[mlir][sparse][GPU] add 2:4 integration test

[K-Wu]

test untracking file

[K-Wu]

not working

[K-Wu]

try again

[K-Wu]

recover

[K-Wu]

try again

[K-Wu]

recover

[K-Wu]

initial

[K-Wu]

rebase origin/main

[K-Wu]

upd test

[K-Wu]

update

[K-Wu]

upd

[K-Wu]

working now

[aartbik]

mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-matmul-2-4-lib.mlir

17

This seems a bit copy-and-past from the sparse-mma-2-4-f16.mlir test (which really uses device code for this method by means of e.g. nvgpu.mma.sp.sync). Here, however ,the library calls are still made from the host. So I would remove the while device/host comments here at L17 and at L62).

Also, the gpu.container_module is not needed, since no method is defined as gpu.module

34

commented out code?

207

avoid commented out code

[K-Wu]

address comments

[K-Wu]

mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-matmul-2-4-lib.mlir

17

Thanks for all these comments! They are all addressed now

[aartbik]

mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-matmul-2-4-lib.mlir

5

it looks like this pipeline can be simplified quite a bit, all the gpu.module(....) can go, right?
Also, the vector to llvm and probably more. Perhaps you can actually get rid of the first mlir-opt call and just start at L7 (bit hard to tell just by looking, but run it by hand and see how far you can strip it)

15

remove gpu.container_module

25

add comment to magic constant here

42

does it work without? in any case, let the TODO jump out a bi t more

119

Copy and paste comment, this is no longer the compressed matrix, but the full 2:4 matrix A

145

empty // line after this comment to seperate it from the CHECK

197

there are no warps in this code, so simply Call the kernel
this all still runs on host

[K-Wu]

rebase origin/main; addressing comments

[K-Wu]

fix test error