The authors implemented RESOLVE, which operates on compiled GPU binaries (language-agnostic) using NVBit for binary instrumentation.
Three-step hybrid pipeline validates GPU kernels by combining concurrency testing with formal proof
RESOLVE uses binary instrumentation to expose races and then proves functional equivalence on reduced, verifiable kernels.
Big Tech
Ashkan Vedadi Gargary · Guido Martínez · Sebastian Burckhardt · Gabriel Ebner · Abhinav Jangda · Madan Musuvathi · +1 more
University of California, Riverside · Math, Inc. · Microsoft Research
Research Digest··3 min read
The authors present RESOLVE, a GPU-kernel validation framework that separates concurrency correctness from functional correctness.
Why this paper
From Microsoft Research and 2 others
In one line
RESOLVE validates GPU kernels by testing for nondeterminism, reducing concurrency, and proving functional equivalence.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§