static inline int ompi_osc_pt2pt_acc_self (ompi_osc_pt2pt_sync_t *pt2pt_sync, const void *source, int source_count, ompi_datatype_t *source_datatype,
OPAL_PTRDIFF_TYPE target_disp, int target_count, ompi_datatype_t *target_datatype,
ompi_op_t *op, ompi_osc_pt2pt_module_t *module, ompi_osc_pt2pt_request_t *request)
{
void *target = (unsigned char*) module->baseptr +
((unsigned long) target_disp * module->disp_unit);
int ret;
/* if we are in active target mode wait until all post messages arrive */
ompi_osc_pt2pt_sync_wait_expected (pt2pt_sync);
ompi_osc_pt2pt_accumulate_lock (module);
if (&ompi_mpi_op_replace.op != op) {
ret = ompi_osc_base_sndrcv_op (source, source_count, source_datatype, target, target_count, target_datatype, op);
} else {
ret = ompi_datatype_sndrcv ((void *)source, source_count, source_datatype, target, target_count, target_datatype);
}
ompi_osc_pt2pt_accumulate_unlock (module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_OUTPUT_VERBOSE((5, ompi_osc_base_framework.framework_output,
"ompi_osc_pt2pt_acc_self: failed performing accumulate operation. ret = %d", ret));
return ret;
}
if (request) {
ompi_osc_pt2pt_request_complete (request, MPI_SUCCESS);
}
return OMPI_SUCCESS;
}
static inline int ompi_osc_rdma_gacc_self (void *source, int source_count, ompi_datatype_t *source_datatype,
void *result, int result_count, ompi_datatype_t *result_datatype,
OPAL_PTRDIFF_TYPE target_disp, int target_count, ompi_datatype_t *target_datatype,
ompi_op_t *op, ompi_osc_rdma_module_t *module, ompi_osc_rdma_request_t *request)
{
void *target = (unsigned char*) module->baseptr +
((unsigned long) target_disp * module->disp_unit);
int ret;
/* if we are in active target mode wait until all post messages arrive */
if (module->sc_group && !module->active_eager_send_active) {
OPAL_THREAD_LOCK(&module->lock);
while (0 != module->num_post_msgs) {
opal_condition_wait(&module->cond, &module->lock);
}
OPAL_THREAD_UNLOCK(&module->lock);
}
if (!(module->passive_target_access_epoch || module->active_eager_send_active)) {
return OMPI_ERR_RMA_SYNC;
}
ompi_osc_rdma_accumulate_lock (module);
do {
ret = ompi_datatype_sndrcv (target, target_count, target_datatype,
result, result_count, result_datatype);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_OUTPUT_VERBOSE((5, ompi_osc_base_framework.framework_output,
"ompi_osc_rdma_gacc_self: failed copying to the target buffer. ret = %d", ret));
break;
}
if (&ompi_mpi_op_no_op.op != op) {
if (&ompi_mpi_op_replace.op != op) {
ret = ompi_osc_base_sndrcv_op (source, source_count, source_datatype, target, target_count, target_datatype, op);
} else {
ret = ompi_datatype_sndrcv (source, source_count, source_datatype, target, target_count, target_datatype);
}
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_OUTPUT_VERBOSE((5, ompi_osc_base_framework.framework_output,
"ompi_osc_rdma_gacc_self: failed performing accumulate operation. ret = %d", ret));
break;
}
} while (0);
ompi_osc_rdma_accumulate_unlock (module);
if (request) {
/* NTH: is it ok to use an ompi error code here? */
ompi_osc_rdma_request_complete (request, ret);
}
return OMPI_SUCCESS;
}
int
ompi_osc_sm_rget_accumulate(void *origin_addr,
int origin_count,
struct ompi_datatype_t *origin_dt,
void *result_addr,
int result_count,
struct ompi_datatype_t *result_dt,
int target,
MPI_Aint target_disp,
int target_count,
struct ompi_datatype_t *target_dt,
struct ompi_op_t *op,
struct ompi_win_t *win,
struct ompi_request_t **ompi_req)
{
int ret;
ompi_osc_sm_request_t *request;
ompi_osc_sm_module_t *module =
(ompi_osc_sm_module_t*) win->w_osc_module;
void *remote_address;
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"rget_accumulate: 0x%lx, %d, %s, %d, %d, %d, %s, %s, 0x%lx",
(unsigned long) origin_addr, origin_count,
origin_dt->name, target, (int) target_disp,
target_count, target_dt->name,
op->o_name,
(unsigned long) win));
OMPI_OSC_SM_REQUEST_ALLOC(win, request);
if (NULL == request) return OMPI_ERR_OUT_OF_RESOURCE;
*ompi_req = &request->super;
remote_address = ((char*) (module->bases[target])) + module->disp_units[target] * target_disp;
opal_atomic_lock(&module->node_states[target].accumulate_lock);
ret = ompi_datatype_sndrcv(remote_address, target_count, target_dt,
result_addr, result_count, result_dt);
if (OMPI_SUCCESS != ret || op == &ompi_mpi_op_no_op.op) goto done;
if (op == &ompi_mpi_op_replace.op) {
ret = ompi_datatype_sndrcv(origin_addr, origin_count, origin_dt,
remote_address, target_count, target_dt);
} else {
ret = ompi_osc_base_sndrcv_op(origin_addr, origin_count, origin_dt,
remote_address, target_count, target_dt,
op);
}
done:
opal_atomic_unlock(&module->node_states[target].accumulate_lock);
OMPI_OSC_SM_REQUEST_COMPLETE(request);
return ret;
}
static int ompi_osc_rdma_gacc_local (const void *source_buffer, int source_count, ompi_datatype_t *source_datatype,
void *result_buffer, int result_count, ompi_datatype_t *result_datatype,
ompi_osc_rdma_peer_t *peer, uint64_t target_address,
mca_btl_base_registration_handle_t *target_handle, int target_count,
ompi_datatype_t *target_datatype, ompi_op_t *op, ompi_osc_rdma_module_t *module,
ompi_osc_rdma_request_t *request)
{
int ret = OMPI_SUCCESS;
do {
if (!ompi_osc_rdma_peer_is_exclusive (peer)) {
(void) ompi_osc_rdma_lock_acquire_exclusive (module, peer, offsetof (ompi_osc_rdma_state_t, accumulate_lock));
}
if (NULL != result_buffer) {
/* get accumulate */
ret = ompi_datatype_sndrcv ((void *) (intptr_t) target_address, target_count, target_datatype,
result_buffer, result_count, result_datatype);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
break;
}
}
if (&ompi_mpi_op_no_op.op != op) {
if (&ompi_mpi_op_replace.op != op) {
ret = ompi_osc_base_sndrcv_op (source_buffer, source_count, source_datatype, (void *) (intptr_t) target_address,
target_count, target_datatype, op);
} else {
ret = ompi_datatype_sndrcv (source_buffer, source_count, source_datatype, (void *) (intptr_t) target_address,
target_count, target_datatype);
}
}
if (!ompi_osc_rdma_peer_is_exclusive (peer)) {
(void) ompi_osc_rdma_lock_release_exclusive (module, peer, offsetof (ompi_osc_rdma_state_t, accumulate_lock));
}
} while (0);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_OUTPUT_VERBOSE((10, ompi_osc_base_framework.framework_output,
"ompi_osc_rdma_gacc_self: failed performing accumulate operation. ret = %d", ret));
return ret;
}
if (request) {
/* NTH: is it ok to use an ompi error code here? */
ompi_osc_rdma_request_complete (request, ret);
}
return ret;
}
static inline int ompi_osc_pt2pt_gacc_self (ompi_osc_pt2pt_sync_t *pt2pt_sync, const void *source, int source_count, ompi_datatype_t *source_datatype,
void *result, int result_count, ompi_datatype_t *result_datatype,
OPAL_PTRDIFF_TYPE target_disp, int target_count, ompi_datatype_t *target_datatype,
ompi_op_t *op, ompi_osc_pt2pt_module_t *module, ompi_osc_pt2pt_request_t *request)
{
void *target = (unsigned char*) module->baseptr +
((unsigned long) target_disp * module->disp_unit);
int ret;
OPAL_OUTPUT_VERBOSE((MCA_BASE_VERBOSE_TRACE, ompi_osc_base_framework.framework_output, "ompi_osc_pt2pt_gacc_self: starting local "
"get accumulate"));
ompi_osc_pt2pt_accumulate_lock (module);
do {
ret = ompi_datatype_sndrcv (target, target_count, target_datatype,
result, result_count, result_datatype);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_OUTPUT_VERBOSE((5, ompi_osc_base_framework.framework_output,
"ompi_osc_pt2pt_gacc_self: failed copying to the target buffer. ret = %d", ret));
break;
}
if (&ompi_mpi_op_no_op.op != op) {
if (&ompi_mpi_op_replace.op != op) {
ret = ompi_osc_base_sndrcv_op (source, source_count, source_datatype, target, target_count, target_datatype, op);
} else {
ret = ompi_datatype_sndrcv ((void *)source, source_count, source_datatype, target, target_count, target_datatype);
}
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
OPAL_OUTPUT_VERBOSE((5, ompi_osc_base_framework.framework_output,
"ompi_osc_pt2pt_gacc_self: failed performing accumulate operation. ret = %d", ret));
break;
}
} while (0);
ompi_osc_pt2pt_accumulate_unlock (module);
OPAL_OUTPUT_VERBOSE((MCA_BASE_VERBOSE_TRACE, ompi_osc_base_framework.framework_output, "ompi_osc_pt2pt_gacc_self: local get "
"accumulate complete"));
if (request) {
/* NTH: is it ok to use an ompi error code here? */
ompi_osc_pt2pt_request_complete (request, ret);
}
return OMPI_SUCCESS;
}
int
ompi_osc_sm_accumulate(void *origin_addr,
int origin_count,
struct ompi_datatype_t *origin_dt,
int target,
OPAL_PTRDIFF_TYPE target_disp,
int target_count,
struct ompi_datatype_t *target_dt,
struct ompi_op_t *op,
struct ompi_win_t *win)
{
int ret;
ompi_osc_sm_module_t *module =
(ompi_osc_sm_module_t*) win->w_osc_module;
void *remote_address;
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
"accumulate: 0x%lx, %d, %s, %d, %d, %d, %s, %s, 0x%lx",
(unsigned long) origin_addr, origin_count,
origin_dt->name, target, (int) target_disp,
target_count, target_dt->name,
op->o_name,
(unsigned long) win));
remote_address = ((char*) (module->bases[target])) + module->disp_units[target] * target_disp;
opal_atomic_lock(&module->node_states[target].accumulate_lock);
if (op == &ompi_mpi_op_replace.op) {
ret = ompi_datatype_sndrcv(origin_addr, origin_count, origin_dt,
remote_address, target_count, target_dt);
} else {
ret = ompi_osc_base_sndrcv_op(origin_addr, origin_count, origin_dt,
remote_address, target_count, target_dt,
op);
}
opal_atomic_unlock(&module->node_states[target].accumulate_lock);
return ret;
}