xref: /external/tensorflow/tensorflow/core/protobuf/rewriter_config.proto

syntax = "proto3";

package tensorflow;

import "tensorflow/core/framework/attr_value.proto";
import "tensorflow/core/protobuf/verifier_config.proto";

option cc_enable_arenas = true;
option java_outer_classname = "RewriterConfigProtos";
option java_multiple_files = true;
option java_package = "org.tensorflow.framework";
option go_package = "github.com/tensorflow/tensorflow/tensorflow/go/core/protobuf/for_core_protos_go_proto";

message AutoParallelOptions {
  bool enable = 1;
  int32 num_replicas = 2;
}

message ScopedAllocatorOptions {
  // If present, only perform optimization for these ops.
  repeated string enable_op = 1;
}

message RewriterConfig {
  // Graph rewriting is experimental and subject to change, not covered by any
  // API stability guarantees.

  // Configuration options for the meta-optimizer. Unless otherwise noted, these
  // configuration options do not apply to explicitly triggered optimization
  // passes in the optimizers field.

  enum Toggle {
    DEFAULT = 0;
    ON = 1;
    OFF = 2;
    // Enable some aggressive optimizations that use assumptions that TF graphs
    // may break. For example, assume the shape of a placeholder matches its
    // actual feed.
    AGGRESSIVE = 3;
    // Run MLIR pass if there's one implemented in TFG, do nothing otherwise.
    // I.e., if there's no corresponding TFG pass, it's an OFF. This is supposed
    // to be mapped with `ON` and there's no `AGGRESSIVE` in MLIR pass now.
    EXPERIMENTAL_MLIR = 4;
    // Run both MLIR and Grappler passes consecutively and MLIR pass will come
    // first.
    EXPERIMENTAL_BOTH = 5;
  }

  // Enum for layout conversion between NCHW and NHWC on CPU. Default is OFF.
  enum CpuLayout {
    NO_CONVERSION_ON_CPU = 0;
    NCHW_TO_NHWC = 1;
    NHWC_TO_NCHW = 2;
  }

  // Enum controlling the number of times to run optimizers. The default is to
  // run them twice.
  enum NumIterationsType {
    DEFAULT_NUM_ITERS = 0;
    ONE = 1;
    TWO = 2;
  }

  // CPU Conversion settings between NHCW and NCHW.
  CpuLayout cpu_layout_conversion = 50;

  // Optimize tensor layouts (default is ON)
  // e.g. This will try to use NCHW layout on GPU which is faster.
  Toggle layout_optimizer = 1;
  // Fold constants (default is ON)
  // Statically infer the value of tensors when possible, and materialize the
  // result using constants.
  Toggle constant_folding = 3;
  // Shape optimizations (default is ON)
  // Simplify computations made on shapes.
  Toggle shape_optimization = 13;
  // Remapping (default is ON)
  // Remap subgraphs onto more efficient implementations.
  Toggle remapping = 14;
  // Common subgraph elimination (default is ON)
  // e.g. Simplify arithmetic ops; merge ops with same value (like constants).
  Toggle common_subgraph_elimination = 24;
  // Arithmetic optimizations (default is ON)
  // e.g. Simplify arithmetic ops; merge ops with same value (like constants).
  Toggle arithmetic_optimization = 7;
  // Control dependency optimizations (default is ON).
  // Remove redundant control dependencies, which may enable other optimization.
  Toggle dependency_optimization = 8;
  // Loop optimizations (default is ON).
  Toggle loop_optimization = 9;
  // Function optimizations (default is ON).
  Toggle function_optimization = 10;
  // Strips debug-related nodes from the graph (off by default).
  Toggle debug_stripper = 11;
  // If true, don't remove unnecessary ops from the graph
  bool disable_model_pruning = 2;
  // Try to allocate some independent Op outputs contiguously in order to
  // merge or eliminate downstream Ops (off by default).
  Toggle scoped_allocator_optimization = 15;
  // Force small ops onto the CPU (default is OFF).
  Toggle pin_to_host_optimization = 18;
  // Enable the swap of kernel implementations based on the device placement
  // (default is ON).
  Toggle implementation_selector = 22;
  // Optimize data types for CUDA (default is OFF).
  // This will try to use float16 on GPU which is faster.
  // Note that this can change the numerical stability of the graph and may
  // require the use of loss scaling to maintain model convergence.
  Toggle auto_mixed_precision = 23;
  // Optimize data types for oneDNN (default is OFF).
  // This will try to use bfloat16 on CPUs, which is faster.
  // Note that this can change the numerical stability of the graph.
  // Note: this is deprecated.
  // It is replaced by auto_mixed_precision_onednn_bfloat16
  Toggle auto_mixed_precision_mkl = 25;
  // Optimize data types for oneDNN (default is OFF).
  // This will try to use bfloat16 on CPUs, which is faster.
  // Note that this can change the numerical stability of the graph.
  // Note: this is equivalent to the deprecated option auto_mixed_precision_mkl
  Toggle auto_mixed_precision_onednn_bfloat16 = 31;
  // Emulate a model using data type float16 on CPU (default is OFF).
  // This will try to emulate the float16 inputs and outputs of an operator
  // on CPU to have better correlation with float16 on GPU; however the
  // computation in the operator is based on float32.
  // Note that this can change the numerical stability of the graph.
  Toggle auto_mixed_precision_cpu = 29;
  // Disable the entire meta optimizer (off by default).
  bool disable_meta_optimizer = 19;
  // Optimizers registered by plugin (default is ON)
  Toggle use_plugin_optimizers = 28;
  // Conditional code motion (default is ON).
  Toggle experimental_conditional_code_motion = 30;

  // Controls how many times we run the optimizers in meta optimizer (default
  // is once).
  NumIterationsType meta_optimizer_iterations = 12;

  // The minimum number of nodes in a graph to optimizer. For smaller graphs,
  // optimization is skipped.
  // 0 means the system picks an appropriate number.
  // < 0 means do not skip optimization.
  int32 min_graph_nodes = 17;

  // Disable optimizations that assume compressed tensors. Note that this flag
  // is experimental and may be removed in the future.
  bool experimental_disable_compressed_tensor_optimization = 26;

  // Disable folding quantization emulation ops such as FakeQuantWithMinMax* and
  // QuantizeAndDequantize*. Some compilers (e.g. the TF-to-tflite converter)
  // have to extract quantization configs (e.g. min/max range, number of bits,
  // and per-channel) from the quantization emulation ops. Note that this flag
  // is experimental and may be removed in the future. See b/174138564 for more
  // details.
  bool experimental_disable_folding_quantization_emulation = 27;

  enum MemOptType {
    // The default setting (SCHEDULING and SWAPPING HEURISTICS only)
    DEFAULT_MEM_OPT = 0;
    // Disabled in the meta-optimizer.
    NO_MEM_OPT = 1;
    // Driven by manual op-level annotations.
    MANUAL = 2;

    // Driven by heuristics. The behavior of these heuristics is subject to
    // change. Currently includes an experimental recomputation and swapping
    // heuristics. Manual annotations are respected, but additional nodes are
    // selected automatically.

    // Swapping heuristic will move a tensor from the GPU to the CPU and move
    // it back when needed to reduce peak memory usage.
    SWAPPING_HEURISTICS = 4;
    // Recomputation heuristics will recompute ops (such as Relu activation)
    // during backprop instead of storing them, reducing peak memory usage.
    RECOMPUTATION_HEURISTICS = 5;
    // Scheduling will split big ops such as AddN and try to enforce a schedule
    // of the new computations that decreases peak memory usage.
    SCHEDULING_HEURISTICS = 6;
    // Use any combination of swapping and recomputation heuristics.
    HEURISTICS = 3;
  }
  // Configures memory optimization passes through the meta-optimizer. Has no
  // effect on manually requested memory optimization passes in the optimizers
  // field.
  MemOptType memory_optimization = 4;
  // A node name scope for node names which are valid outputs of recomputations.
  // Inputs to nodes that match this scope may be recomputed (subject either to
  // manual annotation of those input nodes or to manual annotation and
  // heuristics depending on memory_optimization), but the nodes themselves will
  // not be recomputed. This matches any sub-scopes as well, meaning the scope
  // can appear not just as a top-level scope. For example, if the value is
  // "gradients/", the default, it will match node name "gradients/foo",
  // "foo/gradients/bar", but not "foo_gradients/"
  string memory_optimizer_target_node_name_scope = 6;
  // Maximum number of milliseconds to spend optimizing a single graph before
  // timing out. If less than or equal to 0 (default value) the optimizer will
  // never time out.
  int64 meta_optimizer_timeout_ms = 20;

  // Configures AutoParallel optimization passes either through the
  // meta-optimizer or when manually specified through the optimizers field.
  AutoParallelOptions auto_parallel = 5;

  // If true, any optimization pass failing will cause the MetaOptimizer to
  // stop with an error. By default - or when set to false, failing passes are
  // skipped silently.
  bool fail_on_optimizer_errors = 21;

  ScopedAllocatorOptions scoped_allocator_opts = 16;

  // If non-empty, will use this as an alternative way to specify a list of
  // optimizations to turn on and the order of the optimizations (replacing the
  // meta-optimizer).
  //
  // Of the RewriterConfig options, only the AutoParallel configuration options
  // (the auto_parallel field) apply to manually requested optimization passes
  // ("autoparallel"). Memory optimization passes ("memory") invoked here are
  // not configurable (in contrast to memory optimization passes through the
  // meta-optimizer) and act only on manual op annotations.
  //
  // Custom optimizers (see custom_optimizers) that are not part of this
  // schedule will be run after - in the order that they were specified.
  repeated string optimizers = 100;

  // Message to describe custom graph optimizer and its parameters
  message CustomGraphOptimizer {
    string name = 1;
    map<string, AttrValue> parameter_map = 2;
  }

  // list of CustomGraphOptimizers to apply.
  repeated CustomGraphOptimizer custom_optimizers = 200;

  // VerifierConfig specifying the verifiers to be run after every optimizer.
  VerifierConfig inter_optimizer_verifier_config = 300;

  // VerifierConfig specifying the verifiers to be run at the end, after all
  // optimizers have run.
  VerifierConfig post_optimization_verifier_config = 301;
}