package xrt

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: XLAComputationConfig

• repeated DeviceAssignment.ComputationDevice computation_devices = 1

  As many ComputationDevice as many there are computations (number of cores per replica).

Used in: DeviceAssignment

• repeated ComputationDevice.DeviceMeshCoordinates replica_devices = 1

  As many replicas as there are in the replicated computation.

Used in: ComputationDevice

• repeated int32 value = 1

  The mesh coordinates for the device. Usually (X, Y, Core), in the order in which they are returned in the TopologyProto. X = value(0) Y = value(1) Core = value(2)

Literal to allocate space for, and transfer to, device memory.

• optional xla.LiteralProto value = 2

Options and XLA computation for a compilation.

• optional XLAComputationConfig config = 1

• optional xla.HloSnapshot hlo_snapshot = 2

Options for an XLA compilation.

Used in: XLAComputation

• int32 num_replicas = 1

  The number of replicas the computation will be run on. If this is default (0) it is interpreted as 1.

• int32 num_cores_per_replica = 2

  The number of "model-parallel" cores per replica. If this is default (0) it is interpreted as 1.

• optional tensorflow.tf2xla.HostComputeMetadata host_compute_metadata = 3

  Optional metadata about host sends and recvs.

• optional xla.ProgramShapeProto program_shape = 4

  The arg/result shapes for the whole computation.

• repeated xla.ProgramShapeProto per_core_program_shape = 5

  The arg/result shapes for each core of a model-parallel computation. per_core_args_and_result_shapes is optional for a single-core computation.

• optional DeviceAssignment device_assignment = 6

  Describes how replicated computation instances should be assigned to devices. There are num_cores_per_replica computations, and each one will be sent and executed to the set of replica device numbers described in the DeviceAssignment proto.

• optional xla.DebugOptions debug_options = 7

  The debugging options to be passed to the XLA compilation process.

Node in a tree describing a tuple constructed from input handles. A node is an internal node if tuples is non-empty, in which case input_index and release_input_handle are ignored. Otherwise a node is a leaf node. Each leaf XLATupleNode is the index of an input which corresponds to a handle that will be grafted onto the output tuple at that location. If release_input_handle is true that input handle will be released and become invalid. Inputs may be repeated in which case leaves of the output tuple will alias. If an input is repeated, release_input_handle must be false for every leaf where that input appears. For example, if input 0 has shape {} and input 1 has shape {2,3} then the XLATupleNode with structure {1,{0,1}} corresponds to a tuple with shape {{2,3},{{},{2,3}}}.

• int32 input_index = 1

• bool release_input_handle = 2

• repeated XLATupleNode tuples = 3

• uint32 rng_seed = 1

  If non-zero, rng_seed to reset the core with.

• int32 core_index_in_replica = 2

  Which model-parallel computation to run from the compiled bundle.

• string execution_instance_key = 3

  Optional key to disambiguate between executions. This is only needed if multiple host send/recvs may be outstanding concurrently with executions.

A single chained execute operation. An operation can either be a device data load, or an existing (as in, previously compiled and accessible via its int64 handle) XLA computation execution.

Used in: XRTChainedExecutePlan

• oneof op_oneof

  • int64 data_handle = 1

    The handle to an existing XRT device data.

  • int64 computation_handle = 2

    The handle to an existing XRT compiled computation.

• repeated XRTChainedExecuteOp.Output outputs = 3

  The outputs of this XRTChainedExecuteOp operation.

• repeated XRTChainedExecuteOp.Input inputs = 4

  The inputs of this XRTChainedExecuteOp operation. If data_handle is set, there are no inputs.

Represents an input for this operation.

Used in: XRTChainedExecuteOp

• int64 op_index = 1

  The index within the XRTChainedExecutePlan.ops post-order of the source operation for this input.

• int64 output_index = 2

  The output index of the value generated by the operation at op_index. Zero (default value) means no index ({}) while if an indexing is required, output_index needs to be set to index+1. Thanks proto3!

Represents an output of the XRTChainedExecute operation, which should originate by the output of this operation.

Used in: XRTChainedExecuteOp

• int64 output_index = 1

  The index in the value generated by this operation, which should be forwarded as XRTChainedExecute output. If output_index is zero (default value) the whole output will be used as result. This means that if the output shape is a tuple, the result will be the full tuple. Otherwise the real sub-tuple index will be output_index - 1.

• int64 result_index = 2

  The index in the vector of the results returned by the XRTChainedExecute operation, where this output should be forwarded.

Execution plan for the XRTChainedExecute operation.

• repeated XRTChainedExecuteOp ops = 1

  The post order with the XRT computations to be executed.

Options for an XLA execution.

• int32 device_ordinal = 1

  Local device to run on. This is present because the execute Op may be placed on a device such as CPU or TPU_SYSTEM that logically manages multiple cores.

• int32 core_index_in_replica = 2

  Which model-parallel computation to run from the compiled bundle.

• string execution_instance_key = 3

  Optional key to disambiguate between executions. This is only needed if multiple host send/recvs may be outstanding concurrently with executions.

• uint32 rng_seed = 4

  If non-zero, rng_seed to reset the core with.

• bool release_input_handles = 5

  If true, release allocation handles on the inputs after running.

• bool release_compilation_handle = 6

  If true, release the handle to the computation after running.

• bool return_exploded_tuple = 7

  If set to true, and the result shape is a tuple, then instead of returning a single tuple allocation the execution will return a vector of allocations, one for each of the first-level elements of the result tuple.