static void partition_mult_data(void)
{
gettimeofday(&start, NULL);
starpu_matrix_data_register(&A_handle, 0, (uintptr_t)A,
ydim, ydim, zdim, sizeof(float));
starpu_matrix_data_register(&B_handle, 0, (uintptr_t)B,
zdim, zdim, xdim, sizeof(float));
starpu_matrix_data_register(&C_handle, 0, (uintptr_t)C,
ydim, ydim, xdim, sizeof(float));
starpu_data_set_wt_mask(C_handle, 1<<0);
conf.k = zdim;
conf.m = ydim/nslicesy;
conf.n = xdim/nslicesx;
struct starpu_data_filter f;
f.filter_func = starpu_vertical_block_filter_func;
f.nchildren = nslicesx;
f.get_nchildren = NULL;
f.get_child_ops = NULL;
struct starpu_data_filter f2;
f2.filter_func = starpu_block_filter_func;
f2.nchildren = nslicesy;
f2.get_nchildren = NULL;
f2.get_child_ops = NULL;
starpu_data_partition(B_handle, &f);
starpu_data_partition(A_handle, &f2);
starpu_data_map_filters(C_handle, 2, &f, &f2);
}
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);
}
int STARPU_LU(lu_decomposition)(TYPE *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(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);
int ret = dw_codelet_facto_v3(dataA, nblocks);
/* gather all the data */
starpu_data_unpartition(dataA, STARPU_MAIN_RAM);
starpu_data_unregister(dataA);
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);
}
int main(int argc, char **argv)
{
int ret;
starpu_init(NULL);
starpu_data_malloc_pinned_if_possible((void **)&buffer, VECTORSIZE);
starpu_vector_data_register(&v_handle, 0, (uintptr_t)buffer, VECTORSIZE, sizeof(char));
struct starpu_data_filter f = {
.filter_func = starpu_vector_divide_in_2_filter_func,
/* there are only 2 children */
.nchildren = 2,
/* the length of the first part */
.filter_arg = VECTORSIZE/2,
.get_nchildren = NULL,
.get_child_ops = NULL
};
unsigned iter;
for (iter = 0; iter < NITER; iter++)
{
starpu_data_map_filters(v_handle, 1, &f);
ret = use_handle(starpu_data_get_sub_data(v_handle, 1, 0));
if (ret == -ENODEV)
goto enodev;
ret = use_handle(starpu_data_get_sub_data(v_handle, 1, 1));
if (ret == -ENODEV)
goto enodev;
starpu_task_wait_for_all();
starpu_data_unpartition(v_handle, 0);
ret = use_handle(v_handle);
if (ret == -ENODEV)
goto enodev;
starpu_task_wait_for_all();
}
starpu_data_unregister(v_handle);
starpu_shutdown();
return 0;
enodev:
fprintf(stderr, "WARNING: No one can execute this task\n");
/* yes, we do not perform the computation but we did detect that no one
* could perform the kernel, so this is not an error from StarPU */
return 0;
}
void dw_factoLU_tag(float *matA, unsigned size, unsigned ld, unsigned nblocks, unsigned _no_prio)
{
#ifdef CHECK_RESULTS
FPRINTF(stderr, "Checking results ...\n");
float *Asaved;
Asaved = malloc((size_t)ld*ld*sizeof(float));
memcpy(Asaved, matA, (size_t)ld*ld*sizeof(float));
#endif
no_prio = _no_prio;
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));
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);
dw_codelet_facto_v3(dataA, nblocks);
/* gather all the data */
starpu_data_unpartition(dataA, STARPU_MAIN_RAM);
starpu_data_unregister(dataA);
#ifdef CHECK_RESULTS
compare_A_LU(Asaved, matA, size, ld);
#endif
}
static void partition_mult_data(void)
{
starpu_matrix_data_register(&A_handle, STARPU_MAIN_RAM, (uintptr_t)A,
ydim, ydim, zdim, sizeof(TYPE));
starpu_matrix_data_register(&B_handle, STARPU_MAIN_RAM, (uintptr_t)B,
zdim, zdim, xdim, sizeof(TYPE));
starpu_matrix_data_register(&C_handle, STARPU_MAIN_RAM, (uintptr_t)C,
ydim, ydim, xdim, sizeof(TYPE));
struct starpu_data_filter vert;
memset(&vert, 0, sizeof(vert));
vert.filter_func = starpu_matrix_filter_vertical_block;
vert.nchildren = nslicesx;
struct starpu_data_filter horiz;
memset(&horiz, 0, sizeof(horiz));
horiz.filter_func = starpu_matrix_filter_block;
horiz.nchildren = nslicesy;
starpu_data_partition(B_handle, &vert);
starpu_data_partition(A_handle, &horiz);
starpu_data_map_filters(C_handle, 2, &vert, &horiz);
}
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;
}
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);
}
}
