Documentation of GPU Computing

theory - lab - project

Installation

Install OpenCL

NVIDIA compiler/workbench: see NVIDIA website
AMD compiler/workbench: see AMD website
Intel OpenCL SDK for Windows or Linux

installs in C:\\Program Files\IntelSWTools\system_studio_2020
or download the header files and dlls here. Put the right path in the project!

JOCL (download): java bindings if you want to use OpenCL from within Java.

JCuda (within java)

Available GPUs

See online excel form sent to you via canvas for all details
See powerpoint of first lesson to see the different Nvidia architectures

On machine 'parallel1' at the lab
On the CUDA server: howto-access it will be provided
At the VUB/ULB supercomputing center:

HYDRA hardware
getting an account:

ULB-students: apparently it is not possible anymore... Normally you all have a VUB account too
VUB-students: via the VSC and this generic procedure (I have sent a motivation mail)

You will need a public-private key pair

create it with puttygen as explained here (no need to convert to the openssh format)
the public key you will add to your HYDRA account (on the webpage)
the private key you add to putty and winscp - you will have to pass the passphrase each time you make a connection

accessing HYDRA

use putty: host internet address: <vscxxxx>@login.hpc.vub.ac.be; provide private key (not password) for autenthication (although pass phrase is also necessary)
use WinSCP to transfer files (provide private key under Advanced)
when you log in you arrive at a visitor computer on which you are not supposed to execute programns. You launch your work on the actual computers with qsub which is explained next.

working on HYDRA (thanks to Gian Marco, Ahmed and Matteo)

example GPU code (thanks to Gian Marco, Ahmed and Matteo)
documentation on .pbs job scripts
best is to rerun cmake each time (for me it didn't work otherwise). you can do this by deleting the build directory: put rmdir -f build before the mkdir -p build command
my example code compiles and runs on HYDRA (PBS-FILE - change extension into pbs), except for the listErosion project. To be added to the cpp file: #include <climits>
if qsub is complaining that "script is written in DOS/Windows text format" use dos2unix to convert the text file from DOS/Windows formatting to Linux formatting
basic linux commands

ls or dir: files and subdirectories of current directory

ls -al: list detailed contents of current directory

cd directoryname: change directory, cd .. to go back to the parent directory

pwd: show the path of the current directory

hostname: name of computer

mkdir directoryname: create a directory

ssh hostname: go to another machine (eg parallel1), use exit to go back

cp pathname-filename pathname: copy file to a different location, ~/ is your home directory

Control-c: stops a running program.

use the Up & Down Arrows to repeat commands (!)

copy: select a text, paste: click on right mouse button (or both mouse buttons at the same time)

TAB: to complete the name of a program, file or directory.

grep <options> <word> <files> to look for a word in files, use * to look in all files, use option -R for recursive search (in subdirectories)

use nano to edit text files in a terminal: nano <file>
with 'make clean' all compiled files (object files and executables) are removed

MPI on HYDRA

example MPI code
Counting3s MPI and Multithreaded: works on Visual Studio and Hydra (linux)

read the cpp-file for more info on this version

Make a terminal connection with putty
First run module load buildenv/default-foss-2019b to load the MPI environment through the toochain foss
Second run cmake CMakeLists.txt to create the makefile

a makefile contains the rules of how to compile and link the code
CMakeLists.txt contains all the information for the makefile, such as the MPI include & library directories, and the MPI libraries

Thirdly, run make all to compile the code

make all: everything is compiled
make clean: removes all files generated by the compilation process

Fourthly, submit a job to the cluster with qsub <pbs-file>

examine the status of the job with qstat
output and error files are automatically created and become available when the job is finished

the pbs-file contains the command to start several MPI processes

note that the job script will only run once on the first node

Demo code

Cmake-based solution (intro of cmake) VERSION of 26 February 2020 - OK for Windows and Linux

see readme.txt for install instructions

Cmake for windows
Projects:

deviceQuery: queries all OpenCL-enabled platforms and devices
copyFloat: OpenCL code to copy an array of floats to another array - memory throughput is measured
sumInts: adds two arrays into a third array - speedup and operations per second is measured
sumIntsMultipleRuns: measures statistics after running sumInts multiple times
sumIntsOccupancy: measures runtime in function of array size which is also the number of work items (occupancy)
listErosion: implements a one-dimensional erosion operation. Erosion replaces each value of a list with the minimal value of the surrounding value (= mask with certain size). Contains 2 versions: one with only global memory, a second using local memory.
the other code is not yet standardized and added to cmake: reduction, MxM, matrixVector

Java code: will be added on request.
Visual Studio:

make sure that the ocl-file is located with the exectutable. Best is to automatically copy it:

Project Properties -> Build Events -> Post-Build Event -> Command Line: xcopy /y $(ProjectDir)kernelsOnVectors.ocl $(OutDir)

you might have to copy
Opencl.dll to both C:/windows/System32 and C:/windows/SysWow64
the dll can be found on https://www.dllme.com/dll/files/opencl_dll.html (pick the Khronos one)