This is an experiment in using MLIR to automatically select the best Tensor Processing Primitives for linear algebra.
This repository contains an out-of-tree MLIR dialect as well as an opt-like tool to operate on that dialect and a runner-like tool to execute and benchmark MLIR kernels.
It also contains the recipes to use LIBXSMM from inside MLIR and can be used by other tools to drive our passes.
There's work in progress inside IREE to use this work on their pipeline.
This repository was previously called tpp-sandbox.
If you have a checkout with the previous name, please follow these instructions to rename the remote locally.
| Build | Status |
|---|---|
| Tests |  |
| Benchmarks |  |
In order to build LLVM and TPP-MLIR, several software development tools such as git, cmake, compilers, etc. are needed. As each operating system has its own package
manager and package names, we opted for providing instructions for the user-level package manager conda. This environment has been successfully tested on top of a Fedora Server
minimal installation with less than 400 system-wide packages being installed.
Initial Setup (using Conda):
export TPPMLIR_WORKSPACE_DIR=/foo
cd ${TPPMLIR_WORKSPACE_DIR}
export ARCH_NAME=$(uname -m)
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-${ARCH_NAME}.sh
bash Miniconda3-latest-Linux-${ARCH_NAME}.sh -b -p ${TPPMLIR_WORKSPACE_DIR}/miniconda3
eval "$(${TPPMLIR_WORKSPACE_DIR}/miniconda3/bin/conda shell.bash hook)"
conda activate
conda install -y cmake ninja git clang clangxx llvm lld llvm-openmp llvm-tools binutils
if [ "${ARCH_NAME}" == "aarch64" ]; then
conda install -y gcc_linux-aarch64 gxx_linux-aarch64
elif [ "${ARCH_NAME}" == "x86_64" ]; then
conda install -y gcc_linux-64 gxx_linux-64
fi
python -m pip install coloredlogsReloading the environment after conda deactivate/logout/reboot:
export TPPMLIR_WORKSPACE_DIR=/foo
cd ${TPPMLIR_WORKSPACE_DIR}
eval "$(${TPPMLIR_WORKSPACE_DIR}/miniconda3/bin/conda shell.bash hook)"
conda activate# Clone
git clone https://github.com/llvm/llvm-project.git
# checking out a tpp-mlir compatible version of llvm-project
wget https://raw.githubusercontent.com/plaidml/tpp-mlir/main/build_tools/llvm_version.txt
pushd llvm-project
git checkout `cat ../llvm_version.txt`
popd
rm llvm_version.txt
# create build dir
mkdir llvm-project/build
pushd llvm-project/build
# This is important for the next step
export CUSTOM_LLVM_ROOT=`pwd`
echo $CUSTOM_LLVM_ROOT
export PATH=$CUSTOM_LLVM_ROOT/bin:$PATH
# Configure Build
cmake -G Ninja ../llvm \
-DLLVM_ENABLE_PROJECTS="mlir" \
-DLLVM_BUILD_EXAMPLES=ON \
-DLLVM_INSTALL_UTILS=ON \
-DLLVM_TARGETS_TO_BUILD="host" \
-DCMAKE_BUILD_TYPE=RelWithDebInfo \
-DLLVM_ENABLE_ASSERTIONS=ON \
-DCMAKE_C_COMPILER=clang \
-DCMAKE_CXX_COMPILER=clang++ \
-DLLVM_USE_LINKER=lld
# Build
ninja
popdThis setup assumes that you have built LLVM and MLIR in $CUSTOM_LLVM_ROOT as above.
Note: OpenMP is a requirement to get multi-threaded performance on our code.
If you don't want to build with OpenMP, disable with the CMake flag -DUSE_OpenMP=False.
# Clone
git clone https://github.com/plaidml/tpp-mlir.git
mkdir tpp-mlir/build
pushd tpp-mlir/build
# Build & test
# Please, make sure to use clang to build TPP-MLIR
cmake -G Ninja .. \
-DCMAKE_BUILD_TYPE=RelWithDebInfo \
-DMLIR_DIR=$CUSTOM_LLVM_ROOT/lib/cmake/mlir \
-DLLVM_EXTERNAL_LIT=$CUSTOM_LLVM_ROOT/bin/llvm-lit \
-DCMAKE_C_COMPILER=clang \
-DCMAKE_CXX_COMPILER=clang++
cmake --build . --target check-all
popdTo build the documentation from the TableGen description of the dialect operations, run:
cmake --build . --target mlir-docTo enable experimental GPU support see: GPU/README.md
This dialect template is made available under the Apache License 2.0 with LLVM Exceptions. See the LICENSE.txt file for more details.
BRGEMM: High-Performance Deep Learning via a Single Building Block (2019)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent