void UN_C
( T alpha,
const AbstractDistMatrix<T>& APre,
AbstractDistMatrix<T>& CPre,
bool conjugate=false )
{
EL_DEBUG_CSE
const Int r = APre.Width();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MC, STAR> A1_MC_STAR(g);
DistMatrix<T,VR, STAR> A1_VR_STAR(g);
DistMatrix<T,STAR,MR > A1Trans_STAR_MR(g);
A1_MC_STAR.AlignWith( C );
A1_VR_STAR.AlignWith( C );
A1Trans_STAR_MR.AlignWith( C );
for( Int k=0; k<r; k+=bsize )
{
const Int nb = Min(bsize,r-k);
auto A1 = A( ALL, IR(k,k+nb) );
A1_VR_STAR = A1_MC_STAR = A1;
Transpose( A1_VR_STAR, A1Trans_STAR_MR, conjugate );
LocalTrrk( UPPER, alpha, A1_MC_STAR, A1Trans_STAR_MR, T(1), C );
}
}
void LT_Dot
( T alpha,
const AbstractDistMatrix<T>& APre,
AbstractDistMatrix<T>& CPre,
const bool conjugate,
Int blockSize=2000 )
{
EL_DEBUG_CSE
const Int n = CPre.Height();
const Grid& g = APre.Grid();
const Orientation orient = ( conjugate ? ADJOINT : TRANSPOSE );
DistMatrixReadProxy<T,T,VC,STAR> AProx( APre );
auto& A = AProx.GetLocked();
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& C = CProx.Get();
DistMatrix<T,STAR,STAR> Z( blockSize, blockSize, g );
Zero( Z );
for( Int kOuter=0; kOuter<n; kOuter+=blockSize )
{
const Int nbOuter = Min(blockSize,n-kOuter);
const Range<Int> indOuter( kOuter, kOuter+nbOuter );
auto A1 = A( ALL, indOuter );
auto C11 = C( indOuter, indOuter );
Z.Resize( nbOuter, nbOuter );
Syrk( LOWER, TRANSPOSE, alpha, A1.Matrix(), Z.Matrix(), conjugate );
AxpyContract( T(1), Z, C11 );
for( Int kInner=kOuter+nbOuter; kInner<n; kInner+=blockSize )
{
const Int nbInner = Min(blockSize,n-kInner);
const Range<Int> indInner( kInner, kInner+nbInner );
auto A2 = A( ALL, indInner );
auto C21 = C( indInner, indOuter );
LocalGemm( orient, NORMAL, alpha, A1, A2, Z );
AxpyContract( T(1), Z, C21 );
}
}
}
void SUMMA_NTDot
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre,
Int blockSize=2000 )
{
EL_DEBUG_CSE
const Int m = CPre.Height();
const Int n = CPre.Width();
const Grid& g = APre.Grid();
DistMatrixReadProxy<T,T,STAR,VC> AProx( APre );
auto& A = AProx.GetLocked();
ElementalProxyCtrl BCtrl;
BCtrl.rowConstrain = true;
BCtrl.rowAlign = A.RowAlign();
DistMatrixReadProxy<T,T,STAR,VC> BProx( BPre, BCtrl );
auto& B = BProx.GetLocked();
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& C = CProx.Get();
DistMatrix<T,STAR,STAR> C11_STAR_STAR(g);
for( Int kOuter=0; kOuter<m; kOuter+=blockSize )
{
const Int nbOuter = Min(blockSize,m-kOuter);
const Range<Int> indOuter( kOuter, kOuter+nbOuter );
auto A1 = A( indOuter, ALL );
for( Int kInner=0; kInner<n; kInner+=blockSize )
{
const Int nbInner = Min(blockSize,n-kInner);
const Range<Int> indInner( kInner, kInner+nbInner );
auto B1 = B( indInner, ALL );
auto C11 = C( indOuter, indInner );
LocalGemm( NORMAL, orientB, alpha, A1, B1, C11_STAR_STAR );
AxpyContract( T(1), C11_STAR_STAR, C11 );
}
}
}
void SUMMA_NTC
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
EL_DEBUG_CSE
const Int sumDim = APre.Width();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
const bool conjugate = ( orientB == ADJOINT );
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MC,STAR> A1_MC_STAR(g);
DistMatrix<T,VR,STAR> B1_VR_STAR(g);
DistMatrix<T,STAR,MR> B1Trans_STAR_MR(g);
A1_MC_STAR.AlignWith( C );
B1_VR_STAR.AlignWith( C );
B1Trans_STAR_MR.AlignWith( C );
for( Int k=0; k<sumDim; k+=bsize )
{
const Int nb = Min(bsize,sumDim-k);
auto A1 = A( ALL, IR(k,k+nb) );
auto B1 = B( ALL, IR(k,k+nb) );
A1_MC_STAR = A1;
B1_VR_STAR = B1;
Transpose( B1_VR_STAR, B1Trans_STAR_MR, conjugate );
// C[MC,MR] += alpha A1[MC,*] (B1[MR,*])^T
LocalGemm
( NORMAL, NORMAL, alpha, A1_MC_STAR, B1Trans_STAR_MR, T(1), C );
}
}
void L1DistanceMatrix(direction_t dirA, direction_t dirB, T alpha,
const El::ElementalMatrix<T> &APre, const El::ElementalMatrix<T> &BPre,
T beta, El::ElementalMatrix<T> &CPre) {
if (dirA == base::COLUMNS && dirB == base::COLUMNS) {
// Use a SUMMA-like routine, with C as stationary
// Basically an adaptation of Elementals TN case for stationary C.
const El::Int m = CPre.Height();
const El::Int n = CPre.Width();
const El::Int sumDim = BPre.Height();
const El::Int bsize = El::Blocksize();
const El::Grid& g = APre.Grid();
El::DistMatrixReadProxy<T, T, El::MC, El::MR> AProx(APre);
El::DistMatrixReadProxy<T, T, El::MC, El::MR> BProx(BPre);
El::DistMatrixReadWriteProxy<T, T, El::MC, El::MR> CProx(CPre);
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
El::DistMatrix<T, El::STAR, El::MC> A1_STAR_MC(g);
El::DistMatrix<T, El::STAR, El::MR> B1_STAR_MR(g);
A1_STAR_MC.AlignWith(C);
B1_STAR_MR.AlignWith(C);
El::Scale(beta, C);
for(El::Int k = 0; k < sumDim; k += bsize) {
const El::Int nb = std::min(bsize,sumDim-k);
auto A1 = A(El::IR(k,k+nb), El::IR(0,m));
auto B1 = B(El::IR(k,k+nb), El::IR(0,n));
A1_STAR_MC = A1;
B1_STAR_MR = B1;
L1DistanceMatrix(base::COLUMNS, base::COLUMNS, alpha,
A1_STAR_MC.LockedMatrix(), B1_STAR_MR.LockedMatrix(),
T(1.0), C.Matrix());
}
}
// TODO the rest of the cases.
}
void SUMMA_NTB
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
EL_DEBUG_CSE
const Int m = CPre.Height();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MR,STAR> A1Trans_MR_STAR(g);
DistMatrix<T,STAR,MC> D1_STAR_MC(g);
DistMatrix<T,MR,MC> D1_MR_MC(g);
A1Trans_MR_STAR.AlignWith( B );
D1_STAR_MC.AlignWith( B );
for( Int k=0; k<m; k+=bsize )
{
const Int nb = Min(bsize,m-k);
auto A1 = A( IR(k,k+nb), ALL );
auto C1 = C( IR(k,k+nb), ALL );
// D1[*,MC] := alpha A1[*,MR] (B[MC,MR])^T
// = alpha (A1^T)[MR,*] (B^T)[MR,MC]
Transpose( A1, A1Trans_MR_STAR );
LocalGemm( TRANSPOSE, orientB, alpha, A1Trans_MR_STAR, B, D1_STAR_MC );
// C1[MC,MR] += scattered & transposed D1[*,MC] summed over grid rows
Contract( D1_STAR_MC, D1_MR_MC );
Axpy( T(1), D1_MR_MC, C1 );
}
}
void SUMMA_TNA
( Orientation orientA,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
DEBUG_CSE
const Int n = CPre.Width();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MC,STAR> B1_MC_STAR(g);
DistMatrix<T,MR,STAR> D1_MR_STAR(g);
DistMatrix<T,MR,MC > D1_MR_MC(g);
B1_MC_STAR.AlignWith( A );
D1_MR_STAR.AlignWith( A );
for( Int k=0; k<n; k+=bsize )
{
const Int nb = Min(bsize,n-k);
auto B1 = B( ALL, IR(k,k+nb) );
auto C1 = C( ALL, IR(k,k+nb) );
// D1[MR,*] := alpha (A1[MC,MR])^T B1[MC,*]
// = alpha (A1^T)[MR,MC] B1[MC,*]
B1_MC_STAR = B1;
LocalGemm( orientA, NORMAL, alpha, A, B1_MC_STAR, D1_MR_STAR );
// C1[MC,MR] += scattered & transposed D1[MR,*] summed over grid cols
Contract( D1_MR_STAR, D1_MR_MC );
Axpy( T(1), D1_MR_MC, C1 );
}
}
void SUMMA_NTA
( Orientation orientB,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
EL_DEBUG_CSE
const Int n = CPre.Width();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
const bool conjugate = ( orientB == ADJOINT );
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MR,STAR> B1Trans_MR_STAR(g);
DistMatrix<T,MC,STAR> D1_MC_STAR(g);
B1Trans_MR_STAR.AlignWith( A );
D1_MC_STAR.AlignWith( A );
for( Int k=0; k<n; k+=bsize )
{
const Int nb = Min(bsize,n-k);
auto B1 = B( IR(k,k+nb), ALL );
auto C1 = C( ALL, IR(k,k+nb) );
// C1[MC,*] := alpha A[MC,MR] (B1^[T/H])[MR,*]
Transpose( B1, B1Trans_MR_STAR, conjugate );
LocalGemm( NORMAL, NORMAL, alpha, A, B1Trans_MR_STAR, D1_MC_STAR );
// C1[MC,MR] += scattered result of D1[MC,*] summed over grid rows
AxpyContract( T(1), D1_MC_STAR, C1 );
}
}
void SUMMA_TNB
( Orientation orientA,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
DEBUG_CSE
const Int m = CPre.Height();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
const bool conjugate = ( orientA == ADJOINT );
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MC,STAR> A1_MC_STAR(g);
DistMatrix<T,MR,STAR> D1Trans_MR_STAR(g);
A1_MC_STAR.AlignWith( B );
D1Trans_MR_STAR.AlignWith( B );
for( Int k=0; k<m; k+=bsize )
{
const Int nb = Min(bsize,m-k);
auto A1 = A( ALL, IR(k,k+nb) );
auto C1 = C( IR(k,k+nb), ALL );
// D1[*,MR] := alpha (A1[MC,*])^[T/H] B[MC,MR]
// = alpha (A1^[T/H])[*,MC] B[MC,MR]
A1_MC_STAR = A1; // A1[MC,*] <- A1[MC,MR]
LocalGemm( orientA, NORMAL, T(1), B, A1_MC_STAR, D1Trans_MR_STAR );
TransposeAxpyContract( alpha, D1Trans_MR_STAR, C1, conjugate );
}
}
void SUMMA_TNC
( Orientation orientA,
T alpha,
const AbstractDistMatrix<T>& APre,
const AbstractDistMatrix<T>& BPre,
AbstractDistMatrix<T>& CPre )
{
DEBUG_CSE
const Int sumDim = BPre.Height();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadProxy<T,T,MC,MR> BProx( BPre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& B = BProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,STAR,MC> A1_STAR_MC(g);
DistMatrix<T,MR,STAR> B1Trans_MR_STAR(g);
A1_STAR_MC.AlignWith( C );
B1Trans_MR_STAR.AlignWith( C );
for( Int k=0; k<sumDim; k+=bsize )
{
const Int nb = Min(bsize,sumDim-k);
auto A1 = A( IR(k,k+nb), ALL );
auto B1 = B( IR(k,k+nb), ALL );
// C[MC,MR] += alpha (A1[*,MC])^T B1[*,MR]
// = alpha (A1^T)[MC,*] B1[*,MR]
A1_STAR_MC = A1;
Transpose( B1, B1Trans_MR_STAR );
LocalGemm
( orientA, TRANSPOSE, alpha, A1_STAR_MC, B1Trans_MR_STAR, T(1), C );
}
}
void LT_C
( T alpha,
const AbstractDistMatrix<T>& APre,
AbstractDistMatrix<T>& CPre,
bool conjugate=false )
{
EL_DEBUG_CSE
const Int r = APre.Height();
const Int bsize = Blocksize();
const Grid& g = APre.Grid();
const Orientation orientation = ( conjugate ? ADJOINT : TRANSPOSE );
DistMatrixReadProxy<T,T,MC,MR> AProx( APre );
DistMatrixReadWriteProxy<T,T,MC,MR> CProx( CPre );
auto& A = AProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
DistMatrix<T,MR, STAR> A1Trans_MR_STAR(g);
DistMatrix<T,STAR,VR > A1_STAR_VR(g);
DistMatrix<T,STAR,MC > A1_STAR_MC(g);
A1Trans_MR_STAR.AlignWith( C );
A1_STAR_MC.AlignWith( C );
for( Int k=0; k<r; k+=bsize )
{
const Int nb = Min(bsize,r-k);
auto A1 = A( IR(k,k+nb), ALL );
Transpose( A1, A1Trans_MR_STAR );
Transpose( A1Trans_MR_STAR, A1_STAR_VR );
A1_STAR_MC = A1_STAR_VR;
LocalTrrk
( LOWER, orientation, TRANSPOSE,
alpha, A1_STAR_MC, A1Trans_MR_STAR, T(1), C );
}
}
void SymmetricL1DistanceMatrix(El::UpperOrLower uplo, direction_t dir, T alpha,
const El::ElementalMatrix<T> &APre, T beta, El::ElementalMatrix<T> &CPre) {
if (dir == base::COLUMNS) {
const El::Int r = APre.Height();
const El::Int bsize = El::Blocksize();
const El::Grid& g = APre.Grid();
El::DistMatrixReadProxy<T, T, El::MC, El::MR> AProx(APre);
El::DistMatrixReadWriteProxy<T, T, El::MC, El::MR> CProx(CPre);
auto& A = AProx.GetLocked();
auto& C = CProx.Get();
// Temporary distributions
El::DistMatrix<T, El::STAR, El::MR> A1_STAR_MR(g);
El::DistMatrix<T, El::STAR, El::MC> A1_STAR_MC(g);
A1_STAR_MC.AlignWith(C);
A1_STAR_MR.AlignWith(C);
El::ScaleTrapezoid(beta, uplo, C);
for(El::Int k = 0; k < r; k += bsize) {
const El::Int nb = std::min(bsize, r - k);
auto A1 = A(El::IR(k, k + nb), El::ALL);
A1_STAR_MC = A1;
A1_STAR_MR = A1;
internal::L1DistanceMatrixTU(uplo, base::COLUMNS, base::COLUMNS,
alpha, A1_STAR_MC, A1_STAR_MR,
T(1.0), C);
}
}
// TODO the rest of the cases.
}
