static void dispatch_rts (
pami_context_t context, /**< IN: PAMI context */
void * cookie, /**< IN: dispatch cookie */
const void * header_addr, /**< IN: header address */
size_t header_size, /**< IN: header size */
const void * pipe_addr, /**< IN: address of PAMI pipe buffer */
size_t pipe_size, /**< IN: size of PAMI pipe buffer */
pami_endpoint_t origin,
pami_recv_t * recv) /**< OUT: receive message structure */
{
volatile size_t * active = (volatile size_t *) cookie;
fprintf (stderr, ">> 'rts' dispatch function. cookie = %p (active: %zu), header_size = %zu, pipe_size = %zu, recv = %p\n", cookie, *active, header_size, pipe_size, recv);
rts_info_t * rts = (rts_info_t *) header_addr;
fprintf (stderr, " 'rts' dispatch function. rts->origin = 0x%08x, rts->bytes = %zu\n", rts->origin, rts->bytes);
/*assert(pipe_addr!=NULL); */
/*pami_memregion_t * origin_memregion = (pami_memregion_t *) pipe_addr; */
get_info_t * get = (get_info_t *) malloc (sizeof(get_info_t));
get->value = active;
get->origin = rts->origin;
get->bytes = rts->bytes;
get->pad = 16;
initialize_data (get->buffer, 0, 6);
print_data (get->buffer, 12*4);
/* Create a memregion for the data buffer. */
size_t bytes = 0;
pami_result_t pami_rc = PAMI_Memregion_create (context, get->buffer, 12*4, &bytes, &(get->memregion));
if (PAMI_SUCCESS != pami_rc) {
fprintf (stderr, "PAMI_Memregion_create failed with rc = %d\n", pami_rc) ;
exit(1);
}
/* Perform the rdma get operation */
pami_rget_simple_t parameters;
parameters.rma.dest = rts->origin;
parameters.rma.bytes = rts->bytes;
parameters.rma.cookie = get;
parameters.rma.done_fn = get_done;
parameters.rdma.local.mr = &(get->memregion);
parameters.rdma.local.offset = 16;
parameters.rdma.remote.mr = &(rts->memregion);
parameters.rdma.remote.offset = 0;
fprintf (stderr, " 'rts' dispatch function. Before PAMI_Rget()\n");
pami_result_t status = PAMI_Rget (context, ¶meters);
fprintf (stderr, " 'rts' dispatch function. After PAMI_Rget(), status = %d\n", status);
if (status != PAMI_SUCCESS)
get_done (context, (void *) get, status);
fprintf (stderr, "<< 'rts' dispatch function.\n");
return;
}
int
MPIDI_Win_allgather( MPI_Aint size, MPID_Win **win_ptr )
{
int mpi_errno = MPI_SUCCESS;
MPID_Win *win;
int rank;
MPID_Comm *comm_ptr;
size_t length_out = 0;
pami_result_t rc;
MPIDI_Win_info *winfo;
static char FCNAME[] = "MPIDI_Win_allgather";
win = *win_ptr;
comm_ptr = win->comm_ptr;
rank = comm_ptr->rank;
winfo = &win->mpid.info[rank];
if (size != 0 && win->create_flavor != MPI_WIN_FLAVOR_SHARED)
{
#ifndef USE_PAMI_RDMA
if (!MPIDI_Process.mp_s_use_pami_get)
{
#endif
/* --------------------------------------- */
/* Setup the PAMI sections of the window */
/* --------------------------------------- */
rc = PAMI_Memregion_create(MPIDI_Context[0], win->mpid.info[rank].base_addr, win->size, &length_out, &winfo->memregion);
#ifdef USE_PAMI_RDMA
MPIU_ERR_CHKANDJUMP((rc != PAMI_SUCCESS), mpi_errno, MPI_ERR_OTHER, "**nomem");
MPIU_ERR_CHKANDJUMP((win->size < length_out), mpi_errno, MPI_ERR_OTHER, "**nomem");
#else
if (rc == PAMI_SUCCESS)
{
winfo->memregion_used = 1;
MPID_assert(win->size == length_out);
}
}
#endif
}
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE,
0,
MPI_DATATYPE_NULL,
win->mpid.info,
sizeof(struct MPIDI_Win_info),
MPI_BYTE,
comm_ptr,
&mpi_errno);
fn_fail:
return mpi_errno;
}
int main (int argc, char ** argv)
{
volatile size_t _rts_active = 1;
volatile size_t _ack_active = 1;
memset(&null_send_hint, 0, sizeof(null_send_hint));
pami_client_t client;
pami_context_t context[2];
char cl_string[] = "TEST";
pami_result_t result = PAMI_ERROR;
result = PAMI_Client_create (cl_string, &client, NULL, 0);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami client. result = %d\n", result);
return 1;
}
#ifdef TEST_CROSSTALK
size_t num = 2;
#else
size_t num = 1;
#endif
result = PAMI_Context_createv(client, NULL, 0, context, num);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami context(s). result = %d\n", result);
return 1;
}
pami_configuration_t configuration;
configuration.name = PAMI_CLIENT_TASK_ID;
result = PAMI_Client_query(client, &configuration,1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
pami_task_t task_id = configuration.value.intval;
fprintf (stderr, "My task id = %d\n", task_id);
configuration.name = PAMI_CLIENT_NUM_TASKS;
result = PAMI_Client_query(client, &configuration,1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
size_t num_tasks = configuration.value.intval;
fprintf (stderr, "Number of tasks = %zu\n", num_tasks);
if (num_tasks < 2)
{
fprintf (stderr, "Error. This test requires at least 2 tasks. Number of tasks in this job: %zu\n", num_tasks);
return 1;
}
pami_dispatch_hint_t options={};
#ifdef USE_SHMEM_OPTION
options.use_shmem = PAMI_HINT_ENABLE;
fprintf (stderr, "##########################################\n");
fprintf (stderr, "shared memory optimizations forced ON\n");
fprintf (stderr, "##########################################\n");
#elif defined(NO_SHMEM_OPTION)
options.use_shmem = PAMI_HINT_DISABLE;
fprintf (stderr, "##########################################\n");
fprintf (stderr, "shared memory optimizations forced OFF\n");
fprintf (stderr, "##########################################\n");
#endif
size_t i = 0;
#ifdef TEST_CROSSTALK
for (i=0; i<2; i++)
#endif
{
pami_dispatch_callback_function fn;
fprintf (stderr, "Before PAMI_Dispatch_set(%d) .. &_rts_active = %p, _rts_active = %zu\n", DISPATCH_ID_RTS, &_rts_active, _rts_active);
fn.p2p = dispatch_rts;
result = PAMI_Dispatch_set (context[i],
DISPATCH_ID_RTS,
fn,
(void *)&_rts_active,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami dispatch. result = %d\n", result);
return 1;
}
fprintf (stderr, "Before PAMI_Dispatch_set(%d) .. &_ack_active = %p, _ack_active = %zu\n", DISPATCH_ID_ACK, &_ack_active, _ack_active);
fn.p2p = dispatch_ack;
result = PAMI_Dispatch_set (context[i],
DISPATCH_ID_ACK,
fn,
(void *)&_ack_active,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami dispatch. result = %d\n", result);
return 1;
}
}
if (task_id == 0)
{
pami_send_immediate_t parameters;
#ifdef TEST_CROSSTALK
fprintf (stdout, "PAMI_Rget('simple') functional test [crosstalk]\n");
fprintf (stdout, "\n");
PAMI_Endpoint_create (client, num_tasks-1, 1, ¶meters.dest);
#else
fprintf (stdout, "PAMI_Rget('simple') functional test\n");
fprintf (stdout, "\n");
PAMI_Endpoint_create (client, num_tasks-1, 0, ¶meters.dest);
#endif
/* Allocate some memory from the heap. */
void * send_buffer = malloc (BUFFERSIZE);
/* Initialize the memory for validation. */
initialize_data ((uint32_t *)send_buffer, BUFFERSIZE, 0);
print_data (send_buffer, BUFFERSIZE);
/* Send an 'rts' message to the target task and provide the memory region */
rts_info_t rts_info;
PAMI_Endpoint_create (client, 0, 0, &rts_info.origin);
rts_info.bytes = BUFFERSIZE;
/* Create a memory region for this memoru buffer */
size_t bytes = 0;
pami_result_t pami_rc = PAMI_Memregion_create (context[0], send_buffer, BUFFERSIZE, &bytes, &(rts_info.memregion));
if (PAMI_SUCCESS != pami_rc) {
fprintf (stderr, "PAMI_Memregion_create failed with rc = %d\n", pami_rc) ;
exit(1);
}
parameters.dispatch = DISPATCH_ID_RTS;
parameters.header.iov_base = &rts_info;
parameters.header.iov_len = sizeof(rts_info_t);
parameters.data.iov_base = NULL;
parameters.data.iov_len = 0;
fprintf (stderr, "Before PAMI_Send_immediate()\n");
PAMI_Send_immediate (context[0], ¶meters);
/* wait for the 'ack' */
fprintf (stderr, "Wait for 'ack', _ack_active = %zu\n", _ack_active);
while (_ack_active != 0)
{
result = PAMI_Context_advance (context[0], 100);
if (result != PAMI_SUCCESS && result != PAMI_EAGAIN)
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
/* Destroy the local memory region */
PAMI_Memregion_destroy (context[0], &(rts_info.memregion));
free (send_buffer);
switch (_ack_status)
{
case 0:
fprintf (stdout, "Test PASSED\n");
break;
case 1:
fprintf (stdout, "Test FAILED (rget error)\n");
break;
case 2:
fprintf (stdout, "Test FAILED (data error)\n");
break;
default:
fprintf (stdout, "Test FAILED (unknown error)\n");
break;
}
}
else if (task_id == num_tasks-1)
{
#ifdef TEST_CROSSTALK
size_t contextid = 1;
#else
size_t contextid = 0;
#endif
/* wait for the 'rts' */
fprintf (stderr, "Wait for 'rts', _rts_active = %zu, contextid = %zu\n", _rts_active, contextid);
while (_rts_active != 0)
{
result = PAMI_Context_advance (context[contextid], 100);
if (result != PAMI_SUCCESS && result != PAMI_EAGAIN)
{
fprintf (stderr, "Error. Unable to advance pami context. result = %d\n", result);
return 1;
}
}
}
fprintf (stderr, "Test completed .. cleanup\n");
result = PAMI_Context_destroyv (context, num);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami context. result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy(&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami client. result = %d\n", result);
return 1;
}
/*fprintf (stdout, "Success (%d)\n", task_id); */
return 0;
};
int main(int argc, char* argv[])
{
pami_result_t result = PAMI_ERROR;
/* initialize the second client */
char * clientname = "";
pami_client_t client;
result = PAMI_Client_create(clientname, &client, NULL, 0);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_create");
/* query properties of the client */
pami_configuration_t config[4];
config[0].name = PAMI_CLIENT_NUM_TASKS;
config[1].name = PAMI_CLIENT_TASK_ID;
config[2].name = PAMI_CLIENT_NUM_CONTEXTS;
config[3].name = PAMI_CLIENT_NUM_LOCAL_TASKS;
result = PAMI_Client_query(client, config, 4);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
const size_t world_size = config[0].value.intval;
const size_t world_rank = config[1].value.intval;
const size_t num_contexts = config[2].value.intval;
const size_t num_local_tasks = config[3].value.intval;
TEST_ASSERT(num_contexts>1,"num_contexts>1");
int ppn = (int)num_local_tasks;
int nnodes = world_size/ppn;
int mycore = world_size%nnodes;
int mynode = (world_rank-mycore)/ppn;
if (world_rank==0)
{
printf("hello world from rank %ld of %ld, node %d of %d, core %d of %d \n",
world_rank, world_size, mynode, nnodes, mycore, ppn );
fflush(stdout);
}
/* initialize the contexts */
contexts = (pami_context_t *) safemalloc( num_contexts * sizeof(pami_context_t) );
result = PAMI_Context_createv( client, NULL, 0, contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_createv");
/* setup the world geometry */
pami_geometry_t world_geometry;
result = PAMI_Geometry_world(client, &world_geometry );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Geometry_world");
#ifdef PROGRESS_THREAD
int status = pthread_create(&Progress_thread, NULL, &Progress_function, NULL);
TEST_ASSERT(status==0, "pthread_create");
#endif
/************************************************************************/
for (int n=1; n<=(256*1024); n*=2)
{
if (world_rank==0)
{
printf("starting n = %d \n", n);
fflush(stdout);
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
double * sbuf = safemalloc(world_size*n*sizeof(double));
double * rbuf = safemalloc(world_size*n*sizeof(double));
for (int s=0; s<world_size; s++ )
for (int k=0; k<n; k++)
sbuf[s*n+k] = world_rank*n+k;
for (int s=0; s<world_size; s++ )
for (int k=0; k<n; k++)
rbuf[s*n+k] = -1.0;
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
size_t bytes = world_size * n * sizeof(double), bytes_out;
pami_memregion_t shared_mr;
result = PAMI_Memregion_create(contexts[1], rbuf, bytes, &bytes_out, &shared_mr);
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
pami_memregion_t local_mr;
result = PAMI_Memregion_create(contexts[0], sbuf, bytes, &bytes_out, &local_mr);
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
pami_endpoint_t * target_eps = (pami_endpoint_t *) safemalloc( world_size * sizeof(pami_endpoint_t) );
for (int target=0; target<world_size; target++)
{
result = PAMI_Endpoint_create(client, (pami_task_t) target, 1 /* async context*/, &(target_eps[target]) );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Endpoint_create");
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
pami_memregion_t * shmrs = (pami_memregion_t *) safemalloc( world_size * sizeof(pami_memregion_t) );
result = allgather(world_geometry, contexts[0], sizeof(pami_memregion_t), &shared_mr, shmrs);
TEST_ASSERT(result == PAMI_SUCCESS,"allgather");
if (world_rank==0)
{
printf("starting A2A \n");
fflush(stdout);
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
#ifdef SEPARATE_COMPLETION
done_t active = { .local = world_size,
.remote = world_size };
#else
int active = world_size;
#endif
uint64_t t0 = GetTimeBase();
for (int count=0; count<world_size; count++)
{
int t = world_rank+count;
int target = t%world_size;
//printf("%ld: attempting Rput to %ld (bytes=%ld,loff=%ld, roff=%ld) \n",
// (long)world_rank, (long)target, bytes, n*sizeof(double),
// target*n*sizeof(double), world_rank*n*sizeof(double));
//printf("%ld: attempting Rput to %ld \n", (long)world_rank, (long)target),
//fflush(stdout);
pami_rput_simple_t parameters;
parameters.rma.dest = target_eps[target];
//parameters.rma.hints = ;
parameters.rma.bytes = n*sizeof(double);
parameters.rma.cookie = &active;
#ifdef SEPARATE_COMPLETION
parameters.rma.done_fn = cb_done_local;
parameters.put.rdone_fn = cb_done_remote;
#else
parameters.rma.done_fn = NULL;
parameters.put.rdone_fn = cb_done;
#endif
parameters.rdma.local.mr = &local_mr;
parameters.rdma.local.offset = target*n*sizeof(double);
parameters.rdma.remote.mr = &shmrs[target];
parameters.rdma.remote.offset = world_rank*n*sizeof(double);
result = PAMI_Rput(contexts[0], ¶meters);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Rput");
}
#ifdef SEPARATE_COMPLETION
while (active.local>0)
{
result = PAMI_Context_trylock_advancev(&(contexts[0]), 1, 1000);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_trylock_advancev");
}
#endif
uint64_t t1 = GetTimeBase();
double dt1 = (t1-t0)*tic;
#ifdef SEPARATE_COMPLETION
while (active.remote>0)
#else
while (active>0)
#endif
{
result = PAMI_Context_trylock_advancev(&(contexts[0]), 1, 1000);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_trylock_advancev");
}
uint64_t t2 = GetTimeBase();
double dt2 = (t2-t0)*tic;
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
double megabytes = 1.e-6*bytes;
printf("%ld: PAMI_Rput A2A: %ld bytes per rank, local %lf seconds (%lf MB/s), remote %lf seconds (%lf MB/s) \n",
(long)world_rank, n*sizeof(double),
dt1, megabytes/dt1,
dt2, megabytes/dt2 );
fflush(stdout);
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
for (int s=0; s<world_size; s++ )
for (int k=0; k<n; k++)
{
if (rbuf[s*n+k]!=(1.0*s*n+1.0*k))
printf("%4d: rbuf[%d] = %lf (%lf) \n", (int)world_rank, s*n+k, rbuf[s*n+k], (1.0*s*n+1.0*k) );
}
fflush(stdout);
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
result = PAMI_Memregion_destroy(contexts[0], &shared_mr);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
result = PAMI_Memregion_destroy(contexts[0], &local_mr);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
free(target_eps);
free(shmrs);
free(rbuf);
free(sbuf);
}
/************************************************************************/
#ifdef PROGRESS_THREAD
void * rv;
status = pthread_cancel(Progress_thread);
TEST_ASSERT(status==0, "pthread_cancel");
status = pthread_join(Progress_thread, &rv);
TEST_ASSERT(status==0, "pthread_join");
#endif
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
/* finalize the contexts */
result = PAMI_Context_destroyv( contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_destroyv");
free(contexts);
/* finalize the client */
result = PAMI_Client_destroy( &client );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_destroy");
if (world_rank==0)
printf("%ld: end of test \n", world_rank );
fflush(stdout);
return 0;
}
int main(int argc, char ** argv)
{
pami_client_t client;
pami_context_t context;
pami_result_t status = PAMI_ERROR;
pami_configuration_t pami_config;
pami_geometry_t world_geo;
size_t barrier_alg_num[2];
pami_algorithm_t* bar_always_works_algo = NULL;
pami_metadata_t* bar_always_works_md = NULL;
pami_algorithm_t* bar_must_query_algo = NULL;
pami_metadata_t* bar_must_query_md = NULL;
pami_xfer_t barrier;
int my_id;
volatile int is_fence_done = 0;
volatile int is_barrier_done = 0;
/* create PAMI client */
RC( PAMI_Client_create("TEST", &client, NULL, 0) );
DBG_FPRINTF((stderr,"Client created successfully at 0x%p\n",client));
/* create PAMI context */
RC( PAMI_Context_createv(client, NULL, 0, &context, 1) );
DBG_FPRINTF((stderr,"Context created successfully at 0x%p\n",context));
/* query my task id */
bzero(&pami_config, sizeof(pami_configuration_t));
pami_config.name = PAMI_CLIENT_TASK_ID;
RC( PAMI_Client_query(client, &pami_config, 1) );
my_id = pami_config.value.intval;
DBG_FPRINTF((stderr,"My task id is %d\n", my_id));
/* get the world geometry */
RC( PAMI_Geometry_world(client, &world_geo) );
DBG_FPRINTF((stderr,"World geometry is at 0x%p\n",world_geo));
/* query number of barrier algorithms */
RC( PAMI_Geometry_algorithms_num(world_geo, PAMI_XFER_BARRIER,
barrier_alg_num) );
DBG_FPRINTF((stderr,"%d-%d algorithms are available for barrier op\n",
barrier_alg_num[0], barrier_alg_num[1]));
if (barrier_alg_num[0] <= 0) {
fprintf (stderr, "Error. No (%lu) algorithm is available for barrier op\n",
barrier_alg_num[0]);
return 1;
}
/* query barrier algorithm list */
bar_always_works_algo =
(pami_algorithm_t*)malloc(sizeof(pami_algorithm_t)*barrier_alg_num[0]);
bar_always_works_md =
(pami_metadata_t*)malloc(sizeof(pami_metadata_t)*barrier_alg_num[0]);
bar_must_query_algo =
(pami_algorithm_t*)malloc(sizeof(pami_algorithm_t)*barrier_alg_num[1]);
bar_must_query_md =
(pami_metadata_t*)malloc(sizeof(pami_metadata_t)*barrier_alg_num[1]);
RC( PAMI_Geometry_algorithms_query(world_geo, PAMI_XFER_BARRIER,
bar_always_works_algo, bar_always_works_md, barrier_alg_num[0],
bar_must_query_algo, bar_must_query_md, barrier_alg_num[1]) );
DBG_FPRINTF((stderr,"Algorithm [%s] at 0x%p will be used for barrier op\n",
bar_always_works_md[0].name, bar_always_works_algo[0]));
/* begin PAMI fence */
RC( PAMI_Fence_begin(context) );
DBG_FPRINTF((stderr,"PAMI fence begins\n"));
/* ------------------------------------------------------------------------ */
pami_extension_t extension;
const char ext_name[] = "EXT_hfi_extension";
const char sym_name[] = "hfi_remote_update";
hfi_remote_update_fn remote_update = NULL;
hfi_remote_update_info_t remote_info;
pami_memregion_t mem_region;
size_t mem_region_sz = 0;
unsigned long long operand = 1234;
unsigned long long orig_val = 0;
int offset = (operand)%MAX_TABLE_SZ;
/* initialize table for remote update operation */
int i;
for (i = 0; i < MAX_TABLE_SZ; i ++) {
table[i] = (unsigned long long) i;
}
orig_val = table[offset];
/* open PAMI extension */
RC( PAMI_Extension_open (client, ext_name, &extension) );
DBG_FPRINTF((stderr,"Open %s successfully.\n", ext_name));
/* load PAMI extension function */
remote_update = (hfi_remote_update_fn)
PAMI_Extension_symbol (extension, sym_name);
if (remote_update == (void *)NULL)
{
fprintf (stderr, "Error. Failed to load %s function in %s\n",
sym_name, ext_name);
return 1;
} else {
DBG_FPRINTF((stderr,"Loaded function %s in %s successfully.\n",
sym_name, ext_name));
}
/* create a memory region for remote update operation */
RC( PAMI_Memregion_create(context, table,
MAX_TABLE_SZ*sizeof(unsigned long long),
&mem_region_sz, &mem_region) );
DBG_FPRINTF((stderr,"%d-byte PAMI memory region created successfully.\n",
mem_region_sz));
/* perform a PAMI barrier */
is_barrier_done = 0;
barrier.cb_done = barrier_done;
barrier.cookie = (void*)&is_barrier_done;
barrier.algorithm = bar_always_works_algo[0];
RC( PAMI_Collective(context, &barrier) );
DBG_FPRINTF((stderr,"PAMI barrier op invoked successfully.\n"));
while (is_barrier_done == 0)
PAMI_Context_advance(context, 1000);
DBG_FPRINTF((stderr,"PAMI barrier op finished successfully.\n"));
RC( PAMI_Context_lock(context) );
/* prepare remote update info */
remote_info.dest = my_id^1;
remote_info.op = 0; /* op_add */
remote_info.atomic_operand = operand;
remote_info.dest_buf = (unsigned long long)(&(table[offset]));
/* invoke remote update PAMI extension function */
RC( remote_update(context, 1, &remote_info) );
DBG_FPRINTF((stderr,"Function %s invoked successfully.\n",
sym_name));
RC( PAMI_Context_unlock(context) );
/* perform a PAMI fence */
is_fence_done = 0;
RC( PAMI_Fence_all(context, fence_done, (void*)&is_fence_done) );
DBG_FPRINTF((stderr,"PAMI_Fence_all invoked successfully.\n"));
while (is_fence_done == 0)
PAMI_Context_advance(context, 1000);
DBG_FPRINTF((stderr,"PAMI_Fence_all finished successfully.\n"));
/* perform a PAMI barrier */
is_barrier_done = 0;
barrier.cb_done = barrier_done;
barrier.cookie = (void*)&is_barrier_done;
barrier.algorithm = bar_always_works_algo[0];
RC( PAMI_Collective(context, &barrier) );
DBG_FPRINTF((stderr,"PAMI barrier op invoked successfully.\n"));
while (is_barrier_done == 0)
PAMI_Context_advance(context, 1000);
DBG_FPRINTF((stderr,"PAMI barrier op finished successfully.\n"));
/* verify data after remote update operation */
if (table[offset] != orig_val + operand) {
printf("Data verification at offset %d with operand %lu failed: "
"[%lu expected with %lu updated]\n",
offset, operand, orig_val+operand, table[offset]);
} else {
printf("Data verification at offset %d with operand %lu passed: "
"[%lu expected with %lu updated].\n",
offset, operand, orig_val+operand, table[offset]);
}
/* destroy the memory region after remote update operation */
RC( PAMI_Memregion_destroy(context, &mem_region) );
DBG_FPRINTF((stderr,"PAMI memory region removed successfully.\n"));
/* close PAMI extension */
RC( PAMI_Extension_close (extension) );
DBG_FPRINTF((stderr,"Close %s successfully.\n", ext_name));
/* ------------------------------------------------------------------------ */
/* end PAMI fence */
RC( PAMI_Fence_end(context) );
DBG_FPRINTF((stderr,"PAMI fence ends\n"));
/* destroy PAMI context */
RC( PAMI_Context_destroyv(&context, 1) );
DBG_FPRINTF((stderr, "PAMI context destroyed successfully\n"));
/* destroy PAMI client */
RC( PAMI_Client_destroy(&client) );
DBG_FPRINTF((stderr, "PAMI client destroyed successfully\n"));
return 0;
}
void test_fn (int argc, char * argv[], pami_client_t client, pami_context_t context[])
{
int num_doubles = 16;
if (argc > 1) num_doubles = atoi (argv[1]);
size_t num_tasks = size (client);
pami_task_t my_task_id = task (client);
pami_task_t target_task_id = num_tasks - 1;
pami_task_t origin_task_id = 0;
/*
* Allocate a 'window' of memory region information, one for each task in the
* client. Only the 'local' memory region information for this task will
* contain a valid data buffer. The memory region information is marked
* 'active' when the memory regions are received from each remote task.
*/
memregion_information_t * mr_info =
(memregion_information_t *) malloc (sizeof(memregion_information_t) * num_tasks);
unsigned i;
for (i = 0; i < num_tasks; i++)
{
mr_info[i].data.iov_len = 0;
mr_info[i].data.iov_base = NULL;
mr_info[i].active = 0;
}
/*
* Create a local memregion for each context.
*
* Note that both memregions will describe the same memory location. This is
* necessary when writing portable, platform independent code as the physical
* hardware underlying the contexts may, or may not, require separate memory
* pinning.
*/
size_t actual_memregion_bytes = 0;
mr_info[my_task_id].data.iov_base = malloc (sizeof(double) * num_doubles);
mr_info[my_task_id].data.iov_len = sizeof(double) * num_doubles;
PAMI_Memregion_create (context[0],
mr_info[my_task_id].data.iov_base,
mr_info[my_task_id].data.iov_len,
& actual_memregion_bytes,
& mr_info[my_task_id].memregion[0]);
PAMI_Memregion_create (context[1],
mr_info[my_task_id].data.iov_base,
mr_info[my_task_id].data.iov_len,
& actual_memregion_bytes,
& mr_info[my_task_id].memregion[1]);
mr_info[my_task_id].active = 1;
/*
* Register the memory region exchange dispatch; only needed on the
* first context of each task.
*/
pami_dispatch_hint_t mr_hint = {0};
pami_dispatch_callback_function mr_dispatch;
mr_dispatch.p2p = exchange_memregion_recv_cb;
PAMI_Dispatch_set (context[0], MEMREGION_EXCHANGE_DISPATCH_ID, mr_dispatch, (void *) mr_info, mr_hint);
accumulate_test_information_t test_info;
test_info.data_buffer.iov_base = malloc (sizeof(double) * num_doubles);
test_info.data_buffer.iov_len = sizeof(double) * num_doubles;
test_info.scalar = 1.2;
test_info.data_fn[ACCUMULATE_TEST_SCALAR_SUM] = accumulate_scalar_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_SCALAR_SUM] = (void *) & test_info.scalar;
test_info.data_fn[ACCUMULATE_TEST_VECTOR_SUM] = accumulate_vector_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_SUM] = malloc (sizeof(double) * num_doubles);
test_info.data_fn[ACCUMULATE_TEST_SCALAR_SUBTRACT] = accumulate_scalar_subtract_data_function;
test_info.data_cookie[ACCUMULATE_TEST_SCALAR_SUBTRACT] = (void *) & test_info.scalar;
test_info.data_fn[ACCUMULATE_TEST_VECTOR_SUBTRACT] = accumulate_vector_subtract_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_SUBTRACT] = malloc (sizeof(double) * num_doubles);
test_info.data_fn[ACCUMULATE_TEST_VECTOR_MAX_SUM] = accumulate_vector_max_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_MAX_SUM] = malloc (sizeof(double) * num_doubles);
test_info.data_fn[ACCUMULATE_TEST_VECTOR_MIN_SUM] = accumulate_vector_min_sum_data_function;
test_info.data_cookie[ACCUMULATE_TEST_VECTOR_MIN_SUM] = malloc (sizeof(double) * num_doubles);
/*
* Register the accumulate dispatch; needed on both
* contexts to enable "crosstalk".
*/
pami_dispatch_hint_t acc_hint = {0};
acc_hint.recv_immediate = PAMI_HINT_DISABLE;
pami_dispatch_callback_function acc_dispatch;
acc_dispatch.p2p = accumulate_test_recv_cb;
PAMI_Dispatch_set (context[0], ACCUMULATE_TEST_DISPATCH_ID, acc_dispatch, (void *) & test_info, acc_hint);
PAMI_Dispatch_set (context[1], ACCUMULATE_TEST_DISPATCH_ID, acc_dispatch, (void *) & test_info, acc_hint);
simple_barrier(client, context[0]);
/*
* Exchange the memory regions
*/
volatile unsigned mr_exchange_active = 0;
pami_send_t mr_exchange_parameters = {0};
mr_exchange_parameters.send.dispatch = MEMREGION_EXCHANGE_DISPATCH_ID;
mr_exchange_parameters.send.header.iov_base = NULL;
mr_exchange_parameters.send.header.iov_len = 0;
mr_exchange_parameters.send.data.iov_base = (void *) mr_info[my_task_id].memregion;
mr_exchange_parameters.send.data.iov_len = sizeof(pami_memregion_t) * 2;
mr_exchange_parameters.events.cookie = (void *) & mr_exchange_active;
mr_exchange_parameters.events.local_fn = decrement;
for (i = 0; i < num_tasks; i++)
{
if (i == my_task_id) continue;
PAMI_Endpoint_create (client, i, 0, & mr_exchange_parameters.send.dest);
mr_exchange_active++;
PAMI_Send (context[0], & mr_exchange_parameters);
}
/*
* Advance until local memory regions have been sent and
* all memory regions have been received.
*/
unsigned num_memregions_active;
do
{
num_memregions_active = 0;
for (i = 0; i < num_tasks; i++)
num_memregions_active += mr_info[i].active;
PAMI_Context_advance (context[0], 1);
}
while (num_memregions_active < num_tasks);
while (mr_exchange_active > 0)
PAMI_Context_advance (context[0], 1);
#ifdef ASYNC_PROGRESS
async_progress_t async_progress;
async_progress_open (client, &async_progress);
async_progress_enable (&async_progress, context[1]);
#endif
if (my_task_id == target_task_id)
{
/*
* This is the "passive target" task.
*/
#ifdef ASYNC_PROGRESS
/*
* Do "something" besides communication for a little bit.
*/
sleep(1);
#else
/*
* Advance the second context for a little bit.
*/
fprintf (stdout, "(%03d) spoofing async progress\n", __LINE__);
for (i=0; i<10; i++)
{
fprintf (stdout, "(%03d) 'async progress context' advancing\n", __LINE__);
PAMI_Context_advance (context[1], 100000);
fprintf (stdout, "(%03d) 'async progress context' sleeping\n", __LINE__);
sleep(1);
}
#endif
}
else if (my_task_id == origin_task_id)
{
/*
* This is the "active origin" task.
*/
{
/*
* Use rdma put to initialize the remote buffer with the local data.
*/
volatile unsigned rput_active = 1;
pami_rput_simple_t rput_parameters = {0};
PAMI_Endpoint_create (client, target_task_id, 1, & rput_parameters.rma.dest);
rput_parameters.rma.bytes = num_doubles * sizeof(double);
rput_parameters.rdma.local.mr = mr_info[origin_task_id].memregion;
rput_parameters.rdma.local.offset = 0;
rput_parameters.rdma.remote.mr = mr_info[target_task_id].memregion;
rput_parameters.rdma.remote.offset = 0;
rput_parameters.put.rdone_fn = decrement;
rput_parameters.rma.cookie = (void *) & rput_active;
PAMI_Rput (context[0], & rput_parameters);
while (rput_active > 0)
PAMI_Context_advance (context[0], 1);
}
{
volatile unsigned send_active = 0;
accumulate_test_t test_id = ACCUMULATE_TEST_SCALAR_SUM;
pami_send_t send_parameters = {0};
PAMI_Endpoint_create (client, target_task_id, 1, & send_parameters.send.dest);
send_parameters.send.dispatch = ACCUMULATE_TEST_DISPATCH_ID;
send_parameters.send.header.iov_len = sizeof (accumulate_test_t);
send_parameters.send.header.iov_base = (void *) & test_id;
send_parameters.send.data.iov_base = test_info.data_buffer.iov_base;
send_parameters.send.data.iov_len = test_info.data_buffer.iov_len;
send_parameters.events.remote_fn = decrement;
send_parameters.events.cookie = (void *) & send_active;
for (test_id = ACCUMULATE_TEST_SCALAR_SUM; test_id < ACCUMULATE_TEST_COUNT; test_id++)
{
send_active = 1;
fprintf (stdout, "(%03d) sending data buffer for accumulate test \"%s\"\n", __LINE__, accumulate_test_name[test_id]);
PAMI_Send (context[0], & send_parameters);
fprintf (stdout, "(%03d) waiting for remote completion of data buffer sent for accumulate test \"%s\"\n", __LINE__, accumulate_test_name[test_id]);
while (send_active > 0)
PAMI_Context_advance (context[0], 1);
fprintf (stdout, "(%03d) data buffer received on remote for accumulate test \"%s\"\n", __LINE__, accumulate_test_name[test_id]);
}
}
{
/*
* Use rdma get to retrieve the remote buffer and compare results.
*/
volatile unsigned rget_active = 1;
pami_rget_simple_t rget_parameters = {0};
PAMI_Endpoint_create (client, target_task_id, 1, & rget_parameters.rma.dest);
rget_parameters.rma.done_fn = decrement;
rget_parameters.rma.cookie = (void *) & rget_active;
rget_parameters.rma.bytes = sizeof(double) * num_doubles;
rget_parameters.rdma.local.mr = mr_info[origin_task_id].memregion;
rget_parameters.rdma.local.offset = 0;
rget_parameters.rdma.remote.mr = mr_info[target_task_id].memregion;
rget_parameters.rdma.remote.offset = 0;
PAMI_Rget (context[0], & rget_parameters);
while (rget_active > 0)
PAMI_Context_advance (context[0], 1);
}
}
else
{
/*
* All other tasks, if any, do nothing and simply enter the barrier.
*/
}
simple_barrier (client, context[0]);
#ifdef ASYNC_PROGRESS
async_progress_disable (&async_progress, context[1]);
async_progress_close (&async_progress);
#endif
/*
* Do cleanup ?
*/
return;
}
int main(int argc, char* argv[])
{
pami_result_t result = PAMI_ERROR;
/* initialize the second client */
char * clientname = "";
pami_client_t client;
result = PAMI_Client_create(clientname, &client, NULL, 0);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_create");
/* query properties of the client */
pami_configuration_t config[3];
size_t num_contexts;
config[0].name = PAMI_CLIENT_NUM_TASKS;
config[1].name = PAMI_CLIENT_TASK_ID;
config[2].name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query(client, config, 3);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
world_size = config[0].value.intval;
world_rank = config[1].value.intval;
num_contexts = config[2].value.intval;
TEST_ASSERT(num_contexts>1,"num_contexts>1");
if (world_rank==0)
{
printf("hello world from rank %ld of %ld \n", world_rank, world_size );
fflush(stdout);
}
/* initialize the contexts */
contexts = (pami_context_t *) safemalloc( num_contexts * sizeof(pami_context_t) );
result = PAMI_Context_createv( client, NULL, 0, contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_createv");
/* setup the world geometry */
pami_geometry_t world_geometry;
result = PAMI_Geometry_world(client, &world_geometry );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Geometry_world");
int status = pthread_create(&Progress_thread, NULL, &Progress_function, NULL);
TEST_ASSERT(status==0, "pthread_create");
/************************************************************************/
int n = (argc>1 ? atoi(argv[1]) : 1000000);
size_t bytes = n * sizeof(int), bytes_out;
int * shared = (int *) safemalloc(bytes);
for (int i=0; i<n; i++)
shared[i] = -1;
pami_memregion_t shared_mr;
result = PAMI_Memregion_create(contexts[1], shared, bytes, &bytes_out, &shared_mr);
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
int * local = (int *) safemalloc(bytes);
for (int i=0; i<n; i++)
local[i] = world_rank;
pami_memregion_t local_mr;
result = PAMI_Memregion_create(contexts[0], local, bytes, &bytes_out, &local_mr);
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
pami_memregion_t * shmrs = (pami_memregion_t *) safemalloc( world_size * sizeof(pami_memregion_t) );
result = allgather(world_geometry, contexts[0], sizeof(pami_memregion_t), &shared_mr, shmrs);
TEST_ASSERT(result == PAMI_SUCCESS,"allgather");
int target = (world_rank>0 ? world_rank-1 : world_size-1);
pami_endpoint_t target_ep;
result = PAMI_Endpoint_create(client, (pami_task_t) target, 1, &target_ep);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Endpoint_create");
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
int active = 2;
pami_rput_simple_t parameters;
parameters.rma.dest = target_ep;
//parameters.rma.hints = ;
parameters.rma.bytes = bytes;
parameters.rma.cookie = &active;
parameters.rma.done_fn = cb_done;
parameters.rdma.local.mr = &local_mr;
parameters.rdma.local.offset = 0;
parameters.rdma.remote.mr = &shmrs[target];
parameters.rdma.remote.offset = 0;
parameters.put.rdone_fn = cb_done;
uint64_t t0 = GetTimeBase();
result = PAMI_Rput(contexts[0], ¶meters);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Rput");
while (active)
{
//result = PAMI_Context_advance( contexts[0], 100);
//TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_advance");
result = PAMI_Context_trylock_advancev(&(contexts[0]), 1, 1000);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_trylock_advancev");
}
uint64_t t1 = GetTimeBase();
uint64_t dt = t1-t0;
/* barrier on non-progressing context to make sure CHT does its job */
barrier(world_geometry, contexts[0]);
printf("%ld: PAMI_Rput of %ld bytes achieves %lf MB/s \n", (long)world_rank, bytes, 1.6e9*1e-6*(double)bytes/(double)dt );
fflush(stdout);
int errors = 0;
target = (world_rank<(world_size-1) ? world_rank+1 : 0);
for (int i=0; i<n; i++)
if (shared[i] != target)
errors++;
if (errors>0)
for (int i=0; i<n; i++)
if (shared[i] != target)
printf("%ld: shared[%d] = %d (%d) \n", (long)world_rank, i, shared[i], target);
else
printf("%ld: no errors :-) \n", (long)world_rank);
fflush(stdout);
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
result = PAMI_Memregion_destroy(contexts[0], &shared_mr);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
result = PAMI_Memregion_destroy(contexts[0], &local_mr);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
free(shmrs);
free(local);
free(shared);
/************************************************************************/
void * rv;
status = pthread_cancel(Progress_thread);
TEST_ASSERT(status==0, "pthread_cancel");
status = pthread_join(Progress_thread, &rv);
TEST_ASSERT(status==0, "pthread_join");
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
/* finalize the contexts */
result = PAMI_Context_destroyv( contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_destroyv");
free(contexts);
/* finalize the client */
result = PAMI_Client_destroy( &client );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_destroy");
if (world_rank==0)
printf("%ld: end of test \n", world_rank );
fflush(stdout);
return 0;
}
int main(int argc, char* argv[])
{
pami_result_t result = PAMI_ERROR;
/* initialize the second client */
char * clientname = "";
pami_client_t client;
result = PAMI_Client_create(clientname, &client, NULL, 0);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_create");
/* query properties of the client */
pami_configuration_t config[4];
config[0].name = PAMI_CLIENT_NUM_TASKS;
config[1].name = PAMI_CLIENT_TASK_ID;
config[2].name = PAMI_CLIENT_NUM_CONTEXTS;
config[3].name = PAMI_CLIENT_NUM_LOCAL_TASKS;
result = PAMI_Client_query(client, config, 4);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
const size_t world_size = config[0].value.intval;
const size_t world_rank = config[1].value.intval;
const size_t num_contexts = (config[2].value.intval > 32) ? 32 : config[2].value.intval; /* because I only need 16+16 contexts in c1 mode */
const size_t num_local_tasks = config[3].value.intval;
TEST_ASSERT(num_contexts>1,"num_contexts>1");
const int ppn = (int)num_local_tasks;
const int nnodes = world_size/ppn;
const int mycore = world_size%nnodes;
const int mynode = (world_rank-mycore)/ppn;
const int num_sync = num_contexts/2;
const int num_async = num_contexts/2;
const int async_context_begin = num_sync+1;
const int async_context_end = num_contexts;
if (world_rank==0)
{
printf("hello world from rank %ld of %ld, node %d of %d, core %d of %d \n",
world_rank, world_size, mynode, nnodes, mycore, ppn );
printf("num_contexts = %ld, async_context_begin = %d, async_context_end = %d \n",
num_contexts, async_context_begin, async_context_end);
fflush(stdout);
}
/* initialize the contexts */
contexts = (pami_context_t *) safemalloc( num_contexts * sizeof(pami_context_t) );
result = PAMI_Context_createv( client, NULL, 0, contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_createv");
/* setup the world geometry */
pami_geometry_t world_geometry;
result = PAMI_Geometry_world(client, &world_geometry );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Geometry_world");
/************************************************************************/
for (int n=1; n<=(256*1024); n*=2)
{
if (world_rank==0)
{
printf("starting n = %d \n", n);
fflush(stdout);
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
double * sbuf = safemalloc(world_size*n*sizeof(double));
double * rbuf = safemalloc(world_size*n*sizeof(double));
for (int s=0; s<world_size; s++ )
for (int k=0; k<n; k++)
sbuf[s*n+k] = world_rank*n+k;
for (int s=0; s<world_size; s++ )
for (int k=0; k<n; k++)
rbuf[s*n+k] = -1.0;
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
size_t bytes = world_size * n * sizeof(double), bytes_out;
pami_memregion_t * local_mr = safemalloc(num_sync * sizeof(pami_memregion_t) );
pami_memregion_t * shared_mr = safemalloc(num_sync * sizeof(pami_memregion_t) );
for (int i=0; i<num_sync; i++)
{
result = PAMI_Memregion_create(contexts[i], rbuf, bytes, &bytes_out, &(local_mr[i]));
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
result = PAMI_Memregion_create(contexts[async_context_begin+i], sbuf, bytes, &bytes_out, &(shared_mr[i]));
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
pami_endpoint_t * target_eps = (pami_endpoint_t *) safemalloc( num_async * world_size * sizeof(pami_endpoint_t) );
for (int target=0; target<world_size; target++)
for (int i=0; i<num_async; i++)
{
result = PAMI_Endpoint_create(client, (pami_task_t) target, i, &(target_eps[target*num_async+i]) );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Endpoint_create");
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
pami_memregion_t * shmrs = (pami_memregion_t *) safemalloc( num_async * world_size * sizeof(pami_memregion_t) );
result = allgather(world_geometry, contexts[0], num_async * sizeof(pami_memregion_t), shared_mr, shmrs);
TEST_ASSERT(result == PAMI_SUCCESS,"allgather");
/* check now that count will not iterate over an incomplete iteration space */
int remote_targets_per_thread = world_size/num_sync;
assert((world_size%num_sync)==0);
if (world_rank==0)
{
printf("starting A2A \n");
fflush(stdout);
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
int active = world_size;
uint64_t t0 = GetTimeBase();
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
/* GCC prior to 4.7 will not permit const variables to be private i.e. firstprivate */
#ifdef _OPENMP
#pragma omp parallel default(shared) firstprivate(n, num_async, num_sync)
#endif
{
#ifdef _OPENMP
int tid = omp_get_thread_num();
#else
int tid = 0;
#endif
for (int count=0; count<remote_targets_per_thread; count++)
{
int target = remote_targets_per_thread*tid + count;
target += world_rank;
target = target % world_size;
//printf("%ld: attempting Rget to %ld \n", (long)world_rank, (long)target);
//fflush(stdout);
int local_context = tid; /* each thread uses its own context so this is thread-safe */
int remote_context = target % num_async;
pami_rget_simple_t parameters;
parameters.rma.dest = target_eps[target*num_async+remote_context];
//parameters.rma.hints = ;
parameters.rma.bytes = n*sizeof(double);
parameters.rma.cookie = &active;
parameters.rma.done_fn = cb_done;
parameters.rdma.local.mr = &local_mr[local_context];
parameters.rdma.local.offset = target*n*sizeof(double);
parameters.rdma.remote.mr = &shmrs[target*num_async+remote_context];
parameters.rdma.remote.offset = world_rank*n*sizeof(double);
result = PAMI_Rget(contexts[local_context], ¶meters);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Rget");
}
}
uint64_t t1 = GetTimeBase();
double dt1 = (t1-t0)*tic;
while (active>0)
{
result = PAMI_Context_trylock_advancev(&(contexts[0]), num_sync+num_async, 1000);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_trylock_advancev");
}
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
uint64_t t2 = GetTimeBase();
double dt2 = (t2-t0)*tic;
//result = barrier(world_geometry, contexts[0]);
//TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
double megabytes = 1.e-6*bytes;
printf("%ld: PAMI_Rget A2A: %ld bytes per rank, local %lf seconds (%lf MB/s), remote %lf seconds (%lf MB/s) \n",
(long)world_rank, n*sizeof(double),
dt1, megabytes/dt1,
dt2, megabytes/dt2 );
fflush(stdout);
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
for (int s=0; s<world_size; s++ )
for (int k=0; k<n; k++)
{
if (rbuf[s*n+k]!=(1.0*s*n+1.0*k))
printf("%4d: rbuf[%d] = %lf (%lf) \n", (int)world_rank, s*n+k, rbuf[s*n+k], (1.0*s*n+1.0*k) );
}
fflush(stdout);
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
for (int i=0; i<num_async; i++)
{
result = PAMI_Memregion_destroy(contexts[i], &(local_mr[i]) );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
result = PAMI_Memregion_destroy(contexts[async_context_begin+i], &(shared_mr[i]) );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
}
free(shared_mr);
free(local_mr);
free(target_eps);
free(shmrs);
free(rbuf);
free(sbuf);
}
/************************************************************************/
result = barrier(world_geometry, contexts[0]);
TEST_ASSERT(result == PAMI_SUCCESS,"barrier");
/* finalize the contexts */
result = PAMI_Context_destroyv( contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_destroyv");
free(contexts);
/* finalize the client */
result = PAMI_Client_destroy( &client );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_destroy");
if (world_rank==0)
printf("%ld: end of test \n", world_rank );
fflush(stdout);
return 0;
}
int main(int argc, char* argv[])
{
int status = MPI_SUCCESS;
pami_result_t result = PAMI_ERROR;
int provided = MPI_THREAD_SINGLE;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
/* IBM: --ranks-per-node 64 fails to init threads but this */
/* IBM: testcase doesn't really care so don't exit */
TEST_ASSERT((provided>=MPI_THREAD_MULTIPLE),"MPI_Init_thread");
/* initialize the second client */
char * clientname = "test"; /* IBM: PE PAMI requires a client name */
pami_client_t client;
result = PAMI_Client_create(clientname, &client, NULL, 0);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_create");
/* query properties of the client */
pami_configuration_t config[3];
size_t num_contexts;
config[0].name = PAMI_CLIENT_NUM_TASKS;
config[1].name = PAMI_CLIENT_TASK_ID;
config[2].name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query(client, config, 3);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
world_size = config[0].value.intval;
world_rank = config[1].value.intval;
num_contexts = config[2].value.intval;
if (world_rank==0)
printf("hello world from rank %ld of %ld, number of contexts %zu \n", world_rank, world_size, num_contexts );/*IBM: debug num_contexts */
fflush(stdout);
/* initialize the contexts */
pami_context_t * contexts;
contexts = (pami_context_t *) safemalloc( num_contexts * sizeof(pami_context_t) );
result = PAMI_Context_createv( client, NULL, 0, contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_createv");
/************************************************************************/
/* IBM: Updating the test with the assumption that we will Rput the */
/* IBM: local byte array to our neighbor's shared byte array. */
int n = (argc>1 ? atoi(argv[1]) : 1000);
size_t bytes = n * sizeof(int), bytes_out;/* IBM: debug - scale up testing */
int * shared = (int *) safemalloc(bytes);
for (int i=0; i<n; i++)
shared[i] = -1; /*IBM: initialize with -1, replaced with neighbor's rank */
pami_memregion_t shared_mr;
result = PAMI_Memregion_create(contexts[0], shared, bytes, &bytes_out, &shared_mr);
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
int * local = (int *) safemalloc(bytes);
for (int i=0; i<n; i++)
local[i] = world_rank; /*IBM: initialize with our rank */
pami_memregion_t local_mr;
result = PAMI_Memregion_create(contexts[0], local, bytes, &bytes_out, &local_mr); /* IBM: local */
TEST_ASSERT(result == PAMI_SUCCESS && bytes==bytes_out,"PAMI_Memregion_create");
status = MPI_Barrier(MPI_COMM_WORLD);
TEST_ASSERT(result == MPI_SUCCESS,"MPI_Barrier");
pami_memregion_t * shmrs = (pami_memregion_t *) safemalloc( world_size * sizeof(pami_memregion_t) );
status = MPI_Allgather(&shared_mr, sizeof(pami_memregion_t), MPI_BYTE,
shmrs, sizeof(pami_memregion_t), MPI_BYTE, MPI_COMM_WORLD);
TEST_ASSERT(result == MPI_SUCCESS,"MPI_Allgather");
int target = (world_rank>0 ? world_rank-1 : world_size-1);
pami_endpoint_t target_ep;
result = PAMI_Endpoint_create(client, (pami_task_t) target, 0, &target_ep);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Endpoint_create");
int active = 2;
pami_rput_simple_t parameters;
parameters.rma.dest = target_ep;
parameters.rma.bytes = bytes;
parameters.rma.cookie = &active;
parameters.rma.done_fn = cb_done;
parameters.rdma.local.mr = &local_mr;
parameters.rdma.local.offset = 0;
parameters.rdma.remote.mr = &shmrs[target]; /*IBM: target's mem region */
parameters.rdma.remote.offset = 0;
parameters.put.rdone_fn = cb_done;
result = PAMI_Rput(contexts[0], ¶meters);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Rput");
while (active)
{
//result = PAMI_Context_advance( contexts[0], 100);
//TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_advance");
result = PAMI_Context_trylock_advancev(contexts, 1, 1000);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_trylock_advancev");
}
/* IBM: I'm done with Rput but my world_rank + 1 neighbor might not be so need to advance */
/* IBM: Could do a barrier or send/recv a completion message instead ....*/
active = 10; /* IBM: Arbitrary - advance some more - 10*10000 good enough? */
while (--active) /* IBM*/
{ /* IBM*/
result = PAMI_Context_trylock_advancev(contexts, 1, 10000); /* IBM*/
/*TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_trylock_advancev");*/ /* IBM*/
} /* IBM*/
int errors = 0;
target = (world_rank<(world_size-1) ? world_rank+1 : 0);
for (int i=0; i<n; i++)
if ((shared[i] != target) ||
(local[i] != world_rank)) /*IBM: also verify didn't change local */
errors++;
if (errors>0)
{
printf("%ld: %d errors :-( \n", (long)world_rank, errors); /*IBM: grep "errors" in scaled up output */
for (int i=0; i<n; i++)
printf("%ld: local[%d] = %d , shared[%d] = %d (%d) \n", (long)world_rank, i, local[i], i, shared[i], target); /*IBM: print both arrays */
}
else
printf("%ld: no errors :-) \n", (long)world_rank);
MPI_Barrier(MPI_COMM_WORLD);
result = PAMI_Memregion_destroy(contexts[0], &shared_mr);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
result = PAMI_Memregion_destroy(contexts[0], &local_mr);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Memregion_destroy");
free(shmrs);
free(local);
free(shared);
/************************************************************************/
/* finalize the contexts */
result = PAMI_Context_destroyv( contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_destroyv");
free(contexts);
/* finalize the client */
result = PAMI_Client_destroy( &client );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_destroy");
status = MPI_Barrier(MPI_COMM_WORLD);
TEST_ASSERT(result == MPI_SUCCESS,"MPI_Barrier");
MPI_Finalize();
if (world_rank==0)
printf("%ld: end of test \n", world_rank );
fflush(stdout);
return 0;
}
int main (int argc, char ** argv)
{
pami_client_t client;
pami_context_t context[2];
pami_task_t task_id;
size_t num_tasks = 0;
size_t ncontexts = 0;
size_t errors = 0;
pami_result_t result = PAMI_ERROR;
pami_type_t subtype;
pami_type_t compound_type;
pami_type_t simple_type;
info_t exchange[MAX_TASKS];
double data[BUFFERSIZE];
volatile unsigned ready;
{ /* init */
ready = 0;
unsigned i;
for (i = 0; i < MAX_TASKS; i++)
exchange[i].active = 0;
for (i = 0; i < BUFFERSIZE; i++)
data[i] = E;
result = PAMI_Client_create ("TEST", &client, NULL, 0);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami client. result = %d\n", result);
return 1;
}
pami_configuration_t configuration;
configuration.name = PAMI_CLIENT_TASK_ID;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
task_id = configuration.value.intval;
/*fprintf (stderr, "My task id = %d\n", task_id);*/
configuration.name = PAMI_CLIENT_NUM_TASKS;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
num_tasks = configuration.value.intval;
/*if (task_id == 0)
fprintf (stderr, "Number of tasks = %zu\n", num_tasks);*/
if ((num_tasks < 2) || (num_tasks > MAX_TASKS))
{
fprintf (stderr, "Error. This test requires 2-%d tasks. Number of tasks in this job: %zu\n", MAX_TASKS, num_tasks);
return 1;
}
if (task_id == num_tasks - 1)
{
for (i = 0; i < 320; i++)
data[i] = PI;
}
configuration.name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query(client, &configuration, 1);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable query configuration (%d). result = %d\n", configuration.name, result);
return 1;
}
ncontexts = (configuration.value.intval < 2) ? 1 : 2;
/*if (task_id == 0)
fprintf (stderr, "maximum contexts = %zu, number of contexts used in this test = %zu\n", configuration.value.intval, ncontexts);*/
result = PAMI_Context_createv(client, NULL, 0, context, ncontexts);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create pami context(s). result = %d\n", result);
return 1;
}
pami_dispatch_hint_t options = {};
for (i = 0; i < ncontexts; i++)
{
pami_dispatch_callback_function fn;
fn.p2p = dispatch_exchange;
result = PAMI_Dispatch_set (context[i],
EXCHANGE_DISPATCH_ID,
fn,
(void *) exchange,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami 'exchange' dispatch. result = %d\n", result);
return 1;
}
fn.p2p = dispatch_notify;
result = PAMI_Dispatch_set (context[i],
NOTIFY_DISPATCH_ID,
fn,
(void *) & ready,
options);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable register pami 'notify' dispatch. result = %d\n", result);
return 1;
}
}
/* ***********************************
* Create the pami types
************************************/
/* This compound noncontiguous type is composed of one double, skips a double,
* two doubles, skips a double, three doubles, skips a double, five doubles,
* skips a double, six doubles, skips a double, seven doubles, skips a double,
* eight doubles, then skips two doubles.
*
* This results in a type with 32 doubles that is 40 doubles
* 'wide'.
*/
PAMI_Type_create (&subtype);
PAMI_Type_add_simple (subtype,
sizeof(double), /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 2); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 2, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 3); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 3, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 4); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 5, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 6); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 6, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 7); /* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 7, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 8);/* stride */
PAMI_Type_add_simple (subtype,
sizeof(double) * 8, /* bytes */
0, /* offset */
1, /* count */
sizeof(double) * 10);/* stride */
PAMI_Type_complete (subtype, sizeof(double));
/* This noncontiguous type is composed of the above compound type, repeated
* ten times with no stride.
*
* This results in a type with 320 doubles that is 400 doubles
* 'wide'.
*/
PAMI_Type_create (&compound_type);
PAMI_Type_add_typed (compound_type,
subtype, /* subtype */
0, /* offset */
10, /* count */
sizeof(double) * 32); /* stride */
PAMI_Type_complete (compound_type, sizeof(double));
/* This simple noncontiguous type is composed of eight contiguous doubles,
* then skips a _single_ double, repeated 40 times.
*
* This results in a type with 320 doubles that is 360 doubles 'wide'.
*/
PAMI_Type_create (&simple_type);
PAMI_Type_add_simple (simple_type,
sizeof(double) * 8, /* bytes */
0, /* offset */
40, /* count */
sizeof(double) * 9); /* stride */
PAMI_Type_complete (simple_type, sizeof(double));
/* Create a memory region for the local data buffer. */
size_t bytes;
result = PAMI_Memregion_create (context[0], (void *) data, BUFFERSIZE,
&bytes, &exchange[task_id].mr);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to create memory region. result = %d\n", result);
return 1;
}
else if (bytes < BUFFERSIZE)
{
fprintf (stderr, "Error. Unable to create memory region of a large enough size. result = %d\n", result);
return 1;
}
/* Broadcast the memory region to all tasks - including self. */
for (i = 0; i < num_tasks; i++)
{
pami_send_immediate_t parameters;
parameters.dispatch = EXCHANGE_DISPATCH_ID;
parameters.header.iov_base = (void *) & bytes;
parameters.header.iov_len = sizeof(size_t);
parameters.data.iov_base = (void *) & exchange[task_id].mr;
parameters.data.iov_len = sizeof(pami_memregion_t);
PAMI_Endpoint_create (client, i, 0, ¶meters.dest);
result = PAMI_Send_immediate (context[0], ¶meters);
}
/* Advance until all memory regions have been received. */
for (i = 0; i < num_tasks; i++)
{
while (exchange[i].active == 0)
PAMI_Context_advance (context[0], 100);
}
} /* init done */
pami_send_immediate_t notify;
notify.dispatch = NOTIFY_DISPATCH_ID;
notify.header.iov_base = NULL;
notify.header.iov_len = 0;
notify.data.iov_base = NULL;
notify.data.iov_len = 0;
volatile size_t active = 1;
pami_rget_typed_t parameters;
parameters.rma.hints = (pami_send_hint_t) {0};
parameters.rma.cookie = (void *) & active;
parameters.rma.done_fn = decrement;
parameters.rma.bytes = 320 * sizeof(double);
if (task_id == 0)
{
fprintf (stdout, "PAMI_Rget('typed') functional test %s\n", (ncontexts < 2) ? "" : "[crosstalk]");
fprintf (stdout, "\n");
parameters.rdma.local.mr = &exchange[0].mr;
parameters.rdma.remote.mr = &exchange[num_tasks - 1].mr;
PAMI_Endpoint_create (client, num_tasks - 1, ncontexts - 1, ¶meters.rma.dest);
PAMI_Endpoint_create (client, num_tasks - 1, ncontexts - 1, ¬ify.dest);
}
else
{
parameters.rdma.local.mr = &exchange[num_tasks - 1].mr;
parameters.rdma.remote.mr = &exchange[0].mr;
PAMI_Endpoint_create (client, 0, 0, ¶meters.rma.dest);
PAMI_Endpoint_create (client, 0, 0, ¬ify.dest);
}
/* ******************************************************************** */
/* contiguous -> contiguous transfer test */
/* ******************************************************************** */
if (task_id == 0)
{
parameters.rdma.local.offset = 0;
parameters.rdma.remote.offset = 0;
parameters.type.local = PAMI_TYPE_DOUBLE;
parameters.type.remote = PAMI_TYPE_DOUBLE;
active = 1;
PAMI_Rget_typed (context[0], ¶meters);
while (active > 0)
PAMI_Context_advance (context[0], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[0], ¬ify);
}
else if (task_id == num_tasks - 1)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
ready = 0;
}
/* ******************************************************************** */
/* contiguous -> non-contiguous transfer test */
/* ******************************************************************** */
if (task_id == num_tasks - 1)
{
parameters.rdma.local.offset = 4 * 1024;
parameters.rdma.remote.offset = 0;
parameters.type.local = simple_type;
parameters.type.remote = PAMI_TYPE_DOUBLE;
active = 1;
PAMI_Rget_typed (context[ncontexts - 1], ¶meters);
while (active > 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[ncontexts - 1], ¬ify);
}
else if (task_id == 0)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[0], 100);
ready = 0;
}
/* ******************************************************************** */
/* non-contiguous -> non-contiguous transfer test */
/* ******************************************************************** */
if (task_id == 0)
{
parameters.rdma.local.offset = 4 * 1024;
parameters.rdma.remote.offset = 4 * 1024;
parameters.type.local = compound_type;
parameters.type.remote = simple_type;
active = 1;
PAMI_Rget_typed (context[0], ¶meters);
while (active > 0)
PAMI_Context_advance (context[0], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[0], ¬ify);
}
else if (task_id == num_tasks - 1)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
ready = 0;
}
/* ******************************************************************** */
/* non-contiguous -> contiguous transfer test */
/* ******************************************************************** */
if (task_id == num_tasks - 1)
{
parameters.rdma.local.offset = 8 * 1024;
parameters.rdma.remote.offset = 4 * 1024;
parameters.type.local = PAMI_TYPE_DOUBLE;
parameters.type.remote = compound_type;
active = 1;
PAMI_Rget_typed (context[ncontexts - 1], ¶meters);
while (active > 0)
PAMI_Context_advance (context[ncontexts - 1], 100);
/* Notify the remote task that the data has been transfered. */
PAMI_Send_immediate (context[ncontexts - 1], ¬ify);
}
else if (task_id == 0)
{
/* Wait for notification that the data has been transfered. */
while (ready == 0)
PAMI_Context_advance (context[0], 100);
ready = 0;
}
/* ******************************************************************** */
/* VERIFY data buffers */
/* ******************************************************************** */
if (task_id == num_tasks - 1)
{
if (task_id == 0)
{
unsigned i = 0;
for (; i < 320; i++)
{
if (data[i] != PI)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, PI, data[i]);
}
}
for (; i < 512; i++)
{
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
}
unsigned j = 0;
for (; j < 40; j++)
{
unsigned n = 0;
for (; n < 8; n++)
{
if (data[i] != PI)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, PI, data[i]);
}
i++;
}
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
i++;
}
for (; i < 1024; i++)
{
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
}
for (; i < 1024 + 320; i++)
{
if (data[i] != PI)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, PI, data[i]);
}
}
for (; i < BUFFERSIZE; i++)
{
if (data[i] != E)
{
errors++;
fprintf (stderr, "Error. data[%d] != %g ..... (%g)\n", i, E, data[i]);
}
}
}
}
{ /* cleanup */
result = PAMI_Context_destroyv(context, ncontexts);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami context. result = %d\n", result);
return 1;
}
result = PAMI_Client_destroy(&client);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to destroy pami client. result = %d\n", result);
return 1;
}
if (task_id == num_tasks - 1)
{
if (errors)
fprintf (stdout, "Test completed with errors (%zu)\n", errors);
else
fprintf (stdout, "Test completed with success\n");
}
} /* cleanup done */
return (errors != 0);
};
