onert-micro/luci-interpreter/src/kernels/Sub.cpp

   1 /*
   2  * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
   3  * Copyright 2019 The TensorFlow Authors. All Rights Reserved.
   4  *
   5  * Licensed under the Apache License, Version 2.0 (the "License");
   6  * you may not use this file except in compliance with the License.
   7  * You may obtain a copy of the License at
   8  *
   9  *    http://www.apache.org/licenses/LICENSE-2.0
  10  *
  11  * Unless required by applicable law or agreed to in writing, software
  12  * distributed under the License is distributed on an "AS IS" BASIS,
  13  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  14  * See the License for the specific language governing permissions and
  15  * limitations under the License.
  16  */
  17
  18 #include "kernels/Sub.h"
  19 #include "kernels/Utils.h"
  20
  21 #include "PALSub.h"
  22
  23 #include <tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h>
  24
  25 namespace luci_interpreter
  26 {
  27 namespace kernels
  28 {
  29
  30 Sub::Sub(const Tensor *input1, const Tensor *input2, Tensor *output, const SubParams &params)
  31   : KernelWithParams<SubParams>({input1, input2}, {output}, params)
  32 {
  33 }
  34
  35 void Sub::configure()
  36 {
  37   LUCI_INTERPRETER_CHECK(!(input1()->element_type() != input2()->element_type()))
  38   LUCI_INTERPRETER_CHECK(!(input1()->element_type() != output()->element_type()))
  39   // TODO: enable it only if kernel with dynamic shapes
  40   output()->resize(calculateShapeForBroadcast(input1()->shape(), input2()->shape()));
  41 }
  42
  43 void Sub::execute() const
  44 {
  45   switch (input1()->element_type())
  46   {
  47     case DataType::FLOAT32:
  48       evalFloat();
  49       break;
  50     case DataType::S64:
  51       evalInteger<int64_t>();
  52       break;
  53     case DataType::S32:
  54       evalInteger<int32_t>();
  55       break;
  56     case DataType::U8:
  57       evalQuantized();
  58       break;
  59     default:
  60       assert(false && "Unsupported type.");
  61   }
  62 }
  63
  64 void Sub::evalFloat() const
  65 {
  66   tflite::ArithmeticParams params{};
  67   fillArithmeticActivationRange<float>(params, _params.activation);
  68
  69   const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
  70     getTensorShape(input1()), getTensorShape(input2()), &params);
  71
  72   if (need_broadcast)
  73   {
  74     tflite::reference_ops::BroadcastSubSlow(
  75       params, getTensorShape(input1()), getTensorData<float>(input1()), getTensorShape(input2()),
  76       getTensorData<float>(input2()), getTensorShape(output()), getTensorData<float>(output()));
  77   }
  78   else
  79   {
  80     luci_interpreter_pal::Sub(params, getTensorShape(input1()), getTensorData<float>(input1()),
  81                               getTensorShape(input2()), getTensorData<float>(input2()),
  82                               getTensorShape(output()), getTensorData<float>(output()));
  83   }
  84 }
  85
  86 template <typename T> void Sub::evalInteger() const
  87 {
  88   tflite::ArithmeticParams params{};
  89   fillArithmeticActivationRange<T>(params, _params.activation);
  90
  91   const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
  92     getTensorShape(input1()), getTensorShape(input2()), &params);
  93
  94   if (need_broadcast)
  95   {
  96     tflite::reference_ops::BroadcastSubSlow(
  97       params, getTensorShape(input1()), getTensorData<T>(input1()), getTensorShape(input2()),
  98       getTensorData<T>(input2()), getTensorShape(output()), getTensorData<T>(output()));
  99   }
 100   else
 101   {
 102     tflite::reference_ops::Sub(params, getTensorShape(input1()), getTensorData<T>(input1()),
 103                                getTensorShape(input2()), getTensorData<T>(input2()),
 104                                getTensorShape(output()), getTensorData<T>(output()));
 105   }
 106 }
 107
 108 void Sub::evalQuantized() const
 109 {
 110   const auto input1_scale = static_cast<double>(input1()->scale());
 111   const auto input2_scale = static_cast<double>(input2()->scale());
 112   const auto output_scale = static_cast<double>(output()->scale());
 113
 114   const int left_shift = 20;
 115   const double twice_max_input_scale = 2 * std::max(input1_scale, input2_scale);
 116   const double real_input1_multiplier = input1_scale / twice_max_input_scale;
 117   const double real_input2_multiplier = input2_scale / twice_max_input_scale;
 118   const double real_output_multiplier = twice_max_input_scale / ((1 << left_shift) * output_scale);
 119
 120   int32_t input1_multiplier{}, input2_multiplier{}, output_multiplier{};
 121   int input1_shift{}, input2_shift{}, output_shift{};
 122   quantizeMultiplierSmallerThanOneExp(real_input1_multiplier, &input1_multiplier, &input1_shift);
 123   quantizeMultiplierSmallerThanOneExp(real_input2_multiplier, &input2_multiplier, &input2_shift);
 124   quantizeMultiplierSmallerThanOneExp(real_output_multiplier, &output_multiplier, &output_shift);
 125
 126   int32_t activation_min{};
 127   int32_t activation_max{};
 128   calculateActivationRangeQuantized(_params.activation, output(), &activation_min, &activation_max);
 129
 130   tflite::ArithmeticParams params{};
 131   params.left_shift = left_shift;
 132   // The kernel expects inputs' zero points to be negated.
 133   params.input1_offset = -input1()->zero_point(); // Note the '-'.
 134   params.input1_multiplier = input1_multiplier;
 135   params.input1_shift = input1_shift;
 136   params.input2_offset = -input2()->zero_point(); // Note the '-'.
 137   params.input2_multiplier = input2_multiplier;
 138   params.input2_shift = input2_shift;
 139   params.output_offset = output()->zero_point();
 140   params.output_multiplier = output_multiplier;
 141   params.output_shift = output_shift;
 142   params.quantized_activation_min = activation_min;
 143   params.quantized_activation_max = activation_max;
 144
 145   const bool need_broadcast = tflite::reference_ops::ProcessBroadcastShapes(
 146     getTensorShape(input1()), getTensorShape(input2()), &params);
 147
 148   if (need_broadcast)
 149   {
 150     tflite::reference_ops::BroadcastSubSlow(
 151       params, getTensorShape(input1()), getTensorData<uint8_t>(input1()), getTensorShape(input2()),
 152       getTensorData<uint8_t>(input2()), getTensorShape(output()), getTensorData<uint8_t>(output()));
 153   }
 154   else
 155   {
 156     tflite::reference_ops::Sub(params, getTensorShape(input1()), getTensorData<uint8_t>(input1()),
 157                                getTensorShape(input2()), getTensorData<uint8_t>(input2()),
 158                                getTensorShape(output()), getTensorData<uint8_t>(output()));
 159   }
 160 }
 161
 162 } // namespace kernels
 163 } // namespace luci_interpreter