compiler/luci-interpreter/src/kernels/TransposeConv.cpp

   1 /*
   2  * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
   3  * Copyright 2017 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/TransposeConv.h"
  19
  20 #include "kernels/Utils.h"
  21
  22 #include <tensorflow/lite/kernels/internal/reference/reference_ops.h>
  23
  24 #include <stdexcept>
  25
  26 namespace luci_interpreter
  27 {
  28
  29 namespace kernels
  30 {
  31
  32 TransposeConv::TransposeConv(const Tensor *output_shape, const Tensor *filter, const Tensor *input,
  33                              const Tensor *bias, Tensor *output, const TransposeConvParams &params)
  34     : KernelWithParams<TransposeConvParams>({output_shape, filter, input, bias}, {output}, params)
  35 {
  36 }
  37
  38 TransposeConv::~TransposeConv()
  39 {
  40   // Define destructor here, to delete vector of qunatized multipliers properly
  41 }
  42
  43 void TransposeConv::configure()
  44 {
  45   assert(output_shape()->shape().num_dims() == 1);
  46   assert(input()->shape().num_dims() == 4);
  47   assert(filter()->shape().num_dims() == 4);
  48   assert(input()->element_type() == DataType::FLOAT32 || input()->element_type() == DataType::U8 ||
  49          input()->element_type() == DataType::S16);
  50   assert(input()->element_type() == output()->element_type());
  51   assert(input()->shape().dim(3) == filter()->shape().dim(3));
  52
  53   const int num_dims = output_shape()->shape().dim(0);
  54   Shape out_shape(num_dims);
  55   const auto *shape_data = getTensorData<int32_t>(output_shape());
  56   for (int i = 0; i < num_dims; i++)
  57     out_shape.dim(i) = shape_data[i];
  58   output()->resize(out_shape);
  59
  60   const int32_t filter_height = filter()->shape().dim(1);
  61   const int32_t filter_width = filter()->shape().dim(2);
  62   const int32_t output_height = out_shape.dim(1);
  63   const int32_t output_width = out_shape.dim(2);
  64
  65   const int32_t unused_output_height =
  66       computeOutputSize(params().padding, output_height, filter_height, params().stride_height, 1);
  67   const int32_t unused_output_width =
  68       computeOutputSize(params().padding, output_width, filter_width, params().stride_width, 1);
  69
  70   _padding_height =
  71       computePadding(params().stride_height, 1, output_height, filter_height, unused_output_height);
  72   _padding_width =
  73       computePadding(params().stride_width, 1, output_width, filter_width, unused_output_width);
  74
  75   if (input()->element_type() == DataType::U8 || input()->element_type() == DataType::S16)
  76   {
  77     DataType scratch_data_type =
  78         input()->element_type() == DataType::S16 ? DataType::S64 : DataType::S32;
  79     _scratch_tensor =
  80         std::make_unique<Tensor>(scratch_data_type, output()->shape(), AffineQuantization{}, "");
  81     const std::vector<double> real_multipliers =
  82         getQuantizedConvolutionMultiplers(input()->scale(), filter()->scales(), output()->scale());
  83
  84     _quant_multipliers = quantizeMultipliers(real_multipliers);
  85   }
  86 }
  87
  88 void TransposeConv::execute() const
  89 {
  90   switch (input()->element_type())
  91   {
  92     case DataType::FLOAT32:
  93       evalFloat();
  94       break;
  95     case DataType::U8:
  96       evalQuantized();
  97       break;
  98     case DataType::S16:
  99       evalQuantizedS16();
 100       break;
 101     default:
 102       throw std::runtime_error("Unsupported type.");
 103   }
 104   if (!!_scratch_tensor)
 105     _scratch_tensor->deallocate();
 106 }
 107
 108 void TransposeConv::evalFloat() const
 109 {
 110   tflite::ConvParams op_params{};
 111   op_params.padding_type = tflite::PaddingType::kSame;
 112   op_params.padding_values.height = _padding_height;
 113   op_params.padding_values.width = _padding_width;
 114   op_params.stride_height = params().stride_height;
 115   op_params.stride_width = params().stride_width;
 116   tflite::reference_ops::TransposeConv(op_params,                                                //
 117                                        getTensorShape(input()), getTensorData<float>(input()),   //
 118                                        getTensorShape(filter()), getTensorData<float>(filter()), //
 119                                        getTensorShape(bias()), getTensorData<float>(bias()),     //
 120                                        getTensorShape(output()), getTensorData<float>(output()), //
 121                                        tflite::RuntimeShape(), nullptr);
 122 }
 123
 124 void TransposeConv::evalQuantized() const
 125 {
 126   tflite::ConvParams op_params{};
 127   op_params.padding_type = tflite::PaddingType::kSame;
 128   op_params.padding_values.height = _padding_height;
 129   op_params.padding_values.width = _padding_width;
 130   op_params.stride_height = params().stride_height;
 131   op_params.stride_width = params().stride_width;
 132   // The kernel expects input and filter zero points to be negated.
 133   op_params.input_offset = -input()->zero_point();    // Note the '-'.
 134   op_params.weights_offset = -filter()->zero_point(); // Note the '-'.
 135   op_params.output_offset = output()->zero_point();
 136   op_params.output_multiplier = _quant_multipliers[0].multiplier;
 137   op_params.output_shift = _quant_multipliers[0].shift;
 138   op_params.quantized_activation_min = std::numeric_limits<uint8_t>::min();
 139   op_params.quantized_activation_max = std::numeric_limits<uint8_t>::max();
 140
 141   tflite::reference_ops::TransposeConv(op_params,                                                //
 142                                        getTensorShape(input()), getTensorData<uint8>(input()),   //
 143                                        getTensorShape(filter()), getTensorData<uint8>(filter()), //
 144                                        getTensorShape(bias()), getTensorData<int32_t>(bias()),   //
 145                                        getTensorShape(output()), getTensorData<uint8>(output()), //
 146                                        tflite::RuntimeShape(), nullptr,                          //
 147                                        getTensorData<int32_t>(_scratch_tensor.get()));
 148 }
 149
 150 void TransposeConv::evalQuantizedS16() const
 151 {
 152   const auto *input_data = getTensorData<int16_t>(input());
 153   const auto *filter_data = getTensorData<int16_t>(filter());
 154   const auto *bias_data = getTensorData<int64_t>(bias());
 155   auto *output_data = getTensorData<int16_t>(output());
 156   auto *scratch_data = getTensorData<int64_t>(_scratch_tensor.get());
 157
 158   const Shape &input_shape = input()->shape();
 159   const Shape &filter_shape = filter()->shape();
 160   const Shape &output_shape = output()->shape();
 161
 162   const int32_t batches = input_shape.dim(0);
 163   const int32_t input_height = input_shape.dim(1);
 164   const int32_t input_width = input_shape.dim(2);
 165   const int32_t input_depth = input_shape.dim(3);
 166   const int32_t output_depth = filter_shape.dim(0);
 167   const int32_t filter_height = filter_shape.dim(1);
 168   const int32_t filter_width = filter_shape.dim(2);
 169   const int32_t output_height = output_shape.dim(1);
 170   const int32_t output_width = output_shape.dim(2);
 171
 172   const int32_t stride_height = _params.stride_height;
 173   const int32_t stride_width = _params.stride_width;
 174
 175   int32_t activation_min{};
 176   int32_t activation_max{};
 177   calculateActivationRangeQuantized(Activation::NONE, output(), &activation_min, &activation_max);
 178
 179   std::memset(scratch_data, 0, _scratch_tensor->shape().num_elements() * sizeof(int64_t));
 180
 181   BroadcastableWrapper<ChannelQuantMultipliers> output_multipliers(_quant_multipliers);
 182   for (int32_t batch = 0; batch < batches; ++batch)
 183   {
 184     for (int32_t in_y = 0; in_y < input_height; ++in_y)
 185     {
 186       for (int32_t in_x = 0; in_x < input_width; ++in_x)
 187       {
 188         for (int32_t in_c = 0; in_c < input_depth; ++in_c)
 189         {
 190           const int32_t out_y_origin = in_y * stride_height - _padding_height;
 191           const int32_t out_x_origin = in_x * stride_width - _padding_width;
 192           for (int32_t filter_y = 0; filter_y < filter_height; ++filter_y)
 193           {
 194             for (int32_t filter_x = 0; filter_x < filter_width; ++filter_x)
 195             {
 196               const int32_t out_x = out_x_origin + filter_x;
 197               const int32_t out_y = out_y_origin + filter_y;
 198               if ((out_y >= 0 && out_y < output_height) && (out_x >= 0 && out_x < output_width))
 199               {
 200                 for (int32_t out_c = 0; out_c < output_depth; ++out_c)
 201                 {
 202                   const int16_t input_val =
 203                       input_data[calcOffset(input_shape, batch, in_y, in_x, in_c)];
 204                   const int16_t filter_val =
 205                       filter_data[calcOffset(filter_shape, out_c, filter_y, filter_x, in_c)];
 206                   scratch_data[calcOffset(output_shape, batch, out_y, out_x, out_c)] +=
 207                       static_cast<int64_t>(input_val) * static_cast<int64_t>(filter_val);
 208                 }
 209               }
 210             }
 211           }
 212         }
 213       }
 214     }
 215     for (int32_t out_y = 0; out_y < output_height; ++out_y)
 216     {
 217       for (int32_t out_x = 0; out_x < output_width; ++out_x)
 218       {
 219         for (int32_t out_c = 0; out_c < output_depth; ++out_c)
 220         {
 221           int64_t acc = scratch_data[calcOffset(output_shape, batch, out_y, out_x, out_c)];
 222           if (bias_data)
 223           {
 224             acc += bias_data[out_c];
 225           }
 226           int32_t scaled_acc = tflite::MultiplyByQuantizedMultiplier(
 227               acc, output_multipliers[out_c].multiplier, output_multipliers[out_c].shift);
 228
 229           scaled_acc = std::max(scaled_acc, activation_min);
 230           scaled_acc = std::min(scaled_acc, activation_max);
 231
 232           output_data[calcOffset(output_shape, batch, out_y, out_x, out_c)] = scaled_acc;
 233         }
 234       }
 235     }
 236   }
 237 }
 238
 239 } // namespace kernels
 240 } // namespace luci_interpreter