void ilu_transpose(compressed_matrix<NumericT> const & A,
compressed_matrix<NumericT> & B)
{
viennacl::context orig_ctx = viennacl::traits::context(A);
viennacl::context cpu_ctx(viennacl::MAIN_MEMORY);
(void)orig_ctx;
(void)cpu_ctx;
viennacl::compressed_matrix<NumericT> A_host(0, 0, 0, cpu_ctx);
(void)A_host;
switch (viennacl::traits::handle(A).get_active_handle_id())
{
case viennacl::MAIN_MEMORY:
viennacl::linalg::host_based::ilu_transpose(A, B);
break;
#ifdef VIENNACL_WITH_OPENCL
case viennacl::OPENCL_MEMORY:
A_host = A;
B.switch_memory_context(cpu_ctx);
viennacl::linalg::host_based::ilu_transpose(A_host, B);
B.switch_memory_context(orig_ctx);
break;
#endif
#ifdef VIENNACL_WITH_CUDA
case viennacl::CUDA_MEMORY:
A_host = A;
B.switch_memory_context(cpu_ctx);
viennacl::linalg::host_based::ilu_transpose(A_host, B);
B.switch_memory_context(orig_ctx);
break;
#endif
case viennacl::MEMORY_NOT_INITIALIZED:
throw memory_exception("not initialised!");
default:
throw memory_exception("not implemented");
}
}
int exampleDenseGemmByBlocks(const ordinal_type mmin,
const ordinal_type mmax,
const ordinal_type minc,
const ordinal_type k,
const ordinal_type mb,
const int max_concurrency,
const int max_task_dependence,
const int team_size,
const int mkl_nthreads,
const bool check,
const bool verbose) {
typedef typename
Kokkos::Impl::is_space<DeviceSpaceType>::host_mirror_space::execution_space HostSpaceType ;
const bool detail = false;
std::cout << "DeviceSpace:: "; DeviceSpaceType::print_configuration(std::cout, detail);
std::cout << "HostSpace:: "; HostSpaceType::print_configuration(std::cout, detail);
typedef Kokkos::Experimental::TaskPolicy<DeviceSpaceType> PolicyType;
typedef DenseMatrixBase<value_type,ordinal_type,size_type,HostSpaceType> DenseMatrixBaseHostType;
typedef DenseMatrixView<DenseMatrixBaseHostType> DenseMatrixViewHostType;
typedef DenseMatrixBase<value_type,ordinal_type,size_type,DeviceSpaceType> DenseMatrixBaseDeviceType;
typedef DenseMatrixView<DenseMatrixBaseDeviceType> DenseMatrixViewDeviceType;
typedef TaskView<DenseMatrixViewDeviceType> DenseTaskViewDeviceType;
typedef DenseMatrixBase<DenseTaskViewDeviceType,ordinal_type,size_type,DeviceSpaceType> DenseHierMatrixBaseDeviceType;
typedef DenseMatrixView<DenseHierMatrixBaseDeviceType> DenseHierMatrixViewDeviceType;
typedef TaskView<DenseHierMatrixViewDeviceType> DenseHierTaskViewDeviceType;
int r_val = 0;
Kokkos::Impl::Timer timer;
double t = 0.0;
std::cout << "DenseGemmByBlocks:: test matrices "
<<":: mmin = " << mmin << " , mmax = " << mmax << " , minc = " << minc
<< " , k = "<< k << " , mb = " << mb << std::endl;
const size_t max_task_size = (3*sizeof(DenseTaskViewDeviceType)+sizeof(PolicyType)+128);
PolicyType policy(max_concurrency,
max_task_size,
max_task_dependence,
team_size);
std::ostringstream os;
os.precision(3);
os << std::scientific;
for (ordinal_type m=mmin;m<=mmax;m+=minc) {
os.str("");
// host matrices
DenseMatrixBaseHostType AA_host, BB_host, CC_host("CC_host", m, m), CB_host("CB_host", m, m);
{
if (ArgTransA == Trans::NoTranspose)
AA_host = DenseMatrixBaseHostType("AA_host", m, k);
else
AA_host = DenseMatrixBaseHostType("AA_host", k, m);
if (ArgTransB == Trans::NoTranspose)
BB_host = DenseMatrixBaseHostType("BB_host", k, m);
else
BB_host = DenseMatrixBaseHostType("BB_host", m, k);
for (ordinal_type j=0;j<AA_host.NumCols();++j)
for (ordinal_type i=0;i<AA_host.NumRows();++i)
AA_host.Value(i,j) = 2.0*((value_type)rand()/(RAND_MAX)) - 1.0;
for (ordinal_type j=0;j<BB_host.NumCols();++j)
for (ordinal_type i=0;i<BB_host.NumRows();++i)
BB_host.Value(i,j) = 2.0*((value_type)rand()/(RAND_MAX)) - 1.0;
for (ordinal_type j=0;j<CC_host.NumCols();++j)
for (ordinal_type i=0;i<CC_host.NumRows();++i)
CC_host.Value(i,j) = 2.0*((value_type)rand()/(RAND_MAX)) - 1.0;
DenseMatrixTools::copy(CB_host, CC_host);
}
const double flop = DenseFlopCount<value_type>::Gemm(m, m, k);
#ifdef HAVE_SHYLUTACHO_MKL
mkl_set_num_threads(mkl_nthreads);
#endif
os << "DenseGemmByBlocks:: m = " << m << " n = " << m << " k = " << k << " ";
if (check) {
timer.reset();
DenseMatrixViewHostType A_host(AA_host), B_host(BB_host), C_host(CB_host);
Gemm<ArgTransA,ArgTransB,AlgoGemm::ExternalBlas,Variant::One>::invoke
(policy, policy.member_single(),
1.0, A_host, B_host, 1.0, C_host);
t = timer.seconds();
os << ":: Serial Performance = " << (flop/t/1.0e9) << " [GFLOPs] ";
}
DenseMatrixBaseDeviceType AA_device("AA_device"), BB_device("BB_device"), CC_device("CC_device");
{
timer.reset();
AA_device.mirror(AA_host);
BB_device.mirror(BB_host);
CC_device.mirror(CC_host);
t = timer.seconds();
os << ":: Mirror = " << t << " [sec] ";
}
{
DenseHierMatrixBaseDeviceType HA_device("HA_device"), HB_device("HB_device"), HC_device("HC_device");
DenseMatrixTools::createHierMatrix(HA_device, AA_device, mb, mb);
DenseMatrixTools::createHierMatrix(HB_device, BB_device, mb, mb);
DenseMatrixTools::createHierMatrix(HC_device, CC_device, mb, mb);
DenseHierTaskViewDeviceType TA_device(HA_device), TB_device(HB_device), TC_device(HC_device);
timer.reset();
auto future = policy.proc_create_team
(Gemm<ArgTransA,ArgTransB,AlgoGemm::DenseByBlocks,ArgVariant>
::createTaskFunctor(policy, 1.0, TA_device, TB_device, 1.0, TC_device),
0);
policy.spawn(future);
Kokkos::Experimental::wait(policy);
t = timer.seconds();
os << ":: Parallel Performance = " << (flop/t/1.0e9) << " [GFLOPs] ";
}
CC_host.mirror(CC_device);
if (check) {
double err = 0.0, norm = 0.0;
for (ordinal_type j=0;j<CC_host.NumCols();++j)
for (ordinal_type i=0;i<CC_host.NumRows();++i) {
const double diff = abs(CC_host.Value(i,j) - CB_host.Value(i,j));
const double val = CB_host.Value(i,j);
err += diff*diff;
norm += val*val;
}
os << ":: Check result ::norm = " << sqrt(norm) << ", error = " << sqrt(err);
}
std::cout << os.str() << std::endl;
}
return r_val;
}
int exampleDenseCholByBlocks(const ordinal_type mmin,
const ordinal_type mmax,
const ordinal_type minc,
const ordinal_type mb,
const int max_concurrency,
const int max_task_dependence,
const int team_size,
const int mkl_nthreads,
const bool check,
const bool verbose) {
typedef typename
Kokkos::Impl::is_space<DeviceSpaceType>::host_mirror_space::execution_space HostSpaceType ;
const bool detail = false;
std::cout << "DeviceSpace:: "; DeviceSpaceType::print_configuration(std::cout, detail);
std::cout << "HostSpace:: "; HostSpaceType::print_configuration(std::cout, detail);
typedef Kokkos::Experimental::TaskPolicy<DeviceSpaceType> PolicyType;
typedef DenseMatrixBase<value_type,ordinal_type,size_type,HostSpaceType> DenseMatrixBaseHostType;
typedef DenseMatrixView<DenseMatrixBaseHostType> DenseMatrixViewHostType;
typedef DenseMatrixBase<value_type,ordinal_type,size_type,DeviceSpaceType> DenseMatrixBaseDeviceType;
typedef DenseMatrixView<DenseMatrixBaseDeviceType> DenseMatrixViewDeviceType;
typedef TaskView<DenseMatrixViewDeviceType> DenseTaskViewDeviceType;
typedef DenseMatrixBase<DenseTaskViewDeviceType,ordinal_type,size_type,DeviceSpaceType> DenseHierMatrixBaseDeviceType;
typedef DenseMatrixView<DenseHierMatrixBaseDeviceType> DenseHierMatrixViewDeviceType;
typedef TaskView<DenseHierMatrixViewDeviceType> DenseHierTaskViewDeviceType;
int r_val = 0;
Kokkos::Impl::Timer timer;
double t = 0.0;
std::cout << "DenseCholByBlocks:: test matrices "
<<":: mmin = " << mmin << " , mmax = " << mmax << " , minc = " << minc
<< " , mb = " << mb << std::endl;
const size_t max_task_size = (3*sizeof(DenseTaskViewDeviceType)+sizeof(PolicyType)+128);
PolicyType policy(max_concurrency,
max_task_size,
max_task_dependence,
team_size);
std::ostringstream os;
os.precision(3);
os << std::scientific;
for (ordinal_type m=mmin;m<=mmax;m+=minc) {
os.str("");
// host matrices
DenseMatrixBaseHostType AA_host("AA_host", m, m), AB_host("AB_host"), TT_host("TT_host");
// random T matrix
{
TT_host.createConfTo(AA_host);
for (ordinal_type j=0;j<TT_host.NumCols();++j) {
for (ordinal_type i=0;i<TT_host.NumRows();++i)
TT_host.Value(i,j) = 2.0*((value_type)rand()/(RAND_MAX)) - 1.0;
TT_host.Value(j,j) = std::fabs(TT_host.Value(j,j));
}
}
// create SPD matrix
{
Teuchos::BLAS<ordinal_type,value_type> blas;
blas.HERK(ArgUplo == Uplo::Upper ? Teuchos::UPPER_TRI : Teuchos::LOWER_TRI,
Teuchos::CONJ_TRANS,
m, m,
1.0,
TT_host.ValuePtr(), TT_host.ColStride(),
0.0,
AA_host.ValuePtr(), AA_host.ColStride());
// preserve a copy of A
AB_host.createConfTo(AA_host);
DenseMatrixTools::copy(AB_host, AA_host);
}
const double flop = DenseFlopCount<value_type>::Chol(m);
#ifdef HAVE_SHYLUTACHO_MKL
mkl_set_num_threads(mkl_nthreads);
#endif
os << "DenseCholByBlocks:: m = " << m << " ";
int ierr = 0;
if (check) {
timer.reset();
DenseMatrixViewHostType A_host(AB_host);
ierr = Chol<ArgUplo,AlgoChol::ExternalLapack,Variant::One>::invoke
(policy, policy.member_single(),
A_host);
t = timer.seconds();
TACHO_TEST_FOR_ABORT( ierr, "Fail to perform Cholesky (serial)");
os << ":: Serial Performance = " << (flop/t/1.0e9) << " [GFLOPs] ";
}
DenseMatrixBaseDeviceType AA_device("AA_device");
{
timer.reset();
AA_device.mirror(AA_host);
t = timer.seconds();
os << ":: Mirror = " << t << " [sec] ";
}
{
DenseHierMatrixBaseDeviceType HA_device("HA_device");
DenseMatrixTools::createHierMatrix(HA_device, AA_device, mb, mb);
DenseHierTaskViewDeviceType TA_device(HA_device);
timer.reset();
auto future = policy.proc_create_team
(Chol<ArgUplo,AlgoChol::DenseByBlocks,ArgVariant>
::createTaskFunctor(policy, TA_device),
0);
policy.spawn(future);
Kokkos::Experimental::wait(policy);
t = timer.seconds();
os << ":: Parallel Performance = " << (flop/t/1.0e9) << " [GFLOPs] ";
}
AA_host.mirror(AA_device);
if (!ierr && check) {
double err = 0.0, norm = 0.0;
for (ordinal_type j=0;j<AA_host.NumCols();++j)
for (ordinal_type i=0;i<=j;++i) {
const double diff = abs(AA_host.Value(i,j) - AB_host.Value(i,j));
const double val = AB_host.Value(i,j);
err += diff*diff;
norm += val*val;
}
os << ":: Check result ::norm = " << sqrt(norm) << ", error = " << sqrt(err);
}
std::cout << os.str() << std::endl;
}
return r_val;
}
