ei-cpp-export/edge-impulse-sdk/tensorflow/lite/micro/kernels/add.cc

// Patched by Edge Impulse to include reference and hardware-accelerated kernels
#include "../../../../classifier/ei_classifier_config.h"
#if 0 == 1
/* noop */
#elif EI_CLASSIFIER_TFLITE_ENABLE_CMSIS_NN == 1
/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/add.h"

#include "edge-impulse-sdk/CMSIS/NN/Include/arm_nnfunctions.h"
#include "edge-impulse-sdk/tensorflow/lite/c/builtin_op_data.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/quantization_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/integer_ops/add.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/op_macros.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/memory_helpers.h"

namespace tflite {
namespace ops {
namespace micro {
namespace add {

constexpr int kInputTensor1 = 0;
constexpr int kInputTensor2 = 1;
constexpr int kOutputTensor = 0;

struct OpData {
  bool requires_broadcast;

  // These fields are used in both the general 8-bit -> 8bit quantized path,
  // and the special 16-bit -> 16bit quantized path
  int input1_shift;
  int input2_shift;
  int32_t output_activation_min;
  int32_t output_activation_max;

  // These fields are used only in the general 8-bit -> 8bit quantized path
  int32_t input1_multiplier;
  int32_t input2_multiplier;
  int32_t output_multiplier;
  int output_shift;
  int left_shift;
  int32_t input1_offset;
  int32_t input2_offset;
  int32_t output_offset;

  // Used only for float evals:
  float output_activation_min_f32;
  float output_activation_max_f32;
};

TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteAddParams* params,
                             const TfLiteTensor* input1,
                             const TfLiteTensor* input2, TfLiteTensor* output,
                             OpData* data) {
  data->requires_broadcast = !HaveSameShapes(input1, input2);

  if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
    // 8bit -> 8bit general quantized path, with general rescalings
    data->input1_offset = -input1->params.zero_point;
    data->input2_offset = -input2->params.zero_point;
    data->output_offset = output->params.zero_point;
    data->left_shift = 20;
    const double twice_max_input_scale =
        2 * static_cast<double>(
                std::max(input1->params.scale, input2->params.scale));
    const double real_input1_multiplier =
        static_cast<double>(input1->params.scale) / twice_max_input_scale;
    const double real_input2_multiplier =
        static_cast<double>(input2->params.scale) / twice_max_input_scale;
    const double real_output_multiplier =
        twice_max_input_scale /
        ((1 << data->left_shift) * static_cast<double>(output->params.scale));

    QuantizeMultiplierSmallerThanOneExp(
        real_input1_multiplier, &data->input1_multiplier, &data->input1_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_input2_multiplier, &data->input2_multiplier, &data->input2_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_output_multiplier, &data->output_multiplier, &data->output_shift);

    TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
        context, params->activation, output, &data->output_activation_min,
        &data->output_activation_max));
  } else if (output->type == kTfLiteFloat32) {
    CalculateActivationRange(params->activation,
                             &data->output_activation_min_f32,
                             &data->output_activation_max_f32);
  }

  return kTfLiteOk;
}

void EvalAdd(TfLiteContext* context, TfLiteNode* node, TfLiteAddParams* params,
             const OpData* data, const TfLiteEvalTensor* input1,
             const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) {
  tflite::ArithmeticParams op_params;
  SetActivationParams(data->output_activation_min_f32,
                      data->output_activation_max_f32, &op_params);
#define TF_LITE_ADD(opname)                                               \
  reference_ops::opname(op_params, tflite::micro::GetTensorShape(input1), \
                        tflite::micro::GetTensorData<float>(input1),      \
                        tflite::micro::GetTensorShape(input2),            \
                        tflite::micro::GetTensorData<float>(input2),      \
                        tflite::micro::GetTensorShape(output),            \
                        tflite::micro::GetTensorData<float>(output))
  if (data->requires_broadcast) {
    TF_LITE_ADD(BroadcastAdd4DSlow);
  } else {
    TF_LITE_ADD(Add);
  }
#undef TF_LITE_ADD
}

TfLiteStatus EvalAddQuantized(TfLiteContext* context, TfLiteNode* node,
                              TfLiteAddParams* params, const OpData* data,
                              const TfLiteEvalTensor* input1,
                              const TfLiteEvalTensor* input2,
                              TfLiteEvalTensor* output) {
  if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
    tflite::ArithmeticParams op_params;
    op_params.left_shift = data->left_shift;
    op_params.input1_offset = data->input1_offset;
    op_params.input1_multiplier = data->input1_multiplier;
    op_params.input1_shift = data->input1_shift;
    op_params.input2_offset = data->input2_offset;
    op_params.input2_multiplier = data->input2_multiplier;
    op_params.input2_shift = data->input2_shift;
    op_params.output_offset = data->output_offset;
    op_params.output_multiplier = data->output_multiplier;
    op_params.output_shift = data->output_shift;
    SetActivationParams(data->output_activation_min,
                        data->output_activation_max, &op_params);
    bool need_broadcast = reference_ops::ProcessBroadcastShapes(
        tflite::micro::GetTensorShape(input1),
        tflite::micro::GetTensorShape(input2), &op_params);
#define TF_LITE_ADD(type, opname, dtype)                         \
  type::opname(op_params, tflite::micro::GetTensorShape(input1), \
               tflite::micro::GetTensorData<dtype>(input1),      \
               tflite::micro::GetTensorShape(input2),            \
               tflite::micro::GetTensorData<dtype>(input2),      \
               tflite::micro::GetTensorShape(output),            \
               tflite::micro::GetTensorData<dtype>(output));
    if (output->type == kTfLiteInt8) {
      if (need_broadcast) {
        TF_LITE_ADD(reference_integer_ops, BroadcastAdd4DSlow, int8_t);
      } else {
        arm_elementwise_add_s8(
            tflite::micro::GetTensorData<int8_t>(input1),
            tflite::micro::GetTensorData<int8_t>(input2),
            op_params.input1_offset, op_params.input1_multiplier,
            op_params.input1_shift, op_params.input2_offset,
            op_params.input2_multiplier, op_params.input2_shift,
            op_params.left_shift, tflite::micro::GetTensorData<int8_t>(output),
            op_params.output_offset, op_params.output_multiplier,
            op_params.output_shift, op_params.quantized_activation_min,
            op_params.quantized_activation_max,
            MatchingElementsSize(tflite::micro::GetTensorShape(input1),
                                 tflite::micro::GetTensorShape(input2),
                                 tflite::micro::GetTensorShape(output)));
      }
    } else {
      if (need_broadcast) {
        TF_LITE_ADD(reference_ops, BroadcastAdd4DSlow, uint8_t);
      } else {
        TF_LITE_ADD(reference_ops, Add, uint8_t);
      }
    }
#undef TF_LITE_ADD
  }

  return kTfLiteOk;
}

void* Init(TfLiteContext* context, const char* buffer, size_t length) {
  TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr);
  return context->AllocatePersistentBuffer(context, sizeof(OpData));
}

TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
  TFLITE_DCHECK(node->user_data != nullptr);
  TFLITE_DCHECK(node->builtin_data != nullptr);

  const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
  TF_LITE_ENSURE(context, input1 != nullptr);
  const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
  TF_LITE_ENSURE(context, input2 != nullptr);
  TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
  TF_LITE_ENSURE(context, output != nullptr);

  OpData* data = static_cast<OpData*>(node->user_data);
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  TF_LITE_ENSURE_STATUS(
      CalculateOpData(context, params, input1, input2, output, data));

  return kTfLiteOk;
}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  const TfLiteEvalTensor* input1 =
      tflite::micro::GetEvalInput(context, node, kInputTensor1);
  const TfLiteEvalTensor* input2 =
      tflite::micro::GetEvalInput(context, node, kInputTensor2);
  TfLiteEvalTensor* output =
      tflite::micro::GetEvalOutput(context, node, kOutputTensor);

  TFLITE_DCHECK(node->user_data != nullptr);
  const OpData* data = static_cast<const OpData*>(node->user_data);

  if (output->type == kTfLiteFloat32) {
    EvalAdd(context, node, params, data, input1, input2, output);
  } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
    TF_LITE_ENSURE_OK(context, EvalAddQuantized(context, node, params, data,
                                                input1, input2, output));
  } else {
    TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.",
                       TfLiteTypeGetName(output->type), output->type);
    return kTfLiteError;
  }

  return kTfLiteOk;
}

}  // namespace add

TfLiteRegistration Register_ADD() {
  return {/*init=*/add::Init,
          /*free=*/nullptr,
          /*prepare=*/add::Prepare,
          /*invoke=*/add::Eval,
          /*profiling_string=*/nullptr,
          /*builtin_code=*/0,
          /*custom_name=*/nullptr,
          /*version=*/0};
}

}  // namespace micro
}  // namespace ops
}  // namespace tflite

#elif EI_CLASSIFIER_TFLITE_ENABLE_SILABS_MVP == 1
/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/add.h"

#include "edge-impulse-sdk/tensorflow/lite/c/builtin_op_data.h"
#include "edge-impulse-sdk/tensorflow/lite/c/common.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/quantization_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/integer_ops/add.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/op_macros.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/kernels/kernel_util.h"
#include "sl_mvp_ml_add.h"

namespace tflite {
namespace sl {
namespace add {

constexpr int kInputTensor1 = 0;
constexpr int kInputTensor2 = 1;
constexpr int kOutputTensor = 0;

struct OpData {
  bool requires_broadcast;

  int input1_shift;
  int input2_shift;
  int32_t input1_multiplier;
  int32_t input2_multiplier;
  int32_t output_multiplier;
  int output_shift;
  int left_shift;

  sli_mvp_ml_add_s8_params_t params;

  // Used only for float evals:
  float output_activation_min_f32;
  float output_activation_max_f32;
};

TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteAddParams* params,
                             const TfLiteTensor* input1,
                             const TfLiteTensor* input2, TfLiteTensor* output,
                             OpData* data) {
  data->requires_broadcast = !HaveSameShapes(input1, input2);

  if (output->type == kTfLiteInt8) {
    data->params.input1_offset = -input1->params.zero_point;
    data->params.input2_offset = -input2->params.zero_point;
    data->params.output_offset = output->params.zero_point;
    data->params.input1_multiplier = input1->params.scale;
    data->params.input2_multiplier = input2->params.scale;
    data->params.output_multiplier = 1.0 / output->params.scale;
    data->params.length = GetTensorShape(input1).FlatSize();

    int32_t activation_min;
    int32_t activation_max;
    TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
        context, params->activation, output, &activation_min,
        &activation_max));
    data->params.activation_min = static_cast<int8_t>(activation_min);
    data->params.activation_max = static_cast<int8_t>(activation_max);

    // These multipliers and parameters are not used by the MVP codepath,
    // however are needed in cases where broadcast is used.
    data->left_shift = 20;
    const double twice_max_input_scale =
        2 * static_cast<double>(
                std::max(input1->params.scale, input2->params.scale));
    const double real_input1_multiplier =
        static_cast<double>(input1->params.scale) / twice_max_input_scale;
    const double real_input2_multiplier =
        static_cast<double>(input2->params.scale) / twice_max_input_scale;
    const double real_output_multiplier =
        twice_max_input_scale /
        ((1 << data->left_shift) * static_cast<double>(output->params.scale));

    QuantizeMultiplierSmallerThanOneExp(
        real_input1_multiplier, &data->input1_multiplier, &data->input1_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_input2_multiplier, &data->input2_multiplier, &data->input2_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_output_multiplier, &data->output_multiplier, &data->output_shift);

  } else if (output->type == kTfLiteFloat32) {
    CalculateActivationRange(params->activation,
                             &data->output_activation_min_f32,
                             &data->output_activation_max_f32);
  }

  return kTfLiteOk;
}

void EvalAdd(TfLiteContext* context, TfLiteNode* node, TfLiteAddParams* params,
             const OpData* data, const TfLiteEvalTensor* input1,
             const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) {
  tflite::ArithmeticParams op_params;
  SetActivationParams(data->output_activation_min_f32,
                      data->output_activation_max_f32, &op_params);
  if (data->requires_broadcast) {
    reference_ops::BroadcastAdd4DSlow(op_params, tflite::micro::GetTensorShape(input1), tflite::micro::GetTensorData<float>(input1),
                                      tflite::micro::GetTensorShape(input2), tflite::micro::GetTensorData<float>(input2),
                                      tflite::micro::GetTensorShape(output), tflite::micro::GetTensorData<float>(output));
  } else {
    reference_ops::Add(op_params,
                       tflite::micro::GetTensorShape(input1), tflite::micro::GetTensorData<float>(input1),
                       tflite::micro::GetTensorShape(input2), tflite::micro::GetTensorData<float>(input2),
                       tflite::micro::GetTensorShape(output), tflite::micro::GetTensorData<float>(output));
  }
}

TfLiteStatus EvalAddQuantized(TfLiteContext* context, TfLiteNode* node,
                              TfLiteAddParams* params, const OpData* data,
                              const TfLiteEvalTensor* input1,
                              const TfLiteEvalTensor* input2,
                              TfLiteEvalTensor* output) {
  TfLiteStatus status = kTfLiteOk;
  tflite::ArithmeticParams op_params;
  op_params.left_shift = data->left_shift;
  op_params.input1_offset = data->params.input1_offset;
  op_params.input1_multiplier = data->input1_multiplier;
  op_params.input1_shift = data->input1_shift;
  op_params.input2_offset = data->params.input2_offset;
  op_params.input2_multiplier = data->input2_multiplier;
  op_params.input2_shift = data->input2_shift;
  op_params.output_offset = data->params.output_offset;
  op_params.output_multiplier = data->output_multiplier;
  op_params.output_shift = data->output_shift;
  op_params.quantized_activation_min = data->params.activation_min;
  op_params.quantized_activation_max = data->params.activation_max;

  // TODO: Do we need to support the broadcast scenario?
  bool need_broadcast = reference_ops::ProcessBroadcastShapes(tflite::micro::GetTensorShape(input1), tflite::micro::GetTensorShape(input2), &op_params);

  if (need_broadcast) {
    reference_integer_ops::BroadcastAdd4DSlow(op_params,
                                              tflite::micro::GetTensorShape(input1), tflite::micro::GetTensorData<int8_t>(input1),
                                              tflite::micro::GetTensorShape(input2), tflite::micro::GetTensorData<int8_t>(input2),
                                              tflite::micro::GetTensorShape(output), tflite::micro::GetTensorData<int8_t>(output));
  } else {
    sli_mvp_ml_add_s8_params_t params = data->params;
    params.input1 = tflite::micro::GetTensorData<int8_t>(input1);
    params.input2 = tflite::micro::GetTensorData<int8_t>(input2);
    params.output = tflite::micro::GetTensorData<int8_t>(output);
    sl_status_t ret = sli_mvp_ml_add_s8(&params);
    if (ret != SL_STATUS_OK) {
        status = kTfLiteError;
    }
  }

  return status;
}

void* Init(TfLiteContext* context, const char* buffer, size_t length) {
  TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr);
  return context->AllocatePersistentBuffer(context, sizeof(OpData));
}

TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
  TFLITE_DCHECK(node->user_data != nullptr);
  TFLITE_DCHECK(node->builtin_data != nullptr);

  const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
  TF_LITE_ENSURE(context, input1 != nullptr);
  const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
  TF_LITE_ENSURE(context, input2 != nullptr);
  TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
  TF_LITE_ENSURE(context, output != nullptr);

  OpData* data = static_cast<OpData*>(node->user_data);
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  TF_LITE_ENSURE_STATUS(
      CalculateOpData(context, params, input1, input2, output, data));

  return kTfLiteOk;
}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  TFLITE_DCHECK(node->user_data != nullptr);
  const OpData* data = static_cast<const OpData*>(node->user_data);

  const TfLiteEvalTensor* input1 = tflite::micro::GetEvalInput(context, node, kInputTensor1);
  const TfLiteEvalTensor* input2 = tflite::micro::GetEvalInput(context, node, kInputTensor2);
  TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, kOutputTensor);

  if (output->type == kTfLiteFloat32) {
    EvalAdd(context, node, params, data, input1, input2, output);
  } else if (output->type == kTfLiteInt8) {
    TF_LITE_ENSURE_OK(context, EvalAddQuantized(context, node, params, data,
                                                input1, input2, output));
  } else {
    TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.",
                       TfLiteTypeGetName(output->type), output->type);
    return kTfLiteError;
  }

  return kTfLiteOk;
}

}  // namespace add
}  // namespace sl

namespace ops {
namespace micro {
TfLiteRegistration Register_ADD() {
  return {/*init=*/sl::add::Init,
          /*free=*/nullptr,
          /*prepare=*/sl::add::Prepare,
          /*invoke=*/sl::add::Eval,
          /*profiling_string=*/nullptr,
          /*builtin_code=*/0,
          /*custom_name=*/nullptr,
          /*version=*/0};
}

}  // namespace micro
}  // namespace ops
}  // namespace tflite

#elif EI_CLASSIFIER_TFLITE_ENABLE_ESP_NN == 1
/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "edge-impulse-sdk/tensorflow/lite/c/builtin_op_data.h"
#include "edge-impulse-sdk/tensorflow/lite/c/common.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/quantization_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/add.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/integer_ops/add.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/op_macros.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/memory_helpers.h"

#include <esp_timer.h>
#include "edge-impulse-sdk/porting/espressif/ESP-NN/include/esp_nn.h"

long long add_total_time = 0;

namespace tflite {
namespace ops {
namespace micro {
namespace add {

constexpr int kInputTensor1 = 0;
constexpr int kInputTensor2 = 1;
constexpr int kOutputTensor = 0;

struct OpData {
  bool requires_broadcast;

  // These fields are used in both the general 8-bit -> 8bit quantized path,
  // and the special 16-bit -> 16bit quantized path
  int input1_shift;
  int input2_shift;
  int32_t output_activation_min;
  int32_t output_activation_max;

  // These fields are used only in the general 8-bit -> 8bit quantized path
  int32_t input1_multiplier;
  int32_t input2_multiplier;
  int32_t output_multiplier;
  int output_shift;
  int left_shift;
  int32_t input1_offset;
  int32_t input2_offset;
  int32_t output_offset;

  // Used only for float evals:
  float output_activation_min_f32;
  float output_activation_max_f32;
};

void EvalAdd(TfLiteContext* context, TfLiteNode* node, TfLiteAddParams* params,
             const OpData* data, const TfLiteEvalTensor* input1,
             const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) {
  tflite::ArithmeticParams op_params;
  SetActivationParams(data->output_activation_min_f32,
                      data->output_activation_max_f32, &op_params);
  if (data->requires_broadcast) {
    reference_ops::BroadcastAdd4DSlow(
        op_params, tflite::micro::GetTensorShape(input1),
        tflite::micro::GetTensorData<float>(input1),
        tflite::micro::GetTensorShape(input2),
        tflite::micro::GetTensorData<float>(input2),
        tflite::micro::GetTensorShape(output),
        tflite::micro::GetTensorData<float>(output));
  } else {
    reference_ops::Add(op_params, tflite::micro::GetTensorShape(input1),
                       tflite::micro::GetTensorData<float>(input1),
                       tflite::micro::GetTensorShape(input2),
                       tflite::micro::GetTensorData<float>(input2),
                       tflite::micro::GetTensorShape(output),
                       tflite::micro::GetTensorData<float>(output));
  }
}

TfLiteStatus EvalAddQuantized(TfLiteContext* context, TfLiteNode* node,
                              TfLiteAddParams* params, const OpData* data,
                              const TfLiteEvalTensor* input1,
                              const TfLiteEvalTensor* input2,
                              TfLiteEvalTensor* output) {
  tflite::ArithmeticParams op_params;
  op_params.left_shift = data->left_shift;
  op_params.input1_offset = data->input1_offset;
  op_params.input1_multiplier = data->input1_multiplier;
  op_params.input1_shift = data->input1_shift;
  op_params.input2_offset = data->input2_offset;
  op_params.input2_multiplier = data->input2_multiplier;
  op_params.input2_shift = data->input2_shift;
  op_params.output_offset = data->output_offset;
  op_params.output_multiplier = data->output_multiplier;
  op_params.output_shift = data->output_shift;
  SetActivationParams(data->output_activation_min, data->output_activation_max,
                      &op_params);
  bool need_broadcast = reference_ops::ProcessBroadcastShapes(
      tflite::micro::GetTensorShape(input1),
      tflite::micro::GetTensorShape(input2), &op_params);

  switch (output->type) {
    case kTfLiteInt8: {
      if (need_broadcast) {
        reference_integer_ops::BroadcastAdd4DSlow(
            op_params, tflite::micro::GetTensorShape(input1),
            tflite::micro::GetTensorData<int8_t>(input1),
            tflite::micro::GetTensorShape(input2),
            tflite::micro::GetTensorData<int8_t>(input2),
            tflite::micro::GetTensorShape(output),
            tflite::micro::GetTensorData<int8_t>(output));
      } else {
        const int8_t *input1_data = tflite::micro::GetTensorData<int8_t>(input1);
        const int8_t *input2_data = tflite::micro::GetTensorData<int8_t>(input2);
        int8_t *out_data = tflite::micro::GetTensorData<int8_t>(output);

        esp_nn_add_elementwise_s8(input1_data,
                                  input2_data,
                                  data->input1_offset,
                                  data->input2_offset,
                                  data->input1_multiplier,
                                  data->input2_multiplier,
                                  data->input1_shift,
                                  data->input2_shift,
                                  data->left_shift,
                                  out_data,
                                  data->output_offset,
                                  data->output_multiplier,
                                  data->output_shift,
                                  data->output_activation_min,
                                  data->output_activation_max,
                                  MatchingElementsSize(tflite::micro::GetTensorShape(input1),
                                                       tflite::micro::GetTensorShape(input2),
                                                       tflite::micro::GetTensorShape(output))
                                  );
      }
      break;
    }
    case kTfLiteInt16: {
      if (need_broadcast) {
        reference_ops::BroadcastAdd4DSlow(
            op_params, tflite::micro::GetTensorShape(input1),
            tflite::micro::GetTensorData<int16_t>(input1),
            tflite::micro::GetTensorShape(input2),
            tflite::micro::GetTensorData<int16_t>(input2),
            tflite::micro::GetTensorShape(output),
            tflite::micro::GetTensorData<int16_t>(output));
      } else {
        reference_ops::Add(op_params, tflite::micro::GetTensorShape(input1),
                           tflite::micro::GetTensorData<int16_t>(input1),
                           tflite::micro::GetTensorShape(input2),
                           tflite::micro::GetTensorData<int16_t>(input2),
                           tflite::micro::GetTensorShape(output),
                           tflite::micro::GetTensorData<int16_t>(output),
                           false);
      }
      break;
    }
    default:
      TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.",
                       TfLiteTypeGetName(output->type), output->type);
      return kTfLiteError;
  }

  return kTfLiteOk;
}

TfLiteStatus CalculateOpDataAdd(TfLiteContext* context, TfLiteAddParams* params,
                                const TfLiteTensor* input1,
                                const TfLiteTensor* input2,
                                TfLiteTensor* output, OpData* data) {
  data->requires_broadcast = !HaveSameShapes(input1, input2);

  if (output->type == kTfLiteInt8 || output->type == kTfLiteInt16) {
    // 8bit -> 8bit general quantized path, with general rescalings
    data->input1_offset = -input1->params.zero_point;
    data->input2_offset = -input2->params.zero_point;
    data->output_offset = output->params.zero_point;
    data->left_shift = (output->type == kTfLiteInt16) ? 15 : 20;
    const double twice_max_input_scale =
        2 * static_cast<double>(
                std::max(input1->params.scale, input2->params.scale));
    const double real_input1_multiplier =
        static_cast<double>(input1->params.scale) / twice_max_input_scale;
    const double real_input2_multiplier =
        static_cast<double>(input2->params.scale) / twice_max_input_scale;
    const double real_output_multiplier =
        twice_max_input_scale /
        ((1 << data->left_shift) * static_cast<double>(output->params.scale));

    QuantizeMultiplierSmallerThanOneExp(
        real_input1_multiplier, &data->input1_multiplier, &data->input1_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_input2_multiplier, &data->input2_multiplier, &data->input2_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_output_multiplier, &data->output_multiplier, &data->output_shift);

    TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
        context, params->activation, output, &data->output_activation_min,
        &data->output_activation_max));
  } else if (output->type == kTfLiteFloat32) {
    CalculateActivationRange(params->activation,
                             &data->output_activation_min_f32,
                             &data->output_activation_max_f32);
  }

  return kTfLiteOk;
}

void* Init(TfLiteContext* context, const char* buffer, size_t length) {
  TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr);
  return context->AllocatePersistentBuffer(context, sizeof(OpData));
}

TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
  TFLITE_DCHECK(node->user_data != nullptr);
  TFLITE_DCHECK(node->builtin_data != nullptr);

  const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
  TF_LITE_ENSURE(context, input1 != nullptr);
  const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
  TF_LITE_ENSURE(context, input2 != nullptr);
  TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
  TF_LITE_ENSURE(context, output != nullptr);

  OpData* data = static_cast<OpData*>(node->user_data);
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  TF_LITE_ENSURE_STATUS(
      CalculateOpDataAdd(context, params, input1, input2, output, data));

  return kTfLiteOk;
}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  TFLITE_DCHECK(node->user_data != nullptr);
  const OpData* data = static_cast<const OpData*>(node->user_data);

  const TfLiteEvalTensor* input1 =
      tflite::micro::GetEvalInput(context, node, kInputTensor1);
  const TfLiteEvalTensor* input2 =
      tflite::micro::GetEvalInput(context, node, kInputTensor2);
  TfLiteEvalTensor* output =
      tflite::micro::GetEvalOutput(context, node, kOutputTensor);

  long long start_time = esp_timer_get_time();

  if (output->type == kTfLiteFloat32) {
    EvalAdd(context, node, params, data, input1, input2, output);
  } else if (output->type == kTfLiteInt8 || output->type == kTfLiteInt16) {
    TF_LITE_ENSURE_OK(context, EvalAddQuantized(context, node, params, data,
                                                input1, input2, output));
  } else {
      TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.",
                       TfLiteTypeGetName(output->type), output->type);
    return kTfLiteError;
  }
  add_total_time += esp_timer_get_time() - start_time;

  return kTfLiteOk;
}

}  // namespace add

TfLiteRegistration Register_ADD() {
  return {/*init=*/add::Init,
          /*free=*/nullptr,
          /*prepare=*/add::Prepare,
          /*invoke=*/add::Eval,
          /*profiling_string=*/nullptr,
          /*builtin_code=*/0,
          /*custom_name=*/nullptr,
          /*version=*/0};
}

}  // namespace micro
}  // namespace ops
}  // namespace tflite

#else
/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/add.h"

#include "edge-impulse-sdk/tensorflow/lite/c/builtin_op_data.h"
#include "edge-impulse-sdk/tensorflow/lite/c/common.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/quantization_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/integer_ops/add.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/op_macros.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/memory_helpers.h"

namespace tflite {
namespace ops {
namespace micro {
namespace add {

constexpr int kInputTensor1 = 0;
constexpr int kInputTensor2 = 1;
constexpr int kOutputTensor = 0;

struct OpData {
  bool requires_broadcast;

  // These fields are used in both the general 8-bit -> 8bit quantized path,
  // and the special 16-bit -> 16bit quantized path
  int input1_shift;
  int input2_shift;
  int32_t output_activation_min;
  int32_t output_activation_max;

  // These fields are used only in the general 8-bit -> 8bit quantized path
  int32_t input1_multiplier;
  int32_t input2_multiplier;
  int32_t output_multiplier;
  int output_shift;
  int left_shift;
  int32_t input1_offset;
  int32_t input2_offset;
  int32_t output_offset;

  // Used only for float evals:
  float output_activation_min_f32;
  float output_activation_max_f32;
};

TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteAddParams* params,
                             const TfLiteTensor* input1,
                             const TfLiteTensor* input2, TfLiteTensor* output,
                             OpData* data) {
  data->requires_broadcast = !HaveSameShapes(input1, input2);

  if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
    // 8bit -> 8bit general quantized path, with general rescalings
    data->input1_offset = -input1->params.zero_point;
    data->input2_offset = -input2->params.zero_point;
    data->output_offset = output->params.zero_point;
    data->left_shift = 20;
    const double twice_max_input_scale =
        2 * static_cast<double>(
                std::max(input1->params.scale, input2->params.scale));
    const double real_input1_multiplier =
        static_cast<double>(input1->params.scale) / twice_max_input_scale;
    const double real_input2_multiplier =
        static_cast<double>(input2->params.scale) / twice_max_input_scale;
    const double real_output_multiplier =
        twice_max_input_scale /
        ((1 << data->left_shift) * static_cast<double>(output->params.scale));

    QuantizeMultiplierSmallerThanOneExp(
        real_input1_multiplier, &data->input1_multiplier, &data->input1_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_input2_multiplier, &data->input2_multiplier, &data->input2_shift);

    QuantizeMultiplierSmallerThanOneExp(
        real_output_multiplier, &data->output_multiplier, &data->output_shift);

    TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
        context, params->activation, output, &data->output_activation_min,
        &data->output_activation_max));
  } else if (output->type == kTfLiteFloat32) {
    CalculateActivationRange(params->activation,
                             &data->output_activation_min_f32,
                             &data->output_activation_max_f32);
  }

  return kTfLiteOk;
}

void EvalAdd(TfLiteContext* context, TfLiteNode* node, TfLiteAddParams* params,
             const OpData* data, const TfLiteEvalTensor* input1,
             const TfLiteEvalTensor* input2, TfLiteEvalTensor* output) {
  tflite::ArithmeticParams op_params;
  SetActivationParams(data->output_activation_min_f32,
                      data->output_activation_max_f32, &op_params);
  if (data->requires_broadcast) {
    reference_ops::BroadcastAdd4DSlow(
        op_params, tflite::micro::GetTensorShape(input1),
        tflite::micro::GetTensorData<float>(input1),
        tflite::micro::GetTensorShape(input2),
        tflite::micro::GetTensorData<float>(input2),
        tflite::micro::GetTensorShape(output),
        tflite::micro::GetTensorData<float>(output));
  } else {
    reference_ops::Add(op_params, tflite::micro::GetTensorShape(input1),
                       tflite::micro::GetTensorData<float>(input1),
                       tflite::micro::GetTensorShape(input2),
                       tflite::micro::GetTensorData<float>(input2),
                       tflite::micro::GetTensorShape(output),
                       tflite::micro::GetTensorData<float>(output));
  }
}

TfLiteStatus EvalAddQuantized(TfLiteContext* context, TfLiteNode* node,
                              TfLiteAddParams* params, const OpData* data,
                              const TfLiteEvalTensor* input1,
                              const TfLiteEvalTensor* input2,
                              TfLiteEvalTensor* output) {
  if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
    tflite::ArithmeticParams op_params;
    op_params.left_shift = data->left_shift;
    op_params.input1_offset = data->input1_offset;
    op_params.input1_multiplier = data->input1_multiplier;
    op_params.input1_shift = data->input1_shift;
    op_params.input2_offset = data->input2_offset;
    op_params.input2_multiplier = data->input2_multiplier;
    op_params.input2_shift = data->input2_shift;
    op_params.output_offset = data->output_offset;
    op_params.output_multiplier = data->output_multiplier;
    op_params.output_shift = data->output_shift;
    SetActivationParams(data->output_activation_min,
                        data->output_activation_max, &op_params);
    bool need_broadcast = reference_ops::ProcessBroadcastShapes(
        tflite::micro::GetTensorShape(input1),
        tflite::micro::GetTensorShape(input2), &op_params);
    if (output->type == kTfLiteInt8) {
      if (need_broadcast) {
        reference_integer_ops::BroadcastAdd4DSlow(
            op_params, tflite::micro::GetTensorShape(input1),
            tflite::micro::GetTensorData<int8_t>(input1),
            tflite::micro::GetTensorShape(input2),
            tflite::micro::GetTensorData<int8_t>(input2),
            tflite::micro::GetTensorShape(output),
            tflite::micro::GetTensorData<int8_t>(output));
      } else {
        reference_integer_ops::Add(
            op_params, tflite::micro::GetTensorShape(input1),
            tflite::micro::GetTensorData<int8_t>(input1),
            tflite::micro::GetTensorShape(input2),
            tflite::micro::GetTensorData<int8_t>(input2),
            tflite::micro::GetTensorShape(output),
            tflite::micro::GetTensorData<int8_t>(output));
      }
    } else {
      if (need_broadcast) {
        reference_ops::BroadcastAdd4DSlow(
            op_params, tflite::micro::GetTensorShape(input1),
            tflite::micro::GetTensorData<uint8_t>(input1),
            tflite::micro::GetTensorShape(input2),
            tflite::micro::GetTensorData<uint8_t>(input2),
            tflite::micro::GetTensorShape(output),
            tflite::micro::GetTensorData<uint8_t>(output));
      } else {
        reference_ops::Add(op_params, tflite::micro::GetTensorShape(input1),
                           tflite::micro::GetTensorData<uint8_t>(input1),
                           tflite::micro::GetTensorShape(input2),
                           tflite::micro::GetTensorData<uint8_t>(input2),
                           tflite::micro::GetTensorShape(output),
                           tflite::micro::GetTensorData<uint8_t>(output));
      }
    }
  }

  return kTfLiteOk;
}

void* Init(TfLiteContext* context, const char* buffer, size_t length) {
  TFLITE_DCHECK(context->AllocatePersistentBuffer != nullptr);
  return context->AllocatePersistentBuffer(context, sizeof(OpData));
}

TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
  TFLITE_DCHECK(node->user_data != nullptr);
  TFLITE_DCHECK(node->builtin_data != nullptr);

  const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1);
  TF_LITE_ENSURE(context, input1 != nullptr);
  const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2);
  TF_LITE_ENSURE(context, input2 != nullptr);
  TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
  TF_LITE_ENSURE(context, output != nullptr);

  OpData* data = static_cast<OpData*>(node->user_data);
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  TF_LITE_ENSURE_STATUS(
      CalculateOpData(context, params, input1, input2, output, data));

  return kTfLiteOk;
}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  auto* params = reinterpret_cast<TfLiteAddParams*>(node->builtin_data);

  TFLITE_DCHECK(node->user_data != nullptr);
  const OpData* data = static_cast<const OpData*>(node->user_data);

  const TfLiteEvalTensor* input1 =
      tflite::micro::GetEvalInput(context, node, kInputTensor1);
  const TfLiteEvalTensor* input2 =
      tflite::micro::GetEvalInput(context, node, kInputTensor2);
  TfLiteEvalTensor* output =
      tflite::micro::GetEvalOutput(context, node, kOutputTensor);

  if (output->type == kTfLiteFloat32) {
    EvalAdd(context, node, params, data, input1, input2, output);
  } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) {
    TF_LITE_ENSURE_OK(context, EvalAddQuantized(context, node, params, data,
                                                input1, input2, output));
  } else {
    TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.",
                       TfLiteTypeGetName(output->type), output->type);
    return kTfLiteError;
  }

  return kTfLiteOk;
}

}  // namespace add

TfLiteRegistration Register_ADD() {
  return {/*init=*/add::Init,
          /*free=*/nullptr,
          /*prepare=*/add::Prepare,
          /*invoke=*/add::Eval,
          /*profiling_string=*/nullptr,
          /*builtin_code=*/0,
          /*custom_name=*/nullptr,
          /*version=*/0};
}

}  // namespace micro
}  // namespace ops
}  // namespace tflite

#endif