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)
{
int size;
int rank;
int ret=0;
int sdetached, rdetached;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (size < 2)
{
FPRINTF_MPI(stderr, "We need at least 2 processes.\n");
MPI_Finalize();
return STARPU_TEST_SKIPPED;
}
for(sdetached=0 ; sdetached<=1 ; sdetached++)
{
for(rdetached=0 ; rdetached<=1 ; rdetached++)
{
ret += do_test(rank, sdetached, rdetached);
}
}
MPI_Finalize();
return ret;
}
/*
* Codelet to perform the reduction of two elements
*/
void redux_cpu_func(void *descr[], void *cl_arg)
{
long int *dota = (long int *)STARPU_VARIABLE_GET_PTR(descr[0]);
long int *dotb = (long int *)STARPU_VARIABLE_GET_PTR(descr[1]);
*dota = *dota + *dotb;
FPRINTF_MPI(stderr, "Calling redux %ld=%ld+%ld\n", *dota, *dota-*dotb, *dotb);
}
void func_cpu(void *descr[], STARPU_ATTRIBUTE_UNUSED void *_args)
{
int *value = (int *)STARPU_VARIABLE_GET_PTR(descr[0]);
int rank;
starpu_codelet_unpack_args(_args, &rank);
FPRINTF_MPI(stderr, "Executing codelet with value %d and rank %d\n", *value, rank);
STARPU_ASSERT_MSG(*value == rank, "Received value %d is not the expected value %d\n", *value, rank);
}
void callback(void *arg)
{
unsigned *received = arg;
STARPU_PTHREAD_MUTEX_LOCK(&mutex);
*received = *received + 1;
FPRINTF_MPI(stderr, "Requests %d received\n", *received);
STARPU_PTHREAD_COND_SIGNAL(&cond);
STARPU_PTHREAD_MUTEX_UNLOCK(&mutex);
}
void func_cpu(void *descr[], STARPU_ATTRIBUTE_UNUSED void *_args)
{
int node;
int rank;
starpu_codelet_unpack_args(_args, &node);
starpu_mpi_comm_rank(MPI_COMM_WORLD, &rank);
FPRINTF_MPI(stderr, "Expected node: %d - Actual node: %d\n", node, rank);
assert(node == rank);
}
void check_variable(starpu_data_handle_t handle, int i, int rank, int *error)
{
int other_rank = rank%2 == 0 ? rank+1 : rank-1;
int *rvalue = (int *)starpu_data_get_local_ptr(handle);
if (*rvalue != i*other_rank)
{
FPRINTF_MPI(stderr, "Incorrect received value: %d != %d\n", *rvalue, i*other_rank);
*error = 1;
}
}
void check_complex(starpu_data_handle_t handle, int i, int rank, int *error)
{
double *real = starpu_complex_get_real(handle);
double *imaginary = starpu_complex_get_imaginary(handle);
int other_rank = rank%2 == 0 ? rank+1 : rank-1;
if ((*real != ((i*other_rank)+12)) || (*imaginary != ((i*other_rank)+45)))
{
FPRINTF_MPI(stderr, "Incorrect received value: %f != %d || %f != %d\n", *real, ((i*other_rank)+12), *imaginary, ((i*other_rank)+45));
*error = 1;
}
}
int exchange_void(int rank, int detached)
{
int ret, i;
starpu_data_handle_t tab_handle[NB];
FPRINTF_MPI(stderr, "Exchanging void data with detached=%d\n", detached);
for(i=0 ; i<NB ; i++)
{
starpu_void_data_register(&tab_handle[i]);
starpu_mpi_data_register(tab_handle[i], i, rank);
}
ret = exchange(rank, tab_handle, check_void, detached);
for(i=0 ; i<NB ; i++)
starpu_data_unregister(tab_handle[i]);
return ret;
}
int exchange_variable(int rank, int detached)
{
int ret, i;
starpu_data_handle_t tab_handle[NB];
int value[NB];
FPRINTF_MPI(stderr, "Exchanging variable data with detached=%d\n", detached);
for(i=0 ; i<NB ; i++)
{
value[i]=i*rank;
starpu_variable_data_register(&tab_handle[i], STARPU_MAIN_RAM, (uintptr_t)&value[i], sizeof(int));
starpu_mpi_data_register(tab_handle[i], i, rank);
}
ret = exchange(rank, tab_handle, check_variable, detached);
for(i=0 ; i<NB ; i++)
starpu_data_unregister(tab_handle[i]);
return ret;
}
int exchange_complex(int rank, int detached)
{
int ret, i;
starpu_data_handle_t handle[NB];
double real[NB];
double imaginary[NB];
FPRINTF_MPI(stderr, "Exchanging complex data with detached=%d\n", detached);
for(i=0 ; i<NB ; i++)
{
real[i] = (i*rank)+12;
imaginary[i] = (i*rank)+45;
starpu_complex_data_register(&handle[i], STARPU_MAIN_RAM, &real[i], &imaginary[i], 1);
starpu_mpi_data_register(handle[i], i, rank);
}
ret = exchange(rank, handle, check_complex, detached);
for(i=0 ; i<NB ; i++)
starpu_data_unregister(handle[i]);
return ret;
}
/*
* Display codelet
*/
void display_cpu_func(void *descr[], void *cl_arg)
{
long int *local_x = (long int *)STARPU_VECTOR_GET_PTR(descr[0]);
FPRINTF_MPI(stderr, "Local=%ld\n", *local_x);
}
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;
}
/*
* Codelet to create a neutral element
*/
void init_cpu_func(void *descr[], void *cl_arg)
{
long int *dot = (long int *)STARPU_VARIABLE_GET_PTR(descr[0]);
*dot = 0;
FPRINTF_MPI(stderr, "Init dot\n");
}
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 main(int argc, char **argv)
{
int size, n, x=789;
int rank, other_rank;
int ret;
starpu_data_handle_t data[2];
MPI_Init(&argc, &argv);
starpu_mpi_comm_rank(MPI_COMM_WORLD, &rank);
starpu_mpi_comm_size(MPI_COMM_WORLD, &size);
if (size % 2)
{
FPRINTF(stderr, "We need a even number of processes.\n");
MPI_Finalize();
return STARPU_TEST_SKIPPED;
}
other_rank = rank%2 == 0 ? rank+1 : rank-1;
FPRINTF_MPI(stderr, "rank %d exchanging with rank %d\n", rank, other_rank);
if (rank % 2)
{
MPI_Send(&rank, 1, MPI_INT, other_rank, 10, MPI_COMM_WORLD);
FPRINTF(stderr, "[%d] sending %d\n", rank, rank);
}
else
{
MPI_Recv(&x, 1, MPI_INT, other_rank, 10, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
FPRINTF(stderr, "[%d] received %d\n", rank, 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 % 2)
{
starpu_variable_data_register(&data[0], STARPU_MAIN_RAM, (uintptr_t)&rank, sizeof(unsigned));
starpu_variable_data_register(&data[1], STARPU_MAIN_RAM, (uintptr_t)&rank, sizeof(unsigned));
starpu_mpi_data_register(data[1], 22, 0);
}
else
starpu_variable_data_register(&data[0], -1, (uintptr_t)NULL, sizeof(unsigned));
starpu_mpi_data_register(data[0], 12, 0);
if (rank % 2)
{
starpu_mpi_req req;
starpu_mpi_issend(data[1], &req, other_rank, 22, MPI_COMM_WORLD);
starpu_mpi_send(data[0], other_rank, 12, MPI_COMM_WORLD);
starpu_mpi_wait(&req, MPI_STATUS_IGNORE);
}
else
{
int *xx;
starpu_mpi_recv(data[0], other_rank, 12, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
xx = (int *)starpu_variable_get_local_ptr(data[0]);
FPRINTF_MPI(stderr, "received %d\n", *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(unsigned));
starpu_mpi_data_register(data[1], 22, 0);
starpu_mpi_recv(data[0], other_rank, 22, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
xx = (int *)starpu_variable_get_local_ptr(data[0]);
STARPU_ASSERT_MSG(x==*xx, "Received value %d is incorrect (should be %d)\n", *xx, x);
}
starpu_data_unregister(data[0]);
starpu_data_unregister(data[1]);
starpu_mpi_shutdown();
starpu_shutdown();
MPI_Finalize();
return 0;
}
int exchange(int rank, starpu_data_handle_t *handles, check_func func, int detached)
{
int other_rank = rank%2 == 0 ? rank+1 : rank-1;
int i;
if (rank%2)
{
starpu_mpi_send(handles[0], other_rank, 0, MPI_COMM_WORLD);
starpu_mpi_send(handles[NB-1], other_rank, NB-1, MPI_COMM_WORLD);
for(i=1 ; i<NB-1 ; i++)
{
starpu_mpi_send(handles[i], other_rank, i, MPI_COMM_WORLD);
}
return 0;
}
else
{
int ret=0;
starpu_mpi_req req[NB];
int received = 0;
if (detached)
{
starpu_mpi_irecv_detached(handles[0], other_rank, 0, MPI_COMM_WORLD, callback, &received);
}
else
{
memset(req, 0, NB*sizeof(starpu_mpi_req));
starpu_mpi_irecv(handles[0], &req[0], other_rank, 0, MPI_COMM_WORLD);
STARPU_ASSERT(req[0] != NULL);
}
// We sleep to make sure that the data for the tag 9 will be received before the recv is posted
usleep(2000000);
for(i=1 ; i<NB ; i++)
{
if (detached)
{
starpu_mpi_irecv_detached(handles[i], other_rank, i, MPI_COMM_WORLD, callback, &received);
}
else
{
starpu_mpi_irecv(handles[i], &req[i], other_rank, i, MPI_COMM_WORLD);
STARPU_ASSERT(req[i] != NULL);
}
}
if (detached)
{
STARPU_PTHREAD_MUTEX_LOCK(&mutex);
while (received != NB)
{
FPRINTF_MPI(stderr, "Received %d messages\n", received);
STARPU_PTHREAD_COND_WAIT(&cond, &mutex);
}
STARPU_PTHREAD_MUTEX_UNLOCK(&mutex);
}
else
{
for(i=0 ; i<NB ; i++)
{
starpu_mpi_wait(&req[i], MPI_STATUS_IGNORE);
func(handles[i], i, rank, &ret);
}
}
return ret;
}
}
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 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 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;
}
