static void cholesky(float *matA, 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(float));
starpu_data_set_sequential_consistency_flag(dataA, 0);
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);
_cholesky(dataA, nblocks);
starpu_data_unregister(dataA);
}
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]);
}
}
void starpu_data_set_sequential_consistency_flag(starpu_data_handle handle, unsigned flag)
{
_starpu_spin_lock(&handle->header_lock);
unsigned child;
for (child = 0; child < handle->nchildren; child++)
{
/* make sure that the flags are applied to the children as well */
struct starpu_data_state_t *child_handle = &handle->children[child];
if (child_handle->nchildren > 0)
starpu_data_set_sequential_consistency_flag(child_handle, flag);
}
PTHREAD_MUTEX_LOCK(&handle->sequential_consistency_mutex);
handle->sequential_consistency = flag;
PTHREAD_MUTEX_UNLOCK(&handle->sequential_consistency_mutex);
_starpu_spin_unlock(&handle->header_lock);
}
static int cholesky_grain_rec(float *matA, unsigned size, unsigned ld, unsigned nblocks, unsigned nbigblocks, unsigned reclevel)
{
int ret;
/* create a new codelet */
struct starpu_task *entry_task = NULL;
/* create all the DAG nodes */
unsigned i,j,k;
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(float));
starpu_data_set_sequential_consistency_flag(dataA, 0);
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);
for (k = 0; k < nbigblocks; k++)
{
struct starpu_task *task = create_task_11(dataA, k, reclevel);
/* we defer the launch of the first task */
if (k == 0)
{
entry_task = task;
}
else
{
ret = starpu_task_submit(task);
if (ret == -ENODEV) return 77;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
for (j = k+1; j<nblocks; j++)
{
ret = create_task_21(dataA, k, j, reclevel);
if (ret == -ENODEV) return 77;
for (i = k+1; i<nblocks; i++)
{
if (i <= j)
{
ret = create_task_22(dataA, k, i, j, reclevel);
if (ret == -ENODEV) return 77;
}
}
}
}
/* schedule the codelet */
ret = starpu_task_submit(entry_task);
if (STARPU_UNLIKELY(ret == -ENODEV))
{
FPRINTF(stderr, "No worker may execute this task\n");
return 77;
}
if (nblocks == nbigblocks)
{
/* stall the application until the end of computations */
starpu_tag_wait(TAG11_AUX(nblocks-1, reclevel));
starpu_data_unpartition(dataA, STARPU_MAIN_RAM);
starpu_data_unregister(dataA);
return 0;
}
else
{
STARPU_ASSERT(reclevel == 0);
unsigned ndeps_tags = (nblocks - nbigblocks)*(nblocks - nbigblocks);
starpu_tag_t *tag_array = malloc(ndeps_tags*sizeof(starpu_tag_t));
STARPU_ASSERT(tag_array);
unsigned ind = 0;
for (i = nbigblocks; i < nblocks; i++)
for (j = nbigblocks; j < nblocks; j++)
{
if (i <= j)
tag_array[ind++] = TAG22_AUX(nbigblocks - 1, i, j, reclevel);
}
starpu_tag_wait_array(ind, tag_array);
free(tag_array);
starpu_data_unpartition(dataA, STARPU_MAIN_RAM);
starpu_data_unregister(dataA);
float *newmatA = &matA[nbigblocks*(size/nblocks)*(ld+1)];
return cholesky_grain_rec(newmatA, size/nblocks*(nblocks - nbigblocks), ld, (nblocks - nbigblocks)*2, (nblocks - nbigblocks)*2, reclevel+1);
}
}
