/* 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.
==============================================================================*/

/* Copyright 2020 The Qualcomm Innovation Center, Inc. All Rights Reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted (subject to the limitations in the disclaimer
below) provided that the following conditions are met:

* Redistributions of source code must retain the above copyright notice,
  this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
  this list of conditions and the following disclaimer in the documentation
  and/or other materials provided with the distribution.
* Neither the name of Qualcomm Innovation Center, Inc. nor the names of its
  contributors may be used to endorse or promote products derived from this
  software without specific prior written permission.

NO EXPRESS OR IMPLIED LICENSES TO ANY PARTY'S PATENT RIGHTS ARE GRANTED BY
THIS LICENSE. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT
NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER
OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
==============================================================================*/

#include "tensorflow/lite/micro/kernels/fully_connected.h"

#include "tensorflow/lite/c/builtin_op_data.h"
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/quantization_util.h"
#include "tensorflow/lite/kernels/internal/reference/fully_connected.h"
#include "tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h"
#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "tensorflow/lite/kernels/kernel_util.h"
#include "tensorflow/lite/micro/kernels/kernel_util.h"
#include "third_party/hexagon/hexagon_tflm_translation_fully_connected.h"

namespace tflite {
namespace {

// Input tensors.
constexpr int kInputTensor = 0;
constexpr int kWeightsTensor = 1;
constexpr int kBiasTensor = 2;
// Output tensor.
constexpr int kOutputTensor = 0;

struct OpData {
  // The scaling factor from input to output (aka the 'real multiplier') can
  // be represented as a fixed point multiplier plus a left shift.
  int32_t output_multiplier;
  int output_shift;
  // The range of the fused activation layer. For example for kNone and
  // uint8_t these would be 0 and 255.
  int32_t output_activation_min;
  int32_t output_activation_max;
  // The index of the temporary tensor where the quantized inputs are cached.
  int input_quantized_index;
  // Cached zero point values of tensors.
  int32_t input_zero_point;
  int32_t filter_zero_point;
  int32_t output_zero_point;

  void* hexagon_data;
};

TfLiteStatus CalculateOpData(TfLiteContext* context,
                             TfLiteFusedActivation activation,
                             TfLiteType data_type, const TfLiteTensor* input,
                             const TfLiteTensor* filter,
                             const TfLiteTensor* bias, TfLiteTensor* output,
                             OpData* data) {
  TfLiteStatus status = kTfLiteOk;
  if (data_type != kTfLiteFloat32) {
    double real_multiplier = 0.0;
    TF_LITE_ENSURE_STATUS(GetQuantizedConvolutionMultipler(
        context, input, filter, bias, output, &real_multiplier));
    int exponent;
    QuantizeMultiplier(real_multiplier, &data->output_multiplier, &exponent);
    data->output_shift = -exponent;
    TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized(
        context, activation, output, &data->output_activation_min,
        &data->output_activation_max));

    data->input_zero_point = input->params.zero_point;
    data->filter_zero_point = filter->params.zero_point;
    data->output_zero_point = output->params.zero_point;
  }
  return status;
}

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

  if (data == nullptr) {
    return nullptr;
  }
  OpData* opdata = static_cast<OpData*>(data);
  opdata->hexagon_data =
      tflite::hexagon_fully_connected::HexagonInit(context, buffer, length);

  return data;
}

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

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

  const TfLiteTensor* input = GetInput(context, node, kInputTensor);
  TF_LITE_ENSURE(context, input != nullptr);
  const TfLiteTensor* filter = GetInput(context, node, kWeightsTensor);
  TF_LITE_ENSURE(context, filter != nullptr);
  const TfLiteTensor* bias = GetOptionalInputTensor(context, node, kBiasTensor);
  TfLiteTensor* output = GetOutput(context, node, kOutputTensor);
  TF_LITE_ENSURE(context, output != nullptr);

  TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type);
  TF_LITE_ENSURE_MSG(context, input->type == filter->type,
                     "Hybrid models are not supported on TFLite Micro.");

  tflite::hexagon_fully_connected::HexagonOptimizationEvaluation(context, node);

  if (tflite::hexagon_fully_connected::HexagonOptimizable(context, node)) {
    return tflite::hexagon_fully_connected::HexagonPrepare(context, node);
  } else {
    return CalculateOpData(context, params->activation, input->type, input,
                           filter, bias, output, data);
  }
}

TfLiteStatus EvalQuantizedInt8(TfLiteContext* context, TfLiteNode* node,
                               const OpData& data,
                               const TfLiteEvalTensor* input,
                               const TfLiteEvalTensor* filter,
                               const TfLiteEvalTensor* bias,
                               TfLiteEvalTensor* output) {
  tflite::FullyConnectedParams op_params;
  op_params.input_offset = -data.input_zero_point;
  op_params.weights_offset = -data.filter_zero_point;
  op_params.output_offset = data.output_zero_point;
  op_params.output_multiplier = data.output_multiplier;
  // TODO(b/138810107): Figure out whether output shift should be inverted
  op_params.output_shift = -data.output_shift;
  op_params.quantized_activation_min = data.output_activation_min;
  op_params.quantized_activation_max = data.output_activation_max;

  reference_integer_ops::FullyConnected(
      op_params, tflite::micro::GetTensorShape(input),
      tflite::micro::GetTensorData<int8_t>(input),
      tflite::micro::GetTensorShape(filter),
      tflite::micro::GetTensorData<int8_t>(filter),
      tflite::micro::GetTensorShape(bias),
      tflite::micro::GetTensorData<int32_t>(bias),
      tflite::micro::GetTensorShape(output),
      tflite::micro::GetTensorData<int8_t>(output));

  return kTfLiteOk;
}

TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node,
                           const OpData& data, const TfLiteEvalTensor* input,
                           const TfLiteEvalTensor* filter,
                           const TfLiteEvalTensor* bias,
                           TfLiteEvalTensor* output) {
  const int32_t input_offset = -data.input_zero_point;
  const int32_t filter_offset = -data.filter_zero_point;
  const int32_t output_offset = data.output_zero_point;

  tflite::FullyConnectedParams op_params;
  op_params.input_offset = input_offset;
  op_params.weights_offset = filter_offset;
  op_params.output_offset = output_offset;
  op_params.output_multiplier = data.output_multiplier;
  // Legacy ops used mixed left and right shifts. Now all are +ve-means-left.
  op_params.output_shift = -data.output_shift;
  op_params.quantized_activation_min = data.output_activation_min;
  op_params.quantized_activation_max = data.output_activation_max;

#define TF_LITE_FULLY_CONNECTED(output_data_type)      \
  reference_ops::FullyConnected(                       \
      op_params, tflite::micro::GetTensorShape(input), \
      tflite::micro::GetTensorData<uint8_t>(input),    \
      tflite::micro::GetTensorShape(filter),           \
      tflite::micro::GetTensorData<uint8_t>(filter),   \
      tflite::micro::GetTensorShape(bias),             \
      tflite::micro::GetTensorData<int32_t>(bias),     \
      tflite::micro::GetTensorShape(output),           \
      tflite::micro::GetTensorData<output_data_type>(output))
  switch (output->type) {
    case kTfLiteUInt8:
      TF_LITE_FULLY_CONNECTED(uint8_t);
      break;
    case kTfLiteInt16:
      TF_LITE_FULLY_CONNECTED(int16_t);
      break;
    default:
      TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.",
                         TfLiteTypeGetName(output->type), output->type);
      return kTfLiteError;
  }

  return kTfLiteOk;
}

TfLiteStatus EvalFloat(TfLiteContext* context, TfLiteNode* node,
                       TfLiteFusedActivation activation,
                       const TfLiteEvalTensor* input,
                       const TfLiteEvalTensor* filter,
                       const TfLiteEvalTensor* bias, TfLiteEvalTensor* output) {
  float output_activation_min, output_activation_max;
  CalculateActivationRange(activation, &output_activation_min,
                           &output_activation_max);
  tflite::FullyConnectedParams op_params;
  op_params.float_activation_min = output_activation_min;
  op_params.float_activation_max = output_activation_max;
  tflite::reference_ops::FullyConnected(
      op_params, tflite::micro::GetTensorShape(input),
      tflite::micro::GetTensorData<float>(input),
      tflite::micro::GetTensorShape(filter),
      tflite::micro::GetTensorData<float>(filter),
      tflite::micro::GetTensorShape(bias),
      tflite::micro::GetTensorData<float>(bias),
      tflite::micro::GetTensorShape(output),
      tflite::micro::GetTensorData<float>(output));
  return kTfLiteOk;
}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  TFLITE_DCHECK(node->builtin_data != nullptr);
  const auto* params =
      static_cast<const TfLiteFullyConnectedParams*>(node->builtin_data);

  const TfLiteEvalTensor* input =
      tflite::micro::GetEvalInput(context, node, kInputTensor);
  const TfLiteEvalTensor* filter =
      tflite::micro::GetEvalInput(context, node, kWeightsTensor);
  const TfLiteEvalTensor* bias =
      tflite::micro::GetEvalInput(context, node, kBiasTensor);
  TfLiteEvalTensor* output =
      tflite::micro::GetEvalOutput(context, node, kOutputTensor);

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

  // Checks in Prepare ensure input, output and filter types are all the same.
  switch (input->type) {
    case kTfLiteFloat32:
      return EvalFloat(context, node, params->activation, input, filter, bias,
                       output);

    case kTfLiteInt8:
      if (tflite::hexagon_fully_connected::HexagonOptimizable(context, node)) {
        return tflite::hexagon_fully_connected::HexagonEvalQuantizedInt8(
            context, node, node->user_data, input, filter, bias, output);
      } else {
        return EvalQuantizedInt8(context, node, data, input, filter, bias,
                                 output);
      }

    case kTfLiteUInt8:
      return EvalQuantized(context, node, data, input, filter, bias, output);

    default:
      TF_LITE_KERNEL_LOG(context, "Type %s (%d) not supported.",
                         TfLiteTypeGetName(input->type), input->type);
      return kTfLiteError;
  }
  return kTfLiteOk;
}

}  // namespace

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

}  // namespace tflite