Shawn Zhong 8befc91a79 Add Triton language and compiler (#7919) ...

Close #5530. Infra:
https://github.com/compiler-explorer/infra/pull/1711. Previous work by
@siboehm at #5531


## Summary

This pull request introduces support for the
[Triton](https://github.com/triton-lang/triton) language, a Python-based
DSL for writing highly efficient GPU kernels.


- [x] **New Language Support**: I've added comprehensive support for the
Triton programming language, allowing users to compile and inspect
Triton kernels within Compiler Explorer. (c.f.,
`lib/compilers/triton.ts`)
- [x] **Python Wrapper for Compilation**: A new Python wrapper script
(`triton_wrapper.py`) has been introduced to manage Triton compilation,
patching its behavior to dump compiled kernels and intermediate
representations without requiring actual execution, and consolidating
the output for Compiler Explorer.
- [x] **Device Assembly View**: Enables viewing of generated device
assembly code (e.g., PTX, AMDGCN) and various intermediate
representations (MLIR, LLVM IR) produced by the Triton compiler.
- [x] **MLIR Parsing**: New parsers (`asm-parser-mlir.ts` and
`mlir-pass-dump-parser.ts`) have been added to correctly interpret and
display MLIR assembly and optimization pass dumps, including source
location information.
- [x] **Multi-Version & Multi-Backend Support**: Painstakingly includes
all 8 versions (from 2.2.0 to 3.3.1) of Triton that supports Python
3.12. Supports both CUDA and HIP backend for Triton 3.

## Screenshots

Source and assembly:
<img width="1354" height="789" alt="image"
src="https://github.com/user-attachments/assets/c29650ff-2073-40e0-a9e6-ff8377094b5e"
/>



Device view for MLIR and LLVM IR: 
<img width="1402" height="670" alt="image"
src="https://github.com/user-attachments/assets/43dd5c68-ca78-41b1-9865-e97ffe3ef73c"
/>

Opt pipeline viewer:
<img width="1408" height="668" alt="image"
src="https://github.com/user-attachments/assets/429eef8c-aaac-4781-aafa-39ef0ffc7241"
/>

Diff of TTIR in Triton 3.3.1 vs 2.3.0:
<img width="1580" height="726" alt="image"
src="https://github.com/user-attachments/assets/a928c893-dd9a-4c3a-a048-14046e56a14c"
/>

CUDA & HIP:
<img width="1596" height="800" alt="image"
src="https://github.com/user-attachments/assets/c18800c3-cfad-4e5e-96de-ba92c9f236ea"
/>

## Implementation Details (and Notes for Reviewers)

- For Device Assembly View, I Implemented `MlirAsmParser` for parsing
MLIR assembly. Technically MLIR is not an assembly language, but there
is no better choice to make the source line map work w/ device view.
- I Implemented `MlirPassDumpParser` for processing MLIR optimization
pass dumps. I tried to subclass `LlvmPassDumpParser`, but they turn out
to be too different to worth doing it.
- `LlvmPassDumpParser` made some assumptions that do not hold true for
MLIR passed. Some effort is put to make sure that the passes are
properly diff-ed, since some passes can run multiple times and also
sometimes pass can be nested (i.e., some number of `before`s followed by
some number of `after`s)
- A lot of effort is put into `patch_triton` to make sure that the we
only compile the kernel without actually running it, and that needs to
work across all the versions we support.

## Steps to Run Locally

1. Clone https://github.com/ShawnZhong/compiler-explorer-infra.git
2. Install Triton to `/opt/compiler-explorer/triton`:
```sh
$ cd compiler-explorer-infra
$ ./bin/ce_install install triton
$ ls /opt/compiler-explorer/triton
# v2.2.0  v2.3.0  v2.3.1  v3.0.0  v3.1.0  v3.2.0  v3.3.0  v3.3.1
```
3. Clone https://github.com/ShawnZhong/compiler-explorer.git and
checkout branch `triton`
4. Run Compiler Explorer
```sh
make EXTRA_ARGS='--language triton' dev
```
5. Enjoy

---------

Co-authored-by: Matt Godbolt <matt@godbolt.org>

2025-07-30 10:15:28 -05:00

7.3 KiB

640x728px

Raw History