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);
}
});
int main (int argc, char *argv[])
{
#define amat(I,J) a[I*n + J]
#define bmat(I,J) b[I*n + J]
#define cmat(I,J) c[I*n + J]
int n, nthreads, i, j, k;
double *a, *b, *c;
double t0, t1;
n = 1000;
nthreads = atoi(argv[1]);
omp_set_num_threads (nthreads);
a = (double *) malloc (n * n * sizeof (double));
b = (double *) malloc (n * n * sizeof (double));
c = (double *) malloc (n * n * sizeof (double));
t0 = omp_get_wtime();
#pragma omp parallel for private (i, j, k)
for (j=0; j<n; j++)
for (i=0; i<n; i++)
for (k=0; k<n; k++)
cmat(i,j) = cmat(i,j) + amat(i,k) * bmat(k,j);
t1 = omp_get_wtime();
printf("nthreads, time: %d %6.2f\n", nthreads, t1-t0);
}
/* Used to avoid having really messy code. */
void calc_c(int j2, int k2)
{
//Ck = Ck + Ak*Bjk
int sum;
int x,y,z;
for (x=0; x<n; x++)
{
for (y=0; y<nc; y++)
{
double sum = cmat(x,y);
for (z=0; z<nc; z++)
{
sum = sum + atempmat(x,z) * bmat(z+j2*nc,y);
}
cmat(x,y)=sum;
}
}
}
