how to run Gpu and Cpu in parallel in for loop?

Question

Naveen kumar Elumalai 2019-1-14

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/439740-how-to-run-gpu-and-cpu-in-parallel-in-for-loop

评论： Naveen kumar Elumalai 2019-1-14

for i=1:10
    A(:,:,i)=rand(5,5);
    [q1,r1]=qr(A(:,:,i));
end

I wanted cpu and gpu's to communicate with itself about the loop index. For iteration 1, cpu is working, gpu1 should also simultaneously work on iteration 2. gpu 2 should also simultaneously work on iteration 2. If gpu1 completes its task early, it should start iteration 3. When cpu is done with iteration 1, it should proceed with iteration 4 and so on. How should we implement this. The cpu, gpu1,gpu2 should only communiate about the loop index 'i' and not the A matrix. Once cpu or either of gpu's completes its iteration, it should parallely update the output array. How should this be implemented ?

6 个评论
显示 4更早的评论隐藏 4更早的评论

OCDER 2019-1-14

在 MATLAB Online 中打开

Just curious, why do you want to split the job as such? Do you want GPU and CPU to be treated as indistinguishable processors? I foresee an issue where you must continuously exchange an array from a normal array to gpuArray, which can create a lot of overhead and slow things down.

Example of job tasks you want?

Task#   CPU1   CPU2   GPU1   GPU2
      X
             X
                    X
                           X 
             X
      X
             X
      X
                    X
            X      
                          X

Naveen kumar Elumalai 2019-1-14

Yeah, the above example you referred, is what I want. The conversion of normalArray to gpuArray does bring a lot of overhead. Thats why, I wanted a work around for executing above task.

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Walter Roberson 2019-1-14

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/439740-how-to-run-gpu-and-cpu-in-parallel-in-for-loop#answer_356396

You cannot do that with parfor or spmd.

What you can perhaps do is use parfeval to submit tasks associated with various resources. Monitor task completion, and upon completion upon a resource, collect the results and start a new task on the same resource.