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");
}
}
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);
}
}
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);
}
}
