We can mix Matlab GPUArray code with CUDA through Kernel object. Is there a benefit/cost to have c-mex layer in between.
I think name mangling and converting from Column-major order matrices in Matlab to Row major order matrices in c-mex will be costly.
What is the best combination Matlab GPU, CUDA, C-MEX?
Start with the CUDAKernel Object launcher available in the distcomp toolbox. No need to go further than that unless your code has poor efficiency from data transfer overhead eg frequent launches of a short-running kernel, each time copying the same input data.
In that case, you'd probably want to introduce a Mex shim to manage the data movement explicity, but that'll come with substantially more code complexity.
Use the CUDA profilers to guide your decision.
Matlab CUDA is easy to implement using ptx format. visual C++ CUDA combination always yields high throughput compare than matlab cuda implementation.
- Badges
- Science topic