MATLAB: Arrayfun GPU in “Game of Life” works slower than CPU

gpu

Hello! I've run the demo "paralleldemo_gpu_stencil" and have such a result:
CPU: 2.815ms per generation.
Simple GPU: 2.650ms per generation (1.1x faster).
Arrayfun GPU: 13.253ms per generation (0.2x faster).
I've used ThinkPad P50 with Quadro M1000M and Matlab R2019b with appropriate drivers. Why Arrayfun works such slow ??
Uladzislau.

Best Answer

Check out this Answer. The arrayfun version is rather dependent on good memory performance since the kernel is accessing global GPU memory in a non-coalesced way (multiple threads accessing overlapping regions of memory that aren't contiguous). For your chip, the version that runs multiple kernels on multiple shifted copies of the grid is actually more efficient, despite the kernel launch overhead and the extra memory allocation needed.
Related Question