compiler/locomotiv/src/Node/TensorBroadcast.cpp

   1 /*
   2  * Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at
   7  *
   8  *    http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 #include "NodeExecution.h"
  18
  19 #include "NodeDataImpl.h"
  20 #include "NodeDomain.h"
  21
  22 #include <nncc/core/ADT/tensor/Shape.h>
  23 #include <nncc/core/ADT/tensor/Buffer.h>
  24 #include <nncc/core/ADT/tensor/Index.h>
  25 #include <nncc/core/ADT/tensor/IndexEnumerator.h>
  26 #include <nncc/core/ADT/tensor/LexicalLayout.h>
  27
  28 using nncc::core::ADT::tensor::Index;
  29 using nncc::core::ADT::tensor::IndexEnumerator;
  30 using nncc::core::ADT::tensor::LexicalLayout;
  31 using nncc::core::ADT::tensor::make_buffer;
  32 using nncc::core::ADT::tensor::Shape;
  33
  34 #include <cassert>
  35 #include <stdexcept>
  36
  37 namespace locomotiv
  38 {
  39
  40 void NodeExecution::execute(loco::TensorBroadcast *tensor_broadcast)
  41 {
  42   auto input_data = annot_data(tensor_broadcast->input());
  43
  44   // Calculate output shape
  45   Shape input_shape = *(input_data->shape());
  46
  47   // TODO Reuse "ShapeInferenceService"
  48   Shape output_shape;
  49
  50   output_shape.resize(input_shape.rank());
  51   for (uint32_t axis = 0; axis < input_shape.rank(); ++axis)
  52   {
  53     if (tensor_broadcast->mapping()->defined(axis))
  54     {
  55       assert(input_shape.dim(axis) == 1); // Required by TensorBroadcast definition
  56       output_shape.dim(axis) = tensor_broadcast->mapping()->dim(axis).value();
  57     }
  58     else
  59     {
  60       output_shape.dim(axis) = input_shape.dim(axis);
  61     }
  62   }
  63
  64   assert(input_shape.rank() == output_shape.rank());
  65
  66   uint32_t const rank = input_shape.rank();
  67
  68   std::unique_ptr<NodeData> output_data = nullptr;
  69
  70   switch (input_data->dtype())
  71   {
  72     // TODO Use type-generic implementation!
  73     case loco::DataType::FLOAT32:
  74     {
  75       auto input_bufptr = input_data->as_f32_bufptr();
  76       auto output_buf = make_buffer<float, LexicalLayout>(output_shape);
  77
  78       for (IndexEnumerator e{output_shape}; e.valid(); e.advance())
  79       {
  80         auto input_index = e.current();
  81         const auto &output_index = e.current();
  82
  83         for (uint32_t axis = 0; axis < rank; ++axis)
  84         {
  85           if (tensor_broadcast->mapping()->defined(axis))
  86           {
  87             input_index.at(axis) = 0;
  88           }
  89         }
  90
  91         output_buf.at(output_index) = input_bufptr->at(input_index);
  92       }
  93
  94       output_data = make_data(output_buf);
  95       break;
  96     }
  97     default:
  98       throw std::runtime_error("Not yet supported");
  99   }
 100
 101   assert(output_data != nullptr);
 102   annot_data(tensor_broadcast, std::move(output_data));
 103   annot_domain(tensor_broadcast, loco::Domain::Tensor);
 104 }
 105
 106 } // namespace locomotiv