MATLAB: GpuDevice command very slow

gpudeviceParallel Computing Toolbox

I am running CUDA kernels using the parallel computing toolbox and r2012a. Recently upgraded to a 600 series (Kepler) gpu. To setup the CUDA kernel we extract the maximum threads per block using: gpu_han=gpuDevice(1); k = parallel.gpu.CUDAKernel('gpu_tfm_linear_arb.ptx', gpu_tfm_linear_arb.cu'); k.ThreadBlockSize = gpu_han.MaxThreadsPerBlock;
This is now executing very slowly (order 2mins). If I specify the threadblocksize manually to the max of the card (1024 in this case), it executes in 0.1 s.
This used to run quickly with a 400 series card. Any help gratefully received

Best Answer

MATLAB R2012a doesn't include code for the Kepler series GPUs. This means that the very first time you call any GPU command after upgrading to a Kepler card, be it gpuDevice or something else, MATLAB will wait for the CUDA driver to just-in-time compile all the PTX code that ships with MATLAB for the Kepler device. This behavior allows MATLAB to work with cards that weren't available when that version of MATLAB was released.
The good news is that this should be a one-time hit. The next time you start MATLAB the JIT'd code should be cached and you should not get the performance hit.
The other thing to point out is that you should consider recompiling your CUDA kernel and producing PTX for the new card, if you haven't already done so, or you may see a similar one-time hit the first time you launch your own kernel for the same reason.
Related Question