TheRock/rocm-libraries/shared/stinkytofu

StinkyTofu

StinkyTofu is an LLVM-inspired pass-based IR optimizer for AMD GPU assembly kernels. It is used by hipBLASLt/TensileLite to schedule and optimize generated GPU code for gfx1250.

Features

• Two-level IR: High-level Logical IR (architecture-agnostic) and low-level Asm IR (architecture-specific)
• Pass pipeline: DAG scheduling, wait count insertion, dead code elimination, redundant mov elimination, peephole optimization
• TableGen-based instruction definitions: New architectures require only .def files, no C++ changes
• FileCheck-style testing: LLVM-style lit tests with stinkytofu-check
• Python bindings: Full IR construction and optimization accessible from Python via nanobind

Prerequisites

• CMake >= 3.16
• Ninja
• ROCm SDK (provides amdclang++)

Windows additional setup

Enable long paths (run as Administrator):

reg add HKLM\SYSTEM\CurrentControlSet\Control\FileSystem /v LongPathsEnabled /t REG_DWORD /d 1 /f

Install Visual Studio Build Tools:

winget install --id Microsoft.VisualStudio.2022.BuildTools --source winget --override "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64 --add Microsoft.VisualStudio.Component.VC.CMake.Project --add Microsoft.VisualStudio.Component.VC.ATL --add Microsoft.VisualStudio.Component.Windows11SDK.22621"

Build

Python environment

Python 3.12 with a virtual environment is recommended for both platforms:

python3.12 -m venv .venv
source .venv/bin/activate   # Linux
.venv\Scripts\activate      # Windows

Note: Avoid installing Python from the Microsoft Store on Windows — its path contains spaces, which breaks Ninja. The build falls back to NMake (single-threaded). Install Python from python.org instead.

When using ROCm SDK, install it into the venv:

pip install rocm-sdk
rocm-sdk init

Standalone (Linux)

The quickest way is invoke, which wraps the CMake steps:

pip install invoke
invoke build

Or manually with CMake:

cmake -S . -B build -GNinja \
  -DCMAKE_CXX_COMPILER=amdclang++ \
  -DCMAKE_C_COMPILER=amdclang \
  -DCMAKE_BUILD_TYPE=Debug

cmake --build build

Standalone (Windows)

pip install invoke
invoke build

As part of hipBLASLt

StinkyTofu is automatically built when hipBLASLt includes it via add_subdirectory. Tests and Python bindings are disabled in sub-project mode.

Rebuilding Python bindings after C++ changes

If you modify C++ sources, the installed .so becomes stale. Importing stinkytofu will raise an ImportError listing the modified files.

Standalone — rebuild with invoke or CMake directly:

invoke build
# or:
cmake --build build --target stinkytofu_python

As part of hipBLASLt — rebuild the Python bindings target in your hipBLASLt build directory:

cmake --build <build_dir> --target stinkytofu_python

Run clang-tidy

We use clang-tidy-20, so additional installation may required.

1. Download and Run the LLVM Setup Script

This script automatically detects your OS distribution, adds the correct LLVM stable/nightly APT repository, and updates your package listsspan_2span_2.

# Get the official script and make it executable
wget [https://apt.llvm.org/llvm.sh](https://apt.llvm.org/llvm.sh)
chmod +x llvm.sh

# Run the script to add repositories
sudo ./llvm.sh 20 reponly

Install clang-tidy-20

sudo apt-get update
sudo apt-get install -y clang-tidy-20

Run clang-tidy

invoke build
invoke tidy

Test

# All tests
cd build && ctest

# Unit tests only
./build/tests/unit_tests

# Single test
./build/tests/unit_tests --gtest_filter="DAGSchedulerPassTest.*"

# FileCheck tests
cd build && ctest -R FileCheck

Architecture

hardware/*.def  -->  TableGen  -->  generated .inc files
                                        |
                                    gfxisa library
                                        |
                                  stinkytofu library  -->  tools / Python bindings

• hardware/ - Per-architecture instruction definitions (.def files)
• include/stinkytofu/ - Public headers
• src/ - Library implementation (IR, passes, serialization, pipeline)
• tools/ - stinkytofu-opt, stinkytofu-check, intrinsic-compiler
• tests/ - Unit tests and FileCheck tests
• python_module/ - Python bindings

Documentation

See docs/ for detailed documentation including user guides, developer guides, and design documents.

License

This project is licensed under the MIT License. See LICENSE.md for details.