package xla.poplarplugin

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: PoplarUseDescription

• USE_NO_ALIAS = 0

  A use of an HloPoplarBuffer which reads a buffer, but it is not aliased by any outputs.

• USE_ALIAS_READ_ONLY = 1

  A use of an HloPoplarBuffer which reads a buffer and outputs alias that buffer without modifying the values in it.

• USE_ALIAS_READ_WRITE = 2

  A use of an HloPoplarBuffer which reads a buffer, modifies the values and outputs alias that buffer.

An enum representing copy method for each tensor copied.

Used in: PoplarBackendConfig.CopyConfig

• CloneMethod_PreserveOrderAndAliases = 0

  Preserve order and aliases.

• CloneMethod_PreserveOrderUnlessAliases = 1

  Preserve order unless there are aliases, otherwise expand them.

• CloneMethod_Bypass = 2

  Copy nothing and return input tensor.

• CloneMethod_DeduceNewOrderOrPreserveAliases = 3

  Try to defer copy to get better layout or do alias-preserving copy.

• CloneMethod_DeduceNewOrderOrExpandAliases = 4

  Try to defer copy to get better layout or do alias-expanding copy.

• CloneMethod_DeduceNewOrderOrBypass = 5

  Try to defer copy to get better layout or return input tensor.

Used in: PoplarExecutableProto

• int64 id = 1

• int64 total_elements_per_replica = 2

• int64 replication_factor = 3

• repeated CollectiveBalancedRefSlice gathered_to_ref_slices = 4

• repeated uint32 element_map = 5

Used in: CollectiveBalancedHostRearrangement

• int64 begin = 1

• int64 end = 2

• COLLECTIVE_OP_ADD = 0

• COLLECTIVE_OP_MUL = 1

• COLLECTIVE_OP_MIN = 2

• COLLECTIVE_OP_MAX = 3

• COLLECTIVE_OP_LOGICAL_AND = 4

• COLLECTIVE_OP_LOGICAL_OR = 5

• COLLECTIVE_OP_SQUARE_ADD = 6

• COLLECTIVE_OP_LOCAL = 7

• COLLECTIVE_OP_MEAN = 8

An enum used for representing a convolution type in a multi-convolution.

• Conv = 0

• ConvWithReverse = 1

  Retired2 = 2; Retired3 = 3;

Used in: PoplarExecutableProto

• optional Signature signature = 1

  The functional signature of the poplar executable.

• int64 num_IPUs = 2

  Information about runtime used to compile executable

• string target_type = 3

• string target_arch = 4

• bool gateway_mode = 5

• bool supports_remote_buffers = 6

• bool remote_memory_entry_params = 7

• bool executable_can_stall = 8

  Whether this is model can stall without more data arriving. This can currently happen when using pipelining with Grouped or Interleaved schedule or IO tiles.

Configuration for one of the infeed/outfeed streams

Used in: PoplarExecutableProto

• optional PoplarFeedConfig config = 1

• optional ShapeProto shape = 2

Enums to use as string keys in the FrontendAttributes map from tensorflow/compiler/xla/xla_data.proto for the Poplar specific attributes.

• STOCHASTIC_ROUNDING = 0

  Scope override for stochastic rounding (See tensorflow/python/ipu/scopes.py)

• CALL_CONFIG_TYPE = 1

  Custom call's type (Used in pipelining).

• PARTIALS_TYPE = 2

  Type to use to store the partials results.

• GRADIENT_ACCUMULATION_COUNT = 3

  Used in pipelining to indicate how many times each stage is executed.

• PIPELINE_STAGE_ID = 4

  ID of a pipeline stage.

• PIPELINE_REPEAT_COUNT = 5

  Used in pipelining to indicate how many times the pipeline should be executed.

• PIPELINE_SCHEDULE = 6

  Used to select the pipeline schedule.

• PIPELINE_POPLAR_CONFIG = 7

  Used to configure the pipeline lowering to Poplar.

• OFFLOAD_WEIGHT_UPDATE_VARIABLES = 8

  Used to indicate whether to offload variables which are only used by the resource update.

• NUM_BATCHES_TO_ACCUMULATE = 9

  Used to indicate how many batches need to be accumulated during gradient accumulation.

• ML_TYPE = 10

  The "Machine Learning" type of the instruction

• OPTION_FLAGS = 11

  Used to set Poplar OptionFlags of an instruction.

• PIPELINE_BATCH_SERIALIZATION_ITERATIONS = 12

  The number of times a loop executes to compute a batch on each pipeline stage execution.

• PARTITION_OFFLOADED_WEIGHT_UPDATE_VARIABLES = 13

  Whether to partition offloaded weight update variables across the replicas.

• OFFLOAD_ACTIVATIONS = 14

  Used in pipelining to indicate whether activations should be offloaded into remote memory

• OFFLOAD_GRADIENT_ACCUMULATION_BUFFERS = 15

  Used in pipelining to indicate whether the gradient accumulation buffers should be offloaded.

• PARTITION_VARIABLES = 16

  Used in pipelining to indicate whether weights used by the forward/backward stages should be partitioned between replicas.

• OFFLOAD_VARIABLES = 17

  Used in pipelining to indicate whether weights used by the forward/backward stages should be offloaded.

• RECOMPUTATION_MODE = 18

  Used in pipelining to indicate the recomputation mode.

• UNIQUE_SHARDING = 19

  Used for functions to indicate whether the callsite should have unique sharding.

• KEEP_INPUT_LAYOUTS = 20

  Used for functions to indicate whether the callsite should keep the input layouts or whether they can be reallocated.

Used in: PoplarExecutableProto

• string stream_handle = 1

• string embedding_id = 2

• optional ShapeProto indices_shape = 3

• optional ShapeProto activations_shape = 4

Used in: Signature

• string name = 1

• string handle = 2

• int64 argument = 3

• int64 tuple_index = 4

• optional ShapeProto shape = 5

When to attach to the device.

Used in: IpuOptions

• ALWAYS = 0

  Attach when configuring the device.

• ON_DEMAND = 1

  Defer connection to when the IPU is needed. (i.e when running an executable).

• NEVER = 2

  Never attach to the device. (Device can only be used for compilation).

• PRE_COMPILE = 3

  Do not attach to a device and any execution will just return zeros.

The type of execution profiling to use. See poplar docs for the meaning of the profiling types.

Used in: IpuOptions.Profiling

• NO_PROFILE = 0

  Do not do execution profiling.

• DEVICE_PROFILE = 1

  Do execution profiling at the device level.

• IPU_PROFILE = 2

  Do execution profiling at the IPU level.

• TILE_PROFILE = 3

  Do execution profiling at the Tile level.

• optional IpuOptions.IpuModelConfig ipu_model_config = 1

• optional IpuOptions.Profiling profiling = 2

• repeated PoplarOptionFlag compilation_options = 3

  Options to control Poplar compilation

• repeated IpuOptions.DeviceConfig device_config = 4

• optional IpuOptions.SpeedSizeConfig speed_size_config = 5

• repeated PoplarOptionFlag convolution_options = 6

  Options controlling the configuration of convolutions

• repeated PoplarOptionFlag matmul_options = 7

  Options controlling the configuration of matmuls

• repeated PoplarOptionFlag pooling_options = 8

  Options controlling the configuration of pooling operations

• bool clear_matmul_pass_type = 9

  If set, then don't include MatMul pass type in the matmul options

• bool deprecated = 10

  Removed feature

• int64 max_cross_replica_sum_buffer_size = 11

  The maximum number of bytes to wait before scheduling an all-reduce

• int64 max_reduce_scatter_buffer_size = 12

  The maximum number of bytes to wait before scheduling a reduce-scatter

• optional IpuOptions.FloatingPointBehaviour floating_point_behaviour = 13

• int64 max_inter_ipu_copies_buffer_size = 14

  The maximum number of bytes to wait before a inter IPU copy between IPUs is scheduled.

• int64 max_send_recv_cluster_size = 15

  The maximum number of bytes that can be waiting before a send/recv instruction cluster is scheduled.

• int64 max_scheduler_lookahead_depth = 16

  The maximum distance to look into the future for valid schedules.

• int64 max_scheduler_search_space_size = 17

  The maximum number of nodes to consider when building the tree of future schedules.

• bool prefetch_data_streams = 18

  Whether to prefetch data for data streams.

• bool enable_multi_slice_combiner = 19

  Whether to fuse multi_slice operations that use the same input.

• bool enable_matmul_combiner = 20

  Whether to fuse matmul operations that use the same input or the same weights.

• IpuSelectionOrder selection_order = 21

  Which selection order to use when creating the virtual graphs.

• bool disable_gather_simplifier = 22

  Whether to converts gathers into multiSlice operations.

• string serialization_folder = 23

  Where to save the compiled executable.

• IpuDeviceConnectionType device_connection_type = 24

  When to attach to the device.

• string ipu_version = 25

  Version of the IPU hardware used.

• IpuOptionsCreator creator_id = 27

• bool use_stable_norm_statistics = 28

  Use stable statistics calculation in norms

• int64 num_io_tiles = 29

  How many tiles to reserve per IPU for IO operations.

• repeated PoplarOptionFlag gcl_options = 30

  Collective operations options for the Graphcore Communication Library.

• int64 triangular_solve_expander_block_size = 31

  Block size for triangular solve expander pass.

• bool enable_experimental_remote_buffer_embedding = 32

  Enable remote buffer embeddings

• bool enable_fast_math = 33

  Enable math optimizations which might not adhere to IEEE.

• int64 multi_replica_process_count = 34

  Options for multi-replica distribution.

• int64 multi_replica_process_index = 35

• int64 minimum_remote_tensor_size = 36

  The minimum size a tensor (in bytes) has to be in order to be consider for being stored in remote memory.

• bool place_ops_on_io_tiles = 37

  Whether to place TensorFlow operations on IO tiles.

• bool enable_remote_buffers_without_device = 38

  Whether to enable remote buffers for the compilations without devices attached.

• ThreeState remote_buffer_merging_mode = 39

  Control the merging mode for compatible remote buffers. THREESTATE_OFF: Do not attempt to merge any remote buffers. THREESTATE_ON: Attempt to merge all compatible remote buffers. THREESTATE_UNDEFINED: Merge remote buffers only when it is considered beneficial to enable code re-use.

• int64 cholesky_block_size = 40

  Block size for triangular solve expander pass.

• int64 experimental_distributed_batch_norm_replica_group_size = 41

  Group size for distributed batch norm.

• double io_tile_available_memory_proportion = 42

• repeated PoplarOptionFlag slice_options = 43

  Options controlling the configuration of slice operations.

• int64 max_reduce_many_buffer_size = 44

  The maximum number of bytes to wait before scheduling a reduce-many

• optional IpuOptions.IpuAlgebraicSimplifierConfig algebraic_simplifier_config = 45

• bool enable_experimental_prng_stability = 46

  Option for controlling whether experimental prng work is enabled

• int64 max_all_gather_buffer_size = 47

  The maximum number of bytes to wait before scheduling an all-gather

• bool enable_dynamic_slice_replacement = 48

  Option for controlling whether the DynamicSliceReplacer pass is run.

Options controlling the configuration of each IPU device

Used in: IpuOptions

• oneof selection

  Set one of these fields to determine the way to select IPUs

  • int32 auto_count = 1

    Set this to the number of IPUs which are required. The system will acquire a free IPU configuration containing this many physical IPUs.

  • int32 cfg_index = 2

    Set this to acquire a specific IPU hardware configuration.

Used in: IpuOptions

• bool inv = 2

  These flags match the ones provided by the poplar::FloatingPointBehaviour structure

• bool div0 = 3

• bool oflo = 4

• StochasticRoundingBehaviour esr = 5

• bool nanoo = 6

Used in: IpuOptions

• bool enable_fast_math = 1

• bool enable_dot_strength = 2

Options controlling the software IPU model (see IPUModel in poplar)

Used in: IpuOptions

• bool compile_ipu_code = 1

• int64 tiles_per_ipu = 2

• string ipu_model_version = 3

Options to control IPU profiling

Used in: IpuOptions

• bool enable_compilation_trace = 1

  Set to enable compilation trace.

• bool enable_poplar_reports_text = 2

  Set to have the poplar reports in text, not JSON.

• bool enable_io_trace = 3

  Set to enable I/O trace.

• IpuExecutionProfileType execution_trace_type = 4

  Execution trace type.

• int64 report_every_nth_execution = 5

  Report every N'th execution (0=once only).

• bool enable_ipu_trace_events = 7

  Enable IPU trace events - indpendently of poplar reporting.

• bool enable_poplar_reports_cbor = 8

  Set to have the poplar reports in CBOR, not JSON.

• string report_directory = 9

  When non-empty, reports will be written into this directory, instead of into the Tensorboard events.

• int64 max_report_size = 10

  Reports over this size (in bytes) will be discarded.

• repeated PoplarOptionFlag graph_options = 11

  Options for controlling graph profile report generation.

• bool enable_poplar_graph = 13

  Set to have the Poplar serialized graph to be included with the compiler report.

Options to control the memory size and speed of the execution trade-off.

Used in: IpuOptions

• bool always_rearrange_copies_on_the_host = 1

  Option controlling whether to rearrange streamed copies on the host

• bool disable_graph_outlining = 3

  Option to disable caching of sub-graphs containing convolutions

• bool allow_recompute = 4

  Option to enable re-computation.

• bool merge_infeed_io_copies = 6

  Merge all streaming reads into a block to reduce the host-sync round-trip count. This may decrease the device->host read latency at the expense of having more live tensors in memory on the device.

• IpuSchedulingAlgorithm scheduler_selection = 7

  Specifies the scheduling algorithm to use

Who instantiated the IpuOptions object.

Used in: IpuOptions

• INVALID = 0

• IPU_UTILS = 1

Used in: IpuOptions.SpeedSizeConfig

• CHOOSE_BEST = 0

  Tries to automatically determine the best scheduler based on TensorFlow's native heap simulator.

• CLUSTERING = 1

  Groups clusters of operations together in order to look through stretches of instructions with potentially high liveness.

• POST_ORDER = 2

  Schedules the instructions in the order which is obtained by walking the graph in 'post order'.

• LOOK_AHEAD = 3

  Looks ahead a number of operations from any schedulable one, as given by the `max_scheduler_lookahead_depth` and `max_scheduler_search_space_size` options. It attempts to look through areas of high liveness.

• SHORTEST_PATH = 4

  Schedules the graph giving priority to the shortest path to the root.

Used in: IpuOptions

• AUTO = 0

• ZIGZAG = 1

• SNAKE = 2

• HOOF = 3

• string handle = 1

• int64 key = 2

• int64 start_id = 3

An enum representing the "Machine Learning" type of the instruction.

Used in: PoplarBackendConfig

• NONE = 0

• INFERENCE_FWD = 1

• TRAINING_FWD = 2

• TRAINING_BWD = 3

• TRAINING_WU = 4

Used in: Signature

• string name = 1

• string handle = 2

• int64 tuple_index = 3

• optional ShapeProto shape = 4

• repeated PipelineStagePoplarConfig forward_stages = 1

• repeated PipelineStagePoplarConfig backward_stages = 2

• optional PipelineStagePoplarConfig resource_update = 3

Used in: PipelinePoplarConfig

• repeated PoplarOptionFlag convolution_options = 1

• repeated PoplarOptionFlag matmul_options = 2

• repeated PoplarOptionFlag slice_options = 3

Backend specific HloInstruction config

• oneof Config

  • PoplarBackendConfig.FusionConfig fusion_config = 1

  • PoplarBackendConfig.CallConfig call_config = 2

  • PoplarBackendConfig.CopyConfig copy_config = 14

• bool is_inplace = 3

• int64 hash_of_custom_attributes = 4

• ThreeState stochastic_rounding = 5

• MLType ml_type = 6

• PrimitiveType partials_type = 7

• repeated PoplarOptionFlag convolution_options = 8

  Flags to be used by the convolution.

• repeated PoplarOptionFlag matmul_options = 9

  Flags to be used by the matmul.

• repeated PoplarOptionFlag slice_options = 10

  Flags to be used by the slice operation.

• Tileset tileset = 11

  The tile subset that should be used by this instruction.

• bool is_replica_identical = 12

  Tag that says whether an instruction produces a value thats identical across replicas.

• StochasticRoundingMethod stochastic_rounding_method = 13

  Tag for controlling how we perform SR.

Instruction config for a Call instruction

Used in: PoplarBackendConfig

• CallConfig.Type type = 1

• oneof Config

  • CallConfig.CallConfig call_config = 2

  • CallConfig.RepeatConfig repeat_config = 3

  • CallConfig.PipelineStageConfig pipeline_stage_config = 4

  • CallConfig.PipelineConfig pipeline_config = 5

  • CallConfig.FunctionConfig function_config = 6

  • CallConfig.ResourceUpdateConfig resource_update_config = 7

Used in: CallConfig

(message has no fields)

Used in: CallConfig

• bool keep_input_layouts = 1

• bool unique_sharding = 2

• int64 num_modified_remote_buffer_inputs = 3

• int64 num_unmodified_remote_buffer_inputs = 4

• bool partitioned_elementwise_cluster = 5

Used in: CallConfig

• int64 batch_serialization_iterations = 1

• int64 repeat_count = 2

• PipelineConfig.Schedule schedule = 3

• ThreeState offload_activations = 4

• ThreeState offload_gradient_accumulation_buffers = 5

• ThreeState partition_variables = 6

• ThreeState offload_variables = 7

• PipelineConfig.RecomputationMode recomputation_mode = 8

• int64 gradient_accumulation_index = 9

Used in: PipelineConfig

• Auto = 0

• Recompute_then_backpropagate = 1

• Recompute_and_backpropagate_interleaved = 2

Used in: PipelineConfig

• Grouped = 0

• Interleaved = 1

• Sequential = 2

Used in: CallConfig

• int64 stage_id = 1

Used in: CallConfig

• int64 repeat_count = 1

• bool allow_finer_alias_analysis = 2

  TODO(T25039): always enable this option.

Used in: CallConfig

• ThreeState offload_variables = 1

• ThreeState partition_offloaded_variables = 2

Used in: CallConfig

• Call = 0

  Call by default.

• RepeatLoop = 1

• PipelineStage = 2

• PipelineStageBackward = 3

• Pipeline = 4

• PipelineStageRecomputation = 5

• ResourceUpdate = 6

• Function = 7

• MultiConv = 8

Copy config for a copy instruction

Used in: PoplarBackendConfig

• repeated CloneMethod clone_method = 1

Instruction config for a fusion instruction

Used in: PoplarBackendConfig

• optional Window window = 1

  Convolutions

• optional ConvolutionDimensionNumbers dimension_numbers = 2

• int64 feature_group_count = 3

• int64 batch_group_count = 4

• repeated PoplarUseDescription inplace_descriptions = 5

• bool allow_non_inplace = 6

• uint32 replication_factor = 1

  The number of replicas

• repeated FeedConfig infeeds = 2

  Names of the infeed instructions

• repeated FeedConfig outfeeds = 3

  Names of the outfeed instructions

• repeated PoplarOptionFlag option_flags = 4

  The OptionFlags which were used during engine compilation

• repeated SendRecvConfig sends = 5

• repeated SendRecvConfig recvs = 6

• repeated HostEmbeddingConfig lookups = 7

• repeated HostEmbeddingConfig updates = 8

• repeated HostEmbeddingConfig notifications = 9

• repeated RemoteParameterConfig remote_parameters = 10

• bool logging_cycle_count = 11

• optional EmbeddedRuntimeConfig embedded_runtime_config = 12

• uint32 tf_major_version = 13

• uint32 tf_minor_version = 14

• string tf_git_version = 15

• repeated CollectiveBalancedHostRearrangement collective_balanced_host_rearrangements = 16

• string poplar_package_hash = 17

Message which is stored as the config field for an HloInstruction for an HloInfeedInstruction/HloOutfeedInstruction instruction.

Used in: FeedConfig

• string feed_id = 2

  The unique name of this feed.

• PoplarFeedConfig.Mode mode = 3

  Optional mode.

• repeated tensorflow.DataType tf_data_types = 4

  We store the TF DataType for each shape as it is not safe to convert from XLA PrimitveType to TF DataType.

• int64 prefetch_depth = 5

  How many elements to prefetch.

• bool reusable = 6

  Whether the feed sequence is reusable because this feed id has multiple uses.

• bool optimise_latency = 7

  Whether to focus on latency over throughput

Used in: PoplarFeedConfig

• GetAll = 0

• GetLast = 1

Structure for setting an individual Poplar OptionFlag for OptionFlags

Used in: IpuOptions, IpuOptions.Profiling, PipelineStagePoplarConfig, PoplarBackendConfig, PoplarExecutableProto, PoplarOptionFlags

• string option = 1

• string value = 2

Structure for holding multiple Poplar OptionFlags

• repeated PoplarOptionFlag flags = 1

Used to describe how an input buffer from an operand for an instruction is used in the output of that instruction.

Used in: PoplarBackendConfig.FusionConfig

• int64 operand_number = 1

• repeated int64 operand_index = 2

• repeated int64 output_index = 3

• BufferUseKind kind = 4

Used in: PoplarExecutableProto

• int64 parameter_number = 1

• bool is_replica_partitioned = 2

• string buffer_name = 3

• int64 buffer_offset = 4

• int64 num_merged = 5

• repeated int64 merged_params = 6

• int64 host_rearrangement_id = 7

Used in: PoplarExecutableProto

• string stream_handle = 1

• string rendezvous_key = 2

• optional ShapeProto shape = 3

Used in: EmbeddedRuntimeConfig

• repeated Input inputs = 1

• repeated Output outputs = 2

• repeated Input streamed_inputs = 3

• repeated Output streamed_outputs = 4

Used in: IpuOptions.FloatingPointBehaviour

• StochasticRounding_Off = 0

• StochasticRounding_On = 1

• StochasticRounding_ReplicaIdenticalOnly = 2

Used in: PoplarBackendConfig

• StochasticRoundingMethod_Undefined = 0

• StochasticRoundingMethod_DifferingSeeds = 1

• StochasticRoundingMethod_IdenticalSeeds = 2

• StochasticRoundingMethod_Any = 3

• StochasticRoundingMethod_None = 4

• Seed = 0

• Infeed = 1

• Outfeed = 2

• HostEmbedding = 3

• Parameters = 4

• Unknown = 5

An enum for helping with setting a state of something, where the state can be on, off or automatic/not set/undefined.

Used in: IpuOptions, PoplarBackendConfig, PoplarBackendConfig.CallConfig.PipelineConfig, PoplarBackendConfig.CallConfig.ResourceUpdateConfig

• THREESTATE_OFF = 0

• THREESTATE_ON = 1

• THREESTATE_UNDEFINED = 2

An enum representing a subset of tiles per IPU.

Used in: PoplarBackendConfig

• TILESET_COMPUTE_TILES = 0

• TILESET_IO_TILES = 1