Is this analogy valid: Writing PTX is like writing assembly instead of a higher-level language (C, C++, rust etc) for CPU code? E.g. normally the higher level code compiles to it, but you can do optimizations by going lower?
For context, like the opening paragraph in the article goes into, I generate PTX code regularly, but have no idea what the actual code in the PTX file means!
I'm curious about the forward compatibility the article goes into. I only experience that to a point: Code compiled on Cuda 12 does not seem to work on machines with Cuda 13.