Merge pull request #37 from facebookresearch/update-docker-build

Update install TC from docker instructions

Author: prigoyal

docs/source/index.rst

@@ -34,7 +34,7 @@ Tensor Comprehensions provides framework-Agnostic Abstractions for High-Performa
    :caption: Installation
 
    installation
-   docker_image
+   installation_docker_image
    installation_conda_dep
    installation_conda
    installation_non_conda

docs/source/installation_conda.rst

@@ -169,6 +169,8 @@ Now check your proto version by running:
 
     $ protoc --version
 
+.. _conda_install_tc:
+
 Step 7: Installing TC
 ^^^^^^^^^^^^^^^^^^^^^
 

docs/source/installation_conda_dep.rst

@@ -131,6 +131,8 @@ Set environment variables:
     $ export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/targets/x86_64-linux/lib/stubs/:$LD_LIBRARY_PATH
     $ export PATH=/usr/local/bin:/usr/local/cuda/bin:$PATH
 
+.. _conda_dep_install_tc:
+
 Step 6: Install TC
 ^^^^^^^^^^^^^^^^^^
 

docs/source/docker_image.rst renamed to docs/source/installation_docker_image.rst

@@ -1,4 +1,4 @@
-Installing TC from docker image
+Installing TC from Docker image
 ===============================
 
 We provide docker runtime images for :code:`conda` and :code:`non-conda` environments both. TC officially supports
@@ -7,30 +7,53 @@ for Tensor Comprehensions at the `dockerhub <https://hub.docker.com/u/tensorcomp
 
 The conda and non-conda images for each setup are below:
 
+.. note::
+
+    Docker images below require 12Gb to 16Gb of RAM to run. If you need to adjust the
+    memory requirements, then please add :code:`--memory=16g` to the docker run command.
+    If your memory is not enough, TC will not build and you will see the errors like
+    `g++: internal compiler error: Killed (program cc1plus)`
+
+
 * :code:`Ubuntu 14.04 conda environment`
 
 .. code-block:: bash
 
     $ docker run -i -t tensorcomprehensions/linux-trusty-gcc4.8-tapir5.0-cuda8-cudnn6-py3-conda:x86
 
+Now, follow the instructions below to build TC:
+
+* **Option1:** If you want to build everything from scratch including TC dependencies, follow :ref:`conda_install_tc`.
+* **Option2:** If you want to build TC but using pre-built conda packages, follow :ref:`conda_dep_install_tc`.
+
+
 * :code:`Ubuntu 14.04 non-conda environment`
 
 .. code-block:: bash
 
     $ docker run -i -t tensorcomprehensions/linux-trusty-gcc4.8-tapir5.0-cuda8-cudnn6-py3:x86
 
+Now, to install TC, follow :ref:`non_conda_install_tc`
+
+
 * :code:`Ubuntu 16.04 conda environment`
 
 .. code-block:: bash
 
     $ docker run -i -t tensorcomprehensions/linux-xenial-gcc5-tapir5.0-cuda9-cudnn7-py3-conda:x86
 
+We don't ship the conda packages for TC dependencies that are compatible with gcc5 yet, so you have to
+build them from source. To install TC, see instructions here: :ref:`conda_install_tc`
+
+
 * :code:`Ubuntu 16.04 non-conda environment`
 
 .. code-block:: bash
 
     $ docker run -i -t tensorcomprehensions/linux-xenial-gcc5-tapir5.0-cuda9-cudnn7-py3:x86
 
+Now, to install TC, follow :ref:`non_conda_install_tc`
+
 
 TC runtime image with nvidia-docker
 -----------------------------------
@@ -44,3 +67,8 @@ have :code:`nvidia-docker` installed. Then run the image using following command
 .. code-block:: bash
 
     $ nvidia-docker run --rm -i -t tensorcomprehensions/trusty-gcc4.8-py3-conda-cuda:x86
+
+Now, follow the instructions below to build TC:
+
+* **Option1:** If you want to build everything from scratch including TC dependencies, follow :ref:`conda_install_tc`.
+* **Option2:** If you want to build TC but using pre-built conda packages, follow :ref:`conda_dep_install_tc`.

docs/source/installation_non_conda.rst

@@ -146,6 +146,7 @@ Now, check your yaml import works. If it doesn't, make sure you ran earlier step
 
     $ python3 -c 'import yaml'
 
+.. _non_conda_install_tc:
 
 Step 7: Install TC
 ^^^^^^^^^^^^^^^^^^