Feature/misc fixes

[mgehre-amd]

Uses: Xilinx/llvm-project#35

To build:

python3.8 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

cmake -GNinja -Bbuild \
  -DCMAKE_BUILD_TYPE=Release \
  -DPython3_FIND_VIRTUALENV=ONLY \
  -DLLVM_ENABLE_PROJECTS=mlir \
  -DLLVM_EXTERNAL_PROJECTS="torch-mlir;torch-mlir-dialects" \
  -DLLVM_EXTERNAL_TORCH_MLIR_SOURCE_DIR="$PWD" \
  -DLLVM_EXTERNAL_TORCH_MLIR_DIALECTS_SOURCE_DIR="$PWD"/externals/llvm-external-projects/torch-mlir-dialects \
  -DMLIR_ENABLE_BINDINGS_PYTHON=ON \
  -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
  -DLLVM_TARGETS_TO_BUILD=host  \
  -DCMAKE_C_COMPILER=clang-15 -DCMAKE_CXX_COMPILER=clang++-15 \
  -DLLVM_ENABLE_ASSERTIONS=ON \
  -DCMAKE_EXE_LINKER_FLAGS_INIT="-fuse-ld=lld" -DCMAKE_MODULE_LINKER_FLAGS_INIT="-fuse-ld=lld" -DCMAKE_SHARED_LINKER_FLAGS_INIT="-fuse-ld=lld" \
  externals/llvm-project/llvm

cmake --build build --target tools/torch-mlir/all

To run a script with this torch-mlir version:

export PYTHONPATH=$PWD/build/tools/torch-mlir/python_packages/torch_mlir
python some_model.py

Run a single test with tosa backend:

./tools/e2e_test.sh --filter Quantize -v -s -c tosa

Run all tests:

cmake --build build --target check-torch-mlir-all

echo ":::: Check that update_abstract_interp_lib.sh has been run"
./build_tools/update_abstract_interp_lib.sh

echo ":::: Check that update_torch_ods.sh has been run"
./build_tools/update_torch_ods.sh

echo ":::: Run Linalg e2e integration tests"
./tools/e2e_test.sh --config=linalg -v

echo ":::: Run MHLO e2e integration tests"
./tools/e2e_test.sh --config=mhlo -v

echo ":::: Run TOSA e2e integration tests"
./tools/e2e_test.sh --config=tosa -v

echo ":::: Run Lazy Tensor Core e2e integration tests"
./tools/e2e_test.sh --config=lazy_tensor_core -v

echo ":::: Run TorchDynamo e2e integration tests"
./tools/e2e_test.sh --config=torchdynamo -v

How to release a package:

• Go to https://github.com/Xilinx/torch-mlir/actions/workflows/releaseSnapshotPackage.yml
• Trigger new workflow with branch misc_fixes

How to install torch-mlir from a github release:

# This will also install an exact torch nightly version as required by torch-mlir
pip install <path-to-wheel.whl> -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html

[mgehre-amd]

Has been integrated into #105