void Gemm(const int NN) {
typedef Kokkos::Schedule<Kokkos::Static> ScheduleType;
constexpr int VectorLength = DefaultVectorLength<value_type,typename HostSpaceType::memory_space>::value;
const int N = NN/VectorLength;
{
std::string value_type_name;
if (std::is_same<value_type,double>::value) value_type_name = "double";
if (std::is_same<value_type,Kokkos::complex<double> >::value) value_type_name = "Kokkos::complex<double>";
#if defined(__AVX512F__)
std::cout << "AVX512 is defined: datatype " << value_type_name << " a vector length " << VectorLength << "\n";
#elif defined(__AVX__) || defined(__AVX2__)
std::cout << "AVX or AVX2 is defined: datatype " << value_type_name << " a vector length " << VectorLength << "\n";
#else
std::cout << "SIMD (compiler vectorization) is defined: datatype " << value_type_name << " a vector length " << VectorLength << "\n";
#endif
}
const double flop = (N*VectorLength)*FlopCount(BlkSize,BlkSize,BlkSize);
const double tmax = 1.0e15;
const int iter_begin = -10, iter_end = 100;
Kokkos::Impl::Timer timer;
Kokkos::View<value_type***,Kokkos::LayoutRight,HostSpaceType> cref;
Kokkos::View<value_type***,Kokkos::LayoutRight,HostSpaceType>
amat("amat", N*VectorLength, BlkSize, BlkSize),
bmat("bmat", N*VectorLength, BlkSize, BlkSize);
Kokkos::Random_XorShift64_Pool<HostSpaceType> random(13718);
Kokkos::fill_random(amat, random, value_type(1.0));
Kokkos::fill_random(bmat, random, value_type(1.0));
typedef Vector<SIMD<value_type>,VectorLength> VectorType;
Kokkos::View<VectorType***,Kokkos::LayoutRight,HostSpaceType>
amat_simd("amat_simd", N, BlkSize, BlkSize),
bmat_simd("bmat_simd", N, BlkSize, BlkSize);
Kokkos::parallel_for
(Kokkos::RangePolicy<HostSpaceType>(0, N*VectorLength),
KOKKOS_LAMBDA(const int k) {
const int k0 = k/VectorLength, k1 = k%VectorLength;
for (int i=0;i<BlkSize;++i)
for (int j=0;j<BlkSize;++j) {
amat_simd(k0, i, j)[k1] = amat(k, i, j);
bmat_simd(k0, i, j)[k1] = bmat(k, i, j);
}
});
void LU(const int NN) {
typedef Kokkos::Schedule<Kokkos::Static> ScheduleType;
//typedef Kokkos::Schedule<Kokkos::Dynamic> ScheduleType;
constexpr int VectorLength = DefaultVectorLength<value_type,typename HostSpaceType::memory_space>::value;
const int N = NN/VectorLength;
{
std::string value_type_name;
if (std::is_same<value_type,double>::value) value_type_name = "double";
if (std::is_same<value_type,Kokkos::complex<double> >::value) value_type_name = "Kokkos::complex<double>";
#if defined(__AVX512F__)
std::cout << "AVX512 is defined: datatype " << value_type_name << " a vector length " << VectorLength << "\n";
#elif defined(__AVX__) || defined(__AVX2__)
std::cout << "AVX or AVX2 is defined: datatype " << value_type_name << " a vector length " << VectorLength << "\n";
#else
std::cout << "SIMD (compiler vectorization) is defined: datatype " << value_type_name << " a vector length " << VectorLength << "\n";
#endif
}
const double flop = (N*VectorLength)*FlopCount(BlkSize,BlkSize);
const double tmax = 1.0e15;
const int iter_begin = -10, iter_end = 100;
Kokkos::Impl::Timer timer;
///
/// Reference version using MKL DGETRF
///
Kokkos::View<value_type***,Kokkos::LayoutRight,HostSpaceType> aref;
Kokkos::View<value_type***,Kokkos::LayoutRight,HostSpaceType>
amat("amat", N*VectorLength, BlkSize, BlkSize);
Random<value_type> random;
for (int k=0;k<N*VectorLength;++k) {
// use tridiagonal matrices; for now we just check elementwise l/u factors
// do not allow pivots
for (int i=0;i<BlkSize;++i) {
amat(k, i, i) = random.value() + 10.0;
if ((i+1) < BlkSize) {
amat(k, i, i+1) = random.value() + 1.0;
amat(k, i+1, i) = random.value() + 1.0;
}
}
}
typedef Vector<SIMD<value_type>,VectorLength> VectorType;
Kokkos::View<VectorType***,Kokkos::LayoutRight,HostSpaceType>
amat_simd("amat_simd", N, BlkSize, BlkSize); //, a("a", N, BlkSize, BlkSize);
Kokkos::parallel_for("KokkosBatched::PerfTest::LUHost::Pack",
Kokkos::RangePolicy<HostSpaceType>(0, N*VectorLength),
KOKKOS_LAMBDA(const int k) {
const int k0 = k/VectorLength, k1 = k%VectorLength;
for (int i=0;i<BlkSize;++i)
for (int j=0;j<BlkSize;++j) {
amat_simd(k0, i, j)[k1] = amat(k0*VectorLength+k1, i, j);
}
});
