Important! The code must be compiled with Clang. Using GCC causes the program to run 1.5 to 2 times slower on my machine. Please don’t expect peak performance without fine-tuning the hyperparameters, such as the number of threads, kernel and block sizes, unless you are running it on a Ryzen 7700(X). More on this in the tutorial.
- Simple and scalable C code (<150 LOC)
- Step by step tutorial
- Targets x86 processors with AVX and FMA instructions (=all modern Intel Core and AMD Ryzen CPUs)
- Faster than OpenBLAS when fine-tuned for Ryzen 7700
- Efficiently parallelized with just 3 lines of OpenMP directives
- Follows the BLIS design
- Works for arbitrary matrix sizes
- Intuitive API
void matmul(float* A, float* B, float* C, const int M, const int N, const int K)
Tested on:
- CPU: Ryzen 7 7700 8 Cores, 16 Threads
- RAM: 32GB DDR5 6000 MHz CL36
- Numpy 1.26.4
- Compiler:
clang-17 - Compiler flags:
-O2 -mno-avx512f -march=native - OS: Ubuntu 22.04.4 LTS
To reproduce the results, run:
python benchmark_numpy.py
clang-17 -O2 -mno-avx512f -march=native -fopenmp benchmark.c -o benchmark.out && ./benchmark.out
python plot_benchmark.py
