onnxruntime/winml/test/scenario/cppwinrt/ReluCpu.h

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

#pragma once

#include "core/providers/dml/DmlExecutionProvider/inc/MLOperatorAuthor.h"
#include "core/common//common.h"

struct ReluShapeInferrer : winrt::implements<ReluShapeInferrer, IMLOperatorShapeInferrer> {
  STDMETHOD(InferOutputShapes)(IMLOperatorShapeInferenceContext* context) noexcept {
    uint32_t inputDimsSize;
    context->GetInputTensorDimensionCount(0, &inputDimsSize);

    auto inputDims = new uint32_t[inputDimsSize];
    context->GetInputTensorShape(0, inputDimsSize, inputDims);

    context->SetOutputTensorShape(0, inputDimsSize, inputDims);
    return S_OK;
  }
};

struct ReluOperator : winrt::implements<ReluOperator, IMLOperatorKernel> {
  ReluOperator() {}

  // Computes the outputs of the kernel.  In this case, the output will represent
  // the Rectified Linear Unit (Relu) output.
  //
  // Based on the operators location in the model graph this operator may be called multiple times
  // or simultaneously within the same instance of the class during evaluation.  Implementations
  // of this method must be thread-safe.
  STDMETHOD(Compute)(IMLOperatorKernelContext* context) {
    // Get the input tensor
    winrt::com_ptr<IMLOperatorTensor> inputTensor;
    context->GetInputTensor(0, inputTensor.put());

    // Get the output tensor
    winrt::com_ptr<IMLOperatorTensor> outputTensor;
    context->GetOutputTensor(0, outputTensor.put());

    // Get the input and output shape sizes
    uint32_t inputDimsSize = inputTensor->GetDimensionCount();
    uint32_t outputDimsSize = outputTensor->GetDimensionCount();
    if (inputDimsSize != outputDimsSize) {
      return E_UNEXPECTED;
    }

    // Get the input shape
    std::vector<uint32_t> inputDims(inputDimsSize);
    outputTensor->GetShape(inputDimsSize, inputDims.data());

    // Get the output shape
    std::vector<uint32_t> outputDims(outputDimsSize);
    outputTensor->GetShape(outputDimsSize, outputDims.data());

    // For the number of total elements in the input and output shapes
    auto outputDataSize = std::accumulate(outputDims.begin(), outputDims.end(), 1, std::multiplies<uint32_t>());
    auto inputDataSize = std::accumulate(inputDims.begin(), inputDims.end(), 1, std::multiplies<uint32_t>());
    if (outputDataSize != inputDataSize) {
      return E_UNEXPECTED;
    }

    // If the tensor types are both float type
    if (outputTensor->GetTensorDataType() == MLOperatorTensorDataType::Float &&
        inputTensor->GetTensorDataType() == MLOperatorTensorDataType::Float) {
      // For cpu data
      if (outputTensor->IsCpuData() && inputTensor->IsCpuData()) {
        ComputeInternal<float>(inputTensor.get(), outputTensor.get(), inputDataSize);
      }
    } else if (outputTensor->GetTensorDataType() == MLOperatorTensorDataType::Double &&
               inputTensor->GetTensorDataType() == MLOperatorTensorDataType::Double) {
      // For cpu data
      if (outputTensor->IsCpuData() && inputTensor->IsCpuData()) {
        ComputeInternal<double>(inputTensor.get(), outputTensor.get(), inputDataSize);
      }
    }

    return S_OK;
  }

  template <typename T, typename U = T>
  void ComputeInternal(IMLOperatorTensor* pInputTensor, IMLOperatorTensor* pOutputTensor, uint32_t size) {
    auto inputData = static_cast<T*>(pInputTensor->GetData());
    auto outputData = static_cast<U*>(pOutputTensor->GetData());

    for (uint32_t i = 0; i < size; i++) {
      outputData[i] = static_cast<U>(std::max<T>(0, inputData[i]));
    }
  }
};

struct ReluOperatorFactory : winrt::implements<ReluOperatorFactory, IMLOperatorKernelFactory> {
  STDMETHOD(CreateKernel)(IMLOperatorKernelCreationContext* context, IMLOperatorKernel** kernel) {
    ORT_UNUSED_PARAMETER(context);
    auto reluOperator = winrt::make<ReluOperator>();
    reluOperator.copy_to(kernel);
    return S_OK;
  }

  static MLOperatorEdgeDescription CreateEdgeDescriptor(MLOperatorEdgeType type, MLOperatorTensorDataType dataType) {
    ORT_UNUSED_PARAMETER(type);
    MLOperatorEdgeDescription desc;
    desc.edgeType = MLOperatorEdgeType::Tensor;
    desc.tensorDataType = dataType;
    return desc;
  }

  static void RegisterReluKernel(winrt::com_ptr<IMLOperatorRegistry> registry) {
    MLOperatorKernelDescription kernelDescription;
    kernelDescription.domain = "";
    kernelDescription.name = "Relu";
    kernelDescription.minimumOperatorSetVersion = 1;
    kernelDescription.executionType = MLOperatorExecutionType::Cpu;

    MLOperatorEdgeTypeConstrant typeConstraint;
    typeConstraint.typeLabel = "T";
    std::vector<MLOperatorEdgeDescription> allowedEdges{
      CreateEdgeDescriptor(MLOperatorEdgeType::Tensor, MLOperatorTensorDataType::Double),
      CreateEdgeDescriptor(MLOperatorEdgeType::Tensor, MLOperatorTensorDataType::Float),
      CreateEdgeDescriptor(MLOperatorEdgeType::Tensor, MLOperatorTensorDataType::Float16)
    };
    typeConstraint.allowedTypes = allowedEdges.data();
    typeConstraint.allowedTypeCount = static_cast<uint32_t>(allowedEdges.size());

    std::vector<MLOperatorEdgeTypeConstrant> typeConstraints{typeConstraint};
    kernelDescription.typeConstraints = typeConstraints.data();
    kernelDescription.typeConstraintCount = static_cast<uint32_t>(typeConstraints.size());

    kernelDescription.defaultAttributes = nullptr;
    kernelDescription.defaultAttributeCount = 0;
    kernelDescription.options = MLOperatorKernelOptions::None;
    kernelDescription.executionOptions = 0;

    auto factory = winrt::make<ReluOperatorFactory>();
    auto shareInferrer = winrt::make<ReluShapeInferrer>();

    registry->RegisterOperatorKernel(&kernelDescription, factory.get(), shareInferrer.get());
  }
};