Beyond the Vector API - A Quest for a Lower Level API
summary
Proposal of a hardware intrinsics API exposes one-to-one machine instructions from Java via Project Panama, jextract and generated native libraries, aiming for raw-instruction performance; experiments use machine-code snippets and native entry stubs then let the JVM inline or patch encodings to remove call overhead. Example: an x86 VP DQ pattern runs ~3x faster (5 vs 15 cycles) when mapped to the hardware. Major problems: tens of thousands of opcodes, register-move/code-density inflation, toolchain miscompiles and the need for VM support for vector calling conventions.
AI-written summary · about
chapters (10)
- 0:00Vector API journey
- 1:49Vector API goals and evolution
- 7:52Vector API adoption lessons
- 13:17Scalar to vector bridging
- 17:43Hardware intrinsics thought experiment
- 21:44Hardware intrinsics API design
- 29:44Hardware intrinsics with Jextract
- 36:43Optimizing native calls
- 41:07Hardware intrinsics challenges
- 47:58Q&A
