This repository aims to benchmark Matrix Multiply (SGEMM) hand-tuned libraries and code generation stacks on a single thread on one CPU core. The focus will be on machine learning workloads so FP32 or smaller and irregular sizes of matrices. The goal is to expose high performance atomic kernels that can then be used to build highly efficient higher level implemenations spanning multiple cores or distributed across systems.

Note: 8GB Mac Mini runs roughly 25% slower than the 16GB version on other tests.

Clone the repo along with submodules.

git clone --recurse-submodules https://github.com/mmperf/mmperf.git

Create a virtual environment and install requirements. Python 3.8 is required.

cd mmperf
python3 -m venv ./mmperf_env
source mmperf_env/bin/activate
pip install -r requirements.txt
pip install -r ./external/llvm-project/mlir/python/requirements.txt

Build the project specifying the backend(s) to run matmul. Below is a command to build mmperf with MLIR backend.

cmake -GNinja \
    -DCMAKE_CXX_COMPILER=clang++-11 \
    -DCMAKE_C_COMPILER=clang-11 \
    -DUSE_MLIR=ON \
    -B build .

cmake --build build

Another example to build with all available backends. Assumes you have MKL, OpenBLAS, and Halide installed (see below for installation details)

HALIDE_DIR=/home/foo/lokal/halide/ MKL_DIR=/opt/intel/oneapi/mkl/latest/ cmake -GNinja \
    -DCMAKE_CXX_COMPILER=clang++-11 \
    -DCMAKE_C_COMPILER=clang-11 \
    -DMKL_DIR=/opt/intel/oneapi/mkl/latest/ \
    -DUSE_MLIR=ON \
    -DUSE_MKL=ON \
    -DUSE_RUY=ON \
    -DUSE_HALIDE=ON \
    -DUSE_OPENBLAS=ON \
    -DUSE_IREE=ON \
    -DIREE_LLVMCPU=ON \
    -B build .

cmake --build build

Benchmarking with any GPU backends, NVIDIA CUDA 11 should be pre-installed on your system. To enable CUDA compiler, -DCMAKE_CUDA_COMPILER=nvcc should be set in the command line. For example, to compile the IREE-CUDA and compare with cuBLAS and TVM-CUDA (TVM Auto-scheduler, a.k.a. Ansor) run this command line:

cmake -GNinja \
    -DCMAKE_CXX_COMPILER=clang++-11 \
    -DCMAKE_C_COMPILER=clang-11 \
    -DCMAKE_CUDA_COMPILER=nvcc \
    -DUSE_IREE=ON \
    -DIREE_CUDA=ON \
    -DUSE_CUBLAS=ON \
    -DUSE_TVM_CUDA=ON \
    -DTVM_ENABLE_CUDA=ON \
    -DUSE_TVM_TUNED=ON \
    -DTVM_LIB_DIR=/path/to/tvm-tuner
    -DSIZE_FILE=benchmark_sizes/bert_large_matmul.txt 
    -B build .

Note: -DTVM_LIB_DIR should be set as the absolute path of where TVM binaries located. For how to run TVM auto-scheduler, please refer to this README.

The building of submodule external/llvm-project can be space and time consuming. If you already have your own standalone llvm and don't want to fetch and compile this submodule, you scan specify the llvm on your system with PREBUILT_LLVM_PATH compilation flag:

cmake -GNinja \
    -DCMAKE_CXX_COMPILER=clang++-11 \
    -DCMAKE_C_COMPILER=clang-11 \
    -DPREBUILT_LLVM_PATH=$HOME/opt/llvm \
    -DUSE_MLIR=ON \
    -B build .

cmake --build build

To compile llvm from scratch, you might want all of these as well:

echo "deb http://apt.llvm.org/DISTRO_NAME/ llvm-toolchain-DISTRO_NAME main" >> /etc/apt/sources.list
wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key | apt-key add -
apt-get update && apt-get upgrade -y
apt-get install -y clang-11 clang-tools-11 libc++1-11 libc++-11-dev \
    libc++abi1-11 libc++abi-11-dev libclang1-11 libclang-11-dev \
    libclang-common-11-dev libclang-cpp11 libclang-cpp11-dev liblld-11 \
    liblld-11-dev liblldb-11 liblldb-11-dev libllvm11 libomp-11-dev \
    libomp5-11 lld-11 lldb-11 llvm-11 llvm-11-dev llvm-11-runtime \
    llvm-11-tools libfuzzer-11-dev

We use AOT compilation to generate binaries for matrix multiplication for specified backends and run them to generate the benchmarking numbers. To generate performance numbers and get a comparison plot run:

python3 mmperf.py ./build/matmul/ results

results folder will be created in the mmperf top-level directory which will contain GLOPS for every matmul size and every backend. A plot comparing performances of backends will also be generated in matmul.png.

Each generated binary can also be executed individually. To run a specific matrix size (say 24x64x512) for a backend run:

./build/matmul/matmul_<LIBRARY>_24x64x512

Matrix sizes: benchmark_sizes folder has text files containing the matrix sizes that mmperf runs on. You can change the matrix size input file by editing SIZE_FILE option in cmake/common.cmake. Default is benchmark_all_sizes.txt.

Precision: The default precision for all backends is FP32. FP16 benchmark has been added to cuBLAS, IREE, and triton backends. To enable FP16, use flag -DUSE_FP16=ON

To test the performance of openai-triton with half precision, one has to pip install triton, and then run the following command

python mmperf.py ./build/matmul results -triton -dtype f16 -benchmark_path=./benchmark_sizes/bert_large_matmul.txt

We have uploaded the sample of configuration files after running the Nod-ai SHARK auto-tuner. To test the performance of nodai_shark_cuda on Nvidia A100, one has to build IREE-CUDA (as shown in previous section) and then run:

python mmperf.py ./build/matmul/ results -nodai_shark_cuda -nodai_shark_configs=official_results/best_configs_minilm/

Note: The configuration is particularly optimized for Nvidia A100 GPU, and it may be not working well for other GPU architectures.

IREE: Intermediate Representation Execution Environment

Nod-ai SHARK

git clone https://github.com/halide/Halide.git --recurse-submodules
cd Halide/
sudo apt install libclang-11-dev clang-11 liblld-11-dev
LLD_DIR=/usr/lib/llvm-11/lib/cmake/lld cmake . -GNinja \
    -DCMAKE_BUILD_TYPE=Release \
    -DTARGET_WEBASSEMBLY=OFF \
    -DCMAKE_INSTALL_PREFIX=/home/<foo>/lokal/
ninja
ninja install
export HALIDE_DIR=/home/<foo>/lokal/halide

sudo apt install libopenblas-dev

git clone https://github.com/flame/blis
cd blis
./configure --prefix=/home/foo/lokal/ --enable-cblas -c amd64
make -j 16
make install

Download and install from https://software.intel.com/content/www/us/en/develop/articles/installation-guide-for-intel-oneapi-toolkits.html

This benchmark was run on an Intel Xeon CPU running at 3.1GHz. The machine has 256Kb L1 cache, 8Mb L2 cache and 24.8Mb L3 cache. It supports AVX-512 instructions. The peak performance of the machine is 3.1 x 8 x 2 x 2 = 99.2 GFLOPS for double precision and 198.4 GFLOPS for single precision.