我是 MPI 新手,只学习了基础知识。在基于以下算法执行矩阵向量乘法时,出现一条我无法理解的错误消息,因此无法修复。显示的错误消息截图如下:
/* Matrix-vector product Ab = c with parallel inner products*/
/* Row-oriented blockwise distribution of A */
/* Replicated distribution of vectors b and c */
local n = n/p;
for (i=0; i<local n; i++) local c[i] = 0;
for (i=0; i<local n; i++)
for (j=0; j<m; j++)
local c[i] = local c[i] + local A[i][j] * b[j];
multi broadcast(local c,local n,c);
/* Multi-broadcast operation of (c[0]; :::; c[localn]) to globalc*/
这是我的代码:
#include<stdio.h>
#include<stdlib.h>
#include "mpi.h"
#define n 4
#define m 4
void matrix_vector_product(double **matrix, double *vector, double *result, int rows, int cols)
{
for(int i=0; i<rows; i++)
result[i] = 0;
for(int i=0; i<rows; i++)
for(int j=0; j<cols; j++)
result[i] += matrix[i][j] * vector[j];
}
int main(int argc, char *argv[])
{
int rank, size;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
double **matrix;
matrix = (double **)calloc(n,sizeof(double *));
for (int i=0; i<n; i++)
{
matrix[i] = (double *)calloc(m, sizeof(double));
}
double *b;
b = (double *)calloc(m, sizeof(double));
double *c = NULL;
double *local_c = (double *)calloc(n, sizeof(double));
int local_n = n/size;
if (rank == 0)
{
for(int i=0; i<n; i++)
for(int j =0; j<m; j++)
matrix[i][j] = i+j;
for(int j =0; j<m; j++)
b[j] = j+1;
c = (double *)calloc(m, sizeof(double));
}
MPI_Scatter(&matrix, local_n*m, MPI_DOUBLE, &matrix, local_n*m, MPI_DOUBLE, 0, MPI_COMM_WORLD);
MPI_Bcast(b, m, MPI_DOUBLE, 0, MPI_COMM_WORLD);
matrix_vector_product(matrix, b, local_c, local_n, m);
MPI_Gather(local_c, local_n, MPI_DOUBLE, c, local_n, MPI_DOUBLE, 0, MPI_COMM_WORLD);
if(rank==0){
printf(" Result = ");
for(int i=0; i<n; i++)
printf(" .2%f", c[i]);
}
MPI_Finalize();
free(matrix);
free(b);
free(c);
free(local_c);
return 0;
}
如果您也能解释为什么会发生这种情况,那就太好了,因为我想澄清我的概念。
我尝试将大小为n x m的*矩阵*与大小为m的向量相乘。由于假设采用分布式内存分布,因此我使用 MPI_Scatter 操作来分配矩阵行的内存块,以实现 A 的按行按块分布。使用 MPI_Bcast 操作复制 b。结果是使用函数matrix_vector_product()计算的,本地缓冲区(这里是数组)c的内容聚集在c中。我期望输出是
Result = 8 10 12 14
simpletype*
,在您的情况下是 double*
。所以你不能通过matrix
,这是一个double**
。double *matrix
并使用从 2D 索引到 1D 的转换. Then pass
MPI_Scatter(matrix,....`