Dyno MLIR

🚧 Work in progress

Dyno is an MLIR-based prototype compiler framework designed for dynamic sized vector architectures.

Get Started

1. Clone and initialize the repository:

git clone <repo-url> dyno-mlir
cd dyno-mlir
git submodule update --init --recursive

2. Install dependencies:

sudo apt install build-essential cmake ninja-build clang lld

Optional: to build the RISC-V GNU Toolchain, first apply the workaround patch:

./scripts/workaround-gnutoolchain.sh

Then check the README in third_party/riscv-gnu-toolchain to install dependencies.

Note that the GNU Toolchain is by default enabled in the script, and it also enables an LLVM build to support OpenMP header and libraries, so it will take longer to build.

3. Compile LLVM & MLIR

./scripts/build-llvm.sh

4. Compile Dyno

./scripts/build-dyno.sh

Local Development

For normal host-side development, the existing LLVM and Dyno build flow is unchanged:

./scripts/build-llvm.sh
./scripts/build-dyno.sh

RISC-V Benchmarking

To build the benchmarks for RISC-V platforms, cross-compile the project.

source ./scripts/build-env.sh
./scripts/build-riscv-llvm.sh
./scripts/setup-triton.sh build-riscv
./scripts/build-riscv-dyno.sh

Set ENABLE_TRITON_BENCHMARKS=OFF before ./scripts/build-riscv-dyno.sh if you want to skip the Triton benchmarks entirely.

The Triton setup is split on purpose:

• third_party/triton is pinned to the commit required by third_party/triton_shared
• third_party/triton_shared provides the shared lowering and Python backend
• third_party/triton-cpu is kept on its own head and is only used for triton-opt

This avoids having to patch any Triton source tree locally just to make the benchmarks compile.

Running Tests and Benchmarks

IR lit tests

Run the compiler IR regression suite from the host build:

cmake --build build --target check-dyno

or use ninja

ninja -C build check-dyno

Runtime regressions

Run the runtime regression suite from the RISC-V build:

ctest --test-dir build-riscv -L runtime-regression --output-on-failure

Or use the convenience target:

cmake --build build-riscv --target check-dyno-runtime-regressions

Benchmarks

Benchmarks are built as executables under build-riscv/benchmarks/.... Run them directly, for example:

build-riscv/benchmarks/linalg/matmul/linalg-matmul-benchmark
build-riscv/benchmarks/dyno/matmul/dyno-matmul-benchmark
build-riscv/benchmarks/triton/matmul/triton-matmul-benchmark

qemu-riscv64 may be required to run these benchmarks on a non-RISC-V machine.