Skip to content

An implementation of a workgroup occupancy discovery protocol and an inter-workgroup barrier. Also example applications.

Notifications You must be signed in to change notification settings

mc-imperial/gpu_discovery_barrier

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 

Repository files navigation

Author: Tyler Sorensen

Last revised: 1 Sept. 2016

Table of Contents

  • Introduction

  • Licenses and Copyright

  • Running the Code

    • Obtaining Inputs
    • Running
  • Known Issues

Introduction

This is the code for the paper:

"Portable Inter-Workgroup Barrier Synchronisation for GPUs" appearing in OOPSLA'16.

Please cite the above paper if you use this code in academic work.

If you use our Pannotia or LonestarGPU code, consider citing the papers for those works as well, which are respectively:

"Pannotia: Understanding irregular GPGPU graph applications" appearing in IISWC'13

"A quantitative study of irregular programs on GPUs" appearing in IISW'12

Please contact Ally or Tyler if have questions or comments about this code:

Tyler Sorensen t.sorensen15 AT imperial DOT ac DOT uk

Alastair F. Donaldson alastair.donaldson AT imperial DOT ac DOT uk

Licenses and Copyright

This code consists of 3 different code bases, all with different licenses and copyright holders.

Discovery Protocol

The original code written for this project consists of the discovery protocol, the XF barrier (written using the discovery protocol and OpenCL 2.0 atomics) and the occupancy micro-benchmarks.

This code is located in code/discovery_protocol /code/experiments/occupancy_tests

The copyright holders of this code are: Tyler Sorensen and Alastair Donaldson.

The code is released under the BSD license which can be found in licenses/discovery_protocol

LonestarGPU OpenCL

The original LonestarGPU code is released under the University of Texas research license, which it maintains.

The LonestarGPU code is confined to the directory: code/experiments/lonestar_gpu_opencl

A copy of the license can be found in: licenses/lonestar_gpu

The original LonestarGPU code can be obtained from: http://iss.ices.utexas.edu/?p=projects/galois/lonestargpu

Pannotia

The Pannotia applications are provided with a license by AMD.

A copy of the license can be found in: licenses/pannotia

The Pannotia code is confined to the directory: code/experiments/pannotia

The original Pannotia code can be obtained from: https://github.com/pannotia/pannotia

Running the Code

Obtaining Inputs

To run the Pannotia or LonestarGPU applications, you will need to obtain the (large) graph files as input.

We provide instructions for obtaining these graphs in: code/experiments/pannotia/datasets/README.txt and code/experiments/lonestar_gpu_opencl/inputs/README.txt

We have provided google drive links containing these datasets.

Running

We believe the code is straightforward to run. It uses a cross-platform (Windows and Linux) CMake build system and the applications give examples of the command line arguments if run with no arguments.

Detailed instructions for running the code can be found in our OOPSLA artifact evaluation guide: OOPSLA_AE.txt. This guide was reviewed and accepted by the OOPSLA'16 artifact evaluation committee, so we believe it should be sufficient.

Known Issues

It was very difficult to get these applications running reliably across all the GPUs we tested. We encountered many interesting issues, and we even wrote a workshop paper about it that you should check out:

"The Hitchhiker's Guide to Cross-Platform OpenCL Application Development" appearing in IWOCL'16

Here are some issues we have found that are still present in this code:

Intel Compiler Crash

When running on Intel GPUs on Windows, the compiler seems to non-deterministically crash (which crashes the application).

The application we find that suffers from this problem the worse is the occupancy microbenchmarks

LonestarGPU DMR Application

We have only been able to get this application working reliably on Nvidia chips. And even on these chips, it crashes about 1% of the time (same as the original CUDA application). Other chips deadlock, crash, or produce wrong results 100% of the time.

The original DMR comes with a warning. We pass this warning remains valid (and perhaps even more so) in our OpenCL port.

Structs and Global Memory Pointers

Our LonestarGPU ports pass structures to kernels which contain global memory pointers. This is technically undefined behaviour in OpenCL and we know this to cause issues on ARM GPUs. However, it appears to work fine on Nvidia, Intel and AMD GPUs.

This issue is fixable but requires a significant code refactoring. Its on our todo list.

About

An implementation of a workgroup occupancy discovery protocol and an inter-workgroup barrier. Also example applications.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published