Check the generated CUTLASS code

[YukeWang96]

Is there any way to check/print the generated CUTLASS code?

[masahi]

If you run test_cutlass.py, you will see generated cutlass code under the tmp directory (check lib0.cu file)

[YukeWang96]

It works in test_cutlass.py.
However, when I want to figure out which statement that generate the *.cu, I use the following code

lib = relay.build(mod, target, params=params)
lib = build_cutlass_kernels(lib, sm=75,  tmp_dir="./tmp", lib_path="compile.so")

It will only keep the *.o files while no *.cu under the tmp/ , am I missing something?

[masahi]

I think this should be export_library() call at

https://github.com/apache/tvm/blob/6d35f0bbdf656d393b0722cb93cd213217781c9d/python/tvm/contrib/cutlass/build.py#L353

This is where we dump the .cu file and invoke nvcc.

[YukeWang96]

I also found that when I comment out tune_cutlass_kernels such as https://github.com/apache/tvm/blob/4babd36481b7108bf50df5c3b256c95c0d9c3291/tests/python/contrib/test_cutlass.py#L181-L183 there will be no .cu file generated.

[YukeWang96]

Finally, I figure out the reason: The annotated needs to identify the operators that will be lowering to CUTLASS backend.
Thanks!

[masahi]

You are welcome to join tlcpack slack (aka TVM slack) if you want more high-bandwidth communication with me or others. The invitation link is below.

https://join.slack.com/t/tvmai/shared_invite/zt-foemgff5-9pC0jateEAdpoZImfxxMew

[YukeWang96]

Sure, will do. Thanks for your nice suggestion!