OpenCL Torch

This is a distro of torch library enabled for OpenCL

Installation

Pre-requisites

python 2.7 installed: python command should point to python 2.7, during build (this is necessary for building clBLAS )
have an OpenCL-1.2 enabled GPU device available, and appropriate OpenCL-1.2 enabled drivers installed

Procedure

Please proceed as follows:

git clone --recursive https://github.com/hughperkins/distro -b distro-cl ~/torch-cl
cd ~/torch-cl
bash install-deps
./install.sh

Thats it! To test, you can do for example:

source ~/torch-cl/install/bin/torch-activate
luajit -l torch -e 'torch.test()'
luajit -l nn -e 'nn.test()'
luajit -l cltorch -e 'cltorch.test()'
luajit -l clnn -e 'clnn.test()'

If you're using CUDA, you can also run:

luajit -l cutorch -e 'cutorch.test()'
luajit -l cunn -e 'nn.testcuda()'

Alternative minimal no-gui install-deps

If you are using Ubuntu, and you dont need qt, itorch, or anything gui-like, then, instead of the line bash install-deps in the above instructions, you can do instead, according to your ubuntu version one of:

bash install-deps-nogui-ubuntu1404.sh

or:

bash install-deps-nogui-ubuntu1604.sh

This will install faster, since no qt packages will be installed.

Updating

Note that nn, torch, cutorch, and cunn are pinned, via the rocks-cl repository in your ~/torch-cl/install/etc/luarocks/config.lua file. So, doing any of luarocks install nn, luarocks install torch, luarocks install cltorch, luarocks install clnn, luarocks install cutorch, or luarocks install cunn should no longer break your installation (though they will, if you remove the pinning). However, on the whole, the recommended way of updating distro-cl is:

cd ~/torch-cl
git pull
git submodule update --init --recursive
./install.sh

If any errors like fatal: reference is not a tree, you have two options:

easy option: remove ~/torch-cl completely, reinstall
harder, hacky option:
- in the error message, you should see which submodule is broken. Let's say it is extra/nn
- so edit .git/modules/extra/nn/config, and in the url part, change torch to hughperkins
- if it's not extra/nn, then modify the path of this file appropriatel
- that's it!
- now rerun git submodule update --init --recursive, and the updates should pull down ok (otherwise raise an issue)

Unit-tests

To run, do:

source ~/torch-cl/install/bin/torch-activate
luajit -l torch -e 'torch.test()'
luajit -l nn -e 'nn.test()'
luajit -l cltorch -e 'cltorch.test()'
luajit -l clnn -e 'clnn.test()'

If you're using CUDA, you can also run:

luajit -l cutorch -e 'cutorch.test()'
luajit -l cunn -e 'nn.testcuda()'

Requests for additional operations etc

Please raise an issue for any operations etc which you particularly need, or you feel are not working for some reason.
(Ditto for any build errors)

FAQ

How does this relate to mainline torch?

It's a stabilized version of torch mainline. Torch mainline is kind of in permanent 'sid'-style experimental mode. This is great for rapidly evolving torch, but it kind of sucks to develop solid libraries against :-D This distro holds Torch stable, and allows for porting new features as and when, without getting emails at 4am because something has changed in Torch mainline, and broken clnn :-D

Wont this be behind main Torch bleeding edge?

Yes. But hopefully stable. And working. Please file an issue for any features you want from upstream.

Can I use `cunn` and `cutorch` from this distro?

Yes

Do you support ubuntu 16.04?

Yes. I am running Ubuntu 16.04 :-)

How to request new features, or pull new features from upstream?

Please file an issue.

Why dont you have any github stars?

They're all on the project pages (for now), ie:

cltorch
clnn

Related projects

The OpenCL is enabled by using the following two projects, which are installed implicitly by this distro:

cltorch
clnn

An hcc implementation for Torch is in progress here:

hctorch

Recent changes

22nd August:
- added :maskedSelect (thanks to Jacob Szuppe for showing me how to do this, using boost compute)
- behind the scenes:
  - added boost compute
  - this opens the door to implementing a few other missing methods, basically anything involving various types of scan operation
21st August:
- fixed unit test errors and warnings in nn module
20th August:
- bunch of patches to get neuralconvo to work:
  - in cltorch, apply, map, map2 have now become apply_on_gpu, map_on_gpu, map2_on_gpu2, and apply` is now the same as the cutorch version, ie copies to main memory, does the calcs on the cpu, then copies back to gpu, for compatibility with cutorch
  - updated Tester.lua in torch to the same as upstream main torch distro, so that rnn tests can run now
  - created rocks-cl rocks sources, to override upstream rocks for: rnn, torch7, nn, cutorch, cunn
2nd May:
- Re-applied:
  - 26th March:
    - added TemporalConvolution2: same API and usage as TemporalConvolution, but faster on GPUs
1st May:
- Re-applied:
  - 10th March:
    - @pawni (Nick Pawlowski) added SpatialUpSamplingNearest. Thank you Nick
  - 20th February:
    - @gloine (Jaehyung Lee) added support for non-batched input to ClassNLLCriterion. Thank you Jaehyung
  - 27 March 2016:
    - migrated from clBLAS 2.4 to clBLAS 2.11/develop. This migration is not set in stone, depends on how well that works. However, there is a bug in 2.4 for certain configurations of matrix multiplication, and its not obvious how to fix that, so maybe using 2.11/develop is the easiest way forward?
30th April:
- distro-cl first created, to stabilize Torch distribution

hughperkins/distro-cl