Verifying GPU kernels by test amplification

Abstract

We present a novel technique for verifying properties of data parallel GPU programs via test amplification. The key insight behind our work is that we can use the technique of static information flow to amplify the result of a single test execution over the set of all inputs and interleavings that affect the property being verified. We empirically demonstrate the effectiveness of test amplification for verifying race-freedom and determinism over a large number of standard GPU kernels, by showing that the result of verifying a single dynamic execution can be amplified over the massive space of possible data inputs and thread interleavings.

Interleaving and Lock-Step Semantics for Analysis and Verification of GPU Kernels

“…Interleaving semantics for GPU kernels has been defined by [15,17,12]. These are similar to our semantics except that [15,12] do not give a semantics for barriers.…”

“…Interleaving semantics for GPU kernels has been defined by [15,17,12]. These are similar to our semantics except that [15,12] do not give a semantics for barriers. Contrary to our lock-step approach, [15,17] battle the state space explosion due to arbitrary interleavings of threads by considering one particular schedule.…”

“…Substantial effort has been put into the design of tools for rigorous analysis of GPU kernels [7,16,8,17,15]. In prior work [7], we presented a verification technique and tool, GPUVerify, for analysis of data races and barrier divergence in OpenCL [13] and CUDA [21] kernels.…”

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

“…Several recent works have focused on GPU kernel verification using SMT solving [13,3], combined static and dynamic analysis [12] and separation logic with permissions [8]. The closest work to GPUVerify is the PUG technique and tool [13], and the methods have been compared qualitatively and experimentally [3].…”

“…Data races lead to non-determinism, incorrect computation and undefined behavior. There has been recent interest in the program analysis community on methods for formal or semi-formal analysis of GPU kernels, leading to methods for finding bugs in [14,5] or proving correctness properties of [13,3,8,12] GPU kernels, principally focused on data races.…”