CUDA Matrix Multiplication - Submission by meteharun

What I did

• Implemented two CUDA kernels:
  • naive_cuda_matmul: a basic version that multiplies matrices without any optimization.
  • tiled_cuda_matmul: an improved version using shared memory and tiling to make it faster.

Optimizations

• Used shared memory to store tiles of A and B, which reduces the number of slow global memory accesses.
• Applied tiling (block-level matrix multiplication), so threads in the same block work together on small submatrices.
• Added proper synchronization between threads to make sure shared memory is used correctly.

Challenges

• CUDA compatibility issues in puhti
• Avoiding reuse of variables like cudaEvent_t start in both kernels, which caused compilation errors.

Results

• Added a table to README.md showing the timing results for both CUDA versions, and how much faster the tiled version is compared to the naive one and the CPU version.