Choose an optimal thread count for CUDA kernels

[sjsprecious]

Currently the CUDA thread count per thread block is set to 32. A different value should be set to achieve an optimal performance.

Acceptance Criteria

• Find out the optimal CUDA thread count for MICM GPU kernels.

[sjsprecious]

Based on the Nsight Systems profiling results, it seems varying the thread count per thread block affects the overall GPU performance marginally.

[sjsprecious]

I was wrong before that the CUDA thread count per thread block does matter for the performance. Previously we had an unoptimized GPU version so the difference was not easily visible.