From 8eaa9292e57675aa7ff1d7b890c63a4b506dc06d Mon Sep 17 00:00:00 2001
From: Frank Bossen <fbossen@gmail.com>
Date: Fri, 8 Jul 2022 23:17:32 -0400
Subject: [PATCH] Clean up filter management for MC interpolation
- Enumerate filter types
- Define 6-tap vertical affine filter to reduce number of operations
---
source/Lib/CommonLib/CommonDef.h | 10 +-
source/Lib/CommonLib/InterPrediction.cpp | 96 ++++-----
source/Lib/CommonLib/InterpolationFilter.cpp | 189 +++++++-----------
source/Lib/CommonLib/InterpolationFilter.h | 22 +-
.../CommonLib/x86/InterpolationFilterX86.h | 48 +++--
source/Lib/EncoderLib/InterSearch.cpp | 12 +-
6 files changed, 171 insertions(+), 206 deletions(-)
diff --git a/source/Lib/CommonLib/CommonDef.h b/source/Lib/CommonLib/CommonDef.h
index 039a162f4d..10e2d56e9f 100644
--- a/source/Lib/CommonLib/CommonDef.h
+++ b/source/Lib/CommonLib/CommonDef.h
@@ -408,10 +408,12 @@ static constexpr int MAX_TESTED_QPs = ( 1 + 1 + ( MAX_DELTA_QP << 1 ) );
static constexpr int COM16_C806_TRANS_PREC = 0;
-static constexpr int NTAPS_LUMA = 8; // Number of taps for luma
-static constexpr int NTAPS_CHROMA = 4; // Number of taps for chroma
-static constexpr int NTAPS_BILINEAR = 2; // Number of taps for bilinear filter
-static constexpr int MAX_FILTER_SIZE = NTAPS_LUMA > NTAPS_CHROMA ? NTAPS_LUMA : NTAPS_CHROMA;
+static constexpr int NTAPS_LUMA = 8; // Number of taps for luma
+static constexpr int NTAPS_LUMA_AFFINE = 6; // Number of taps for luma affine
+static constexpr int NTAPS_CHROMA = 4; // Number of taps for chroma
+static constexpr int NTAPS_CHROMA_AFFINE = 4; // Number of taps for chroma affine
+static constexpr int NTAPS_BILINEAR = 2; // Number of taps for bilinear filter
+static constexpr int MAX_FILTER_SIZE = NTAPS_LUMA > NTAPS_CHROMA ? NTAPS_LUMA : NTAPS_CHROMA;
#if LUMA_ADAPTIVE_DEBLOCKING_FILTER_QP_OFFSET
static constexpr int MAX_LADF_INTERVALS = 5; /// max number of luma adaptive deblocking filter qp offset intervals
diff --git a/source/Lib/CommonLib/InterPrediction.cpp b/source/Lib/CommonLib/InterPrediction.cpp
index 21721a8bec..3385f49b04 100644
--- a/source/Lib/CommonLib/InterPrediction.cpp
+++ b/source/Lib/CommonLib/InterPrediction.cpp
@@ -760,15 +760,17 @@ void InterPrediction::xPredInterBlk(const ComponentID &compID, const PredictionU
dstBuf.buf = m_filteredBlockTmp[2 + m_iRefListIdx][compID] + 2 * dstBuf.stride + 2;
}
+ const int filterIdx = bilinearMC ? InterpolationFilter::FILTER_DMVR : InterpolationFilter::FILTER_DEFAULT;
+
if (yFrac == 0)
{
m_if.filterHor(compID, (Pel *) refBuf.buf, refBuf.stride, dstBuf.buf, dstBuf.stride, backupWidth, backupHeight,
- xFrac, rndRes, clpRng, bilinearMC, bilinearMC, useAltHpelIf);
+ xFrac, rndRes, clpRng, filterIdx, useAltHpelIf);
}
else if (xFrac == 0)
{
m_if.filterVer(compID, (Pel *) refBuf.buf, refBuf.stride, dstBuf.buf, dstBuf.stride, backupWidth, backupHeight,
- yFrac, true, rndRes, clpRng, bilinearMC, bilinearMC, useAltHpelIf);
+ yFrac, true, rndRes, clpRng, filterIdx, useAltHpelIf);
}
else
{
@@ -785,12 +787,11 @@ void InterPrediction::xPredInterBlk(const ComponentID &compID, const PredictionU
vFilterSize = NTAPS_BILINEAR;
}
m_if.filterHor(compID, (Pel *) refBuf.buf - ((vFilterSize >> 1) - 1) * refBuf.stride, refBuf.stride, tmpBuf.buf,
- tmpBuf.stride, backupWidth, backupHeight + vFilterSize - 1, xFrac, false, clpRng, bilinearMC,
- bilinearMC, useAltHpelIf);
+ tmpBuf.stride, backupWidth, backupHeight + vFilterSize - 1, xFrac, false, clpRng, filterIdx,
+ useAltHpelIf);
JVET_J0090_SET_CACHE_ENABLE(false);
m_if.filterVer(compID, (Pel *) tmpBuf.buf + ((vFilterSize >> 1) - 1) * tmpBuf.stride, tmpBuf.stride, dstBuf.buf,
- dstBuf.stride, backupWidth, backupHeight, yFrac, false, rndRes, clpRng, bilinearMC, bilinearMC,
- useAltHpelIf);
+ dstBuf.stride, backupWidth, backupHeight, yFrac, false, rndRes, clpRng, filterIdx, useAltHpelIf);
}
JVET_J0090_SET_CACHE_ENABLE(
(srcPadStride == 0)
@@ -1100,12 +1101,14 @@ void InterPrediction::xPredAffineBlk(const ComponentID &compID, const Prediction
}
#endif
+ const int filterIdx = InterpolationFilter::FILTER_AFFINE;
+
if( isRefScaled )
{
CHECK(enableProf, "PROF should be disabled with RPR");
xPredInterBlkRPR(scalingRatio, pps,
CompArea(compID, chFmt, pu.blocks[compID].offset(w, h), Size(sbWidth, sbHeight)), refPic,
- curMv, dstBuf.buf + w + h * dstBuf.stride, dstBuf.stride, bi, wrapRef, clpRng, 2);
+ curMv, dstBuf.buf + w + h * dstBuf.stride, dstBuf.stride, bi, wrapRef, clpRng, filterIdx);
}
else
{
@@ -1139,21 +1142,23 @@ void InterPrediction::xPredAffineBlk(const ComponentID &compID, const Prediction
if (yFrac == 0)
{
- m_if.filterHor(compID, ref, refStride, dst, dstStride, sbWidth, sbHeight, xFrac, isLast, clpRng);
+ m_if.filterHor(compID, ref, refStride, dst, dstStride, sbWidth, sbHeight, xFrac, isLast, clpRng, filterIdx);
}
else if (xFrac == 0)
{
- m_if.filterVer(compID, ref, refStride, dst, dstStride, sbWidth, sbHeight, yFrac, true, isLast, clpRng);
+ m_if.filterVer(compID, ref, refStride, dst, dstStride, sbWidth, sbHeight, yFrac, true, isLast, clpRng,
+ filterIdx);
}
else
{
- const int filterSize = isLuma(compID) ? NTAPS_LUMA : NTAPS_CHROMA;
+ const int filterSize = isLuma(compID) ? NTAPS_LUMA_AFFINE : NTAPS_CHROMA_AFFINE;
+ const int rowsAbove = (filterSize - 1) >> 1;
- m_if.filterHor(compID, ref - ((filterSize >> 1) - 1) * refStride, refStride, tmpBuf.buf, tmpBuf.stride,
- sbWidth, sbHeight + filterSize - 1, xFrac, false, clpRng);
+ m_if.filterHor(compID, ref - rowsAbove * refStride, refStride, tmpBuf.buf, tmpBuf.stride, sbWidth,
+ sbHeight + filterSize - 1, xFrac, false, clpRng, filterIdx);
JVET_J0090_SET_CACHE_ENABLE(false);
- m_if.filterVer(compID, tmpBuf.buf + ((filterSize >> 1) - 1) * tmpBuf.stride, tmpBuf.stride, dst, dstStride,
- sbWidth, sbHeight, yFrac, false, isLast, clpRng);
+ m_if.filterVer(compID, tmpBuf.buf + rowsAbove * tmpBuf.stride, tmpBuf.stride, dst, dstStride, sbWidth,
+ sbHeight, yFrac, false, isLast, clpRng, filterIdx);
JVET_J0090_SET_CACHE_ENABLE(true);
}
@@ -2300,67 +2305,44 @@ bool InterPrediction::xPredInterBlkRPR( const std::pair<int, int>& scalingRatio,
int yFilter = filterIndex;
const int rprThreshold1 = ( 1 << SCALE_RATIO_BITS ) * 5 / 4;
const int rprThreshold2 = ( 1 << SCALE_RATIO_BITS ) * 7 / 4;
- if( filterIndex == 0 )
+ if (filterIndex == InterpolationFilter::FILTER_DEFAULT || !isLuma(compID))
{
if( scalingRatio.first > rprThreshold2 )
{
- xFilter = 4;
+ xFilter = InterpolationFilter::FILTER_RPR2;
}
else if( scalingRatio.first > rprThreshold1 )
{
- xFilter = 3;
+ xFilter = InterpolationFilter::FILTER_RPR1;
}
if( scalingRatio.second > rprThreshold2 )
{
- yFilter = 4;
+ yFilter = InterpolationFilter::FILTER_RPR2;
}
else if( scalingRatio.second > rprThreshold1 )
{
- yFilter = 3;
+ yFilter = InterpolationFilter::FILTER_RPR1;
}
}
- if (filterIndex == 2)
+ else if (filterIndex == InterpolationFilter::FILTER_AFFINE)
{
- if (isLuma(compID))
+ if (scalingRatio.first > rprThreshold2)
{
- if (scalingRatio.first > rprThreshold2)
- {
- xFilter = 6;
- }
- else if (scalingRatio.first > rprThreshold1)
- {
- xFilter = 5;
- }
-
- if (scalingRatio.second > rprThreshold2)
- {
- yFilter = 6;
- }
- else if (scalingRatio.second > rprThreshold1)
- {
- yFilter = 5;
- }
+ xFilter = InterpolationFilter::FILTER_AFFINE_RPR2;
}
- else
+ else if (scalingRatio.first > rprThreshold1)
{
- if (scalingRatio.first > rprThreshold2)
- {
- xFilter = 4;
- }
- else if (scalingRatio.first > rprThreshold1)
- {
- xFilter = 3;
- }
+ xFilter = InterpolationFilter::FILTER_AFFINE_RPR1;
+ }
- if (scalingRatio.second > rprThreshold2)
- {
- yFilter = 4;
- }
- else if (scalingRatio.second > rprThreshold1)
- {
- yFilter = 3;
- }
+ if (scalingRatio.second > rprThreshold2)
+ {
+ yFilter = InterpolationFilter::FILTER_AFFINE_RPR2;
+ }
+ else if (scalingRatio.second > rprThreshold1)
+ {
+ yFilter = InterpolationFilter::FILTER_AFFINE_RPR1;
}
}
@@ -2433,7 +2415,7 @@ bool InterPrediction::xPredInterBlkRPR( const std::pair<int, int>& scalingRatio,
Pel *const tempBuf = m_filteredBlockTmpRPR + col;
m_if.filterHor(compID, (Pel *) refBuf.buf - ((vFilterSize >> 1) - 1) * refBuf.stride, refBuf.stride, tempBuf,
- tmpStride, 1, refHeight + vFilterSize - 1 + extSize, xFrac, false, clpRng, xFilter, false,
+ tmpStride, 1, refHeight + vFilterSize - 1 + extSize, xFrac, false, clpRng, xFilter,
useAltHpelIf && scalingRatio.first == 1 << SCALE_RATIO_BITS);
}
@@ -2450,7 +2432,7 @@ bool InterPrediction::xPredInterBlkRPR( const std::pair<int, int>& scalingRatio,
JVET_J0090_SET_CACHE_ENABLE( false );
m_if.filterVer(compID, tempBuf + ((vFilterSize >> 1) - 1) * tmpStride, tmpStride, dst + row * dstStride,
- dstStride, width, 1, yFrac, false, rndRes, clpRng, yFilter, false,
+ dstStride, width, 1, yFrac, false, rndRes, clpRng, yFilter,
useAltHpelIf && scalingRatio.second == 1 << SCALE_RATIO_BITS);
JVET_J0090_SET_CACHE_ENABLE( true );
}
diff --git a/source/Lib/CommonLib/InterpolationFilter.cpp b/source/Lib/CommonLib/InterpolationFilter.cpp
index 88304e6919..431123e6b6 100644
--- a/source/Lib/CommonLib/InterpolationFilter.cpp
+++ b/source/Lib/CommonLib/InterpolationFilter.cpp
@@ -54,6 +54,7 @@ CacheModel* InterpolationFilter::m_cacheModel;
// ====================================================================================================================
// Tables
// ====================================================================================================================
+// TODO: implement 6-tap horizontal filtering in SIMD code such that m_affineLumaFilter can be used instead of m_lumaFilter4x4
const TFilterCoeff InterpolationFilter::m_lumaFilter4x4[LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS][NTAPS_LUMA] =
{
{ 0, 0, 0, 64, 0, 0, 0, 0 },
@@ -74,6 +75,31 @@ const TFilterCoeff InterpolationFilter::m_lumaFilter4x4[LUMA_INTERPOLATION_FILTE
{ 0, 1, -2, 4, 63, -3, 1, 0 }
};
+// clang-format off
+const TFilterCoeff InterpolationFilter::m_affineLumaFilter[LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS][NTAPS_LUMA] =
+{
+ { 0, 0, 64, 0, 0, 0, 0, 0 },
+ { 1, -3, 63, 4, -2, 1, 0, 0 },
+ { 1, -5, 62, 8, -3, 1, 0, 0 },
+ { 2, -8, 60, 13, -4, 1, 0, 0 },
+
+ { 3, -10, 58, 17, -5, 1, 0, 0 },
+ { 3, -11, 52, 26, -8, 2, 0, 0 },
+ { 2, -9, 47, 31, -10, 3, 0, 0 },
+ { 3, -11, 45, 34, -10, 3, 0, 0 },
+
+ { 3, -11, 40, 40, -11, 3, 0, 0 },
+ { 3, -10, 34, 45, -11, 3, 0, 0 },
+ { 3, -10, 31, 47, -9, 2, 0, 0 },
+ { 2, -8, 26, 52, -11, 3, 0, 0 },
+
+ { 1, -5, 17, 58, -10, 3, 0, 0 },
+ { 1, -4, 13, 60, -8, 2, 0, 0 },
+ { 1, -3, 8, 62, -5, 1, 0, 0 },
+ { 1, -2, 4, 63, -3, 1, 0, 0 }
+};
+// clang-format on
+
const TFilterCoeff InterpolationFilter::m_lumaFilter[LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS][NTAPS_LUMA] =
{
{ 0, 0, 0, 64, 0, 0, 0, 0 },
@@ -344,6 +370,11 @@ InterpolationFilter::InterpolationFilter()
m_filterVer[2][1][0] = filter<2, true, true, false>;
m_filterVer[2][1][1] = filter<2, true, true, true>;
+ m_filterVer[3][0][0] = filter<6, true, false, false>;
+ m_filterVer[3][0][1] = filter<6, true, false, true>;
+ m_filterVer[3][1][0] = filter<6, true, true, false>;
+ m_filterVer[3][1][1] = filter<6, true, true, true>;
+
m_filterCopy[0][0] = filterCopy<false, false>;
m_filterCopy[0][1] = filterCopy<false, true>;
m_filterCopy[1][0] = filterCopy<true, false>;
@@ -625,128 +656,65 @@ void InterpolationFilter::filter(const ClpRng& clpRng, Pel const *src, int srcSt
}
}
-/**
- * \brief Filter a block of samples (horizontal)
- *
- * \tparam N Number of taps
- * \param bitDepth Bit depth of samples
- * \param src Pointer to source samples
- * \param srcStride Stride of source samples
- * \param dst Pointer to destination samples
- * \param dstStride Stride of destination samples
- * \param width Width of block
- * \param height Height of block
- * \param isLast Flag indicating whether it is the last filtering operation
- * \param coeff Pointer to filter taps
- */
+static constexpr int tapToIdx(const int N)
+{
+ return N == 8 ? 0 : (N == 4 ? 1 : (N == 2 ? 2 : (N == 6 ? 3 : 4)));
+}
+
template<int N>
void InterpolationFilter::filterHor(const ClpRng& clpRng, Pel const *src, int srcStride, Pel *dst, int dstStride, int width, int height, bool isLast, TFilterCoeff const *coeff, bool biMCForDMVR)
{
-//#if ENABLE_SIMD_OPT_MCIF
- if( N == 8 )
- {
- m_filterHor[0][1][isLast](clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
- }
- else if( N == 4 )
- {
- m_filterHor[1][1][isLast](clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
- }
- else if( N == 2 )
- {
- m_filterHor[2][1][isLast](clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
- }
- else
- {
- THROW( "Invalid tap number" );
- }
+ constexpr int IDX = tapToIdx(N);
+ static_assert(IDX < 3, "Unsupported tap count");
+ m_filterHor[IDX][1][isLast](clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
}
-/**
- * \brief Filter a block of samples (vertical)
- *
- * \tparam N Number of taps
- * \param bitDepth Bit depth
- * \param src Pointer to source samples
- * \param srcStride Stride of source samples
- * \param dst Pointer to destination samples
- * \param dstStride Stride of destination samples
- * \param width Width of block
- * \param height Height of block
- * \param isFirst Flag indicating whether it is the first filtering operation
- * \param isLast Flag indicating whether it is the last filtering operation
- * \param coeff Pointer to filter taps
- */
template<int N>
void InterpolationFilter::filterVer(const ClpRng& clpRng, Pel const *src, int srcStride, Pel *dst, int dstStride, int width, int height, bool isFirst, bool isLast, TFilterCoeff const *coeff, bool biMCForDMVR)
{
-//#if ENABLE_SIMD_OPT_MCIF
- if( N == 8 )
- {
- m_filterVer[0][isFirst][isLast]( clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
- }
- else if( N == 4 )
- {
- m_filterVer[1][isFirst][isLast]( clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
- }
- else if( N == 2 )
- {
- m_filterVer[2][isFirst][isLast]( clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
- }
- else{
- THROW( "Invalid tap number" );
- }
+ constexpr int IDX = tapToIdx(N);
+ static_assert(IDX < 4, "Unsupported tap count");
+ m_filterVer[IDX][isFirst][isLast](clpRng, src, srcStride, dst, dstStride, width, height, coeff, biMCForDMVR);
}
// ====================================================================================================================
// Public member functions
// ====================================================================================================================
-/**
- * \brief Filter a block of Luma/Chroma samples (horizontal)
- *
- * \param compID Chroma component ID
- * \param src Pointer to source samples
- * \param srcStride Stride of source samples
- * \param dst Pointer to destination samples
- * \param dstStride Stride of destination samples
- * \param width Width of block
- * \param height Height of block
- * \param frac Fractional sample offset
- * \param isLast Flag indicating whether it is the last filtering operation
- * \param bitDepth Bit depth
- */
void InterpolationFilter::filterHor(const ComponentID compID, Pel const *src, int srcStride, Pel *dst, int dstStride,
int width, int height, int frac, bool isLast, const ClpRng &clpRng, int nFilterIdx,
- bool biMCForDMVR, bool useAltHpelIf)
+ bool useAltHpelIf)
{
- if( frac == 0 && nFilterIdx < 2 )
+ const bool biMCForDMVR = nFilterIdx == FILTER_DMVR;
+
+ if (frac == 0 && nFilterIdx <= FILTER_AFFINE)
{
m_filterCopy[true][isLast]( clpRng, src, srcStride, dst, dstStride, width, height, biMCForDMVR );
}
else if( isLuma( compID ) )
{
CHECK( frac < 0 || frac >= LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS, "Invalid fraction" );
- if( nFilterIdx == 1 )
+ if (nFilterIdx == FILTER_DMVR)
{
filterHor<NTAPS_BILINEAR>( clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_bilinearFilterPrec4[frac], biMCForDMVR );
}
- else if( nFilterIdx == 2 )
+ else if (nFilterIdx == FILTER_AFFINE)
{
filterHor<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_lumaFilter4x4[frac], biMCForDMVR );
}
- else if( nFilterIdx == 3 )
+ else if (nFilterIdx == FILTER_RPR1)
{
filterHor<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_lumaFilterRPR1[frac], biMCForDMVR );
}
- else if( nFilterIdx == 4 )
+ else if (nFilterIdx == FILTER_RPR2)
{
filterHor<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_lumaFilterRPR2[frac], biMCForDMVR );
}
- else if (nFilterIdx == 5)
+ else if (nFilterIdx == FILTER_AFFINE_RPR1)
{
filterHor<NTAPS_LUMA>(clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_affineLumaFilterRPR1[frac], biMCForDMVR);
}
- else if (nFilterIdx == 6)
+ else if (nFilterIdx == FILTER_AFFINE_RPR2)
{
filterHor<NTAPS_LUMA>(clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_affineLumaFilterRPR2[frac], biMCForDMVR);
}
@@ -754,10 +722,6 @@ void InterpolationFilter::filterHor(const ComponentID compID, Pel const *src, in
{
filterHor<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_lumaAltHpelIFilter, biMCForDMVR );
}
- else if( ( width == 4 && height == 4 ) || ( width == 4 && height == ( 4 + NTAPS_LUMA - 1 ) ) )
- {
- filterHor<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_lumaFilter4x4[frac], biMCForDMVR );
- }
else
{
filterHor<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_lumaFilter[frac], biMCForDMVR );
@@ -766,12 +730,12 @@ void InterpolationFilter::filterHor(const ComponentID compID, Pel const *src, in
else
{
CHECK(frac < 0 || frac >= CHROMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS, "Invalid fraction");
- if( nFilterIdx == 3 )
+ if (nFilterIdx == FILTER_RPR1)
{
filterHor<NTAPS_CHROMA>(clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_chromaFilterRPR1[frac],
biMCForDMVR);
}
- else if( nFilterIdx == 4 )
+ else if (nFilterIdx == FILTER_RPR2)
{
filterHor<NTAPS_CHROMA>(clpRng, src, srcStride, dst, dstStride, width, height, isLast, m_chromaFilterRPR2[frac],
biMCForDMVR);
@@ -784,54 +748,41 @@ void InterpolationFilter::filterHor(const ComponentID compID, Pel const *src, in
}
}
-
-/**
- * \brief Filter a block of Luma/Chroma samples (vertical)
- *
- * \param compID Colour component ID
- * \param src Pointer to source samples
- * \param srcStride Stride of source samples
- * \param dst Pointer to destination samples
- * \param dstStride Stride of destination samples
- * \param width Width of block
- * \param height Height of block
- * \param frac Fractional sample offset
- * \param isFirst Flag indicating whether it is the first filtering operation
- * \param isLast Flag indicating whether it is the last filtering operation
- * \param bitDepth Bit depth
- */
void InterpolationFilter::filterVer(const ComponentID compID, Pel const *src, int srcStride, Pel *dst, int dstStride,
int width, int height, int frac, bool isFirst, bool isLast, const ClpRng &clpRng,
- int nFilterIdx, bool biMCForDMVR, bool useAltHpelIf)
+ int nFilterIdx, bool useAltHpelIf)
{
- if( frac == 0 && nFilterIdx < 2 )
+ const bool biMCForDMVR = nFilterIdx == FILTER_DMVR;
+
+ if (frac == 0 && nFilterIdx <= FILTER_AFFINE)
{
m_filterCopy[isFirst][isLast]( clpRng, src, srcStride, dst, dstStride, width, height, biMCForDMVR );
}
else if( isLuma( compID ) )
{
CHECK( frac < 0 || frac >= LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS, "Invalid fraction" );
- if( nFilterIdx == 1 )
+ if (nFilterIdx == FILTER_DMVR)
{
filterVer<NTAPS_BILINEAR>( clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_bilinearFilterPrec4[frac], biMCForDMVR );
}
- else if( nFilterIdx == 2 )
+ else if (nFilterIdx == FILTER_AFFINE)
{
- filterVer<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_lumaFilter4x4[frac], biMCForDMVR );
+ filterVer<NTAPS_LUMA_AFFINE>(clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast,
+ m_affineLumaFilter[frac], biMCForDMVR);
}
- else if( nFilterIdx == 3 )
+ else if (nFilterIdx == FILTER_RPR1)
{
filterVer<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_lumaFilterRPR1[frac], biMCForDMVR );
}
- else if( nFilterIdx == 4 )
+ else if (nFilterIdx == FILTER_RPR2)
{
filterVer<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_lumaFilterRPR2[frac], biMCForDMVR );
}
- else if (nFilterIdx == 5)
+ else if (nFilterIdx == FILTER_AFFINE_RPR1)
{
filterVer<NTAPS_LUMA>(clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_affineLumaFilterRPR1[frac], biMCForDMVR);
}
- else if (nFilterIdx == 6)
+ else if (nFilterIdx == FILTER_AFFINE_RPR2)
{
filterVer<NTAPS_LUMA>(clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_affineLumaFilterRPR2[frac], biMCForDMVR);
}
@@ -839,10 +790,6 @@ void InterpolationFilter::filterVer(const ComponentID compID, Pel const *src, in
{
filterVer<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_lumaAltHpelIFilter, biMCForDMVR );
}
- else if( width == 4 && height == 4 )
- {
- filterVer<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_lumaFilter4x4[frac], biMCForDMVR );
- }
else
{
filterVer<NTAPS_LUMA>( clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast, m_lumaFilter[frac], biMCForDMVR );
@@ -851,12 +798,12 @@ void InterpolationFilter::filterVer(const ComponentID compID, Pel const *src, in
else
{
CHECK(frac < 0 || frac >= CHROMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS, "Invalid fraction");
- if( nFilterIdx == 3 )
+ if (nFilterIdx == FILTER_RPR1)
{
filterVer<NTAPS_CHROMA>(clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast,
m_chromaFilterRPR1[frac], biMCForDMVR);
}
- else if( nFilterIdx == 4 )
+ else if (nFilterIdx == FILTER_RPR2)
{
filterVer<NTAPS_CHROMA>(clpRng, src, srcStride, dst, dstStride, width, height, isFirst, isLast,
m_chromaFilterRPR2[frac], biMCForDMVR);
diff --git a/source/Lib/CommonLib/InterpolationFilter.h b/source/Lib/CommonLib/InterpolationFilter.h
index 2a99646f22..abdbc51ab7 100644
--- a/source/Lib/CommonLib/InterpolationFilter.h
+++ b/source/Lib/CommonLib/InterpolationFilter.h
@@ -57,6 +57,8 @@ static inline int IF_INTERNAL_FRAC_BITS(const int bd) { return std::max(2, IF_IN
class InterpolationFilter
{
static const TFilterCoeff m_lumaFilter4x4[LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS][NTAPS_LUMA];
+ static const TFilterCoeff m_affineLumaFilter[LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS][NTAPS_LUMA];
+
public:
static const TFilterCoeff m_lumaFilter[LUMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS][NTAPS_LUMA]; ///< Luma filter taps
static const TFilterCoeff m_chromaFilter[CHROMA_INTERPOLATION_FILTER_SUB_SAMPLE_POSITIONS][NTAPS_CHROMA]; ///< Chroma filter taps
@@ -88,10 +90,22 @@ protected:
static CacheModel* m_cacheModel;
#endif
public:
+ enum
+ {
+ FILTER_DEFAULT = 0,
+ FILTER_DMVR,
+ FILTER_AFFINE,
+ FILTER_RPR1,
+ FILTER_RPR2,
+ FILTER_AFFINE_RPR1,
+ FILTER_AFFINE_RPR2,
+ };
+
InterpolationFilter();
~InterpolationFilter() {}
void( *m_filterHor[3][2][2] )( const ClpRng& clpRng, Pel const *src, int srcStride, Pel *dst, int dstStride, int width, int height, TFilterCoeff const *coeff, bool biMCForDMVR);
- void( *m_filterVer[3][2][2] )( const ClpRng& clpRng, Pel const *src, int srcStride, Pel *dst, int dstStride, int width, int height, TFilterCoeff const *coeff, bool biMCForDMVR);
+ void (*m_filterVer[4][2][2])(const ClpRng &clpRng, Pel const *src, int srcStride, Pel *dst, int dstStride, int width,
+ int height, TFilterCoeff const *coeff, bool biMCForDMVR);
void( *m_filterCopy[2][2] ) ( const ClpRng& clpRng, Pel const *src, int srcStride, Pel *dst, int dstStride, int width, int height, bool biMCForDMVR);
void( *m_weightedGeoBlk )(const PredictionUnit &pu, const uint32_t width, const uint32_t height, const ComponentID compIdx, const uint8_t splitDir, PelUnitBuf& predDst, PelUnitBuf& predSrc0, PelUnitBuf& predSrc1);
@@ -102,11 +116,11 @@ public:
void _initInterpolationFilterX86();
#endif
void filterHor(const ComponentID compID, Pel const *src, int srcStride, Pel *dst, int dstStride, int width,
- int height, int frac, bool isLast, const ClpRng &clpRng, int nFilterIdx = 0, bool biMCForDMVR = false,
+ int height, int frac, bool isLast, const ClpRng &clpRng, int nFilterIdx = FILTER_DEFAULT,
bool useAltHpelIf = false);
void filterVer(const ComponentID compID, Pel const *src, int srcStride, Pel *dst, int dstStride, int width,
- int height, int frac, bool isFirst, bool isLast, const ClpRng &clpRng, int nFilterIdx = 0,
- bool biMCForDMVR = false, bool useAltHpelIf = false);
+ int height, int frac, bool isFirst, bool isLast, const ClpRng &clpRng, int nFilterIdx = FILTER_DEFAULT,
+ bool useAltHpelIf = false);
#if JVET_J0090_MEMORY_BANDWITH_MEASURE
void cacheAssign( CacheModel *cache ) { m_cacheModel = cache; }
#endif
diff --git a/source/Lib/CommonLib/x86/InterpolationFilterX86.h b/source/Lib/CommonLib/x86/InterpolationFilterX86.h
index d61d1415cb..f721ed9064 100644
--- a/source/Lib/CommonLib/x86/InterpolationFilterX86.h
+++ b/source/Lib/CommonLib/x86/InterpolationFilterX86.h
@@ -2131,27 +2131,31 @@ static void simdFilter( const ClpRng& clpRng, Pel const *src, int srcStride, Pel
}
}
{
- if( N == 8 && !( width & 0x07 ) )
+ if ((N == 8 || N == 6) && !(width & 0x07))
{
if( !isVertical )
{
#if RExt__HIGH_BIT_DEPTH_SUPPORT
if (vext >= AVX2)
{
- simdInterpolateHorM8_HBD_AVX2<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
+ simdInterpolateHorM8_HBD_AVX2<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset,
+ clpRng, c);
}
else
{
- simdInterpolateHorM8_HBD<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
+ simdInterpolateHorM8_HBD<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset,
+ clpRng, c);
}
#else
if( vext>= AVX2 )
{
- simdInterpolateHorM8_AVX2<vext, 8, isLast>( src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c );
+ simdInterpolateHorM8_AVX2<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset,
+ clpRng, c);
}
else
{
- simdInterpolateHorM8<vext, 8, isLast>( src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c );
+ simdInterpolateHorM8<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng,
+ c);
}
#endif
}
@@ -2160,41 +2164,47 @@ static void simdFilter( const ClpRng& clpRng, Pel const *src, int srcStride, Pel
#if RExt__HIGH_BIT_DEPTH_SUPPORT
if (vext >= AVX2)
{
- simdInterpolateVerM8_HBD_AVX2<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
+ simdInterpolateVerM8_HBD_AVX2<vext, N, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset,
+ clpRng, c);
}
else
{
- simdInterpolateVerM8_HBD<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
+ simdInterpolateVerM8_HBD<vext, N, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset,
+ clpRng, c);
}
#else
if( vext>= AVX2 )
{
- simdInterpolateVerM8_AVX2<vext, 8, isLast>( src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c );
+ simdInterpolateVerM8_AVX2<vext, N, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset,
+ clpRng, c);
}
else
{
- simdInterpolateVerM8<vext, 8, isLast>( src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c );
+ simdInterpolateVerM8<vext, N, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng,
+ c);
}
#endif
}
return;
}
- else if( N == 8 && !( width & 0x03 ) )
+ else if ((N == 8 || N == 6) && !(width & 0x03))
{
if( !isVertical )
{
#if RExt__HIGH_BIT_DEPTH_SUPPORT
- simdInterpolateHorM4_HBD<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
+ simdInterpolateHorM4_HBD<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng,
+ c);
#else
- simdInterpolateHorM4<vext, 8, isLast>( src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c );
+ simdInterpolateHorM4<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
#endif
}
else
{
#if RExt__HIGH_BIT_DEPTH_SUPPORT
- simdInterpolateVerM4_HBD<vext, 8, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
+ simdInterpolateVerM4_HBD<vext, N, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng,
+ c);
#else
- simdInterpolateVerM4<vext, 8, isLast>( src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c );
+ simdInterpolateVerM4<vext, N, isLast>(src, srcStride, dst, dstStride, width, height, shift, offset, clpRng, c);
#endif
}
return;
@@ -2585,6 +2595,11 @@ void InterpolationFilter::_initInterpolationFilterX86()
m_filterVer[2][1][0] = simdFilter<vext, 2, true, true, false>;
m_filterVer[2][1][1] = simdFilter<vext, 2, true, true, true>;
+ m_filterVer[3][0][0] = simdFilter<vext, 6, true, false, false>;
+ m_filterVer[3][0][1] = simdFilter<vext, 6, true, false, true>;
+ m_filterVer[3][1][0] = simdFilter<vext, 6, true, true, false>;
+ m_filterVer[3][1][1] = simdFilter<vext, 6, true, true, true>;
+
m_filterCopy[0][0] = simdFilterCopy_HBD<vext, false, false>;
m_filterCopy[0][1] = simdFilterCopy_HBD<vext, false, true>;
m_filterCopy[1][0] = simdFilterCopy_HBD<vext, true, false>;
@@ -2623,6 +2638,11 @@ void InterpolationFilter::_initInterpolationFilterX86()
m_filterVer[2][1][0] = simdFilter<vext, 2, true, true, false>;
m_filterVer[2][1][1] = simdFilter<vext, 2, true, true, true>;
+ m_filterVer[3][0][0] = simdFilter<vext, 6, true, false, false>;
+ m_filterVer[3][0][1] = simdFilter<vext, 6, true, false, true>;
+ m_filterVer[3][1][0] = simdFilter<vext, 6, true, true, false>;
+ m_filterVer[3][1][1] = simdFilter<vext, 6, true, true, true>;
+
m_filterCopy[0][0] = simdFilterCopy<vext, false, false>;
m_filterCopy[0][1] = simdFilterCopy<vext, false, true>;
m_filterCopy[1][0] = simdFilterCopy<vext, true, false>;
diff --git a/source/Lib/EncoderLib/InterSearch.cpp b/source/Lib/EncoderLib/InterSearch.cpp
index 258d33a7b3..e21c55060b 100644
--- a/source/Lib/EncoderLib/InterSearch.cpp
+++ b/source/Lib/EncoderLib/InterSearch.cpp
@@ -9030,17 +9030,17 @@ void InterSearch::xExtDIFUpSamplingH(CPelBuf* pattern, bool useAltHpelIf)
const Pel *srcPtr = pattern->buf - halfFilterSize*srcStride - 1;
m_if.filterHor(COMPONENT_Y, srcPtr, srcStride, m_filteredBlockTmp[0][0], intStride, width + 1, height + filterSize,
- 0 << MV_FRACTIONAL_BITS_DIFF, false, clpRng, 0, false, useAltHpelIf);
+ 0 << MV_FRACTIONAL_BITS_DIFF, false, clpRng, InterpolationFilter::FILTER_DEFAULT, useAltHpelIf);
if (!m_skipFracME)
{
m_if.filterHor(COMPONENT_Y, srcPtr, srcStride, m_filteredBlockTmp[2][0], intStride, width + 1, height + filterSize,
- 2 << MV_FRACTIONAL_BITS_DIFF, false, clpRng, 0, false, useAltHpelIf);
+ 2 << MV_FRACTIONAL_BITS_DIFF, false, clpRng, InterpolationFilter::FILTER_DEFAULT, useAltHpelIf);
}
intPtr = m_filteredBlockTmp[0][0] + halfFilterSize * intStride + 1;
dstPtr = m_filteredBlock[0][0][0];
m_if.filterVer(COMPONENT_Y, intPtr, intStride, dstPtr, dstStride, width + 0, height + 0, 0 << MV_FRACTIONAL_BITS_DIFF,
- false, true, clpRng, 0, false, useAltHpelIf);
+ false, true, clpRng, InterpolationFilter::FILTER_DEFAULT, useAltHpelIf);
if (m_skipFracME)
{
return;
@@ -9049,17 +9049,17 @@ void InterSearch::xExtDIFUpSamplingH(CPelBuf* pattern, bool useAltHpelIf)
intPtr = m_filteredBlockTmp[0][0] + (halfFilterSize - 1) * intStride + 1;
dstPtr = m_filteredBlock[2][0][0];
m_if.filterVer(COMPONENT_Y, intPtr, intStride, dstPtr, dstStride, width + 0, height + 1, 2 << MV_FRACTIONAL_BITS_DIFF,
- false, true, clpRng, 0, false, useAltHpelIf);
+ false, true, clpRng, InterpolationFilter::FILTER_DEFAULT, useAltHpelIf);
intPtr = m_filteredBlockTmp[2][0] + halfFilterSize * intStride;
dstPtr = m_filteredBlock[0][2][0];
m_if.filterVer(COMPONENT_Y, intPtr, intStride, dstPtr, dstStride, width + 1, height + 0, 0 << MV_FRACTIONAL_BITS_DIFF,
- false, true, clpRng, 0, false, useAltHpelIf);
+ false, true, clpRng, InterpolationFilter::FILTER_DEFAULT, useAltHpelIf);
intPtr = m_filteredBlockTmp[2][0] + (halfFilterSize - 1) * intStride;
dstPtr = m_filteredBlock[2][2][0];
m_if.filterVer(COMPONENT_Y, intPtr, intStride, dstPtr, dstStride, width + 1, height + 1, 2 << MV_FRACTIONAL_BITS_DIFF,
- false, true, clpRng, 0, false, useAltHpelIf);
+ false, true, clpRng, InterpolationFilter::FILTER_DEFAULT, useAltHpelIf);
}
--
GitLab