Modular GPU Programming with Typed Perspectives
summary
Prism is a low-level GPU language that tracks thread32 perspective and other granularities in types, exposes constructs like split and group and partition (plus a claim for async memory sharding), and verifies at call sites that collective ops (tensor cores, warp syncs) are only invoked with the required perspective. A formal core, the Bundl calculus, proves this soundness. Implementations include prefix scan and warp-specialized matmul; benchmarks match speed-of-light kernels with a reported 10% difference tied to a compiler pass mismatch.
AI-written summary · about
chapters (7)
- 0:00Modular GPU programming intro
- 3:52CUDA collaboration challenges
- 6:17Prism language and tracking
- 8:21Enforcing Prism contracts
- 13:19Memory hierarchy in Prism
- 14:44Prism performance validation
- 15:46Q&A
related talks
15mSpecifying and Testing GPU Workgroup Progress Models
25mA Modular Static Cost Analysis for GPU Warp-Level Parallelism
49mQuantitative program reasoning in Granule via graded modal types
10mLifting PT to New Heights with MAX and Mojo
30mGenerating High Performance Code for Irregular Data Structures using Dependent Types
30mData-Parallel Flattening by Expansion
22mA Vision of Miking: Interactive Programmatic Modeling, Language Composition, and Compilation
28m"Making a Programming Language for 3D Design"