void
Stokhos::GramSchmidtBasis<ordinal_type, value_type>::
transformCoeffs(const value_type *in, value_type *out) const
{
Teuchos::BLAS<ordinal_type, value_type> blas;
for (ordinal_type i=0; i<sz; i++)
out[i] = in[i];
blas.TRSM(Teuchos::LEFT_SIDE, Teuchos::LOWER_TRI, Teuchos::TRANS,
Teuchos::UNIT_DIAG, sz, 1, 1.0, gs_mat.values(), sz, out, sz);
}
KOKKOS_INLINE_FUNCTION
int
Trsm<Side::Left,Uplo::Upper,Trans::NoTranspose,
AlgoTrsm::ExternalBlas,Variant::One>
::invoke(PolicyType &policy,
const MemberType &member,
const int diagA,
const ScalarType alpha,
DenseExecViewTypeA &A,
DenseExecViewTypeB &B) {
// static_assert( Kokkos::Impl::is_same<
// typename DenseMatrixTypeA::space_type,
// Kokkos::Cuda
// >::value,
// "Cuda space is not available for calling external BLAS" );
// static_assert( Kokkos::Impl::is_same<
// typename DenseMatrixTypeA::space_type,
// typename DenseMatrixTypeB::space_type
// >::value,
// "Space type of input matrices does not match" );
//typedef typename DenseExecViewTypeA::space_type space_type;
typedef typename DenseExecViewTypeA::ordinal_type ordinal_type;
typedef typename DenseExecViewTypeA::value_type value_type;
if (member.team_rank() == 0) {
#ifdef HAVE_SHYLUTACHO_TEUCHOS
Teuchos::BLAS<ordinal_type,value_type> blas;
const ordinal_type m = A.NumRows();
const ordinal_type n = B.NumCols();
blas.TRSM(Teuchos::LEFT_SIDE, Teuchos::UPPER_TRI, Teuchos::NO_TRANS,
(diagA == Diag::Unit ? Teuchos::UNIT_DIAG : Teuchos::NON_UNIT_DIAG),
m, n,
alpha,
A.ValuePtr(), A.BaseObject().ColStride(),
B.ValuePtr(), B.BaseObject().ColStride());
#else
TACHO_TEST_FOR_ABORT( true, MSG_NOT_HAVE_PACKAGE("Teuchos") );
#endif
}
return 0;
}
KOKKOS_INLINE_FUNCTION
int exampleDenseTrsmMKL(const OrdinalType mmin,
const OrdinalType mmax,
const OrdinalType minc,
const OrdinalType k,
const bool verbose) {
typedef ValueType value_type;
typedef OrdinalType ordinal_type;
typedef SizeType size_type;
typedef DenseMatrixBase<value_type,ordinal_type,size_type,SpaceType,MemoryTraits> DenseMatrixBaseType;
int r_val = 0;
Kokkos::Impl::Timer timer;
double t = 0.0;
cout << "DenseGemmMKL:: test matrices "
<<":: mmin = " << mmin << " , mmax = " << mmax << " , minc = " << minc << " , k = "<< k << endl;
ostringstream os;
os.precision(3);
os << scientific;
for (ordinal_type m=mmin;m<=mmax;m+=minc) {
os.str("");
DenseMatrixBaseType AA("AA", m, m), BB("BB", m, k), BC("BC", m, k);
// setup upper triangular
for (ordinal_type j=0;j<AA.NumCols();++j) {
AA.Value(j,j) = 10.0;
for (ordinal_type i=0;i<j;++i)
AA.Value(i,j) = 2.0*((value_type)rand()/(RAND_MAX)) - 1.0;
}
// setup one and right hand side is going to be overwritten by the product of AB
for (ordinal_type j=0;j<BB.NumCols();++j)
for (ordinal_type i=0;i<BB.NumRows();++i)
BB.Value(i,j) = 1.0;
Teuchos::BLAS<ordinal_type,value_type> blas;
blas.GEMM(Teuchos::CONJ_TRANS, Teuchos::NO_TRANS,
m, k, m,
1.0,
AA.ValuePtr(), AA.ColStride(),
BB.ValuePtr(), BB.ColStride(),
0.0,
BC.ValuePtr(), BC.ColStride());
BB.copy(BC);
const double flop = get_flop_trsm_upper<value_type>(m, k);
os << "DenseTrsmMKL:: m = " << m << " k = " << k;
{
timer.reset();
Teuchos::BLAS<ordinal_type,value_type> blas;
const ordinal_type mm = AA.NumRows();
const ordinal_type nn = BB.NumCols();
blas.TRSM(Teuchos::LEFT_SIDE, Teuchos::UPPER_TRI, Teuchos::CONJ_TRANS,
Teuchos::NON_UNIT_DIAG,
mm, nn,
1.0,
AA.ValuePtr(), AA.ColStride(),
BB.ValuePtr(), BB.ColStride());
t = timer.seconds();
os << ":: MKL Performance = " << (flop/t/1.0e9) << " [GFLOPs] ";
}
cout << os.str() << endl;
}
return r_val;
}