int STARPU_LU(lu_decomposition_pivot_no_stride)(TYPE **matA, unsigned *ipiv, unsigned size, unsigned ld, unsigned nblocks)
{
starpu_data_handle_t *dataAp = malloc(nblocks*nblocks*sizeof(starpu_data_handle_t));
/* monitor and partition the A matrix into blocks :
* one block is now determined by 2 unsigned (i,j) */
unsigned bi, bj;
for (bj = 0; bj < nblocks; bj++)
for (bi = 0; bi < nblocks; bi++)
{
starpu_matrix_data_register(&dataAp[bi+nblocks*bj], STARPU_MAIN_RAM,
(uintptr_t)matA[bi+nblocks*bj], size/nblocks,
size/nblocks, size/nblocks, sizeof(TYPE));
}
unsigned i;
for (i = 0; i < size; i++)
ipiv[i] = i;
struct piv_s *piv_description = malloc(nblocks*sizeof(struct piv_s));
unsigned block;
for (block = 0; block < nblocks; block++)
{
piv_description[block].piv = ipiv;
piv_description[block].first = block * (size / nblocks);
piv_description[block].last = (block + 1) * (size / nblocks);
}
double timing;
int ret = dw_codelet_facto_pivot(dataAp, piv_description, nblocks, get_block_with_no_striding, &timing);
if (ret)
return ret;
unsigned n = starpu_matrix_get_nx(dataAp[0])*nblocks;
double flop = (2.0f*n*n*n)/3.0f;
PRINTF("# size\tms\tGFlops");
if (bound)
PRINTF("\tTms\tTGFlops");
PRINTF("\n");
PRINTF("%u\t%.0f\t%.1f", n, timing/1000, flop/timing/1000.0f);
if (bound)
{
double min;
starpu_bound_compute(&min, NULL, 0);
PRINTF("\t%.0f\t%.1f", min, flop/min/1000000.0f);
}
PRINTF("\n");
for (bj = 0; bj < nblocks; bj++)
for (bi = 0; bi < nblocks; bi++)
{
starpu_data_unregister(dataAp[bi+nblocks*bj]);
}
free(dataAp);
return ret;
}
void STARPU_LU(lu_decomposition_pivot)(TYPE *matA, unsigned *ipiv, unsigned size, unsigned ld, unsigned nblocks)
{
starpu_data_handle dataA;
/* monitor and partition the A matrix into blocks :
* one block is now determined by 2 unsigned (i,j) */
starpu_matrix_data_register(&dataA, 0, (uintptr_t)matA, ld, size, size, sizeof(TYPE));
/* We already enforce deps by hand */
starpu_data_set_sequential_consistency_flag(dataA, 0);
struct starpu_data_filter f;
f.filter_func = starpu_vertical_block_filter_func;
f.filter_arg = nblocks;
struct starpu_data_filter f2;
f2.filter_func = starpu_block_filter_func;
f2.filter_arg = nblocks;
starpu_data_map_filters(dataA, 2, &f, &f2);
unsigned i;
for (i = 0; i < size; i++)
ipiv[i] = i;
struct piv_s *piv_description = malloc(nblocks*sizeof(struct piv_s));
unsigned block;
for (block = 0; block < nblocks; block++)
{
piv_description[block].piv = ipiv;
piv_description[block].first = block * (size / nblocks);
piv_description[block].last = (block + 1) * (size / nblocks);
}
#if 0
unsigned j;
for (j = 0; j < nblocks; j++)
for (i = 0; i < nblocks; i++)
{
printf("BLOCK %d %d %p\n", i, j, &matA[i*(size/nblocks) + j * (size/nblocks)*ld]);
}
#endif
double timing;
timing = dw_codelet_facto_pivot(&dataA, piv_description, nblocks, get_block_with_striding);
fprintf(stderr, "Computation took (in ms)\n");
fprintf(stderr, "%2.2f\n", timing/1000);
unsigned n = starpu_matrix_get_nx(dataA);
double flop = (2.0f*n*n*n)/3.0f;
fprintf(stderr, "Synthetic GFlops : %2.2f\n", (flop/timing/1000.0f));
/* gather all the data */
starpu_data_unpartition(dataA, 0);
}
void STARPU_LU(lu_decomposition_pivot_no_stride)(TYPE **matA, unsigned *ipiv, unsigned size, unsigned ld, unsigned nblocks)
{
starpu_data_handle *dataAp = malloc(nblocks*nblocks*sizeof(starpu_data_handle));
/* monitor and partition the A matrix into blocks :
* one block is now determined by 2 unsigned (i,j) */
unsigned bi, bj;
for (bj = 0; bj < nblocks; bj++)
for (bi = 0; bi < nblocks; bi++)
{
starpu_matrix_data_register(&dataAp[bi+nblocks*bj], 0,
(uintptr_t)matA[bi+nblocks*bj], size/nblocks,
size/nblocks, size/nblocks, sizeof(TYPE));
/* We already enforce deps by hand */
starpu_data_set_sequential_consistency_flag(dataAp[bi+nblocks*bj], 0);
}
unsigned i;
for (i = 0; i < size; i++)
ipiv[i] = i;
struct piv_s *piv_description = malloc(nblocks*sizeof(struct piv_s));
unsigned block;
for (block = 0; block < nblocks; block++)
{
piv_description[block].piv = ipiv;
piv_description[block].first = block * (size / nblocks);
piv_description[block].last = (block + 1) * (size / nblocks);
}
double timing;
timing = dw_codelet_facto_pivot(dataAp, piv_description, nblocks, get_block_with_no_striding);
fprintf(stderr, "Computation took (in ms)\n");
fprintf(stderr, "%2.2f\n", timing/1000);
unsigned n = starpu_matrix_get_nx(dataAp[0])*nblocks;
double flop = (2.0f*n*n*n)/3.0f;
fprintf(stderr, "Synthetic GFlops : %2.2f\n", (flop/timing/1000.0f));
for (bj = 0; bj < nblocks; bj++)
for (bi = 0; bi < nblocks; bi++)
{
starpu_data_unregister(dataAp[bi+nblocks*bj]);
}
}
int STARPU_LU(lu_decomposition_pivot)(TYPE *matA, unsigned *ipiv, unsigned size, unsigned ld, unsigned nblocks)
{
starpu_data_handle_t dataA;
/* monitor and partition the A matrix into blocks :
* one block is now determined by 2 unsigned (i,j) */
starpu_matrix_data_register(&dataA, STARPU_MAIN_RAM, (uintptr_t)matA, ld, size, size, sizeof(TYPE));
struct starpu_data_filter f =
{
.filter_func = starpu_matrix_filter_vertical_block,
.nchildren = nblocks
};
struct starpu_data_filter f2 =
{
.filter_func = starpu_matrix_filter_block,
.nchildren = nblocks
};
starpu_data_map_filters(dataA, 2, &f, &f2);
unsigned i;
for (i = 0; i < size; i++)
ipiv[i] = i;
struct piv_s *piv_description = malloc(nblocks*sizeof(struct piv_s));
unsigned block;
for (block = 0; block < nblocks; block++)
{
piv_description[block].piv = ipiv;
piv_description[block].first = block * (size / nblocks);
piv_description[block].last = (block + 1) * (size / nblocks);
}
double timing;
int ret = dw_codelet_facto_pivot(&dataA, piv_description, nblocks, get_block_with_striding, &timing);
if (ret)
return ret;
unsigned n = starpu_matrix_get_nx(dataA);
double flop = (2.0f*n*n*n)/3.0f;
PRINTF("# size\tms\tGFlops");
if (bound)
PRINTF("\tTms\tTGFlops");
PRINTF("\n");
PRINTF("%u\t%.0f\t%.1f", n, timing/1000, flop/timing/1000.0f);
if (bound)
{
double min;
starpu_bound_compute(&min, NULL, 0);
PRINTF("\t%.0f\t%.1f", min, flop/min/1000000.0f);
}
PRINTF("\n");
/* gather all the data */
starpu_data_unpartition(dataA, STARPU_MAIN_RAM);
starpu_data_unregister(dataA);
free(piv_description);
return ret;
}
