baidu-research / persistent-rnn

Fast Recurrent Networks Library

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Various issues with GTX 1080

SeanNaren opened this issue · comments

Hey!

I'm having difficulties getting the library to work correctly with my GTX 1080.

It doesn't seem to recognize that my graphics card is of sm_60 and arch major 6. If i manually change variables in the nvcc.py file to force sm_60 it does compile.

This goes into my next issue which is the cudaDevAttrMultiProcessorCount is returning 20 when calling getMaximumSizeRNNForThisGPU, which means I don't get the tile sizes associated with pascal.

Also I had to make a modification here, I assume line 54 in recurrent_ops.cu:

typedef TileConfig<56, 2432, 2560, 352, 320, 11, 16, direction, T> TileSize;

is supposed to be:

typedef TileConfig<56, 2432, 2560, 352, 320, 11, 16, direction, float16> TileSize;

Hopefully can solve the issue, let me know if any more information is needed!

I'm running Ubuntu 14.04 using CUDA 8.0, V8.0.26 (RC) with driver (installed after CUDA as suggested for 1080 drivers) 367.35.

Just wanted to note I have the same issues with a Titan X (non-pascal) with CUDA 7.5.