Add modular transform library with reusable named sequences

[erwei-xilinx]

Summary

• Add transform_library.mlir with 17 reusable transform.named_sequence definitions covering canonicalization, elementwise fusion, tiling, pad+promote, bufferization, post-bufferize cleanup, AIR herd mapping, and vector type casts
• Refactor all elementwise transform scripts (relu, sigmoid, silu, gelu, leaky_relu, swiglu, axpy, vec-add) to use transform.include — reducing each from 144-195 lines to 42-71 lines (~70% reduction)
• Partially refactor complex scripts (softmax, layernorm, matmul, rms_norm, weighted_rms_norm, average_pool, matvec) where boilerplate canonicalize/bufferize blocks match library sequences
• driver.py auto-expands transform.include calls at load time by inlining library sequence bodies with SSA renaming, since mlir-air's run_transform does not natively support transform.include
• Net: -1321 lines across 28 files, with no behavioral changes

Test plan

• Clear triton cache and run all 8 elementwise examples on NPU2 (AIE2P/Strix) — all pass
• Run softmax, layernorm, matmul on NPU2 — all pass
• CI build validation
• Run on NPU1 (AIE2/Phoenix) if available

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR introduces a shared MLIR Transform Dialect “library” and refactors many per-example transform scripts to reuse it via transform.include, with driver-side inlining to work around missing/unstable transform.include support in the underlying runtime.

Changes:

• Add amd_triton_npu/backend/transform_library.mlir with reusable transform.named_sequence blocks for canonicalization, tiling, pad+promote, bufferization, AIR herd mapping, and common vector type casts.
• Update amd_triton_npu/backend/driver.py to expand transform.include by inlining sequence bodies from transform_library.mlir.
• Refactor multiple example transform scripts to use transform.include (reducing duplicated boilerplate).

Reviewed changes

Copilot reviewed 27 out of 27 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
examples/weighted_rms_norm/transform_aie2p.mlir	Replace inline canonicalize/CSE + bufferize blocks with library transform.include calls.
examples/weighted_rms_norm/transform_aie2.mlir	Add AIE2 transform script; uses library bufferize include in one phase.
examples/vec-add/transform_aie2p.mlir	Rewrite vec-add transform to use library sequences and a small custom vector tiling step.
examples/vec-add/transform_aie2.mlir	Same refactor for AIE2 (different vector tile size).
examples/test_softmax/transform_aie2p.mlir	Replace repeated canonicalize/bufferize blocks with library includes.
examples/test_softmax/transform_aie2.mlir	Same include-based refactor for AIE2.
examples/test_layernorm/transform_aie2p.mlir	Replace repeated canonicalize/bufferize blocks with library includes; keep custom cleanup where needed.
examples/test_layernorm/transform_aie2.mlir	Same include-based refactor for AIE2.
examples/swiglu/transform_aie2p.mlir	Large reduction of boilerplate via library sequences + shared cast helper.
examples/swiglu/transform_aie2.mlir	Same refactor for AIE2, using extern-linking herd mapping helper.
examples/silu/transform_aie2p.mlir	Refactor to shared elementwise pipeline + sigmoid-family cast helper.
examples/silu/transform_aie2.mlir	Same refactor for AIE2, using extern-linking herd mapping helper.
examples/sigmoid/transform_aie2p.mlir	Refactor to shared elementwise pipeline + sigmoid-family cast helper.
examples/sigmoid/transform_aie2.mlir	Same refactor for AIE2, using extern-linking herd mapping helper.
examples/rms_norm/transform_aie2p.mlir	Replace one-shot bufferize block with @one_shot_bufferize include.
examples/relu/transform_aie2p.mlir	Refactor to shared elementwise pipeline + maxnumf cast helper.
examples/relu/transform_aie2.mlir	Same refactor for AIE2.
examples/matvec/transform_aie2p.mlir	Add GEMV transform script; uses includes for canonicalize+CSE and bufferize.
examples/matvec/transform_aie2.mlir	Add GEMV transform script (AIE2 variant); uses includes for canonicalize+CSE and bufferize.
examples/matmul/transform_aie2p.mlir	Replace repeated canonicalize/CSE + bufferize blocks with library includes.
examples/matmul/transform_aie2.mlir	Same include-based refactor for AIE2.
examples/leaky_relu/transform_aie2p.mlir	Refactor to shared elementwise pipeline + leaky-relu cast helper.
examples/leaky_relu/transform_aie2.mlir	Same refactor for AIE2.
examples/gelu/transform_aie2p.mlir	Refactor to shared elementwise pipeline + sigmoid-family cast helper.
examples/axpy/transform_aie2p.mlir	Refactor to shared binary pipeline + addf/mulf cast helpers.
examples/axpy/transform_aie2.mlir	Same refactor for AIE2.
examples/average_pool/transform_aie2p.mlir	Replace one-shot bufferize block with @one_shot_bufferize include.
examples/average_pool/transform_aie2.mlir	Same include replacement for AIE2.
amd_triton_npu/backend/transform_library.mlir	New shared sequence library for canonicalize/tiling/promotion/bufferize/AIR mapping/casts.
amd_triton_npu/backend/driver.py	Add include-expansion logic to inline library sequences into user transform scripts.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.