void soclShutdown() {
static int shutdown = 0;
if (!shutdown) {
shutdown = 1;
starpu_pthread_mutex_lock(&_socl_mutex);
if( _starpu_init )
starpu_task_wait_for_all();
gc_stop();
if( _starpu_init )
starpu_task_wait_for_all();
int active_entities = gc_active_entity_count();
if (active_entities != 0) {
DEBUG_MSG("Unreleased entities: %d\n", active_entities);
gc_print_remaining_entities();
}
if( _starpu_init && _starpu_init_failed != -ENODEV)
starpu_shutdown();
starpu_pthread_mutex_unlock(&_socl_mutex);
if (socl_devices != NULL) {
free(socl_devices);
socl_devices = NULL;
}
}
}
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;
}
int main(int argc, char **argv)
{
double timing;
struct timeval start;
struct timeval end;
starpu_init(NULL);
fprintf(stderr, "#tasks : %d\n", ntasks);
unsigned i;
for (i = 0; i < ntasks; i++)
{
struct starpu_task *task = starpu_task_create();
/* We check if the function is valid from the codelet or from
* the callback */
task->cl = &dummy_cl;
task->cl_arg = task;
task->callback_func = check_task_callback;
task->callback_arg = task;
int ret = starpu_task_submit(task, NULL);
STARPU_ASSERT(!ret);
}
starpu_task_wait_for_all();
fprintf(stderr, "#empty tasks : %d\n", ntasks);
/* We repeat the same experiment with null codelets */
for (i = 0; i < ntasks; i++)
{
struct starpu_task *task = starpu_task_create();
task->cl = NULL;
/* We check if the function is valid from the callback */
task->callback_func = check_task_callback;
task->callback_arg = task;
int ret = starpu_task_submit(task, NULL);
STARPU_ASSERT(!ret);
}
starpu_task_wait_for_all();
starpu_shutdown();
return 0;
}
int main(int argc, char **argv)
{
int ntasks = NTASKS;
int ret;
struct starpu_conf conf;
starpu_conf_init(&conf);
conf.sched_policy = &dummy_sched_policy,
ret = starpu_init(&conf);
if (ret == -ENODEV)
return 77;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
#ifdef STARPU_QUICK_CHECK
ntasks /= 100;
#endif
int i;
for (i = 0; i < ntasks; i++)
{
struct starpu_task *task = starpu_task_create();
task->cl = &dummy_codelet;
task->cl_arg = NULL;
ret = starpu_task_submit(task);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
starpu_task_wait_for_all();
starpu_shutdown();
return 0;
}
static int
run(struct starpu_sched_policy *policy)
{
int ret;
struct starpu_conf conf;
int i;
starpu_conf_init(&conf);
conf.sched_policy = policy;
ret = starpu_init(&conf);
if (ret != 0)
exit(STARPU_TEST_SKIPPED);
starpu_profiling_status_set(1);
struct starpu_codelet clA =
{
.cpu_funcs = {A},
.nbuffers = 0
};
struct starpu_codelet clB =
{
.cpu_funcs = {B},
.nbuffers = 0
};
starpu_srand48(0);
for (i = 0; i < NTASKS; i++)
{
struct starpu_task *task = starpu_task_create();
if (((int)(starpu_drand48()*2))%2)
{
task->cl = &clA;
task->priority=STARPU_MIN_PRIO;
}
else
{
task->cl = &clB;
task->priority=STARPU_MAX_PRIO;
}
task->detach=1;
ret = starpu_task_submit(task);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
}
starpu_task_wait_for_all();
FPRINTF(stdout,"\n");
starpu_shutdown();
return 0;
enodev:
starpu_shutdown();
return -ENODEV;
}
int main(int argc, char **argv)
{
int ret;
unsigned loop, nloops=NLOOPS;
ret = starpu_initialize(NULL, &argc, &argv);
if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
struct starpu_task *taskA, *taskB;
for (loop = 0; loop < nloops; loop++)
{
taskA = create_dummy_task();
taskB = create_dummy_task();
/* By default, dynamically allocated tasks are destroyed at
* termination, we cannot declare a dependency on something
* that does not exist anymore. */
taskA->destroy = 0;
taskA->synchronous = 1;
ret = starpu_task_submit(taskA);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
starpu_task_declare_deps_array(taskB, 1, &taskA);
taskB->synchronous = 1;
ret = starpu_task_submit(taskB);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
starpu_task_destroy(taskA);
}
ret = starpu_task_wait_for_all();
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_wait_for_all");
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;
}
/*main program*/
int main(int argc, char * argv[])
{
/* Init */
int ret;
int mpi_rank, mpi_size;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &mpi_rank);
MPI_Comm_size(MPI_COMM_WORLD, &mpi_size);
ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
ret = starpu_mpi_init(NULL, NULL, 0);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
/*element initialization : domains are connected as a ring for this test*/
int num_elements=NUM_EL;
struct element * el_left=malloc(num_elements*sizeof(el_left[0]));
struct element * el_right=malloc(num_elements*sizeof(el_right[0]));
int i;
for(i=0; i<num_elements; i++)
{
init_element(el_left+i,i+1,((mpi_rank-1)+mpi_size)%mpi_size);
init_element(el_right+i,i+1,(mpi_rank+1)%mpi_size);
}
/* Communication loop */
for (i=0; i<NUM_LOOPS; i++) //number of "computations loops"
{
int e;
for (e=0; e<num_elements; e++) //Do something for each elements
{
insert_work_for_one_element(el_right+e);
insert_work_for_one_element(el_left+e);
}
}
/* End */
starpu_task_wait_for_all();
for(i=0; i<num_elements; i++)
{
free_element(el_left+i);
free_element(el_right+i);
}
starpu_mpi_shutdown();
starpu_shutdown();
MPI_Finalize();
FPRINTF(stderr, "No assert until end\n");
return 0;
}
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");
unsigned b;
for (b = 0; b < NBUFFERS; b++)
{
buffers[b].index = b;
starpu_variable_data_register(&buffers[b].handle, STARPU_MAIN_RAM, (uintptr_t)&buffers[b].val, sizeof(unsigned));
}
unsigned iter;
for (iter = 0; iter < NITER; iter++)
{
for (b = 0; b < NBUFFERS; b++)
{
ret = starpu_data_acquire_cb(buffers[b].handle, STARPU_RW,
callback_sync_data, &buffers[b]);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_data_acquire_cb");
}
}
ret = starpu_task_wait_for_all();
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_wait_for_all");
/* do some cleanup */
ret = EXIT_SUCCESS;
for (b = 0; b < NBUFFERS; b++)
{
starpu_data_unregister(buffers[b].handle);
/* check result */
if (buffers[b].val != NITER)
{
FPRINTF(stderr, "buffer[%u] = %u should be %d\n", b, buffers[b].val, NITER);
ret = EXIT_FAILURE;
}
}
starpu_shutdown();
return ret;
}
int main(int argc, char **argv)
{
int v=40;
int ret;
ret = starpu_init(NULL);
if (ret == -ENODEV)
return 77;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
starpu_variable_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)&v, sizeof(int));
double *x = (double*)malloc(sizeof(double));
struct starpu_task *task = starpu_task_create();
task->cl = &cl;
task->prologue_callback_func = callback_func;
task->prologue_callback_arg = NULL;
task->prologue_callback_pop_func = pop_prologue_callback_func;
task->prologue_callback_pop_arg = (void*) 5;
task->handles[0] = handle;
ret = starpu_task_submit(task);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_submit");
*x = -999.0;
ret = starpu_task_insert(&cl,
STARPU_RW, handle,
STARPU_PROLOGUE_CALLBACK, prologue_callback_func,
STARPU_PROLOGUE_CALLBACK_ARG, x,
STARPU_PROLOGUE_CALLBACK_POP, pop_prologue_callback_func,
STARPU_PROLOGUE_CALLBACK_POP_ARG, 5,
0);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
starpu_task_wait_for_all();
enodev:
starpu_data_unregister(handle);
free(x);
FPRINTF(stderr, "v -> %d\n", v);
starpu_shutdown();
return (ret == -ENODEV) ? 77 : 0;
}
int main(int argc, char **argv)
{
starpu_init(&conf);
unsigned i;
for (i = 0; i < NTASKS; i++)
{
struct starpu_task *task = starpu_task_create();
task->cl = &dummy_codelet;
task->cl_arg = NULL;
starpu_task_submit(task, NULL);
}
starpu_task_wait_for_all();
starpu_shutdown();
return 0;
}
int main(int argc, char **argv)
{
starpu_init(NULL);
/* Allocate all buffers and register them to StarPU */
unsigned b;
for (b = 0; b < NBUFFERS; b++)
{
starpu_data_malloc_pinned_if_possible((void **)&buffer[b], VECTORSIZE);
starpu_vector_data_register(&v_handle[b], 0,
(uintptr_t)buffer[b], VECTORSIZE, sizeof(char));
}
unsigned 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++)
use_handle(v_handle[b]);
starpu_task_wait_for_all();
/* Grab the different pieces of data into main memory */
for (b = 0; b < NBUFFERS; b++)
starpu_data_acquire(v_handle[b], STARPU_RW);
/* 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_shutdown();
return 0;
}
int main(int argc, char **argv)
{
int ret;
int var = 42;
starpu_data_handle_t handle;
ret = starpu_init(NULL);
if (ret == -ENODEV) return STARPU_TEST_SKIPPED;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
int copy = starpu_asynchronous_copy_disabled();
FPRINTF(stderr, "copy %d\n", copy);
starpu_variable_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)&var, sizeof(var));
ret = starpu_task_insert(&cl,
STARPU_R, handle,
0);
if (ret == -ENODEV) goto enodev;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
starpu_task_wait_for_all();
starpu_data_unregister(handle);
starpu_shutdown();
return 0;
enodev:
starpu_data_unregister(handle);
starpu_shutdown();
/* 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 */
fprintf(stderr, "WARNING: No one can execute this task\n");
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)
{
unsigned i;
float foo;
starpu_data_handle float_array_handle;
starpu_codelet cl;
starpu_init(NULL);
if (argc == 2) niter = atoi(argv[1]);
foo = 0.0f;
starpu_variable_data_register(&float_array_handle, 0 /* home node */,
(uintptr_t)&foo, sizeof(float));
#ifdef STARPU_USE_OPENCL
starpu_opencl_load_opencl_from_file("examples/basic_examples/variable_kernels_opencl_codelet.cl", &opencl_code);
#endif
cl.where = STARPU_CPU|STARPU_CUDA|STARPU_OPENCL;
cl.cpu_func = cpu_codelet;
#ifdef STARPU_USE_CUDA
cl.cuda_func = cuda_codelet;
#endif
#ifdef STARPU_USE_OPENCL
cl.opencl_func = opencl_codelet;
#endif
cl.nbuffers = 1;
cl.model = NULL;
for (i = 0; i < niter; i++)
{
struct starpu_task *task = starpu_task_create();
int ret;
task->cl = &cl;
task->callback_func = NULL;
task->buffers[0].handle = float_array_handle;
task->buffers[0].mode = STARPU_RW;
ret = starpu_task_submit(task, NULL);
if (STARPU_UNLIKELY(ret == -ENODEV))
{
fprintf(stderr, "No worker may execute this task\n");
exit(0);
}
}
starpu_task_wait_for_all();
/* update the array in RAM */
starpu_data_acquire(float_array_handle, STARPU_R);
fprintf(stderr, "variable -> %f\n", foo);
starpu_data_release(float_array_handle);
starpu_shutdown();
return 0;
}
int main(int argc, char **argv)
{
int size, x;
int color;
MPI_Comm newcomm;
int rank, newrank;
int ret;
starpu_data_handle_t data[3];
int value = 90;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (size < 4)
{
FPRINTF(stderr, "We need at least 4 processes.\n");
MPI_Finalize();
return STARPU_TEST_SKIPPED;
}
color = rank%2;
MPI_Comm_split(MPI_COMM_WORLD, color, rank, &newcomm);
MPI_Comm_rank(newcomm, &newrank);
FPRINTF(stderr, "[%d][%d] color %d\n", rank, newrank, color);
if (newrank == 0)
{
FPRINTF(stderr, "[%d][%d] sending %d\n", rank, newrank, rank);
MPI_Send(&rank, 1, MPI_INT, 1, 10, newcomm);
}
else if (newrank == 1)
{
MPI_Recv(&x, 1, MPI_INT, 0, 10, newcomm, MPI_STATUS_IGNORE);
FPRINTF(stderr, "[%d][%d] received %d\n", rank, newrank, x);
}
ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
ret = starpu_mpi_init(NULL, NULL, 0);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
if (rank == 0)
{
starpu_variable_data_register(&data[2], STARPU_MAIN_RAM, (uintptr_t)&value, sizeof(int));
}
else
starpu_variable_data_register(&data[2], -1, (uintptr_t)NULL, sizeof(int));
starpu_mpi_data_register_comm(data[2], 44, 0, MPI_COMM_WORLD);
if (newrank == 0)
{
starpu_variable_data_register(&data[0], STARPU_MAIN_RAM, (uintptr_t)&rank, sizeof(int));
starpu_variable_data_register(&data[1], STARPU_MAIN_RAM, (uintptr_t)&rank, sizeof(int));
starpu_mpi_data_register_comm(data[1], 22, 0, newcomm);
}
else
starpu_variable_data_register(&data[0], -1, (uintptr_t)NULL, sizeof(int));
starpu_mpi_data_register_comm(data[0], 12, 0, newcomm);
if (newrank == 0)
{
starpu_mpi_req req[2];
starpu_mpi_issend(data[1], &req[0], 1, 22, newcomm);
starpu_mpi_isend(data[0], &req[1], 1, 12, newcomm);
starpu_mpi_wait(&req[0], MPI_STATUS_IGNORE);
starpu_mpi_wait(&req[1], MPI_STATUS_IGNORE);
}
else if (newrank == 1)
{
int *xx;
starpu_mpi_recv(data[0], 0, 12, newcomm, MPI_STATUS_IGNORE);
starpu_data_acquire(data[0], STARPU_RW);
xx = (int *)starpu_variable_get_local_ptr(data[0]);
starpu_data_release(data[0]);
FPRINTF(stderr, "[%d][%d] received %d\n", rank, newrank, *xx);
STARPU_ASSERT_MSG(x==*xx, "Received value %d is incorrect (should be %d)\n", *xx, x);
starpu_variable_data_register(&data[1], -1, (uintptr_t)NULL, sizeof(int));
starpu_mpi_data_register_comm(data[1], 22, 0, newcomm);
starpu_mpi_recv(data[0], 0, 22, newcomm, MPI_STATUS_IGNORE);
starpu_data_acquire(data[0], STARPU_RW);
xx = (int *)starpu_variable_get_local_ptr(data[0]);
starpu_data_release(data[0]);
FPRINTF(stderr, "[%d][%d] received %d\n", rank, newrank, *xx);
STARPU_ASSERT_MSG(x==*xx, "Received value %d is incorrect (should be %d)\n", *xx, x);
}
if (rank == 0)
{
starpu_data_acquire(data[2], STARPU_RW);
int rvalue = *((int *)starpu_variable_get_local_ptr(data[2]));
starpu_data_release(data[2]);
FPRINTF_MPI(stderr, "sending value %d to %d and receiving from %d\n", rvalue, 1, size-1);
starpu_mpi_send(data[2], 1, 44, MPI_COMM_WORLD);
starpu_mpi_recv(data[2], size-1, 44, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
starpu_data_acquire(data[2], STARPU_RW);
int *xx = (int *)starpu_variable_get_local_ptr(data[2]);
starpu_data_release(data[2]);
FPRINTF_MPI(stderr, "Value back is %d\n", *xx);
STARPU_ASSERT_MSG(*xx == rvalue + (2*(size-1)), "Received value %d is incorrect (should be %d)\n", *xx, rvalue + (2*(size-1)));
}
else
{
int next = (rank == size-1) ? 0 : rank+1;
starpu_mpi_recv(data[2], rank-1, 44, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
starpu_data_acquire(data[2], STARPU_RW);
int *xx = (int *)starpu_variable_get_local_ptr(data[2]);
FPRINTF_MPI(stderr, "receiving %d from %d and sending %d to %d\n", *xx, rank-1, *xx+2, next);
*xx = *xx + 2;
starpu_data_release(data[2]);
starpu_mpi_send(data[2], next, 44, MPI_COMM_WORLD);
}
if (newrank == 0 || newrank == 1)
{
starpu_mpi_insert_task(newcomm, &mycodelet,
STARPU_RW, data[0],
STARPU_VALUE, &x, sizeof(x),
STARPU_EXECUTE_ON_NODE, 1,
0);
starpu_task_wait_for_all();
starpu_data_unregister(data[0]);
starpu_data_unregister(data[1]);
}
starpu_data_unregister(data[2]);
starpu_mpi_shutdown();
starpu_shutdown();
MPI_Comm_free(&newcomm);
MPI_Finalize();
return 0;
}
int main(int argc, char **argv)
{
int ret, rank, size, err, node;
long x0=32;
int x1=23;
starpu_data_handle_t data_handlesx0;
starpu_data_handle_t data_handlesx1;
ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
ret = starpu_mpi_init(&argc, &argv, 1);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (rank != 0 && rank != 1) goto end;
if (rank == 0)
{
starpu_variable_data_register(&data_handlesx0, STARPU_MAIN_RAM, (uintptr_t)&x0, sizeof(x0));
starpu_mpi_data_register(data_handlesx0, 0, rank);
starpu_variable_data_register(&data_handlesx1, -1, (uintptr_t)NULL, sizeof(x1));
starpu_mpi_data_register(data_handlesx1, 1, 1);
}
else if (rank == 1)
{
starpu_variable_data_register(&data_handlesx1, STARPU_MAIN_RAM, (uintptr_t)&x1, sizeof(x1));
starpu_mpi_data_register(data_handlesx1, 1, rank);
starpu_variable_data_register(&data_handlesx0, -1, (uintptr_t)NULL, sizeof(x0));
starpu_mpi_data_register(data_handlesx0, 0, 0);
}
node = starpu_mpi_data_get_rank(data_handlesx1);
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_r_w,
STARPU_VALUE, &node, sizeof(node),
STARPU_R, data_handlesx0, STARPU_W, data_handlesx1,
0);
assert(err == 0);
node = starpu_mpi_data_get_rank(data_handlesx0);
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_rw_r,
STARPU_VALUE, &node, sizeof(node),
STARPU_RW, data_handlesx0, STARPU_R, data_handlesx1,
0);
assert(err == 0);
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_rw_rw,
STARPU_VALUE, &node, sizeof(node),
STARPU_RW, data_handlesx0, STARPU_RW, data_handlesx1,
0);
assert(err == 0);
node = 1;
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_rw_rw,
STARPU_VALUE, &node, sizeof(node),
STARPU_RW, data_handlesx0, STARPU_RW, data_handlesx1, STARPU_EXECUTE_ON_NODE, node,
0);
assert(err == 0);
node = 0;
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_rw_rw,
STARPU_VALUE, &node, sizeof(node),
STARPU_RW, data_handlesx0, STARPU_RW, data_handlesx1, STARPU_EXECUTE_ON_NODE, node,
0);
assert(err == 0);
node = 0;
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_r_r,
STARPU_VALUE, &node, sizeof(node),
STARPU_R, data_handlesx0, STARPU_R, data_handlesx1, STARPU_EXECUTE_ON_NODE, node,
0);
assert(err == 0);
/* Here the value specified by the property STARPU_EXECUTE_ON_NODE is
going to overwrite the node even though the data model clearly specifies
which node is going to execute the codelet */
node = 0;
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_r_w,
STARPU_VALUE, &node, sizeof(node),
STARPU_R, data_handlesx0, STARPU_W, data_handlesx1, STARPU_EXECUTE_ON_NODE, node,
0);
assert(err == 0);
/* Here the value specified by the property STARPU_EXECUTE_ON_NODE is
going to overwrite the node even though the data model clearly specifies
which node is going to execute the codelet */
node = 0;
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet_w_r,
STARPU_VALUE, &node, sizeof(node),
STARPU_W, data_handlesx0, STARPU_R, data_handlesx1, STARPU_EXECUTE_ON_NODE, node,
0);
assert(err == 0);
FPRINTF_MPI(stderr, "Waiting ...\n");
starpu_task_wait_for_all();
starpu_data_unregister(data_handlesx0);
starpu_data_unregister(data_handlesx1);
end:
starpu_mpi_shutdown();
starpu_shutdown();
return 0;
}
int main(int argc, char **argv)
{
int ret = 0;
starpu_data_handle_t handle1;
starpu_data_handle_t handle2;
double real = 45.0;
double imaginary = 12.0;
double copy_real = 78.0;
double copy_imaginary = 78.0;
int compare;
int *compare_ptr = &compare;
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/interface/complex_kernels.cl",
&opencl_program, NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file");
#endif
starpu_complex_data_register(&handle1, STARPU_MAIN_RAM, &real, &imaginary, 1);
starpu_complex_data_register(&handle2, STARPU_MAIN_RAM, ©_real, ©_imaginary, 1);
ret = starpu_task_insert(&cl_display, STARPU_VALUE, "handle1", strlen("handle1")+1, STARPU_R, handle1, 0);
if (ret == -ENODEV) goto end;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
ret = starpu_task_insert(&cl_display, STARPU_VALUE, "handle2", strlen("handle2")+1, STARPU_R, handle2, 0);
if (ret == -ENODEV) goto end;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
ret = starpu_task_insert(&cl_compare,
STARPU_R, handle1,
STARPU_R, handle2,
STARPU_VALUE, &compare_ptr, sizeof(compare_ptr),
0);
if (ret == -ENODEV) goto end;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
starpu_task_wait_for_all();
if (compare != 0)
{
FPRINTF(stderr, "Complex numbers should NOT be similar\n");
goto end;
}
ret = starpu_task_insert(&cl_copy,
STARPU_R, handle1,
STARPU_W, handle2,
0);
if (ret == -ENODEV) goto end;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
ret = starpu_task_insert(&cl_display, STARPU_VALUE, "handle1", strlen("handle1")+1, STARPU_R, handle1, 0);
if (ret == -ENODEV) goto end;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
ret = starpu_task_insert(&cl_display, STARPU_VALUE, "handle2", strlen("handle2")+1, STARPU_R, handle2, 0);
if (ret == -ENODEV) goto end;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
ret = starpu_task_insert(&cl_compare,
STARPU_R, handle1,
STARPU_R, handle2,
STARPU_VALUE, &compare_ptr, sizeof(compare_ptr),
0);
if (ret == -ENODEV) goto end;
STARPU_CHECK_RETURN_VALUE(ret, "starpu_task_insert");
starpu_task_wait_for_all();
if (compare != 1)
{
FPRINTF(stderr, "Complex numbers should be similar\n");
}
end:
#ifdef STARPU_USE_OPENCL
{
int ret2 = starpu_opencl_unload_opencl(&opencl_program);
STARPU_CHECK_RETURN_VALUE(ret2, "starpu_opencl_unload_opencl");
}
#endif
starpu_data_unregister(handle1);
starpu_data_unregister(handle2);
starpu_shutdown();
if (ret == -ENODEV) return 77; else return !compare;
}
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");
#ifdef STARPU_USE_OPENCL
ret = starpu_opencl_load_opencl_from_file("tests/datawizard/scratch_opencl_kernel.cl",
&opencl_program, NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_load_opencl_from_file");
#endif
A = (unsigned *) calloc(VECTORSIZE, sizeof(unsigned));
starpu_vector_data_register(&A_handle, STARPU_MAIN_RAM, (uintptr_t)A, VECTORSIZE, sizeof(unsigned));
starpu_vector_data_register(&B_handle, -1, (uintptr_t)NULL, VECTORSIZE, sizeof(unsigned));
unsigned loop;
for (loop = 0; loop < NLOOPS; loop++)
{
struct starpu_task *task_f = starpu_task_create();
task_f->cl = &cl_f;
task_f->handles[0] = A_handle;
task_f->handles[1] = B_handle;
ret = starpu_task_submit(task_f);
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_unregister(A_handle);
starpu_data_unregister(B_handle);
#ifdef STARPU_USE_OPENCL
ret = starpu_opencl_unload_opencl(&opencl_program);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_unload_opencl");
#endif
starpu_shutdown();
/* Check result */
unsigned i;
ret = EXIT_SUCCESS;
for (i = 0; i < VECTORSIZE; i++)
{
if (A[i] != NLOOPS)
{
FPRINTF(stderr, "Error: Incorrect value A[%u] = %u != %d\n", i, A[i], NLOOPS);
ret = EXIT_FAILURE;
break;
}
}
free(A);
STARPU_RETURN(ret);
enodev:
starpu_data_unregister(A_handle);
starpu_data_unregister(B_handle);
#ifdef STARPU_USE_OPENCL
ret = starpu_opencl_unload_opencl(&opencl_program);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_opencl_unload_opencl");
#endif
starpu_shutdown();
/* 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 */
fprintf(stderr, "WARNING: No one can execute this task\n");
return STARPU_TEST_SKIPPED;
}
int main(int argc, char **argv)
{
int rank, nodes;
int ret=0;
int compare=0;
ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
ret = starpu_mpi_init(&argc, &argv, 1);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
starpu_mpi_comm_rank(MPI_COMM_WORLD, &rank);
starpu_mpi_comm_size(MPI_COMM_WORLD, &nodes);
if (nodes < 2)
{
fprintf(stderr, "This program needs at least 2 nodes (%d available)\n", nodes);
ret = 77;
}
else
{
starpu_data_handle_t handle;
starpu_data_handle_t handle2;
double real[2] = {4.0, 2.0};
double imaginary[2] = {7.0, 9.0};
double real2[2] = {14.0, 12.0};
double imaginary2[2] = {17.0, 19.0};
if (rank == 1)
{
real[0] = 0.0;
real[1] = 0.0;
imaginary[0] = 0.0;
imaginary[1] = 0.0;
}
starpu_complex_data_register(&handle, STARPU_MAIN_RAM, real, imaginary, 2);
starpu_complex_data_register(&handle2, -1, real2, imaginary2, 2);
if (rank == 0)
{
int *compare_ptr = &compare;
starpu_task_insert(&cl_display, STARPU_VALUE, "node0 initial value", strlen("node0 initial value")+1, STARPU_R, handle, 0);
starpu_mpi_isend_detached(handle, 1, 10, MPI_COMM_WORLD, NULL, NULL);
starpu_mpi_irecv_detached(handle2, 1, 20, MPI_COMM_WORLD, NULL, NULL);
starpu_task_insert(&cl_display, STARPU_VALUE, "node0 received value", strlen("node0 received value")+1, STARPU_R, handle2, 0);
starpu_task_insert(&cl_compare, STARPU_R, handle, STARPU_R, handle2, STARPU_VALUE, &compare_ptr, sizeof(compare_ptr), 0);
}
else if (rank == 1)
{
starpu_mpi_irecv_detached(handle, 0, 10, MPI_COMM_WORLD, NULL, NULL);
starpu_task_insert(&cl_display, STARPU_VALUE, "node1 received value", strlen("node1 received value")+1, STARPU_R, handle, 0);
starpu_mpi_isend_detached(handle, 0, 20, MPI_COMM_WORLD, NULL, NULL);
}
starpu_task_wait_for_all();
starpu_data_unregister(handle);
starpu_data_unregister(handle2);
}
starpu_mpi_shutdown();
starpu_shutdown();
if (rank == 0) return !compare; else return ret;
}
int do_test(int rank, int sdetached, int rdetached)
{
int ret, i;
int val[2];
starpu_data_handle_t data[2];
ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
ret = starpu_mpi_init(NULL, NULL, 0);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
if (rank == 1)
{
val[0] = VAL0;
val[1] = VAL1;
}
else
{
val[0] = -1;
val[1] = -1;
}
starpu_variable_data_register(&data[0], STARPU_MAIN_RAM, (uintptr_t)&val[0], sizeof(val[0]));
starpu_variable_data_register(&data[1], STARPU_MAIN_RAM, (uintptr_t)&val[1], sizeof(val[1]));
starpu_mpi_data_register(data[0], 77, 1);
starpu_mpi_data_register(data[1], 88, 1);
if (rank == 1)
{
for(i=1 ; i>=0 ; i--)
{
if (sdetached)
starpu_mpi_isend_detached(data[i], 0, starpu_data_get_tag(data[i]), MPI_COMM_WORLD, NULL, NULL);
else
starpu_mpi_send(data[i], 0, starpu_data_get_tag(data[i]), MPI_COMM_WORLD);
}
}
else if (rank == 0)
{
int received = 0;
for(i=0 ; i<2 ; i++)
FPRINTF_MPI(stderr, "Value[%d] = %d\n", i, val[i]);
for(i=0 ; i<2 ; i++)
{
if (rdetached)
starpu_mpi_irecv_detached(data[i], 1, starpu_data_get_tag(data[i]), MPI_COMM_WORLD, callback, &received);
else
starpu_mpi_recv(data[i], 1, starpu_data_get_tag(data[i]), MPI_COMM_WORLD, MPI_STATUS_IGNORE);
}
if (rdetached)
{
STARPU_PTHREAD_MUTEX_LOCK(&mutex);
while (received != 2)
{
FPRINTF_MPI(stderr, "Received %d messages\n", received);
STARPU_PTHREAD_COND_WAIT(&cond, &mutex);
}
STARPU_PTHREAD_MUTEX_UNLOCK(&mutex);
}
for(i=0 ; i<2 ; i++)
starpu_data_acquire(data[i], STARPU_R);
for(i=0 ; i<2 ; i++)
FPRINTF_MPI(stderr, "Value[%d] = %d\n", i, val[i]);
for(i=0 ; i<2 ; i++)
starpu_data_release(data[i]);
}
FPRINTF_MPI(stderr, "Waiting ...\n");
starpu_task_wait_for_all();
starpu_data_unregister(data[0]);
starpu_data_unregister(data[1]);
if (rank == 0)
{
ret = (val[0] == VAL0 && val[1] == VAL1) ? 0 : 1;
}
starpu_mpi_shutdown();
starpu_shutdown();
return ret;
}
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 main(int argc, char **argv)
{
int ret, rank, size;
ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
ret = starpu_mpi_init(NULL, NULL, 1);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (size < 2)
{
if (rank == 0)
FPRINTF(stderr, "We need at least 2 processes.\n");
MPI_Finalize();
return STARPU_TEST_SKIPPED;
}
starpu_vector_data_register(&token_handle, STARPU_MAIN_RAM, (uintptr_t)&token, 1, sizeof(token));
int nloops = NITER;
int loop;
int last_loop = nloops - 1;
int last_rank = size - 1;
for (loop = 0; loop < nloops; loop++)
{
int tag = loop*size + rank;
if (loop == 0 && rank == 0)
{
token = 0;
FPRINTF(stdout, "Start with token value %u\n", token);
}
else
{
starpu_mpi_irecv_detached(token_handle, (rank+size-1)%size, tag, MPI_COMM_WORLD, NULL, NULL);
}
increment_token();
if (loop == last_loop && rank == last_rank)
{
starpu_data_acquire(token_handle, STARPU_R);
FPRINTF(stdout, "Finished : token value %u\n", token);
starpu_data_release(token_handle);
}
else
{
starpu_mpi_isend_detached(token_handle, (rank+1)%size, tag+1, MPI_COMM_WORLD, NULL, NULL);
}
}
starpu_task_wait_for_all();
starpu_data_unregister(token_handle);
starpu_mpi_shutdown();
starpu_shutdown();
if (rank == last_rank)
{
FPRINTF(stderr, "[%d] token = %u == %u * %d ?\n", rank, token, nloops, size);
STARPU_ASSERT(token == nloops*size);
}
return 0;
}
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 rank, size, err;
int x[2];
int ret, i;
starpu_data_handle_t data_handles[2];
int values[2];
ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
ret = starpu_mpi_init(&argc, &argv, 1);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_mpi_init");
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (rank == 0)
{
x[0] = 11;
starpu_variable_data_register(&data_handles[0], STARPU_MAIN_RAM, (uintptr_t)&x[0], sizeof(x[0]));
starpu_variable_data_register(&data_handles[1], -1, (uintptr_t)NULL, sizeof(x[1]));
}
else if (rank == 1)
{
x[1] = 12;
starpu_variable_data_register(&data_handles[0], -1, (uintptr_t)NULL, sizeof(x[0]));
starpu_variable_data_register(&data_handles[1], STARPU_MAIN_RAM, (uintptr_t)&x[1], sizeof(x[1]));
}
else
{
starpu_variable_data_register(&data_handles[0], -1, (uintptr_t)NULL, sizeof(x[0]));
starpu_variable_data_register(&data_handles[1], -1, (uintptr_t)NULL, sizeof(x[1]));
}
starpu_mpi_data_register(data_handles[0], 0, 0);
starpu_mpi_data_register(data_handles[1], 1, 1);
err = starpu_mpi_task_insert(MPI_COMM_WORLD, &mycodelet,
STARPU_RW, data_handles[0], STARPU_RW, data_handles[1],
STARPU_EXECUTE_ON_DATA, data_handles[1],
0);
assert(err == 0);
starpu_task_wait_for_all();
for(i=0 ; i<2 ; i++)
{
starpu_mpi_get_data_on_node_detached(MPI_COMM_WORLD, data_handles[i], 0, NULL, NULL);
if (rank == 0)
{
starpu_data_acquire(data_handles[i], STARPU_R);
values[i] = *((int *)starpu_data_get_local_ptr(data_handles[i]));
starpu_data_release(data_handles[i]); }
}
ret = 0;
if (rank == 0)
{
FPRINTF(stderr, "[%d][local ptr] VALUES: %d %d\n", rank, values[0], values[1]);
if (values[0] != 12 || values[1] != 144)
{
ret = EXIT_FAILURE;
}
}
starpu_data_unregister(data_handles[0]);
starpu_data_unregister(data_handles[1]);
starpu_mpi_shutdown();
starpu_shutdown();
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 dw_codelet_facto_v3(starpu_data_handle_t dataA, unsigned nblocks)
{
double start;
double end;
int ret;
/* create all the DAG nodes */
unsigned i,j,k;
if (bound)
starpu_bound_start(bounddeps, boundprio);
start = starpu_timing_now();
for (k = 0; k < nblocks; k++)
{
ret = create_task_11(dataA, k);
if (ret == -ENODEV) return ret;
for (i = k+1; i<nblocks; i++)
{
ret = create_task_12(dataA, k, i);
if (ret == -ENODEV) return ret;
ret = create_task_21(dataA, k, i);
if (ret == -ENODEV) return ret;
}
starpu_data_wont_use(starpu_data_get_sub_data(dataA, 2, k, k));
for (i = k+1; i<nblocks; i++)
for (j = k+1; j<nblocks; j++)
{
ret = create_task_22(dataA, k, i, j);
if (ret == -ENODEV) return ret;
}
for (i = k+1; i<nblocks; i++)
{
starpu_data_wont_use(starpu_data_get_sub_data(dataA, 2, k, i));
starpu_data_wont_use(starpu_data_get_sub_data(dataA, 2, i, k));
}
}
/* stall the application until the end of computations */
starpu_task_wait_for_all();
end = starpu_timing_now();
if (bound)
starpu_bound_stop();
double timing = end - start;
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");
return 0;
}
static int dw_codelet_facto_pivot(starpu_data_handle_t *dataAp,
struct piv_s *piv_description,
unsigned nblocks,
starpu_data_handle_t (* get_block)(starpu_data_handle_t *, unsigned, unsigned, unsigned),
double *timing)
{
double start;
double end;
int ret;
/* create all the DAG nodes */
unsigned i,j,k;
if (bound)
starpu_bound_start(bounddeps, boundprio);
start = starpu_timing_now();
for (k = 0; k < nblocks; k++)
{
ret = create_task_11_pivot(dataAp, nblocks, k, piv_description, get_block);
if (ret == -ENODEV) return ret;
for (i = 0; i < nblocks; i++)
{
if (i != k)
{
ret = create_task_pivot(dataAp, nblocks, piv_description, k, i, get_block);
if (ret == -ENODEV) return ret;
}
}
for (i = k+1; i<nblocks; i++)
{
ret = create_task_12(dataAp, nblocks, k, i, get_block);
if (ret == -ENODEV) return ret;
ret = create_task_21(dataAp, nblocks, k, i, get_block);
if (ret == -ENODEV) return ret;
}
starpu_data_wont_use(get_block(dataAp, nblocks, k, k));
for (i = k+1; i<nblocks; i++)
for (j = k+1; j<nblocks; j++)
{
ret = create_task_22(dataAp, nblocks, k, i, j, get_block);
if (ret == -ENODEV) return ret;
}
for (i = k+1; i<nblocks; i++)
{
starpu_data_wont_use(get_block(dataAp, nblocks, k, i));
starpu_data_wont_use(get_block(dataAp, nblocks, i, k));
}
}
/* stall the application until the end of computations */
starpu_task_wait_for_all();
end = starpu_timing_now();
if (bound)
starpu_bound_stop();
*timing = end - start;
return 0;
}
int main(int argc, char **argv)
{
int my_rank, size, x, y, loop;
float mean=0;
float matrix[X][Y];
starpu_data_handle_t data_handles[X][Y];
int ret = starpu_init(NULL);
STARPU_CHECK_RETURN_VALUE(ret, "starpu_init");
starpu_mpi_init(&argc, &argv, 1);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
parse_args(argc, argv);
/* Initial data values */
starpu_srand48((long int)time(NULL));
for(x = 0; x < X; x++)
{
for (y = 0; y < Y; y++)
{
matrix[x][y] = (float)starpu_drand48();
mean += matrix[x][y];
}
}
mean /= (X*Y);
if (display)
{
FPRINTF_MPI(stdout, "mean=%2.2f\n", mean);
for(x = 0; x < X; x++)
{
fprintf(stdout, "[%d] ", my_rank);
for (y = 0; y < Y; y++)
{
fprintf(stdout, "%2.2f ", matrix[x][y]);
}
fprintf(stdout, "\n");
}
}
/* Initial distribution */
for(x = 0; x < X; x++)
{
for (y = 0; y < Y; y++)
{
int mpi_rank = my_distrib(x, y, size);
if (mpi_rank == my_rank)
{
//FPRINTF(stderr, "[%d] Owning data[%d][%d]\n", my_rank, x, y);
starpu_variable_data_register(&data_handles[x][y], 0, (uintptr_t)&(matrix[x][y]), sizeof(float));
}
else if (my_rank == my_distrib(x+1, y, size) || my_rank == my_distrib(x-1, y, size)
|| my_rank == my_distrib(x, y+1, size) || my_rank == my_distrib(x, y-1, size))
{
/* I don't own that index, but will need it for my computations */
//FPRINTF(stderr, "[%d] Neighbour of data[%d][%d]\n", my_rank, x, y);
starpu_variable_data_register(&data_handles[x][y], -1, (uintptr_t)NULL, sizeof(float));
}
else
{
/* I know it's useless to allocate anything for this */
data_handles[x][y] = NULL;
}
if (data_handles[x][y])
{
starpu_mpi_data_register(data_handles[x][y], (y*X)+x, mpi_rank);
}
}
}
/* First computation with initial distribution */
for(loop=0 ; loop<niter; loop++)
{
for (x = 1; x < X-1; x++)
{
for (y = 1; y < Y-1; y++)
{
starpu_mpi_task_insert(MPI_COMM_WORLD, &stencil5_cl, STARPU_RW, data_handles[x][y],
STARPU_R, data_handles[x-1][y], STARPU_R, data_handles[x+1][y],
STARPU_R, data_handles[x][y-1], STARPU_R, data_handles[x][y+1],
0);
}
}
}
FPRINTF(stderr, "Waiting ...\n");
starpu_task_wait_for_all();
/* Now migrate data to a new distribution */
/* First register newly needed data */
for(x = 0; x < X; x++)
{
for (y = 0; y < Y; y++)
{
int mpi_rank = my_distrib2(x, y, size);
if (!data_handles[x][y] && (mpi_rank == my_rank
|| my_rank == my_distrib2(x+1, y, size) || my_rank == my_distrib2(x-1, y, size)
|| my_rank == my_distrib2(x, y+1, size) || my_rank == my_distrib2(x, y-1, size)))
{
/* Register newly-needed data */
starpu_variable_data_register(&data_handles[x][y], -1, (uintptr_t)NULL, sizeof(float));
starpu_mpi_data_register(data_handles[x][y], (y*X)+x, mpi_rank);
}
if (data_handles[x][y] && mpi_rank != starpu_mpi_data_get_rank(data_handles[x][y]))
{
/* Migrate the data */
starpu_mpi_get_data_on_node_detached(MPI_COMM_WORLD, data_handles[x][y], mpi_rank, NULL, NULL);
/* And register new rank of the matrix */
starpu_mpi_data_set_rank(data_handles[x][y], mpi_rank);
}
}
}
/* Second computation with new distribution */
for(loop=0 ; loop<niter; loop++)
{
for (x = 1; x < X-1; x++)
{
for (y = 1; y < Y-1; y++)
{
starpu_mpi_task_insert(MPI_COMM_WORLD, &stencil5_cl, STARPU_RW, data_handles[x][y],
STARPU_R, data_handles[x-1][y], STARPU_R, data_handles[x+1][y],
STARPU_R, data_handles[x][y-1], STARPU_R, data_handles[x][y+1],
0);
}
}
}
FPRINTF(stderr, "Waiting ...\n");
starpu_task_wait_for_all();
/* Unregister data */
for(x = 0; x < X; x++)
{
for (y = 0; y < Y; y++)
{
if (data_handles[x][y])
{
int mpi_rank = my_distrib(x, y, size);
/* Get back data to original place where the user-provided buffer is. */
starpu_mpi_get_data_on_node_detached(MPI_COMM_WORLD, data_handles[x][y], mpi_rank, NULL, NULL);
/* Register original rank of the matrix (although useless) */
starpu_mpi_data_set_rank(data_handles[x][y], mpi_rank);
/* And unregister it */
starpu_data_unregister(data_handles[x][y]);
}
}
}
starpu_mpi_shutdown();
starpu_shutdown();
if (display)
{
FPRINTF(stdout, "[%d] mean=%2.2f\n", my_rank, mean);
for(x = 0; x < X; x++)
{
FPRINTF(stdout, "[%d] ", my_rank);
for (y = 0; y < Y; y++)
{
FPRINTF(stdout, "%2.2f ", matrix[x][y]);
}
FPRINTF(stdout, "\n");
}
}
return 0;
}
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;
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;
}
