Comment by unnouinceput
4 years ago
The one I wrote 2 years ago? That's intellectual property of said data scientist, not mine. Al I can say is that I parallelize it a lot, hence the entire month. From programming point of view is a mess and hard to follow its ~5k lines. Usually parallel programming is a mess, you should take a look at any parallelization CUDA code available on GitHub.
Was the original just regular python, or numpy?
The C version wasn't GPU-targeted though from your description. I'm curious what other implementations would be capable of, for instance julia, maybe gpu-targeted.
We discussed, since we already agreed on parallelization, if he wanted CUDA, since that would've been even faster. But after discussing with his team, he said no GPU dependent implementation and I started the work. He never shared why no GPU implementation and I didn't pressed the matter further since I was already knee deep in trying to understand the algorithm which was the bigger stone to crack at the time.