int main(int argc, char **argv)
{
int ret;
ret = starpu_initialize(NULL, &argc, &argv);
if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
starpu_malloc((void**)&data, sizeof(*data));
*data = 42;
/* register a piece of data */
starpu_vector_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)data,
1, sizeof(unsigned));
struct starpu_task *task = starpu_task_create();
task->cl = &wrong_codelet;
task->handles[0] = handle;
task->use_tag = 1;
task->tag_id = TAG;
task->callback_func = wrong_callback;
task->detach = 0;
ret = starpu_task_submit(task);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
ret = starpu_tag_wait(TAG);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_tag_wait");
/* This call is valid as it is done by the application outside a
* callback */
ret = starpu_data_acquire(handle, STARPU_RW);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_data_acquire");
starpu_data_release(handle);
ret = starpu_task_wait(task);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_wait");
starpu_data_unregister(handle);
starpu_free(data);
starpu_shutdown();
return EXIT_SUCCESS;
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 */
starpu_shutdown();
return STARPU_TEST_SKIPPED;
}
static void shutdown_system(float **matA, unsigned pinned)
{
if (pinned)
{
starpu_free(*matA);
}
else
{
free(*matA);
}
starpu_cublas_shutdown();
starpu_shutdown();
}
void matrix_free(float ****bmat, int rank, int nodes, int alloc_everywhere)
{
unsigned x, y;
for(x=0 ; x<nblocks ; x++)
{
for(y=0 ; y<nblocks ; y++)
{
int mpi_rank = my_distrib(x, y, nodes);
if (alloc_everywhere || (mpi_rank == rank))
{
starpu_free((void *)(*bmat)[x][y]);
}
}
free((*bmat)[x]);
}
free(*bmat);
}
int main(int argc, char **argv)
{
int i, j, ret;
ret = starpu_initialize(NULL, &argc, &argv);
if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_initialize");
float *data;
starpu_malloc((void**)&data, sizeof(*data) * NB_BUNDLE);
float factors[NB_BUNDLE];
starpu_data_handle_t handles[NB_BUNDLE];
struct starpu_task *task[NB_ITERATION];
starpu_task_bundle_t bundles[NB_BUNDLE];
for (i = 0; i < NB_BUNDLE; i++)
{
data[i] = i + 1;
factors[i] = NB_BUNDLE - i;
}
for (i = 0; i < NB_BUNDLE; i++)
starpu_variable_data_register(&handles[i], STARPU_MAIN_RAM, (uintptr_t)&data[i], sizeof(float));
FPRINTF(stderr, "VALUES:");
for (i = 0; i < NB_BUNDLE; i++)
FPRINTF(stderr, " %f (%f)", data[i], factors[i]);
FPRINTF(stderr, "\n");
for (i = 0; i < NB_BUNDLE; i++)
{
starpu_task_bundle_create(&bundles[i]);
for (j = 0; j < NB_ITERATION; j++)
{
task[j] = starpu_task_create();
task[j]->cl = &codelet;
task[j]->cl_arg = &factors[i];
task[j]->cl_arg_size = sizeof(float);
task[j]->handles[0] = handles[i];
ret = starpu_task_bundle_insert(bundles[i], task[j]);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
}
ret = starpu_task_bundle_remove(bundles[i], task[NB_ITERATION / 2]);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_remove");
for (j = 0; j < NB_ITERATION; j++)
{
ret = starpu_task_submit(task[j]);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
starpu_task_bundle_close(bundles[i]);
}
ret = starpu_task_wait_for_all();
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_wait_for_all");
for(i = 0; i < NB_BUNDLE ; i++)
{
ret = starpu_data_acquire(handles[i], STARPU_R);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_data_acquire");
}
FPRINTF(stderr, "VALUES:");
for (i = 0; i < NB_BUNDLE; i++)
FPRINTF(stderr, " %f (%f)", data[i], factors[i]);
FPRINTF(stderr, "\n");
for(i = 0; i < NB_BUNDLE ; i++)
{
starpu_data_release(handles[i]);
starpu_data_unregister(handles[i]);
}
starpu_free(data);
starpu_shutdown();
return EXIT_SUCCESS;
enodev:
starpu_shutdown();
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 STARPU_TEST_SKIPPED;
}
int main(int argc, char **argv)
{
int ret;
#ifdef STARPU_QUICK_CHECK
nbuffers /= 4;
niter /= 4;
vectorsize /= 8;
#endif
ret = starpu_initialize(NULL, &argc, &argv);
if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
/* Allocate all buffers and register them to StarPU */
int b;
for (b = 0; b < nbuffers; b++)
{
ret = starpu_malloc((void **)&buffer[b], vectorsize);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_malloc");
starpu_vector_data_register(&v_handle[b], STARPU_MAIN_RAM,
(uintptr_t)buffer[b], vectorsize, sizeof(char));
}
int iter;
for (iter = 0; iter < niter; iter++)
{
/* Use the buffers on the different workers so that it may not
* be in main memory anymore */
for (b = 0; b < nbuffers; b++)
{
ret = use_handle(v_handle[b]);
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");
/* Grab the different pieces of data into main memory */
for (b = 0; b < nbuffers; b++)
{
ret = starpu_data_acquire(v_handle[b], STARPU_RW);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_data_acquire");
}
/* Release them */
for (b = 0; b < nbuffers; b++)
starpu_data_release(v_handle[b]);
}
/* do some cleanup */
for (b = 0; b < nbuffers; b++)
{
starpu_data_unregister(v_handle[b]);
starpu_free(buffer[b]);
}
starpu_shutdown();
return EXIT_SUCCESS;
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 */
starpu_shutdown();
return STARPU_TEST_SKIPPED;
}
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;
}
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;
}
