Hacker News new | comments | show | ask | jobs | submit login

was the performance gain worth it in the end? In my experience numpy is pretty tight on its own, but I've seen some excellent speed gains from using cython + numpy.

Yeah, I got the runtime down from 40min to 10min. Then I implemented a Python->CUDA compiler and got it down to 30sec. :)

I implemented a Python->CUDA compiler

Wow, interesting. Is it released somewhere? Googling found me the Copperhead project[1], was that what you used?

I'm not sure if "implemented" means you implemented your code for it, or you implemented the entire compiler. :)

[1] http://code.google.com/p/copperhead/

No, it's a custom implementation of a simple compiler. It's nothing complicated. It converts Python to C++ and compiles that with nvcc. It also supports numpy arrays. It doesn't do any complex optimization steps like a full compiler. It's more like Cython, actually (with type annotations via an @gpu decorator). This allowed me to take my Python image processing code almost literally and annotate it with @gpu. The code isn't released, yet.

I originally wanted use Copperhead and got in contact with the developer a year ago, but it was too early even for "private" beta testing, so I never got access to their code. Also, my compiler is specialized on image processing, so probably Copperhead wouldn't have worked, anyway. I'm only jealous of Copperhead's type inferencer. :) But then again, I have to get finished with my thesis and a type inferencer wouldn't help with that goal. ;)

Interesting, thanks for explaining. :)

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact