Marcus LewisNew blog post. I tried something new, got good results, and learned some surprising nuances of GPU programming.
https://probablymarcus.com/blocks/2023/10/19/vectorizing-wide-pytorch-expressions.html
In machine learning and scientific computing, there are a lot of scenarios where code could be much more vectorized, but there's seemingly no way to exploit that opportunity while keeping the code readable.
I created a library called Vexpr to try to fill this gap.
In this blog post, I demonstrate Vexpr in one of these scenarios, running a customized Gaussian Process on a GPU.
I was excited to write this up. It's fun to learn the nuances and build up a cartoon mental model of what vectorization really does.
For more info on Vexpr, check out https://vexpr.org
