Re: XLA GPU Code Generation using MLIR

[Mehdi AMINI]

+xla...@googlegroups.com : cross-posting from ml...@tensorflow.org as this is quite relevant!

On Tue, Jun 23, 2020 at 8:23 PM Mehdi AMINI <joke...@gmail.com> wrote:

Hi all,

Here is some information we just published about one of the subproject involving MLIR in the TensorFlow land: https://www.tensorflow.org/mlir/xla_gpu_codegen

This document describes the high-level plan to integrate the MLIR codegen (using Linalg/Affine/GPU dialect/...) with XLA on the short term. We are starting this transition by hooking MLIR starting at the LHLO dialect level: this dialect models "HLO operating on buffers", instead of immutable tensor values. We convert the entire XLA module to an LHLO module after XLA completes the buffer assignment. The main entry point at the moment for this HLO dialect -> LHLO dialect conversion by involving XLA is https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/xla/transforms/xla_hlo_to_lhlo_with_xla.cc ; note that it is incomplete and actively under development, mainly by Tim (CC).

This is just one subproject that we documented here, hopefully we'll make progress on documenting other aspects of MLIR in TensorFlow. In particular we're also having some pieces to go from HLO->LHLO directly without using XLA buffer-assignment here: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/xla/transforms/hlo_legalize_to_lhlo.cc

Cheers,

-- 

Mehdi