Unavoidable memory leaks in MEX

2 次查看（过去 30 天）

Ayal 2013-8-26

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/85601-unavoidable-memory-leaks-in-mex

I'm trying to write a MEX running some CUDA code (can't use feval). Unfortunately, no matter what I do, I can't avoid the memory leaks.

I give, as a parameter to the function, a gpuArray of 0-s. I tried making the prhs non-constant, so I could overwrite it:

double * outPtr=(double* ) mxGetData(prhs[2]);

The data is a gpuArray, so It should be stored on the GPU, but when I pass outPtr as a parameter to a CUDA kernel, I get an error (CUDA-related).

so, overwriting didn't work, I tried:

    mxGPUArray * tmp=mxGPUCopyFromMxArray(prhs[2]);
    double * outPtr=(double* ) mxGPUGetData(tmp);

and then

plhs[0]=mxGPUCreateMxArrayOnGPU(tmp);

which causes memory leaks on the GPU. I suspect the memory on the GPU isn't freed later when I clear/overwrite the variable in plhs[0].

I tried many things, but none of them solved this problem. I even tried making a persistent variable "out", and destroying it at the beginning of the MEX code (I figured maybe if I explicitly told MATLAB to clear it, instead of expecting MATLAB to clear it as soon as I overwrite it). But it seems mxArrayDestroy doesn't clear memory on GPU.

So, Is there ANY way for me to free the space stored in plhs[0] when I'm done with it?

I am using MATLAB 2013a

3 个评论
显示 1更早的评论隐藏 1更早的评论

Ayal 2013-8-26

Jill - I asked a little bit more here (for a possible alternative), but you are correct - the questions are very similar. I didn't notice your answer (and come to think of it, it's likely I accidentally thought the e-mail update I got was about my own answer). I will try to have someone apply the patch (i'm using a computer in the university...) This doesn't look like the problem I've been having (the error for being out of memory seems different - like the memory problem is not with the GPU's memory, but is brought on by the use of the GPU)

But i'll give it a try. Thanks!

Ayal 2013-8-27