Skip to content

mlir-lowering.md

Latest commit

 

History

History
66 lines (47 loc) · 2.87 KB

mlir-lowering.md

File metadata and controls

66 lines (47 loc) · 2.87 KB

MLIR Lowering (Normative)

This chapter specifies the deterministic lowering of canonical Core IR into MLIR for the feature-gate that ships with the public compiler. The rules cover only the stable, publicly implemented subset.

Scope and prerequisites

  • Lowering is feature-gated (e.g. via an mlir-lowering feature flag in the compiler).
  • The input module MUST be verified and canonicalised per Core IR.
  • Lowering produces deterministic MLIR textual IR suitable for snapshot testing. Minor changes may occur between MIND versions but are stable within a release.
  • Core v1 MLIR lowering rules target the CPU backend. Any GPU-specific MLIR dialects or pipelines are experimental and outside the scope of this chapter (see Runtime).

Lowering patterns

Constants

  • Scalars and tensors lower to arith.constant with explicit MLIR tensor types.

Tensor allocation and fill

  • Zero-initialised tensors are represented using tensor.empty followed by linalg.fill fed by a separate arith.constant for the fill value.

Matrix multiplication

  • MatMul lowers to linalg.matmul in destination-passing style:
    • A tensor.empty destination is created with the inferred result shape.
    • The destination is passed via outs and the filled tensor is the operation result.

Convolution

  • Conv2d lowers to linalg.conv_2d_nhwc_hwcf with NHWC input and HWCF filter.
  • Destination-passing mirrors matmul: allocate with tensor.empty, feed through outs.
  • Verification ensures input and filter channels match before lowering.

Elementwise and reductions

  • Elementwise BinOp instructions lower to MLIR arithmetic ops on broadcasted tensors using MLIR canonical broadcasting utilities.
  • Sum and Mean lower to reduction patterns that preserve keepdims semantics; Mean divides by the reduced element count explicitly.

Determinism and stability

  • Given canonical IR, the emitted MLIR text is deterministic. Ordering follows IR instruction order and canonical operand ordering rules.
  • The textual form is stable within a compiler release. Future releases MAY evolve op selections or attributes but MUST preserve semantics for the defined Core v1 operations.