xGeMM

Accelerated General (FP32) Matrix Multiplication. Tested on NVIDIA RTX 3090 using Ubuntu 24.04.1 LTS with nvidia-driver-550 and CUDA 12.4.

Watch the YouTube video (click the image below)

Dependencies

• Eigen 3.4.0 (Put it in lib)

Running Benchmarks

1. Eigen (CPU) matrix multiplication

Compile: make 00a_benchmark_cpu.out

Execute: ./00a_benchmark_cpu.out

2. cuBLAS (GPU) matrix multiplication:

Compile: make 00b_benchmark_cuBLAS.out

Execute: ./00b_benchmark_cuBLAS.out

3. Naive (GPU) matrix multiplication:

Compile: make 01_benchmark_naive.out

Execute: ./01_benchmark_naive.out

4. Coalesced (GPU) matrix multiplication:

Compile: make 02_benchmark_coalesced.out

Execute: ./02_benchmark_coalesced.out

5. Tiled (GPU) matrix multiplication:

Compile: make 03_benchmark_tiled.out

Execute: ./03_benchmark_tiled.out

6. 1D thread coarsening (GPU) matrix multiplication:

Compile: make 04_benchmark_coarse_1d.out

Execute: ./04_benchmark_coarse_1d.out

7. 2D thread coarsening (GPU) matrix multiplication:

Compile: make 05_benchmark_coarse_2d.out

Execute: ./05_benchmark_coarse_2d.out

8. Vectorized Mmemory accesses (GPU) matrix multiplication:

Compile: make 06_benchmark_coarse_2d_vec.out

Execute: ./06_benchmark_coarse_2d_vec.out