Skip to content

Latest commit

 

History

History
44 lines (38 loc) · 2.35 KB

INVOCATION.md

File metadata and controls

44 lines (38 loc) · 2.35 KB
Raw

mlirlus - command line options

Compilation produces the command line tool mlirlus. It takes one input file and prints its output on the standard output.

The command : ./mlirlus file.lus just parses file.lus, verifies its syntactic correction and prints it unchanged.

mlirlus takes several options, which activate the various Lustre-specific compilation steps described in this paper:

  • --normalize performes the lus dialect normalization step. It makes all fby and pre operations work on the base clock (at every cycle) and then moves their representation in the node signature, as loop-carried dependences. This ensures the respect of dominance rules.
  • --inline-nodes. When requested using the inline keyword in the program text, this step inlines instantiated nodes to avoid creating a new thread in the implementation.
  • --convert-lus-to-sync converts lus operations to a combination of operations of the dialects sync and scf (structured control flow).
  • --convert-sync-to-std converts sync operations to std operations, including calls to the run-time functions implementing the reactive execution machine.

mlirprime - command line options

Recall that mlirprime is an auxiliary tool which exposes a number of MLIR code transformations (and corrects some bugs).

The tool takes an input file and prints its output on the standard output. The command : ./mlir-prime file.prime just parses file.prime, verify its correction and prints it.

Several options can be composed with the builtin options of MLIR :

  • --bufferize-linalg-reshape corrects a limitation of the bufferization passes on the version of MLIR we used for the use cases (this limitation is corrected in the current MLIR version).
  • --remove-copy-prime removes linalg.copy operations on which we had problems for lowering.
  • --prime-linalg-to-affine bypasses the builtin lowering of the dialect linalg to the dialect affine using packing, tiling... These optimisations are machine-dependent and correspond to our benchmark architecture.
  • --loop-permutation-prime extends the built-in loop-permutation option which just operates on the first loop of the program. We are currently submitting this extension to the MLIR project.