Skip to content

BuildOnSG2044.md

Latest commit

 

History

History
129 lines (93 loc) · 4.33 KB

BuildOnSG2044.md

File metadata and controls

129 lines (93 loc) · 4.33 KB
Raw

Build and Run on SG2044

This document describes the validated native bring-up flow for the structured RISC-V backend on an SG2044 machine. It is intentionally separate from the generic installation guide because this path uses a real RISC-V host, a Buddy / LLVM toolchain, and a native GCC sysroot instead of the usual x86 or GPU development environment.

Scope

The validated flow is:

  • TileLang -> MLIR Linalg -> MLIR vector / RVV lowering
  • native host shared library build with the RISC-V GCC toolchain
  • execution on SG2044

This document is about native execution on SG2044.

Get the Sources

Clone the repository with its bundled dependencies, including the TileKernels submodule used by the RISC-V operator-library tests:

git clone --recursive https://github.com/tile-ai/tilelang.git tilelang-riscv
cd tilelang-riscv

If you already cloned the repository without submodules, run:

git submodule update --init --recursive

Environment Setup

Activate the Python environment and export the toolchain variables:

source ~/.venv-buddy/bin/activate

export TILELANG_RISCV_LLVM_ROOT=/path/to/buddy-mlir/llvm/build
export Z3_ROOT=/path/to/.local/z3
export CC=/path/to/gcc
export CXX=/path/to/g++
export PATH=/path/to/gcc-bin:$TILELANG_RISCV_LLVM_ROOT/bin:$Z3_ROOT/bin:$PATH
export LIBRARY_PATH=/path/to/gcc-lib64/lp64d:/path/to/gcc-libgcc/lp64d${LIBRARY_PATH:+:$LIBRARY_PATH}
export TVM_FFI_DISABLE_TORCH_C_DLPACK=1
export CMAKE_ARGS="-DUSE_CUDA=OFF -DUSE_ROCM=OFF"

Notes:

  • Z3_ROOT is required when z3-solver wheels are unavailable on riscv64
  • if the native toolchain uses a split lib64/lp64d layout, include it in LIBRARY_PATH or CMake's compiler sanity check can fail on -lgcc_s
  • TVM_FFI_DISABLE_TORCH_C_DLPACK=1 avoids first-import JIT building of the optional Torch DLPack extension
  • torch-c-dlpack-ext is skipped on riscv64; the validated SG2044 flow uses TVM FFI without that optional prebuilt extension

Install Dependencies

Install the Python-side build dependencies first:

pip install scikit-build-core cython patchelf setuptools_scm cloudpickle pytest

Install the vendored TVM FFI package from the source tree:

cd tilelang-riscv
pip install ./3rdparty/tvm/3rdparty/tvm-ffi --no-build-isolation --no-deps

Then install TileLang itself in editable mode:

cd tilelang-riscv
pip install -e . --no-build-isolation -v

Build Notes

The SG2044 path depends on a few implementation details that are already wired into this tree:

  • use target="riscv" as the backend name
  • toolchain discovery covers sibling Buddy builds such as ../buddy-mlir/llvm/build
  • the RISC-V host wrapper passes --gcc-toolchain and --sysroot when it detects /opt/gcc-native
  • RISC-V builds disable TVM's alternative linker selection because ld.lld failed on the validated SG2044 toolchain

If you need to rebuild after source changes, the same editable install command is sufficient:

cd tilelang-riscv
pip install -e . --no-build-isolation -v

Validation

The validated native test set is:

cd tilelang-riscv
pytest testing/python/riscv/test_riscv_target_parse.py -q
pytest testing/python/riscv/test_riscv_toolchain.py -q
pytest testing/python/riscv/test_riscv_tladapter_pipeline.py -q
pytest testing/python/riscv/test_riscv_cache.py -q
pytest testing/python/riscv/codegen_ops -q
pytest testing/python/riscv/TileKernels -q

Useful extended runtime checks:

pytest testing/python/riscv/test_riscv_jit_runtime.py -q
pytest testing/python/riscv/test_riscv_artifact_export.py -q

test_riscv_jit_runtime.py and test_riscv_artifact_export.py exercise broader runtime and artifact-export paths and may expose additional toolchain or runtime issues beyond the smoke set above.

Troubleshooting

  • If configuration fails with a Z3 lookup error, check that Z3_ROOT points to a prefix containing include and lib
  • If the first CMake configure fails on -lgcc_s, verify that the toolchain's lib64/lp64d directory is present in LIBRARY_PATH
  • If runtime shared library linking fails, verify that CC, CXX, and /opt/gcc-native/sysroot are all present
  • If tvm_ffi import fails, reinstall the vendored package from 3rdparty/tvm/3rdparty/tvm-ffi
  • If first import becomes slow or tries to build extra Torch extensions, keep TVM_FFI_DISABLE_TORCH_C_DLPACK=1