Warps and Atomics: Beyond Barrier Synchronization in the Verification of GPU Kernels

The paper

• Source
  • GPUVerify source
  • Bugle source
• Binaries
  • Linux x86-64
  • Mac OSX (Mavericks)
• Benchmark set

AWESOME NEWS!

We got confirmation the ParBoil bugs (mentioned in §5, p13) are genuine bugs!

How to Build

1. Download the above sources
2. Build as per for regular GPUVerify (http://multicore.doc.ic.ac.uk/tools/GPUVerify/)

How to Use

Intra-warp synchronization

There are 3 additional flags for warps:

• --warp-sync=32 uses the re-sync method with a warp size of 32
• --warp-sync=32 --only-warp runs the intra warp pass of the two-pass method with a warp size of 32
• --warp-sync=32 --no-warp runs the inter-warp pass of the two-pass method with a warp size of 32

Atomics

Atomic support, with the refined atomic abstraction, is enabled by default. To turn of the refined abstraction and force adversarial abstraction, use --no-refined-atomics

Contact

Ethel Bardsley and Alastair Donaldson