Add a CPU lowering pass

[shintaro-iwasaki]

This demonstrates how triton-shared can potentially run its lowered linalg + its friends on CPU as a reference.

1. We don't need any changes in the Triton runtime
2. A very basic empty kernel runs, and it succeeds
3. Unfortunately I could not make the standard MLIR lowering passes work for triton-shared generated code (it causes SEGV).

Recently Triton updated its test suite to specify the device type iirc, so we can naturally run test_core.py using a reference CPU backend if the problem above will be solved. Would it be something interesting?

[shintaro-iwasaki]

examples.yml runs Triton code using GitHub Actions.

[manbearian]

when i ran this locally there's a warning from Python using is operator in this situation. I'm definitely not Python programmer but i think the correct code you want might be something more like this:

def _get_triton_shared_opt_path():
    path = os.getenv("TRITON_SHARED_OPT_PATH", "")
    if path == "":
        raise Exception("TRITON_SHARED_OPT_PATH is not set.")
    return path


def _get_llvm_bin_path(bin_name: str):
    path = os.getenv("LLVM_BINARY_DIR", "")
    if path == "":
        raise Exception("LLVM_BINARY_DIR is not set.")
    return f"{path}/{bin_name}"

[shintaro-iwasaki]

Thank you! I don't remember why I used is, but I fixed the issue in the latest commit.

[danikhan632]

keep gettting this error on latest version on triton which odd since Intel's XPU backend does the same thing

/triton/python/triton/third_party/cpu/__init__.py", line 13, in <module>
    from triton.runtime.jit import version_key  # noqa:E402
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ImportError: cannot import name 'version_key' from 'triton.runtime.jit' 

[manbearian]

@danikhan632 i'm pretty sure there's some kind of versioning mismatch going on as the code around this changed recently.

[shintaro-iwasaki]

Thanks for taking a look at this PR!

• From a reproducibility perspective, it worked at least on GitHub Actions below, so if you follow exactly the same way as what it does, it should work.
• If you are using the very latest Triton (not the version specified by ".github/workflows/integration-tests.yml"), maybe the upstream has reorganized files and function names etc. I will investigate it when the version is updated.

[danikhan632]

was working on a similar thing, do you plan on making the compiler with gcc/clang?
worked on the intel extenison so most of my experience building a launcher comes from that

[shintaro-iwasaki]

I might not fully understand your question, but this PR uses the following binaries.

• Triton's default LLVM binaries (llc and mlir-opt)
• The default C++ compiler for the pybind logic and linking etc; now it uses g++ (hardcoded, but can be changed).

[danikhan632]

Which triton commit hash are you using?s till running into the same issue

[shintaro-iwasaki]

See .github/workflows/integration-tests.yml. It uses 05dc28be0e72dd496300a31b99a21a5a5118f8e9.
Please also check the installation method written in the GitHub Action script.

[danikhan632]

missed that, thank you