/*
* Wait for termination
*/
void wait_end_tasks(int rank)
{
int bz;
int nbz = get_nbz();
for (bz = 0; bz < nbz; bz++)
{
if (get_block_mpi_node(bz) == rank)
{
/* Wait for the task producing block "bz" */
starpu_tag_wait(TAG_FINISH(bz));
/* Get the result back to memory */
struct block_description *block = get_block_description(bz);
starpu_data_acquire(block->layers_handle[0], STARPU_R);
starpu_data_acquire(block->layers_handle[1], STARPU_R);
/* the data_acquire here is done to make sure
* the data is sent back to the ram memory, we
* can safely do a data_release, to avoid the
* data_unregister to block later on
*/
starpu_data_release(block->layers_handle[0]);
starpu_data_release(block->layers_handle[1]);
}
}
}
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 callback_sync_data(void *arg)
{
struct data *data = (struct data *) arg;
data->val++;
starpu_data_release(data->handle);
}
int main(int argc, char **argv)
{
int ret, rank, size;
starpu_data_handle_t handle;
int var;
MPI_Init(&argc, &argv);
starpu_mpi_comm_rank(MPI_COMM_WORLD, &rank);
starpu_mpi_comm_size(MPI_COMM_WORLD, &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");
if (size<3)
{
FPRINTF(stderr, "We need more than 2 processes.\n");
starpu_mpi_shutdown();
starpu_shutdown();
MPI_Finalize();
return STARPU_TEST_SKIPPED;
}
if (rank == 0)
{
int n;
for(n=1 ; n<size ; n++)
{
MPI_Status status;
FPRINTF_MPI(stderr, "receiving from node %d\n", n);
starpu_variable_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)&var, sizeof(var));
starpu_mpi_recv(handle, n, 42, MPI_COMM_WORLD, &status);
starpu_data_acquire(handle, STARPU_R);
STARPU_ASSERT_MSG(var == n, "Received incorrect value <%d> from node <%d>\n", var, n);
FPRINTF_MPI(stderr, "received <%d> from node %d\n", var, n);
starpu_data_release(handle);
starpu_data_unregister(handle);
}
}
else
{
FPRINTF_MPI(stderr, "sending to node %d\n", 0);
var = rank;
starpu_variable_data_register(&handle, STARPU_MAIN_RAM, (uintptr_t)&var, sizeof(var));
starpu_mpi_send(handle, 0, 42, MPI_COMM_WORLD);
starpu_data_unregister(handle);
}
starpu_mpi_shutdown();
starpu_shutdown();
MPI_Finalize();
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 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)
{
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 ret, rank, size;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (size<3)
{
FPRINTF(stderr, "We need more than 2 processes.\n");
MPI_Finalize();
return STARPU_TEST_SKIPPED;
}
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)
{
int n;
for(n=1 ; n<size ; n++)
{
int i, var[2];
MPI_Status status[3];
starpu_data_handle_t handle[2];
FPRINTF_MPI(stderr, "receiving from node %d\n", n);
for(i=0 ; i<2 ; i++)
starpu_variable_data_register(&handle[i], STARPU_MAIN_RAM, (uintptr_t)&var[i], sizeof(var[i]));
starpu_mpi_recv(handle[0], n, 42, MPI_COMM_WORLD, &status[0]);
starpu_data_acquire(handle[0], STARPU_R);
STARPU_ASSERT_MSG(var[0] == n, "Received incorrect value <%d> from node <%d>\n", var[0], n);
FPRINTF_MPI(stderr, "received <%d> from node %d\n", var[0], n);
starpu_data_release(handle[0]);
starpu_mpi_recv(handle[0], n, 42, MPI_COMM_WORLD, &status[1]);
starpu_mpi_recv(handle[1], n, 44, MPI_COMM_WORLD, &status[2]);
for(i=0 ; i<2 ; i++)
starpu_data_acquire(handle[i], STARPU_R);
STARPU_ASSERT_MSG(var[0] == n*2, "Received incorrect value <%d> from node <%d>\n", var[0], n);
STARPU_ASSERT_MSG(var[1] == n*4, "Received incorrect value <%d> from node <%d>\n", var[0], n);
FPRINTF_MPI(stderr, "received <%d> and <%d> from node %d\n", var[0], var[1], n);
for(i=0 ; i<2 ; i++)
starpu_data_release(handle[i]);
for(i=0 ; i<2 ; i++)
starpu_data_unregister(handle[i]);
}
}
else
{
int i, var[3];
starpu_data_handle_t handle[3];
FPRINTF_MPI(stderr, "sending to node %d\n", 0);
var[0] = rank;
var[1] = var[0] * 2;
var[2] = var[0] * 4;
for(i=0 ; i<3 ; i++)
starpu_variable_data_register(&handle[i], STARPU_MAIN_RAM, (uintptr_t)&var[i], sizeof(var[i]));
starpu_mpi_send(handle[0], 0, 42, MPI_COMM_WORLD);
starpu_mpi_send(handle[1], 0, 42, MPI_COMM_WORLD);
starpu_mpi_send(handle[2], 0, 44, MPI_COMM_WORLD);
for(i=0 ; i<3 ; i++)
starpu_data_unregister(handle[i]);
}
starpu_mpi_shutdown();
starpu_shutdown();
MPI_Finalize();
return 0;
}
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)
{
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)
{
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 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)
{
MPI_Init(NULL, NULL);
int rank, size;
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 processes.\n");
MPI_Finalize();
return 0;
}
/* We only use 2 nodes for that test */
if (rank >= 2)
{
MPI_Finalize();
return 0;
}
starpu_init(NULL);
starpu_mpi_initialize();
/* Node 0 will allocate a big block and only register an inner part of
* it as the block data, Node 1 will allocate a block of small size and
* register it directly. Node 0 and 1 will then exchange the content of
* their blocks. */
float *block;
starpu_data_handle block_handle;
if (rank == 0)
{
block = calloc(BIGSIZE*BIGSIZE*BIGSIZE, sizeof(float));
assert(block);
/* fill the inner block */
unsigned i, j, k;
for (k = 0; k < SIZE; k++)
for (j = 0; j < SIZE; j++)
for (i = 0; i < SIZE; i++)
{
block[i + j*BIGSIZE + k*BIGSIZE*BIGSIZE] = 1.0f;
}
starpu_block_data_register(&block_handle, 0,
(uintptr_t)block, BIGSIZE, BIGSIZE*BIGSIZE,
SIZE, SIZE, SIZE, sizeof(float));
}
else /* rank == 1 */
{
block = calloc(SIZE*SIZE*SIZE, sizeof(float));
assert(block);
starpu_block_data_register(&block_handle, 0,
(uintptr_t)block, SIZE, SIZE*SIZE,
SIZE, SIZE, SIZE, sizeof(float));
}
if (rank == 0)
{
starpu_mpi_send(block_handle, 1, 0x42, MPI_COMM_WORLD);
MPI_Status status;
starpu_mpi_recv(block_handle, 1, 0x1337, MPI_COMM_WORLD, &status);
/* check the content of the block */
starpu_data_acquire(block_handle, STARPU_R);
unsigned i, j, k;
for (k = 0; k < SIZE; k++)
for (j = 0; j < SIZE; j++)
for (i = 0; i < SIZE; i++)
{
assert(block[i + j*BIGSIZE + k*BIGSIZE*BIGSIZE] == 33.0f);
}
starpu_data_release(block_handle);
}
else /* rank == 1 */
{
MPI_Status status;
starpu_mpi_recv(block_handle, 0, 0x42, MPI_COMM_WORLD, &status);
/* check the content of the block and modify it */
starpu_data_acquire(block_handle, STARPU_RW);
unsigned i, j, k;
for (k = 0; k < SIZE; k++)
for (j = 0; j < SIZE; j++)
for (i = 0; i < SIZE; i++)
{
assert(block[i + j*SIZE + k*SIZE*SIZE] == 1.0f);
block[i + j*SIZE + k*SIZE*SIZE] = 33.0f;
}
starpu_data_release(block_handle);
starpu_mpi_send(block_handle, 0, 0x1337, MPI_COMM_WORLD);
}
fprintf(stdout, "Rank %d is done\n", rank);
fflush(stdout);
starpu_mpi_shutdown();
starpu_shutdown();
MPI_Finalize();
return 0;
}