#kernels

Every CUDA kernel launch has overhead. Fusing three operations into one can be 3x faster. Here's where fusion helps and how to get it.