[Mlir-commits] [mlir] [mlir][gpu][NVPTX] Enable NVIDIA GPU JIT compilation path (PR #66220)

[Fabian Mora]

fabianmcg wrote:

> Since I don't see changes in the sparse code, I assume you want some feedback, but I need a bit more context on what you had in mind. In general, we have a lot of "knobs" in the sparse pipeline setup, so generally I am not opposed to adding one more ;-)

With this patch we have 3 ways to compile code, `JIT->format=assembly`, `Cubin->format=bin`, `Fatbin->format=fatbin`, JIT will obviously add a performance hit at runtime, so the questions are:
1. Is it okay to add another option to the sparse compiler to specify which format to use? 
2. Is there a preference to which option to use by default?

JIT will make the test work, but fatbin is preferable for runtime performance as it can be used for AOT.


https://github.com/llvm/llvm-project/pull/66220