pami_result_t MPIDI_Pami_post_wrapper(pami_context_t context, void *cookie)
{
TRACE_ERR("In post wrapper\n");
TRACE_ERR("About to call collecive\n");
PAMI_Collective(context, (pami_xfer_t *)cookie);
TRACE_ERR("Done calling collective\n");
return PAMI_SUCCESS;
}
static void
gasnete_coll_pami_scatt(const gasnet_team_handle_t team, void *dst,
gasnet_image_t srcimage, const void *src,
size_t nbytes, int flags GASNETI_THREAD_FARG)
{
const int i_am_root = gasnete_coll_image_is_local(team, srcimage);
#if GASNET_PAR
int i_am_leader = gasnete_coll_pami_images_barrier(team); /* XXX: over-synced for IN_NO and IN_MY */
if ((flags & GASNET_COLL_LOCAL) && i_am_root) {
/* root thread must be leader for its node */
const gasnete_coll_threaddata_t * const td = GASNETE_COLL_MYTHREAD_NOALLOC;
i_am_leader = (srcimage == td->my_image);
}
#else
const int i_am_leader = 1;
#endif
if (i_am_leader) {
volatile unsigned int done = 0;
pami_result_t rc;
pami_xfer_t op;
if (flags & GASNET_COLL_IN_ALLSYNC) gasnetc_fast_barrier();
op = gasnete_op_template_scatt;
op.cookie = (void *)&done;
op.algorithm = team->pami.scatt_alg;
op.cmd.xfer_scatter.root = gasnetc_endpoint(GASNETE_COLL_REL2ACT(team,gasnete_coll_image_node(team, srcimage)));
op.cmd.xfer_scatter.sndbuf = (/*not-const*/ void *)src;
op.cmd.xfer_scatter.stypecount = nbytes;
op.cmd.xfer_scatter.rcvbuf = dst;
op.cmd.xfer_scatter.rtypecount = nbytes;
GASNETC_PAMI_LOCK(gasnetc_context);
rc = PAMI_Collective(gasnetc_context, &op);
GASNETC_PAMI_UNLOCK(gasnetc_context);
GASNETC_PAMI_CHECK(rc, "initiating blocking scatter");
gasneti_polluntil(done);
}
if (flags & GASNET_COLL_OUT_ALLSYNC) {
if (i_am_leader) gasnetc_fast_barrier();
(void) gasnete_coll_pami_images_barrier(team);
}
}
static void native_collective(pami_xfer_t *op_p, int need_lock) {
pami_result_t rc;
volatile unsigned int counter = 0;
op_p->cb_done = &gasnetc_cb_inc_uint;
op_p->cookie = (void *)&counter;
op_p->options.multicontext = PAMI_HINT_DISABLE;
if (need_lock) GASNETC_PAMI_LOCK(gasnetc_context);
rc = PAMI_Collective(gasnetc_context, op_p);
GASNETC_PAMI_CHECK(rc, "initiating a native collective");
if (need_lock) GASNETC_PAMI_UNLOCK(gasnetc_context);
if (gasneti_attach_done) {
gasneti_polluntil(counter);
} else {
rc = gasnetc_wait_uint(gasnetc_context, &counter, 1);
GASNETC_PAMI_CHECK(rc, "polling a native collective");
}
}
/**
* \brief Blocking 'world geometry' barrier
*
* This function is provided for illustrative purposes only. One would never
* include the retrieval of the world geometry and the query of the barrier
* algorithm in a performance critical code.
*
* \param[in] client The PAMI client; needed to obtain the geometry
* \param[in] context The PAMI context; used for the barrier communication
*/
void simple_barrier (pami_client_t client, pami_context_t context)
{
pami_result_t result;
pami_geometry_t world_geometry;
pami_xfer_t xfer;
pami_algorithm_t algorithm;
pami_metadata_t metadata;
/* Retrieve the PAMI 'world' geometry */
result = PAMI_ERROR;
result = PAMI_Geometry_world (client, &world_geometry);
assert (result == PAMI_SUCCESS);
/* Query the 'always works' barrier algorithm in the geometry */
result = PAMI_ERROR;
result = PAMI_Geometry_algorithms_query (world_geometry, PAMI_XFER_BARRIER,
&algorithm, &metadata, 1,
NULL, NULL, 0);
assert (result == PAMI_SUCCESS);
/* Set up the barrier */
volatile unsigned active = 1;
xfer.cb_done = simple_barrier_decrement;
xfer.cookie = (void *) & active;
xfer.algorithm = algorithm;
/* Issue the barrier collective */
result = PAMI_ERROR;
result = PAMI_Collective (context, &xfer);
assert (result == PAMI_SUCCESS);
/* Advance until the barrier has completed */
while (active)
{
result = PAMI_ERROR;
result = PAMI_Context_advance (context, 1);
assert (result == PAMI_SUCCESS);
}
return;
}
static void
gasnete_coll_pami_allga(const gasnet_team_handle_t team,
void *dst, const void *src,
size_t nbytes, int flags GASNETE_THREAD_FARG)
{
#if GASNET_PAR
int i_am_leader = gasnete_coll_pami_images_barrier(team); /* XXX: over-synced for IN_NO and IN_MY */
#else
const int i_am_leader = 1;
#endif
if (i_am_leader) {
volatile unsigned int done = 0;
pami_result_t rc;
pami_xfer_t op;
if (flags & GASNET_COLL_IN_ALLSYNC) gasnetc_fast_barrier();
op = gasnete_op_template_allga;
op.cookie = (void *)&done;
op.algorithm = team->pami.allga_alg;
op.cmd.xfer_allgather.sndbuf = (/*not-const*/ void *)src;
op.cmd.xfer_allgather.stypecount = nbytes;
op.cmd.xfer_allgather.rcvbuf = dst;
op.cmd.xfer_allgather.rtypecount = nbytes;
GASNETC_PAMI_LOCK(gasnetc_context);
rc = PAMI_Collective(gasnetc_context, &op);
GASNETC_PAMI_UNLOCK(gasnetc_context);
GASNETC_PAMI_CHECK(rc, "initiating blocking allgather");
gasneti_polluntil(done);
}
if (flags & GASNET_COLL_OUT_ALLSYNC) {
if (i_am_leader) gasnetc_fast_barrier();
(void) gasnete_coll_pami_images_barrier(team);
}
}
int main(int argc, char* argv[])
{
pami_result_t result = PAMI_ERROR;
if (Kernel_GetRank()==0)
print_meminfo(stdout, "before PAMI_Client_create");
/* initialize the client */
char * clientname = "";
pami_client_t client;
result = PAMI_Client_create( clientname, &client, NULL, 0 );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_create");
if (Kernel_GetRank()==0)
print_meminfo(stdout, "after PAMI_Client_create");
/* query properties of the client */
pami_configuration_t config;
size_t num_contexts;
config.name = PAMI_CLIENT_TASK_ID;
result = PAMI_Client_query( client, &config, 1);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
world_rank = config.value.intval;
config.name = PAMI_CLIENT_NUM_TASKS;
result = PAMI_Client_query( client, &config, 1);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
world_size = config.value.intval;
if ( world_rank == 0 )
{
printf("starting test on %ld ranks \n", world_size);
fflush(stdout);
}
config.name = PAMI_CLIENT_PROCESSOR_NAME;
result = PAMI_Client_query( client, &config, 1);
assert(result == PAMI_SUCCESS);
//printf("rank %ld is processor %s \n", world_rank, config.value.chararray);
//fflush(stdout);
config.name = PAMI_CLIENT_NUM_CONTEXTS;
result = PAMI_Client_query( client, &config, 1);
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_query");
num_contexts = config.value.intval;
/* initialize the contexts */
pami_context_t * contexts = NULL;
contexts = (pami_context_t *) malloc( num_contexts * sizeof(pami_context_t) ); assert(contexts!=NULL);
if (Kernel_GetRank()==0)
fprintf(stdout, "num_contexts = %ld \n", (long)num_contexts);
result = PAMI_Context_createv( client, &config, 0, contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_createv");
if (Kernel_GetRank()==0)
print_meminfo(stdout, "after PAMI_Context_createv");
/* setup the world geometry */
pami_geometry_t world_geometry;
pami_xfer_type_t barrier_xfer = PAMI_XFER_BARRIER;
size_t num_alg[2];
pami_algorithm_t * safe_barrier_algs = NULL;
pami_metadata_t * safe_barrier_meta = NULL;
pami_algorithm_t * fast_barrier_algs = NULL;
pami_metadata_t * fast_barrier_meta = NULL;
result = PAMI_Geometry_world( client, &world_geometry );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Geometry_world");
if (Kernel_GetRank()==0)
print_meminfo(stdout, "after PAMI_Geometry_world");
result = PAMI_Geometry_algorithms_num( world_geometry, barrier_xfer, num_alg );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Geometry_algorithms_num");
if ( world_rank == 0 ) printf("number of barrier algorithms = {%ld,%ld} \n", num_alg[0], num_alg[1] );
if (Kernel_GetRank()==0)
print_meminfo(stdout, "after PAMI_Geometry_algorithms_num");
safe_barrier_algs = (pami_algorithm_t *) malloc( num_alg[0] * sizeof(pami_algorithm_t) ); assert(safe_barrier_algs!=NULL);
safe_barrier_meta = (pami_metadata_t *) malloc( num_alg[0] * sizeof(pami_metadata_t) ); assert(safe_barrier_meta!=NULL);
fast_barrier_algs = (pami_algorithm_t *) malloc( num_alg[1] * sizeof(pami_algorithm_t) ); assert(fast_barrier_algs!=NULL);
fast_barrier_meta = (pami_metadata_t *) malloc( num_alg[1] * sizeof(pami_metadata_t) ); assert(fast_barrier_meta!=NULL);
result = PAMI_Geometry_algorithms_query( world_geometry, barrier_xfer,
safe_barrier_algs, safe_barrier_meta, num_alg[0],
fast_barrier_algs, fast_barrier_meta, num_alg[1] );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Geometry_algorithms_query");
if (Kernel_GetRank()==0)
print_meminfo(stdout, "after PAMI_Geometry_algorithms_query");
/* perform a barrier */
size_t b;
pami_xfer_t barrier;
volatile int active = 0;
for ( b = 0 ; b < num_alg[0] ; b++ )
{
barrier.cb_done = cb_done;
barrier.cookie = (void*) &active;
barrier.algorithm = safe_barrier_algs[b];
uint64_t t0 = GetTimeBase();
active = 1;
result = PAMI_Collective( contexts[0], &barrier );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Collective - barrier");
while (active)
result = PAMI_Context_advance( contexts[0], 1 );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_advance - barrier");
uint64_t t1 = GetTimeBase();
if ( world_rank == 0 ) printf("safe barrier algorithm %ld (%s) - took %llu cycles \n", b, safe_barrier_meta[b].name, (long long unsigned int)t1-t0 );
fflush(stdout);
}
for ( b = 0 ; b < num_alg[1] ; b++ )
{
barrier.cb_done = cb_done;
barrier.cookie = (void*) &active;
barrier.algorithm = fast_barrier_algs[b];
uint64_t t0 = GetTimeBase();
active = 1;
result = PAMI_Collective( contexts[0], &barrier );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Collective - barrier");
while (active)
result = PAMI_Context_advance( contexts[0], 1 );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_advance - barrier");
uint64_t t1 = GetTimeBase();
if ( world_rank == 0 ) printf("fast barrier algorithm %ld (%s) - took %llu cycles \n", b, fast_barrier_meta[b].name, (long long unsigned int)t1-t0 );
fflush(stdout);
}
if (Kernel_GetRank()==0)
print_meminfo(stdout, "after barrier tests");
/* finalize the contexts */
result = PAMI_Context_destroyv( contexts, num_contexts );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Context_destroyv");
free(contexts);
if (Kernel_GetRank()==0)
print_meminfo(stdout, "before PAMI_Client_destroy");
/* finalize the client */
result = PAMI_Client_destroy( &client );
TEST_ASSERT(result == PAMI_SUCCESS,"PAMI_Client_destroy");
if (Kernel_GetRank()==0)
print_meminfo(stdout, "after PAMI_Client_destroy");
if ( world_rank == 0 )
{
printf("end of test \n");
fflush(stdout);
}
return 0;
}
int main(int argc, char*argv[])
{
pami_client_t client;
pami_context_t *context;
pami_geometry_t world_geometry;
pami_task_t root_task = 0;
/* Barrier variables */
size_t barrier_num_algorithm[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_type_t barrier_xfer = PAMI_XFER_BARRIER;
pami_xfer_t barrier;
volatile unsigned bar_poll_flag = 0;
/* Amscatter variables */
size_t amscatter_num_algorithm[2];
pami_algorithm_t *amscatter_always_works_algo = NULL;
pami_metadata_t *amscatter_always_works_md = NULL;
pami_algorithm_t *amscatter_must_query_algo = NULL;
pami_metadata_t *amscatter_must_query_md = NULL;
pami_xfer_type_t amscatter_xfer = PAMI_XFER_AMSCATTER;
pami_xfer_t amscatter;
volatile unsigned amscatter_total_count = 0;
int nalg = 0, i;
double ti, tf, usec;
/* Process environment variables and setup globals */
setup_env();
assert(gNum_contexts > 0);
context = (pami_context_t*)malloc(sizeof(pami_context_t) * gNum_contexts);
/* \note Test environment variable" TEST_ROOT=N, defaults to 0.*/
char* sRoot = getenv("TEST_ROOT");
/* Override ROOT */
if (sRoot) root_task = (pami_task_t) atoi(sRoot);
/* Initialize PAMI */
int rc = pami_init(&client, /* Client */
context, /* Context */
NULL, /* Clientname=default */
&gNum_contexts, /* gNum_contexts */
NULL, /* null configuration */
0, /* no configuration */
&my_task_id, /* task id */
&num_tasks); /* number of tasks */
if (rc == 1)
return 1;
if (gNumRoots > num_tasks) gNumRoots = num_tasks;
/* Allocate buffer(s) */
int err = 0;
void *sbuf = NULL;
err = posix_memalign(&sbuf, 128, (gMax_byte_count * num_tasks) + gBuffer_offset);
assert(err == 0);
sbuf = (char*)sbuf + gBuffer_offset;
void* rbuf = NULL;
err = posix_memalign(&rbuf, 128, gMax_byte_count + gBuffer_offset);
assert(err == 0);
rbuf = (char*)rbuf + gBuffer_offset;
void *headers = NULL;
err = posix_memalign((void **)&headers, 128, (num_tasks * sizeof(user_header_t)) + gBuffer_offset);
headers = (char*)headers + gBuffer_offset;
void *validation = NULL;
err = posix_memalign((void **)&validation, 128, (num_tasks * sizeof(validation_t)) + gBuffer_offset);
validation = (char*)validation + gBuffer_offset;
/* Initialize the headers */
for(i = 0; i < num_tasks; ++i)
{
((user_header_t *)headers)[i].dst_rank = i;
}
unsigned iContext = 0;
for (; iContext < gNum_contexts; ++iContext)
{
if (my_task_id == 0)
printf("# Context: %u\n", iContext);
/* Query the world geometry for barrier algorithms */
rc |= query_geometry_world(client,
context[iContext],
&world_geometry,
barrier_xfer,
barrier_num_algorithm,
&bar_always_works_algo,
&bar_always_works_md,
&bar_must_query_algo,
&bar_must_query_md);
if (rc == 1)
return 1;
/* Query the world geometry for amscatter algorithms */
rc |= query_geometry_world(client,
context[iContext],
&world_geometry,
amscatter_xfer,
amscatter_num_algorithm,
&amscatter_always_works_algo,
&amscatter_always_works_md,
&amscatter_must_query_algo,
&amscatter_must_query_md);
if (rc == 1)
return 1;
_g_recv_buffer = rbuf;
_g_send_buffer = sbuf;
_g_val_buffer = validation;
barrier.cb_done = cb_done;
barrier.cookie = (void*) & bar_poll_flag;
barrier.algorithm = bar_always_works_algo[0];
blocking_coll(context[iContext], &barrier, &bar_poll_flag);
amscatter.algorithm = amscatter_always_works_algo[0];
amscatter.cmd.xfer_amscatter.headers = headers;
amscatter.cmd.xfer_amscatter.headerlen = sizeof(user_header_t);
amscatter.cmd.xfer_amscatter.sndbuf = sbuf;
amscatter.cmd.xfer_amscatter.stype = PAMI_TYPE_BYTE;
amscatter.cmd.xfer_amscatter.stypecount = 0;
for (nalg = 0; nalg < amscatter_num_algorithm[0]; nalg++)
{
gProtocolName = amscatter_always_works_md[nalg].name;
if (my_task_id == root_task)
{
printf("# AMScatter Bandwidth Test(size:%zu) -- context = %d, root = %d, protocol: %s\n",num_tasks,
iContext, root_task, amscatter_always_works_md[nalg].name);
printf("# Size(bytes) iterations bytes/sec usec\n");
printf("# ----------- ----------- ----------- ---------\n");
fflush(stdout);
}
if (((strstr(amscatter_always_works_md[nalg].name,gSelected) == NULL) && gSelector) ||
((strstr(amscatter_always_works_md[nalg].name,gSelected) != NULL) && !gSelector)) continue;
int j;
pami_collective_hint_t h = {0};
pami_dispatch_callback_function fn;
lgContext = context[iContext];
fn.amscatter = cb_amscatter_recv;
PAMI_AMCollective_dispatch_set(context[iContext],
amscatter_always_works_algo[nalg],
root_task,/* Set the dispatch id, can be any arbitrary value */
fn,
(void*) &amscatter_total_count,
h);
amscatter.cmd.xfer_amscatter.dispatch = root_task;
amscatter.algorithm = amscatter_always_works_algo[nalg];
volatile unsigned *nscatter = &amscatter_total_count;
for (i = gMin_byte_count; i <= gMax_byte_count; i *= 2)
{
size_t dataSent = i;
int niter;
pami_result_t result;
if (dataSent < CUTOFF)
niter = gNiterlat;
else
niter = NITERBW;
*nscatter = 0;
memset(rbuf, 0xFF, i);
scatter_initialize_sndbuf (sbuf, i, num_tasks);
blocking_coll(context[iContext], &barrier, &bar_poll_flag);
ti = timer();
for (j = 0; j < niter; j++)
{
root_task = (root_task + num_tasks - 1) % num_tasks;
if (my_task_id == root_task)
{
amscatter.cmd.xfer_amscatter.stypecount = i;
result = PAMI_Collective(context[iContext], &amscatter);
if (result != PAMI_SUCCESS)
{
fprintf (stderr, "Error. Unable to issue collective. result = %d\n", result);
return 1;
}
}
while (*nscatter <= j)
result = PAMI_Context_advance (context[iContext], 1);
rc |= _gRc; /* validation return code done in cb_amscatter_done */
}
assert(*nscatter == niter);
tf = timer();
blocking_coll(context[iContext], &barrier, &bar_poll_flag);
usec = (tf - ti) / (double)niter;
if(my_task_id == root_task)
{
printf(" %11lld %16d %14.1f %12.2f\n",
(long long)dataSent,
niter,
(double)1e6*(double)dataSent / (double)usec,
usec);
fflush(stdout);
}
}
lgContext = NULL;
}
free(bar_always_works_algo);
free(bar_always_works_md);
free(bar_must_query_algo);
free(bar_must_query_md);
free(amscatter_always_works_algo);
free(amscatter_always_works_md);
free(amscatter_must_query_algo);
free(amscatter_must_query_md);
} /*for(unsigned iContext = 0; iContext < gNum_contexts; ++iContexts)*/
sbuf = (char*)sbuf - gBuffer_offset;
free(sbuf);
rbuf = (char*)rbuf - gBuffer_offset;
free(rbuf);
headers = (char*)headers - gBuffer_offset;
free(headers);
validation = (char*)validation - gBuffer_offset;
free(validation);
rc |= pami_shutdown(&client, context, &gNum_contexts);
return rc;
}
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;
}
static void
gasnete_coll_pami_allgavi(const gasnet_team_handle_t team,
void *dst, const void *src,
size_t nbytes, int flags GASNETE_THREAD_FARG)
{
int i_am_leader = gasnete_coll_pami_images_barrier(team); /* XXX: over-synced for IN_NO and IN_MY */
const gasnete_coll_threaddata_t * const td = GASNETE_COLL_MYTHREAD_NOALLOC;
if (flags & GASNET_COLL_IN_ALLSYNC) {
if (i_am_leader) gasnetc_fast_barrier();
(void) gasnete_coll_pami_images_barrier(team);
}
GASNETE_FAST_UNALIGNED_MEMCPY(gasnete_coll_scale_ptr(team->pami.scratch_space,
td->my_local_image,
nbytes),
src, nbytes);
(void) gasnete_coll_pami_images_barrier(team);
if (i_am_leader) {
volatile unsigned int done = 0;
pami_result_t rc;
pami_xfer_t op;
op = gasnete_op_template_allgavi; /* allgatherv_int */
op.cookie = (void *)&done;
op.algorithm = team->pami.allgavi_alg;
op.cmd.xfer_allgatherv_int.sndbuf = team->pami.scratch_space;
op.cmd.xfer_allgatherv_int.stypecount = nbytes * team->my_images;
op.cmd.xfer_allgatherv_int.rcvbuf = dst;
op.cmd.xfer_allgatherv_int.rtypecounts = team->pami.counts;
op.cmd.xfer_allgatherv_int.rdispls = team->pami.displs;
if (team->pami.prev_nbytes != nbytes) {
int i;
for (i = 0; i < team->total_ranks; ++i) {
op.cmd.xfer_allgatherv_int.rtypecounts[i] = nbytes * team->all_images[i];
op.cmd.xfer_allgatherv_int.rdispls[i] = nbytes * team->all_offset[i];
}
team->pami.prev_nbytes = nbytes;
}
GASNETC_PAMI_LOCK(gasnetc_context);
rc = PAMI_Collective(gasnetc_context, &op);
GASNETC_PAMI_UNLOCK(gasnetc_context);
GASNETC_PAMI_CHECK(rc, "initiating blocking allgatherv_int");
gasneti_polluntil(done);
gasneti_assert(NULL == team->pami.tmp_addr);
gasneti_sync_writes(); /* XXX: is this necessary? */
team->pami.tmp_addr = dst; /* wakes pollers, below */
(void) gasnete_coll_pami_images_barrier(team); /* matches instance below vvvv */
team->pami.tmp_addr = NULL;
} else {
gasneti_waitwhile(NULL == team->pami.tmp_addr);
GASNETE_FAST_UNALIGNED_MEMCPY(dst, team->pami.tmp_addr, nbytes * team->total_images);
(void) gasnete_coll_pami_images_barrier(team); /* matches instance above ^^^^ */
}
if (flags & GASNET_COLL_OUT_ALLSYNC) {
if (i_am_leader) gasnetc_fast_barrier();
(void) gasnete_coll_pami_images_barrier(team);
}
}
static void
gasnete_coll_pami_scattvi(const gasnet_team_handle_t team, void *dst,
gasnet_image_t srcimage, const void *src,
size_t nbytes, int flags GASNETI_THREAD_FARG)
{
const int i_am_root = gasnete_coll_image_is_local(team, srcimage);
int i_am_leader = gasnete_coll_pami_images_barrier(team); /* XXX: over-synced for IN_NO and IN_MY */
const gasnete_coll_threaddata_t * const td = GASNETE_COLL_MYTHREAD_NOALLOC;
if ((flags & GASNET_COLL_LOCAL) && i_am_root) {
/* root thread must be leader for its node */
i_am_leader = (srcimage == td->my_image);
}
if (i_am_leader) {
volatile unsigned int done = 0;
pami_result_t rc;
pami_xfer_t op;
if (flags & GASNET_COLL_IN_ALLSYNC) gasnetc_fast_barrier();
op = gasnete_op_template_scattvi; /* scatterv_int */
op.cookie = (void *)&done;
op.algorithm = team->pami.scattvi_alg;
op.cmd.xfer_scatterv_int.root = gasnetc_endpoint(GASNETE_COLL_REL2ACT(team,gasnete_coll_image_node(team, srcimage)));
op.cmd.xfer_scatterv_int.rcvbuf = team->pami.scratch_space;
op.cmd.xfer_scatterv_int.rtypecount = nbytes * team->my_images;
if (i_am_root) {
op.cmd.xfer_scatterv_int.sndbuf = (/*not-const*/ void *)src;
op.cmd.xfer_scatterv_int.stypecounts = team->pami.counts;
op.cmd.xfer_scatterv_int.sdispls = team->pami.displs;
if (team->pami.prev_nbytes != nbytes) {
int i;
for (i = 0; i < team->total_ranks; ++i) {
op.cmd.xfer_scatterv_int.stypecounts[i] = nbytes * team->all_images[i];
op.cmd.xfer_scatterv_int.sdispls[i] = nbytes * team->all_offset[i];
}
team->pami.prev_nbytes = nbytes;
}
}
GASNETC_PAMI_LOCK(gasnetc_context);
rc = PAMI_Collective(gasnetc_context, &op);
GASNETC_PAMI_UNLOCK(gasnetc_context);
GASNETC_PAMI_CHECK(rc, "initiating blocking scatterv_int");
gasneti_polluntil(done);
gasneti_assert(NULL == team->pami.tmp_addr);
gasneti_sync_writes();
team->pami.tmp_addr = team->pami.scratch_space; /* wakes pollers, below */
} else {
gasneti_waitwhile(NULL == team->pami.tmp_addr);
}
GASNETI_MEMCPY (dst,
gasnete_coll_scale_ptr(team->pami.tmp_addr,
td->my_local_image,
nbytes),
nbytes);
(void) gasnete_coll_pami_images_barrier(team);
if (i_am_leader) {
team->pami.tmp_addr = NULL;
}
if (flags & GASNET_COLL_OUT_ALLSYNC) {
if (i_am_leader) gasnetc_fast_barrier();
(void) gasnete_coll_pami_images_barrier(team);
}
}
