segtest.c

mpi并行计算的c++代码 可用vc或gcc编译通过 可以用来搭建并行计算试验环境

/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
 *  (C) 2001 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */
#include "mpi.h"
#include "mpiimpl.h"
#include <stdio.h>

/* 
 * Simple segment test, including timing code
 */

/* 
 * Build datatype structures
 *
 * Contiguous
 *     n = 1, 4, 16, 64, 128, 512, 2048, 8196 ints
 * Vector
 *     blocksize = 1, 4, 64 ints
 *     stride    = 1, 64, 127
 * Block Indexed
 *     blocksize = 1, 4 ints
 *     offsets   = i*24 for i = 0 to n, n = 0, 64, 512
 * Indexed
 *     blocksizes = 1, 2, 4, 3, 7, 5, 6
 *     offsets    = i*24 for i = 0 to n, n = 0, 4, 7, 64, 512 
 *     (Wrap blocksizes to match offsets)
 *
 * Also need a few nested datatypes, such as vector of vectors
 * Do the versions in Using MPI
 * 
 */

/*
 * Routines to create dataloops for basic dataloops
 */
/*
 *  Contig
 */
MPID_Dataloop *MPID_Dataloop_init_contig( int count )
{
    MPID_Dataloop *ct;
    
    ct = (MPID_Dataloop *)MPIU_Malloc( sizeof(MPID_Dataloop ) );
    ct->kind                     = MPID_DTYPE_CONTIG | DATALOOP_FINAL_MASK;
    ct->loop_params.c_t.count    = count;
    ct->loop_params.c_t.dataloop = 0;
    ct->extent                   = count;
    ct->handle                       = 0;

    return ct;
}

/*
 * Vector
 */
MPID_Dataloop *MPID_Dataloop_init_vector( int count, int blocksize, 
					  int stride )
{
    MPID_Dataloop *v;

    v = (MPID_Dataloop *)MPIU_Malloc( sizeof(MPID_Dataloop) );
    v->kind                      = MPID_DTYPE_VECTOR | DATALOOP_FINAL_MASK;
    v->loop_params.v_t.count     = count;
    v->loop_params.v_t.blocksize = blocksize;
    v->loop_params.v_t.stride    = stride;
    v->loop_params.v_t.dataloop  = 0;
    v->extent                    = (count-1)*stride + blocksize;
    v->handle                        = 0;

    return v;
}

/* 
 * Block indexed
 */
MPID_Dataloop *MPID_Dataloop_init_blockindexed( int count, int blocksize, 
						MPI_Aint *offset )
{
    MPID_Dataloop *bi;
    MPI_Aint      extent;
    int           i;

    bi = (MPID_Dataloop *)MPIU_Malloc( sizeof(MPID_Dataloop) );
    bi->kind                       = MPID_DTYPE_BLOCKINDEXED | DATALOOP_FINAL_MASK;
    bi->loop_params.bi_t.count     = count;
    bi->loop_params.bi_t.blocksize = blocksize;
    bi->loop_params.bi_t.offset    = 
	(MPI_Aint *)MPIU_Malloc( sizeof(MPI_Aint) * count );
    for (i=0; i<count; i++) {
	bi->loop_params.bi_t.offset[i] = offset[i];
	if (offset[i] + blocksize > extent) 
	    extent = offset[i] + blocksize;
    }
    bi->loop_params.bi_t.dataloop  = 0;
    bi->extent                     = extent;
    bi->handle                         = 0;

    return bi;
}

/*
 * Indexed 
 */
MPID_Dataloop *MPID_Dataloop_init_indexed( int count, int *blocksize, 
					   MPI_Aint *offset )
{
    MPID_Dataloop *it;
    MPI_Aint      extent = 0;
    int           i;

    it = (MPID_Dataloop *)MPIU_Malloc( sizeof(MPID_Dataloop) );
    it->kind                      = MPID_DTYPE_INDEXED | DATALOOP_FINAL_MASK;
    it->loop_params.i_t.count     = count;
    it->loop_params.i_t.blocksize = (int *)MPIU_Malloc( sizeof(int) * count );
    it->loop_params.i_t.offset    = 
	(MPI_Aint *)MPIU_Malloc( sizeof(MPI_Aint) * count );
    for (i=0; i<count; i++) {
	it->loop_params.i_t.offset[i]    = offset[i];
	it->loop_params.i_t.blocksize[i] = blocksize[i];
	if (offset[i] + blocksize[i] > extent) 
	    extent = offset[i] + blocksize[i];
    }
    it->loop_params.i_t.dataloop  = 0;
    it->extent                    = extent;
    it->handle                        = 0;

    return it;
}

int main( int argc, char **argv )
{
    /* MPID_Dataloop *vecloop; */
    MPI_Datatype vectype;
    int count=200, blocksize=4, stride = 7*4;
    char *src_buf, *dest_buf;
    int  i,j,k;
    double r1, r2;

    MPI_Init( &argc, &argv );
    
/*    vecloop = MPID_Dataloop_init_vector( count, blocksize, stride ); */

    MPI_Type_vector( count, 1, 7, MPI_INT, &vectype );

    /* Initialize the data */
    src_buf = (char *)MPIU_Malloc( (count - 1) * stride + blocksize );
    for (i=0; i<(count-1)*stride+blocksize; i++) 
	src_buf[i] = -i;
    for (i=0; i<count; i++) {
	for (j=0; j<blocksize; j++) 
	    src_buf[i*stride+j] = i*blocksize + j;
    }
    dest_buf = (char *)MPIU_Malloc( count*blocksize );
    for (i=0; i<count*blocksize; i++) {
	dest_buf[i] = -i;
    }
    r1 = MPI_Wtime();
    for (i=0; i<100; i++) {
	int position = 0;
	/*MPID_Segment_pack( vecloop, src_buf, dest_buf );*/
	MPI_Pack( src_buf, count, vectype, dest_buf, count*blocksize, 
		  &position, MPI_COMM_WORLD );
    }
    r2 = MPI_Wtime();
    printf( "Timer for vector pack is %e\n", (r2-r1)/100 );
    for (i=0; i<count*blocksize; i++) {
	if (dest_buf[i] != (char)i) { 
	    printf( "Error at location %d\n", i );
	}
    }
    r1 = MPI_Wtime();
    for (k=0; k<100; k++) {
	char *dest=dest_buf, *src=src_buf;
	for (i=0; i<count; i++) {
	    for (j=0; j<blocksize; j++) 
		*dest++ = src[j];
	    src+= stride;
	}
    }
    r2 = MPI_Wtime();
    printf( "Timer for hand vector pack is %e\n", (r2-r1)/100 );

    r1 = MPI_Wtime();
    for (k=0; k<100; k++) {
	int *dest=(int*)dest_buf, *src=(int*)src_buf;
	int bsize = blocksize >> 2;
	int istride = stride >> 2;
	if (bsize == 1) { 
	    for (i=0; i<count; i++) {
		*dest++ = *src;
		src+= istride;
	    }
	}
	else {
	    for (i=0; i<count; i++) {
		for (j=0; j<bsize; j++) 
		    *dest++ = src[j];
		src+= istride;
	    }
	}
    }
    r2 = MPI_Wtime();
    printf( "Timer for hand vector pack (int) is %e\n", (r2-r1)/100 );
    
    MPI_Finalize();
    return 0;
}

/*
 * Nested vector.
 *   The y-z subface is
 *   Type_vector( ey-sy+1, 1, nx, MPI_DOUBLE, &newx1 );
 *   Type_hvector( ez-sz+1, 1, nx*ny_sizeof(double), newx1, &newx );
 * This gives the a(i,sy:ey,sz:ez) of a(nx,ny,nz) (in Fortran notation)
 */