矩阵乘法并行实现（OpenMP）

Question

I ran a matrix multiplication code serially and parallelized.There was no significant improvement with the parallel version.

    dimension =4000;

    //#pragma omp parallel for shared(A,B,C) private(i,j,k)
    {
    for(int i=0; i<dimension; i++){
        for(int j=0; j<dimension; j++){
           for(int k=0; k<dimension; k++){
             C[i][j] += A[i][k] * B[k][j];
            }
        }       
    }
    }

Output: time ./a.out

real    4m58,760s
user    4m58,706s
sys     0m0,036s

for serial code (I put the #pragma... in a comment,rest of code is same) I got following output

real    4m51,240s
user    4m51,210s
sys     0m0,024s

Answer 1

You need to compile code with -fopenmp for the pragma to work. Also, you don't need to comment the pragma to run without OpenMP, just don't compile with OpenMP.

With OpenMP: gcc -fopenmp -o a.out code.c
Without OpenMP: gcc -o a.out code.c

Answer 2

Unless this is a (poorly chosen) educational example, please don't write your own matrix multiply and parallelize it. Getting near optimal performance from matrix multiplication involves other optimizations (vectorization, cache-blocking) which take time to get write and are hard to get correct.

As ever, "The best code is the code I don't have to write", so go grab a copy of a BLAS library (for instance, Intel MKL is now free).

I know it's more fun to write code than read a manual, but sometimes the latter is more productive!

Full disclosre: I work for Intel, but not on MKL.