inference-engine/thirdparty/clDNN/src/include/fused_conv_bn_scale_inst.h

   1 /*
   2 // Copyright (c) 2018 Intel Corporation
   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 ///////////////////////////////////////////////////////////////////////////////////////////////////
  18 #pragma once
  19 #include "api_extension/CPP/fused_conv_bn_scale.hpp"
  20 #include "primitive_inst.h"
  21
  22 #include <memory>
  23
  24 namespace cldnn
  25 {
  26
  27 template <>
  28 struct typed_program_node<fused_conv_bn_scale> : public typed_program_node_base<fused_conv_bn_scale>
  29 {
  30     using parent = typed_program_node_base<fused_conv_bn_scale>;
  31
  32 public:
  33     typed_program_node(std::shared_ptr<primitive> prim, program_impl& prog)
  34         : parent(prim, prog)
  35         , split(this->get_primitive()->split())
  36     {
  37     }
  38
  39     void set_split(int32_t node_split) { split = node_split; }
  40     int32_t get_split() const { return split; }
  41
  42     program_node& input(size_t idx = 0) const
  43     {
  44         if (static_cast<int32_t>(idx) >= static_cast<int32_t>(desc->input.size()))
  45             throw std::range_error("input index too big");
  46
  47         return get_dependency(idx);
  48     }
  49
  50     program_node& weights(size_t idx = 0) const
  51     {
  52         if (static_cast<int32_t>(idx) >= this->get_split())
  53             throw std::range_error("weights offset too big");
  54
  55         return get_dependency(desc->input.size() + idx);
  56     }
  57
  58     program_node& bias(size_t idx = 0) const
  59     {
  60         if (static_cast<int32_t>(idx) >= this->get_split())
  61             throw std::range_error("bias offset too big");
  62
  63         return get_dependency(desc->input.size() + this->get_split() + idx);
  64     }
  65
  66     program_node& weights_quantization_factors(size_t idx = 0) const
  67     {
  68         if (static_cast<int32_t>(idx) >= this->get_split())
  69             throw std::range_error("quantization factor offset too big");
  70
  71         return get_dependency(desc->input.size() + 2*this->get_split() + idx);
  72     }
  73
  74     program_node& output_calibration_factors(size_t idx = 0) const
  75     {
  76         if (static_cast<int32_t>(idx) >= this->get_split())
  77             throw std::range_error("calibration factor offset too big");
  78
  79         return get_dependency(desc->input.size() + 3 * this->get_split() + idx);
  80     }
  81
  82     bool bias_term() const
  83     {
  84         return get_primitive()->bias.size() > 0;
  85     }
  86
  87     bool scale_bias_term() const
  88     {
  89         return !get_primitive()->scale_bias.empty();
  90     }
  91
  92     bool is_fused_in_training() const
  93     {
  94         return !get_primitive()->inv_variance.empty();
  95     }
  96
  97 private:
  98     int32_t split;
  99 };
 100
 101 using fused_conv_bn_scale_node = typed_program_node<fused_conv_bn_scale>;
 102
 103 template <>
 104 class typed_primitive_inst<fused_conv_bn_scale> : public typed_primitive_inst_base<fused_conv_bn_scale>
 105 {
 106     using parent = typed_primitive_inst_base<fused_conv_bn_scale>;
 107
 108 public:
 109     static layout calc_output_layout(fused_conv_bn_scale_node const& node);
 110     static std::string to_string(fused_conv_bn_scale_node const& node);
 111
 112 public:
 113     typed_primitive_inst(network_impl& network, fused_conv_bn_scale_node const& node);
 114
 115     memory_impl& weights_memory(size_t index) const
 116     {
 117         if (static_cast<int32_t>(index) >= node.get_split())
 118             throw std::range_error("weights offset too big");
 119
 120         return dep_memory(inputs_memory_count() + index);
 121     }
 122
 123     memory_impl& bias_memory(size_t index) const
 124     {
 125         if (static_cast<int32_t>(index) >= node.get_split())
 126             throw std::range_error("bias offset too big");
 127
 128         return dep_memory(inputs_memory_count() + node.get_split() + index);
 129     }
 130
 131     bool bias_term() const
 132     {
 133         return node.bias_term();
 134     }
 135
 136     bool scale_bias_term() const
 137     {
 138         return node.scale_bias_term();
 139     }
 140
 141     bool is_fused_in_training() const
 142     {
 143         return node.is_fused_in_training();
 144     }
 145 };
 146
 147 using fused_conv_bn_scale_inst = typed_primitive_inst<fused_conv_bn_scale>;
 148
 149 }