void unregister_data(void)
{
starpu_data_unpartition(sparse_matrix, STARPU_MAIN_RAM);
starpu_data_unregister(sparse_matrix);
starpu_data_unpartition(vector_in, STARPU_MAIN_RAM);
starpu_data_unregister(vector_in);
starpu_data_unpartition(vector_out, STARPU_MAIN_RAM);
starpu_data_unregister(vector_out);
}
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 terminate(void)
{
fprintf(stderr, "unpartition !!\n");
starpu_data_unpartition(C_handle, 0);
starpu_data_unregister(C_handle);
gettimeofday(&end, NULL);
double timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
display_stats(timing);
#ifdef CHECK_OUTPUT
/* check results */
/* compute C = C - AB */
SGEMM("N", "N", ydim, xdim, zdim, -1.0f, A, ydim, B, zdim, 1.0f, C, ydim);
/* make sure C = 0 */
float err;
err = SASUM(xdim*ydim, C, 1);
if (err < xdim*ydim*0.001) {
fprintf(stderr, "Results are OK\n");
}
else {
fprintf(stderr, "There were errors ... err = %f\n", err);
}
#endif // CHECK_OUTPUT
}
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 _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));
}
void starpu_data_unpartition(starpu_data_handle_t root_handle, unsigned gathering_node)
{
unsigned child;
unsigned worker;
unsigned nworkers = starpu_worker_get_count();
unsigned node;
unsigned sizes[root_handle->nchildren];
_STARPU_TRACE_START_UNPARTITION(root_handle, gathering_node);
_starpu_spin_lock(&root_handle->header_lock);
STARPU_ASSERT_MSG(root_handle->nchildren != 0, "data %p is not partitioned, can not unpartition it", root_handle);
/* first take all the children lock (in order !) */
for (child = 0; child < root_handle->nchildren; child++)
{
starpu_data_handle_t child_handle = starpu_data_get_child(root_handle, child);
/* make sure the intermediate children is unpartitionned as well */
if (child_handle->nchildren > 0)
starpu_data_unpartition(child_handle, gathering_node);
/* If this is a multiformat handle, we must convert the data now */
#ifdef STARPU_DEVEL
#warning TODO: _starpu_fetch_data_on_node should be doing it
#endif
if (_starpu_data_is_multiformat_handle(child_handle) &&
starpu_node_get_kind(child_handle->mf_node) != STARPU_CPU_RAM)
{
struct starpu_codelet cl =
{
.where = STARPU_CPU,
.cpu_funcs = { _starpu_empty_codelet_function },
.modes = { STARPU_RW },
.nbuffers = 1
};
struct starpu_task *task = starpu_task_create();
task->name = "convert_data";
STARPU_TASK_SET_HANDLE(task, child_handle, 0);
task->cl = &cl;
task->synchronous = 1;
if (_starpu_task_submit_internally(task) != 0)
_STARPU_ERROR("Could not submit the conversion task while unpartitionning\n");
}
int ret;
/* for now we pretend that the RAM is almost unlimited and that gathering
* data should be possible from the node that does the unpartionning ... we
* don't want to have the programming deal with memory shortage at that time,
* really */
/* Acquire the child data on the gathering node. This will trigger collapsing any reduction */
ret = starpu_data_acquire_on_node(child_handle, gathering_node, STARPU_RW);
STARPU_ASSERT(ret == 0);
starpu_data_release_on_node(child_handle, gathering_node);
_starpu_spin_lock(&child_handle->header_lock);
child_handle->busy_waiting = 1;
_starpu_spin_unlock(&child_handle->header_lock);
/* Wait for all requests to finish (notably WT requests) */
STARPU_PTHREAD_MUTEX_LOCK(&child_handle->busy_mutex);
while (1)
{
/* Here helgrind would shout that this an unprotected access,
* but this is actually fine: all threads who do busy_count--
* are supposed to call _starpu_data_check_not_busy, which will
* wake us up through the busy_mutex/busy_cond. */
if (!child_handle->busy_count)
break;
/* This is woken by _starpu_data_check_not_busy, always called
* after decrementing busy_count */
STARPU_PTHREAD_COND_WAIT(&child_handle->busy_cond, &child_handle->busy_mutex);
}
STARPU_PTHREAD_MUTEX_UNLOCK(&child_handle->busy_mutex);
_starpu_spin_lock(&child_handle->header_lock);
sizes[child] = _starpu_data_get_size(child_handle);
_starpu_data_unregister_ram_pointer(child_handle);
for (worker = 0; worker < nworkers; worker++)
{
struct _starpu_data_replicate *local = &child_handle->per_worker[worker];
STARPU_ASSERT(local->state == STARPU_INVALID);
if (local->allocated && local->automatically_allocated)
_starpu_request_mem_chunk_removal(child_handle, local, starpu_worker_get_memory_node(worker), sizes[child]);
}
_starpu_memory_stats_free(child_handle);
}
int main(int argc, char **argv)
{
unsigned *foo;
starpu_data_handle_t handle;
int ret;
unsigned n, i, size;
ret = starpu_initialize(NULL, &argc, &argv);
if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
#ifdef STARPU_USE_OPENCL
ret = starpu_opencl_load_opencl_from_file("tests/datawizard/scal_opencl.cl", &opencl_program, NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file");
#endif
n = starpu_worker_get_count();
if (n == 1)
{
starpu_shutdown();
return STARPU_TEST_SKIPPED;
}
size = 10 * n;
foo = (unsigned *) calloc(size, sizeof(*foo));
for (i = 0; i < size; i++)
foo[i] = i;
starpu_vector_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)foo, size, sizeof(*foo));
/* Broadcast the data to force in-place partitioning */
for (i = 0; i < n; i++)
starpu_data_prefetch_on_node(handle, starpu_worker_get_memory_node(i), 0);
struct starpu_data_filter f =
{
.filter_func = starpu_vector_filter_block,
.nchildren = n,
};
starpu_data_partition(handle, &f);
for (i = 0; i < f.nchildren; i++)
{
struct starpu_task *task = starpu_task_create();
task->handles[0] = starpu_data_get_sub_data(handle, 1, i);
task->cl = &scal_codelet;
task->execute_on_a_specific_worker = 1;
task->workerid = i;
ret = starpu_task_submit(task);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
ret = starpu_task_wait_for_all();
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_wait_for_all");
starpu_data_unpartition(handle, STARPU_MAIN_RAM);
starpu_data_unregister(handle);
starpu_shutdown();
ret = EXIT_SUCCESS;
for (i = 0; i < size; i++)
{
if (foo[i] != i*2)
{
FPRINTF(stderr,"value %u is %u instead of %u\n", i, foo[i], 2*i);
ret = EXIT_FAILURE;
}
}
return ret;
enodev:
starpu_data_unregister(handle);
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 */
starpu_shutdown();
return STARPU_TEST_SKIPPED;
}
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;
}
int main(int argc, char **argv)
{
int ret, exit_value = 0;
/* Initialize StarPU */
ret = starpu_init(NULL);
if (ret == -ENODEV)
return 77;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
#ifdef STARPU_USE_OPENCL
ret = starpu_opencl_load_opencl_from_file("examples/axpy/axpy_opencl_kernel.cl",
&opencl_program, NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file");
#endif
starpu_cublas_init();
/* This is equivalent to
vec_a = malloc(N*sizeof(TYPE));
vec_b = malloc(N*sizeof(TYPE));
*/
starpu_malloc((void **)&_vec_x, N*sizeof(TYPE));
assert(_vec_x);
starpu_malloc((void **)&_vec_y, N*sizeof(TYPE));
assert(_vec_y);
unsigned i;
for (i = 0; i < N; i++)
{
_vec_x[i] = 1.0f; /*(TYPE)starpu_drand48(); */
_vec_y[i] = 4.0f; /*(TYPE)starpu_drand48(); */
}
FPRINTF(stderr, "BEFORE x[0] = %2.2f\n", _vec_x[0]);
FPRINTF(stderr, "BEFORE y[0] = %2.2f\n", _vec_y[0]);
/* Declare the data to StarPU */
starpu_vector_data_register(&_handle_x, STARPU_MAIN_RAM, (uintptr_t)_vec_x, N, sizeof(TYPE));
starpu_vector_data_register(&_handle_y, STARPU_MAIN_RAM, (uintptr_t)_vec_y, N, sizeof(TYPE));
/* Divide the vector into blocks */
struct starpu_data_filter block_filter =
{
.filter_func = starpu_vector_filter_block,
.nchildren = NBLOCKS
};
starpu_data_partition(_handle_x, &block_filter);
starpu_data_partition(_handle_y, &block_filter);
double start;
double end;
start = starpu_timing_now();
unsigned b;
for (b = 0; b < NBLOCKS; b++)
{
struct starpu_task *task = starpu_task_create();
task->cl = &axpy_cl;
task->cl_arg = &_alpha;
task->cl_arg_size = sizeof(_alpha);
task->handles[0] = starpu_data_get_sub_data(_handle_x, 1, b);
task->handles[1] = starpu_data_get_sub_data(_handle_y, 1, b);
task->tag_id = b;
ret = starpu_task_submit(task);
if (ret == -ENODEV)
{
exit_value = 77;
goto enodev;
}
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
starpu_task_wait_for_all();
enodev:
starpu_data_unpartition(_handle_x, STARPU_MAIN_RAM);
starpu_data_unpartition(_handle_y, STARPU_MAIN_RAM);
starpu_data_unregister(_handle_x);
starpu_data_unregister(_handle_y);
end = starpu_timing_now();
double timing = end - start;
FPRINTF(stderr, "timing -> %2.2f us %2.2f MB/s\n", timing, 3*N*sizeof(TYPE)/timing);
FPRINTF(stderr, "AFTER y[0] = %2.2f (ALPHA = %2.2f)\n", _vec_y[0], _alpha);
if (exit_value != 77)
exit_value = check();
starpu_free((void *)_vec_x);
starpu_free((void *)_vec_y);
#ifdef STARPU_USE_OPENCL
ret = starpu_opencl_unload_opencl(&opencl_program);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_unload_opencl");
#endif
/* Stop StarPU */
starpu_shutdown();
return exit_value;
}
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);
}
}
int main(int argc, char **argv)
{
double start, end;
int ret;
parse_args(argc, argv);
#ifdef STARPU_QUICK_CHECK
niter /= 10;
#endif
ret = starpu_init(NULL);
if (ret == -ENODEV)
return 77;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
starpu_cublas_init();
init_problem_data();
partition_mult_data();
if (bound)
starpu_bound_start(0, 0);
start = starpu_timing_now();
unsigned x, y, iter;
for (iter = 0; iter < niter; iter++)
{
for (x = 0; x < nslicesx; x++)
for (y = 0; y < nslicesy; y++)
{
struct starpu_task *task = starpu_task_create();
task->cl = &cl;
task->handles[0] = starpu_data_get_sub_data(A_handle, 1, y);
task->handles[1] = starpu_data_get_sub_data(B_handle, 1, x);
task->handles[2] = starpu_data_get_sub_data(C_handle, 2, x, y);
task->flops = 2ULL * (xdim/nslicesx) * (ydim/nslicesy) * zdim;
ret = starpu_task_submit(task);
if (ret == -ENODEV)
{
ret = 77;
goto enodev;
}
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
starpu_task_wait_for_all();
}
end = starpu_timing_now();
if (bound)
starpu_bound_stop();
double timing = end - start;
double min, min_int;
double flops = 2.0*((unsigned long long)niter)*((unsigned long long)xdim)
*((unsigned long long)ydim)*((unsigned long long)zdim);
if (bound)
starpu_bound_compute(&min, &min_int, 1);
PRINTF("# x\ty\tz\tms\tGFlops");
if (bound)
PRINTF("\tTms\tTGFlops\tTims\tTiGFlops");
PRINTF("\n");
PRINTF("%u\t%u\t%u\t%.0f\t%.1f", xdim, ydim, zdim, timing/niter/1000.0, flops/timing/1000.0);
if (bound)
PRINTF("\t%.0f\t%.1f\t%.0f\t%.1f", min, flops/min/1000000.0, min_int, flops/min_int/1000000.0);
PRINTF("\n");
enodev:
starpu_data_unpartition(C_handle, STARPU_MAIN_RAM);
starpu_data_unpartition(B_handle, STARPU_MAIN_RAM);
starpu_data_unpartition(A_handle, STARPU_MAIN_RAM);
starpu_data_unregister(A_handle);
starpu_data_unregister(B_handle);
starpu_data_unregister(C_handle);
if (check)
check_output();
starpu_free(A);
starpu_free(B);
starpu_free(C);
starpu_cublas_shutdown();
starpu_shutdown();
return ret;
}
int main(int argc, char **argv)
{
int ret;
unsigned part;
double timing;
double start, end;
unsigned row, pos;
unsigned ind;
/* CSR matrix description */
float *nzval;
uint32_t nnz;
uint32_t *colind;
uint32_t *rowptr;
/* Input and Output vectors */
float *vector_in_ptr;
float *vector_out_ptr;
/*
* Parse command-line arguments
*/
parse_args(argc, argv);
/*
* Launch StarPU
*/
ret = starpu_init(NULL);
if (ret == -ENODEV)
return 77;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
/*
* Create a 3-band sparse matrix as input example
*/
nnz = 3*size-2;
starpu_malloc((void **)&nzval, nnz*sizeof(float));
starpu_malloc((void **)&colind, nnz*sizeof(uint32_t));
starpu_malloc((void **)&rowptr, (size+1)*sizeof(uint32_t));
assert(nzval && colind && rowptr);
/* fill the matrix */
for (row = 0, pos = 0; row < size; row++)
{
rowptr[row] = pos;
if (row > 0)
{
nzval[pos] = 1.0f;
colind[pos] = row-1;
pos++;
}
nzval[pos] = 5.0f;
colind[pos] = row;
pos++;
if (row < size - 1)
{
nzval[pos] = 1.0f;
colind[pos] = row+1;
pos++;
}
}
STARPU_ASSERT(pos == nnz);
rowptr[size] = nnz;
/* initiate the 2 vectors */
starpu_malloc((void **)&vector_in_ptr, size*sizeof(float));
starpu_malloc((void **)&vector_out_ptr, size*sizeof(float));
assert(vector_in_ptr && vector_out_ptr);
/* fill them */
for (ind = 0; ind < size; ind++)
{
vector_in_ptr[ind] = 2.0f;
vector_out_ptr[ind] = 0.0f;
}
/*
* Register the CSR matrix and the 2 vectors
*/
starpu_csr_data_register(&sparse_matrix, STARPU_MAIN_RAM, nnz, size, (uintptr_t)nzval, colind, rowptr, 0, sizeof(float));
starpu_vector_data_register(&vector_in, STARPU_MAIN_RAM, (uintptr_t)vector_in_ptr, size, sizeof(float));
starpu_vector_data_register(&vector_out, STARPU_MAIN_RAM, (uintptr_t)vector_out_ptr, size, sizeof(float));
/*
* Partition the CSR matrix and the output vector
*/
csr_f.nchildren = nblocks;
vector_f.nchildren = nblocks;
starpu_data_partition(sparse_matrix, &csr_f);
starpu_data_partition(vector_out, &vector_f);
/*
* If we use OpenCL, we need to compile the SpMV kernel
*/
#ifdef STARPU_USE_OPENCL
compile_spmv_opencl_kernel();
#endif
start = starpu_timing_now();
/*
* Create and submit StarPU tasks
*/
for (part = 0; part < nblocks; part++)
{
struct starpu_task *task = starpu_task_create();
task->cl = &spmv_cl;
task->handles[0] = starpu_data_get_sub_data(sparse_matrix, 1, part);
task->handles[1] = vector_in;
task->handles[2] = starpu_data_get_sub_data(vector_out, 1, part);
ret = starpu_task_submit(task);
if (STARPU_UNLIKELY(ret == -ENODEV))
{
FPRINTF(stderr, "No worker may execute this task\n");
exit(0);
}
}
starpu_task_wait_for_all();
end = starpu_timing_now();
/*
* Unregister the CSR matrix and the output vector
*/
starpu_data_unpartition(sparse_matrix, STARPU_MAIN_RAM);
starpu_data_unpartition(vector_out, STARPU_MAIN_RAM);
/*
* Unregister data
*/
starpu_data_unregister(sparse_matrix);
starpu_data_unregister(vector_in);
starpu_data_unregister(vector_out);
/*
* Display the result
*/
for (row = 0; row < STARPU_MIN(size, 16); row++)
{
FPRINTF(stdout, "%2.2f\t%2.2f\n", vector_in_ptr[row], vector_out_ptr[row]);
}
starpu_free(nzval);
starpu_free(colind);
starpu_free(rowptr);
starpu_free(vector_in_ptr);
starpu_free(vector_out_ptr);
/*
* Stop StarPU
*/
starpu_shutdown();
timing = end - start;
FPRINTF(stderr, "Computation took (in ms)\n");
FPRINTF(stdout, "%2.2f\n", timing/1000);
return 0;
}