Skip to content

Dialect op wrappers (aie.dialects.*)

This is the low-level Python layer: thin wrappers around the individual MLIR operations of the aie, aiex, and aievec dialects. These are what the high-level aie.iron objects lower to — when a Worker, ObjectFifo, or Runtime resolves, it emits ops from exactly these modules.

Most designs never import these directly. Reach for them when you are writing a design op-by-op, building a custom lowering, or reading MLIR that the JIT emitted and want to know which Python call produced a given op.

How the wrappers are generated

Each op wrapper corresponds one-to-one with an operation defined in the dialect's TableGen (*.td) files. The bindings come in two forms:

  • Generated bindings (_aie_ops_gen, _aiex_ops_gen, _aievec_ops_gen) are produced at build time by mlir-tblgen directly from the op definitions. There is one class per op, named after the op (e.g. ObjectFifoCreateOp, NpuDmaMemcpyNdOp).
  • Hand-written conveniences in aie.py / aiex.py wrap those generated classes with friendlier constructors and Python-side helpers — for example object_fifo (wrapping ObjectFifoCreateOp), core (wrapping CoreOp), buffer (wrapping BufferOp), and runtime_sequence.

Because the generated *_ops_gen modules only exist inside a built tree, this page is a conceptual guide rather than an auto-generated symbol dump. For the authoritative, op-by-op reference, use the dialect pages under MLIR / C++ (below).

The three dialects

Module Dialect What it wraps
aie.dialects.aie aie Core structural ops: tiles, buffers, locks, cores, ObjectFifos, flows, and packet flows.
aie.dialects.aiex aiex Extended / runtime ops: NPU DMA memcpy, sync, RTP writes, DMA task configuration, and the runtime_sequence.
aie.dialects.aievec aievec AIE vector ops used by the vectorizing compiler passes. 
# Op-by-op construction with the wrappers (rarely needed directly):
from aie.dialects.aie import tile, core, object_fifo, end
from aie.dialects.aiex import runtime_sequence, npu_dma_memcpy_nd

Reference

The op wrappers are a Python surface over the dialects. For the definitive per-op documentation — operands, results, attributes, and assembly format — see the dialect and pass references in the MLIR / C++ group of this section: