/** * Copyright (c) 2021 Mellanox Technologies. All rights reserved. * Copyright (c) 2022 NVIDIA Corporation. All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow * */ #include "coll_ucc_common.h" static inline ucc_status_t mca_coll_ucc_reduce_scatter_block_init(const void *sbuf, void *rbuf, int rcount, struct ompi_datatype_t *dtype, struct ompi_op_t *op, mca_coll_ucc_module_t *ucc_module, ucc_coll_req_h *req, mca_coll_ucc_req_t *coll_req) { ucc_datatype_t ucc_dt; ucc_reduction_op_t ucc_op; int comm_size = ompi_comm_size(ucc_module->comm); if (MPI_IN_PLACE == sbuf) { /* TODO: UCC defines inplace differently: data in rbuf of rank R is shifted by R * rcount */ UCC_VERBOSE(5, "inplace reduce_scatter_block is not supported"); return UCC_ERR_NOT_SUPPORTED; } ucc_dt = ompi_dtype_to_ucc_dtype(dtype); ucc_op = ompi_op_to_ucc_op(op); if (OPAL_UNLIKELY(COLL_UCC_DT_UNSUPPORTED == ucc_dt)) { UCC_VERBOSE(5, "ompi_datatype is not supported: dtype = %s", dtype->super.name); goto fallback; } if (OPAL_UNLIKELY(COLL_UCC_OP_UNSUPPORTED == ucc_op)) { UCC_VERBOSE(5, "ompi_op is not supported: op = %s", op->o_name); goto fallback; } ucc_coll_args_t coll = { .mask = 0, .flags = 0, .coll_type = UCC_COLL_TYPE_REDUCE_SCATTER, .src.info = { .buffer = (void*)sbuf, .count = ((size_t)rcount) * comm_size, .datatype = ucc_dt, .mem_type = UCC_MEMORY_TYPE_UNKNOWN }, .dst.info = { .buffer = rbuf, .count = rcount, .datatype = ucc_dt, .mem_type = UCC_MEMORY_TYPE_UNKNOWN }, .op = ucc_op, }; COLL_UCC_REQ_INIT(coll_req, req, coll, ucc_module); return UCC_OK; fallback: return UCC_ERR_NOT_SUPPORTED; } int mca_coll_ucc_reduce_scatter_block(const void *sbuf, void *rbuf, int rcount, struct ompi_datatype_t *dtype, struct ompi_op_t *op, struct ompi_communicator_t *comm, mca_coll_base_module_t *module) { mca_coll_ucc_module_t *ucc_module = (mca_coll_ucc_module_t*)module; ucc_coll_req_h req; UCC_VERBOSE(3, "running ucc reduce scatter block"); COLL_UCC_CHECK(mca_coll_ucc_reduce_scatter_block_init(sbuf, rbuf, rcount, dtype, op, ucc_module, &req, NULL)); COLL_UCC_POST_AND_CHECK(req); COLL_UCC_CHECK(coll_ucc_req_wait(req)); return OMPI_SUCCESS; fallback: UCC_VERBOSE(3, "running fallback reduce_scatter_block"); return ucc_module->previous_reduce_scatter_block(sbuf, rbuf, rcount, dtype, op, comm, ucc_module->previous_reduce_scatter_block_module); } int mca_coll_ucc_ireduce_scatter_block(const void *sbuf, void *rbuf, int rcount, struct ompi_datatype_t *dtype, struct ompi_op_t *op, struct ompi_communicator_t *comm, ompi_request_t** request, mca_coll_base_module_t *module) { mca_coll_ucc_module_t *ucc_module = (mca_coll_ucc_module_t*)module; ucc_coll_req_h req; mca_coll_ucc_req_t *coll_req = NULL; UCC_VERBOSE(3, "running ucc ireduce_scatter_block"); COLL_UCC_GET_REQ(coll_req); COLL_UCC_CHECK(mca_coll_ucc_reduce_scatter_block_init(sbuf, rbuf, rcount, dtype, op, ucc_module, &req, coll_req)); COLL_UCC_POST_AND_CHECK(req); *request = &coll_req->super; return OMPI_SUCCESS; fallback: UCC_VERBOSE(3, "running fallback ireduce_scatter_block"); if (coll_req) { mca_coll_ucc_req_free((ompi_request_t **)&coll_req); } return ucc_module->previous_ireduce_scatter_block(sbuf, rbuf, rcount, dtype, op, comm, request, ucc_module->previous_ireduce_scatter_block_module); }