coll_tuned_reduce.c
来自「MPI stands for the Message Passing Inter」· C语言 代码 · 共 549 行 · 第 1/2 页
C
549 行
/* * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana * University Research and Technology * Corporation. All rights reserved. * Copyright (c) 2004-2006 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. * $COPYRIGHT$ * * Additional copyrights may follow * * $HEADER$ */#include "ompi_config.h"#include "mpi.h"#include "ompi/constants.h"#include "ompi/datatype/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 "coll_tuned.h"#include "coll_tuned_topo.h"/** * 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 * th number of datatype to the original count (original_count) */int ompi_coll_tuned_reduce_generic( void* sendbuf, void* recvbuf, int original_count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, ompi_coll_tree_t* tree, int count_by_segment ){ char *inbuf[2] = {(char*)NULL, (char*)NULL}; char *local_op_buffer = NULL, *accumbuf = NULL, *sendtmpbuf = NULL; ptrdiff_t extent, lower_bound; size_t typelng, realsegsize; ompi_request_t* reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL}; int num_segments, line, ret, segindex, i, rank; int recvcount, prevcount, inbi, previnbi; /** * Determine number of segments and number of elements * sent per operation */ ompi_ddt_get_extent( datatype, &lower_bound, &extent ); ompi_ddt_type_size( datatype, &typelng ); num_segments = (original_count + count_by_segment - 1) / count_by_segment; realsegsize = count_by_segment * extent; sendtmpbuf = (char*) sendbuf; if( sendbuf == MPI_IN_PLACE ) { sendtmpbuf = (char *)recvbuf; } rank = ompi_comm_rank(comm); /* non-leaf nodes - wait for children to send me data & forward up (if needed) */ if( tree->tree_nextsize > 0 ) { /* 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) ) { accumbuf = (char*)malloc(realsegsize * num_segments); /* TO BE OPTIMIZED */ if (accumbuf == NULL) { line = __LINE__; ret = -1; goto error_hndl; } } /* Allocate two buffers for incoming segments */ inbuf[0] = (char*) malloc(realsegsize); if( inbuf[0] == NULL ) { line = __LINE__; ret = -1; goto error_hndl; } /* if there is chance to overlap communication - allocate second buffer */ if( (num_segments > 1) || (tree->tree_nextsize > 1) ) { inbuf[1] = (char*) malloc(realsegsize); if( inbuf[1] == NULL ) { line = __LINE__; ret = -1; goto error_hndl;} } else { inbuf[1] = NULL; } /* 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 - count_by_segment * 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) 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 we are root and are USING MPI_IN_PLACE this is wrong ek! * check for root might not be needed as it should be checked higher up */ if( !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) { local_recvbuf = accumbuf + segindex * realsegsize; } } 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 */ previnbi = (inbi+1) % 2; /* wait on data from last child for previous segment */ ret = ompi_request_wait_all( 1, &reqs[previnbi], MPI_STATUSES_IGNORE ); if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; } local_op_buffer = inbuf[previnbi]; if( i > 0 ) { /* our first operation is to combine our own [sendbuf] data with the data * we recvd from down stream (but only if we are not root and not using * MPI_IN_PLACE) */ if( 1 == i ) { if( !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) { local_op_buffer = sendtmpbuf + segindex * realsegsize; } } /* apply operation */ ompi_op_reduce(op, local_op_buffer, accumbuf+segindex*realsegsize, recvcount, datatype ); } else if ( segindex > 0 ) { void* accumulator = accumbuf + (segindex-1) * realsegsize; if( tree->tree_nextsize <= 1 ) { if( !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) { local_op_buffer = sendtmpbuf+(segindex-1)*realsegsize; } } 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 your 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 to allow for pipelining */ if (segindex == num_segments) break; } /* update input buffer index */ inbi = previnbi; } /* end of for each child */ } /* end of for each segment */ /* clean up */ if( inbuf[0] != NULL) free(inbuf[0]); if( inbuf[1] != NULL) free(inbuf[1]); if( (NULL == recvbuf) || (root != rank) ) free(accumbuf); } /* leaf nodes */ else { /* Send segmented data to parents */ segindex = 0; while( original_count > 0 ) { if( original_count < count_by_segment ) count_by_segment = original_count; ret = MCA_PML_CALL( send((char*)sendbuf + segindex * realsegsize, 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; } } return OMPI_SUCCESS; error_hndl: /* error handler */ OPAL_OUTPUT (( ompi_coll_tuned_stream, "ERROR_HNDL: node %d file %s line %d error %d\n", rank, __FILE__, line, ret )); if( inbuf[0] != NULL ) free(inbuf[0]); if( inbuf[1] != NULL ) free(inbuf[1]); if( (NULL == recvbuf) && (NULL != accumbuf) ) free(accumbuf); 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_tuned_reduce_intra_chain( void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, uint32_t segsize, int fanout){ int segcount = count; size_t typelng; OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_intra_chain rank %d fo %d ss %5d", ompi_comm_rank(comm), fanout, segsize)); COLL_TUNED_UPDATE_CHAIN( comm, root, fanout ); /** * Determine number of segments and number of elements * sent per operation */ ompi_ddt_type_size( datatype, &typelng ); COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount ); return ompi_coll_tuned_reduce_generic( sendbuf, recvbuf, count, datatype, op, root, comm, comm->c_coll_selected_data->cached_chain, segcount );}int ompi_coll_tuned_reduce_intra_pipeline( void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, uint32_t segsize ){ int segcount = count; size_t typelng; OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_intra_pipeline rank %d ss %5d", ompi_comm_rank(comm), segsize)); COLL_TUNED_UPDATE_PIPELINE( comm, root ); /** * Determine number of segments and number of elements * sent per operation */ ompi_ddt_type_size( datatype, &typelng ); COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount ); return ompi_coll_tuned_reduce_generic( sendbuf, recvbuf, count, datatype, op, root, comm, comm->c_coll_selected_data->cached_pipeline, segcount );}int ompi_coll_tuned_reduce_intra_binary( void *sendbuf, void *recvbuf, int count, ompi_datatype_t* datatype, ompi_op_t* op, int root, ompi_communicator_t* comm, uint32_t segsize ){ int segcount = count; size_t typelng; OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_intra_binary rank %d ss %5d", ompi_comm_rank(comm), segsize)); COLL_TUNED_UPDATE_BINTREE( comm, root ); /** * Determine number of segments and number of elements * sent per operation */
⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?