https://github.com/ollycassidy13/cmatmul
A cache-based matrix multiplication kernel
https://github.com/ollycassidy13/cmatmul
cache cpp matrix-multiplication openmp-parallelization
Last synced: 3 months ago
JSON representation
A cache-based matrix multiplication kernel
- Host: GitHub
- URL: https://github.com/ollycassidy13/cmatmul
- Owner: ollycassidy13
- License: mit
- Created: 2025-03-26T13:54:46.000Z (3 months ago)
- Default Branch: main
- Last Pushed: 2025-03-26T14:23:24.000Z (3 months ago)
- Last Synced: 2025-03-26T15:37:50.471Z (3 months ago)
- Topics: cache, cpp, matrix-multiplication, openmp-parallelization
- Language: C++
- Homepage:
- Size: 517 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE.txt
Awesome Lists containing this project
README
# CMATMUL - Cache Based Matrix Multiplication Kernel
CMATMUL is an optimised C++ implementation of matrix multiplication that leverages modern CPU features such as AVX2/FMA SIMD vectorisation and OpenMP multi-threading for increased throughput. This project examines a naïve triple-loop algorithm and uses the findings, presented [here](docs.md), to implement a high-performance kernel through techniques like cache tiling, register blocking, and memory access optimisation.
## Overview
Matrix multiplication is fundamental to many applications in computing, data analysis, and machine learning. However, the naïve approach underutilises modern hardware's capabilities significantly. CMATMUL tackles this by:
- **Memory Access Optimisation:** Reordering accesses to maximise cache line usage.
- **Register Blocking:** Keeping small blocks of data in CPU registers for rapid reuse.
- **Tiling for Cache Locality:** Dividing matrices into cache-friendly blocks.
- **SIMD Vectorisation:** Using AVX2/FMA intrinsics to process multiple floats concurrently.
- **OpenMP Parallelisation:** Distributing work across multiple cores to scale performance.
By combining these techniques, the optimised kernel significantly outperforms a naïve implementation, achieving over 100× the throughput in tested configurations.
## Getting Started
```bash
git clone https://github.com/ollycassidy13/CMATMUL
cd CMATMUL
```
### Compilation
Compile the code using the following command:
```bash
g++ -O3 -mavx2 -mfma -fopenmp -std=c++11 cmatmul.cpp -o cmatmul
```
> Note: A modern C++ compiler with support for C++11 (or later) and OpenMP, and a CPU with AVX2 and FMA support are needed
This command enables high optimisation, AVX2, FMA, and OpenMP to fully exploit the hardware capabilities.
### Use
#### With OpenMP (Multi-threaded):
```bash
export OMP_NUM_THREADS=4 # Set number of threads (e.g., 4)
./cmatmul
```
#### Without OpenMP (Single-threaded):
```bash
export OMP_NUM_THREADS=1
./cmatmul
```
## Documentation
Full documentation on how the example implementation's methodology works is provided in [docs.md](docs.md)
---
*Note: The code provided in this repository is intended for educational purposes. Users are encouraged to experiment with and modify the code to suit their specific hardware configurations and application requirements.*