void _rmw ( pami_context_t context,
void * src_ptr_in,
void * src_ptr_out,
void * target_ptr,
pami_atomic_t op,
pami_type_t type,
pami_endpoint_t target )
{
pami_rmw_t rmw;
memset(&rmw, 0, sizeof(pami_rmw_t));
rmw.cookie = (void *)&_send_active;
rmw.done_fn = cb_done;
int test = 0;
rmw.local = src_ptr_out;
rmw.remote = target_ptr;
rmw.value = src_ptr_in;
rmw.test = &test;
rmw.type = type;
rmw.operation = op;
rmw.dest = target;
#if ENABLE_PROGRESS
pami_work_t work;
PAMI_Context_post (context, &work, (pami_work_function)PAMI_Rmw,
(void*)&rmw);
while (_send_active == 1);
#else
RC( PAMI_Rmw (context, &rmw) );
while (_send_active) PAMI_Context_advance (context, POLL_CNT);
#endif
_send_active = 1;
}
void get_contig ( pami_context_t context,
void * lbuf,
void * rbuf,
void * lbase,
void * rbase,
pami_memregion_t * lmr,
pami_memregion_t * rmr,
size_t sndlen,
pami_endpoint_t target )
{
pami_rget_simple_t rget;
rget.rma.dest = target;
rget.rma.bytes = sndlen;
rget.rma.cookie = (void*)&_send_active;
rget.rma.done_fn = cb_done;
rget.rma.hints.buffer_registered = PAMI_HINT_ENABLE;
rget.rma.hints.use_rdma = PAMI_HINT_ENABLE;
rget.rdma.local.mr = lmr;
rget.rdma.local.offset = (size_t)lbuf - (size_t)lbase;
rget.rdma.remote.mr = rmr;
rget.rdma.remote.offset = (size_t)rbuf - (size_t)rbase;
assert (_send_active == 1);
#if ENABLE_PROGRESS
pami_work_t work;
PAMI_Context_post (context, &work, (pami_work_function)PAMI_Rget,
(void*)&rget);
while (_send_active == 1);
#else
RC( PAMI_Rget (context, &rget) );
while (_send_active) PAMI_Context_advance (context, POLL_CNT);
#endif
_send_active = 1;
}
void send_contig ( pami_context_t context,
size_t dispatch,
void * metadata,
size_t hdrsize,
char * buffer,
size_t sndlen,
pami_endpoint_t target )
{
//fprintf(stderr, "Calling send dispatch %ld, hdrsize %ld, bytes %ld\n",
//dispatch, hdrsize, sndlen);
pami_send_t parameters;
parameters.send.dispatch = dispatch;
parameters.send.header.iov_base = metadata;
parameters.send.header.iov_len = hdrsize;
parameters.send.data.iov_base = buffer;
parameters.send.data.iov_len = sndlen;
parameters.events.cookie = (void *) & _send_active;
parameters.events.local_fn = NULL; //cb_done;
parameters.events.remote_fn = cb_done;
parameters.send.dest = target;
memset(¶meters.send.hints, 0, sizeof(parameters.send.hints));
#if ENABLE_PROGRESS
pami_work_t work;
PAMI_Context_post (context, &work, (pami_work_function)PAMI_Send,
(void*)¶meters);
while (_send_active == 1);
#else
RC( PAMI_Send (context, ¶meters) );
while (_send_active) PAMI_Context_advance (context, POLL_CNT);
#endif
_send_active = 1;
}
/**
* \brief Initialize support for MPIR_* nbc implementation.
*
* The MPIR_* non-blocking collectives only work if the schedule is advanced.
* This is done by posting a work function to context 0 that invokes the
* schedule progress function.
*
* Because this is a persistent work function and will negatively impact the
* performance of all other MPI operations - even when mpir non-blocking
* collectives are not used - the work function is only posted if explicitly
* requested.
*/
void MPIDI_NBC_init ()
{
if (MPIDI_Process.mpir_nbc != 0)
{
PAMI_Context_post(MPIDI_Context[0],
&mpidi_nbc_work_object,
mpidi_nbc_work_function,
NULL);
}
return;
}
/**
* This calculates a context to use and posts the short-send work to
* it. It then advances all contexts until the send is complete.
*/
static void *SendShort(void *c)
{
pami_task_t remote_task = 1-task;
size_t local_context = (remote_task+SHORT_DISPATCH)&(num_contexts-1);
pami_work_t state;
PAMI_Context_post(contexts[local_context], &state, SendShortHandoff, NULL);
while (!done.sshort.send)
//PAMI_Context_advancev(contexts, num_contexts, 1);
PAMI_Context_trylock_advancev(contexts, num_contexts, 1);
return NULL;
}
/**
* This calculates a context to use and posts the long-send work to
* it. It then advances all contexts until the send is complete.
*
* Because SendLongHandoff may complete before sending the msginfo, I
* declare it here and pass it in the cookie. Since this function
* will not exit until the local sen is complete, it is safe to put
* the msginfo on the stack for this one.
*/
static void *SendLong(void *c)
{
unsigned quad[] = {(unsigned)task, 0x11111, 0x22222, 0x33333};
pami_task_t remote_task = 1-task;
size_t local_context = (remote_task+LONG_DISPATCH)&(num_contexts-1);
pami_work_t state;
PAMI_Context_post(contexts[local_context], &state, SendLongHandoff, quad);
while (!done.slong.send)
//PAMI_Context_advancev(contexts, num_contexts, 1);
PAMI_Context_trylock_advancev(contexts, num_contexts, 1);
return NULL;
}
void put_contig ( pami_context_t context,
void * srcbuf,
void * dstbuf,
void * src_base,
void * dst_base,
pami_memregion_t * src_mr,
pami_memregion_t * dst_mr,
size_t sndlen,
pami_endpoint_t target )
{
//fprintf(stderr, "Calling send dispatch %ld, hdrsize %ld, bytes %ld\n",
// dispatch, hdrsize, sndlen);
pami_rput_simple_t rput;
rput.rma.dest = target;
rput.rma.hints.buffer_registered = PAMI_HINT_ENABLE;
rput.rma.hints.use_rdma = PAMI_HINT_ENABLE;
rput.rma.bytes = sndlen;
rput.rma.cookie = (void*)&_send_active;
rput.rma.done_fn = NULL; //cb_done;
rput.rdma.local.mr = src_mr;
rput.rdma.local.offset = (size_t)srcbuf - (size_t)src_base;
rput.rdma.remote.mr = dst_mr;
rput.rdma.remote.offset = (size_t)dstbuf - (size_t)dst_base;
rput.put.rdone_fn = cb_done;
#if ENABLE_PROGRESS
pami_work_t work;
PAMI_Context_post (context, &work, (pami_work_function)PAMI_Rput,
(void*)&rput);
while (_send_active == 1);
#else
RC( PAMI_Rput (context, &rput) );
while (_send_active) PAMI_Context_advance (context, POLL_CNT);
#endif
_send_active = 1;
}
