xla::UserComputation Class Reference

#include <user_computation.h>

Public Member Functions
StatusOr< ComputationDataHandle >	AddParameterInstruction (const ParameterRequest &parameter_request)
	Enqueue a parameter-retrieving instruction onto this UserComputation. More...
VersionedComputationHandle	GetVersionedHandle () const
	Lock mutex and call GetVersionedHandleInternal() More...
StatusOr< std::unique_ptr< HloComputation > >	BuildHloComputation (VersionedComputationHandle::Version version, HloComputationResolver hlo_resolver, const DebugOptions &debug_options, bool include_unreachable_instructions=true) const
	Build a HLO computation from the UserComputation More...

Private Member Functions
VersionedComputationHandle	GetVersionedHandleInternal () const EXCLUSIVE_LOCKS_REQUIRED(mutex_)
	Get the VersionedComputationHandle recorded in UserComputation object. More...

Detailed Description

Google Docs:

A UserComputation is the built-up computation that users create via the XLA Service interface.

The XLA service adds instructions to a user computation via this interface. The state of the computation is stored as a SessionComputation proto which holds a record of all operation-building requests received by the XLA service.

UserComputations are lowered to HloComputations which are passed to the high level compiler interface.

Member Function Documentation

AddParameterInstruction()

StatusOr< ComputationDataHandle > xla::UserComputation::AddParameterInstruction

(

const ParameterRequest &

parameter_request

)

Enqueue a parameter-retrieving instruction onto this UserComputation.

1. Check whether this parameter has been registered by member parameter_requests (map<parameter_number, OperationRequest>).
2. Create a new ComputationDataHandle object which will use the next_handle_value_ as value.
3. Register an OperationRequest object in member session_computation_'s requests map by using the value of ComputationDataHandle as a key.
4. Register the ComputationDataHandle object created in step 2 in the map mentioned in step 1.
5. Return the ComputationDataHandle which represents the index of OperationRequest object.

Here is the call graph for this function:

Here is the caller graph for this function:

BuildHloComputation()

StatusOr< std::unique_ptr< HloComputation > > xla::UserComputation::BuildHloComputation	(	VersionedComputationHandle::Version	version,
		HloComputationResolver	hlo_resolver,
		const DebugOptions &	debug_options,
		bool	include_unreachable_instructions = true
	)		const

Build a HLO computation from the UserComputation

1. Call xla::anonymous_namespace{user_computation.cc}::ComputationLowerer::Lower() to build a HloComputation.

• Implementation

Here is the caller graph for this function:

GetVersionedHandle()

VersionedComputationHandle xla::UserComputation::GetVersionedHandle

(

)

const

Lock mutex and call GetVersionedHandleInternal()

• VersionedComputationHandle
  • Contains
    • ComputationHandle which represents an UserComputation
    • Version (int64) which represents an specific ComputationDataHandle which represents an OpRequest recorded in session_computation_

Here is the call graph for this function:

Here is the caller graph for this function:

GetVersionedHandleInternal()

VersionedComputationHandle xla::UserComputation::GetVersionedHandleInternal ( ) const

private

Get the VersionedComputationHandle recorded in UserComputation object.

• Check the member handle_to_return_
  • Default to be 0 which means no specific ComputationDataHandle has been assigned. Therefore this function will return the last enqueued OpRequest's ComputationDataHandle which is usually the output node of a graph.
  • If it's non-zero, just return the specific ComputationDataHandle.

Here is the caller graph for this function:

---

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

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