int MPIDI_CH3I_Accumulate(const void *origin_addr, int origin_count, MPI_Datatype
origin_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPID_Win * win_ptr, MPID_Request * ureq)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_msg_sz_t data_sz;
int dt_contig ATTRIBUTE((unused)), rank;
MPI_Aint dt_true_lb ATTRIBUTE((unused));
MPID_Datatype *dtp;
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
int made_progress = 0;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_ACCUMULATE);
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state == MPIDI_RMA_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
MPIDI_Datatype_get_info(origin_count, origin_datatype, dt_contig, data_sz, dtp, dt_true_lb);
if (data_sz == 0) {
goto fn_exit;
}
rank = win_ptr->comm_ptr->rank;
if (win_ptr->shm_allocated == TRUE && target_rank != rank &&
win_ptr->create_flavor != MPI_WIN_FLAVOR_SHARED) {
/* check if target is local and shared memory is allocated on window,
* if so, we directly perform this operation on shared memory region. */
/* FIXME: Here we decide whether to perform SHM operations by checking if origin and target are on
* the same node. However, in ch3:sock, even if origin and target are on the same node, they do
* not within the same SHM region. Here we filter out ch3:sock by checking shm_allocated flag first,
* which is only set to TRUE when SHM region is allocated in nemesis.
* In future we need to figure out a way to check if origin and target are in the same "SHM comm".
*/
MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &target_vc);
}
/* Do =! rank first (most likely branch?) */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED ||
(win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id)) {
mpi_errno = MPIDI_CH3I_Shm_acc_op(origin_addr, origin_count, origin_datatype,
target_rank, target_disp, target_count, target_datatype,
op, win_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (ureq) {
/* Complete user request and release the ch3 ref */
mpi_errno = MPID_Request_complete(ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
}
else {
MPIDI_RMA_Op_t *op_ptr = NULL;
MPIDI_CH3_Pkt_accum_t *accum_pkt = NULL;
int use_immed_pkt = FALSE;
int is_origin_contig, is_target_contig;
MPI_Aint stream_elem_count, stream_unit_count;
MPI_Aint predefined_dtp_size, predefined_dtp_count, predefined_dtp_extent;
MPID_Datatype *origin_dtp = NULL, *target_dtp = NULL;
int i;
/* queue it up */
mpi_errno = MPIDI_CH3I_Win_get_op(win_ptr, &op_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
/******************** Setting operation struct areas ***********************/
op_ptr->origin_addr = (void *) origin_addr;
op_ptr->origin_count = origin_count;
op_ptr->origin_datatype = origin_datatype;
op_ptr->target_rank = target_rank;
/* Remember user request */
op_ptr->ureq = ureq;
/* if source or target datatypes are derived, increment their
* reference counts */
if (!MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) {
MPID_Datatype_get_ptr(origin_datatype, origin_dtp);
}
if (!MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
MPID_Datatype_get_ptr(target_datatype, target_dtp);
}
/* Get size and count for predefined datatype elements */
if (MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) {
MPID_Datatype_get_size_macro(origin_datatype, predefined_dtp_size);
predefined_dtp_count = origin_count;
MPID_Datatype_get_extent_macro(origin_datatype, predefined_dtp_extent);
}
else {
MPIU_Assert(origin_dtp->basic_type != MPI_DATATYPE_NULL);
MPID_Datatype_get_size_macro(origin_dtp->basic_type, predefined_dtp_size);
predefined_dtp_count = data_sz / predefined_dtp_size;
MPID_Datatype_get_extent_macro(origin_dtp->basic_type, predefined_dtp_extent);
}
MPIU_Assert(predefined_dtp_count > 0 && predefined_dtp_size > 0 &&
predefined_dtp_extent > 0);
/* Calculate number of predefined elements in each stream unit, and
* total number of stream units. */
stream_elem_count = MPIDI_CH3U_Acc_stream_size / predefined_dtp_extent;
stream_unit_count = (predefined_dtp_count - 1) / stream_elem_count + 1;
MPIU_Assert(stream_elem_count > 0 && stream_unit_count > 0);
for (i = 0; i < stream_unit_count; i++) {
if (origin_dtp != NULL) {
MPID_Datatype_add_ref(origin_dtp);
}
if (target_dtp != NULL) {
MPID_Datatype_add_ref(target_dtp);
}
}
MPID_Datatype_is_contig(origin_datatype, &is_origin_contig);
MPID_Datatype_is_contig(target_datatype, &is_target_contig);
/* Judge if we can use IMMED data packet */
if (MPIR_DATATYPE_IS_PREDEFINED(origin_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype) && is_origin_contig && is_target_contig) {
if (data_sz <= MPIDI_RMA_IMMED_BYTES)
use_immed_pkt = TRUE;
}
/* Judge if this operation is an piggyback candidate. */
if (MPIR_DATATYPE_IS_PREDEFINED(origin_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
/* FIXME: currently we only piggyback LOCK flag with op using predefined datatypes
* for both origin and target data. We should extend this optimization to derived
* datatypes as well. */
if (data_sz <= MPIR_CVAR_CH3_RMA_OP_PIGGYBACK_LOCK_DATA_SIZE)
op_ptr->piggyback_lock_candidate = 1;
}
/************** Setting packet struct areas in operation ****************/
accum_pkt = &(op_ptr->pkt.accum);
if (use_immed_pkt) {
MPIDI_Pkt_init(accum_pkt, MPIDI_CH3_PKT_ACCUMULATE_IMMED);
}
else {
MPIDI_Pkt_init(accum_pkt, MPIDI_CH3_PKT_ACCUMULATE);
}
accum_pkt->addr = (char *) win_ptr->basic_info_table[target_rank].base_addr +
win_ptr->basic_info_table[target_rank].disp_unit * target_disp;
accum_pkt->count = target_count;
accum_pkt->datatype = target_datatype;
accum_pkt->info.dataloop_size = 0;
accum_pkt->op = op;
accum_pkt->target_win_handle = win_ptr->basic_info_table[target_rank].win_handle;
accum_pkt->source_win_handle = win_ptr->handle;
accum_pkt->flags = MPIDI_CH3_PKT_FLAG_NONE;
if (use_immed_pkt) {
void *src = (void *) origin_addr, *dest = (void *) (accum_pkt->info.data);
mpi_errno = immed_copy(src, dest, data_sz);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
mpi_errno = MPIDI_CH3I_Win_enqueue_op(win_ptr, op_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, target_rank, &made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD >= 0 &&
MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
while (MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_ACCUMULATE);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_Accumulate(const void *origin_addr, int origin_count, MPI_Datatype
origin_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPID_Win *win_ptr)
{
int mpi_errno=MPI_SUCCESS;
MPIDI_msg_sz_t data_sz;
int dt_contig ATTRIBUTE((unused)), rank;
MPI_Aint dt_true_lb ATTRIBUTE((unused));
MPID_Datatype *dtp;
MPIDI_VC_t *orig_vc, *target_vc;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_ACCUMULATE);
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
if (win_ptr->epoch_state == MPIDI_EPOCH_NONE && win_ptr->fence_issued) {
win_ptr->epoch_state = MPIDI_EPOCH_FENCE;
}
MPIU_ERR_CHKANDJUMP(win_ptr->epoch_state == MPIDI_EPOCH_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
MPIDI_Datatype_get_info(origin_count, origin_datatype,
dt_contig, data_sz, dtp, dt_true_lb);
if (data_sz == 0) {
goto fn_exit;
}
rank = win_ptr->comm_ptr->rank;
if (win_ptr->shm_allocated == TRUE && target_rank != rank && win_ptr->create_flavor != MPI_WIN_FLAVOR_SHARED) {
/* check if target is local and shared memory is allocated on window,
if so, we directly perform this operation on shared memory region. */
/* FIXME: Here we decide whether to perform SHM operations by checking if origin and target are on
the same node. However, in ch3:sock, even if origin and target are on the same node, they do
not within the same SHM region. Here we filter out ch3:sock by checking shm_allocated flag first,
which is only set to TRUE when SHM region is allocated in nemesis.
In future we need to figure out a way to check if origin and target are in the same "SHM comm".
*/
MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &target_vc);
}
/* Do =! rank first (most likely branch?) */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED ||
(win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id))
{
mpi_errno = MPIDI_CH3I_Shm_acc_op(origin_addr, origin_count, origin_datatype,
target_rank, target_disp, target_count, target_datatype,
op, win_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
else
{
MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank);
MPIDI_RMA_Op_t *new_ptr = NULL;
/* queue it up */
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_alloc);
mpi_errno = MPIDI_CH3I_RMA_Ops_alloc_tail(ops_list, &new_ptr);
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_alloc);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
/* If predefined and contiguous, use a simplified element */
if (MPIR_DATATYPE_IS_PREDEFINED(origin_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype) && enableShortACC) {
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
new_ptr->type = MPIDI_RMA_ACC_CONTIG;
/* Only the information needed for the contig/predefined acc */
/* Cast away const'ness for origin_address as
* MPIDI_RMA_Op_t contain both PUT and GET like ops */
new_ptr->origin_addr = (void *) origin_addr;
new_ptr->origin_count = origin_count;
new_ptr->origin_datatype = origin_datatype;
new_ptr->target_rank = target_rank;
new_ptr->target_disp = target_disp;
new_ptr->target_count = target_count;
new_ptr->target_datatype = target_datatype;
new_ptr->op = op;
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
goto fn_exit;
}
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
new_ptr->type = MPIDI_RMA_ACCUMULATE;
/* Cast away const'ness for origin_address as MPIDI_RMA_Op_t
* contain both PUT and GET like ops */
new_ptr->origin_addr = (void *) origin_addr;
new_ptr->origin_count = origin_count;
new_ptr->origin_datatype = origin_datatype;
new_ptr->target_rank = target_rank;
new_ptr->target_disp = target_disp;
new_ptr->target_count = target_count;
new_ptr->target_datatype = target_datatype;
new_ptr->op = op;
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
/* if source or target datatypes are derived, increment their
reference counts */
if (!MPIR_DATATYPE_IS_PREDEFINED(origin_datatype))
{
MPID_Datatype_get_ptr(origin_datatype, dtp);
MPID_Datatype_add_ref(dtp);
}
if (!MPIR_DATATYPE_IS_PREDEFINED(target_datatype))
{
MPID_Datatype_get_ptr(target_datatype, dtp);
MPID_Datatype_add_ref(dtp);
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_ACCUMULATE);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
