+ rawCoefTmp[0][l][op0] = rawTmp0;
+ rawCoefTmp[0][l][op1] = rawTmp1;
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[0][l][op0], s0Tmp[0]), _mm_shuffle_epi8(rawClipTmp[0][l][op1], s1Tmp[0]));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[0][l][op0], s2Tmp[0]), _mm_shuffle_epi8(rawClipTmp[0][l][op1], s3Tmp[0]));
+ rawClipTmp[0][l][op0] = rawTmp0;
+ rawClipTmp[0][l][op1] = rawTmp1;
+ }
+
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ for (int m = 0; m < shuffleTime13FixedBasedLongLength[transposeIdx1]; m++)
+#else
+ for (int m = 0; m < shuffleTime13LongLength[transposeIdx1]; m++)
+#endif
+ {
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ int op0 = shuffleOp13FixedBasedLongLength[transposeIdx1][m][0];
+ int op1 = shuffleOp13FixedBasedLongLength[transposeIdx1][m][1];
+#else
+ int op0 = shuffleOp13LongLength[transposeIdx1][m][0];
+ int op1 = shuffleOp13LongLength[transposeIdx1][m][1];
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ s0Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx1][m][0]);
+ s1Tmp[1] = _mm_xor_si128(s0Tmp[1], _mm_set1_epi8((char) 0x80));
+ s2Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13FixedBasedLongLength[transposeIdx1][m][1]);
+ s3Tmp[1] = _mm_xor_si128(s2Tmp[1], _mm_set1_epi8((char) 0x80));
+#else
+ s0Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx1][m][0]);
+ s1Tmp[1] = _mm_xor_si128(s0Tmp[1], _mm_set1_epi8((char) 0x80));
+ s2Tmp[1] = _mm_loadu_si128((const __m128i *) shuffleTab13LongLength[transposeIdx1][m][1]);
+ s3Tmp[1] = _mm_xor_si128(s2Tmp[1], _mm_set1_epi8((char) 0x80));
+#endif
+
+ __m128i rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[1][l][op0], s0Tmp[1]), _mm_shuffle_epi8(rawCoefTmp[1][l][op1], s1Tmp[1]));
+ __m128i rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawCoefTmp[1][l][op0], s2Tmp[1]), _mm_shuffle_epi8(rawCoefTmp[1][l][op1], s3Tmp[1]));
+ rawCoefTmp[1][l][op0] = rawTmp0;
+ rawCoefTmp[1][l][op1] = rawTmp1;
+
+ rawTmp0 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[1][l][op0], s0Tmp[1]), _mm_shuffle_epi8(rawClipTmp[1][l][op1], s1Tmp[1]));
+ rawTmp1 = _mm_or_si128(_mm_shuffle_epi8(rawClipTmp[1][l][op0], s2Tmp[1]), _mm_shuffle_epi8(rawClipTmp[1][l][op1], s3Tmp[1]));
+ rawClipTmp[1][l][op0] = rawTmp0;
+ rawClipTmp[1][l][op1] = rawTmp1;
+ }
+
+ rawCoef[l][0] = _mm256_castsi128_si256(rawCoefTmp[0][l][0] );
+ rawCoef[l][0] = _mm256_insertf128_si256(rawCoef[l][0], rawCoefTmp[1][l][0], 1 );
+ rawCoef[l][1] = _mm256_castsi128_si256(rawCoefTmp[0][l][1]);
+ rawCoef[l][1] = _mm256_insertf128_si256(rawCoef[l][1], rawCoefTmp[1][l][1], 1);
+ rawCoef[l][2] = _mm256_castsi128_si256(rawCoefTmp[0][l][2]);
+ rawCoef[l][2] = _mm256_insertf128_si256(rawCoef[l][2], rawCoefTmp[1][l][2], 1);
+ rawCoef[l][3] = _mm256_castsi128_si256(rawCoefTmp[0][l][3]);
+ rawCoef[l][3] = _mm256_insertf128_si256(rawCoef[l][3], rawCoefTmp[1][l][3], 1);
+ rawCoef[l][4] = _mm256_castsi128_si256(rawCoefTmp[0][l][4]);
+ rawCoef[l][4] = _mm256_insertf128_si256(rawCoef[l][4], rawCoefTmp[1][l][4], 1);
+
+ rawClip[l][0] = _mm256_castsi128_si256(rawClipTmp[0][l][0]);
+ rawClip[l][0] = _mm256_insertf128_si256(rawClip[l][0], rawClipTmp[1][l][0], 1);
+ rawClip[l][1] = _mm256_castsi128_si256(rawClipTmp[0][l][1]);
+ rawClip[l][1] = _mm256_insertf128_si256(rawClip[l][1], rawClipTmp[1][l][1], 1);
+ rawClip[l][2] = _mm256_castsi128_si256(rawClipTmp[0][l][2]);
+ rawClip[l][2] = _mm256_insertf128_si256(rawClip[l][2], rawClipTmp[1][l][2], 1);
+ rawClip[l][3] = _mm256_castsi128_si256(rawClipTmp[0][l][3]);
+ rawClip[l][3] = _mm256_insertf128_si256(rawClip[l][3], rawClipTmp[1][l][3], 1);
+ rawClip[l][4] = _mm256_castsi128_si256(rawClipTmp[0][l][4]);
+ rawClip[l][4] = _mm256_insertf128_si256(rawClip[l][4], rawClipTmp[1][l][4], 1);
+ } // for l
+
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ int limR, lim0, lim1, lim2, lim3;
+#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ limR = 5, lim0 = 5, lim1 = 5, lim2 = 4, lim3 = 4;
+#elif JVET_AD0222_ALF_LONG_FIXFILTER
+ limR = 5, lim0 = 5, lim1 = 5, lim2 = 4, lim3 = 4;
+#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ limR = 5, lim0 = 5, lim1 = 4, lim2 = 4, lim3 = 4;
+#else
+ limR = 5, lim0 = 5, lim1 = 4, lim2 = 4, lim3 = 4;
+#endif
+ for (unsigned char l = 0; l < limR; l++)
+#else
+ for (unsigned char l = 0; l < 5; l++)
+#endif
+ {
+ int m = l << 2;
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ if (l < lim0)
+ {
+#endif
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0x00), _mm256_shuffle_epi32(rawCoef[1][l], 0x00));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0x00), _mm256_shuffle_epi32(rawCoef[3][l], 0x00));
+ params[k][0][0 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0x00), _mm256_shuffle_epi32(rawClip[1][l], 0x00));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0x00), _mm256_shuffle_epi32(rawClip[3][l], 0x00));
+ params[k][1][0 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+ if (l < lim1)
+ {
+#endif
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0x55), _mm256_shuffle_epi32(rawCoef[1][l], 0x55));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0x55), _mm256_shuffle_epi32(rawCoef[3][l], 0x55));
+ params[k][0][1 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0x55), _mm256_shuffle_epi32(rawClip[1][l], 0x55));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0x55), _mm256_shuffle_epi32(rawClip[3][l], 0x55));
+ params[k][1][1 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+ if (l < lim2)
+ {
+#endif
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0xaa), _mm256_shuffle_epi32(rawCoef[1][l], 0xaa));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0xaa), _mm256_shuffle_epi32(rawCoef[3][l], 0xaa));
+ params[k][0][2 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0xaa), _mm256_shuffle_epi32(rawClip[1][l], 0xaa));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0xaa), _mm256_shuffle_epi32(rawClip[3][l], 0xaa));
+ params[k][1][2 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+ if (l < lim3)
+ {
+#endif
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0xff), _mm256_shuffle_epi32(rawCoef[1][l], 0xff));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0xff), _mm256_shuffle_epi32(rawCoef[3][l], 0xff));
+ params[k][0][3 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0xff), _mm256_shuffle_epi32(rawClip[1][l], 0xff));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0xff), _mm256_shuffle_epi32(rawClip[3][l], 0xff));
+ params[k][1][3 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ }
+#endif
+ } // for l
+ } // for k
+
+ const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
+#if !JVET_AD0222_ALF_LONG_FIXFILTER
+ const Pel *pImg9, *pImg10, *pImg11, *pImg12;
+#endif
+ const Pel *pImgP0;
+
+ pImg0 = src + j;
+ pImg1 = pImg0 + srcStride;
+ pImg2 = pImg0 - srcStride;
+ pImg3 = pImg1 + srcStride;
+ pImg4 = pImg2 - srcStride;
+ pImg5 = pImg3 + srcStride;
+ pImg6 = pImg4 - srcStride;
+ pImg7 = pImg5 + srcStride;
+ pImg8 = pImg6 - srcStride;
+#if !JVET_AD0222_ALF_LONG_FIXFILTER
+ pImg9 = pImg7 + srcStride;
+ pImg10 = pImg8 - srcStride;
+ pImg11 = pImg9 + srcStride;
+ pImg12 = pImg10 - srcStride;
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
+ int filterSetIdx = 2 + fixedFilterSetIdx;
+#else
+ int filterSetIdx = 0 + fixedFilterSetIdx;
+#endif
+ const Pel *pImg0FixedBased, *pImg1FixedBased, *pImg2FixedBased, *pImg3FixedBased, *pImg4FixedBased;
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ const Pel *pImg5FixedBased, *pImg6FixedBased, *pImg7FixedBased, *pImg8FixedBased, *pImg9FixedBased,
+ *pImg10FixedBased, *pImg11FixedBased, *pImg12FixedBased;
+#endif
+ if (isFixedFilterPaddedPerCtu)
+ {
+ pImg0FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 0] + j + padSize;
+ pImg1FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 1] + j + padSize;
+ pImg2FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 1] + j + padSize;
+ pImg3FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 2] + j + padSize;
+ pImg4FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 2] + j + padSize;
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ pImg5FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 3] + j + padSize;
+ pImg6FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 3] + j + padSize;
+ pImg7FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 4] + j + padSize;
+ pImg8FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 4] + j + padSize;
+ pImg9FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 5] + j + padSize;
+ pImg10FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 5] + j + padSize;
+ pImg11FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize + 6] + j + padSize;
+ pImg12FixedBased = fixedFilterResultsPerCtu[filterSetIdx][i + padSize - 6] + j + padSize;
+#endif
+ }
+ else
+ {
+ pImg0FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 0] + blkDst.x + j + padSize;
+ pImg1FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 1] + blkDst.x + j + padSize;
+ pImg2FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 1] + blkDst.x + j + padSize;
+ pImg3FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 2] + blkDst.x + j + padSize;
+ pImg4FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 2] + blkDst.x + j + padSize;
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ pImg5FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 3] + blkDst.x + j + padSize;
+ pImg6FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 3] + blkDst.x + j + padSize;
+ pImg7FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 4] + blkDst.x + j + padSize;
+ pImg8FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 4] + blkDst.x + j + padSize;
+ pImg9FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 5] + blkDst.x + j + padSize;
+ pImg10FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 5] + blkDst.x + j + padSize;
+ pImg11FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize + 6] + blkDst.x + j + padSize;
+ pImg12FixedBased = fixedFilterResults[filterSetIdx][blkDst.y + i + padSize - 6] + blkDst.x + j + padSize;
+#endif
+ }
+#endif
+#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
+ const Pel *pImg0Gauss[NUM_GAUSS_FILTERED_SOURCE];
+
+ for (int gaussIdx = 0; gaussIdx < NUM_GAUSS_FILTERED_SOURCE; gaussIdx++)
+ {
+ if (isFixedFilterPaddedPerCtu)
+ {
+ pImg0Gauss[gaussIdx] = gaussCtu[gaussIdx][i + padSizeGauss + 0] + j + padSizeGauss;
+ }
+ else
+ {
+ pImg0Gauss[gaussIdx] = gaussPic[gaussIdx][blkDst.y + i + padSizeGauss + 0] + blkDst.x + j + padSizeGauss;
+ }
+ }
+#endif
+ __m256i cur = _mm256_loadu_si256((const __m256i *) pImg0);
+ __m256i accumA = mmOffset;
+ __m256i accumB = mmOffset;
+
+ auto process2coeffs = [&](const int i, const Pel *ptr0, const Pel *ptr1, const Pel *ptr2, const Pel *ptr3)
+ {
+ const __m256i val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr0), cur);
+ const __m256i val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr2), cur);
+ const __m256i val01 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr1), cur);
+ const __m256i val11 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr3), cur);
+
+ __m256i val01A = _mm256_unpacklo_epi16(val00, val10);
+ __m256i val01B = _mm256_unpackhi_epi16(val00, val10);
+ __m256i val01C = _mm256_unpacklo_epi16(val01, val11);
+ __m256i val01D = _mm256_unpackhi_epi16(val01, val11);
+
+ __m256i limit01A = params[0][1][i];
+ __m256i limit01B = params[1][1][i];
+
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ val01C = _mm256_min_epi16(val01C, limit01A);
+ val01D = _mm256_min_epi16(val01D, limit01B);
+
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+ val01C = _mm256_max_epi16(val01C, limit01A);
+ val01D = _mm256_max_epi16(val01D, limit01B);
+
+ val01A = _mm256_add_epi16(val01A, val01C);
+ val01B = _mm256_add_epi16(val01B, val01D);
+
+ const __m256i coeff01A = params[0][0][i];
+ const __m256i coeff01B = params[1][0][i];
+
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+ };
#if JVET_AD0222_ALF_LONG_FIXFILTER && JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- coeff01A = params[0][0][18];
- coeff01B = params[1][0][18];
+ process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
+ process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
+ process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
+ process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
+ process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
+ process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
+ process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
+ process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
+ process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
+ process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
+ process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
#elif JVET_AD0222_ALF_LONG_FIXFILTER
- coeff01A = params[0][0][16];
- coeff01B = params[1][0][16];
+ process2coeffs(0, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(1, pImg3 + 0, pImg4 - 0, pImg1 + 1, pImg2 - 1);
+ process2coeffs(2, pImg1 + 0, pImg2 - 0, pImg1 - 1, pImg2 + 1);
+ process2coeffs(3, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(4, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(5, pImg12FixedBased - 0, pImg11FixedBased + 0, pImg10FixedBased - 0, pImg9FixedBased + 0);
+ process2coeffs(6, pImg8FixedBased - 0, pImg7FixedBased + 0, pImg6FixedBased - 0, pImg5FixedBased + 0);
+ process2coeffs(7, pImg4FixedBased - 1, pImg3FixedBased + 1, pImg4FixedBased - 0, pImg3FixedBased + 0);
+ process2coeffs(8, pImg4FixedBased + 1, pImg3FixedBased - 1, pImg2FixedBased - 2, pImg1FixedBased + 2);
+ process2coeffs(9, pImg2FixedBased - 1, pImg1FixedBased + 1, pImg2FixedBased - 0, pImg1FixedBased + 0);
+ process2coeffs(10, pImg2FixedBased + 1, pImg1FixedBased - 1, pImg2FixedBased + 2, pImg1FixedBased - 2);
+ process2coeffs(11, pImg0FixedBased - 6, pImg0FixedBased + 6, pImg0FixedBased - 5, pImg0FixedBased + 5);
+ process2coeffs(12, pImg0FixedBased - 4, pImg0FixedBased + 4, pImg0FixedBased - 3, pImg0FixedBased + 3);
+ process2coeffs(13, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
#elif JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- coeff01A = params[0][0][17];
- coeff01B = params[1][0][17];
+ process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
+ process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
+ process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
+ process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
+ process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
+ process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
+ process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
+ process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+ process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
+ process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
+ process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
#else
- coeff01A = params[0][0][15];
- coeff01B = params[1][0][15];
+ process2coeffs(0, pImg11 + 0, pImg12 - 0, pImg9 + 0, pImg10 - 0);
+ process2coeffs(1, pImg7 + 0, pImg8 - 0, pImg5 + 0, pImg6 - 0);
+ process2coeffs(2, pImg3 + 2, pImg4 - 2, pImg3 + 1, pImg4 - 1);
+ process2coeffs(3, pImg3 + 0, pImg4 - 0, pImg3 - 1, pImg4 + 1);
+ process2coeffs(4, pImg3 - 2, pImg4 + 2, pImg1 + 2, pImg2 - 2);
+ process2coeffs(5, pImg1 + 1, pImg2 - 1, pImg1 + 0, pImg2 - 0);
+ process2coeffs(6, pImg1 - 1, pImg2 + 1, pImg1 - 2, pImg2 + 2);
+ process2coeffs(7, pImg0 + 6, pImg0 - 6, pImg0 + 5, pImg0 - 5);
+ process2coeffs(8, pImg0 + 4, pImg0 - 4, pImg0 + 3, pImg0 - 3);
+ process2coeffs(9, pImg0 + 2, pImg0 - 2, pImg0 + 1, pImg0 - 1);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ process2coeffs(10, pImg4FixedBased - 0, pImg3FixedBased + 0, pImg2FixedBased - 1, pImg1FixedBased + 1);
+ process2coeffs(11, pImg2FixedBased - 0, pImg1FixedBased + 0, pImg2FixedBased + 1, pImg1FixedBased - 1);
+ process2coeffs(12, pImg0FixedBased - 2, pImg0FixedBased + 2, pImg0FixedBased - 1, pImg0FixedBased + 1);
#endif
+#endif
+ pImg0 = srcBeforeDb + j;
+ pImg1 = pImg0 + srcBeforeDbStride;
+ pImg2 = pImg0 - srcBeforeDbStride;
+ pImgP0 = srcResi + j;
- accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
- accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
- // end prediction fixed filter
-#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- val00 = _mm_sub_epi16(_mm_loadu_si128((const __m128i *) (pImg0Gauss[0])), cur);
- val10 = _mm_setzero_si128();
- val01A = _mm_unpacklo_epi16(val00, val10);
- val01B = _mm_unpackhi_epi16(val00, val10);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AD0222_ALF_LONG_FIXFILTER
- limit01A = params[0][1][19];
- limit01B = params[1][1][19];
+ process2coeffs(14, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
#else
- limit01A = params[0][1][18];
- limit01B = params[1][1][18];
+ process2coeffs(13, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
+#endif
+#else
+ process2coeffs(10, pImg1 + 0, pImg2 + 0, pImg0 + 1, pImg0 - 1);
#endif
- val01A = _mm_min_epi16(val01A, limit01A);
- val01B = _mm_min_epi16(val01B, limit01B);
- limit01A = _mm_sub_epi16(_mm_setzero_si128(), limit01A);
- limit01B = _mm_sub_epi16(_mm_setzero_si128(), limit01B);
- val01A = _mm_max_epi16(val01A, limit01A);
- val01B = _mm_max_epi16(val01B, limit01B);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+ __m256i val00 = _mm256_sub_epi16( _mm256_loadu_si256((const __m256i *) (fixedFilterResults[0 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)), cur);
+ __m256i val10 = _mm256_sub_epi16( _mm256_loadu_si256((const __m256i *) (fixedFilterResults[2 + fixedFilterSetIdx][blkDst.y + i + padSize] + blkDst.x + j + padSize)), cur);
+#else
+ __m256i val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) pImg0), cur);
+ __m256i val10 = _mm256_sub_epi16( _mm256_loadu_si256((const __m256i *) (fixedFilterResults[fixedFilterSetIdx][blkDst.y + i] + blkDst.x + j)), cur);
+#endif
+ __m256i val01A = _mm256_unpacklo_epi16(val00, val10);
+ __m256i val01B = _mm256_unpackhi_epi16(val00, val10);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
#if JVET_AD0222_ALF_LONG_FIXFILTER
- coeff01A = params[0][0][19];
- coeff01B = params[1][0][19];
+ __m256i limit01A = params[0][1][15];
+ __m256i limit01B = params[1][1][15];
#else
- coeff01A = params[0][0][18];
- coeff01B = params[1][0][18];
+ __m256i limit01A = params[0][1][14];
+ __m256i limit01B = params[1][1][14];
#endif
- accumA = _mm_add_epi32(accumA, _mm_madd_epi16(val01A, coeff01A));
- accumB = _mm_add_epi32(accumB, _mm_madd_epi16(val01B, coeff01B));
+#else
+ __m256i limit01A = params[0][1][11];
+ __m256i limit01B = params[1][1][11];
#endif
- accumA = _mm_srai_epi32(accumA, shift);
- accumB = _mm_srai_epi32(accumB, shift);
-
- accumA = _mm_packs_epi32(accumA, accumB);
-#if JVET_AI0084_ALF_RESIDUALS_SCALING
- if ( bScalingCorr )
- {
- accumA = _mm_add_epi16(accumA, curBase);
- }
- else
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ __m256i coeff01A = params[0][0][15];
+ __m256i coeff01B = params[1][0][15];
+#else
+ __m256i coeff01A = params[0][0][14];
+ __m256i coeff01B = params[1][0][14];
#endif
- accumA = _mm_add_epi16(accumA, cur);
- accumA = _mm_min_epi16(mmMax, _mm_max_epi16(accumA, mmMin));
+#else
+ __m256i coeff01A = params[0][0][11];
+ __m256i coeff01B = params[1][0][11];
+#endif
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
- _mm_storeu_si128((__m128i *) (dst + j), accumA);
- } // for j
- src += srcStride * stepY;
- dst += dstStride * stepY;
- srcBeforeDb += srcBeforeDbStride * stepY;
- srcResi += srcResiStride * stepY;
- } // for i
-}
+ // start residual fixed filter
+ __m256i zero = _mm256_setzero_si256();
+ val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) (fixedFilterResiResults[1 - fixedFilterSetIdx][blkDst.y + i] + blkDst.x + j)), zero);
+ val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) (pImg0)), cur);
+ val01A = _mm256_unpacklo_epi16(val00, val10);
+ val01B = _mm256_unpackhi_epi16(val00, val10);
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ limit01A = params[0][1][16];
+ limit01B = params[1][1][16];
+#else
+ limit01A = params[0][1][15];
+ limit01B = params[1][1][15];
+#endif
-template<X86_VEXT vext>
-static void simdFilter13x13BlkExtDbResi(
- AlfClassifier * *classifier, const PelUnitBuf &recDst, const PelUnitBuf &recBeforeDb, const PelUnitBuf &resi,
- const CPelUnitBuf &recSrc, const Area &blkDst, const Area &blk, const ComponentID compId, const short *filterSet,
-#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
- const Pel *fClipSet
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ coeff01A = params[0][0][16];
+ coeff01B = params[1][0][16];
#else
- const short *fClipSet
+ coeff01A = params[0][0][15];
+ coeff01B = params[1][0][15];
#endif
- ,const ClpRng &clpRng, CodingStructure &cs, Pel ***fixedFilterResults, Pel ***fixedFilterResiResults, int fixedFilterSetIdx
+
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+ // end residual fixed filter
+
#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- ,Pel ***fixedFilterResultsPerCtu, bool isFixedFilterPaddedPerCtu
-#endif
+ val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) pImgP0), zero);
#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- , Pel ***gaussPic, Pel ***gaussCtu
-#endif
-#if JVET_AG0158_ALF_LUMA_COEFF_PRECISION
- , char coeffBits
+ val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) (pImg0Gauss[0])), cur);
+#else
+ val10 = _mm256_sub_epi16(cur, cur);
#endif
-)
-{
- const CPelBuf srcBuffer = recSrc.get(compId);
- PelBuf dstBuffer = recDst.get(compId);
- const CPelBuf scrBufferBeforeDb = recBeforeDb.get(compId);
- const CPelBuf scrBufferResi = resi.get(compId);
-
- const size_t srcStride = srcBuffer.stride;
- const size_t dstStride = dstBuffer.stride;
- const size_t srcBeforeDbStride = scrBufferBeforeDb.stride;
- const size_t srcResiStride = scrBufferResi.stride;
-#if JVET_AI0084_ALF_RESIDUALS_SCALING
- int adjustShift = coeffBits - 1;
- const bool bScalingCorr = isLuma(compId) && fixedFilterSetIdx < 0;
- if ( bScalingCorr )
- {
- fixedFilterSetIdx = -fixedFilterSetIdx - 1;
- adjustShift -= shiftPrecis; // add more precision
- }
- const int shift = adjustShift;
- const Pel currBase = 512;
- int round = 1 << (shift - 1);
- __m128i curBase = _mm_set_epi16( currBase, currBase, currBase, currBase, currBase, currBase, currBase, currBase );
+ val01A = _mm256_unpacklo_epi16(val00, val10);
+ val01B = _mm256_unpackhi_epi16(val00, val10);
#else
-#if JVET_AG0158_ALF_LUMA_COEFF_PRECISION
- int shift = coeffBits;
- shift -= 1;
- int round = 1 << (shift - 1);
+ __m256i val = _mm256_sub_epi16( _mm256_loadu_si256( (const __m256i *) (fixedFilterResults[EXT_LENGTH + fixedFilterSetIdx][blkDst.y + i] + blkDst.x + j)), cur);
+ val01A = _mm256_shuffle_epi8(val, _mm256_setr_epi8(0, 1, 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, 7, 6, 7, 0, 1, 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6, 7, 6, 7));
+ val01B = _mm256_shuffle_epi8(val, _mm256_setr_epi8(8, 9, 8, 9, 10, 11, 10, 11, 12, 13, 12, 13, 14, 15, 14, 15, 8, 9, 8, 9, 10, 11, 10, 11, 12, 13, 12, 13, 14, 15, 14, 15));
+#endif
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ limit01A = params[0][1][17];
+ limit01B = params[1][1][17];
#else
- constexpr int shift = AdaptiveLoopFilter::m_NUM_BITS - 1;
- constexpr int round = 1 << (shift - 1);
+ limit01A = params[0][1][16];
+ limit01B = params[1][1][16];
+#endif
+#else
+ limit01A = params[0][1][12];
+ limit01B = params[1][1][12];
+#endif
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AD0222_ALF_LONG_FIXFILTER
+ coeff01A = params[0][0][17];
+ coeff01B = params[1][0][17];
+#else
+ coeff01A = params[0][0][16];
+ coeff01B = params[1][0][16];
#endif
+#else
+ coeff01A = params[0][0][12];
+ coeff01B = params[1][0][12];
#endif
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
- const size_t width = blk.width;
- const size_t height = blk.height;
+ accumA = _mm256_srai_epi32(accumA, shift);
+ accumB = _mm256_srai_epi32(accumB, shift);
- constexpr size_t stepX = 8;
- size_t stepY = 1;
+ accumA = _mm256_packs_epi32(accumA, accumB);
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ if( bScalingCorr )
+ {
+ accumA = _mm256_add_epi16(accumA, curBase);
+ }
+ else
+#endif
+ accumA = _mm256_add_epi16(accumA, cur);
+ accumA = _mm256_min_epi16(mmMax, _mm256_max_epi16(accumA, mmMin));
+
+ _mm256_storeu_si256((__m256i *) (dst + j), accumA);
+ } // for j
+ src += srcStride * stepY;
+ dst += dstStride * stepY;
+ srcBeforeDb += srcBeforeDbStride * stepY;
+ srcResi += srcResiStride * stepY;
+ } // for i
+ }
+ else
+ {
const __m128i mmOffset = _mm_set1_epi32(round);
const __m128i mmMin = _mm_set1_epi16(clpRng.min);
const __m128i mmMax = _mm_set1_epi16(clpRng.max);
-
- static_assert(sizeof(*filterSet) == 2, "ALF coeffs must be 16-bit wide");
- static_assert(sizeof(*fClipSet) == 2, "ALF clip values must be 16-bit wide");
-
- const Pel *src = srcBuffer.buf + blk.y * srcStride + blk.x;
- Pel *dst = dstBuffer.buf + blkDst.y * dstStride + blkDst.x;
- const Pel *srcBeforeDb = scrBufferBeforeDb.buf + blk.y * srcBeforeDbStride + blk.x;
- const Pel *srcResi = scrBufferResi.buf + blk.y * srcResiStride + blk.x;
-#if JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
- const int padSize = ALF_PADDING_SIZE_FIXED_RESULTS;
-#endif
-#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
- const int padSizeGauss = ALF_PADDING_SIZE_GAUSS_RESULTS;
+#if JVET_AI0084_ALF_RESIDUALS_SCALING
+ const __m128i curBase = _mm_set1_epi16(currBase);
#endif
+#endif //Use AVX2 SIMD
for (size_t i = 0; i < height; i += stepY)
{
const AlfClassifier *pClass = classifier[blkDst.y + i] + blkDst.x;
@@ -4586,13 +6079,20 @@ static void simdFilter13x13BlkExtDbResi(
srcBeforeDb += srcBeforeDbStride * stepY;
srcResi += srcResiStride * stepY;
} // for i
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ }//Use 256 Bit Simd
+ #endif
}
#endif
#if JVET_AD0222_ADDITONAL_ALF_FIXFILTER
//Gauss Filter
template<X86_VEXT vext>
-static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx )
+static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelBuf &srcLuma, const Area &blkDst, const Area &blk, const ClpRng &clpRng, const Pel clippingValues[4], int filterSetIdx, int storeIdx
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, AlfClassifier** classifierCodingInfo
+#endif
+ )
{
int16_t gaussCoefTable[NUM_GAUSS_FILTERED_SOURCE][25] =
{
@@ -4617,8 +6117,18 @@ static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelB
}
int16_t diffTH = 32;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+
+ const bool useBounCondition = applyCodingInfo && !(!isSpsAdjust && isIntraSlice);
+ const bool useResiCondition = applyCodingInfo && (isSpsAdjust || !isSpsAdjust) && !isIntraSlice && false;
+ const int offsetClipValue = 1 << ( clpRng.bd - 1 );
+#endif
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
const __m128i offsetMax = _mm_set1_epi16(diffTH);
const __m128i offsetMin = _mm_sub_epi16(_mm_setzero_si128(), offsetMax);
+#endif
const CPelBuf srcBuffer = srcLuma;
const int srcStride = srcBuffer.stride;
@@ -4632,9 +6142,11 @@ static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelB
constexpr int stepX = 8;
int stepY = 1;
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
const __m128i mmOffset = _mm_set1_epi32(round);
const __m128i mmMin = _mm_set1_epi16(clpRng.min);
const __m128i mmMax = _mm_set1_epi16(clpRng.max);
+#endif
static_assert(sizeof(*gaussCoefTable[0]) == 2, "ALF coeffs must be 16-bit wide");
static_assert(sizeof(*gaussClipTable ) == 2, "ALF clip values must be 16-bit wide");
@@ -4642,8 +6154,296 @@ static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelB
const Pel *src = srcBuffer.buf + blk.y * srcStride + blk.x;
const int padSizeGauss = ALF_PADDING_SIZE_GAUSS_RESULTS;
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool use256BitSimd = vext >= AVX2 && blkDst.width % 16 == 0 ? true : false;
+
+ if( use256BitSimd )
+ {
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m256i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m256i mmOriOffset;
+ __m256i mmSignOffsetP, mmSignOffsetN;
+ __m256i mmAbsOffset;
+ __m256i mmAdjOffset;
+ __m256i mmZeroVector = _mm256_set1_epi16( 0 );
+ __m256i mm01Vector = _mm256_set1_epi16( 1 );
+ __m256i mm08Vector = _mm256_set1_epi16( 8 );
+ __m256i mm16Vector = _mm256_set1_epi16( 16 );
+ __m256i mmPOffsetClipVector = _mm256_set1_epi16( +offsetClipValue );
+ __m256i mmNOffsetClipVector = _mm256_set1_epi16( -offsetClipValue );
+ // Set Factor
+ __m256i mmBsFactor = isIntraSlice ? _mm256_set1_epi16( 4 + 2 ) : _mm256_set1_epi16( 3 + 2 );
+ __m256i mmResiFactor = isIntraSlice ? _mm256_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0)) : _mm256_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0));
+#endif
+ const __m256i offsetMax = _mm256_set1_epi16(diffTH);
+ const __m256i offsetMin = _mm256_sub_epi16(_mm256_set1_epi16( 0 ), offsetMax);
+ const __m256i mmOffset = _mm256_set1_epi32(round);
+ const __m256i mmMin = _mm256_set1_epi16(clpRng.min);
+ const __m256i mmMax = _mm256_set1_epi16(clpRng.max);
+
+ for (int i = 0; i < height; i += stepY)
+ {
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if( useBounCondition || useResiCondition )
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
+ for (int j = 0; j < width; j += stepX * 2)
+ {
+ __m256i params[2][2][6];
+
+ for (int k = 0; k < 2; k++)
+ {
+ __m256i rawCoef[4][2], rawClip[4][2], s0, s1;
+ __m128i rawCoefTmp[4][2], rawClipTmp[4][2];
+
+ for (int l = 0; l < 4; l++)
+ {
+ rawCoefTmp[l][0] = _mm_loadu_si128((const __m128i *) (gaussCoefTable[filterSetIdx] + 0));
+ rawCoefTmp[l][1] = _mm_loadu_si128((const __m128i *) (gaussCoefTable[filterSetIdx] + 8));
+
+ rawClipTmp[l][0] = _mm_loadu_si128((const __m128i *) (gaussClipTable + 0));
+ rawClipTmp[l][1] = _mm_loadu_si128((const __m128i *) (gaussClipTable + 8));
+
+ rawCoef[l][0] = _mm256_castsi128_si256( rawCoefTmp[l][0]);
+ rawCoef[l][0] = _mm256_insertf128_si256(rawCoef[l][0], rawCoefTmp[l][0], 1);
+ rawCoef[l][1] = _mm256_castsi128_si256( rawCoefTmp[l][1]);
+ rawCoef[l][1] = _mm256_insertf128_si256(rawCoef[l][1], rawCoefTmp[l][1], 1);
+
+ rawClip[l][0] = _mm256_castsi128_si256(rawClipTmp[l][0]);
+ rawClip[l][0] = _mm256_insertf128_si256(rawClip[l][0], rawClipTmp[l][0], 1);
+ rawClip[l][1] = _mm256_castsi128_si256(rawClipTmp[l][1]);
+ rawClip[l][1] = _mm256_insertf128_si256(rawClip[l][1], rawClipTmp[l][1], 1);
+ } // for l
+
+ for (unsigned char l = 0; l < 2; l++)
+ {
+ int m = l << 2;
+
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0x00), _mm256_shuffle_epi32(rawCoef[1][l], 0x00));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0x00), _mm256_shuffle_epi32(rawCoef[3][l], 0x00));
+ params[k][0][0 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0x00), _mm256_shuffle_epi32(rawClip[1][l], 0x00));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0x00), _mm256_shuffle_epi32(rawClip[3][l], 0x00));
+ params[k][1][0 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0x55), _mm256_shuffle_epi32(rawCoef[1][l], 0x55));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0x55), _mm256_shuffle_epi32(rawCoef[3][l], 0x55));
+ params[k][0][1 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0x55), _mm256_shuffle_epi32(rawClip[1][l], 0x55));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0x55), _mm256_shuffle_epi32(rawClip[3][l], 0x55));
+ params[k][1][1 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+
+ if (l < 1)
+ {
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0xaa), _mm256_shuffle_epi32(rawCoef[1][l], 0xaa));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0xaa), _mm256_shuffle_epi32(rawCoef[3][l], 0xaa));
+ params[k][0][2 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0xaa), _mm256_shuffle_epi32(rawClip[1][l], 0xaa));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0xaa), _mm256_shuffle_epi32(rawClip[3][l], 0xaa));
+ params[k][1][2 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[0][l], 0xff), _mm256_shuffle_epi32(rawCoef[1][l], 0xff));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawCoef[2][l], 0xff), _mm256_shuffle_epi32(rawCoef[3][l], 0xff));
+ params[k][0][3 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ s0 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[0][l], 0xff), _mm256_shuffle_epi32(rawClip[1][l], 0xff));
+ s1 = _mm256_unpacklo_epi64(_mm256_shuffle_epi32(rawClip[2][l], 0xff), _mm256_shuffle_epi32(rawClip[3][l], 0xff));
+ params[k][1][3 + m] = _mm256_blend_epi16(_mm256_shuffle_epi32(s0, 0x88), _mm256_shuffle_epi32(s1, 0x88), 0xf0);
+ }
+ } // for l
+ } // for k
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm256_set1_epi16( 0 );
+ mmClassIdxBsN = _mm256_set1_epi16( 0 );
+ if( useBounCondition )
+ {
+ mmClassIdxTmp = _mm256_loadu_si256((const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16(mmClassIdxTmp, 1);
+ mmClassIdxBsN = _mm256_sub_epi16( mm01Vector, mmClassIdxBsP);
+ }
+ mmClassIdxResiP = _mm256_set1_epi16( 0 );
+ mmClassIdxResiN = _mm256_set1_epi16( 0 );
+ if( useResiCondition )
+ {
+ mmClassIdxTmp = _mm256_loadu_si256((const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16(mmClassIdxTmp, 1);
+ mmClassIdxResiP = _mm256_sub_epi16(mmClassIdxTmp, _mm256_add_epi16(mmClassIdxBsP, mmClassIdxBsP));
+ mmClassIdxResiN = _mm256_sub_epi16( mm01Vector, mmClassIdxResiP);
+ }
+#endif
+
+ const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6;
+ pImg0 = src + j;
+ pImg1 = pImg0 + srcStride;
+ pImg2 = pImg0 - srcStride;
+ pImg3 = pImg1 + srcStride;
+ pImg4 = pImg2 - srcStride;
+ pImg5 = pImg3 + srcStride;
+ pImg6 = pImg4 - srcStride;
+
+ __m256i cur = _mm256_loadu_si256((const __m256i *) pImg0);
+ __m256i accumA = mmOffset;
+ __m256i accumB = mmOffset;
+
+ auto process2coeffs = [&](const int i, const Pel *ptr0, const Pel *ptr1, const Pel *ptr2, const Pel *ptr3)
+ {
+ const __m256i val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr0), cur);
+ const __m256i val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr2), cur);
+ const __m256i val01 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr1), cur);
+ const __m256i val11 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr3), cur);
+
+ __m256i val01A = _mm256_unpacklo_epi16(val00, val10);
+ __m256i val01B = _mm256_unpackhi_epi16(val00, val10);
+ __m256i val01C = _mm256_unpacklo_epi16(val01, val11);
+ __m256i val01D = _mm256_unpackhi_epi16(val01, val11);
+
+ __m256i limit01A = params[0][1][i];
+ __m256i limit01B = params[1][1][i];
+
+ val01A = _mm256_min_epi16(val01A, limit01A);
+ val01B = _mm256_min_epi16(val01B, limit01B);
+ val01C = _mm256_min_epi16(val01C, limit01A);
+ val01D = _mm256_min_epi16(val01D, limit01B);
+
+ limit01A = _mm256_sub_epi16(_mm256_setzero_si256(), limit01A);
+ limit01B = _mm256_sub_epi16(_mm256_setzero_si256(), limit01B);
+
+ val01A = _mm256_max_epi16(val01A, limit01A);
+ val01B = _mm256_max_epi16(val01B, limit01B);
+ val01C = _mm256_max_epi16(val01C, limit01A);
+ val01D = _mm256_max_epi16(val01D, limit01B);
+
+ val01A = _mm256_add_epi16(val01A, val01C);
+ val01B = _mm256_add_epi16(val01B, val01D);
+
+ const __m256i coeff01A = params[0][0][i];
+ const __m256i coeff01B = params[1][0][i];
+
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff01A));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff01B));
+ };
+
+ process2coeffs(0, pImg6 - 0, pImg5 + 0, pImg4 - 1, pImg3 + 1);
+ process2coeffs(1, pImg4 - 0, pImg3 + 0, pImg4 + 1, pImg3 - 1);
+ process2coeffs(2, pImg2 - 2, pImg1 + 2, pImg2 - 1, pImg1 + 1);
+ process2coeffs(3, pImg2 - 0, pImg1 + 0, pImg2 + 1, pImg1 - 1);
+ process2coeffs(4, pImg2 + 2, pImg1 - 2, pImg0 - 3, pImg0 + 3);
+ process2coeffs(5, pImg0 - 2, pImg0 + 2, pImg0 - 1, pImg0 + 1);
+
+ accumA = _mm256_srai_epi32(accumA, shift);
+ accumB = _mm256_srai_epi32(accumB, shift);
+
+ accumA = _mm256_packs_epi32(accumA, accumB);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if ( useBounCondition )
+ {
+ accumA = _mm256_min_epi16( mmPOffsetClipVector, accumA);
+ accumA = _mm256_max_epi16( mmNOffsetClipVector, accumA);
+ // accumA is Ori Offset
+ mmOriOffset = accumA;
+ // Calc Sign
+ // P = 1, N = 0
+ mmSignOffsetP = _mm256_abs_epi16( _mm256_cmpgt_epi16(mmOriOffset, mmZeroVector) );
+ // P = 0, N = 1
+ mmSignOffsetN = _mm256_abs_epi16( _mm256_sub_epi16( mm01Vector, mmSignOffsetP));
+ // Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16( mmOriOffset );
+ // BS based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16(mmAbsOffset, _mm256_add_epi16( mm16Vector, mmBsFactor));
+ mmAdjOffset = _mm256_add_epi16(mmAdjOffset, mm08Vector);
+ mmAdjOffset = _mm256_srai_epi16(mmAdjOffset, 4);
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16(mmClassIdxBsP, mmAdjOffset);
+ __m256i mmTmpOrg = _mm256_mullo_epi16(mmClassIdxBsN, mmAbsOffset);
+
+ __m256i mmTmpFin = _mm256_add_epi16(mmTmpAdj, mmTmpOrg);
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16(mmSignOffsetP, mmTmpFin);
+ __m256i mmTmpSignN = _mm256_sub_epi16( mmZeroVector, _mm256_mullo_epi16(mmSignOffsetN, mmTmpFin) );
+
+ accumA = _mm256_add_epi16(mmTmpSignP, mmTmpSignN);
+ }
+
+ if ( useResiCondition )
+ {
+ accumA = _mm256_min_epi16( mmPOffsetClipVector, accumA);
+ accumA = _mm256_max_epi16( mmNOffsetClipVector, accumA);
+ // accumA is Ori Offset
+ mmOriOffset = accumA;
+ // Calc Sign
+ // P = 1, N = 0
+ mmSignOffsetP = _mm256_abs_epi16( _mm256_cmpgt_epi16(mmOriOffset, mmZeroVector));
+ // P = 0, N = 1
+ mmSignOffsetN = _mm256_abs_epi16( _mm256_sub_epi16( mm01Vector, mmSignOffsetP));
+ // Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16(mmOriOffset);
+ // Resi based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16(mmAbsOffset, _mm256_add_epi16( mm16Vector, mmResiFactor));
+ mmAdjOffset = _mm256_add_epi16(mmAdjOffset, mm08Vector);
+ mmAdjOffset = _mm256_srai_epi16(mmAdjOffset, 4);
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16(mmClassIdxResiP, mmAdjOffset);
+ __m256i mmTmpOrg = _mm256_mullo_epi16(mmClassIdxResiN, mmAbsOffset);
+
+ __m256i mmTmpFin = _mm256_add_epi16(mmTmpAdj, mmTmpOrg);
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16(mmSignOffsetP, mmTmpFin);
+ __m256i mmTmpSignN = _mm256_sub_epi16(mmZeroVector, _mm256_mullo_epi16(mmSignOffsetN, mmTmpFin));
+
+ accumA = _mm256_add_epi16(mmTmpSignP, mmTmpSignN);
+ }
+#endif
+ // Clip Offset
+ accumA = _mm256_min_epi16(accumA, offsetMax);
+ accumA = _mm256_max_epi16(accumA, offsetMin);
+
+ accumA = _mm256_add_epi16(accumA, cur);
+ accumA = _mm256_min_epi16(mmMax, _mm256_max_epi16(accumA, mmMin));
+
+ int curY = blkDst.y + i + padSizeGauss;
+ int curX = blkDst.x + j + padSizeGauss;
+
+ _mm256_storeu_si256((__m256i *) (gaussPic[storeIdx][curY] + curX), accumA);
+ } // for j
+ src += srcStride * stepY;
+ } // for i
+ }
+ else //use256BitSimd
+ {
+
+ const __m128i offsetMax = _mm_set1_epi16(diffTH);
+ const __m128i offsetMin = _mm_sub_epi16(_mm_setzero_si128(), offsetMax);
+ const __m128i mmOffset = _mm_set1_epi32(round);
+ const __m128i mmMin = _mm_set1_epi16(clpRng.min);
+ const __m128i mmMax = _mm_set1_epi16(clpRng.max);
+#endif //Use AVX2 SIMD
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m128i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m128i mmOriOffset;
+ __m128i mmSignOffsetP, mmSignOffsetN;
+ __m128i mmAbsOffset;
+ __m128i mmAdjOffset;
+ __m128i mmZeroVector = _mm_set1_epi16( 0 );
+ __m128i mm01Vector = _mm_set1_epi16( 1 );
+ __m128i mm08Vector = _mm_set1_epi16( 8 );
+ __m128i mm16Vector = _mm_set1_epi16( 16 );
+ __m128i mmPOffsetClipVector = _mm_set1_epi16( +offsetClipValue );
+ __m128i mmNOffsetClipVector = _mm_set1_epi16( -offsetClipValue );
+ //Set Factor
+ __m128i mmBsFactor = isIntraSlice ? _mm_set1_epi16( 4 + 2 ) : _mm_set1_epi16( 3 + 2 );
+ __m128i mmResiFactor = isIntraSlice ? _mm_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0) ) : _mm_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0) );
+#endif
for (int i = 0; i < height; i += stepY)
{
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if( useBounCondition || useResiCondition )
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
for (int j = 0; j < width; j += stepX)
{
__m128i params[2][2][6];
@@ -4697,6 +6497,25 @@ static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelB
}
}//for l
}//for k
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm_set1_epi16( 0 );
+ mmClassIdxBsN = _mm_set1_epi16( 0 );
+ if( useBounCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxBsN = _mm_sub_epi16( mm01Vector, mmClassIdxBsP );
+ }
+ mmClassIdxResiP = _mm_set1_epi16( 0 );
+ mmClassIdxResiN = _mm_set1_epi16( 0 );
+ if( useResiCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxResiP = _mm_sub_epi16( mmClassIdxTmp, _mm_add_epi16( mmClassIdxBsP, mmClassIdxBsP) );
+ mmClassIdxResiN = _mm_sub_epi16( mm01Vector, mmClassIdxResiP );
+ }
+#endif
const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6;
pImg0 = src + j;
@@ -4759,6 +6578,65 @@ static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelB
accumB = _mm_srai_epi32(accumB, shift);
accumA = _mm_packs_epi32(accumA, accumB);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //BS based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmBsFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxBsP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxBsN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+
+ if( useResiCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //Resi based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmResiFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxResiP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxResiN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+#endif
//Clip Offset
accumA = _mm_min_epi16(accumA, offsetMax);
@@ -4774,6 +6652,9 @@ static void simdGaussFiltering(CodingStructure &cs, Pel ***gaussPic, const CPelB
}//for j
src += srcStride * stepY;
}//for i
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ }//use256BitSimd
+#endif
}
#endif
@@ -4893,7 +6774,11 @@ static void simdFixFilter13x13Db9Blk( AlfClassifier **classifier, const CPelBuf
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
const CPelBuf &srcLumaBeforeDb,
#endif
- Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4] )
+ Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ )
#else
static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLuma, const Area& curBlk,
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
@@ -4925,6 +6810,15 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
const __m128i mm11 = _mm_set1_epi8(1);
const __m128i mm3 = _mm_set1_epi16(3);
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+ const bool useCodingInfo = true;
+
+ const bool useBounCondition = applyCodingInfo && !( !isSpsAdjust && isIntraSlice ) && useCodingInfo;
+ const bool useResiCondition = applyCodingInfo && (isSpsAdjust || !isSpsAdjust) && !isIntraSlice && useCodingInfo;
+ const int offsetClipValue = 1 << (clpRng.bd - 1);
+#endif
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
const int srcBeforeDbStride = srcLumaBeforeDb.stride;
const Pel *srcBeforeDb = srcLumaBeforeDb.buf + curBlk.y * srcBeforeDbStride + curBlk.x;
@@ -4946,6 +6840,22 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
mmClippingValues256 = _mm256_insertf128_si256(mmClippingValues256, mmClippingValues, 1);
const __m256i mm11 = _mm256_set1_epi8(1);
const __m256i mm3 = _mm256_set1_epi16(3);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m256i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m256i mmOriOffset;
+ __m256i mmSignOffsetP, mmSignOffsetN;
+ __m256i mmAbsOffset;
+ __m256i mmAdjOffset;
+ __m256i mmZeroVector = _mm256_set1_epi16( 0 );
+ __m256i mm01Vector = _mm256_set1_epi16( 1 );
+ __m256i mm08Vector = _mm256_set1_epi16( 8 );
+ __m256i mm16Vector = _mm256_set1_epi16( 16 );
+ __m256i mmPOffsetClipVector = _mm256_set1_epi16( +offsetClipValue );
+ __m256i mmNOffsetClipVector = _mm256_set1_epi16( -offsetClipValue );
+ //Set Factor
+ __m256i mmBsFactor = isIntraSlice ? _mm256_set1_epi16( 4 ) : _mm256_set1_epi16( 3 );
+ __m256i mmResiFactor = isIntraSlice ? _mm256_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0) ) : _mm256_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0) );
+#endif
for (int i = 0; i < height; i += stepY)
{
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
@@ -4953,6 +6863,13 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
#else
const AlfClassifier *pClass = classifier[curBlk.y + i] + curBlk.x;
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if( useBounCondition || useResiCondition )
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
for (int j = 0; j < width; j += stepX * 2)
{
@@ -5125,6 +7042,25 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
params[29] = _mm256_unpackhi_epi64(_mm256_unpacklo_epi32(rawCoef[0][8], rawCoef[1][8]), _mm256_unpacklo_epi32(rawCoef[2][8], rawCoef[3][8]));
params[30] = _mm256_unpacklo_epi64(_mm256_unpackhi_epi32(rawCoef[0][8], rawCoef[1][8]), _mm256_unpackhi_epi32(rawCoef[2][8], rawCoef[3][8]));
params[31] = _mm256_unpackhi_epi64(_mm256_unpackhi_epi32(rawCoef[0][8], rawCoef[1][8]), _mm256_unpackhi_epi32(rawCoef[2][8], rawCoef[3][8]));
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm256_set1_epi16( 0 );
+ mmClassIdxBsN = _mm256_set1_epi16( 0 );
+ if( useBounCondition )
+ {
+ mmClassIdxTmp = _mm256_loadu_si256( (const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxBsN = _mm256_sub_epi16( mm01Vector, mmClassIdxBsP );
+ }
+ mmClassIdxResiP = _mm256_set1_epi16( 0 );
+ mmClassIdxResiN = _mm256_set1_epi16( 0 );
+ if( useResiCondition )
+ {
+ mmClassIdxTmp = _mm256_loadu_si256( (const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxResiP = _mm256_sub_epi16( mmClassIdxTmp, _mm256_add_epi16( mmClassIdxBsP, mmClassIdxBsP ) );
+ mmClassIdxResiN = _mm256_sub_epi16( mm01Vector, mmClassIdxResiP );
+ }
#endif
for (int ii = 0; ii < stepY; ii++)
{
@@ -5269,6 +7205,65 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
accumB = _mm256_srai_epi32(accumB, shift);
accumA = _mm256_blend_epi16(accumA, _mm256_slli_si256(accumB, 2), 0xAA);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ accumA = _mm256_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm256_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm256_abs_epi16( _mm256_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm256_abs_epi16( _mm256_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16( mmOriOffset );
+ //BS based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16( mmAbsOffset, _mm256_add_epi16( mm16Vector, mmBsFactor ) );
+ mmAdjOffset = _mm256_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm256_srai_epi16( mmAdjOffset, 4 );
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16( mmClassIdxBsP, mmAdjOffset );
+ __m256i mmTmpOrg = _mm256_mullo_epi16( mmClassIdxBsN, mmAbsOffset );
+
+ __m256i mmTmpFin = _mm256_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m256i mmTmpSignN = _mm256_sub_epi16( mmZeroVector, _mm256_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm256_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+
+ if( useResiCondition )
+ {
+ accumA = _mm256_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm256_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 0, N = 1
+ mmSignOffsetP = _mm256_abs_epi16( _mm256_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 1, N = 0
+ mmSignOffsetN = _mm256_abs_epi16( _mm256_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16( mmOriOffset );
+ //Resi based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16( mmAbsOffset, _mm256_add_epi16( mm16Vector, mmResiFactor ) );
+ mmAdjOffset = _mm256_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm256_srai_epi16( mmAdjOffset, 4 );
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16( mmClassIdxResiP, mmAdjOffset );
+ __m256i mmTmpOrg = _mm256_mullo_epi16( mmClassIdxResiN, mmAbsOffset );
+
+ __m256i mmTmpFin = _mm256_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m256i mmTmpSignN = _mm256_sub_epi16( mmZeroVector, _mm256_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm256_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+#endif
accumA = _mm256_add_epi16(accumA, cur);
accumA = _mm256_min_epi16(mmMax, _mm256_max_epi16(accumA, mmMin));
@@ -5301,6 +7296,22 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
const __m128i mmClippingValues = _mm_loadl_epi64((const __m128i *)clippingValues);
const __m128i mm11 = _mm_set1_epi8(1);
const __m128i mm3 = _mm_set1_epi16(3);
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m128i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m128i mmOriOffset;
+ __m128i mmSignOffsetP, mmSignOffsetN;
+ __m128i mmAbsOffset;
+ __m128i mmAdjOffset;
+ __m128i mmZeroVector = _mm_set1_epi16( 0 );
+ __m128i mm01Vector = _mm_set1_epi16( 1 );
+ __m128i mm08Vector = _mm_set1_epi16( 8 );
+ __m128i mm16Vector = _mm_set1_epi16( 16 );
+ __m128i mmPOffsetClipVector = _mm_set1_epi16( +offsetClipValue );
+ __m128i mmNOffsetClipVector = _mm_set1_epi16( -offsetClipValue );
+ //Set Factor
+ __m128i mmBsFactor = isIntraSlice ? _mm_set1_epi16( 4 ) : _mm_set1_epi16( 3 );
+ __m128i mmResiFactor = isIntraSlice ? _mm_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0) ) : _mm_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0) );
#endif
for (int i = 0; i < height; i += stepY)
{
@@ -5309,6 +7320,13 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
#else
const AlfClassifier *pClass = classifier[curBlk.y + i] + curBlk.x;
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if( useBounCondition || useResiCondition )
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
for (int j = 0; j < width; j += stepX)
{
@@ -5414,6 +7432,25 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
params[30] = _mm_unpacklo_epi64(_mm_unpackhi_epi32(rawCoef[0][8], rawCoef[1][8]), _mm_unpackhi_epi32(rawCoef[2][8], rawCoef[3][8]));
params[31] = _mm_unpackhi_epi64(_mm_unpackhi_epi32(rawCoef[0][8], rawCoef[1][8]), _mm_unpackhi_epi32(rawCoef[2][8], rawCoef[3][8]));
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm_set1_epi16( 0 );
+ mmClassIdxBsN = _mm_set1_epi16( 0 );
+ if( useBounCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) ( pClassCodingInfo + j ) );
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxBsN = _mm_sub_epi16( mm01Vector, mmClassIdxBsP );
+ }
+ mmClassIdxResiP = _mm_set1_epi16( 0 );
+ mmClassIdxResiN = _mm_set1_epi16( 0 );
+ if( useResiCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) ( pClassCodingInfo + j ) );
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxResiP = _mm_sub_epi16( mmClassIdxTmp , _mm_add_epi16( mmClassIdxBsP, mmClassIdxBsP) );
+ mmClassIdxResiN = _mm_sub_epi16( mm01Vector, mmClassIdxResiP );
+ }
+#endif
for (int ii = 0; ii < stepY; ii++)
{
@@ -5558,6 +7595,65 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
accumB = _mm_srai_epi32(accumB, shift);
accumA = _mm_blend_epi16(accumA, _mm_slli_si128(accumB, 2), 0xAA);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //BS based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmBsFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxBsP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxBsN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+
+ if( useResiCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //Resi based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmResiFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxResiP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxResiN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+#endif
accumA = _mm_add_epi16(accumA, cur);
accumA = _mm_min_epi16(mmMax, _mm_max_epi16(accumA, mmMin));
@@ -5588,7 +7684,11 @@ static void simdFilter13x13Blk( AlfClassifier **classifier, const CPelBuf &srcLu
#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
template<X86_VEXT vext>
-static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area& curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4])
+static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &srcLuma, const Area& curBlk, const Area &blkDst, const CPelBuf &srcLumaBeforeDb, Pel ***fixedFilterResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ )
{
const int srcStride = srcLuma.stride;
constexpr int shift = AdaptiveLoopFilter::m_NUM_BITS_FIXED_FILTER - 1;
@@ -5613,6 +7713,15 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
const __m128i mm11 = _mm_set1_epi8(1);
const __m128i mm3 = _mm_set1_epi16(3);
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+ const bool useCodingInfo = true;
+
+ const bool useBounCondition = applyCodingInfo && !( !isSpsAdjust && isIntraSlice ) && useCodingInfo;
+ const bool useResiCondition = applyCodingInfo && (isSpsAdjust || !isSpsAdjust) && !isIntraSlice && useCodingInfo;
+ const int offsetClipValue = 1 << ( clpRng.bd - 1 );
+#endif
const int srcBeforeDbStride = srcLumaBeforeDb.stride;
const Pel *srcBeforeDb = srcLumaBeforeDb.buf + curBlk.y * srcBeforeDbStride + curBlk.x;
@@ -5630,6 +7739,22 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
mmClippingValues256 = _mm256_insertf128_si256(mmClippingValues256, mmClippingValues, 1);
const __m256i mm11 = _mm256_set1_epi8(1);
const __m256i mm3 = _mm256_set1_epi16(3);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m256i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m256i mmOriOffset;
+ __m256i mmSignOffsetP, mmSignOffsetN;
+ __m256i mmAbsOffset;
+ __m256i mmAdjOffset;
+ __m256i mmZeroVector = _mm256_set1_epi16( 0 );
+ __m256i mm01Vector = _mm256_set1_epi16( 1 );
+ __m256i mm08Vector = _mm256_set1_epi16( 8 );
+ __m256i mm16Vector = _mm256_set1_epi16( 16 );
+ __m256i mmPOffsetClipVector = _mm256_set1_epi16( +offsetClipValue );
+ __m256i mmNOffsetClipVector = _mm256_set1_epi16( -offsetClipValue );
+ //Set Factor
+ __m256i mmBsFactor = isIntraSlice ? _mm256_set1_epi16( 4 ) : _mm256_set1_epi16( 3 );
+ __m256i mmResiFactor = isIntraSlice ? _mm256_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0) ) : _mm256_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0) );
+#endif
for (int i = 0; i < height; i += stepY)
{
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
@@ -5637,6 +7762,13 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
#else
const AlfClassifier *pClass = classifier[curBlk.y + i] + curBlk.x;
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if( useBounCondition || useResiCondition )
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
for (int j = 0; j < width; j += stepX * 2)
{
@@ -5735,7 +7867,25 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
params[18] = _mm256_unpackhi_epi64(_mm256_unpacklo_epi32(rawCoef[0][5], rawCoef[1][5]), _mm256_unpacklo_epi32(rawCoef[2][5], rawCoef[3][5]));
params[19] = _mm256_unpacklo_epi64(_mm256_unpackhi_epi32(rawCoef[0][5], rawCoef[1][5]), _mm256_unpackhi_epi32(rawCoef[2][5], rawCoef[3][5]));
params[20] = _mm256_unpackhi_epi64(_mm256_unpackhi_epi32(rawCoef[0][5], rawCoef[1][5]), _mm256_unpackhi_epi32(rawCoef[2][5], rawCoef[3][5]));
-
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm256_set1_epi16( 0 );
+ mmClassIdxBsN = _mm256_set1_epi16( 0 );
+ if( useBounCondition )
+ {
+ mmClassIdxTmp = _mm256_loadu_si256( (const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxBsN = _mm256_sub_epi16( mm01Vector, mmClassIdxBsP );
+ }
+ mmClassIdxResiP = _mm256_set1_epi16( 0 );
+ mmClassIdxResiN = _mm256_set1_epi16( 0 );
+ if( useResiCondition )
+ {
+ mmClassIdxTmp = _mm256_loadu_si256( (const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxResiP = _mm256_sub_epi16( mmClassIdxTmp, _mm256_add_epi16( mmClassIdxBsP, mmClassIdxBsP ) );
+ mmClassIdxResiN = _mm256_sub_epi16( mm01Vector, mmClassIdxResiP );
+ }
+#endif
for (int ii = 0; ii < stepY; ii++)
{
const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
@@ -5836,6 +7986,65 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
accumB = _mm256_srai_epi32(accumB, shift);
accumA = _mm256_blend_epi16(accumA, _mm256_slli_si256(accumB, 2), 0xAA);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ accumA = _mm256_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm256_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm256_abs_epi16( _mm256_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm256_abs_epi16( _mm256_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16( mmOriOffset );
+ //BS based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16( mmAbsOffset, _mm256_add_epi16( mm16Vector, mmBsFactor ) );
+ mmAdjOffset = _mm256_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm256_srai_epi16( mmAdjOffset, 4 );
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16( mmClassIdxBsP, mmAdjOffset );
+ __m256i mmTmpOrg = _mm256_mullo_epi16( mmClassIdxBsN, mmAbsOffset );
+
+ __m256i mmTmpFin = _mm256_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m256i mmTmpSignN = _mm256_sub_epi16( mmZeroVector, _mm256_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm256_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+
+ if( useResiCondition )
+ {
+ accumA = _mm256_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm256_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm256_abs_epi16( _mm256_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm256_abs_epi16( _mm256_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16( mmOriOffset );
+ //Resi based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16( mmAbsOffset, _mm256_add_epi16( mm16Vector, mmResiFactor ) );
+ mmAdjOffset = _mm256_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm256_srai_epi16( mmAdjOffset, 4 );
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16( mmClassIdxResiP, mmAdjOffset );
+ __m256i mmTmpOrg = _mm256_mullo_epi16( mmClassIdxResiN, mmAbsOffset );
+
+ __m256i mmTmpFin = _mm256_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m256i mmTmpSignN = _mm256_sub_epi16( mmZeroVector, _mm256_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm256_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+#endif
accumA = _mm256_add_epi16(accumA, cur);
accumA = _mm256_min_epi16(mmMax, _mm256_max_epi16(accumA, mmMin));
@@ -5866,6 +8075,22 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
const __m128i mmClippingValues = _mm_loadl_epi64((const __m128i *)clippingValues);
const __m128i mm11 = _mm_set1_epi8(1);
const __m128i mm3 = _mm_set1_epi16(3);
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m128i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m128i mmOriOffset;
+ __m128i mmSignOffsetP, mmSignOffsetN;
+ __m128i mmAbsOffset;
+ __m128i mmAdjOffset;
+ __m128i mmZeroVector = _mm_set1_epi16( 0 );
+ __m128i mm01Vector = _mm_set1_epi16( 1 );
+ __m128i mm08Vector = _mm_set1_epi16( 8 );
+ __m128i mm16Vector = _mm_set1_epi16( 16 );
+ __m128i mmPOffsetClipVector = _mm_set1_epi16( +offsetClipValue );
+ __m128i mmNOffsetClipVector = _mm_set1_epi16( -offsetClipValue );
+ //Set Factor
+ __m128i mmBsFactor = isIntraSlice ? _mm_set1_epi16( 4 ) : _mm_set1_epi16( 3 );
+ __m128i mmResiFactor = isIntraSlice ? _mm_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0) ) : _mm_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0) );
#endif
for (int i = 0; i < height; i += stepY)
{
@@ -5874,6 +8099,13 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
#else
const AlfClassifier *pClass = classifier[curBlk.y + i] + curBlk.x;
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if( useBounCondition || useResiCondition )
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
for (int j = 0; j < width; j += stepX)
{
@@ -5936,6 +8168,25 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
params[19] = _mm_unpacklo_epi64(_mm_unpackhi_epi32(rawCoef[0][5], rawCoef[1][5]), _mm_unpackhi_epi32(rawCoef[2][5], rawCoef[3][5]));
params[20] = _mm_unpackhi_epi64(_mm_unpackhi_epi32(rawCoef[0][5], rawCoef[1][5]), _mm_unpackhi_epi32(rawCoef[2][5], rawCoef[3][5]));
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm_set1_epi16( 0 );
+ mmClassIdxBsN = _mm_set1_epi16( 0 );
+ if( useBounCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxBsN = _mm_sub_epi16( mm01Vector, mmClassIdxBsP );
+ }
+ mmClassIdxResiP = _mm_set1_epi16( 0 );
+ mmClassIdxResiN = _mm_set1_epi16( 0 );
+ if( useResiCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxResiP = _mm_sub_epi16( mmClassIdxTmp, _mm_add_epi16( mmClassIdxBsP, mmClassIdxBsP) );
+ mmClassIdxResiN = _mm_sub_epi16( mm01Vector, mmClassIdxResiP );
+ }
+#endif
for (int ii = 0; ii < stepY; ii++)
{
const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
@@ -6036,6 +8287,65 @@ static void simdFixFilter9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &sr
accumB = _mm_srai_epi32(accumB, shift);
accumA = _mm_blend_epi16(accumA, _mm_slli_si128(accumB, 2), 0xAA);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //BS based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmBsFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxBsP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxBsN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+
+ if( useResiCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //Resi based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmResiFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxResiP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxResiN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+#endif
accumA = _mm_add_epi16(accumA, cur);
accumA = _mm_min_epi16(mmMax, _mm_max_epi16(accumA, mmMin));
@@ -6370,126 +8680,456 @@ static void simdDeriveVariance(const CPelBuf &srcLuma, const Area &blkDst, const
__m128i xx4 = _mm_alignr_epi8(xx8, xx0, 8);
__m128i xx6 = _mm_alignr_epi8(xx8, xx0, 12);
- x0 = _mm_add_epi32(x0, y0);
- s0 = _mm_add_epi32(s0, s2);
+ x0 = _mm_add_epi32(x0, y0);
+ s0 = _mm_add_epi32(s0, s2);
+
+ __m128i x2 = _mm_alignr_epi8(s0, x0, 4);
+ __m128i x4 = _mm_alignr_epi8(s0, x0, 8);
+ __m128i x6 = _mm_alignr_epi8(s0, x0, 12);
+
+ yy0 = _mm_add_epi32(xx0, xx2);
+ xx0 = _mm_add_epi32(xx4, xx6);
+ yy0 = _mm_add_epi32(yy0, xx8);
+
+ y0 = _mm_add_epi32(x0, x2);
+ x4 = _mm_add_epi32(x4, x6);
+ y0 = _mm_add_epi32(y0, s0);
+
+ __m128i sum2 = _mm_add_epi32(yy0, xx0);
+ __m128i sum = _mm_add_epi32(y0, x4);
+
+ x0 = x8;
+ y0 = y8;
+
+ _mm_storeu_si128((__m128i *) &variance[2][iOffset][jOffset], sum);
+ _mm_storeu_si128((__m128i *) &variance[3][iOffset][jOffset], sum2);
+
+ if (i == 8)
+ {
+ x8 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 4][jOffset]);
+ y8 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 4][jOffset]);
+ x6 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 3][jOffset]);
+ __m128i y6 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 3][jOffset]);
+ x4 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 2][jOffset]);
+ __m128i y4 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 2][jOffset]);
+ x2 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 1][jOffset]);
+ __m128i y2 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 1][jOffset]);
+
+ x8 = _mm_add_epi32(sum, x8);
+ y8 = _mm_add_epi32(sum2, y8);
+
+ x4 = _mm_add_epi32(x6, x4);
+ y4 = _mm_add_epi32(y6, y4);
+
+ x2 = _mm_add_epi32(x8, x2);
+ y2 = _mm_add_epi32(y8, y2);
+
+ sum = _mm_add_epi32(x4, x2);
+ sum2 = _mm_add_epi32(y4, y2);
+ _mm_storeu_si128((__m128i *) &variance[0][iOffset - 4][jOffset], sum);
+ _mm_storeu_si128((__m128i *) &variance[1][iOffset - 4][jOffset], sum2);
+
+ sum2 = _mm_mullo_epi32(sum2, n);
+ sum = _mm_mullo_epi32(sum, sum);
+ sum2 = _mm_add_epi32(sum2, o);
+ sum2 = _mm_sub_epi32(sum2, sum);
+ sum2 = _mm_srli_epi32(sum2, 3);
+ sum2 = _mm_mullo_epi32(sum2, m13);
+ sum2 = _mm_srli_epi32(sum2, 14);
+ _mm_storeu_si128((__m128i *) &variance[VARIANCE][iOffset - 4][jOffset], sum2);
+ }
+ else if (i > 8)
+ {
+ x8 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 5][jOffset]);
+ y8 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 5][jOffset]);
+ x6 = _mm_loadu_si128((__m128i *)&variance[0][iOffset - 5][jOffset]);
+ __m128i y6 = _mm_loadu_si128((__m128i *)&variance[1][iOffset - 5][jOffset]);
+
+ x6 = _mm_sub_epi32(x6, x8);
+ y6 = _mm_sub_epi32(y6, y8);
+
+ sum = _mm_add_epi32(x6, sum);
+ sum2 = _mm_add_epi32(y6, sum2);
+ _mm_storeu_si128((__m128i *) &variance[0][iOffset - 4][jOffset], sum);
+ _mm_storeu_si128((__m128i *) &variance[1][iOffset - 4][jOffset], sum2);
+
+ sum2 = _mm_mullo_epi32(sum2, n);
+ sum = _mm_mullo_epi32(sum, sum);
+ sum2 = _mm_add_epi32(sum2, o);
+ sum2 = _mm_sub_epi32(sum2, sum);
+ sum2 = _mm_srli_epi32(sum2, 3);
+ sum2 = _mm_mullo_epi32(sum2, m13);
+ sum2 = _mm_srli_epi32(sum2, 14);
+ _mm_storeu_si128((__m128i *) &variance[VARIANCE][iOffset - 4][jOffset], sum2);
+ }
+ }
+
+ }
+#if USE_AVX2
+ }
+#endif
+}
+#endif
+
+#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
+#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
+template<X86_VEXT vext>
+static void simdFilterResi9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &srcResiLuma, const Area &curBlk, const Area &blkDst, Pel ***fixedFilterResiResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4]
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ , bool applyCodingInfo, CodingStructure &cs, AlfClassifier** classifierCodingInfo
+#endif
+ )
+{
+ const int srcStride = srcResiLuma.stride;
+ constexpr int shift = AdaptiveLoopFilter::m_NUM_BITS_FIXED_FILTER - 1;
+ constexpr int round = 1 << (shift - 1);
+
+ const int width = curBlk.width;
+ const int height = curBlk.height;
+
+ constexpr int stepX = 8;
+ constexpr int stepY = 2;
+
+ const Pel *src = srcResiLuma.buf + curBlk.y * srcStride + curBlk.x;
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool isIntraSlice = cs.slice->isIntra();
+ const bool isSpsAdjust = cs.sps->getAlfLumaFixedFilterAdjust();
+ const bool useCodingInfo = isSpsAdjust ? true : false;
+ const bool useBounCondition = applyCodingInfo && !( !isSpsAdjust && isIntraSlice ) && useCodingInfo;
+ const bool useResiCondition = applyCodingInfo && (isSpsAdjust || !isSpsAdjust) && !isIntraSlice && useCodingInfo;
+ const int offsetClipValue = 1 << ( clpRng.bd - 1 );
+#endif
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
+ const __m128i mmOffset = _mm_set1_epi32(round);
+#endif
+
+ const int clpRngmin = -clpRng.max;
+ const int clpRngmax = clpRng.max;
+#if !( USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION )
+ const __m128i mmMin = _mm_set1_epi16(clpRngmin);
+ const __m128i mmMax = _mm_set1_epi16(clpRngmax);
+
+ const __m128i mmClippingValues = _mm_loadl_epi64((const __m128i *) clippingValues);
+ const __m128i mm11 = _mm_set1_epi8(1);
+ const __m128i mm3 = _mm_set1_epi16(3);
+#endif
+ const std::array<std::array<short, FIX_FILTER_NUM_COEFF_DB_COMBINE_9_DB_9 + 1>, NUM_FIXED_FILTERS>& filterCoeffFixed = packedDataFixedFilters9Db9Combine[fixedFiltQpInd];
+ const Pel zeros[8] = { 0 };
+
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool use256BitSimd = vext >= AVX2 && blkDst.width % 16 == 0;
+
+ if( use256BitSimd )
+ {
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m256i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m256i mmOriOffset;
+ __m256i mmSignOffsetP, mmSignOffsetN;
+ __m256i mmAbsOffset;
+ __m256i mmAdjOffset;
+ __m256i mmZeroVector = _mm256_set1_epi16(0);
+ __m256i mm01Vector = _mm256_set1_epi16(1);
+ __m256i mm08Vector = _mm256_set1_epi16(8);
+ __m256i mm16Vector = _mm256_set1_epi16(16);
+ __m256i mmPOffsetClipVector = _mm256_set1_epi16(+offsetClipValue);
+ __m256i mmNOffsetClipVector = _mm256_set1_epi16(-offsetClipValue);
+ // Set Factor
+ __m256i mmBsFactor = isIntraSlice ? _mm256_set1_epi16( 4 ) : _mm256_set1_epi16( 3 );
+ __m256i mmResiFactor = isIntraSlice ? _mm256_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0)) : _mm256_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0) );
+#endif
+
+ const __m256i mmOffset = _mm256_set1_epi32(round);
+
+ const int clpRngmin = -clpRng.max;
+ const int clpRngmax = clpRng.max;
+ const __m256i mmMin = _mm256_set1_epi16(clpRngmin);
+ const __m256i mmMax = _mm256_set1_epi16(clpRngmax);
+
+ const __m128i mmClippingValues = _mm_loadl_epi64((const __m128i *) clippingValues);
+ __m256i mmClippingValues256 = _mm256_castsi128_si256(mmClippingValues);
+ mmClippingValues256 = _mm256_insertf128_si256(mmClippingValues256, mmClippingValues, 1);
+ const __m256i mm11 = _mm256_set1_epi8(1);
+ const __m256i mm3 = _mm256_set1_epi16(3);
+ const std::array<std::array<short, FIX_FILTER_NUM_COEFF_DB_COMBINE_9_DB_9 + 1>, NUM_FIXED_FILTERS>
+ &filterCoeffFixed = packedDataFixedFilters9Db9Combine[fixedFiltQpInd];
+ const Pel zeros[16] = { 0 };
+
+ for (int i = 0; i < height; i += stepY)
+ {
+#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
+ const AlfClassifier *pClass = classifier[blkDst.y + i] + blkDst.x;
+#else
+ const AlfClassifier *pClass = classifier[curBlk.y + i] + curBlk.x;
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if (useBounCondition || useResiCondition)
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
+
+ for (int j = 0; j < width; j += stepX * 2)
+ {
+ __m256i params[11];
+ __m256i rawCoef[4][3];
+ for (int m = 0; m < 4; m++)
+ {
+ int transposeIdx0 = pClass[j + 2 * m] & 0x3;
+ const int filterIdx0 = classIndFixed[pClass[j + 2 * m] >> 2];
+
+ __m128i rawCoef00 = _mm_loadu_si128((const __m128i *) (filterCoeffFixed[filterIdx0].data()));
+ __m128i rawCoef01 = _mm_loadu_si128((const __m128i *) (filterCoeffFixed[filterIdx0].data() + 8));
+ __m128i rawCoef02 = _mm_loadu_si128((const __m128i *) (filterCoeffFixed[filterIdx0].data() + 14));
+ // transpose0
+ if (transposeIdx0 != 0)
+ {
+ const __m128i s00 = _mm_loadu_si128((const __m128i *) shTab9Db9[transposeIdx0][0]);
+ const __m128i s01 = _mm_loadu_si128((const __m128i *) shTab9Db9[transposeIdx0][1]);
+ const __m128i s02 = _mm_loadu_si128((const __m128i *) shTab9Db9[transposeIdx0][2]);
+
+ rawCoef00 = _mm_shuffle_epi8(rawCoef00, s00);
+ rawCoef01 = _mm_shuffle_epi8(rawCoef01, s01);
+ rawCoef02 = _mm_shuffle_epi8(rawCoef02, s02);
+ }
+
+ int transposeIdx1 = pClass[j + 2 * m + 8] & 0x3;
+ const int filterIdx1 = classIndFixed[pClass[j + 2 * m + 8] >> 2];
+
+ __m128i rawCoef10 = _mm_loadu_si128((const __m128i *) (filterCoeffFixed[filterIdx1].data()));
+ __m128i rawCoef11 = _mm_loadu_si128((const __m128i *) (filterCoeffFixed[filterIdx1].data() + 8));
+ __m128i rawCoef12 = _mm_loadu_si128((const __m128i *) (filterCoeffFixed[filterIdx1].data() + 14));
+ // transpose1
+ if (transposeIdx1 != 0)
+ {
+ const __m128i s10 = _mm_loadu_si128((const __m128i *) shTab9Db9[transposeIdx1][0]);
+ const __m128i s11 = _mm_loadu_si128((const __m128i *) shTab9Db9[transposeIdx1][1]);
+ const __m128i s12 = _mm_loadu_si128((const __m128i *) shTab9Db9[transposeIdx1][2]);
+
+ rawCoef10 = _mm_shuffle_epi8(rawCoef10, s10);
+ rawCoef11 = _mm_shuffle_epi8(rawCoef11, s11);
+ rawCoef12 = _mm_shuffle_epi8(rawCoef12, s12);
+ }
+
+ rawCoef[m][0] = _mm256_castsi128_si256(rawCoef00);
+ rawCoef[m][0] = _mm256_insertf128_si256(rawCoef[m][0], rawCoef10, 1);
+ rawCoef[m][1] = _mm256_castsi128_si256(rawCoef01);
+ rawCoef[m][1] = _mm256_insertf128_si256(rawCoef[m][1], rawCoef11, 1);
+ rawCoef[m][2] = _mm256_castsi128_si256(rawCoef02);
+ rawCoef[m][2] = _mm256_insertf128_si256(rawCoef[m][2], rawCoef12, 1);
+ } // for(m)
+
+ params[0] = _mm256_unpacklo_epi64(_mm256_unpacklo_epi32(rawCoef[0][0], rawCoef[1][0]), _mm256_unpacklo_epi32(rawCoef[2][0], rawCoef[3][0]));
+ params[1] = _mm256_unpackhi_epi64(_mm256_unpacklo_epi32(rawCoef[0][0], rawCoef[1][0]), _mm256_unpacklo_epi32(rawCoef[2][0], rawCoef[3][0]));
+ params[2] = _mm256_unpacklo_epi64(_mm256_unpackhi_epi32(rawCoef[0][0], rawCoef[1][0]), _mm256_unpackhi_epi32(rawCoef[2][0], rawCoef[3][0]));
+ params[3] = _mm256_unpackhi_epi64(_mm256_unpackhi_epi32(rawCoef[0][0], rawCoef[1][0]), _mm256_unpackhi_epi32(rawCoef[2][0], rawCoef[3][0]));
+
+ params[4] = _mm256_unpacklo_epi64(_mm256_unpacklo_epi32(rawCoef[0][1], rawCoef[1][1]), _mm256_unpacklo_epi32(rawCoef[2][1], rawCoef[3][1]));
+ params[5] = _mm256_unpackhi_epi64(_mm256_unpacklo_epi32(rawCoef[0][1], rawCoef[1][1]), _mm256_unpacklo_epi32(rawCoef[2][1], rawCoef[3][1]));
+ params[6] = _mm256_unpacklo_epi64(_mm256_unpackhi_epi32(rawCoef[0][1], rawCoef[1][1]), _mm256_unpackhi_epi32(rawCoef[2][1], rawCoef[3][1]));
+
+ params[7] = _mm256_unpacklo_epi64(_mm256_unpacklo_epi32(rawCoef[0][2], rawCoef[1][2]), _mm256_unpacklo_epi32(rawCoef[2][2], rawCoef[3][2]));
+ params[8] = _mm256_unpackhi_epi64(_mm256_unpacklo_epi32(rawCoef[0][2], rawCoef[1][2]), _mm256_unpacklo_epi32(rawCoef[2][2], rawCoef[3][2]));
+ params[9] = _mm256_unpacklo_epi64(_mm256_unpackhi_epi32(rawCoef[0][2], rawCoef[1][2]), _mm256_unpackhi_epi32(rawCoef[2][2], rawCoef[3][2]));
+ params[10] = _mm256_unpackhi_epi64(_mm256_unpackhi_epi32(rawCoef[0][2], rawCoef[1][2]), _mm256_unpackhi_epi32(rawCoef[2][2], rawCoef[3][2]));
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm256_set1_epi16( 0 );
+ mmClassIdxBsN = _mm256_set1_epi16( 0 );
+ if (useBounCondition)
+ {
+ mmClassIdxTmp = _mm256_loadu_si256((const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16(mmClassIdxTmp, 1);
+ mmClassIdxBsN = _mm256_sub_epi16(mm01Vector, mmClassIdxBsP);
+ }
+ mmClassIdxResiP = _mm256_set1_epi16( 0 );
+ mmClassIdxResiN = _mm256_set1_epi16( 0 );
+ if (useResiCondition)
+ {
+ mmClassIdxTmp = _mm256_loadu_si256((const __m256i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm256_srai_epi16(mmClassIdxTmp, 1);
+ mmClassIdxResiP = _mm256_sub_epi16(mmClassIdxTmp, _mm256_add_epi16(mmClassIdxBsP, mmClassIdxBsP));
+ mmClassIdxResiN = _mm256_sub_epi16(mm01Vector, mmClassIdxResiP);
+ }
+#endif
+ for (int ii = 0; ii < stepY; ii++)
+ {
+ const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
+ pImg0 = src + j + ii * srcStride;
+ pImg1 = pImg0 + srcStride;
+ pImg2 = pImg0 - srcStride;
+ pImg3 = pImg1 + srcStride;
+ pImg4 = pImg2 - srcStride;
+ pImg5 = pImg3 + srcStride;
+ pImg6 = pImg4 - srcStride;
+ pImg7 = pImg5 + srcStride;
+ pImg8 = pImg6 - srcStride;
+
+ __m256i cur = _mm256_loadu_si256((const __m256i *) pImg0);
+ __m256i accumA = mmOffset;
+ __m256i accumB = mmOffset;
- __m128i x2 = _mm_alignr_epi8(s0, x0, 4);
- __m128i x4 = _mm_alignr_epi8(s0, x0, 8);
- __m128i x6 = _mm_alignr_epi8(s0, x0, 12);
+ auto process2coeffs = [&](const int i, const Pel *ptr0, const Pel *ptr1, const Pel *ptr2, const Pel *ptr3)
+ {
+ const __m256i val00 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr0), cur);
+ const __m256i val10 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr2), cur);
+ const __m256i val01 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr1), cur);
+ const __m256i val11 = _mm256_sub_epi16(_mm256_loadu_si256((const __m256i *) ptr3), cur);
- yy0 = _mm_add_epi32(xx0, xx2);
- xx0 = _mm_add_epi32(xx4, xx6);
- yy0 = _mm_add_epi32(yy0, xx8);
+ __m256i val01A = _mm256_blend_epi16(val00, _mm256_slli_si256(val10, 2), 0xAA);
+ __m256i val01B = _mm256_blend_epi16(_mm256_srli_si256(val00, 2), val10, 0xAA);
+ __m256i val01C = _mm256_blend_epi16(val01, _mm256_slli_si256(val11, 2), 0xAA);
+ __m256i val01D = _mm256_blend_epi16(_mm256_srli_si256(val01, 2), val11, 0xAA);
- y0 = _mm_add_epi32(x0, x2);
- x4 = _mm_add_epi32(x4, x6);
- y0 = _mm_add_epi32(y0, s0);
+ __m256i mmClippingFixed = _mm256_and_si256(params[i], mm3);
- __m128i sum2 = _mm_add_epi32(yy0, xx0);
- __m128i sum = _mm_add_epi32(y0, x4);
+ __m256i mmClippingFixed2 = _mm256_packs_epi16(mmClippingFixed, mmClippingFixed);
+ mmClippingFixed2 = _mm256_add_epi8(mmClippingFixed2, mmClippingFixed2);
+ __m256i xmm2 = _mm256_add_epi8(mmClippingFixed2, mm11);
+ __m256i xmmA = _mm256_unpacklo_epi8(mmClippingFixed2, xmm2);
+ __m256i limit = _mm256_shuffle_epi8(mmClippingValues256, xmmA);
- x0 = x8;
- y0 = y8;
+ val01A = _mm256_min_epi16(val01A, limit);
+ val01B = _mm256_min_epi16(val01B, limit);
+ val01C = _mm256_min_epi16(val01C, limit);
+ val01D = _mm256_min_epi16(val01D, limit);
- _mm_storeu_si128((__m128i *) &variance[2][iOffset][jOffset], sum);
- _mm_storeu_si128((__m128i *) &variance[3][iOffset][jOffset], sum2);
+ limit = _mm256_sub_epi16(_mm256_setzero_si256(), limit);
- if (i == 8)
- {
- x8 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 4][jOffset]);
- y8 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 4][jOffset]);
- x6 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 3][jOffset]);
- __m128i y6 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 3][jOffset]);
- x4 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 2][jOffset]);
- __m128i y4 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 2][jOffset]);
- x2 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 1][jOffset]);
- __m128i y2 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 1][jOffset]);
+ val01A = _mm256_max_epi16(val01A, limit);
+ val01B = _mm256_max_epi16(val01B, limit);
+ val01C = _mm256_max_epi16(val01C, limit);
+ val01D = _mm256_max_epi16(val01D, limit);
- x8 = _mm_add_epi32(sum, x8);
- y8 = _mm_add_epi32(sum2, y8);
+ val01A = _mm256_add_epi16(val01A, val01C);
+ val01B = _mm256_add_epi16(val01B, val01D);
- x4 = _mm_add_epi32(x6, x4);
- y4 = _mm_add_epi32(y6, y4);
+ const __m256i coeff = _mm256_srai_epi16(params[i], 2);
- x2 = _mm_add_epi32(x8, x2);
- y2 = _mm_add_epi32(y8, y2);
+ accumA = _mm256_add_epi32(accumA, _mm256_madd_epi16(val01A, coeff));
+ accumB = _mm256_add_epi32(accumB, _mm256_madd_epi16(val01B, coeff));
+ };
- sum = _mm_add_epi32(x4, x2);
- sum2 = _mm_add_epi32(y4, y2);
- _mm_storeu_si128((__m128i *) &variance[0][iOffset - 4][jOffset], sum);
- _mm_storeu_si128((__m128i *) &variance[1][iOffset - 4][jOffset], sum2);
+ process2coeffs(0, pImg8 + 0, pImg7 + 0, pImg6 - 1, pImg5 + 1);
+ process2coeffs(1, pImg4 - 2, pImg3 + 2, pImg2 - 3, pImg1 + 3);
+ process2coeffs(2, pImg0 - 4, pImg0 + 4, pImg6 + 1, pImg5 - 1);
+ process2coeffs(3, pImg4 + 2, pImg3 - 2, pImg2 + 3, pImg1 - 3);
- sum2 = _mm_mullo_epi32(sum2, n);
- sum = _mm_mullo_epi32(sum, sum);
- sum2 = _mm_add_epi32(sum2, o);
- sum2 = _mm_sub_epi32(sum2, sum);
- sum2 = _mm_srli_epi32(sum2, 3);
- sum2 = _mm_mullo_epi32(sum2, m13);
- sum2 = _mm_srli_epi32(sum2, 14);
- _mm_storeu_si128((__m128i *) &variance[VARIANCE][iOffset - 4][jOffset], sum2);
- }
- else if (i > 8)
- {
- x8 = _mm_loadu_si128((__m128i *)&variance[2][iOffset - 5][jOffset]);
- y8 = _mm_loadu_si128((__m128i *)&variance[3][iOffset - 5][jOffset]);
- x6 = _mm_loadu_si128((__m128i *)&variance[0][iOffset - 5][jOffset]);
- __m128i y6 = _mm_loadu_si128((__m128i *)&variance[1][iOffset - 5][jOffset]);
+ process2coeffs(4, pImg6 + 0, pImg5 - 0, pImg4 - 1, pImg3 + 1);
+ process2coeffs(5, pImg2 - 2, pImg1 + 2, pImg0 - 3, pImg0 + 3);
+ process2coeffs(6, pImg4 + 1, pImg3 - 1, pImg2 + 2, pImg1 - 2);
- x6 = _mm_sub_epi32(x6, x8);
- y6 = _mm_sub_epi32(y6, y8);
+ process2coeffs(7, pImg4 + 0, pImg3 - 0, pImg2 - 1, pImg1 + 1);
+ process2coeffs(8, pImg0 - 2, pImg0 + 2, pImg2 + 1, pImg1 - 1);
+ process2coeffs(9, pImg2 + 0, pImg1 - 0, pImg0 - 1, pImg0 + 1);
+ process2coeffs(10, zeros, pImg0, pImg0, pImg0);
- sum = _mm_add_epi32(x6, sum);
- sum2 = _mm_add_epi32(y6, sum2);
- _mm_storeu_si128((__m128i *) &variance[0][iOffset - 4][jOffset], sum);
- _mm_storeu_si128((__m128i *) &variance[1][iOffset - 4][jOffset], sum2);
+ accumA = _mm256_srai_epi32(accumA, shift);
+ accumB = _mm256_srai_epi32(accumB, shift);
- sum2 = _mm_mullo_epi32(sum2, n);
- sum = _mm_mullo_epi32(sum, sum);
- sum2 = _mm_add_epi32(sum2, o);
- sum2 = _mm_sub_epi32(sum2, sum);
- sum2 = _mm_srli_epi32(sum2, 3);
- sum2 = _mm_mullo_epi32(sum2, m13);
- sum2 = _mm_srli_epi32(sum2, 14);
- _mm_storeu_si128((__m128i *) &variance[VARIANCE][iOffset - 4][jOffset], sum2);
- }
- }
+ accumA = _mm256_blend_epi16(accumA, _mm256_slli_si256(accumB, 2), 0xAA);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if (useBounCondition)
+ {
+ accumA = _mm256_min_epi16(mmPOffsetClipVector, accumA);
+ accumA = _mm256_max_epi16(mmNOffsetClipVector, accumA);
+ // accumA is Ori Offset
+ mmOriOffset = accumA;
+ // Calc Sign
+ // P = 1, N = 0
+ mmSignOffsetP = _mm256_abs_epi16(_mm256_cmpgt_epi16(mmOriOffset, mmZeroVector));
+ // P = 0, N = 1
+ mmSignOffsetN = _mm256_abs_epi16(_mm256_sub_epi16(mm01Vector, mmSignOffsetP));
+ // Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16(mmOriOffset);
+ // BS based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16(mmAbsOffset, _mm256_add_epi16(mm16Vector, mmBsFactor));
+ mmAdjOffset = _mm256_add_epi16(mmAdjOffset, mm08Vector);
+ mmAdjOffset = _mm256_srai_epi16(mmAdjOffset, 4);
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16(mmClassIdxBsP, mmAdjOffset);
+ __m256i mmTmpOrg = _mm256_mullo_epi16(mmClassIdxBsN, mmAbsOffset);
+
+ __m256i mmTmpFin = _mm256_add_epi16(mmTmpAdj, mmTmpOrg);
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16(mmSignOffsetP, mmTmpFin);
+ __m256i mmTmpSignN = _mm256_sub_epi16(mmZeroVector, _mm256_mullo_epi16(mmSignOffsetN, mmTmpFin));
+
+ accumA = _mm256_add_epi16(mmTmpSignP, mmTmpSignN);
+ }
- }
-#if USE_AVX2
- }
-#endif
-}
+ if (useResiCondition)
+ {
+ accumA = _mm256_min_epi16(mmPOffsetClipVector, accumA);
+ accumA = _mm256_max_epi16(mmNOffsetClipVector, accumA);
+ // accumA is Ori Offset
+ mmOriOffset = accumA;
+ // Calc Sign
+ // P = 1, N = 0
+ mmSignOffsetP = _mm256_abs_epi16(_mm256_cmpgt_epi16(mmOriOffset, mmZeroVector));
+ // P = 0, N = 1
+ mmSignOffsetN = _mm256_abs_epi16(_mm256_sub_epi16(mm01Vector, mmSignOffsetP));
+ // Calc Abs Offset
+ mmAbsOffset = _mm256_abs_epi16(mmOriOffset);
+ // Resi based Adjustment
+ mmAdjOffset = _mm256_mullo_epi16(mmAbsOffset, _mm256_add_epi16(mm16Vector, mmResiFactor));
+ mmAdjOffset = _mm256_add_epi16(mmAdjOffset, mm08Vector);
+ mmAdjOffset = _mm256_srai_epi16(mmAdjOffset, 4);
+
+ __m256i mmTmpAdj = _mm256_mullo_epi16(mmClassIdxResiP, mmAdjOffset);
+ __m256i mmTmpOrg = _mm256_mullo_epi16(mmClassIdxResiN, mmAbsOffset);
+
+ __m256i mmTmpFin = _mm256_add_epi16(mmTmpAdj, mmTmpOrg);
+
+ __m256i mmTmpSignP = _mm256_mullo_epi16(mmSignOffsetP, mmTmpFin);
+ __m256i mmTmpSignN = _mm256_sub_epi16(mmZeroVector, _mm256_mullo_epi16(mmSignOffsetN, mmTmpFin));
+
+ accumA = _mm256_add_epi16(mmTmpSignP, mmTmpSignN);
+ }
#endif
+ accumA = _mm256_add_epi16(accumA, cur);
+ accumA = _mm256_min_epi16(mmMax, _mm256_max_epi16(accumA, mmMin));
-#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
-#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
-template<X86_VEXT vext>
-static void simdFilterResi9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &srcResiLuma, const Area &curBlk, const Area &blkDst, Pel ***fixedFilterResiResults, int picWidth, const int fixedFiltInd, const short classIndFixed[NUM_CLASSES_FIX], int fixedFiltQpInd, int dirWindSize, const ClpRng &clpRng, const Pel clippingValues[4])
-{
- const int srcStride = srcResiLuma.stride;
- constexpr int shift = AdaptiveLoopFilter::m_NUM_BITS_FIXED_FILTER - 1;
- constexpr int round = 1 << (shift - 1);
-
- const int width = curBlk.width;
- const int height = curBlk.height;
-
- constexpr int stepX = 8;
- constexpr int stepY = 2;
+#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
+ _mm256_storeu_si256((__m256i *) (&(fixedFilterResiResults[fixedFiltInd][blkDst.y + i + ii][blkDst.x + j])), accumA);
+#else
+ _mm256_storeu_si256((__m256i *) (&(fixedFilterResiResults[fixedFiltInd][curBlk.y + i + ii][curBlk.x + j])), accumA);
+#endif
+ } // for (size_t ii = 0; ii < stepY; ii++)
+ } // for (size_t j = 0; j < width; j += stepX)
+ src += srcStride * stepY;
+ }
- const Pel *src = srcResiLuma.buf + curBlk.y * srcStride + curBlk.x;
+ }
+ else
+ {
const __m128i mmOffset = _mm_set1_epi32(round);
- const int clpRngmin = -clpRng.max;
- const int clpRngmax = clpRng.max;
const __m128i mmMin = _mm_set1_epi16(clpRngmin);
const __m128i mmMax = _mm_set1_epi16(clpRngmax);
const __m128i mmClippingValues = _mm_loadl_epi64((const __m128i *) clippingValues);
const __m128i mm11 = _mm_set1_epi8(1);
const __m128i mm3 = _mm_set1_epi16(3);
- const std::array<std::array<short, FIX_FILTER_NUM_COEFF_DB_COMBINE_9_DB_9 + 1>, NUM_FIXED_FILTERS>& filterCoeffFixed = packedDataFixedFilters9Db9Combine[fixedFiltQpInd];
- const Pel zeros[8] = { 0 };
+#endif //Use Avx2 Simd
+
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ __m128i mmClassIdxBsP, mmClassIdxResiP, mmClassIdxBsN, mmClassIdxResiN, mmClassIdxTmp;
+ __m128i mmOriOffset;
+ __m128i mmSignOffsetP, mmSignOffsetN;
+ __m128i mmAbsOffset;
+ __m128i mmAdjOffset;
+ __m128i mmZeroVector = _mm_set1_epi16( 0 );
+ __m128i mm01Vector = _mm_set1_epi16( 1 );
+ __m128i mm08Vector = _mm_set1_epi16( 8 );
+ __m128i mm16Vector = _mm_set1_epi16( 16 );
+ __m128i mmPOffsetClipVector = _mm_set1_epi16( +offsetClipValue );
+ __m128i mmNOffsetClipVector = _mm_set1_epi16( -offsetClipValue );
+ //Set Factor
+ __m128i mmBsFactor = isIntraSlice ? _mm_set1_epi16( 4 ) : _mm_set1_epi16( 3 );
+ __m128i mmResiFactor = isIntraSlice ? _mm_set1_epi16( 0 >> (!isSpsAdjust ? 1 : 0) ) : _mm_set1_epi16( 3 >> (!isSpsAdjust ? 1 : 0) );
+#endif
+
for (int i = 0; i < height; i += stepY)
{
#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
@@ -6497,6 +9137,13 @@ static void simdFilterResi9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &s
#else
const AlfClassifier *pClass = classifier[curBlk.y + i] + curBlk.x;
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ AlfClassifier *pClassCodingInfo = nullptr;
+ if( useBounCondition || useResiCondition )
+ {
+ pClassCodingInfo = classifierCodingInfo[blkDst.y + i] + blkDst.x;
+ }
+#endif
for (int j = 0; j < width; j += stepX)
{
@@ -6536,6 +9183,25 @@ static void simdFilterResi9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &s
params[8] = _mm_unpackhi_epi64(_mm_unpacklo_epi32(rawCoef[0][2], rawCoef[1][2]), _mm_unpacklo_epi32(rawCoef[2][2], rawCoef[3][2]));
params[9] = _mm_unpacklo_epi64(_mm_unpackhi_epi32(rawCoef[0][2], rawCoef[1][2]), _mm_unpackhi_epi32(rawCoef[2][2], rawCoef[3][2]));
params[10] = _mm_unpackhi_epi64(_mm_unpackhi_epi32(rawCoef[0][2], rawCoef[1][2]), _mm_unpackhi_epi32(rawCoef[2][2], rawCoef[3][2]));
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ mmClassIdxBsP = _mm_set1_epi16( 0 );
+ mmClassIdxBsN = _mm_set1_epi16( 0 );
+ if( useBounCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxBsN = _mm_sub_epi16( mm01Vector, mmClassIdxBsP );
+ }
+ mmClassIdxResiP = _mm_set1_epi16( 0 );
+ mmClassIdxResiN = _mm_set1_epi16( 0 );
+ if( useResiCondition )
+ {
+ mmClassIdxTmp = _mm_loadu_si128( (const __m128i *) (pClassCodingInfo + j));
+ mmClassIdxBsP = _mm_srai_epi16( mmClassIdxTmp, 1 );
+ mmClassIdxResiP = _mm_sub_epi16( mmClassIdxTmp, _mm_add_epi16( mmClassIdxBsP, mmClassIdxBsP) );
+ mmClassIdxResiN = _mm_sub_epi16( mm01Vector, mmClassIdxResiP );
+ }
+#endif
for (int ii = 0; ii < stepY; ii++)
{
const Pel *pImg0, *pImg1, *pImg2, *pImg3, *pImg4, *pImg5, *pImg6, *pImg7, *pImg8;
@@ -6612,6 +9278,65 @@ static void simdFilterResi9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &s
accumB = _mm_srai_epi32(accumB, shift);
accumA = _mm_blend_epi16(accumA, _mm_slli_si128(accumB, 2), 0xAA);
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( useBounCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //BS based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmBsFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector );
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxBsP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxBsN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+
+ if( useResiCondition )
+ {
+ accumA = _mm_min_epi16( mmPOffsetClipVector, accumA );
+ accumA = _mm_max_epi16( mmNOffsetClipVector, accumA );
+ //accumA is Ori Offset
+ mmOriOffset = accumA;
+ //Calc Sign
+ //P = 1, N = 0
+ mmSignOffsetP = _mm_abs_epi16( _mm_cmpgt_epi16( mmOriOffset , mmZeroVector ));
+ //P = 0, N = 1
+ mmSignOffsetN = _mm_abs_epi16( _mm_sub_epi16( mm01Vector, mmSignOffsetP ));
+ //Calc Abs Offset
+ mmAbsOffset = _mm_abs_epi16( mmOriOffset );
+ //Resi based Adjustment
+ mmAdjOffset = _mm_mullo_epi16( mmAbsOffset, _mm_add_epi16( mm16Vector, mmResiFactor ) );
+ mmAdjOffset = _mm_add_epi16( mmAdjOffset, mm08Vector);
+ mmAdjOffset = _mm_srai_epi16( mmAdjOffset, 4 );
+
+ __m128i mmTmpAdj = _mm_mullo_epi16( mmClassIdxResiP, mmAdjOffset );
+ __m128i mmTmpOrg = _mm_mullo_epi16( mmClassIdxResiN, mmAbsOffset );
+
+ __m128i mmTmpFin = _mm_add_epi16( mmTmpAdj, mmTmpOrg );
+
+ __m128i mmTmpSignP = _mm_mullo_epi16( mmSignOffsetP, mmTmpFin );
+ __m128i mmTmpSignN = _mm_sub_epi16( mmZeroVector, _mm_mullo_epi16( mmSignOffsetN, mmTmpFin ));
+
+ accumA = _mm_add_epi16( mmTmpSignP, mmTmpSignN );
+ }
+#endif
accumA = _mm_add_epi16(accumA, cur);
accumA = _mm_min_epi16(mmMax, _mm_max_epi16(accumA, mmMin));
@@ -6624,6 +9349,9 @@ static void simdFilterResi9x9Db9Blk(AlfClassifier **classifier, const CPelBuf &s
} // for (size_t j = 0; j < width; j += stepX)
src += srcStride * stepY;
}
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ }//Use 256 Bit Simd
+#endif
}
#else
template<X86_VEXT vext>
@@ -7103,6 +9831,243 @@ static void simdCalcClass0(AlfClassifier **classifier, const Area &blkDst, const
#endif
const __m128i shift = _mm_cvtsi32_si128(9 + bitDepth - 4);
const int multTab[] = { 5628, 1407, 624, 351, 225, 156 };
+
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool use256BitSimd = vext >= AVX2 && blkDst.width % 16 == 0 ? true : false;
+
+ if( use256BitSimd)
+ {
+#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
+ const __m256i mult = _mm256_set1_epi32(multTab[dirWindSize % 10]);
+#else
+ const __m256i mult = _mm256_set1_epi32(multTab[dirWindSize]);
+#endif
+ const __m256i dirOff = _mm256_set1_epi32(noDir * (noDir + 1));
+ const __m256i ones = _mm256_set1_epi32(1);
+ const __m256i zeros = _mm256_setzero_si256();
+ const __m256i scale = _mm256_set1_epi32(192);
+
+ int lapOffset = (dirWindSize == 1) ? 2 : 0;
+ for (int i = 0; i < curBlk.height; i += 2)
+ {
+ int iOffset = (i >> 1) + lapOffset;
+ for (int j = 0; j < curBlk.width; j += 16)
+ {
+ int jOffset = (j >> 1) + lapOffset;
+#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
+ int iOffsetV = i >> 1;
+ int jOffsetV = j >> 1;
+#endif
+ __m256i sumV = _mm256_loadu_si256((const __m256i *) &laplacian[VER][iOffset][jOffset]); // 8 32-bit values
+ __m256i sumH = _mm256_loadu_si256((const __m256i *) &laplacian[HOR][iOffset][jOffset]);
+ __m256i sumD0 = _mm256_loadu_si256((const __m256i *) &laplacian[DIAG0][iOffset][jOffset]);
+ __m256i sumD1 = _mm256_loadu_si256((const __m256i *) &laplacian[DIAG1][iOffset][jOffset]);
+
+ // sum += sumV + sumH;
+ __m256i tempAct = _mm256_add_epi32(sumV, sumH);
+ __m256i activity = _mm256_mullo_epi32(tempAct, mult);
+ activity = _mm256_srl_epi32(activity, shift);
+ activity = _mm256_min_epi32(activity, scale);
+
+ __m256i xmm2 = activity;
+ __m256i xmm0 = _mm256_setzero_si256();
+ __m256i xmm15 = _mm256_cmpeq_epi32(xmm0, xmm0);
+ __m256i xmm1 = _mm256_srli_epi32(xmm15, 31);
+ __m256i xmm7 = _mm256_srli_epi32(xmm15, 29);
+#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
+ __m256i xmm8 = _mm256_srli_epi32(xmm15, 28);
+#endif
+ __m256i xmm9 = _mm256_add_epi32(_mm256_slli_epi32(xmm7, 2), xmm1);
+
+ __m256i LUT192 = _mm256_set_epi32(0x0C020A00, 0x0E040608, 0x0E040608, 0x0C020A00, 0x0C020A00, 0x0E040608, 0x0E040608, 0x0C020A00);
+
+ xmm2 = _mm256_or_si256(xmm2, _mm256_srli_epi32(xmm2, 1));
+ xmm2 = _mm256_or_si256(xmm2, _mm256_srli_epi32(xmm2, 2));
+ xmm2 = _mm256_or_si256(xmm2, _mm256_srli_epi32(xmm2, 4));
+ xmm2 = _mm256_mullo_epi16(xmm2, xmm9);
+ xmm2 = _mm256_and_si256(_mm256_srli_epi32(xmm2, 5), xmm7);
+ xmm2 = _mm256_shuffle_epi8(LUT192, xmm2);
+
+ __m256i xmm4 = _mm256_xor_si256(activity, _mm256_srli_epi32(activity, 1));
+// xmm4 = _mm256_cmplt_epi32(xmm4, activity);
+ xmm4 = _mm256_cmpgt_epi32(xmm4, _mm256_sub_epi32( activity, _mm256_set1_epi32(1) ) );
+ xmm4 = _mm256_add_epi32( _mm256_abs_epi32(xmm4), _mm256_set1_epi32( -1 ) );
+
+ xmm4 = _mm256_or_si256(_mm256_cmpeq_epi32(activity, xmm1), xmm4);
+ xmm4 = _mm256_and_si256(xmm4, xmm1);
+
+ activity = _mm256_or_si256(xmm2, xmm4);
+
+ __m256i hv1 = _mm256_max_epi32(sumV, sumH);
+ __m256i hv0 = _mm256_min_epi32(sumV, sumH);
+
+ __m256i d1 = _mm256_max_epi32(sumD0, sumD1);
+ __m256i d0 = _mm256_min_epi32(sumD0, sumD1);
+
+ // edgeStrengthHV, to optimize
+ __m256i hv0Two = _mm256_slli_epi32(hv0, 1);
+ __m256i hv0Eight = _mm256_slli_epi32(hv0, 3);
+ __m256i hv1Two = _mm256_slli_epi32(hv1, 1);
+ __m256i strength = _mm256_cmpgt_epi32(_mm256_slli_epi32(hv1, 2), _mm256_add_epi32(hv0, _mm256_slli_epi32(hv0, 2))); // 4, 5
+ __m256i edgeStrengthHV = _mm256_and_si256(strength, ones);
+
+ strength = _mm256_cmpgt_epi32(hv1Two, _mm256_add_epi32(hv0, hv0Two)); // 2, 3
+ edgeStrengthHV = _mm256_add_epi32(edgeStrengthHV, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(hv1, hv0Two); // 1, 2
+ edgeStrengthHV = _mm256_add_epi32(edgeStrengthHV, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(hv1, _mm256_add_epi32(hv0, hv0Two)); // 1, 3
+ edgeStrengthHV = _mm256_add_epi32(edgeStrengthHV, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(hv1Two, _mm256_add_epi32(hv0, hv0Eight)); // 2, 9
+ edgeStrengthHV = _mm256_add_epi32(edgeStrengthHV, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(hv1, hv0Eight); // 1, 8
+ edgeStrengthHV = _mm256_add_epi32(edgeStrengthHV, _mm256_and_si256(strength, ones));
+
+ // edgeStrengthD, to optimize
+ __m256i d0Two = _mm256_slli_epi32(d0, 1);
+ __m256i d0Eight = _mm256_slli_epi32(d0, 3);
+ __m256i d1Two = _mm256_slli_epi32(d1, 1);
+ strength = _mm256_cmpgt_epi32(_mm256_slli_epi32(d1, 2), _mm256_add_epi32(d0, _mm256_slli_epi32(d0, 2))); // 4, 5
+ __m256i edgeStrengthD = _mm256_and_si256(strength, ones);
+
+ strength = _mm256_cmpgt_epi32(d1Two, _mm256_add_epi32(d0, d0Two)); // 2, 3
+ edgeStrengthD = _mm256_add_epi32(edgeStrengthD, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(d1, d0Two); // 1, 2
+ edgeStrengthD = _mm256_add_epi32(edgeStrengthD, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(d1, _mm256_add_epi32(d0, d0Two)); // 1, 3
+ edgeStrengthD = _mm256_add_epi32(edgeStrengthD, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(d1Two, _mm256_add_epi32(d0, d0Eight)); // 2, 9
+ edgeStrengthD = _mm256_add_epi32(edgeStrengthD, _mm256_and_si256(strength, ones));
+
+ strength = _mm256_cmpgt_epi32(d1, d0Eight); // 1, 8
+ edgeStrengthD = _mm256_add_epi32(edgeStrengthD, _mm256_and_si256(strength, ones));
+
+ const __m256i hv1Xd0e = _mm256_mul_epi32(hv1, d0);
+ const __m256i hv0Xd1e = _mm256_mul_epi32(hv0, d1);
+ const __m256i hv1Xd0o = _mm256_mul_epi32(_mm256_srli_si256(hv1, 4), _mm256_srli_si256(d0, 4));
+ const __m256i hv0Xd1o = _mm256_mul_epi32(_mm256_srli_si256(hv0, 4), _mm256_srli_si256(d1, 4));
+
+ const __m256i xmme = _mm256_sub_epi64(hv0Xd1e, hv1Xd0e);
+ const __m256i xmmo = _mm256_sub_epi64(hv0Xd1o, hv1Xd0o);
+
+ __m256i dirCondition = _mm256_srai_epi32(_mm256_blend_epi16(_mm256_srli_si256(xmme, 4), xmmo, 0xCC), 31);
+
+ __m256i cx = _mm256_blendv_epi8(edgeStrengthHV, edgeStrengthD, dirCondition); // x
+ __m256i cy = _mm256_blendv_epi8(edgeStrengthD, edgeStrengthHV, dirCondition); // y
+ __m256i dirOffset = _mm256_blendv_epi8(_mm256_set1_epi32(28), zeros, dirCondition);
+ // direction = (y*(y+1))/2 + x
+ __m256i direction = _mm256_mullo_epi32(cy, cy);
+ direction = _mm256_add_epi32(direction, cy);
+ direction = _mm256_srli_epi32(direction, 1);
+ direction = _mm256_add_epi32(direction, cx);
+ direction = _mm256_andnot_si256(_mm256_cmpgt_epi32(cx, cy), direction);
+ direction = _mm256_add_epi32(direction, dirOffset);
+#if JVET_AE0139_ALF_IMPROVED_FIXFILTER
+ __m256i sum2 = _mm256_loadu_si256((const __m256i *) &laplacian[VARIANCE][iOffsetV][jOffsetV]);
+ __m256i shiftLut = _mm256_set_m128i( _mm_loadu_si128( (const __m128i *) divShift2), _mm_loadu_si128( (const __m128i *) divShift2) );
+ __m256i shiftVal = _mm256_shuffle_epi8(shiftLut, activity);
+ shiftVal = _mm256_add_epi32(shiftVal, xmm1);
+ shiftVal = _mm256_add_epi32(shiftVal, xmm1);
+ if (vext >= AVX2)
+ {
+ sum2 = _mm256_srlv_epi32(sum2, shiftVal);
+ }
+ else
+ {
+ __m128i sum2Tmp0 = _mm256_extracti128_si256(sum2, 0);
+ __m128i sum2Tmp1 = _mm256_extracti128_si256(sum2, 1);
+ __m128i shiftValTmp0 = _mm256_extracti128_si256(shiftVal, 0);
+ __m128i shiftValTmp1 = _mm256_extracti128_si256(shiftVal, 1);
+
+ uint64_t tmpVal0[4];
+ int32_t *pVal0 = (int32_t *) tmpVal0;
+ _mm_storeu_si128((__m128i *) tmpVal0, sum2Tmp0);
+ _mm_storeu_si128((__m128i *) (tmpVal0 + 2), shiftValTmp0);
+ pVal0[0] >>= pVal0[4];
+ pVal0[1] >>= pVal0[5];
+ pVal0[2] >>= pVal0[6];
+ pVal0[3] >>= pVal0[7];
+
+ uint64_t tmpVal1[4];
+ int32_t *pVal1 = (int32_t *) tmpVal1;
+ _mm_storeu_si128((__m128i *) tmpVal1, sum2Tmp1);
+ _mm_storeu_si128((__m128i *) (tmpVal1 + 2), shiftValTmp1);
+ pVal1[0] >>= pVal1[4];
+ pVal1[1] >>= pVal1[5];
+ pVal1[2] >>= pVal1[6];
+ pVal1[3] >>= pVal1[7];
+
+ sum2 = _mm256_set_m128i( _mm_loadu_si128((const __m128i *) pVal1 ), _mm_loadu_si128((const __m128i *) pVal0) );
+ }
+
+ __m256i LUT0 = _mm256_set_m128i( _mm_loadu_si128((const __m128i *) sqrtSum), _mm_loadu_si128((const __m128i *) sqrtSum) );
+ __m256i LUT1 = _mm256_set_m128i(_mm_loadu_si128((const __m128i *) &sqrtSum[16]), _mm_loadu_si128((const __m128i *) &sqrtSum[16]) );
+ __m256i xmm16 = _mm256_set_epi32(16, 16, 16, 16, 16, 16, 16, 16);
+ __m256i xmm35 = _mm256_set_epi32(35, 35, 35, 35, 35, 35, 35, 35);
+ __m256i xmm48 = _mm256_set_epi32(48, 48, 48, 48, 48, 48, 48, 48);
+
+ __m256i use1 = _mm256_cmpgt_epi32(sum2, xmm8);
+
+ __m256i idx0 = _mm256_and_si256(sum2, xmm8);
+ __m256i idx1 = _mm256_sub_epi32(sum2, xmm16);
+ idx1 = _mm256_min_epi32(idx1, xmm8);
+
+ idx0 = _mm256_shuffle_epi8(LUT0, idx0);
+ idx1 = _mm256_shuffle_epi8(LUT1, idx1);
+
+ idx1 = _mm256_add_epi32(idx1, _mm256_slli_epi32(xmm1, 2));
+
+ idx0 = _mm256_andnot_si256(use1, idx0);
+ idx1 = _mm256_and_si256(use1, idx1);
+ idx0 = _mm256_add_epi32(idx0, idx1);
+
+ xmm35 = _mm256_cmpgt_epi32(sum2, xmm35);
+ xmm48 = _mm256_cmpgt_epi32(sum2, xmm48);
+
+ xmm35 = _mm256_and_si256(xmm35, xmm1);
+ xmm48 = _mm256_and_si256(xmm48, xmm1);
+
+ xmm35 = _mm256_add_epi32(xmm35, xmm48);
+
+ xmm2 = _mm256_add_epi32(idx0, xmm35);
+ xmm2 = _mm256_slli_epi32(xmm2, 4);
+ activity = _mm256_add_epi32(activity, xmm2);
+#endif
+ __m256i classIdx = _mm256_mullo_epi32(dirOff, activity);
+ classIdx = _mm256_add_epi32(classIdx, direction);
+
+ // transpose
+ __m256i dirTempHVMinus1 = _mm256_cmpgt_epi32(sumV, sumH);
+ __m256i dirTempDMinus1 = _mm256_cmpgt_epi32(sumD0, sumD1);
+ __m256i transposeIdx = _mm256_set1_epi32(3);
+ transposeIdx = _mm256_add_epi32(transposeIdx, dirTempHVMinus1);
+ transposeIdx = _mm256_add_epi32(transposeIdx, dirTempDMinus1);
+ transposeIdx = _mm256_add_epi32(transposeIdx, dirTempDMinus1);
+
+ classIdx = _mm256_slli_epi16(classIdx, 2);
+ classIdx = _mm256_add_epi16(classIdx, transposeIdx);
+ classIdx = _mm256_shuffle_epi8(classIdx, _mm256_setr_epi8(0, 1, 0, 1, 4, 5, 4, 5, 8, 9, 8, 9, 12, 13, 12, 13, 0, 1, 0, 1, 4, 5, 4, 5, 8, 9, 8, 9, 12, 13, 12, 13));
+#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
+ _mm256_storeu_si256((__m256i *) &classifier[blkDst.pos().y + i][blkDst.pos().x + j], classIdx);
+ _mm256_storeu_si256((__m256i *) &classifier[blkDst.pos().y + i + 1][blkDst.pos().x + j], classIdx);
+#else
+ _mm256_storeu_si256((__m256i *) &classifier[curBlk.pos().y + i][curBlk.pos().x + j], classIdx);
+ _mm256_storeu_si256((__m256i *) &classifier[curBlk.pos().y + i + 1][curBlk.pos().x + j], classIdx);
+#endif
+
+ } // for (int j = 0; j < curBlk.width; j += 16)
+ } // for (int i = 0; i < curBlk.height; i += 2)
+
+ }
+ else
+ {
+#endif
#if JVET_AG0157_ALF_CHROMA_FIXED_FILTER
const __m128i mult = _mm_set1_epi32(multTab[dirWindSize % 10]);
#else
@@ -7312,8 +10277,14 @@ static void simdCalcClass0(AlfClassifier **classifier, const Area &blkDst, const
}//for (int j = 0; j < curBlk.width; j += 8)
}//for (int i = 0; i < curBlk.height; i += 2)
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ }// Use 256 Bit Simd
+ #endif
}
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+template <X86_VEXT vext>
+#endif
static void simdCalcClass1(AlfClassifier **classifier, const Area &blkDst, const Area &curBlk, int dirWindSize, int classDir, int noDir, int noAct, int bitDepth, int subBlkSize, int mappingDir[NUM_DIR_FIX][NUM_DIR_FIX], uint32_t **laplacian[NUM_DIRECTIONS])
{
const __m128i shift = _mm_cvtsi32_si128(9 + bitDepth);
@@ -7322,6 +10293,82 @@ static void simdCalcClass1(AlfClassifier **classifier, const Area &blkDst, const
#else
const int multTab[] = { 5628, 1407, 624, 351, 225, 156 };
#endif
+
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ const bool use256BitSimd = vext >= AVX2 && curBlk.width % 16 == 0 ? true : false;
+ if( use256BitSimd )
+ {
+ const __m256i mult = _mm256_set1_epi32(multTab[dirWindSize]);
+ const __m256i scale = _mm256_set1_epi32(15);
+
+ for (int i = 0; i < curBlk.height; i += 2)
+ {
+ int iOffset = i >> 1;
+ for (int j = 0; j < curBlk.width; j += 16)
+ {
+ int jOffset = j >> 1;
+ __m256i sumV = _mm256_loadu_si256((const __m256i *) &laplacian[VER][iOffset][jOffset]); // 8 32-bit values
+ __m256i sumH = _mm256_loadu_si256((const __m256i *) &laplacian[HOR][iOffset][jOffset]);
+ __m256i sumD0 = _mm256_loadu_si256((const __m256i *) &laplacian[DIAG0][iOffset][jOffset]);
+ __m256i sumD1 = _mm256_loadu_si256((const __m256i *) &laplacian[DIAG1][iOffset][jOffset]);
+
+ // sum += sumV + sumH;
+ __m256i tempAct = _mm256_add_epi32(sumV, sumH);
+ __m256i activity = _mm256_mullo_epi32(tempAct, mult);
+ activity = _mm256_srl_epi32(activity, shift);
+ activity = _mm256_min_epi32(activity, scale);
+ __m256i classIdx = _mm256_shuffle_epi8(_mm256_setr_epi8(0, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4, 0, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4), activity);
+ classIdx = _mm256_add_epi32(classIdx, _mm256_slli_epi32(classIdx, 2)); // activity * 5
+
+ __m256i hv1 = _mm256_max_epi32(sumV, sumH);
+ __m256i hv0 = _mm256_min_epi32(sumV, sumH);
+
+ __m256i d1 = _mm256_max_epi32(sumD0, sumD1);
+ __m256i d0 = _mm256_min_epi32(sumD0, sumD1);
+
+ const __m256i hv1Xd0e = _mm256_mul_epi32(hv1, d0);
+ const __m256i hv0Xd1e = _mm256_mul_epi32(hv0, d1);
+ const __m256i hv1Xd0o = _mm256_mul_epi32(_mm256_srli_si256(hv1, 4), _mm256_srli_si256(d0, 4));
+ const __m256i hv0Xd1o = _mm256_mul_epi32(_mm256_srli_si256(hv0, 4), _mm256_srli_si256(d1, 4));
+
+ const __m256i xmme = _mm256_sub_epi64(hv1Xd0e, hv0Xd1e);
+ const __m256i xmmo = _mm256_sub_epi64(hv1Xd0o, hv0Xd1o);
+
+ __m256i dirCondition = _mm256_srai_epi32(_mm256_blend_epi16(_mm256_srli_si256(xmme, 4), xmmo, 0xCC), 31);
+
+ __m256i hvd1 = _mm256_blendv_epi8(hv1, d1, dirCondition);
+ __m256i hvd0 = _mm256_blendv_epi8(hv0, d0, dirCondition);
+ __m256i strength1 = _mm256_cmpgt_epi32(hvd1, _mm256_add_epi32(hvd0, hvd0));
+ __m256i strength2 = _mm256_cmpgt_epi32(_mm256_add_epi32(hvd1, hvd1), _mm256_add_epi32(hvd0, _mm256_slli_epi32(hvd0, 3)));
+ __m256i offset = _mm256_and_si256(strength1, _mm256_set1_epi32(1));
+ __m256i direction = _mm256_add_epi32(offset, _mm256_and_si256(strength2, _mm256_set1_epi32(1)));
+ direction = _mm256_add_epi32(direction, _mm256_andnot_si256(dirCondition, _mm256_set1_epi32(2)));
+ direction = _mm256_and_si256(direction, strength1);
+ classIdx = _mm256_add_epi32(direction, classIdx);
+
+ // transpose
+ __m256i dirTempHVMinus1 = _mm256_cmpgt_epi32(sumV, sumH);
+ __m256i dirTempDMinus1 = _mm256_cmpgt_epi32(sumD0, sumD1);
+ __m256i transposeIdx = _mm256_set1_epi32(3);
+ transposeIdx = _mm256_add_epi32(transposeIdx, dirTempHVMinus1);
+ transposeIdx = _mm256_add_epi32(transposeIdx, dirTempDMinus1);
+ transposeIdx = _mm256_add_epi32(transposeIdx, dirTempDMinus1);
+ classIdx = _mm256_slli_epi16(classIdx, 2);
+ classIdx = _mm256_add_epi16(classIdx, transposeIdx);
+ classIdx = _mm256_shuffle_epi8(classIdx, _mm256_setr_epi8(0, 1, 0, 1, 4, 5, 4, 5, 8, 9, 8, 9, 12, 13, 12, 13, 0, 1, 0, 1, 4, 5, 4, 5, 8, 9, 8, 9, 12, 13, 12, 13));
+#if JVET_Z0105_LOOP_FILTER_VIRTUAL_BOUNDARY
+ _mm256_storeu_si256((__m256i *) &classifier[blkDst.pos().y + i][blkDst.pos().x + j], classIdx);
+ _mm256_storeu_si256((__m256i *) &classifier[blkDst.pos().y + i + 1][blkDst.pos().x + j], classIdx);
+#else
+ _mm256_storeu_si256((__m256i *) &classifier[curBlk.pos().y + i][curBlk.pos().x + j], classIdx);
+ _mm256_storeu_si256((__m256i *) &classifier[curBlk.pos().y + i + 1][curBlk.pos().x + j], classIdx);
+#endif
+ } // for (int j = 0; j < curBlk.width; j += 16)
+ } // for (int i = 0; i < curBlk.height; i += 2)
+ }
+ else
+ {
+#endif
const __m128i mult = _mm_set1_epi32(multTab[dirWindSize]);
const __m128i scale = _mm_set1_epi32(15);
@@ -7389,6 +10436,9 @@ static void simdCalcClass1(AlfClassifier **classifier, const Area &blkDst, const
#endif
}//for (int j = 0; j < curBlk.width; j += 8)
}//for (int i = 0; i < curBlk.height; i += 2)
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ }//Use 256 Bit Simd
+#endif
}
#endif
@@ -7438,7 +10488,11 @@ void AdaptiveLoopFilter::_initAdaptiveLoopFilterX86()
#else
m_calcClass0 = simdCalcClass0;
#endif
+#if USE_AVX2 && JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_calcClass1 = simdCalcClass1<vext>;
+#else
m_calcClass1 = simdCalcClass1;
+#endif
for( int i = 0; i < NUM_SETS_FIXED_FILTERS; i++ )
{
diff --git a/source/Lib/DecoderLib/DecLib.cpp b/source/Lib/DecoderLib/DecLib.cpp
index 82223c1df013107d5a3b10cd4e7d70537795293e..955a7d0b25d7df026bbceda2439ca720cd3ccdfe 100644
--- a/source/Lib/DecoderLib/DecLib.cpp
+++ b/source/Lib/DecoderLib/DecLib.cpp
@@ -815,6 +815,9 @@ void DecLib::executeLoopFilters()
m_cALF.copyResiData(cs);
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_cALF.callCodingInfoBuf( cs ).fill( 0 );
+#endif
// deblocking filter
#if DB_PARAM_TID
@@ -842,8 +845,12 @@ void DecLib::executeLoopFilters()
}
#endif
-
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf codingInfoBuf = m_cALF.callCodingInfoBuf( cs );
+ m_cLoopFilter.loopFilterPic( cs, codingInfoBuf, true );
+#else
m_cLoopFilter.loopFilterPic( cs );
+#endif
#if !MULTI_PASS_DMVR
CS::setRefinedMotionField(cs);
diff --git a/source/Lib/DecoderLib/VLCReader.cpp b/source/Lib/DecoderLib/VLCReader.cpp
index f974db44538cd4898d9ee7e5f2a9a49e3424a7e4..5e9c09f5826f1e7e90d226605facac4490df15aa 100644
--- a/source/Lib/DecoderLib/VLCReader.cpp
+++ b/source/Lib/DecoderLib/VLCReader.cpp
@@ -2182,6 +2182,16 @@ void HLSyntaxReader::parseSPS(SPS* pcSPS)
{
pcSPS->setAlfPrecisionFlag( false );
}
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( pcSPS->getALFEnabledFlag() )
+ {
+ READ_FLAG( uiCode, "sps_alf_luma_fixed_filter_adjust" ); pcSPS->setAlfLumaFixedFilterAdjust( uiCode ? true : false );
+ }
+ else
+ {
+ pcSPS->setAlfLumaFixedFilterAdjust( false );
+ }
#endif
if (pcSPS->getALFEnabledFlag() && pcSPS->getChromaFormatIdc() != CHROMA_400)
{
diff --git a/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp b/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp
index 944ae6daf30388244b9151cea55754bd3cabd6ec..38e77f597248ba1a5d4820adb075179d5200580d 100644
--- a/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp
+++ b/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp
@@ -1421,6 +1421,9 @@ void EncAdaptiveLoopFilter::ALFProcess( CodingStructure& cs, const double *lambd
#endif
const CPelBuf &resiLuma = resiYuv.get(COMPONENT_Y);
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf recYuvCodingInfo = m_tempBufCodingInfo.getBuf( cs.area );
+#endif
#if JVET_AI0166_CCALF_CHROMA_SAO_INPUT
m_tempBufSAO.copyFrom(cs.getRecoBuf());
PelUnitBuf recYuvSAO = m_tempBufSAO.getBuf(cs.area);
@@ -1483,7 +1486,23 @@ void EncAdaptiveLoopFilter::ALFProcess( CodingStructure& cs, const double *lambd
buf.extendBorderPel( MAX_ALF_PADDING_SIZE );
#endif
buf = buf.subBuf( UnitArea( cs.area.chromaFormat, Area( clipL ? 0 : MAX_ALF_PADDING_SIZE, clipT ? 0 : MAX_ALF_PADDING_SIZE, w, h ) ) );
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf bufCodingInfo = m_tempBufCodingInfo2.subBuf( UnitArea( CHROMA_400, Area( 0, 0, wBuf, hBuf ) ) );
+ bufCodingInfo.copyFrom( recYuvCodingInfo.subBuf( UnitArea( CHROMA_400, Area( xStart - ( clipL ? 0 : MAX_ALF_PADDING_SIZE ), yStart - ( clipT ? 0 : MAX_ALF_PADDING_SIZE ), wBuf, hBuf ) ) ) );
+ // pad top-left unavailable samples for raster slice
+ if( xStart == xPos && yStart == yPos && ( rasterSliceAlfPad & 1 ) )
+ {
+ bufCodingInfo.padBorderPel( MAX_ALF_PADDING_SIZE, 1 );
+ }
+ // pad bottom-right unavailable samples for raster slice
+ if( xEnd == xPos + width && yEnd == yPos + height && ( rasterSliceAlfPad & 2 ) )
+ {
+ bufCodingInfo.padBorderPel( MAX_ALF_PADDING_SIZE, 2 );
+ }
+ mirroredPaddingForAlf(cs, bufCodingInfo, MAX_ALF_PADDING_SIZE, true, false);
+ bufCodingInfo = bufCodingInfo.subBuf( UnitArea( CHROMA_400, Area( clipL ? 0 : MAX_ALF_PADDING_SIZE, clipT ? 0 : MAX_ALF_PADDING_SIZE, w, h ) ) );
+#endif
#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
PelUnitBuf bufResi = m_tempBufResi2.subBuf(UnitArea(cs.area.chromaFormat, Area(0, 0, wBuf, hBuf)));
bufResi.copyFrom(resiYuv.subBuf(
@@ -1543,6 +1562,9 @@ void EncAdaptiveLoopFilter::ALFProcess( CodingStructure& cs, const double *lambd
const Area blkDstChroma( xStart >> scaleX, yStart >> scaleY, w >> scaleX, h >> scaleY );
deriveFixedFilterChroma( m_classifier, buf, bufDb, blkDstChroma, blkSrcChroma, cs, -1, MAX_NUM_COMPONENT );
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ calcAlfLumaCodingInfoBlk(cs, m_classifierCodingInfo[0], blkDst, blkSrc, buf.get(COMPONENT_Y), 2, 2, m_inputBitDepth[CHANNEL_TYPE_LUMA], bufResi.get(COMPONENT_Y), m_laplacian[0], bufCodingInfo.get(COMPONENT_Y) );
+#endif
#if JVET_X0071_ALF_BAND_CLASSIFIER
deriveClassification( m_classifier, buf.get( COMPONENT_Y ),
#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
@@ -1577,6 +1599,9 @@ void EncAdaptiveLoopFilter::ALFProcess( CodingStructure& cs, const double *lambd
Area blkChroma(xPos >> scaleX, yPos >> scaleY, width >> scaleX, height >> scaleY);
deriveFixedFilterChroma( m_classifier, recYuv, recYuvBeforeDb, blkChroma, blkChroma, cs, -1, MAX_NUM_COMPONENT );
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ calcAlfLumaCodingInfoBlk(cs, m_classifierCodingInfo[0], blk, blk, recYuv.get(COMPONENT_Y), 2, 2, m_inputBitDepth[CHANNEL_TYPE_LUMA], resiYuv.get(COMPONENT_Y), m_laplacian[0], recYuvCodingInfo.get(COMPONENT_Y) );
+#endif
#if JVET_X0071_ALF_BAND_CLASSIFIER
deriveClassification( m_classifier, recLuma,
#if JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
diff --git a/source/Lib/EncoderLib/EncCfg.h b/source/Lib/EncoderLib/EncCfg.h
index 616369020a22ab7bf9eddaa436d81378558df2ae..3d709252d5ae1ee795e3e449c0a1688c51c48457 100644
--- a/source/Lib/EncoderLib/EncCfg.h
+++ b/source/Lib/EncoderLib/EncCfg.h
@@ -712,6 +712,9 @@ protected:
#endif
#if JVET_AH0057_CCALF_COEFF_PRECISION
bool m_ccalfPrecision;
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ bool m_alfLumaFixedFilterAdjust;
#endif
bool m_bTestSAODisableAtPictureLevel;
double m_saoEncodingRate; // When non-0 SAO early picture termination is enabled for luma and chroma
@@ -2389,6 +2392,10 @@ public:
#if JVET_AH0057_CCALF_COEFF_PRECISION
void setUseCCALFPrecision(bool b) { m_ccalfPrecision = b; }
bool getUseCCALFPrecision() const { return m_ccalfPrecision; }
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ void setAlfLumaFixedFilterAdjust(bool b) { m_alfLumaFixedFilterAdjust = b; }
+ bool getAlfLumaFixedFilterAdjust() const { return m_alfLumaFixedFilterAdjust; }
#endif
void setTestSAODisableAtPictureLevel (bool bVal) { m_bTestSAODisableAtPictureLevel = bVal; }
bool getTestSAODisableAtPictureLevel ( ) const { return m_bTestSAODisableAtPictureLevel; }
diff --git a/source/Lib/EncoderLib/EncCu.cpp b/source/Lib/EncoderLib/EncCu.cpp
index 4b27948fbd737786ae9a2356c6f0043142b9d13e..4f9d962143d3015177a148924f8d24406a4fe4ee 100644
--- a/source/Lib/EncoderLib/EncCu.cpp
+++ b/source/Lib/EncoderLib/EncCu.cpp
@@ -23992,13 +23992,23 @@ void EncCu::xCalDebCost( CodingStructure &cs, Partitioner &partitioner, bool cal
if ( leftEdgeAvai )
{
m_pcLoopFilter->resetFilterLengths();
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ //No Impact on OBMC Buffer
+ m_pcLoopFilter->xDeblockCU( *cu, EDGE_VER, false, m_tempWoOBMCBuffer );
+#else
m_pcLoopFilter->xDeblockCU( *cu, EDGE_VER );
+#endif
}
if (topEdgeAvai)
{
m_pcLoopFilter->resetFilterLengths();
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ //No Impact on OBMC Buffer
+ m_pcLoopFilter->xDeblockCU( *cu, EDGE_HOR, false, m_tempWoOBMCBuffer );
+#else
m_pcLoopFilter->xDeblockCU( *cu, EDGE_HOR );
+#endif
}
//update current CU SSE
diff --git a/source/Lib/EncoderLib/EncGOP.cpp b/source/Lib/EncoderLib/EncGOP.cpp
index 1d8c0cd314828a9799ab8705417a21dcea192484..f23b366bd9612471502010ea8b55c086d0615f5c 100644
--- a/source/Lib/EncoderLib/EncGOP.cpp
+++ b/source/Lib/EncoderLib/EncGOP.cpp
@@ -3771,7 +3771,9 @@ void EncGOP::compressGOP(int iPOCLast, int iNumPicRcvd, PicList &rcListPic, std:
m_pcALF->copyResiData(cs);
}
#endif
-
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ m_pcALF->callCodingInfoBuf( cs ).fill( 0 );
+#endif
// create SAO object based on the picture size
if( pcSlice->getSPS()->getSAOEnabledFlag()
#if JVET_W0066_CCSAO
@@ -3862,7 +3864,12 @@ void EncGOP::compressGOP(int iPOCLast, int iNumPicRcvd, PicList &rcListPic, std:
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf codingInfoBuf = m_pcALF->callCodingInfoBuf( cs );
+ m_pcLoopFilter->loopFilterPic( cs, codingInfoBuf, true );
+#else
m_pcLoopFilter->loopFilterPic( cs );
+#endif
#if !MULTI_PASS_DMVR
CS::setRefinedMotionField(cs);
@@ -5128,7 +5135,12 @@ uint64_t EncGOP::preLoopFilterPicAndCalcDist( Picture* pcPic )
}
#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ PelUnitBuf codingInfoBuf = m_pcALF->callCodingInfoBuf( cs );
+ m_pcLoopFilter->loopFilterPic( cs, codingInfoBuf, false );
+#else
m_pcLoopFilter->loopFilterPic( cs );
+#endif
const CPelUnitBuf picOrg = pcPic->getRecoBuf();
const CPelUnitBuf picRec = cs.getRecoBuf();
diff --git a/source/Lib/EncoderLib/EncLib.cpp b/source/Lib/EncoderLib/EncLib.cpp
index d3209d48c5d0393c47f9c28c089ee38f79578fb2..5a1cad6daf4be6e33ccbe0b45db5e89383cbad44 100644
--- a/source/Lib/EncoderLib/EncLib.cpp
+++ b/source/Lib/EncoderLib/EncLib.cpp
@@ -2225,6 +2225,9 @@ void EncLib::xInitSPS( SPS& sps )
#endif
#if JVET_AH0057_CCALF_COEFF_PRECISION
sps.setCCALFPrecisionFlag( m_ccalfPrecision );
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ sps.setAlfLumaFixedFilterAdjust( m_intraPeriod < 0 ? false : true );
#endif
sps.setJointCbCrEnabledFlag( m_JointCbCrMode );
sps.setMaxTLayers( m_maxTempLayer );
diff --git a/source/Lib/EncoderLib/VLCWriter.cpp b/source/Lib/EncoderLib/VLCWriter.cpp
index bdb2d71d23654628a9d1001ee3451c316293773d..11bb3fbeb9a5785d121f84e6cb6e2a8d8faf6b51 100644
--- a/source/Lib/EncoderLib/VLCWriter.cpp
+++ b/source/Lib/EncoderLib/VLCWriter.cpp
@@ -1344,6 +1344,12 @@ void HLSWriter::codeSPS( const SPS* pcSPS )
}
#endif
}
+#endif
+#if JVET_AJ0188_CODING_INFO_CLASSIFICATION
+ if( pcSPS->getALFEnabledFlag() )
+ {
+ WRITE_FLAG( pcSPS->getAlfLumaFixedFilterAdjust(), "sps_alf_luma_fixed_filter_adjust" );
+ }
#endif
if (pcSPS->getALFEnabledFlag() && pcSPS->getChromaFormatIdc() != CHROMA_400)
{