search

Celebrandil / CudaSift

A CUDA implementation of SIFT for NVidia GPUs (1.2 ms on a GTX 1060)
MIT License
860 stars 286 forks source link

Turing coming on scene #56

Open msububble opened 5 years ago

msububble commented 5 years ago

Hello, Celebrandil!

Thanks for sharing the implementation of the SIFT algorithm on CUDA!

Have you started exploring the Turing architecture, may there occur any difficulties with compiling and running the existing version on CUDA 10.0 (omitting the issue of updating the code to meet new hardware features, just running as it is now)?

msububble commented 5 years ago

Answering my own question: yes, cudaSift.dll does compile and work on RTX2070, on cuda 10.0 toolkit.

Celebrandil commented 5 years ago

It's good to hear that you got it running. As soon as I have time, I will try to explore Turing and see what CUDA 10 has to offer. Most likely though, it will be hard to find any immediate improvements. You might benefit from the tensor cores, but since most operations are memory bound, you benefit more from changes in the memory architecture than from improvements in compute.