yea, i wrote this blogpost rather to show how to use scan in different ways than the canonical example of calculating prefix sum of a vector shown in introductions on gpu programming.
It doesn’t. The batch size is just 8. This is a very good trick and often needed to archive peak performance in memory bound kernels. You can checkout the equivalent code in cuda aswell :)
The kernels I mention in CUDA use all the equivalent logic like the Mojo kernels. You can find them on my GitHub: https://github.com/simveit/effective_transpose
You may want to provide a faster kernel on H100 via PR and I will merge after checking it’s faster.