/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */ /* * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana * University Research and Technology * Corporation. All rights reserved. * Copyright (c) 2004-2017 The University of Tennessee and The University * of Tennessee Research Foundation. All rights * reserved. * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, * University of Stuttgart. All rights reserved. * Copyright (c) 2004-2005 The Regents of the University of California. * All rights reserved. * Copyright (c) 2013 Los Alamos National Security, LLC. All Rights * reserved. * Copyright (c) 2015-2016 Research Organization for Information Science * and Technology (RIST). All rights reserved. * Copyright (c) 2016-2017 IBM Corporation. All rights reserved. * Copyright (c) 2018 Siberian State University of Telecommunications * and Information Science. All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow * * $HEADER$ */ #include "ompi_config.h" #include "mpi.h" #include "opal/util/bit_ops.h" #include "ompi/constants.h" #include "ompi/datatype/ompi_datatype.h" #include "ompi/communicator/communicator.h" #include "ompi/mca/coll/coll.h" #include "ompi/mca/coll/base/coll_tags.h" #include "ompi/mca/pml/pml.h" #include "ompi/op/op.h" #include "ompi/mca/coll/base/coll_base_functions.h" #include "coll_base_topo.h" #include "coll_base_util.h" int mca_coll_base_reduce_local(const void *inbuf, void *inoutbuf, int count, struct ompi_datatype_t * dtype, struct ompi_op_t * op, mca_coll_base_module_t *module) { /* XXX -- CONST -- do not cast away const -- update ompi/op/op.h */ ompi_op_reduce(op, (void *)inbuf, inoutbuf, count, dtype); return OMPI_SUCCESS; } /** * This is a generic implementation of the reduce protocol. It used the tree * provided as an argument and execute all operations using a segment of * count times a datatype. * For the last communication it will update the count in order to limit * the number of datatype to the original count (original_count) * * Note that for non-commutative operations we cannot save memory copy * for the first block: thus we must copy sendbuf to accumbuf on intermediate * to keep the optimized loop happy. */ int ompi_coll_base_reduce_generic( const void* sendbuf, void* recvbuf, int original_count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, mca_coll_base_module_t *module, ompi_coll_tree_t* tree, int count_by_segment, int max_outstanding_reqs ) { char *inbuf[2] = {NULL, NULL}, *inbuf_free[2] = {NULL, NULL}; char *accumbuf = NULL, *accumbuf_free = NULL; char *local_op_buffer = NULL, *sendtmpbuf = NULL; ptrdiff_t extent, size, gap = 0, segment_increment; ompi_request_t **sreq = NULL, *reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL}; int num_segments, line, ret, segindex, i, rank; int recvcount, prevcount, inbi; /** * Determine number of segments and number of elements * sent per operation */ ompi_datatype_type_extent( datatype, &extent ); num_segments = (int)(((size_t)original_count + (size_t)count_by_segment - (size_t)1) / (size_t)count_by_segment); segment_increment = (ptrdiff_t)count_by_segment * extent; sendtmpbuf = (char*) sendbuf; if( sendbuf == MPI_IN_PLACE ) { sendtmpbuf = (char *)recvbuf; } OPAL_OUTPUT((ompi_coll_base_framework.framework_output, "coll:base:reduce_generic count %d, msg size %ld, segsize %ld, max_requests %d", original_count, (unsigned long)((ptrdiff_t)num_segments * (ptrdiff_t)segment_increment), (unsigned long)segment_increment, max_outstanding_reqs)); rank = ompi_comm_rank(comm); /* non-leaf nodes - wait for children to send me data & forward up (if needed) */ if( tree->tree_nextsize > 0 ) { ptrdiff_t real_segment_size; /* handle non existant recv buffer (i.e. its NULL) and protect the recv buffer on non-root nodes */ accumbuf = (char*)recvbuf; if( (NULL == accumbuf) || (root != rank) ) { /* Allocate temporary accumulator buffer. */ size = opal_datatype_span(&datatype->super, original_count, &gap); accumbuf_free = (char*)malloc(size); if (accumbuf_free == NULL) { line = __LINE__; ret = -1; goto error_hndl; } accumbuf = accumbuf_free - gap; } /* If this is a non-commutative operation we must copy sendbuf to the accumbuf, in order to simplfy the loops */ if (!ompi_op_is_commute(op) && MPI_IN_PLACE != sendbuf) { ompi_datatype_copy_content_same_ddt(datatype, original_count, (char*)accumbuf, (char*)sendtmpbuf); } /* Allocate two buffers for incoming segments */ real_segment_size = opal_datatype_span(&datatype->super, count_by_segment, &gap); inbuf_free[0] = (char*) malloc(real_segment_size); if( inbuf_free[0] == NULL ) { line = __LINE__; ret = -1; goto error_hndl; } inbuf[0] = inbuf_free[0] - gap; /* if there is chance to overlap communication - allocate second buffer */ if( (num_segments > 1) || (tree->tree_nextsize > 1) ) { inbuf_free[1] = (char*) malloc(real_segment_size); if( inbuf_free[1] == NULL ) { line = __LINE__; ret = -1; goto error_hndl; } inbuf[1] = inbuf_free[1] - gap; } /* reset input buffer index and receive count */ inbi = 0; recvcount = 0; /* for each segment */ for( segindex = 0; segindex <= num_segments; segindex++ ) { prevcount = recvcount; /* recvcount - number of elements in current segment */ recvcount = count_by_segment; if( segindex == (num_segments-1) ) recvcount = original_count - (ptrdiff_t)count_by_segment * (ptrdiff_t)segindex; /* for each child */ for( i = 0; i < tree->tree_nextsize; i++ ) { /** * We try to overlap communication: * either with next segment or with the next child */ /* post irecv for current segindex on current child */ if( segindex < num_segments ) { void* local_recvbuf = inbuf[inbi]; if( 0 == i ) { /* for the first step (1st child per segment) and * commutative operations we might be able to irecv * directly into the accumulate buffer so that we can * reduce(op) this with our sendbuf in one step as * ompi_op_reduce only has two buffer pointers, * this avoids an extra memory copy. * * BUT if the operation is non-commutative or * we are root and are USING MPI_IN_PLACE this is wrong! */ if( (ompi_op_is_commute(op)) && !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) { local_recvbuf = accumbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment; } } ret = MCA_PML_CALL(irecv(local_recvbuf, recvcount, datatype, tree->tree_next[i], MCA_COLL_BASE_TAG_REDUCE, comm, &reqs[inbi])); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl;} } /* wait for previous req to complete, if any. if there are no requests reqs[inbi ^1] will be MPI_REQUEST_NULL. */ /* wait on data from last child for previous segment */ ret = ompi_request_wait(&reqs[inbi ^ 1], MPI_STATUSES_IGNORE ); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } local_op_buffer = inbuf[inbi ^ 1]; if( i > 0 ) { /* our first operation is to combine our own [sendbuf] data * with the data we recvd from down stream (but only * the operation is commutative and if we are not root and * not using MPI_IN_PLACE) */ if( 1 == i ) { if( (ompi_op_is_commute(op)) && !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) { local_op_buffer = sendtmpbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment; } } /* apply operation */ ompi_op_reduce(op, local_op_buffer, accumbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment, recvcount, datatype ); } else if ( segindex > 0 ) { void* accumulator = accumbuf + (ptrdiff_t)(segindex-1) * (ptrdiff_t)segment_increment; if( tree->tree_nextsize <= 1 ) { if( (ompi_op_is_commute(op)) && !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) { local_op_buffer = sendtmpbuf + (ptrdiff_t)(segindex-1) * (ptrdiff_t)segment_increment; } } ompi_op_reduce(op, local_op_buffer, accumulator, prevcount, datatype ); /* all reduced on available data this step (i) complete, * pass to the next process unless you are the root. */ if (rank != tree->tree_root) { /* send combined/accumulated data to parent */ ret = MCA_PML_CALL( send( accumulator, prevcount, datatype, tree->tree_prev, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_STANDARD, comm) ); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } } /* we stop when segindex = number of segments (i.e. we do num_segment+1 steps for pipelining */ if (segindex == num_segments) break; } /* update input buffer index */ inbi = inbi ^ 1; } /* end of for each child */ } /* end of for each segment */ /* clean up */ if( inbuf_free[0] != NULL) free(inbuf_free[0]); if( inbuf_free[1] != NULL) free(inbuf_free[1]); if( accumbuf_free != NULL ) free(accumbuf_free); } /* leaf nodes Depending on the value of max_outstanding_reqs and the number of segments we have two options: - send all segments using blocking send to the parent, or - avoid overflooding the parent nodes by limiting the number of outstanding requests to max_oustanding_reqs. TODO/POSSIBLE IMPROVEMENT: If there is a way to determine the eager size for the current communication, synchronization should be used only when the message/segment size is smaller than the eager size. */ else { /* If the number of segments is less than a maximum number of oustanding requests or there is no limit on the maximum number of outstanding requests, we send data to the parent using blocking send */ if ((0 == max_outstanding_reqs) || (num_segments <= max_outstanding_reqs)) { segindex = 0; while ( original_count > 0) { if (original_count < count_by_segment) { count_by_segment = original_count; } ret = MCA_PML_CALL( send((char*)sendbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment, count_by_segment, datatype, tree->tree_prev, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_STANDARD, comm) ); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } segindex++; original_count -= count_by_segment; } } /* Otherwise, introduce flow control: - post max_outstanding_reqs non-blocking synchronous send, - for remaining segments - wait for a ssend to complete, and post the next one. - wait for all outstanding sends to complete. */ else { int creq = 0; sreq = ompi_coll_base_comm_get_reqs(module->base_data, max_outstanding_reqs); if (NULL == sreq) { line = __LINE__; ret = -1; goto error_hndl; } /* post first group of requests */ for (segindex = 0; segindex < max_outstanding_reqs; segindex++) { ret = MCA_PML_CALL( isend((char*)sendbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment, count_by_segment, datatype, tree->tree_prev, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_SYNCHRONOUS, comm, &sreq[segindex]) ); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } original_count -= count_by_segment; } creq = 0; while ( original_count > 0 ) { /* wait on a posted request to complete */ ret = ompi_request_wait(&sreq[creq], MPI_STATUS_IGNORE); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } if( original_count < count_by_segment ) { count_by_segment = original_count; } ret = MCA_PML_CALL( isend((char*)sendbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment, count_by_segment, datatype, tree->tree_prev, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_SYNCHRONOUS, comm, &sreq[creq]) ); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } creq = (creq + 1) % max_outstanding_reqs; segindex++; original_count -= count_by_segment; } /* Wait on the remaining request to complete */ ret = ompi_request_wait_all( max_outstanding_reqs, sreq, MPI_STATUSES_IGNORE ); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } } } return OMPI_SUCCESS; error_hndl: /* error handler */ OPAL_OUTPUT (( ompi_coll_base_framework.framework_output, "ERROR_HNDL: node %d file %s line %d error %d\n", rank, __FILE__, line, ret )); (void)line; // silence compiler warning if( inbuf_free[0] != NULL ) free(inbuf_free[0]); if( inbuf_free[1] != NULL ) free(inbuf_free[1]); if( accumbuf_free != NULL ) free(accumbuf); if( NULL != sreq ) { ompi_coll_base_free_reqs(sreq, max_outstanding_reqs); } return ret; } /* Attention: this version of the reduce operations does not work for: - non-commutative operations - segment sizes which are not multiplies of the extent of the datatype meaning that at least one datatype must fit in the segment ! */ int ompi_coll_base_reduce_intra_chain( const void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, mca_coll_base_module_t *module, uint32_t segsize, int fanout, int max_outstanding_reqs ) { int segcount = count; size_t typelng; mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module; mca_coll_base_comm_t *data = base_module->base_data; OPAL_OUTPUT((ompi_coll_base_framework.framework_output,"coll:base:reduce_intra_chain rank %d fo %d ss %5d", ompi_comm_rank(comm), fanout, segsize)); COLL_BASE_UPDATE_CHAIN( comm, base_module, root, fanout ); /** * Determine number of segments and number of elements * sent per operation */ ompi_datatype_type_size( datatype, &typelng ); COLL_BASE_COMPUTED_SEGCOUNT( segsize, typelng, segcount ); return ompi_coll_base_reduce_generic( sendbuf, recvbuf, count, datatype, op, root, comm, module, data->cached_chain, segcount, max_outstanding_reqs ); } int ompi_coll_base_reduce_intra_pipeline( const void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, mca_coll_base_module_t *module, uint32_t segsize, int max_outstanding_reqs ) { int segcount = count; size_t typelng; mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module; mca_coll_base_comm_t *data = base_module->base_data; OPAL_OUTPUT((ompi_coll_base_framework.framework_output,"coll:base:reduce_intra_pipeline rank %d ss %5d", ompi_comm_rank(comm), segsize)); COLL_BASE_UPDATE_PIPELINE( comm, base_module, root ); /** * Determine number of segments and number of elements * sent per operation */ ompi_datatype_type_size( datatype, &typelng ); COLL_BASE_COMPUTED_SEGCOUNT( segsize, typelng, segcount ); return ompi_coll_base_reduce_generic( sendbuf, recvbuf, count, datatype, op, root, comm, module, data->cached_pipeline, segcount, max_outstanding_reqs ); } int ompi_coll_base_reduce_intra_binary( const void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, mca_coll_base_module_t *module, uint32_t segsize, int max_outstanding_reqs ) { int segcount = count; size_t typelng; mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module; mca_coll_base_comm_t *data = base_module->base_data; OPAL_OUTPUT((ompi_coll_base_framework.framework_output,"coll:base:reduce_intra_binary rank %d ss %5d", ompi_comm_rank(comm), segsize)); COLL_BASE_UPDATE_BINTREE( comm, base_module, root ); /** * Determine number of segments and number of elements * sent per operation */ ompi_datatype_type_size( datatype, &typelng ); COLL_BASE_COMPUTED_SEGCOUNT( segsize, typelng, segcount ); return ompi_coll_base_reduce_generic( sendbuf, recvbuf, count, datatype, op, root, comm, module, data->cached_bintree, segcount, max_outstanding_reqs ); } int ompi_coll_base_reduce_intra_binomial( const void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, mca_coll_base_module_t *module, uint32_t segsize, int max_outstanding_reqs ) { int segcount = count; size_t typelng; mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module; mca_coll_base_comm_t *data = base_module->base_data; OPAL_OUTPUT((ompi_coll_base_framework.framework_output,"coll:base:reduce_intra_binomial rank %d ss %5d", ompi_comm_rank(comm), segsize)); COLL_BASE_UPDATE_IN_ORDER_BMTREE( comm, base_module, root ); /** * Determine number of segments and number of elements * sent per operation */ ompi_datatype_type_size( datatype, &typelng ); COLL_BASE_COMPUTED_SEGCOUNT( segsize, typelng, segcount ); return ompi_coll_base_reduce_generic( sendbuf, recvbuf, count, datatype, op, root, comm, module, data->cached_in_order_bmtree, segcount, max_outstanding_reqs ); } /* * reduce_intra_in_order_binary * * Function: Logarithmic reduce operation for non-commutative operations. * Acecpts: same as MPI_Reduce() * Returns: MPI_SUCCESS or error code */ int ompi_coll_base_reduce_intra_in_order_binary( const void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, mca_coll_base_module_t *module, uint32_t segsize, int max_outstanding_reqs ) { int ret, rank, size, io_root, segcount = count; void *use_this_sendbuf = NULL; void *use_this_recvbuf = NULL; char *tmpbuf_free = NULL; size_t typelng; mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module; mca_coll_base_comm_t *data = base_module->base_data; rank = ompi_comm_rank(comm); size = ompi_comm_size(comm); OPAL_OUTPUT((ompi_coll_base_framework.framework_output,"coll:base:reduce_intra_in_order_binary rank %d ss %5d", rank, segsize)); COLL_BASE_UPDATE_IN_ORDER_BINTREE( comm, base_module ); /** * Determine number of segments and number of elements * sent per operation */ ompi_datatype_type_size( datatype, &typelng ); COLL_BASE_COMPUTED_SEGCOUNT( segsize, typelng, segcount ); /* An in-order binary tree must use root (size-1) to preserve the order of operations. Thus, if root is not rank (size - 1), then we must handle 1. MPI_IN_PLACE option on real root, and 2. we must allocate temporary recvbuf on rank (size - 1). Note that generic function must be careful not to switch order of operations for non-commutative ops. */ io_root = size - 1; use_this_sendbuf = (void *)sendbuf; use_this_recvbuf = recvbuf; if (io_root != root) { ptrdiff_t dsize, gap = 0; char *tmpbuf; dsize = opal_datatype_span(&datatype->super, count, &gap); if ((root == rank) && (MPI_IN_PLACE == sendbuf)) { tmpbuf_free = (char *) malloc(dsize); if (NULL == tmpbuf_free) { return MPI_ERR_INTERN; } tmpbuf = tmpbuf_free - gap; ompi_datatype_copy_content_same_ddt(datatype, count, (char*)tmpbuf, (char*)recvbuf); use_this_sendbuf = tmpbuf; } else if (io_root == rank) { tmpbuf_free = (char *) malloc(dsize); if (NULL == tmpbuf_free) { return MPI_ERR_INTERN; } tmpbuf = tmpbuf_free - gap; use_this_recvbuf = tmpbuf; } } /* Use generic reduce with in-order binary tree topology and io_root */ ret = ompi_coll_base_reduce_generic( use_this_sendbuf, use_this_recvbuf, count, datatype, op, io_root, comm, module, data->cached_in_order_bintree, segcount, max_outstanding_reqs ); if (MPI_SUCCESS != ret) { return ret; } /* Clean up */ if (io_root != root) { if (root == rank) { /* Receive result from rank io_root to recvbuf */ ret = MCA_PML_CALL(recv(recvbuf, count, datatype, io_root, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE)); if (MPI_SUCCESS != ret) { return ret; } } else if (io_root == rank) { /* Send result from use_this_recvbuf to root */ ret = MCA_PML_CALL(send(use_this_recvbuf, count, datatype, root, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_STANDARD, comm)); if (MPI_SUCCESS != ret) { return ret; } } } if (NULL != tmpbuf_free) { free(tmpbuf_free); } return MPI_SUCCESS; } /* * Linear functions are copied from the BASIC coll module * they do not segment the message and are simple implementations * but for some small number of nodes and/or small data sizes they * are just as fast as base/tree based segmenting operations * and as such may be selected by the decision functions * These are copied into this module due to the way we select modules * in V1. i.e. in V2 we will handle this differently and so will not * have to duplicate code. * GEF Oct05 after asking Jeff. */ /* * reduce_lin_intra * * Function: - reduction using O(N) algorithm * Accepts: - same as MPI_Reduce() * Returns: - MPI_SUCCESS or error code */ int ompi_coll_base_reduce_intra_basic_linear(const void *sbuf, void *rbuf, int count, struct ompi_datatype_t *dtype, struct ompi_op_t *op, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module) { int i, rank, err, size; ptrdiff_t extent, dsize, gap = 0; char *free_buffer = NULL; char *pml_buffer = NULL; char *inplace_temp_free = NULL; char *inbuf; /* Initialize */ rank = ompi_comm_rank(comm); size = ompi_comm_size(comm); /* If not root, send data to the root. */ if (rank != root) { err = MCA_PML_CALL(send(sbuf, count, dtype, root, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_STANDARD, comm)); return err; } dsize = opal_datatype_span(&dtype->super, count, &gap); ompi_datatype_type_extent(dtype, &extent); if (MPI_IN_PLACE == sbuf) { sbuf = rbuf; inplace_temp_free = (char*)malloc(dsize); if (NULL == inplace_temp_free) { return OMPI_ERR_OUT_OF_RESOURCE; } rbuf = inplace_temp_free - gap; } if (size > 1) { free_buffer = (char*)malloc(dsize); if (NULL == free_buffer) { if (NULL != inplace_temp_free) { free(inplace_temp_free); } return OMPI_ERR_OUT_OF_RESOURCE; } pml_buffer = free_buffer - gap; } /* Initialize the receive buffer. */ if (rank == (size - 1)) { err = ompi_datatype_copy_content_same_ddt(dtype, count, (char*)rbuf, (char*)sbuf); } else { err = MCA_PML_CALL(recv(rbuf, count, dtype, size - 1, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE)); } if (MPI_SUCCESS != err) { if (NULL != free_buffer) { free(free_buffer); } return err; } /* Loop receiving and calling reduction function (C or Fortran). */ for (i = size - 2; i >= 0; --i) { if (rank == i) { inbuf = (char*)sbuf; } else { err = MCA_PML_CALL(recv(pml_buffer, count, dtype, i, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE)); if (MPI_SUCCESS != err) { if (NULL != free_buffer) { free(free_buffer); } return err; } inbuf = pml_buffer; } /* Perform the reduction */ ompi_op_reduce(op, inbuf, rbuf, count, dtype); } if (NULL != inplace_temp_free) { err = ompi_datatype_copy_content_same_ddt(dtype, count, (char*)sbuf, rbuf); free(inplace_temp_free); } if (NULL != free_buffer) { free(free_buffer); } /* All done */ return MPI_SUCCESS; } /* * ompi_coll_base_reduce_intra_redscat_gather * * Function: Reduce using Rabenseifner's algorithm. * Accepts: Same arguments as MPI_Reduce * Returns: MPI_SUCCESS or error code * * Description: an implementation of Rabenseifner's reduce algorithm [1, 2]. * [1] Rajeev Thakur, Rolf Rabenseifner and William Gropp. * Optimization of Collective Communication Operations in MPICH // * The Int. Journal of High Performance Computing Applications. Vol 19, * Issue 1, pp. 49--66. * [2] http://www.hlrs.de/mpi/myreduce.html. * * This algorithm is a combination of a reduce-scatter implemented with * recursive vector halving and recursive distance doubling, followed either * by a binomial tree gather [1]. * * Step 1. If the number of processes is not a power of two, reduce it to * the nearest lower power of two (p' = 2^{\floor{\log_2 p}}) * by removing r = p - p' extra processes as follows. In the first 2r processes * (ranks 0 to 2r - 1), all the even ranks send the second half of the input * vector to their right neighbor (rank + 1), and all the odd ranks send * the first half of the input vector to their left neighbor (rank - 1). * The even ranks compute the reduction on the first half of the vector and * the odd ranks compute the reduction on the second half. The odd ranks then * send the result to their left neighbors (the even ranks). As a result, * the even ranks among the first 2r processes now contain the reduction with * the input vector on their right neighbors (the odd ranks). These odd ranks * do not participate in the rest of the algorithm, which leaves behind * a power-of-two number of processes. The first r even-ranked processes and * the last p - 2r processes are now renumbered from 0 to p' - 1. * * Step 2. The remaining processes now perform a reduce-scatter by using * recursive vector halving and recursive distance doubling. The even-ranked * processes send the second half of their buffer to rank + 1 and the odd-ranked * processes send the first half of their buffer to rank - 1. All processes * then compute the reduction between the local buffer and the received buffer. * In the next log_2(p') - 1 steps, the buffers are recursively halved, and the * distance is doubled. At the end, each of the p' processes has 1 / p' of the * total reduction result. * * Step 3. A binomial tree gather is performed by using recursive vector * doubling and distance halving. In the non-power-of-two case, if the root * happens to be one of those odd-ranked processes that would normally * be removed in the first step, then the role of this process and process 0 * are interchanged. * * Limitations: * count >= 2^{\floor{\log_2 p}} * commutative operations only * intra-communicators only * * Memory requirements (per process): * rank != root: 2 * count * typesize + 4 * \log_2(p) * sizeof(int) = O(count) * rank == root: count * typesize + 4 * \log_2(p) * sizeof(int) = O(count) * * Recommendations: root = 0, otherwise it is required additional steps * in the root process. */ int ompi_coll_base_reduce_intra_redscat_gather( const void *sbuf, void *rbuf, int count, struct ompi_datatype_t *dtype, struct ompi_op_t *op, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module) { int comm_size = ompi_comm_size(comm); int rank = ompi_comm_rank(comm); OPAL_OUTPUT((ompi_coll_base_framework.framework_output, "coll:base:reduce_intra_redscat_gather: rank %d/%d, root %d", rank, comm_size, root)); /* Find nearest power-of-two less than or equal to comm_size */ int nsteps = opal_hibit(comm_size, comm->c_cube_dim + 1); /* ilog2(comm_size) */ assert(nsteps >= 0); int nprocs_pof2 = 1 << nsteps; /* flp2(comm_size) */ if (nprocs_pof2 < 2 || count < nprocs_pof2 || !ompi_op_is_commute(op)) { OPAL_OUTPUT((ompi_coll_base_framework.framework_output, "coll:base:reduce_intra_redscat_gather: rank %d/%d count %d " "switching to basic linear reduce", rank, comm_size, count)); return ompi_coll_base_reduce_intra_basic_linear(sbuf, rbuf, count, dtype, op, root, comm, module); } int err = MPI_SUCCESS; int *rindex = NULL, *rcount = NULL, *sindex = NULL, *scount = NULL; ptrdiff_t lb, extent, dsize, gap; ompi_datatype_get_extent(dtype, &lb, &extent); dsize = opal_datatype_span(&dtype->super, count, &gap); /* Temporary buffers */ char *tmp_buf_raw = NULL, *rbuf_raw = NULL; tmp_buf_raw = malloc(dsize); if (NULL == tmp_buf_raw) { err = OMPI_ERR_OUT_OF_RESOURCE; goto cleanup_and_return; } char *tmp_buf = tmp_buf_raw - gap; if (rank != root) { rbuf_raw = malloc(dsize); if (NULL == rbuf_raw) { err = OMPI_ERR_OUT_OF_RESOURCE; goto cleanup_and_return; } rbuf = rbuf_raw - gap; } if ((rank != root) || (sbuf != MPI_IN_PLACE)) { err = ompi_datatype_copy_content_same_ddt(dtype, count, rbuf, (char *)sbuf); if (MPI_SUCCESS != err) { goto cleanup_and_return; } } /* * Step 1. Reduce the number of processes to the nearest lower power of two * p' = 2^{\floor{\log_2 p}} by removing r = p - p' processes. * 1. In the first 2r processes (ranks 0 to 2r - 1), all the even ranks send * the second half of the input vector to their right neighbor (rank + 1) * and all the odd ranks send the first half of the input vector to their * left neighbor (rank - 1). * 2. All 2r processes compute the reduction on their half. * 3. The odd ranks then send the result to their left neighbors * (the even ranks). * * The even ranks (0 to 2r - 1) now contain the reduction with the input * vector on their right neighbors (the odd ranks). The first r even * processes and the p - 2r last processes are renumbered from * 0 to 2^{\floor{\log_2 p}} - 1. These odd ranks do not participate in the * rest of the algorithm. */ int vrank, step, wsize; int nprocs_rem = comm_size - nprocs_pof2; if (rank < 2 * nprocs_rem) { int count_lhalf = count / 2; int count_rhalf = count - count_lhalf; if (rank % 2 != 0) { /* * Odd process -- exchange with rank - 1 * Send the left half of the input vector to the left neighbor, * Recv the right half of the input vector from the left neighbor */ err = ompi_coll_base_sendrecv(rbuf, count_lhalf, dtype, rank - 1, MCA_COLL_BASE_TAG_REDUCE, (char *)tmp_buf + (ptrdiff_t)count_lhalf * extent, count_rhalf, dtype, rank - 1, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE, rank); if (MPI_SUCCESS != err) { goto cleanup_and_return; } /* Reduce on the right half of the buffers (result in rbuf) */ ompi_op_reduce(op, (char *)tmp_buf + (ptrdiff_t)count_lhalf * extent, (char *)rbuf + count_lhalf * extent, count_rhalf, dtype); /* Send the right half to the left neighbor */ err = MCA_PML_CALL(send((char *)rbuf + (ptrdiff_t)count_lhalf * extent, count_rhalf, dtype, rank - 1, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_STANDARD, comm)); if (MPI_SUCCESS != err) { goto cleanup_and_return; } /* This process does not pariticipate in recursive doubling phase */ vrank = -1; } else { /* * Even process -- exchange with rank + 1 * Send the right half of the input vector to the right neighbor, * Recv the left half of the input vector from the right neighbor */ err = ompi_coll_base_sendrecv((char *)rbuf + (ptrdiff_t)count_lhalf * extent, count_rhalf, dtype, rank + 1, MCA_COLL_BASE_TAG_REDUCE, tmp_buf, count_lhalf, dtype, rank + 1, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE, rank); if (MPI_SUCCESS != err) { goto cleanup_and_return; } /* Reduce on the right half of the buffers (result in rbuf) */ ompi_op_reduce(op, tmp_buf, rbuf, count_lhalf, dtype); /* Recv the right half from the right neighbor */ err = MCA_PML_CALL(recv((char *)rbuf + (ptrdiff_t)count_lhalf * extent, count_rhalf, dtype, rank + 1, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE)); if (MPI_SUCCESS != err) { goto cleanup_and_return; } vrank = rank / 2; } } else { /* rank >= 2 * nprocs_rem */ vrank = rank - nprocs_rem; } /* * Step 2. Reduce-scatter implemented with recursive vector halving and * recursive distance doubling. We have p' = 2^{\floor{\log_2 p}} * power-of-two number of processes with new ranks (vrank) and result in rbuf. * * The even-ranked processes send the right half of their buffer to rank + 1 * and the odd-ranked processes send the left half of their buffer to * rank - 1. All processes then compute the reduction between the local * buffer and the received buffer. In the next \log_2(p') - 1 steps, the * buffers are recursively halved, and the distance is doubled. At the end, * each of the p' processes has 1 / p' of the total reduction result. */ rindex = malloc(sizeof(*rindex) * nsteps); /* O(\log_2(p)) */ sindex = malloc(sizeof(*sindex) * nsteps); rcount = malloc(sizeof(*rcount) * nsteps); scount = malloc(sizeof(*scount) * nsteps); if (NULL == rindex || NULL == sindex || NULL == rcount || NULL == scount) { err = OMPI_ERR_OUT_OF_RESOURCE; goto cleanup_and_return; } if (vrank != -1) { step = 0; wsize = count; sindex[0] = rindex[0] = 0; for (int mask = 1; mask < nprocs_pof2; mask <<= 1) { /* * On each iteration: rindex[step] = sindex[step] -- begining of the * current window. Length of the current window is storded in wsize. */ int vdest = vrank ^ mask; /* Translate vdest virtual rank to real rank */ int dest = (vdest < nprocs_rem) ? vdest * 2 : vdest + nprocs_rem; if (rank < dest) { /* * Recv into the left half of the current window, send the right * half of the window to the peer (perform reduce on the left * half of the current window) */ rcount[step] = wsize / 2; scount[step] = wsize - rcount[step]; sindex[step] = rindex[step] + rcount[step]; } else { /* * Recv into the right half of the current window, send the left * half of the window to the peer (perform reduce on the right * half of the current window) */ scount[step] = wsize / 2; rcount[step] = wsize - scount[step]; rindex[step] = sindex[step] + scount[step]; } /* Send part of data from the rbuf, recv into the tmp_buf */ err = ompi_coll_base_sendrecv((char *)rbuf + (ptrdiff_t)sindex[step] * extent, scount[step], dtype, dest, MCA_COLL_BASE_TAG_REDUCE, (char *)tmp_buf + (ptrdiff_t)rindex[step] * extent, rcount[step], dtype, dest, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE, rank); if (MPI_SUCCESS != err) { goto cleanup_and_return; } /* Local reduce: rbuf[] = tmp_buf[] rbuf[] */ ompi_op_reduce(op, (char *)tmp_buf + (ptrdiff_t)rindex[step] * extent, (char *)rbuf + (ptrdiff_t)rindex[step] * extent, rcount[step], dtype); /* Move the current window to the received message */ if (step + 1 < nsteps) { rindex[step + 1] = rindex[step]; sindex[step + 1] = rindex[step]; wsize = rcount[step]; step++; } } } /* * Assertion: each process has 1 / p' of the total reduction result: * rcount[nsteps - 1] elements in the rbuf[rindex[nsteps - 1], ...]. */ /* * Setup the root process for gather operation. * Case 1: root < 2r and root is odd -- root process was excluded on step 1 * Recv data from process 0, vroot = 0, vrank = 0 * Case 2: root < 2r and root is even: vroot = root / 2 * Case 3: root >= 2r: vroot = root - r */ int vroot = 0; if (root < 2 * nprocs_rem) { if (root % 2 != 0) { vroot = 0; if (rank == root) { /* * Case 1: root < 2r and root is odd -- root process was * excluded on step 1 (newrank == -1). * Recv a data from the process 0. */ rindex[0] = 0; step = 0, wsize = count; for (int mask = 1; mask < nprocs_pof2; mask *= 2) { rcount[step] = wsize / 2; scount[step] = wsize - rcount[step]; rindex[step] = 0; sindex[step] = rcount[step]; step++; wsize /= 2; } err = MCA_PML_CALL(recv(rbuf, rcount[nsteps - 1], dtype, 0, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE)); if (MPI_SUCCESS != err) { goto cleanup_and_return; } vrank = 0; } else if (vrank == 0) { /* Send a data to the root */ err = MCA_PML_CALL(send(rbuf, rcount[nsteps - 1], dtype, root, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_STANDARD, comm)); if (MPI_SUCCESS != err) { goto cleanup_and_return; } vrank = -1; } } else { /* Case 2: root < 2r and a root is even: vroot = root / 2 */ vroot = root / 2; } } else { /* Case 3: root >= 2r: newroot = root - r */ vroot = root - nprocs_rem; } /* * Step 3. Gather result at the vroot by the binomial tree algorithm. * Each process has 1 / p' of the total reduction result: * rcount[nsteps - 1] elements in the rbuf[rindex[nsteps - 1], ...]. * All exchanges are executed in reverse order relative * to recursive doubling (previous step). */ if (vrank != -1) { int vdest_tree, vroot_tree; step = nsteps - 1; /* step = ilog2(p') - 1 */ for (int mask = nprocs_pof2 >> 1; mask > 0; mask >>= 1) { int vdest = vrank ^ mask; /* Translate vdest virtual rank to real rank */ int dest = (vdest < nprocs_rem) ? vdest * 2 : vdest + nprocs_rem; if ((vdest == 0) && (root < 2 * nprocs_rem) && (root % 2 != 0)) dest = root; vdest_tree = vdest >> step; vdest_tree <<= step; vroot_tree = vroot >> step; vroot_tree <<= step; if (vdest_tree == vroot_tree) { /* Send data from rbuf and exit */ err = MCA_PML_CALL(send((char *)rbuf + (ptrdiff_t)rindex[step] * extent, rcount[step], dtype, dest, MCA_COLL_BASE_TAG_REDUCE, MCA_PML_BASE_SEND_STANDARD, comm)); if (MPI_SUCCESS != err) { goto cleanup_and_return; } break; } else { /* Recv and continue */ err = MCA_PML_CALL(recv((char *)rbuf + (ptrdiff_t)sindex[step] * extent, scount[step], dtype, dest, MCA_COLL_BASE_TAG_REDUCE, comm, MPI_STATUS_IGNORE)); if (MPI_SUCCESS != err) { goto cleanup_and_return; } } step--; } } cleanup_and_return: if (NULL != tmp_buf_raw) free(tmp_buf_raw); if (NULL != rbuf_raw) free(rbuf_raw); if (NULL != rindex) free(rindex); if (NULL != sindex) free(sindex); if (NULL != rcount) free(rcount); if (NULL != scount) free(scount); return err; }