kernel_def_builder.h

Go to the documentation of this file.

// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.

#pragma once

#include <limits.h>
#include <memory>
#include <optional>
#include <string>
#include <unordered_map>
#include <vector>

#include "core/common/common.h"
#include "core/framework/allocator.h"
#include "core/framework/data_types.h"
#include "core/graph/basic_types.h"

namespace onnxruntime {
class KernelDefBuilder;

typedef std::map<size_t, OrtMemType> MemTypeMap;

class KernelDef {
 private:
  // note that input/output might be on CPU implicitly when the node is from CPU execution provider
  constexpr static inline bool MemTypeOnCpuExplicitly(OrtMemType mem_type) {
    return mem_type == OrtMemTypeCPUInput || mem_type == OrtMemTypeCPUOutput;
  }

 public:
  explicit KernelDef() = default;

  const std::string& OpName() const {
    return op_name_;
  }

  const std::string& Domain() const {
    return op_domain_;
  }

  void SinceVersion(/*out*/ int* start, /*out*/ int* end) const {
    *start = op_since_version_start_;
    *end = op_since_version_end_;
  }

  const std::pair<int, int> SinceVersion() const {
    return std::pair<int, int>(op_since_version_start_, op_since_version_end_);
  }

  onnxruntime::ProviderType Provider() const {
    return provider_type_;
  }

  // type constraints with types supported in this build
  const std::unordered_map<std::string, std::vector<MLDataType>>& TypeConstraints() const {
    return type_constraints_;
  }

  const std::vector<std::pair<int, int>>& MayInplace() const {
    return inplace_map_;
  }

  const std::vector<std::pair<int, int>>& Alias() const {
    return alias_map_;
  }

  const std::optional<std::pair<int, int>>& VariadicAlias() const {
    return variadic_alias_offsets_;
  }

  OrtMemType InputMemoryType(size_t input_index) const {
    auto it = input_memory_type_args_.find(input_index);
    if (it == input_memory_type_args_.end())
      return default_inputs_mem_type_;
    return it->second;
  }

  bool IsInputOnCpu(size_t input_index) const { return MemTypeOnCpuExplicitly(InputMemoryType(input_index)); }

  bool IsOutputOnCpu(size_t output_index) const { return MemTypeOnCpuExplicitly(OutputMemoryType(output_index)); }

  bool AllocateInputsContiguously() const { return allocate_inputs_contiguously_; }

  bool HasExternalOutputs() const { return external_outputs_; }

#ifdef ENABLE_STRIDED_TENSORS
  const std::vector<int>& MayStridedInput() const { return may_strided_inputs_; }
  const std::vector<std::pair<int, int>>& MayStridedOutput() const { return may_strided_output_map_; }
#endif

  OrtMemType OutputMemoryType(size_t output_index) const {
    auto it = output_memory_type_args_.find(output_index);
    if (it == output_memory_type_args_.end())
      return default_outputs_mem_type_;
    return it->second;
  }

  int ExecQueueId() const {
    return exec_queue_id_;
  }

  bool IsConflict(const KernelDef& other) const;

 private:
  friend class KernelDefBuilder;

  // The operator name supported by <*this> kernel..
  std::string op_name_;

  // The operator since_version range supported by <*this> kernel.
  // A kernel could support an operator definition between <op_since_version_start>
  // and <op_since_version_end> (inclusive).
  int op_since_version_start_ = 1;
  int op_since_version_end_ = INT_MAX;

  // The operator domain supported by <*this> kernel.
  // Default to 'onnxruntime::kOnnxDomain'.
  // Please note the behavior of std::string("") and std::string() are different
  std::string op_domain_;

  // The type of the execution provider.
  std::string provider_type_;

  // The data types that are supported in this build (enabled) for inputs/outputs.
  // Key is input/output/type constraint name defined in op schema, Value is supported types.
  std::unordered_map<std::string, std::vector<MLDataType>> type_constraints_;

  // An element <i, j> means that output j reuses the memory of input i.
  std::vector<std::pair<int, int>> inplace_map_;

  // An element <i, j> means that output j is an alias of input i.
  std::vector<std::pair<int, int>> alias_map_;

  // This variable stores <input_offset, output_offset> for the variadic alias mapping
  // output 'i + output_offset' is an alias of input 'i + input_offset' for all i >= 0
  std::optional<std::pair<int, int>> variadic_alias_offsets_;

  // Require input tensors to be allocated contiguously.
  bool allocate_inputs_contiguously_ = false;

  // Whether the outputs are from external.
  bool external_outputs_ = false;

#ifdef ENABLE_STRIDED_TENSORS
  // An element i means i-th input can be strided tensor.
  std::vector<int> may_strided_inputs_;

  // An element <i, j> means j-th output can be a strided tensor, which share the data from i-th input.
  std::vector<std::pair<int, int>> may_strided_output_map_;
#endif

  // The memory types of inputs/outputs of this kernel
  MemTypeMap input_memory_type_args_;
  MemTypeMap output_memory_type_args_;

  // execution command queue id, 0 for default queue in execution provider
  int exec_queue_id_ = 0;
  // Default memory type for all inputs
  OrtMemType default_inputs_mem_type_{OrtMemTypeDefault};
  // Default memory type for all outputs
  OrtMemType default_outputs_mem_type_{OrtMemTypeDefault};
};

class KernelDefBuilder {
 public:
  static std::unique_ptr<KernelDefBuilder> Create() { return std::make_unique<KernelDefBuilder>(); }

  explicit KernelDefBuilder()
      : kernel_def_(std::make_unique<KernelDef>()) {}

  KernelDefBuilder& SetName(const std::string& op_name);
  KernelDefBuilder& SetName(const char* op_name);

  KernelDefBuilder& SetDomain(const std::string& domain);
  KernelDefBuilder& SetDomain(const char* domain);

  /**
     This kernel supports operator definition since <since_version> (to latest).
  */
  KernelDefBuilder& SinceVersion(int since_version) {
    kernel_def_->op_since_version_start_ = since_version;
    return *this;
  }

  /**
     The start and end version should be set accordingly per version range for
     each domain registered in OpSchemaRegistry::DomainToVersionRange in
     \onnxruntime\onnxruntime\core\graph\op.h as below.
     Key: domain. Value: <lowest version, highest version> pair.
     std::unordered_map<std::string, std::pair<int, int>> map_;
  */
  KernelDefBuilder& SinceVersion(int since_version_start, int since_version_end) {
    kernel_def_->op_since_version_start_ = since_version_start;
    kernel_def_->op_since_version_end_ = since_version_end;
    return *this;
  }

  /**
     The execution provider type of the kernel.
  */
  KernelDefBuilder& Provider(ProviderType provider_type);
  KernelDefBuilder& Provider(const char* provider_type);

  /**
     Specify the set of types that this kernel supports. A further restriction
     of the set of types specified in the op schema.

     @param arg_name The arg name can be either op formal parameter name, say "X", or type
                     argument name specified in op schema, say "T".
     @param types The types that are supported in this build.
  */
  KernelDefBuilder& TypeConstraint(const std::string& arg_name, std::vector<MLDataType> types);
  KernelDefBuilder& TypeConstraint(const char* arg_name, std::vector<MLDataType> types);

  /**
     Like TypeConstraint but supports just a single type.
  */
  KernelDefBuilder& TypeConstraint(const std::string& arg_name, MLDataType type);
  KernelDefBuilder& TypeConstraint(const char* arg_name, MLDataType type);

  /**
     Inplace mapping from inputs to outputs allowed.
     It means that uplayer runtime could do memory in-place optimization
     as it will not impact the correctness of this kernel.
  */
  KernelDefBuilder& MayInplace(const std::vector<std::pair<int, int>>& inplaces);
  KernelDefBuilder& MayInplace(int input_index, int output_index);

  /**
     Alias mapping from inputs to outputs. Different from Inplace that the
     content of the tensor is not changed. This is to take care of operators
     such as Identity and Reshape.
  */
  KernelDefBuilder& Alias(const std::vector<std::pair<int, int>>& aliases);
  KernelDefBuilder& Alias(int input_index, int output_index);

  /**
     Apply variadic number of alias mapping from inputs to outputs.
     This is effectively applying Alias(i + input_offset, i + output_offset) for i >= 0
  */
  KernelDefBuilder& VariadicAlias(int input_offset, int output_offset);

  /**
     Specify that this kernel requires input tensors to be allocated
     contiguously. This allows kernels to execute as a single large
     computation, rather than numerous smaller computations.
  */
  KernelDefBuilder& AllocateInputsContiguously() {
    kernel_def_->allocate_inputs_contiguously_ = true;
    return *this;
  }

  /**
     Specify that this kernel's output buffers are passed from external,
     i.e. not created or managed by ORT's memory allocator.
  */
  KernelDefBuilder& ExternalOutputs() {
    kernel_def_->external_outputs_ = true;
    return *this;
  }

#ifdef ENABLE_STRIDED_TENSORS
  /**
     Specify that the input_index-th input can be strided tensor.
   */
  KernelDefBuilder& MayStridedInput(int input_index);

  /**
     Specify that the output_index-th output can be strided tensor, and share the data
     from input_index-th input.
   */
  KernelDefBuilder& MayStridedOutput(int input_index, int output_index);
#endif

  /**
     Specify that this kernel requires an input arg
     in certain memory type (instead of the default, device memory).
  */
  KernelDefBuilder& InputMemoryType(OrtMemType type, int input_index) {
    kernel_def_->input_memory_type_args_.insert(std::make_pair(input_index, type));
    return *this;
  }

  /**
     Specify that this kernel requires input arguments
     in certain memory type (instead of the default, device memory).
  */
  KernelDefBuilder& InputMemoryType(OrtMemType type, const std::vector<int>& input_indexes) {
    for (auto input_index : input_indexes) {
      kernel_def_->input_memory_type_args_.insert(std::make_pair(input_index, type));
    }
    return *this;
  }

  /**
     Specify that this kernel provides an output arg
     in certain memory type (instead of the default, device memory).
  */
  KernelDefBuilder& OutputMemoryType(OrtMemType type, int output_index) {
    kernel_def_->output_memory_type_args_.insert(std::make_pair(output_index, type));
    return *this;
  }

  /**
     Specify that this kernel provides an output arguments
     in certain memory type (instead of the default, device memory).
  */
  KernelDefBuilder& OutputMemoryType(OrtMemType type, const std::vector<int>& output_indexes) {
    for (auto output_index : output_indexes) {
      kernel_def_->output_memory_type_args_.insert(std::make_pair(output_index, type));
    }
    return *this;
  }

  /**
     Specify that this kernel runs on which execution queue in the provider
  */
  KernelDefBuilder& ExecQueueId(int queue_id) {
    kernel_def_->exec_queue_id_ = queue_id;
    return *this;
  }

  /**
  Specify the default inputs memory type, if not specified, it is DefaultMemory
  */
  KernelDefBuilder& SetDefaultInputsMemoryType(OrtMemType mem_type) {
    kernel_def_->default_inputs_mem_type_ = mem_type;
    return *this;
  }

  /**
  Specify the default outputs memory type, if not specified, it is DefaultMemory
  */
  KernelDefBuilder& SetDefaultOutputMemoryType(OrtMemType mem_type) {
    kernel_def_->default_outputs_mem_type_ = mem_type;
    return *this;
  }

  /**
     Return the kernel definition, passing ownership of the KernelDef to the caller
  */
  std::unique_ptr<KernelDef> Build() {
    return std::move(kernel_def_);
  }

 private:
  // we own the KernelDef until Build() is called.
  std::unique_ptr<KernelDef> kernel_def_;
};

}  // namespace onnxruntime