Add target-specific MmaOp/CopyOp to a FlyDSL backend dialect.
npx skills add https://github.com/ROCm/FlyDSL --skill add-target-atom-op
Please help me install this Agent Skill. Skill: add-target-atom-op Source: https://github.com/ROCm/FlyDSL/tree/main/.claude/skills/add-target-atom-op Command: npx skills add https://github.com/ROCm/FlyDSL --skill add-target-atom-op
This Skill provides a structured recipe to add a new target-specific MmaOp/CopyOp to a FlyDSL backend dialect, enabling support for new hardware backends and the associated lowering and wiring.
Follow the recipe to implement a new backend-specific MmaOp/CopyOp by declaring the payload types in the backend dialect, wiring up interface methods, adding lowering paths, and integrating build and Python bindings.
High-intent search queries and answers about installing and using this skill.
To add a new hardware MmaOp or CopyOp to a FlyDSL backend dialect, you declare backend payload types in TableGen, implement interface methods, wire lowering hooks, and integrate CMake to support custom intrinsics.
The two-level type design for extending FlyDSL backends uses generic payloads in the fly dialect, such as mma_atom, and backend-specific payloads in the backend dialect, such as rocdl cdna3 mfma, to separate abstract and target-specific operations.
Lowering for newly added backend operators in FlyDSL works by implementing Fly_MmaOpTypeInterface or Fly_CopyOpTypeInterface methods and configuring lowering passes like emitAtomCall or emitAtomCallSSA to translate operations into hardware intrinsics.
Yes, you need to validate ThrVal and ThrBit layouts when adding a custom operator to FlyDSL to ensure safe and correct operation of the new target-specific MmaOp or CopyOp during backend lowering.
The best way to wire build system and Python bindings for a new FlyDSL backend op is to follow the end-to-end roadmap, integrating CMake configurations and Python bindings alongside the TableGen declarations and backend code.
Yes, you can introduce optional SSA lowering paths for custom FlyDSL hardware intrinsics by implementing stateful interfaces and utilizing lowering paths like emitAtomCallSSA alongside the standard lowering passes.
ROCmAutomate building and installing FlyDSL from source on remote hosts or containers.
tenstorrentRegister TTIR composite op decomposition patterns for the TTMetal backend.
tenstorrentAutomate TTIR elementwise lowering to D2M for TTMetal in MLIR.
gmh5225Design MLIR dialects, transformation passes, and lowering pipelines for compilers.
