tvm: support external microkernels

[guillon]

Motivation

Allow TVM schedules to delegate an inner computation tile to a user-provided C-ABI microkernel.

Description

Add the external_at schedule operation and lower externalized loop tiles to T.call_extern.

The external ABI passes output and input pointers, inner-loop extents, and projected strides—including zero strides for invariant dimensions. Reduction initialization is separated from externally implemented updates.

Add support for compiling and linking additional C sources into TVM shared-library and archive outputs, or forwarding them with C-source outputs.

Add executable matmul and Conv2D examples covering spatial tiles, strip-mined reductions, multidimensional reductions, OpenMP SIMD, and aligned AVX-512 microkernels.

Commits

• tvm: implement external_at for tvm schedule

Discussion

Current limitations include one output, perfectly nested externalized loops.

TODO:

• Implement external_at for the MLIR backend.
• Consider a memref-compatible external ABI.

Testing

• External matmul and Conv2D FileCheck tests
• pytest -q tests/pytest/unit — 53 passed
• Pyright and Mypy checks

[guillon]

@qaco for information, may be useful if you want to call hand-optimized kernel tiles (or generated with another tool), xtc can mange the external memory tiling and call a C-abi compatible fast L1 resident tile for instance for matmul or conv2d.

This is actually the long waited tentative implementation of the external_call/tensorize.

I would appreciate a first feedback. Refer to the examples in tests/.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[guillon]

Merging pull request, you may comment later @qaco