From 6bef7f0d35e82ac1bc53387e4b79b45d7dba5fb5 Mon Sep 17 00:00:00 2001
From: Nan Hu <nanh@qti.qualcomm.com>
Date: Thu, 19 Oct 2023 09:35:37 -0700
Subject: [PATCH] =?UTF-8?q?JVET=5FAF0177=5FALF=5FCOV=5FFLOAT=EF=BC=9A=20ch?=
=?UTF-8?q?ange=20data=20type=20of=20ALF/CCALF=20covariance=20from=20doubl?=
=?UTF-8?q?e=20to=20float?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
---
source/Lib/CommonLib/TypeDef.h | 1 +
.../Lib/EncoderLib/EncAdaptiveLoopFilter.cpp | 171 +++++++++++++++++-
source/Lib/EncoderLib/EncAdaptiveLoopFilter.h | 69 ++++++-
3 files changed, 230 insertions(+), 11 deletions(-)
diff --git a/source/Lib/CommonLib/TypeDef.h b/source/Lib/CommonLib/TypeDef.h
index ed1323c8e..84cf787f5 100644
--- a/source/Lib/CommonLib/TypeDef.h
+++ b/source/Lib/CommonLib/TypeDef.h
@@ -66,6 +66,7 @@
#define REMOVE_PCM 1 // Remove PCM related code for memory reduction and speedup
#define JVET_Y0152_TT_ENC_SPEEDUP 1 // TT encoding speedup
#define JVET_AE0057_MTT_ET 1 // JVET-AE0057: MTT early termination of 64x64 luma CU based on no-split RD cost
+#define JVET_AF0177_ALF_COV_FLOAT 1 // JVET-AF0177: Use float for ALF and CCALF covariance matrix
// Software optimization
diff --git a/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp b/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp
index 111f94122..0f43b3cc0 100644
--- a/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp
+++ b/source/Lib/EncoderLib/EncAdaptiveLoopFilter.cpp
@@ -102,15 +102,24 @@ void AlfCovariance::reduceClipCost(const AlfFilterShape& alfShape, int *clip) co
}
#if JVET_AD0222_ALF_RESI_CLASS
+#if JVET_AF0177_ALF_COV_FLOAT
+double AlfCovariance::optimizeFilter(const AlfFilterShape& alfShape, int* clip, Ty f, bool optimize_clip, bool enableLessClip) const
+#else
double AlfCovariance::optimizeFilter(const AlfFilterShape& alfShape, int* clip, double *f, bool optimize_clip, bool enableLessClip) const
+#endif
#else
double AlfCovariance::optimizeFilter(const AlfFilterShape& alfShape, int* clip, double *f, bool optimize_clip) const
#endif
{
const int size = alfShape.numCoeff;
int clip_max[MAX_NUM_ALF_LUMA_COEFF];
+#if JVET_AF0177_ALF_COV_FLOAT
+ Ty fBest;
+ int copySize = sizeof( float ) * size;
+#else
double fBest[MAX_NUM_ALF_LUMA_COEFF];
int copySize = sizeof( double ) * size;
+#endif
double err_best, err_last;
TE kE;
@@ -305,7 +314,11 @@ double AlfCovariance::calcErrorForCcAlfCoeffs(const int16_t *coeff, const int nu
return error / factor;
}
+#if JVET_AF0177_ALF_COV_FLOAT
+double AlfCovariance::calculateError( const int *clip, const Ty coeff, const int numCoeff ) const
+#else
double AlfCovariance::calculateError( const int *clip, const double *coeff, const int numCoeff ) const
+#endif
{
double sum = 0;
for( int i = 0; i < numCoeff; i++ )
@@ -327,17 +340,28 @@ double AlfCovariance::calculateError( const int *clip ) const
//********************************
#define ROUND(a) (((a) < 0)? (int)((a) - 0.5) : (int)((a) + 0.5))
+#if JVET_AF0177_ALF_COV_FLOAT
+#define REG ((float)0.0001)
+#define REG_SQR ((float)0.0000001)
+#else
#define REG 0.0001
#define REG_SQR 0.0000001
+#endif
//Find filter coeff related
int AlfCovariance::gnsCholeskyDec( TE inpMatr, TE outMatr, int numEq ) const
{
for( int i = 0; i < numEq; i++ )
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ float* inputM = inpMatr[i];
+ float* outputM = outMatr[i];
+ float scale = inputM[i];
+#else
double* inputM = inpMatr[i];
double* outputM = outMatr[i];
double scale = inputM[i];
+#endif
for( int k = i - 1; k >= 0; k-- )
{
@@ -350,7 +374,11 @@ int AlfCovariance::gnsCholeskyDec( TE inpMatr, TE outMatr, int numEq ) const
}
outputM[i] = sqrt( scale );
+#if JVET_AF0177_ALF_COV_FLOAT
+ float tmp = 1 / outputM[i];
+#else
double tmp = 1 / outputM[i];
+#endif
for( int j = i + 1; j < numEq; j++ )
{
@@ -368,14 +396,22 @@ int AlfCovariance::gnsCholeskyDec( TE inpMatr, TE outMatr, int numEq ) const
return 1; // Signal that Cholesky factorization is successfully performed
}
+#if JVET_AF0177_ALF_COV_FLOAT
+void AlfCovariance::gnsTransposeBacksubstitution( TE U, Ty rhs, Ty x, int order ) const
+#else
void AlfCovariance::gnsTransposeBacksubstitution( TE U, double* rhs, double* x, int order ) const
+#endif
{
/* Backsubstitution starts */
x[0] = rhs[0] / U[0][0]; /* First row of U' */
for( int i = 1; i < order; i++ )
{ /* For the rows 1..order-1 */
+#if JVET_AF0177_ALF_COV_FLOAT
+ float sum = 0; //Holds backsubstitution from already handled rows
+#else
double sum = 0; //Holds backsubstitution from already handled rows
+#endif
for( int j = 0; j < i; j++ ) /* Backsubst already solved unknowns */
{
@@ -386,14 +422,22 @@ void AlfCovariance::gnsTransposeBacksubstitution( TE U, double* rhs, double* x,
}
}
+#if JVET_AF0177_ALF_COV_FLOAT
+void AlfCovariance::gnsBacksubstitution( TE R, Ty z, int size, Ty A ) const
+#else
void AlfCovariance::gnsBacksubstitution( TE R, double* z, int size, double* A ) const
+#endif
{
size--;
A[size] = z[size] / R[size][size];
for( int i = size - 1; i >= 0; i-- )
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ float sum = 0;
+#else
double sum = 0;
+#endif
for( int j = i + 1; j <= size; j++ )
{
@@ -404,7 +448,11 @@ void AlfCovariance::gnsBacksubstitution( TE R, double* z, int size, double* A )
}
}
+#if JVET_AF0177_ALF_COV_FLOAT
+int AlfCovariance::gnsSolveByChol( const int *clip, Ty x, int numEq ) const
+#else
int AlfCovariance::gnsSolveByChol( const int *clip, double *x, int numEq ) const
+#endif
{
TE LHS;
Ty rhs;
@@ -413,7 +461,11 @@ int AlfCovariance::gnsSolveByChol( const int *clip, double *x, int numEq ) const
return gnsSolveByChol( LHS, rhs, x, numEq );
}
+#if JVET_AF0177_ALF_COV_FLOAT
+int AlfCovariance::gnsSolveByChol( TE LHS, Ty rhs, Ty x, int numEq ) const
+#else
int AlfCovariance::gnsSolveByChol( TE LHS, double* rhs, double *x, int numEq ) const
+#endif
{
Ty aux; /* Auxiliary vector */
TE U; /* Upper triangular Cholesky factor of LHS */
@@ -450,7 +502,11 @@ int AlfCovariance::gnsSolveByChol( TE LHS, double* rhs, double *x, int numEq ) c
if( !res )
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ std::memset( x, 0, sizeof( float )*numEq );
+#else
std::memset( x, 0, sizeof( double )*numEq );
+#endif
return 0;
}
@@ -3007,7 +3063,11 @@ double EncAdaptiveLoopFilter::deriveCoeffQuant( int *filterClipp, int *filterCoe
const int min_value = -factor + 1;
const int numCoeff = shape.numCoeff;
+#if JVET_AF0177_ALF_COV_FLOAT
+ float filterCoeff[MAX_NUM_ALF_LUMA_COEFF];
+#else
double filterCoeff[MAX_NUM_ALF_LUMA_COEFF];
+#endif
#if JVET_AD0222_ALF_RESI_CLASS
cov.optimizeFilter( shape, filterClipp, filterCoeff, optimizeClip, m_enableLessClip );
@@ -3062,7 +3122,11 @@ double EncAdaptiveLoopFilter::deriveCoeffQuant( int *filterClipp, int *filterCoe
return errRef;
}
+#if JVET_AF0177_ALF_COV_FLOAT
+void EncAdaptiveLoopFilter::roundFiltCoeff( int *filterCoeffQuant, float *filterCoeff, const int numCoeff, const int factor )
+#else
void EncAdaptiveLoopFilter::roundFiltCoeff( int *filterCoeffQuant, double *filterCoeff, const int numCoeff, const int factor )
+#endif
{
for( int i = 0; i < numCoeff; i++ )
{
@@ -3071,8 +3135,11 @@ void EncAdaptiveLoopFilter::roundFiltCoeff( int *filterCoeffQuant, double *filte
}
}
-void EncAdaptiveLoopFilter::roundFiltCoeffCCALF(int16_t *filterCoeffQuant, double *filterCoeff, const int numCoeff,
- const int factor)
+#if JVET_AF0177_ALF_COV_FLOAT
+void EncAdaptiveLoopFilter::roundFiltCoeffCCALF(int16_t *filterCoeffQuant, float *filterCoeff, const int numCoeff, const int factor)
+#else
+void EncAdaptiveLoopFilter::roundFiltCoeffCCALF(int16_t *filterCoeffQuant, double *filterCoeff, const int numCoeff, const int factor)
+#endif
{
for( int i = 0; i < numCoeff; i++ )
{
@@ -3946,7 +4013,11 @@ void EncAdaptiveLoopFilter::getBlkStats( AlfCovariance* alfCovariance, const Alf
int classIdx = 0;
#if JVET_AD0222_ALF_RESI_CLASS
int classIdxNext = 0;
+#if JVET_AF0177_ALF_COV_FLOAT
+ float curY, curP;
+#else
double curY, curP;
+#endif
#endif
for( int i = 0; i < area.height; i++ )
@@ -3983,10 +4054,18 @@ void EncAdaptiveLoopFilter::getBlkStats( AlfCovariance* alfCovariance, const Alf
CHECK(!alfCovariance[classIdx].sameSizeAs(alfCovarianceNext[classIdxNext]), "Covariance size mismatch");
}
#endif
+#if JVET_AF0177_ALF_COV_FLOAT
+ float weight = 1.0;
+#else
double weight = 1.0;
+#endif
if( m_alfWSSD )
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ weight = (float)m_lumaLevelToWeightPLUT[org[j]];
+#else
weight = m_lumaLevelToWeightPLUT[org[j]];
+#endif
}
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
Intermediate_Int yLocal = org[j] - rec[j];
@@ -4047,7 +4126,11 @@ calcCovariance( ELocal, rec + j, recStride, recBeforeDb + j, recBeforeDbStride,
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
#if JVET_AD0222_ALF_RESI_CLASS
const ptrdiff_t oe = alfCovariance[classIdx].getOffsetEfast(b0, b1, k, l);
+#if JVET_AF0177_ALF_COV_FLOAT
+ const float curE = weight * (ELocal[k][b0] * (float)ELocal[l][b1]);
+#else
const double curE = weight * (ELocal[k][b0] * (double) ELocal[l][b1]);
+#endif
alfCovariance[classIdx].data[oe] += curE;
if (reuse)
{
@@ -4065,7 +4148,11 @@ calcCovariance( ELocal, rec + j, recStride, recBeforeDb + j, recBeforeDbStride,
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
#if JVET_AD0222_ALF_RESI_CLASS
const ptrdiff_t oe = alfCovariance[classIdx].getOffsetEfast(b0, b1, k, l);
+#if JVET_AF0177_ALF_COV_FLOAT
+ const float curE = ELocal[k][b0] * (float)ELocal[l][b1];
+#else
const double curE = ELocal[k][b0] * (double) ELocal[l][b1];
+#endif
alfCovariance[classIdx].data[oe] += curE;
if (reuse)
{
@@ -4095,13 +4182,21 @@ calcCovariance( ELocal, rec + j, recStride, recBeforeDb + j, recBeforeDbStride,
const ptrdiff_t oy = alfCovariance[classIdx].getOffsetY(b, k);
if (reuse)
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ curY = weight * (ELocal[k][b] * (float)yLocal);
+#else
curY = weight * (ELocal[k][b] * (double)yLocal);
+#endif
alfCovariance[classIdx].data[oy] += curY;
alfCovarianceNext[classIdxNext].data[oy] += curY;
}
else
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance[classIdx].data[oy] += weight * (ELocal[k][b] * (float)yLocal);
+#else
alfCovariance[classIdx].data[oy] += weight * (ELocal[k][b] * (double) yLocal);
+#endif
}
#else
alfCovariance[classIdx].y[b][k] += weight * (ELocal[k][b] * (double)yLocal);
@@ -4117,13 +4212,21 @@ calcCovariance( ELocal, rec + j, recStride, recBeforeDb + j, recBeforeDbStride,
const ptrdiff_t oy = alfCovariance[classIdx].getOffsetY(b, k);
if( reuse )
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ curY = ELocal[k][b] * (float)yLocal;
+#else
curY = ELocal[k][b] * (double)yLocal;
+#endif
alfCovariance[classIdx].data[oy] += curY;
alfCovarianceNext[classIdxNext].data[oy] += curY;
}
else
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance[classIdx].data[oy] += ELocal[k][b] * (float)yLocal;
+#else
alfCovariance[classIdx].data[oy] += ELocal[k][b] * (double) yLocal;
+#endif
}
#else
alfCovariance[classIdx].y[b][k] += ELocal[k][b] * (double)yLocal;
@@ -4140,13 +4243,21 @@ calcCovariance( ELocal, rec + j, recStride, recBeforeDb + j, recBeforeDbStride,
#if JVET_AD0222_ALF_RESI_CLASS
if( reuse )
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ curP = weight * (yLocal * (float)yLocal);
+#else
curP = weight * (yLocal * (double)yLocal);
+#endif
alfCovariance[classIdx].pixAcc += curP;
alfCovarianceNext[classIdxNext].pixAcc += curP;
}
else
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance[classIdx].pixAcc += weight * (yLocal * (float)yLocal);
+#else
alfCovariance[classIdx].pixAcc += weight * (yLocal * (double)yLocal);
+#endif
}
#else
alfCovariance[classIdx].pixAcc += weight * (yLocal * (double)yLocal);
@@ -4161,13 +4272,21 @@ calcCovariance( ELocal, rec + j, recStride, recBeforeDb + j, recBeforeDbStride,
#if JVET_AD0222_ALF_RESI_CLASS
if( reuse )
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ curP = yLocal * (float)yLocal;
+#else
curP = yLocal * (double)yLocal;
+#endif
alfCovariance[classIdx].pixAcc += curP;
alfCovarianceNext[classIdxNext].pixAcc += curP;
}
else
{
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance[classIdx].pixAcc += yLocal * (float)yLocal;
+#else
alfCovariance[classIdx].pixAcc += yLocal * (double)yLocal;
+#endif
}
#else
alfCovariance[classIdx].pixAcc += yLocal * (double)yLocal;
@@ -6675,14 +6794,27 @@ void EncAdaptiveLoopFilter::deriveCcAlfFilterCoeff( ComponentID compID, const Pe
}
#if JVET_X0071_LONGER_CCALF && !JVET_AA0095_ALF_WITH_SAMPLES_BEFORE_DBF && !JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS && !JVET_AC0162_ALF_RESIDUAL_SAMPLES_INPUT
+#if JVET_AF0177_ALF_COV_FLOAT
+ using TE = float[MAX_NUM_CC_ALF_CHROMA_COEFF][MAX_NUM_CC_ALF_CHROMA_COEFF];
+ using Ty = float[MAX_NUM_CC_ALF_CHROMA_COEFF];
+#else
using TE = double[MAX_NUM_CC_ALF_CHROMA_COEFF][MAX_NUM_CC_ALF_CHROMA_COEFF];
using Ty = double[MAX_NUM_CC_ALF_CHROMA_COEFF];
+#endif
+#else
+#if JVET_AF0177_ALF_COV_FLOAT
+ using TE = float[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF];
+ using Ty = float[MAX_NUM_ALF_LUMA_COEFF];
#else
using TE = double[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF];
using Ty = double[MAX_NUM_ALF_LUMA_COEFF];
#endif
-
+#endif
+#if JVET_AF0177_ALF_COV_FLOAT
+ Ty filterCoeffDbl;
+#else
double filterCoeffDbl[MAX_NUM_CC_ALF_CHROMA_COEFF];
+#endif
int16_t filterCoeffInt[MAX_NUM_CC_ALF_CHROMA_COEFF];
std::fill_n(filterCoeffInt, MAX_NUM_CC_ALF_CHROMA_COEFF, 0);
@@ -7490,12 +7622,19 @@ void EncAdaptiveLoopFilter::getBlkStatsCcAlf(AlfCovariance &alfCovariance, const
#endif
{
std::memset(ELocal, 0, sizeof(ELocal));
-
+#if JVET_AF0177_ALF_COV_FLOAT
+ float weight = 1.0;
+ if (m_alfWSSD)
+ {
+ weight = (float)m_lumaLevelToWeightPLUT[org[j]];
+ }
+#else
double weight = 1.0;
if (m_alfWSSD)
{
weight = m_lumaLevelToWeightPLUT[org[j]];
}
+#endif
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
Intermediate_Int yLocal = org[j] - rec[compID][j];
@@ -7524,7 +7663,11 @@ void EncAdaptiveLoopFilter::getBlkStatsCcAlf(AlfCovariance &alfCovariance, const
if (m_alfWSSD)
{
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance.E(b0, b1, k, l) += weight * (ELocal[k][b0] * (float)ELocal[l][b1]);
+#else
alfCovariance.E(b0,b1,k,l) += weight * (ELocal[k][b0] * (double)ELocal[l][b1]);
+#endif
#else
alfCovariance.E(b0,b1,k,l) += weight * (double) (ELocal[k][b0] * ELocal[l][b1]);
#endif
@@ -7532,7 +7675,11 @@ void EncAdaptiveLoopFilter::getBlkStatsCcAlf(AlfCovariance &alfCovariance, const
else
{
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance.E(b0, b1, k, l) += ELocal[k][b0] * (float)ELocal[l][b1];
+#else
alfCovariance.E(b0,b1,k,l) += ELocal[k][b0] * (double)ELocal[l][b1];
+#endif
#else
alfCovariance.E(b0,b1,k,l) += ELocal[k][b0] * ELocal[l][b1];
#endif
@@ -7545,7 +7692,11 @@ void EncAdaptiveLoopFilter::getBlkStatsCcAlf(AlfCovariance &alfCovariance, const
if (m_alfWSSD)
{
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance.y(b,k) += weight * (ELocal[k][b] * (float)yLocal);
+#else
alfCovariance.y(b,k) += weight * (ELocal[k][b] * (double)yLocal);
+#endif
#else
alfCovariance.y[b][k] += weight * (double) (ELocal[k][b] * yLocal);
#endif
@@ -7553,7 +7704,11 @@ void EncAdaptiveLoopFilter::getBlkStatsCcAlf(AlfCovariance &alfCovariance, const
else
{
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance.y(b,k) += ELocal[k][b] * (float)yLocal;
+#else
alfCovariance.y(b,k) += ELocal[k][b] * (double)yLocal;
+#endif
#else
alfCovariance.y[b][k] += ELocal[k][b] * yLocal;
#endif
@@ -7563,7 +7718,11 @@ void EncAdaptiveLoopFilter::getBlkStatsCcAlf(AlfCovariance &alfCovariance, const
if (m_alfWSSD)
{
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance.pixAcc += weight * (yLocal * (float)yLocal);
+#else
alfCovariance.pixAcc += weight * (yLocal * (double)yLocal);
+#endif
#else
alfCovariance.pixAcc += weight * (double) (yLocal * yLocal);
#endif
@@ -7571,7 +7730,11 @@ void EncAdaptiveLoopFilter::getBlkStatsCcAlf(AlfCovariance &alfCovariance, const
else
{
#if JVET_R0351_HIGH_BIT_DEPTH_SUPPORT
+#if JVET_AF0177_ALF_COV_FLOAT
+ alfCovariance.pixAcc += yLocal * (float)yLocal;
+#else
alfCovariance.pixAcc += yLocal * (double)yLocal;
+#endif
#else
alfCovariance.pixAcc += yLocal * yLocal;
#endif
diff --git a/source/Lib/EncoderLib/EncAdaptiveLoopFilter.h b/source/Lib/EncoderLib/EncAdaptiveLoopFilter.h
index 73548903a..ad7d650be 100644
--- a/source/Lib/EncoderLib/EncAdaptiveLoopFilter.h
+++ b/source/Lib/EncoderLib/EncAdaptiveLoopFilter.h
@@ -49,15 +49,29 @@ struct AlfCovariance
static constexpr int MaxAlfNumClippingValues = AdaptiveLoopFilter::MaxAlfNumClippingValues;
#if JVET_X0071_LONGER_CCALF && !JVET_AA0095_ALF_WITH_SAMPLES_BEFORE_DBF && !JVET_AB0184_ALF_MORE_FIXED_FILTER_OUTPUT_TAPS
+#if JVET_AF0177_ALF_COV_FLOAT
+ using TE = float[MAX_NUM_CC_ALF_CHROMA_COEFF][MAX_NUM_CC_ALF_CHROMA_COEFF];
+ using Ty = float[MAX_NUM_CC_ALF_CHROMA_COEFF];
+#else
using TE = double[MAX_NUM_CC_ALF_CHROMA_COEFF][MAX_NUM_CC_ALF_CHROMA_COEFF];
using Ty = double[MAX_NUM_CC_ALF_CHROMA_COEFF];
+#endif
+#else
+#if JVET_AF0177_ALF_COV_FLOAT
+ using TE = float[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF];
+ using Ty = float[MAX_NUM_ALF_LUMA_COEFF];
+ using DTy = double[MAX_NUM_ALF_LUMA_COEFF];
#else
using TE = double[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF];
using Ty = double[MAX_NUM_ALF_LUMA_COEFF];
#endif
+#endif
-
+#if JVET_AF0177_ALF_COV_FLOAT
+ std::vector<float> data;
+#else
std::vector<double> data;
+#endif
ptrdiff_t offsetE;
ptrdiff_t stridey;
@@ -84,16 +98,26 @@ struct AlfCovariance
return offsetE + (v1 <= v0 ? (v0 * (v0 + 1) >> 1) + v1 : (v1 * (v1 + 1) >> 1) + v0);
}
+#if JVET_AF0177_ALF_COV_FLOAT
+ float &y(ptrdiff_t i, ptrdiff_t j) { return data[getOffsetY(i, j)]; }
+ const float &y(ptrdiff_t i, ptrdiff_t j) const { return data[getOffsetY(i, j)]; }
+ float &E(ptrdiff_t i, ptrdiff_t j, ptrdiff_t k, ptrdiff_t l) { return data[getOffsetE(i, j, k, l)]; }
+ const float &E(ptrdiff_t i, ptrdiff_t j, ptrdiff_t k, ptrdiff_t l) const { return data[getOffsetE(i, j, k, l)]; }
+#else
double &y(ptrdiff_t i, ptrdiff_t j) { return data[getOffsetY(i, j)]; }
const double &y(ptrdiff_t i, ptrdiff_t j) const { return data[getOffsetY(i, j)]; }
double &E(ptrdiff_t i, ptrdiff_t j, ptrdiff_t k, ptrdiff_t l) { return data[getOffsetE(i, j, k, l)]; }
const double &E(ptrdiff_t i, ptrdiff_t j, ptrdiff_t k, ptrdiff_t l) const { return data[getOffsetE(i, j, k, l)]; }
-
+#endif
int numCoeff;
int numBins;
+#if JVET_AF0177_ALF_COV_FLOAT
+ float pixAcc;
+#else
double pixAcc;
+#endif
AlfCovariance() {}
~AlfCovariance() {}
@@ -106,7 +130,11 @@ struct AlfCovariance
stridey = numCoeff;
offsetE = stridey * numBins;
data.resize(offsetE + (numCols * (numCols + 1) >> 1));
+#if JVET_AF0177_ALF_COV_FLOAT
+ std::fill(data.begin(), data.end(), (float)0.0);
+#else
std::fill(data.begin(), data.end(), 0.0);
+#endif
}
void destroy()
@@ -116,7 +144,11 @@ struct AlfCovariance
void reset()
{
pixAcc = 0;
+#if JVET_AF0177_ALF_COV_FLOAT
+ std::fill(data.begin(), data.end(), (float)0.0);
+#else
std::fill(data.begin(), data.end(), 0.0);
+#endif
}
void reset(int _numBins)
@@ -236,15 +268,22 @@ struct AlfCovariance
}
}
}
-
+#if JVET_AF0177_ALF_COV_FLOAT
+ double optimizeFilter(const int* clip, Ty f, int size) const
+#else
double optimizeFilter(const int* clip, double *f, int size) const
+#endif
{
gnsSolveByChol( clip, f, size );
return calculateError( clip, f );
}
#if JVET_AD0222_ALF_RESI_CLASS
+#if JVET_AF0177_ALF_COV_FLOAT
+ double optimizeFilter(const AlfFilterShape& alfShape, int* clip, Ty f, bool optimize_clip, bool enableLessClip) const;
+#else
double optimizeFilter(const AlfFilterShape& alfShape, int* clip, double *f, bool optimize_clip, bool enableLessClip) const;
+#endif
double optimizeFilterClip(const AlfFilterShape& alfShape, int* clip, bool enableLessClip) const
{
Ty f;
@@ -260,22 +299,34 @@ struct AlfCovariance
#endif
double calculateError( const int *clip ) const;
+#if JVET_AF0177_ALF_COV_FLOAT
+ double calculateError( const int *clip, const Ty coeff ) const { return calculateError(clip, coeff, numCoeff); }
+ double calculateError( const int *clip, const Ty coeff, const int numCoeff ) const;
+#else
double calculateError( const int *clip, const double *coeff ) const { return calculateError(clip, coeff, numCoeff); }
double calculateError( const int *clip, const double *coeff, const int numCoeff ) const;
+#endif
double calcErrorForCoeffs( const int *clip, const int *coeff, const int numCoeff, const int bitDepth ) const;
double calcErrorForCcAlfCoeffs(const int16_t *coeff, const int numCoeff, const int bitDepth) const;
void getClipMax(const AlfFilterShape& alfShape, int *clip_max) const;
void reduceClipCost(const AlfFilterShape& alfShape, int *clip) const;
-
+#if JVET_AF0177_ALF_COV_FLOAT
+ int gnsSolveByChol( TE LHS, Ty rhs, Ty x, int numEq ) const;
+#else
int gnsSolveByChol( TE LHS, double* rhs, double *x, int numEq ) const;
-
+#endif
private:
// Cholesky decomposition
-
+#if JVET_AF0177_ALF_COV_FLOAT
+ int gnsSolveByChol( const int *clip, Ty x, int numEq ) const;
+ void gnsTransposeBacksubstitution( TE U, Ty rhs, Ty x, int order ) const;
+ void gnsBacksubstitution( TE R, Ty z, int size, Ty A ) const;
+#else
int gnsSolveByChol( const int *clip, double *x, int numEq ) const;
void gnsBacksubstitution( TE R, double* z, int size, double* A ) const;
void gnsTransposeBacksubstitution( TE U, double* rhs, double* x, int order ) const;
+#endif
int gnsCholeskyDec( TE inpMatr, TE outMatr, int numEq ) const;
};
@@ -569,9 +620,13 @@ private:
, int fixedFilterSetIdx
#endif
);
+#if JVET_AF0177_ALF_COV_FLOAT
+ void roundFiltCoeff( int *filterCoeffQuant, float *filterCoeff, const int numCoeff, const int factor );
+ void roundFiltCoeffCCALF( int16_t *filterCoeffQuant, float *filterCoeff, const int numCoeff, const int factor );
+#else
void roundFiltCoeff( int *filterCoeffQuant, double *filterCoeff, const int numCoeff, const int factor );
void roundFiltCoeffCCALF(int16_t *filterCoeffQuant, double *filterCoeff, const int numCoeff, const int factor);
-
+#endif
double getDistCoeffForce0( bool* codedVarBins, double errorForce0CoeffTab[MAX_NUM_ALF_CLASSES][2], int* bitsVarBin, int zeroBitsVarBin, const int numFilters);
int lengthUvlc( int uiCode );
#if ALF_IMPROVEMENT
--
GitLab