static struct starpu_task * create_task_11(starpu_data_handle_t dataA, unsigned k)
{
/* FPRINTF(stdout, "task 11 k = %d TAG = %llx\n", k, (TAG11(k))); */
struct starpu_task *task = create_task(TAG11(k));
task->cl = &cl11;
/* which sub-data is manipulated ? */
task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
/* this is an important task */
if (!noprio)
task->priority = STARPU_MAX_PRIO;
/* enforce dependencies ... */
if (k > 0)
{
starpu_tag_declare_deps(TAG11(k), 1, TAG22(k-1, k, k));
}
int n = starpu_matrix_get_nx(task->handles[0]);
task->flops = FLOPS_SPOTRF(n);
return task;
}
static void create_task_12(starpu_data_handle_t dataA, unsigned k, unsigned i)
{
int ret;
/* printf("task 12 k,i = %d,%d TAG = %llx\n", k,i, TAG12(k,i)); */
struct starpu_task *task = create_task(TAG12(k, i));
task->cl = &cl12;
/* which sub-data is manipulated ? */
task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
task->handles[1] = starpu_data_get_sub_data(dataA, 2, i, k);
if (!no_prio && (i == k+1))
{
task->priority = STARPU_MAX_PRIO;
}
/* enforce dependencies ... */
if (k > 0)
{
starpu_tag_declare_deps(TAG12(k, i), 2, TAG11(k), TAG22(k-1, i, k));
}
else
{
starpu_tag_declare_deps(TAG12(k, i), 1, TAG11(k));
}
ret = starpu_task_submit(task);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
static void create_task_21(starpu_data_handle_t dataA, unsigned k, unsigned j)
{
int ret;
struct starpu_task *task = create_task(TAG21(k, j));
task->cl = &cl21;
/* which sub-data is manipulated ? */
task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
task->handles[1] = starpu_data_get_sub_data(dataA, 2, k, j);
if (!no_prio && (j == k+1))
{
task->priority = STARPU_MAX_PRIO;
}
/* enforce dependencies ... */
if (k > 0)
{
starpu_tag_declare_deps(TAG21(k, j), 2, TAG11(k), TAG22(k-1, k, j));
}
else
{
starpu_tag_declare_deps(TAG21(k, j), 1, TAG11(k));
}
ret = starpu_task_submit(task);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
static int create_task_11_pivot(starpu_data_handle_t *dataAp, unsigned nblocks,
unsigned k, struct piv_s *piv_description,
starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned))
{
int ret;
struct starpu_task *task = starpu_task_create();
task->cl = &cl11_pivot;
task->cl_arg = &piv_description[k];
/* which sub-data is manipulated ? */
task->handles[0] = get_block(dataAp, nblocks, k, k);
task->tag_id = TAG11(k);
/* this is an important task */
if (!no_prio)
task->priority = STARPU_MAX_PRIO;
ret = starpu_task_submit(task);
if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
return ret;
}
static void create_task_21(starpu_data_handle_t dataA, unsigned k, unsigned j)
{
struct starpu_task *task = create_task(TAG21(k, j));
task->cl = &cl21;
/* which sub-data is manipulated ? */
task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
task->handles[1] = starpu_data_get_sub_data(dataA, 2, k, j);
if (!noprio && (j == k+1))
{
task->priority = STARPU_MAX_PRIO;
}
/* enforce dependencies ... */
if (k > 0)
{
starpu_tag_declare_deps(TAG21(k, j), 2, TAG11(k), TAG22(k-1, k, j));
}
else
{
starpu_tag_declare_deps(TAG21(k, j), 1, TAG11(k));
}
int n = starpu_matrix_get_nx(task->handles[0]);
task->flops = FLOPS_STRSM(n, n);
int ret = starpu_task_submit(task);
if (STARPU_UNLIKELY(ret == -ENODEV))
{
FPRINTF(stderr, "No worker may execute this task\n");
exit(0);
}
}
static int create_task_11(starpu_data_handle_t dataA, unsigned k)
{
int ret;
struct starpu_task *task = starpu_task_create();
task->cl = &cl11;
/* which sub-data is manipulated ? */
task->handles[0] = starpu_data_get_sub_data(dataA, 2, k, k);
task->tag_id = TAG11(k);
/* this is an important task */
if (!no_prio)
task->priority = STARPU_MAX_PRIO;
ret = starpu_task_submit(task);
if (ret != -ENODEV) STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
return ret;
}
static void _cholesky(starpu_data_handle_t dataA, unsigned nblocks)
{
double start;
double end;
struct starpu_task *entry_task = NULL;
/* create all the DAG nodes */
unsigned i,j,k;
start = starpu_timing_now();
for (k = 0; k < nblocks; k++)
{
struct starpu_task *task = create_task_11(dataA, k);
/* we defer the launch of the first task */
if (k == 0)
{
entry_task = task;
}
else
{
int ret = starpu_task_submit(task);
if (STARPU_UNLIKELY(ret == -ENODEV))
{
FPRINTF(stderr, "No worker may execute this task\n");
exit(0);
}
}
for (j = k+1; j<nblocks; j++)
{
create_task_21(dataA, k, j);
for (i = k+1; i<nblocks; i++)
{
if (i <= j)
create_task_22(dataA, k, i, j);
}
}
}
/* schedule the codelet */
int ret = starpu_task_submit(entry_task);
if (STARPU_UNLIKELY(ret == -ENODEV))
{
FPRINTF(stderr, "No worker may execute this task\n");
exit(0);
}
/* stall the application until the end of computations */
starpu_tag_wait(TAG11(nblocks-1));
starpu_data_unpartition(dataA, STARPU_MAIN_RAM);
end = starpu_timing_now();
double timing = end - start;
unsigned n = starpu_matrix_get_nx(dataA);
double flop = (1.0f*n*n*n)/3.0f;
PRINTF("# size\tms\tGFlops\n");
PRINTF("%u\t%.0f\t%.1f\n", n, timing/1000, (flop/timing/1000.0f));
}
