High-Performance FP32 Matrix Multiplication on CPU

Important note: Please don’t expect peak performance without fine-tuning hyperparameters such as the number of threads, kernel size and block sizes, unless you're running it on a Ryzen 7700(X). Current parallelization strategy is optimized for Intel and AMD x86 desktop CPUs. For many-core server processors, consider using nested parallelism and parallelizing 2-3 loops to increase the performance (e.g., the 5th, 3rd, and 2nd loops around the kernel).

Key Features

• Performance comparable to OpenBLAS and MKL
• Simple and scalable C code
• Supports arbitrary matrix sizes
• Efficiently parallelized with 3 lines of OpenMP directives
• Targets x86 processors with AVX2 and FMA3 instructions
• Follows the BLIS design
• Step by step, beginner-friendly tutorial

Installation

Install the following packages via apt if you are using a Debian-based Linux distribution

sudo apt-get install cmake build-essential python3-dev python3-pip libomp-dev

Create the virtual environment using pip or conda e.g.

python3 -m venv .venv
source .venv/bin/activate

and install the Python dependencies

python -m pip install -r requirements.txt

Optional:

To benchmark OpenBLAS, start by installing it according to the installation guide. During the installation, ensure you set an appropriate TARGET and disable AVX512 instructions. For instance, if you're using Zen4/5 CPUs, compile OpenBLAS with:

make TARGET=ZEN

Otherwise, OpenBLAS defaults to AVX512 instructions available on Zen4/5 CPUs.

Performance

Test enviroment:

• CPU: AMD Ryzen 7 7700 | AMD Ryzen 9 7900X
• CPU locked clock speed: 4.5GHz
• RAM: 32GB DDR5 6000 MHz CL36
• OpenBLAS v.0.3.26
• MKL 2023.1
• Compiler: GCC 11.4.0
• OS: Ubuntu 22.04.4 LTS

First, lock CPU clock speed:

sudo cpupower frequency-set -u CLK
sudo cpupower frequency-set -d CLK

For Ryzen 7700 I use CLK=4500mhz. Ensure that clock speed is stable and doesn't vary during the benchmark. You can check the CPU clock speed by running the command

watch -n 1 grep \"cpu MHz\" /proc/cpuinfo

To benchmark the matmul implementation, run

cmake -B build -S . -DOPENBLAS=OFF -DNTHREADS=X
cmake --build build
./build/benchmark MINSIZE MAXSIZE NPTS WARMUP

and set -DNTHREADS according to your CPU. On Ryzen 7700 I use -DNTHREADS=16. If not manually specified, default benchmark parameters are MINSIZE=200, MAXSIZE=8000, NPTS=40, WARMUP=5.

To benchmark OpenBLAS, run

cmake -B build -S . -DOPENBLAS=ON -DOPENBLAS_PATH=path/to/OpenBLAS/
cmake --build build
./build/benchmark MINSIZE MAXSIZE NPTS WARMUP

Or you can use

bash benchmark.sh /path/to/OpenBLAS NTHREADS

to benchmark both the code and OpenBLAS.

For the visualization of the results, simply run

python plot_benchmark.py

Tests

bash test.sh NTHREADS