WIP: CUDA performances - Githubissues

EconForge / interpolation.py

BSD 2-Clause "Simplified" License

124 stars 35 forks source link

WIP: CUDA performances #6

Open albop opened 8 years ago

albop commented 8 years ago

Currently, the experiments with CUDA yield terrible performances on AWS/g2. It may be because, calculations are made on 64 bits by default. To do :

test 32 bits calculations
test 64 bits on a card which supports it natively (e.g. gtx titan)
try guvectorize(...target='gpu')

tilman-g commented 5 years ago

Dear albop, I started using the interpolation package and I am amazed by the speed-up of jitting. I would like to explore the GPU-side a bit more as I would be happy to speed things up a bit more and contribute code. Has there been any attempts so far to re-activate the cuda code and if so, what is the current status?

albop commented 5 years ago

No, I did not attempt to reactivate the code, mostly for lack of time and other distractions. One idea would be to tweak a bit the output of codegen. From this discussion, https://groups.google.com/a/continuum.io/forum/#!msg/numba-users/8Hn6GagrXXU/ivQa4JUzCgAJ , I gather there is nothing fundamentally wrong with this approach. The other idea, would be to test, whether the current eval_linear function can be called on the gpu inside a cuda kernel. This is not totally impossible and would simplify things a lot. Can you try it ?

tilman-g commented 5 years ago

Okay, that makes sense. I am a new to GPU coding, but would really like to try. So I will look into it the next few weeks and see what I can do. I will focus on the eval_linear function for the moment.