xla::DotDecomposer Class Reference

#include <dot_decomposer.h>

Inherits HloPassInterface.

Public Member Functions
StatusOr< bool >	Run (HloModule *module) override
	Entry point of xla::DotDecomposer More...

Private Attributes
bool	decompose_batch_dot_

Detailed Description

Google docs:

DotDecomposer is a pass which decomposes batch Dot operations into a sequence of smaller (R2) Dot operations.

Member Function Documentation

Run()

StatusOr< bool > xla::DotDecomposer::Run ( HloModule *  module )

override

Entry point of xla::DotDecomposer

Run DotDecomposer pass on computations in 'module'. Returns whether the 'module' was changed.

Gather all batch Dot operations.

1. For All non-fusion computations in hlo module
   1. For all the instructions in the computation
      1. Ignore all HloOpcode::kDot instructions
      2. Store it into batch_dots if xla::DotDecomposer::decompose_batch_dot_ is set
2. xla::anonymous_namespace{dot_decomposer.cc}::DecomposeBatchDot to decompose

Here is the call graph for this function:

Member Data Documentation

decompose_batch_dot_

bool xla::DotDecomposer::decompose_batch_dot_

private

Indicate this instruction should be processed by xla::DotDecomposer

---

The documentation for this class was generated from the following files:

• tensorflow/tensorflow/compiler/xla/service/dot_decomposer.h
• tensorflow/tensorflow/compiler/xla/service/dot_decomposer.cc