Performance improvement of CUDA programs

In this section, we will see some basic guidelines that we can follow to improve the performance of CUDA programs. These are explained one by one.

Using an optimum number of blocks and threads

We have seen two parameters that need to be specified during a kernel call: the number of blocks and the number of threads per block. GPU resources should not be idle during a kernel call; only then it will give the optimum performance. If resources remain idle, then it may degrade the performance of the program. The number of blocks and threads per block help in keeping GPU resources busy. It has been researched that if the number of blocks are double the number of multiprocessors on the GPU...