When using opencl with torch, lua uses integrated graphics not dedicated gpu

Question

I am using Andrej Karpathy's code to train a rnn. When I give the flag "-opencl 1" to tell it to use an opencl gpu, it uses integrated graphics and nothing else.

I tried reinstalling cltorch and using different flags, but nothing has seemed to work. To add to this I can't see if my gpu is under load because I'm on macos. I looked through the code and I could't find any errors, but I have little experience with lua.

Code can be found here: https://github.com/karpathy/char-rnn.

I expect with the flag "opencl 1" or something to the like, my radeon pro 560x will be used to train on my dataset and not my cpu or integrated graphics.

Have you tried different values for opt.gpuid into setDevice? Reading cltorch there is a "getDeviceCount" which probably returns 2 or 3; CPU, integrated graphics and discrete GPU, depending on what you have installed drivers for. — Andreas
I have tried using a 2 instead of 1 but it defaulted to the CPU. I’ll have to try getDeviceCount. Thanks for the comment. — AWOLASAP

AWOLASAP AWOLASAP · Accepted Answer · 2019-04-27T20:42:29

When reading instructions I thought just -opencl flag needed to be used but it turns out -gpuid needs to be used in conjuction with it as well. This was also after a reinstall of torch and opencl drivers, so that could have been a problem as well.

When using opencl with torch, lua uses integrated graphics not dedicated gpu

1 Answers