How can I set shared memory configuration for NVIDIA graphics card to increase L1 cache?

I need to set the shared memory configuration for NVIDIA graphics card by running the following command in C:

cudaDeviceSetCacheConfig(cudaFuncCachePreferL1);

This will set L1 cache to 48KB instead of default 16KB (to benefit from higher L1 cache hit rate in my application).

How do I do this if I don't want to use mex file but run CUDA kernel on GPU using precompiled PTX code?

Answer 1

[Joss Knight]

You can't, but you can write a mex function to do it, that will affect your next CUDAKernel call.

MEX functions do not initialise the GPU on every call so I don't quite understand what is preventing you from using mex functions in this case. The mechanism through which kernels are launched are different so I couldn't say definitively whether using CUDAKernel or launching a kernel yourself using a MEX function would be faster.