void smp_coll_barrier_tree_pull_push(smp_coll_t handle, int flags) {
int i;
int flagset = handle->barrier_flag_set;
gasnett_local_wmb();
for(i=0; i<handle->barrier_num_children; i++) {
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, handle->barrier_children[i], flagset)==0);
}
/*reset old one*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, !flagset, 0);
/*set my flag indicating barrier is done*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset, 1);
if(handle->MYTHREAD!=handle->barrier_root) {
/*singal parent and wait for parent to signal us*/
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, handle->MYTHREAD, 2+flagset)==0);
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, 2+flagset, 0);
}
/*signal all my children*/
for(i=0; i<handle->barrier_num_children; i++) {
SMP_COLL_SET_BARRIER_FLAG(handle, handle->barrier_children[i], 2+flagset, 1);
}
handle->barrier_flag_set = !handle->barrier_flag_set;
gasnett_local_rmb();
}
void smp_coll_barrier_tree_pull_pull(smp_coll_t handle, int flags) {
int i;
int flagset = handle->barrier_flag_set;
gasnett_local_wmb();
for(i=0; i<handle->barrier_num_children; i++) {
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, handle->barrier_children[i], flagset)==0);
}
/*set my flag indicating barrier is done*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, !flagset, 0);
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset, 1);
/*wait for parent to raise flag*/
if(handle->MYTHREAD!=handle->barrier_root) {
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, handle->barrier_parent, 2+flagset)==0);
}
/*parent has now acked my signal so we can clear the up signal*/
/*clear my down flags from previous round*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, 2+(!flagset), 0);
/*set my down flag for this round*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, 2+flagset, 1);
handle->barrier_flag_set = !handle->barrier_flag_set;
gasnett_local_rmb();
}
/*this is a push based implementation since each thread signals on the remote thread when it is ready*/
void smp_coll_barrier_tree_push_push(smp_coll_t handle, int flags) {
int i;
int flagset = handle->barrier_flag_set;
int atomicset = handle->curr_atomic_set;
gasnett_local_wmb();
/*push based tree wait for all children*/
gasneti_waitwhile(SMP_COLL_READ_ATOMIC(handle, handle->MYTHREAD, 0, atomicset)!=handle->barrier_num_children);
SMP_COLL_RESET_ATOMIC(handle, handle->MYTHREAD, 0, atomicset);
/*if i'm not root*/
if(handle->MYTHREAD!=handle->barrier_root) {
/*singal parent and wait for parent to signal us*/
SMP_COLL_INC_ATOMIC(handle, handle->barrier_parent, 0, atomicset);
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset)==0);
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset, 0);
}
/*signal all my children*/
for(i=0; i<handle->barrier_num_children; i++) {
SMP_COLL_SET_BARRIER_FLAG(handle, handle->barrier_children[i], flagset, 1);
}
handle->barrier_flag_set = !handle->barrier_flag_set;
handle->curr_atomic_set = !handle->curr_atomic_set;
gasnett_local_rmb();
}
void smp_coll_barrier_tree_push_pull(smp_coll_t handle, int flags) {
int flagset = handle->barrier_flag_set;
gasnett_local_wmb();
/*push based tree wait for all children*/
gasneti_waitwhile(SMP_COLL_READ_ATOMIC(handle, handle->MYTHREAD, 0, handle->curr_atomic_set)!=handle->barrier_num_children);
SMP_COLL_RESET_ATOMIC(handle, handle->MYTHREAD, 0, handle->curr_atomic_set);
/*signal parent and wiat for parent*/
if(handle->MYTHREAD!=handle->barrier_root) {
SMP_COLL_INC_ATOMIC(handle, handle->barrier_parent, 0, handle->curr_atomic_set);
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, handle->barrier_parent, flagset)==0);
}
/*parent has now acked my signal so we can clear the up signal*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset, 0);
/*clear my down flags from previous round*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, (!flagset), 0);
/*set my down flag for this round*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset, 1);
handle->barrier_flag_set = !handle->barrier_flag_set;
handle->curr_atomic_set = !handle->curr_atomic_set;
gasnett_local_rmb();
}
void smp_coll_barrier_tree_flag(smp_coll_t handle, int flags) {
int idx = 0;
int num_digits = handle->barrier_log_radix_THREADS;
int radixlog2 = handle->barrier_log_2_radix;
int radix = handle->barrier_radix;
int i,j,k;
int parent=-1;
gasnett_local_wmb();
/* reduce data from all the children*/
for(i=num_digits-1,j=0; i>=0; i--,j++) {
/*if my i^th digit is 0 that means that i am a sender for this round*/
if(SMP_COLL_GET_ITH_DIGIT_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0 &&
SMP_COLL_GET_LOWER_K_DIGITS_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0) {
for(k=1;k<radix;k++) {
int dest = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, i, k, radix, radixlog2);
if(dest<handle->THREADS) {
/*wait for dest to be ready before we send*/
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, dest, 0)==0);
}
}
}
}
/*set the flag indicating that my data (and all the data under my subtree is ready)*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, 0, 1);
if(handle->MYTHREAD!=barrier_root) {
/*Wait for parent to signal that my data for the entire tree has arrived*/
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, handle->MYTHREAD, 0)!=0);
}
/*broadcast the data back down my subtree*/
for(i=num_digits-1,j=0; i>=0; i--,j++) {
/*if my i^th digit is 0 that means that i am a sender for this round*/
if(SMP_COLL_GET_ITH_DIGIT_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0 &&
SMP_COLL_GET_LOWER_K_DIGITS_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0) {
for(k=1;k<radix;k++) {
int dest = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, i, k, radix, radixlog2);
if(dest<handle->THREADS) {
/*write memory barrier to ensure data is transfered before we set the flag*/
SMP_COLL_SET_BARRIER_FLAG(handle, dest, 0, 0);
}
}
}
}
gasnett_local_rmb();
}
void smp_coll_barrier_flag_tree_up_tree_down(smp_coll_t handle, int flags) {
int idx = 0;
int num_digits = handle->barrier_log_radix_THREADS;
int radixlog2 = handle->barrier_log_2_radix;
int radix = handle->barrier_radix;
int i,j,k;
int flagset = handle->barrier_flag_set;
int parent = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, 0, 0, radix, radixlog2);
/* reduce data from all the children*/
for(i=num_digits-1,j=0; i>=0; i--,j++) {
/*if my i^th digit is 0 that means that i am a sender for this round*/
if(SMP_COLL_GET_ITH_DIGIT_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0 &&
SMP_COLL_GET_LOWER_K_DIGITS_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0) {
for(k=1;k<radix;k++) {
int dest = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, i, k, radix, radixlog2);
if(dest<handle->THREADS) {
/*wait for dest to be ready before we send*/
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, dest, flagset)==0);
}
}
}
}
/*set the flag indicating that my data (and all the data under my subtree is ready)*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset, 1);
/*wait for parent to raise flag*/
if(handle->MYTHREAD!=barrier_root) {
gasneti_waitwhile(SMP_COLL_GET_BARRIER_FLAG(handle, parent, 2+flagset)==0);
}
/*parent has now acked my signal so we can clear the up signal*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, flagset, 0);
/*clear my down flags from previous round*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, 2+(!flagset), 0);
/*set my down flag for this round*/
SMP_COLL_SET_BARRIER_FLAG(handle, handle->MYTHREAD, 2+flagset, 1);
handle->barrier_flag_set = !handle->barrier_flag_set;
}
void smp_coll_broadcast_tree_flag(smp_coll_t handle, int num_addrs, void * const dstlist[], const void *src,
size_t nbytes, int flags, int radix){
int num_digits = smp_coll_mylogn(handle->THREADS, radix);
int radixlog2 = smp_coll_mylogn(radix,2);
int i,j,k;
if((flags & SMP_COLL_ALL_SYNC)) smp_coll_barrier(handle, flags);
/*first thing all threads do is set their flag to be 1 indicating they have arrived*/
SMP_COLL_SET_BCAST_FLAG(handle, handle->MYTHREAD, 0, 1);
/*
Don't care who my parent is for this algorithm
for(i=0; i<num_digits; i++) {
if(SMP_COLL_GET_ITH_DIGIT_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)!=0) {
parent = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, i, 0, radix, radixlog2);
break;
}
}
*/
/*they then wait for the parent to come around and reset their flag back to 0 indicating the data has also arrived*/
if(handle->MYTHREAD!=0) {
gasneti_waitwhile(SMP_COLL_GET_BCAST_FLAG(handle, handle->MYTHREAD,0)!=0);
} else {
memcpy(dstlist[0], src, nbytes);
}
for(i=num_digits-1,j=0; i>=0; i--,j++) {
/*if my i^th digit is 0 that means that i am a sender for this round*/
if(SMP_COLL_GET_ITH_DIGIT_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0 &&
SMP_COLL_GET_LOWER_K_DIGITS_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0) {
for(k=1;k<radix;k++) {
int dest = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, i, k, radix, radixlog2);
if(dest<handle->THREADS) {
/*wait for dest to be ready before we send*/
gasneti_waitwhile(SMP_COLL_GET_BCAST_FLAG(handle, dest, 0)==0);
memcpy(dstlist[dest], dstlist[handle->MYTHREAD], nbytes);
/*write memory barrier to ensure data is transfered before we set the flag*/
gasnett_local_wmb();
SMP_COLL_SET_BCAST_FLAG(handle, dest, 0, 0);
}
}
}
}
if((flags & SMP_COLL_ALL_SYNC)) smp_coll_barrier(handle, flags);
}
void smp_coll_broadcast_tree_atomic(smp_coll_t handle, int num_addrs, void * const dstlist[], const void *src,
size_t nbytes, int flags, int radix){
int num_digits = smp_coll_mylogn(handle->THREADS, radix);
int radixlog2 = smp_coll_mylogn(radix,2);
int i,j,k;
int parent=-1;
if((flags & SMP_COLL_ALL_SYNC)) smp_coll_barrier(handle, flags);
for(i=0; i<num_digits; i++) {
if(SMP_COLL_GET_ITH_DIGIT_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)!=0) {
parent = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, i, 0, radix, radixlog2);
break;
}
}
if(parent!=-1) {
gasneti_waitwhile(SMP_COLL_READ_ATOMIC(handle, handle->MYTHREAD, 0, handle->curr_atomic_set)!=1);
SMP_COLL_DEC_ATOMIC(handle, handle->MYTHREAD, 0, handle->curr_atomic_set);
gasnett_local_rmb();
} else {
memcpy(dstlist[0], src, nbytes);
}
for(i=num_digits-1,j=0; i>=0; i--,j++) {
/*if my i^th digit is 0 that means that i am a sender for this round*/
if(SMP_COLL_GET_ITH_DIGIT_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0 &&
SMP_COLL_GET_LOWER_K_DIGITS_POWER2RADIX(handle->MYTHREAD, i, radix, radixlog2)==0) {
for(k=1;k<radix;k++) {
int dest = SMP_COLL_MAKE_NUM_POWER2RADIX(handle->MYTHREAD, i, k, radix, radixlog2);
if(dest<handle->THREADS) {
GASNETE_FAST_UNALIGNED_MEMCPY(dstlist[dest], dstlist[handle->MYTHREAD], nbytes);
gasnett_local_wmb();
SMP_COLL_INC_ATOMIC(handle, dest, 0, handle->curr_atomic_set);
}
}
}
}
if((flags & SMP_COLL_ALL_SYNC)) smp_coll_barrier(handle, flags);
handle->curr_atomic_set = !handle->curr_atomic_set;
}
/*log_r(n) rounds with O(n) messages per round*/
void smp_coll_barrier_dissem_atomic(smp_coll_t handle, int flags) {
const dissem_info_t *dissem = handle->dissem_info;
const dissem_vector_t *barrier_order = dissem->barrier_order;
int i,j;
gasnett_local_wmb();
for(i=0; i<dissem->dissemination_phases; i++) {
if(barrier_order[i].n > 0) {
const int* elem_list = barrier_order[i].elem_list;
/*post on the sems*/
for(j=0; j<barrier_order[i].n; j++) {
int dest = elem_list[j];
/*increment counter i on dest by 1*/
SMP_COLL_INC_ATOMIC(handle, dest, i, handle->curr_atomic_set);
}
/*wait for counter i to be barrier_order[i].n*/
gasneti_waitwhile(SMP_COLL_READ_ATOMIC(handle, handle->MYTHREAD, i, handle->curr_atomic_set)!=barrier_order[i].n);
SMP_COLL_RESET_ATOMIC(handle, handle->MYTHREAD, i, handle->curr_atomic_set);
}
}
handle->curr_atomic_set = !handle->curr_atomic_set;
gasnett_local_rmb();
}
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);
}
}
