update codegen and codegen test for diagonal direction

[n0thingNoob]

King Mesh Diagonal Codegen Update

Background

GenerateCodePass previously handled only 4-neighbor directions:

• EAST
• WEST
• NORTH
• SOUTH

When running on a king_mesh topology (8-neighbor connectivity), diagonal links could be resolved as LOCAL or mapped incorrectly in generated outputs. This caused mismatches between mapping/link intent and emitted instruction operands in:

• tmp-generated-instructions.yaml
• tmp-generated-instructions.asm

What Was Changed

1) Direction support in codegen

File updated:

• dataflow/lib/NeuraDialect/Transforms/GenerateCodePass.cpp

Changes:

• Extended Topology::getDirBetween(...) to support diagonal directions:
  • (1, 1) -> NORTHEAST
  • (-1, 1) -> NORTHWEST
  • (1, -1) -> SOUTHEAST
  • (-1, -1) -> SOUTHWEST
• Extended Topology::invertDir(...) accordingly:
  • NORTHEAST <-> SOUTHWEST
  • NORTHWEST <-> SOUTHEAST
• Kept LOCAL semantics unchanged for true local/no-link scenarios.

2) Dedicated king-mesh architecture spec for tests

New file:

• dataflow/test/arch_spec/architecture_king_mesh.yaml

This file is copied from the regular test architecture and sets:

• multi_cgra_defaults.base_topology: "king_mesh"
• per_cgra_defaults.base_topology: "king_mesh"

3) New codegen test

New file:

• dataflow/test/code_gen/test_code_generate_king_mesh.mlir

Test behavior:

• Runs the full codegen pipeline using architecture_king_mesh.yaml
• Checks full mapping IR expectation (MAPPING)
• Checks partial YAML/ASM expectation focused on diagonal path evidence
• Uses --implicit-check-not=LOCAL for YAML/ASM to catch accidental fallback

Verification Summary

The new king-mesh codegen test passes:

• llvm-lit -v dataflow/test/code_gen/test_code_generate_king_mesh.mlir

Manual inspection of generated artifacts confirms diagonal operands are emitted as expected (e.g., NORTHEAST, NORTHWEST, SOUTHEAST, SOUTHWEST) instead of incorrect LOCAL.

[Copilot]

Pull request overview

Updates Neura’s codegen to correctly resolve diagonal link directions when targeting a king-mesh (8-neighbor) topology, and adds a dedicated lit test + architecture spec to validate diagonal operands in generated YAML/ASM outputs.

Changes:

• Extend Topology::getDirBetween / invertDir to handle NORTHEAST/NORTHWEST/SOUTHEAST/SOUTHWEST.
• Add a king-mesh test architecture YAML for lit-based codegen testing.
• Add a new codegen lit test that checks diagonal direction evidence in emitted YAML/ASM.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
lib/NeuraDialect/Transforms/GenerateCodePass.cpp	Adds diagonal direction resolution/inversion for king-mesh link operands.
test/arch_spec/architecture_king_mesh.yaml	Introduces a king-mesh topology architecture spec for tests.
test/code_gen/test_code_generate_king_mesh.mlir	Adds a new lit test ensuring diagonals are emitted (and not incorrectly treated as LOCAL).

---

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

You can also share your feedback on Copilot code review. Take the survey.

[tancheng]

BTW, why do we need king-mesh? I thought our evaluations are only running on mesh.

[n0thingNoob]

BTW, why do we need king-mesh? I thought our evaluations are only running on mesh.

I am currently working on a case study and found that mapping is quite difficult with a mesh-only topology. @ShangkunLi mentioned that it works on a king mesh, but I noticed that the generated ASM and YAML were missing the proper directions for diagonal connections.

[tancheng]

BTW, why do we need king-mesh? I thought our evaluations are only running on mesh.

I am currently working on a case study and found that mapping is quite difficult with a mesh-only topology. @ShangkunLi mentioned that it works on a king mesh, but I noticed that the generated ASM and YAML were missing the proper directions for diagonal connections.

For which kernel? I thought we can now map most kernels you are working on.

[ShangkunLi]

BTW, why do we need king-mesh? I thought our evaluations are only running on mesh.

I am currently working on a case study and found that mapping is quite difficult with a mesh-only topology. @ShangkunLi mentioned that it works on a king mesh, but I noticed that the generated ASM and YAML were missing the proper directions for diagonal connections.

For which kernel? I thought we can now map most kernels you are working on.

@n0thingNoob introduces a synthetic function composed of two gemv kernels and a relu kernel for a case study, which can only be mapped on a 4x4 king-mesh/an 8x8 mesh architecture.

Considering we also need to support codegen for king-mesh architectures, we decided to make this pr.

[tancheng]

The 2 gemm + 1 relu forming a single large DFG (how many nodes) for mapping? Or viewed as 3 tasks, temporarily scheduled?

[ShangkunLi]

The 2 gemm + 1 relu forming a single large DFG (how many nodes) for mapping? Or viewed as 3 tasks, temporarily scheduled?

A single large DFG for mapping, though with bad performance (i.e., compiled II = 28).