Several common methods of matrix multiplication are implemented on CPU and Nvidia GPU using C++11 and CUDA. The performance benefits of each optimization method were simply tested.

CPU

• naive
• reordering
• tiling
• strassen
• coppersmith-winograd

Nvidia GPU

• cublas
• naive
• kahan
• shared_memory

Compile

Environment

• OS: Linux
• Cmake Version: >= 3.8
• GCC Version: >= 4.8
• CUDA Version: 11.4 (best)
• CUDA Driver Version: 470.129.06 (best)

Clone

git clone https://github.com/Bruce-Lee-LY/matrix_multiply.git

Build

cd matrix_multiply
./build.sh -t Release -b OFF
./build.sh -t Debug -b ON

Run Sample

./run_sample.sh

Performance

• OS: Ubuntu 20.04.4
• CPU: i5-9400F
• GPU: NVIDIA GeForce GTX 1080 Ti
• CUDA Version: 11.4
• CUDA Driver Version: 470.129.06
• Matrix (float): A (512 * 512) * B (512 * 512) = C (512 * 512)

CPU

Method	Cost / ms
naive	1238.647
reordering	984.445
tiling	1000.095
strassen	57429.407
coppersmith-winograd	77668.238

Nvidia GPU

Method	Cost / ms
cublas	0.100
naive	0.613
kahan	0.616
shared_memory	0.153