tf_1.8_xla_doc
Classes | Public Member Functions | Static Public Member Functions | List of all members
xla::CompileOnlyService Class Reference

#include <compile_only_service.h>

Inheritance diagram for xla::CompileOnlyService:
[legend]
Collaboration diagram for xla::CompileOnlyService:
[legend]

Classes

struct  AotComputationInstance
 A description of a computation to compile using CompileAheadOfTime. More...
 

Public Member Functions

StatusOr< std::vector< std::unique_ptr< AotCompilationResult > > > CompileAheadOfTime (const tensorflow::gtl::ArraySlice< AotComputationInstance > computations, const AotCompilationOptions &Options)
 
- Public Member Functions inherited from xla::Service
tensorflow::Status Computation (const ComputationRequest *arg, ComputationResponse *result) override
 Populates computation_ with a valid object. Used before any given operation is enqueued. More...
 
tensorflow::Status Op (const OpRequest *arg, OpResponse *result) override
 Enqueue an Op on the computation. More...
 

Static Public Member Functions

static StatusOr< std::unique_ptr< CompileOnlyService > > NewService (const ServiceOptions &options)
 

Detailed Description

Google Docs:

An XLA Service specialization for ahead-of-time compilation. This only instantiates a Compiler object for the relevant platform; it does not instantiate or require an execution backend.

Member Function Documentation

◆ CompileAheadOfTime()

StatusOr< std::vector< std::unique_ptr< AotCompilationResult > > > xla::CompileOnlyService::CompileAheadOfTime ( const tensorflow::gtl::ArraySlice< AotComputationInstance computations,
const AotCompilationOptions &  options 
)

Google Docs:

Compiles a list of computations for ahead-of-time execution. This is intended for use in static compilation.

  1. Create xla::HloModule
    1. Get the UserComputation object and get the xla::VersionedComputationHandle through it (Check xla::UserComputation::GetVersionedHandle()).
    2. Compute the ProgramShape of VersionedComputationHandle and create config of module
    3. Call xla::ComputationTracker::BuildHloModule()
    4. Push HloModule object in member hlo_modules
  2. Call xla::cpu::CpuCompiler::CompileAheadOfTime(Overridden, not shown in call graph)
Here is the call graph for this function:
Here is the caller graph for this function:

◆ NewService()

StatusOr< std::unique_ptr< CompileOnlyService > > xla::CompileOnlyService::NewService ( const ServiceOptions &  options)
static
  1. Get pointer to the compiler singleton for the assigned platform.
  2. Create a compile-only XLA service.

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