C++ AMP backend Implementation for Torch7

##Introduction: ##

This repository hosts the C++ AMP backend implementation project for torch7. Torch7 framework currently has a CUDA backend support in the form of cutorch and cunn packages. The goal of this project is to develop gputorch and gpunn packages that would functionally behave as C++ AMP counterparts for existing cutorch and cunn packages. This project mainly targets the linux platform and makes use of the linux-based C++ AMP compiler implementation hosted here

##Prerequisites: ##

• dGPU: AMD firepro S9150
• OS : Ubuntu 14.04 LTS
• Ubuntu Pack: libc6-dev-i386
• AMD APP SDK : Ver 2.9.1 launched on 18/8/2014 from here
• AMD Driver installer: amd-driver-installer-14.301.1001-x86.x86_64

##Building and set up:
######Need to be a super user AMDAPPSDK on path /opt/AMDAPP (Might require to perform a rename)

(i) Torch7 install:

  *  curl -sk https://raw.githubusercontent.com/torch/ezinstall/master/install-deps | bash

  *  curl -sk https://raw.githubusercontent.com/torch/ezinstall/master/install-luajit+torch | bash

(ii) C++ AMP Compiler installation: Indepth details can be found here

Prepare a directory for work space.

• mkdir mcw_cppamp

• cd mcw_cppamp

• git clone https://bitbucket.org/multicoreware/cppamp-driver-ng.git src

• git checkout gmac-exp-cache-kernel (gmac-exp-cache-kernel branch is tailor made for torch7 use case) (note that you can also use git checkout origin/gmac-exp-cache-kernel)

Create a build directory and configure using CMake.

• mkdir mcw_cppamp/gmac_exp_build_cache

• cd mcw_cppamp/gmac_exp_build_cache

• cmake ../src -DCMAKE_BUILD_TYPE=Release (The gmac-exp-cache-kernel branch expects the AMDAPP SDK in the path /opt/AMDAPP)

Build the whole system. This will build clang and other libraries that require one time build.

• export CXXAMP_NV =1 to choose DMA implementation of clamp.

• make [-j #] world (# is the number of parallel builds. Generally it is # of CPU cores)

• make (this builds llvm utilities)

Note that you might need to manually check updates from C++ AMP Compiler. Please do the following and rebuild the Compiler if any update is available

 #
 # check updates from C++AMP Compiler
 cd mcw_cppamp/src
 git fetch --all
 git checkout origin/gmac-exp-cache-kernel

 # check updates from C++AMP Compiler's dependency
 cd mcw_cppamp/src/compiler/tools/clang
 git fetch --all
 git checkout origin/master

(iii) Bolt Set up:

Bolt binaries are automatically built in (ii)

• git checkout gmac-exp-cache-kernel (Need to get back to this)

(iV) Building gputorch and gpunn:

Prior to building these packages the following environment variables need to be set using export command

• AMDAPPSDKROOT=path to AMD APP SDK (Ideally /opt/AMDAPP)
• MCWCPPAMPROOT=path to mcw_cppamp dir
• CXXAMP_NV=1 (Need to make sure this is set to use DMA)

After this the following are the steps to build gputorch and gpunn

Build Torch7 Core:

• cd mcw_torch7_core

• luarocks make rocks/torch-scm-1.rockspec

Build NN pack

• cd mcw_nn

• luarocks make rocks/nn-scm-1.rockspec

Build gputorch pack

• cd gputorch

• luarocks make rocks/gputorch-scm-1.rockspec

Build gpunn pack

• cd gpunn

• luarocks make rocks/gpunn-scm-1.rockspec

Testing gputorch:

• cd gputorch

• th -lgputorch -e "gputorch.test()"

If fails, run the following to get detailed messages if any,

 th
 require ('gputorch')

Testing gpunn:

• cd gpunn

• th -lgpunn -e "nn.testgpu()"

Running FaceBook Benchmark:

Download the benchmark from here

• tar -xvf faceboook-imagenet-barebones.tar.gz

• cd imagenet-barebones

• th -i runMetaData.lua

• exit Torch prompt

• th -lgpunn -i main.lua