xla::CpuCopyInsertion Class Reference

#include <cpu_copy_insertion.h>

Inherits HloPassInterface.

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

Detailed Description

Google docs:

Besides the modifications made by the generic xla::CopyInsertion, this CPU-specific copy insertion pass also adds copies to values live out of computations satisfying certain conditions (defined by constant or parameter, etc). This is necessary because of deficiencies of buffer assignment. Specifically, buffer assignment is computation-scoped and does not recognized aliasing between arguments and outputs of computations.

TODO(b/62548313): Remove this when buffer assignment is smarter (module-scoped).

Member Function Documentation

Run()

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

override

Entry point of xla::CpuCopyInsertion

1. Call generic version of copy insertion(xla::CopyInsertion())
2. Run addtional copy insertion for CPU version backend(xla::CopyInsertion::AddCopiesForBufferAssignment()) due to deficiencies in buffer assignment

Here is the call graph for this function:

---

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

• tensorflow/tensorflow/compiler/xla/service/cpu/cpu_copy_insertion.h
• tensorflow/tensorflow/compiler/xla/service/cpu/cpu_copy_insertion.cc