int ompi_osc_ucx_get_accumulate(const 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, ptrdiff_t target_disp, int target_count, struct ompi_datatype_t *target_dt, struct ompi_op_t *op, struct ompi_win_t *win) { ompi_osc_ucx_module_t *module = (ompi_osc_ucx_module_t*) win->w_osc_module; ucp_ep_h ep = OSC_UCX_GET_EP(module->comm, target); int ret = OMPI_SUCCESS; ret = check_sync_state(module, target, false); if (ret != OMPI_SUCCESS) { return ret; } ret = start_atomicity(module, ep, target); if (ret != OMPI_SUCCESS) { return ret; } ret = ompi_osc_ucx_get(result_addr, result_count, result_dt, target, target_disp, target_count, target_dt, win); if (ret != OMPI_SUCCESS) { return ret; } if (op != &ompi_mpi_op_no_op.op) { if (op == &ompi_mpi_op_replace.op) { ret = ompi_osc_ucx_put(origin_addr, origin_count, origin_dt, target, target_disp, target_count, target_dt, win); if (ret != OMPI_SUCCESS) { return ret; } } else { void *temp_addr = NULL; uint32_t temp_count; ompi_datatype_t *temp_dt; ptrdiff_t temp_lb, temp_extent; ucs_status_t status; bool is_origin_contig = ompi_datatype_is_contiguous_memory_layout(origin_dt, origin_count); if (ompi_datatype_is_predefined(target_dt)) { temp_dt = target_dt; temp_count = target_count; } else { ret = ompi_osc_base_get_primitive_type_info(target_dt, &temp_dt, &temp_count); if (ret != OMPI_SUCCESS) { return ret; } } ompi_datatype_get_true_extent(temp_dt, &temp_lb, &temp_extent); temp_addr = malloc(temp_extent * temp_count); if (temp_addr == NULL) { return OMPI_ERR_TEMP_OUT_OF_RESOURCE; } ret = ompi_osc_ucx_get(temp_addr, (int)temp_count, temp_dt, target, target_disp, target_count, target_dt, win); if (ret != OMPI_SUCCESS) { return ret; } status = ucp_ep_flush(ep); if (status != UCS_OK) { opal_output_verbose(1, ompi_osc_base_framework.framework_output, "%s:%d: ucp_ep_flush failed: %d\n", __FILE__, __LINE__, status); return OMPI_ERROR; } if (ompi_datatype_is_predefined(origin_dt) || is_origin_contig) { ompi_op_reduce(op, (void *)origin_addr, temp_addr, (int)temp_count, temp_dt); } else { ucx_iovec_t *origin_ucx_iov = NULL; uint32_t origin_ucx_iov_count = 0; uint32_t origin_ucx_iov_idx = 0; ret = create_iov_list(origin_addr, origin_count, origin_dt, &origin_ucx_iov, &origin_ucx_iov_count); if (ret != OMPI_SUCCESS) { return ret; } if ((op != &ompi_mpi_op_maxloc.op && op != &ompi_mpi_op_minloc.op) || ompi_datatype_is_contiguous_memory_layout(temp_dt, temp_count)) { size_t temp_size; ompi_datatype_type_size(temp_dt, &temp_size); while (origin_ucx_iov_idx < origin_ucx_iov_count) { int curr_count = origin_ucx_iov[origin_ucx_iov_idx].len / temp_size; ompi_op_reduce(op, origin_ucx_iov[origin_ucx_iov_idx].addr, temp_addr, curr_count, temp_dt); temp_addr = (void *)((char *)temp_addr + curr_count * temp_size); origin_ucx_iov_idx++; } } else { int i; void *curr_origin_addr = origin_ucx_iov[origin_ucx_iov_idx].addr; for (i = 0; i < (int)temp_count; i++) { ompi_op_reduce(op, curr_origin_addr, (void *)((char *)temp_addr + i * temp_extent), 1, temp_dt); curr_origin_addr = (void *)((char *)curr_origin_addr + temp_extent); origin_ucx_iov_idx++; if (curr_origin_addr >= (void *)((char *)origin_ucx_iov[origin_ucx_iov_idx].addr + origin_ucx_iov[origin_ucx_iov_idx].len)) { origin_ucx_iov_idx++; curr_origin_addr = origin_ucx_iov[origin_ucx_iov_idx].addr; } } } free(origin_ucx_iov); } ret = ompi_osc_ucx_put(temp_addr, (int)temp_count, temp_dt, target, target_disp, target_count, target_dt, win); if (ret != OMPI_SUCCESS) { return ret; } status = ucp_ep_flush(ep); if (status != UCS_OK) { opal_output_verbose(1, ompi_osc_base_framework.framework_output, "%s:%d: ucp_ep_flush failed: %d\n", __FILE__, __LINE__, status); return OMPI_ERROR; } free(temp_addr); } } ret = end_atomicity(module, ep, target); return ret; }
static inline int ompi_osc_rdma_gacc_master (ompi_osc_rdma_sync_t *sync, 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_request_t *request) { ompi_osc_rdma_module_t *module = sync->module; struct iovec source_iovec[OMPI_OSC_RDMA_DECODE_MAX], target_iovec[OMPI_OSC_RDMA_DECODE_MAX]; const size_t acc_limit = (mca_osc_rdma_component.buffer_size >> 3); uint32_t source_primitive_count, target_primitive_count; opal_convertor_t source_convertor, target_convertor; uint32_t source_iov_count, target_iov_count; uint32_t source_iov_index, target_iov_index; ompi_datatype_t *source_primitive, *target_primitive; /* needed for opal_convertor_raw but not used */ size_t source_size, target_size; ompi_osc_rdma_request_t *subreq; size_t result_position; ptrdiff_t lb, extent; int ret, acc_len; bool done; (void) ompi_datatype_get_extent (target_datatype, &lb, &extent); target_address += lb; /* fast path for accumulate on built-in types */ if (OPAL_LIKELY((!source_count || ompi_datatype_is_predefined (source_datatype)) && ompi_datatype_is_predefined (target_datatype) && (!result_count || ompi_datatype_is_predefined (result_datatype)) && (target_datatype->super.size * target_count <= acc_limit))) { if (NULL == request) { OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, request); request->internal = true; request->type = result_datatype ? OMPI_OSC_RDMA_TYPE_GET_ACC : OMPI_OSC_RDMA_TYPE_ACC; } if (source_datatype) { (void) ompi_datatype_get_extent (source_datatype, &lb, &extent); source_buffer = (void *)((intptr_t) source_buffer + lb); } if (result_datatype) { (void) ompi_datatype_get_extent (result_datatype, &lb, &extent); result_buffer = (void *)((intptr_t) result_buffer + lb); } ret = ompi_osc_rdma_gacc_contig (sync, source_buffer, source_count, source_datatype, result_buffer, result_count, result_datatype, peer, target_address, target_handle, target_count, target_datatype, op, request); if (OPAL_LIKELY(OMPI_SUCCESS == ret)) { return OMPI_SUCCESS; } if (source_datatype) { /* the convertors will handle the lb */ (void) ompi_datatype_get_extent (source_datatype, &lb, &extent); source_buffer = (void *)((intptr_t) source_buffer - lb); } if (result_datatype) { (void) ompi_datatype_get_extent (result_datatype, &lb, &extent); result_buffer = (void *)((intptr_t) result_buffer - lb); } } /* the convertor will handle lb from here */ (void) ompi_datatype_get_extent (target_datatype, &lb, &extent); target_address -= lb; /* get the primitive datatype info */ ret = ompi_osc_base_get_primitive_type_info (target_datatype, &target_primitive, &target_primitive_count); if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) { /* target datatype is not made up of a single basic datatype */ return ret; } if (source_datatype) { ret = ompi_osc_base_get_primitive_type_info (source_datatype, &source_primitive, &source_primitive_count); if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) { /* target datatype is not made up of a single basic datatype */ return ret; } if (OPAL_UNLIKELY(source_primitive != target_primitive)) { return MPI_ERR_TYPE; } } /* prepare convertors for the source and target. these convertors will be used to determine the * contiguous segments within the source and target. */ /* the source may be NULL if using MPI_OP_NO_OP with MPI_Get_accumulate */ if (source_datatype) { OBJ_CONSTRUCT(&source_convertor, opal_convertor_t); ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &source_datatype->super, source_count, source_buffer, 0, &source_convertor); if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) { return ret; } } /* target_datatype can never be NULL */ OBJ_CONSTRUCT(&target_convertor, opal_convertor_t); ret = opal_convertor_copy_and_prepare_for_send (ompi_mpi_local_convertor, &target_datatype->super, target_count, (void *) (intptr_t) target_address, 0, &target_convertor); if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) { return ret; } if (request) { /* keep the request from completing until all the transfers have started */ request->outstanding_requests = 1; } target_iov_index = 0; target_iov_count = 0; result_position = 0; do { /* decode segments of the source data */ source_iov_count = OMPI_OSC_RDMA_DECODE_MAX; source_iov_index = 0; /* opal_convertor_raw returns done when it has reached the end of the data */ if (!source_datatype) { done = true; source_iovec[0].iov_len = (size_t) -1; source_iovec[0].iov_base = NULL; source_iov_count = 1; } else { done = opal_convertor_raw (&source_convertor, source_iovec, &source_iov_count, &source_size); } /* loop on the target segments until we have exhaused the decoded source data */ while (source_iov_index != source_iov_count) { if (target_iov_index == target_iov_count) { /* decode segments of the target buffer */ target_iov_count = OMPI_OSC_RDMA_DECODE_MAX; target_iov_index = 0; (void) opal_convertor_raw (&target_convertor, target_iovec, &target_iov_count, &target_size); } /* we already checked that the target was large enough. this should be impossible */ assert (0 != target_iov_count); /* determine how much to put in this operation */ acc_len = min(target_iovec[target_iov_index].iov_len, source_iovec[source_iov_index].iov_len); acc_len = min((size_t) acc_len, acc_limit); /* execute the get */ OMPI_OSC_RDMA_REQUEST_ALLOC(module, peer, subreq); subreq->internal = true; subreq->parent_request = request; if (request) { (void) OPAL_THREAD_ADD32 (&request->outstanding_requests, 1); } if (result_datatype) { /* prepare a convertor for this part of the result */ opal_convertor_copy_and_prepare_for_recv (ompi_mpi_local_convertor, &result_datatype->super, result_count, result_buffer, 0, &subreq->convertor); opal_convertor_set_position (&subreq->convertor, &result_position); subreq->type = OMPI_OSC_RDMA_TYPE_GET_ACC; } else { subreq->type = OMPI_OSC_RDMA_TYPE_ACC; } OPAL_OUTPUT_VERBOSE((60, ompi_osc_base_framework.framework_output, "target index = %d, target = {%p, %lu}, source_index = %d, source = {%p, %lu}, result = %p, result position = %lu, " "acc_len = %d, count = %lu", target_iov_index, target_iovec[target_iov_index].iov_base, (unsigned long) target_iovec[target_iov_index].iov_len, source_iov_index, source_iovec[source_iov_index].iov_base, (unsigned long) source_iovec[source_iov_index].iov_len, result_buffer, (unsigned long) result_position, acc_len, (unsigned long)(acc_len / target_primitive->super.size))); ret = ompi_osc_rdma_gacc_contig (sync, source_iovec[source_iov_index].iov_base, acc_len / target_primitive->super.size, target_primitive, NULL, 0, NULL, peer, (uint64_t) (intptr_t) target_iovec[target_iov_index].iov_base, target_handle, acc_len / target_primitive->super.size, target_primitive, op, subreq); if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) { if (OPAL_UNLIKELY(OMPI_ERR_OUT_OF_RESOURCE != ret)) { /* something bad happened. need to figure out how to handle these errors */ return ret; } /* progress and try again */ ompi_osc_rdma_progress (module); continue; } /* adjust io vectors */ target_iovec[target_iov_index].iov_len -= acc_len; source_iovec[source_iov_index].iov_len -= acc_len; target_iovec[target_iov_index].iov_base = (void *)((intptr_t) target_iovec[target_iov_index].iov_base + acc_len); source_iovec[source_iov_index].iov_base = (void *)((intptr_t) source_iovec[source_iov_index].iov_base + acc_len); result_position += acc_len; source_iov_index += !source_datatype || (0 == source_iovec[source_iov_index].iov_len); target_iov_index += (0 == target_iovec[target_iov_index].iov_len); } } while (!done); if (request) { /* release our reference so the request can complete */ (void) OPAL_THREAD_ADD32 (&request->outstanding_requests, -1); } if (source_datatype) { opal_convertor_cleanup (&source_convertor); OBJ_DESTRUCT(&source_convertor); } opal_convertor_cleanup (&target_convertor); OBJ_DESTRUCT(&target_convertor); return OMPI_SUCCESS; }