README_GPU

The GPU installation requires that the appropriate proprietary drivers for the GPU. The
OpenCL libraries must also be installed.

Notes for AMD

Install Catalyst 12 or newer drivers.

Install AMDAPP2.5 or more recent. Protinfo_cluster expects the path to be /opt/AMDAPP but
this can be changed by changing the compile flags. 

The OpenCL kernels are by default stored in  /opt/AMDAPP/include 
This can be changed using the -p option at runtime

AMD GPUs must spawn an xhost whether or not there is an actual monitor attached. Instructions
on how to run remote jobs in spite of this are described:

http://devgurus.amd.com/message/1284840
http://code.compeng.uni-frankfurt.de/projects/caldgemm/wiki/Headless_system

In addition, some versions of the Catalyst driver will crash very long jobs unless they are 
in an actual xterm. The workaround is to run the command using the xterm's -e option i.e 

xterm -e 'Protinfo_cluster <options>  >& captured_screen'

Hopefully this has been fixed in the newest drivers. The software requires GPUs that handle
double precision.

Notes for NVIDIA

cuda must be installed. As of Aug 2013, OpenCL support for cuda-5 had not been released. In cuda-4
some of the headers are not present and must be downloaded (thanks to Suhail A Islam for this)

cd /usr/local/cuda/include/CL
wget http://www.khronos.org/registry/cl/api/1.1/cl.hpp

Aternatively the cl.hpp from AMDAPP can be used.

The default location for cuda is /usr/local - and can be changed in the Makefile.

The OpenCL kernels are by default stored in  /usr/local/cuda/include 
This can be changed using the -p option at runtime

NVIDIA does not require an xhost to run and instead provides a persistent mode. This is done by

nvidia-smi -pm 1

Alternatively you can set up an xhost as with AMD. However, the xhost can be terminated and the 
GPU software can still be run subsequently without the xhost.

Hong Hung 8-13