Beehive SPIR-V Toolkit: A Composable and Functional API for Runtime SPIR-V Code G...
summary
A toolkit generates a Java API from the SPIR‑V JSON grammar and emits/reads binary modules via an assembler/disassembler; it integrates into TornadoVM to replace OpenCL C codegen and the LLVM subprocess. The generated library is automatically generated Java library, the integrated pipeline compiles kernels about 3x faster than TornadoVM’s OpenCL C JIT and runs up to 1.52x faster on some workloads, and validation uses SPIR‑V tools plus unit tests so generated modules are accepted by drivers.
AI-written summary · about
chapters (17)
- 0:00Beehive SPIR-V Toolkit intro
- 1:05Motivation for SPIR-V toolkit
- 2:27SPIR-V binary challenges
- 3:33OpenCL and SPIR-V basics
- 4:58Gaps in SPIR-V access
- 6:00Beehive SPIR-V overview
- 7:05Automatic toolkit generation
- 9:56Library structure
- 11:42SPIR-V module examples
- 13:29Tornado VM integration
- 15:16JIT compilation in Tornado VM
- 17:04Code generation performance
- 19:14Runtime SPIR-V speedups
- 21:21Summary
- 22:04Q&A: Graph axes
- 23:07Q&A: Benchmark regression
- 24:32Q&A: Formal spec
related talks
36mAccelerate RISC-V Instruction Set Simulation by Tiered JIT Compilation
38mGraal
11mSemantics Lifting for Scientific Kernels
11mLightning Talk: Torch-Spyre: Compiling To a Multi-core Dataflow Accelerator...
14mVerification of Algorithm and Code Generation for Signal Transforms- 14mVerification of Algorithm and Code Generation for Signal Transforms
11mUsing Pyrope To Create Transformable RISC V Architectures
22mRISC-V Perf-Model: An Open Source Cycle Accurate Performance Mo...