Add initial TileArray programming model and Neura compiler flow

[ShangkunLi]

Summary

This PR establishes the initial Synapse programming model and compiler flow for explicitly placed programs on a CGRA tile array.

It introduces a Python TileArray DSL, records typed tile-level operations, lowers them through the MLIR Python bindings into a single Taskflow task containing a Neura kernel, and invokes the Amoeba/Neura backend to produce mapped Neura IR.

Key Changes

• Add the initial Python frontend and AST parser.
• Add the user-facing TileArray spatial abstraction using Neura-compatible (x, y) coordinates.
• Add a typed TileArray program representation with initial i32 and f32 support.
• Add frontend operations for placed constants and additions.
• Lower a captured TileArray program into:

func.func
  taskflow.task
    neura.kernel
      placed Neura operations

• Generate Neura placement attributes directly from Tile coordinates.
• Add a compiler driver that runs:
  • --leverage-predicated-value
  • --insert-data-mov
  • template mapping in spatial-only mode
• Return the final mapped IR separately from backend diagnostic output.
• Replace the standalone Neura dependency with a pinned Amoeba submodule, with Neura managed by Amoeba.
• Add tests for frontend parsing, TileArray semantics, program recording, pre-mapping IR, and final mapped IR.
• Reorganize Python tests into frontend, language, and compiler directories.
• Expand the README with LLVM, Amoeba, Python binding, build, and test instructions.
• Add CI jobs for:
  • Python 3.10 and 3.11 frontend/language tests.
  • Python 3.11 compiler integration using a pinned LLVM revision and locally built Amoeba/Neura bindings.

Current Compiler Flow

Python TileArray program
  → TileArrayProgram
  → Taskflow + placed Neura IR
  → predicated Neura values
  → inserted data movement
  → template mapping
  → mapped tiles, links, registers, and time steps

Current Scope

This PR intentionally targets the first single-task milestone:

• one implicit Taskflow task;
• one Neura kernel;
• explicitly placed TileArray operations;
• constants and additions;
• i32 and f32 scalar values.

Task-level syntax, multiple tasks, task dependencies, inter-core communication, automatic template selection, and general structured lowering remain future work.

Testing

The focused local test suite passes:

9 passed

Run it with:

python -m pytest -q \
  tests/python/frontend \
  tests/python/language \
  tests/python/compiler

The compiler tests require the pinned LLVM/MLIR build and the Amoeba Python bindings described in the README.

[guosran]

Yeah I mean one constructor for each single operations.

________________________________ From: Shangkun LI ***@***.***> Sent: Tuesday, 25 August 2026 19:36:28 To: coredac/synapse ***@***.***> Cc: GUO Shiran ***@***.***>; Review requested ***@***.***> Subject: Re: [coredac/synapse] Add initial TileArray programming model and Neura compiler flow (PR #2) @ShangkunLi commented on this pull request.

________________________________ In python/synapse/frontend/lowering.py<#2 (comment)>:

@@ -103,6 +105,41 @@ def get_placement(tile: Tile) -> DictAttr:

} ) + @singledispatch + def lower_operation(operation: TileArrayOp, operands, result_type): + """Lower one frontend TileArray operation to a Neura operation. + + The caller handles common lowering such as resolving operands, + attaching placement, and recording the resulting SSA value. + """ + raise NotImplementedError( + f"unsupported tile-array operation: {type(operation).__name__}" + ) + + @lower_operation.register + def lower_constant(operation: ConstantOp, operands, result_type): But we may have some additional checking other than just the number of operands, so it would be better to implement operation-specific checking. — Reply to this email directly, view it on GitHub<#2?email_source=notifications&email_token=BHMYV7VBOT3HWINWHAI26N35LV23ZA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKMBRHA2DANZXGUZKM4TFMFZW63VQOJSXM2LFO5PXEZLROVSXG5DFMSSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#discussion_r3852553409>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BHMYV7SHJTIRJR2P3U2M7BD5LV23ZAVCNFSNUABGKJSXA33TNF2G64TZHMYTENZQGMZTGMJZGE5US43TOVSTWNJRHEZDSNBSHA3DHILWAI>. You are receiving this because your review was requested.Message ID: ***@***.***>

[guosran]

Got it

________________________________ From: Shangkun LI ***@***.***> Sent: Tuesday, 25 August 2026 19:42:36 To: coredac/synapse ***@***.***> Cc: GUO Shiran ***@***.***>; Review requested ***@***.***> Subject: Re: [coredac/synapse] Add initial TileArray programming model and Neura compiler flow (PR #2) @ShangkunLi commented on this pull request.

________________________________ In python/synapse/language/tile_array_program.py<#2 (comment)>:

-# This union explicitly lists every operation currently supported by the -# tile-array frontend. Future operations such as MacOp and GatherOp should be -# added here. -TileArrayOp: TypeAlias = ConstantOp | AddOp ***@***.***(frozen=True) +class AddOp(TileArrayOp): + """A scalar addition executed by a tile-array operation.""" But add and constant are used to define the API exposed to users. — Reply to this email directly, view it on GitHub<#2?email_source=notifications&email_token=BHMYV7RD7KA32BRL54OJ4HD5LV3SZA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKMBRHA2DKNRUGYZKM4TFMFZW63VQOJSXM2LFO5PXEZLROVSXG5DFMSSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#discussion_r3852596530>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BHMYV7VSC23KT3IWND7YHYT5LV3SZAVCNFSNUABGKJSXA33TNF2G64TZHMYTENZQGMZTGMJZGE5US43TOVSTWNJRHEZDSNBSHA3DHILWAI>. You are receiving this because your review was requested.Message ID: ***@***.***>