diff --git a/source/Lib/CommonLib/InterPrediction.cpp b/source/Lib/CommonLib/InterPrediction.cpp
index 69e1c5b811b30b8752a604b49226e5fa2f664aab..2d5d737f1fd1668bcdca59f9fa83ab728dabd989 100644
--- a/source/Lib/CommonLib/InterPrediction.cpp
+++ b/source/Lib/CommonLib/InterPrediction.cpp
@@ -798,6 +798,47 @@ void InterPrediction::xPredInterBlk ( const ComponentID& compID, const Predictio
}
}
+#if JVET_N0068_AFFINE_MEM_BW
+bool InterPrediction::isSubblockVectorSpreadOverLimit( int a, int b, int c, int d, int predType )
+{
+ int s4 = ( 4 << 11 );
+ int filterTap = 6;
+
+ if ( predType == 3 )
+ {
+ int refBlkWidth = std::max( std::max( 0, 4 * a + s4 ), std::max( 4 * c, 4 * a + 4 * c + s4 ) ) - std::min( std::min( 0, 4 * a + s4 ), std::min( 4 * c, 4 * a + 4 * c + s4 ) );
+ int refBlkHeight = std::max( std::max( 0, 4 * b ), std::max( 4 * d + s4, 4 * b + 4 * d + s4 ) ) - std::min( std::min( 0, 4 * b ), std::min( 4 * d + s4, 4 * b + 4 * d + s4 ) );
+ refBlkWidth = ( refBlkWidth >> 11 ) + filterTap + 3;
+ refBlkHeight = ( refBlkHeight >> 11 ) + filterTap + 3;
+
+ if ( refBlkWidth * refBlkHeight > ( filterTap + 9 ) * ( filterTap + 9 ) )
+ {
+ return true;
+ }
+ }
+ else
+ {
+ int refBlkWidth = std::max( 0, 4 * a + s4 ) - std::min( 0, 4 * a + s4 );
+ int refBlkHeight = std::max( 0, 4 * b ) - std::min( 0, 4 * b );
+ refBlkWidth = ( refBlkWidth >> 11 ) + filterTap + 3;
+ refBlkHeight = ( refBlkHeight >> 11 ) + filterTap + 3;
+ if ( refBlkWidth * refBlkHeight > ( filterTap + 9 ) * ( filterTap + 5 ) )
+ {
+ return true;
+ }
+
+ refBlkWidth = std::max( 0, 4 * c ) - std::min( 0, 4 * c );
+ refBlkHeight = std::max( 0, 4 * d + s4 ) - std::min( 0, 4 * d + s4 );
+ refBlkWidth = ( refBlkWidth >> 11 ) + filterTap + 3;
+ refBlkHeight = ( refBlkHeight >> 11 ) + filterTap + 3;
+ if ( refBlkWidth * refBlkHeight > ( filterTap + 5 ) * ( filterTap + 9 ) )
+ {
+ return true;
+ }
+ }
+ return false;
+}
+#endif
void InterPrediction::xPredAffineBlk( const ComponentID& compID, const PredictionUnit& pu, const Picture* refPic, const Mv* _mv, PelUnitBuf& dstPic, const bool& bi, const ClpRng& clpRng )
{
if ( (pu.cu->affineType == AFFINEMODEL_6PARAM && _mv[0] == _mv[1] && _mv[0] == _mv[2])
@@ -873,6 +914,13 @@ void InterPrediction::xPredAffineBlk( const ComponentID& compID, const Predictio
const int vFilterSize = isLuma(compID) ? NTAPS_LUMA : NTAPS_CHROMA;
const int shift = iBit - 4 + MV_FRACTIONAL_BITS_INTERNAL;
+#if JVET_N0068_AFFINE_MEM_BW
+ bool subblkMVSpreadOverLimit = false;
+ if ( isSubblockVectorSpreadOverLimit( iDMvHorX, iDMvHorY, iDMvVerX, iDMvVerY, pu.interDir ) )
+ {
+ subblkMVSpreadOverLimit = true;
+ }
+#endif
// get prediction block by block
@@ -888,9 +936,20 @@ void InterPrediction::xPredAffineBlk( const ComponentID& compID, const Predictio
if(compID == COMPONENT_Y)
#endif //JVET_N0671_AFFINE
{
- iMvScaleTmpHor = iMvScaleHor + iDMvHorX * (iHalfBW + w) + iDMvVerX * (iHalfBH + h);
- iMvScaleTmpVer = iMvScaleVer + iDMvHorY * (iHalfBW + w) + iDMvVerY * (iHalfBH + h);
-
+#if JVET_N0068_AFFINE_MEM_BW
+ if ( !subblkMVSpreadOverLimit )
+ {
+#endif
+ iMvScaleTmpHor = iMvScaleHor + iDMvHorX * (iHalfBW + w) + iDMvVerX * (iHalfBH + h);
+ iMvScaleTmpVer = iMvScaleVer + iDMvHorY * (iHalfBW + w) + iDMvVerY * (iHalfBH + h);
+#if JVET_N0068_AFFINE_MEM_BW
+ }
+ else
+ {
+ iMvScaleTmpHor = iMvScaleHor + iDMvHorX * ( cxWidth >> 1 ) + iDMvVerX * ( cxHeight >> 1 );
+ iMvScaleTmpVer = iMvScaleVer + iDMvHorY * ( cxWidth >> 1 ) + iDMvVerY * ( cxHeight >> 1 );
+ }
+#endif
roundAffineMv(iMvScaleTmpHor, iMvScaleTmpVer, shift);
Mv tmpMv(iMvScaleTmpHor, iMvScaleTmpVer);
tmpMv.clipToStorageBitDepth();
diff --git a/source/Lib/CommonLib/InterPrediction.h b/source/Lib/CommonLib/InterPrediction.h
index 760e2268a27fb1831f5bc10f15e0592f6d3791b7..5130f46f8a5f76edbb1b5a33d1bc192594f5c0c8 100644
--- a/source/Lib/CommonLib/InterPrediction.h
+++ b/source/Lib/CommonLib/InterPrediction.h
@@ -190,6 +190,9 @@ public:
#if ENABLE_SPLIT_PARALLELISM
int getShareState() const { return m_shareState; }
#endif
+#if JVET_N0068_AFFINE_MEM_BW
+ bool isSubblockVectorSpreadOverLimit( int a, int b, int c, int d, int predType );
+#endif
};
//! \}
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index 63bd3f0f85c4cf54f759057a2b27ab2769dd233a..90dce7152866f1e8e27d27f2194b062dbdbf0c7b 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -50,6 +50,8 @@
#include <assert.h>
#include <cassert>
+#define JVET_N0068_AFFINE_MEM_BW 1 // memory bandwidth reduction for affine mode
+
#define JVET_N0308_MAX_CU_SIZE_FOR_ISP 1
#define JVET_N0280_RESIDUAL_CODING_TS 1
diff --git a/source/Lib/CommonLib/UnitTools.cpp b/source/Lib/CommonLib/UnitTools.cpp
index 3bab928cf55e03ebdefcb2d3a4c77929d5648118..9d343da3cccb3b16f26ffc67d7a851e3b34b3cd8 100644
--- a/source/Lib/CommonLib/UnitTools.cpp
+++ b/source/Lib/CommonLib/UnitTools.cpp
@@ -4032,14 +4032,30 @@ void PU::setAllAffineMv( PredictionUnit& pu, Mv affLT, Mv affRT, Mv affLB, RefPi
MotionBuf mb = pu.getMotionBuf();
int mvScaleTmpHor, mvScaleTmpVer;
-
+#if JVET_N0068_AFFINE_MEM_BW
+ bool subblkMVSpreadOverLimit = false;
+ InterPrediction temp;
+ subblkMVSpreadOverLimit = temp.isSubblockVectorSpreadOverLimit( deltaMvHorX, deltaMvHorY, deltaMvVerX, deltaMvVerY, pu.interDir );
+#endif
for ( int h = 0; h < pu.Y().height; h += blockHeight )
{
for ( int w = 0; w < pu.Y().width; w += blockWidth )
{
- mvScaleTmpHor = mvScaleHor + deltaMvHorX * (halfBW + w) + deltaMvVerX * (halfBH + h);
- mvScaleTmpVer = mvScaleVer + deltaMvHorY * (halfBW + w) + deltaMvVerY * (halfBH + h);
+#if JVET_N0068_AFFINE_MEM_BW
+ if ( !subblkMVSpreadOverLimit )
+ {
+#endif
+ mvScaleTmpHor = mvScaleHor + deltaMvHorX * (halfBW + w) + deltaMvVerX * (halfBH + h);
+ mvScaleTmpVer = mvScaleVer + deltaMvHorY * (halfBW + w) + deltaMvVerY * (halfBH + h);
+#if JVET_N0068_AFFINE_MEM_BW
+ }
+ else
+ {
+ mvScaleTmpHor = mvScaleHor + deltaMvHorX * ( pu.Y().width >> 1 ) + deltaMvVerX * ( pu.Y().height >> 1 );
+ mvScaleTmpVer = mvScaleVer + deltaMvHorY * ( pu.Y().width >> 1 ) + deltaMvVerY * ( pu.Y().height >> 1 );
+ }
+#endif
roundAffineMv( mvScaleTmpHor, mvScaleTmpVer, shift );
Mv curMv(mvScaleTmpHor, mvScaleTmpVer);
curMv.clipToStorageBitDepth();
diff --git a/source/Lib/EncoderLib/InterSearch.cpp b/source/Lib/EncoderLib/InterSearch.cpp
index af871b34e8c07ac79f4d44957391c7768a07a756..59e74f97b3c7aa574bb10f420160c2fa1e50029e 100644
--- a/source/Lib/EncoderLib/InterSearch.cpp
+++ b/source/Lib/EncoderLib/InterSearch.cpp
@@ -4306,7 +4306,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
for ( int iRefList = 0; iRefList < iNumPredDir; iRefList++ )
{
RefPicList eRefPicList = ( iRefList ? REF_PIC_LIST_1 : REF_PIC_LIST_0 );
-
+#if JVET_N0068_AFFINE_MEM_BW
+ pu.interDir = ( iRefList ? 2 : 1 );
+#endif
for (int iRefIdxTemp = 0; iRefIdxTemp < slice.getNumRefIdx(eRefPicList); iRefIdxTemp++)
{
// Get RefIdx bits
@@ -4645,7 +4647,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
if ( slice.isInterB() && !PU::isBipredRestriction(pu) )
{
tryBipred = 1;
-
+#if JVET_N0068_AFFINE_MEM_BW
+ pu.interDir = 3;
+#endif
// Set as best list0 and list1
iRefIdxBi[0] = iRefIdx[0];
iRefIdxBi[1] = iRefIdx[1];
@@ -4901,7 +4905,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
{
lastMode = 2;
affineCost = uiCostBi;
-
+#if JVET_N0068_AFFINE_MEM_BW
+ pu.interDir = 3;
+#endif
PU::setAllAffineMv( pu, cMvBi[0][0], cMvBi[0][1], cMvBi[0][2], REF_PIC_LIST_0
, changeToHighPrec
);
@@ -4922,7 +4928,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
}
}
+#if !JVET_N0068_AFFINE_MEM_BW
pu.interDir = 3;
+#endif
pu.mvpIdx[REF_PIC_LIST_0] = aaiMvpIdxBi[0][iRefIdxBi[0]];
pu.mvpNum[REF_PIC_LIST_0] = aaiMvpNum[0][iRefIdxBi[0]];
@@ -4933,7 +4941,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
{
lastMode = 0;
affineCost = uiCost[0];
-
+#if JVET_N0068_AFFINE_MEM_BW
+ pu.interDir = 1;
+#endif
PU::setAllAffineMv( pu, aacMv[0][0], aacMv[0][1], aacMv[0][2], REF_PIC_LIST_0
, changeToHighPrec
);
@@ -4947,8 +4957,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
pu.mvdAffi[0][verIdx] = pu.mvdAffi[0][verIdx] - pu.mvdAffi[0][0];
}
}
-
+#if !JVET_N0068_AFFINE_MEM_BW
pu.interDir = 1;
+#endif
pu.mvpIdx[REF_PIC_LIST_0] = aaiMvpIdx[0][iRefIdx[0]];
pu.mvpNum[REF_PIC_LIST_0] = aaiMvpNum[0][iRefIdx[0]];
@@ -4957,7 +4968,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
{
lastMode = 1;
affineCost = uiCost[1];
-
+#if JVET_N0068_AFFINE_MEM_BW
+ pu.interDir = 2;
+#endif
PU::setAllAffineMv( pu, aacMv[1][0], aacMv[1][1], aacMv[1][2], REF_PIC_LIST_1
, changeToHighPrec
);
@@ -4971,8 +4984,9 @@ void InterSearch::xPredAffineInterSearch( PredictionUnit& pu,
pu.mvdAffi[1][verIdx] = pu.mvdAffi[1][verIdx] - pu.mvdAffi[1][0];
}
}
-
+#if !JVET_N0068_AFFINE_MEM_BW
pu.interDir = 2;
+#endif
pu.mvpIdx[REF_PIC_LIST_1] = aaiMvpIdx[1][iRefIdx[1]];
pu.mvpNum[REF_PIC_LIST_1] = aaiMvpNum[1][iRefIdx[1]];