Skip to content

Latest commit

 

History

History
66 lines (52 loc) · 1.83 KB

build-from-source.md

File metadata and controls

66 lines (52 loc) · 1.83 KB
Raw

Building from source

Pre-requisites

  • you need an OpenCL-enabled GPU installed and OpenCL drivers for that GPU installed. Currently, supported OpenCL version is 1.2 or better

    • To check this: run clinfo, and check you have at least one device with:
      • Device Type: 'GPU', and
      • Device OpenCL C Version: 1.2, or higher
    • If you do, then you're good 👍

  • normal non-GPU tensorflow prerequisites for building from source

  • then do:

git clone --recursive https://github.com/hughperkins/tf-coriander
cd tf-coriander
bash ./install_deps.sh

Build

From the root of the cloned tf-coriander repo, do:

bash ./build.sh

Install

source env3/bin/activate
pip install --upgrade soft/tensorflowpkg/tensorflow-0.11.0rc0-py3-none-any.whl

Test/validate

source env3/bin/activate
( cd; python -c 'import tensorflow' )
# check no errors
python tensorflow/stream_executor/cl/test/test_simple.py
# hopefully no errors :-)
# expected result:
# [[  4.   7.   9.]
# [  8.  10.  12.]]
py.test -v
# hopefully no errors :-)

If you do get errors, please collect together as much inforamtion as possible, and save to a gist, then create an issue on the github. I'll want:

  • the github commit of tf-coriander and coriander, that you are using
  • which operating system
  • which GPU(s)
  • the output of clinfo
  • the full output of the command that produced the error, as well as the command

Updating

  • if you pull down new updates from the tf-coriander repository, you need to update the coriander installation:
git submodule update --init --recursive
  • .. and then redo the build

Dockerfile

There is a dockerfile, based on Ubuntu 16.04, at docker. Whilst this probalby wont run as such, it provides a useful reference to how to build, possibly.