Running a Large Number of Small-Scale Optimization on GPU

I'm running a simple optimization process(using Matlab quadprog command) many times (~40K and more). When run on CPU, a single optimization takes about 20s, which means the whole thing takes days. I am thinking about using GPU to perform the computation, however, it seems that quadprog does not support the gpuArray input. Right now, after some research, I believe my only option is to write the CUDA kernel myself. Am I missing something here? Is there another way to run many optimizations processes on GPU?

[Yashvardhan Agarwal]

Hi Jian,

I am facing this same problem and was curious if you did find any solution for it. If not, then did you try to make a CUDA implementation for it and what were the results?

Your help is greatly appreciated.

Thank you.

Answer 1

[Ryan Povall]

Hi Jian,

I do not think you are missing anything. You are correct in that "quadprog" is not currently supported for gpuArray as seen in the following documentation page:

http://www.mathworks.com/help/distcomp/run-built-in-functions-on-a-gpu.html

I can also see that the Optimization Toolbox is not currently supported for GPU computing as it's not listed in this following page:

http://www.mathworks.com/discovery/matlab-gpu.html

There looks to be different options for running other optimization processes in parallel as seen in the below documentation page. However, "quadprog" is not listed and does not seem to be supported:

http://www.mathworks.com/help/optim/ug/parallel-computing-in-optimization-toolbox-functions.html

I hope this helps.

Ryan