Skip to content

t27: data-parallel kernels for GPU as pure array programs plus a separate schedule (research spike) #9005

Open

Description

@gHashTag

Part of the effects-at-the-edge epic. Research basis: Halide (algorithm and schedule separated), Futhark (data-parallel functional arrays, compiled to GPU) and Triton (tiled kernels).

Design:

  • Express a GPU kernel in t27 as a pure map/reduce/scan over arrays. The core semantics stay deterministic and testable on the CPU under the reference and t27b.
  • A separate schedule (tiling, parallel axes) drives the backends: Verilog today; CUDA, Metal or WGSL later.
  • t27o maps simple CUDA/WGSL kernels into this form, and refuses anything else with a reason.

Acceptance: a design note, plus one kernel (a dot product or matmul tile) written as a pure array program in .t27 that passes on the CPU, with a schedule stub.

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions