Skip to content

Emit structured kernel metadata alongside PTX #71

Description

@tetsuo-cpp

Summary

When compiling, emit a machine-readable description of the kernel interface (name, parameters, shared memory, element types/sizes) next to or embedded with the PTX.

Motivation

Hosts currently re-parse Forth headers or hard-code layouts. That is brittle for:

  • Python launch APIs
  • Multi-kernel projects
  • Tooling that binds tensors by name

Metadata should be a compile artifact, not a second parser of source.

Proposed artifact

kernel.ptx + kernel.json (or a single archive / PTX comment section):

{
  "kernel": "attention",
  "arch": "sm_80",
  "params": [
    {"name": "Q", "kind": "buffer", "element": "f32", "count": 32768},
    {"name": "SEQ_LEN", "kind": "scalar", "type": "i64"}
  ],
  "shared": [
    {"name": "SCORES", "element": "f64", "count": 1024}
  ]
}

CLI

warpforthc attention.forth -o attention.ptx --emit-metadata attention.json
# or default: write attention.json next to -o

Acceptance criteria

  • Metadata covers kernel name, params (scalar vs buffer), types, static sizes
  • Stable schema version field
  • Python package / runner can launch using metadata without re-reading .forth
  • LIT or unit test that schema matches a fixture kernel
  • Document schema in docs/

Related

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions