/* 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 "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/kernel_util.h"
#include "tensorflow/lite/kernels/op_macros.h"
#include "tensorflow/lite/micro/kernels/softmax.h"

namespace tflite {

namespace {
// Softmax parameter data that persists in user_data
const int kInt16LUTArraySize = 513;

TfLiteStatus CalculateSoftmaxParams(TfLiteContext* context,
                                    const TfLiteTensor* input,
                                    TfLiteTensor* output,
                                    const TfLiteSoftmaxParams* params,
                                    SoftmaxParams* op_data) {
  if (input->type == kTfLiteInt8 || input->type == kTfLiteInt16) {
    if (input->type == kTfLiteInt16) {
      TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);
      TF_LITE_ENSURE_NEAR(context, output->params.scale, 1.f / 32768,
                          (0.001f * 1.f / 32768));
    } else {  // input->type == kTfLiteInt8
      TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteInt8);
      if (output->type == kTfLiteInt16) {
        TF_LITE_ENSURE_EQ(context, output->params.zero_point, -32768);
        TF_LITE_ENSURE_NEAR(context, output->params.scale, 1.f / 65536,
                            (0.001f * 1.f / 65536));
      } else {  // output->type == kTfLiteint8
        TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteInt8);
        TF_LITE_ENSURE_EQ(context, output->params.zero_point, -128);
        TF_LITE_ENSURE(context, output->params.scale == 1.f / 256);
      }
    }

    static const int kScaledDiffIntegerBits = 5;

    // Calculate input_multiplier and input_left_shift
    if (input->type == kTfLiteInt16) {
      int input_left_shift;
      double input_scale_beta_rescale =
          static_cast<double>(input->params.scale) *
          static_cast<double>(params->beta) /
          (10.0 / 65535.0);  // scale the input_diff such that [-65535, 0]
                             // correspond to [-10.0, 0.0]
      QuantizeMultiplier(input_scale_beta_rescale, &op_data->input_multiplier,
                         &input_left_shift);
      op_data->input_left_shift = input_left_shift;
    } else {
      int input_left_shift;
      tflite::PreprocessSoftmaxScaling(
          static_cast<double>(params->beta),
          static_cast<double>(input->params.scale), kScaledDiffIntegerBits,
          &op_data->input_multiplier, &input_left_shift);
      op_data->input_left_shift = input_left_shift;
      op_data->diff_min =
          -1.0 * tflite::CalculateInputRadius(kScaledDiffIntegerBits,
                                              op_data->input_left_shift);
    }
  } else {
    TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteFloat32);
    TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteFloat32);
    op_data->beta = static_cast<double>(params->beta);
  }
  return kTfLiteOk;
}

}  // namespace

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

TfLiteStatus SoftmaxPrepare(TfLiteContext* context, TfLiteNode* node) {
  TF_LITE_ENSURE_EQ(context, NumInputs(node), 1);
  TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);
  const TfLiteTensor* input = GetInput(context, node, 0);
  TF_LITE_ENSURE(context, input != nullptr);
  TF_LITE_ENSURE(context, NumDimensions(input) >= 1);
  TfLiteTensor* output = GetOutput(context, node, 0);
  TF_LITE_ENSURE(context, output != nullptr);

  TF_LITE_ENSURE(context, node->user_data != nullptr);
  SoftmaxParams* op_data = static_cast<SoftmaxParams*>(node->user_data);
  // Only allocate LUTs for KTfLiteInt16 data type
  if (input->type == kTfLiteInt16) {
    void* raw_exp_lut = context->AllocatePersistentBuffer(
        context, sizeof(int16_t) * kInt16LUTArraySize);
    TF_LITE_ENSURE(context, raw_exp_lut != nullptr);
    op_data->exp_lut = reinterpret_cast<int16_t*>(raw_exp_lut);
    void* one_over_one_plus_x_lut = context->AllocatePersistentBuffer(
        context, sizeof(int16_t) * kInt16LUTArraySize);
    TF_LITE_ENSURE(context, one_over_one_plus_x_lut != nullptr);
    op_data->one_over_one_plus_x_lut =
        reinterpret_cast<int16_t*>(one_over_one_plus_x_lut);
  }

  if (output->type == kTfLiteInt16) {
    TF_LITE_ENSURE(context,
                   input->type == kTfLiteInt8 || input->type == kTfLiteInt16);
  } else {
    TF_LITE_ENSURE_EQ(context, input->type, output->type);
  }

  // Populate LUT if required
  if (input->type == kTfLiteInt16) {
    TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0);
    // exp LUT only used on negative values
    // we consider exp(-10.0) is insignificant to accumulation
    gen_lut([](float value) { return std::exp(value); }, -10.0f, 0.0f,
            op_data->exp_lut, kInt16LUTArraySize);
    gen_lut([](float value) { return 1.0f / (1.0f + value); }, 0.0f, 1.0f,
            op_data->one_over_one_plus_x_lut, kInt16LUTArraySize);
    op_data->zero_point = output->params.zero_point;
    op_data->scale = output->params.scale;
  }

  auto* params = static_cast<TfLiteSoftmaxParams*>(node->builtin_data);
  return CalculateSoftmaxParams(context, input, output, params, op_data);
}

}  // namespace tflite