My final year project for Theoretical Physics.
This project aimed to write CUDA kernels for computing two point correlator functions. For this, I explored implementing matrix multipliation kernels. I managed to beat the CUBLAS CGEMM implementation on the computer I was working on by 35% for batch multiplication jobs.
For the particular problem I was working on, I could also fold some sparse matrix multiplications into my kernels, further improving overall performance.
Finally, by casting the final part of the problem from many matrix trace operations to several matrix multiplication problems, I improved performance in this task by two orders of magnitude.
The code is available in CUDA/TauMulti. The writeup is available in report