inference-engine/thirdparty/clDNN/src/gpu/broadcast_gpu.cpp

   1 // Copyright (c) 2018 Intel Corporation
   2 //
   3 // Licensed under the Apache License, Version 2.0 (the "License");
   4 // you may not use this file except in compliance with the License.
   5 // You may obtain a copy of the License at
   6 //
   7 //      http://www.apache.org/licenses/LICENSE-2.0
   8 //
   9 // Unless required by applicable law or agreed to in writing, software
  10 // distributed under the License is distributed on an "AS IS" BASIS,
  11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12 // See the License for the specific language governing permissions and
  13 // limitations under the License.
  14
  15
  16 #include "broadcast_inst.h"
  17
  18 #include "primitive_gpu_base.h"
  19 #include "implementation_map.h"
  20 #include "kernel_selector_helper.h"
  21 #include "broadcast/broadcast_kernel_selector.h"
  22 #include "broadcast/broadcast_kernel_base.h"
  23 #include "error_handler.h"
  24
  25 namespace cldnn { namespace gpu {
  26
  27 struct broadcast_gpu : typed_primitive_gpu_impl<broadcast>
  28 {
  29     using parent = typed_primitive_gpu_impl<broadcast>;
  30     using parent::parent;
  31
  32
  33     static primitive_impl* create(const broadcast_node& arg)
  34     {
  35         auto bc_params          = get_default_params<kernel_selector::broadcast_params>(arg, 1);
  36         auto bc_optional_params = get_default_optional_params<kernel_selector::broadcast_optional_params>(arg.get_program());
  37
  38         const auto& broadcast_axes = arg.get_primitive()->broadcast_axes;
  39         uint16_t index = (uint16_t) 0;
  40         uint16_t input_index = (uint16_t) broadcast_axes.size();
  41
  42         //bfyx format
  43         for (size_t i = 0; i < 4; ++i)
  44         {
  45             if (std::find(broadcast_axes.begin(), broadcast_axes.end(), i) != broadcast_axes.end())
  46             {
  47                 bc_params.input_order.push_back(index);
  48                 ++index;
  49             }
  50             else
  51             {
  52                 bc_params.input_order.push_back(input_index);
  53                 ++input_index;
  54             }
  55         }
  56
  57         auto& kernel_selector = kernel_selector::broadcast_kernel_selector::Instance();
  58         auto best_kernels = kernel_selector.GetBestKernels(bc_params, bc_optional_params);
  59
  60         CLDNN_ERROR_BOOL(arg.id(), "Best_kernel.empty()", best_kernels.empty(), "Cannot find a proper kernel with this arguments");
  61
  62         return new broadcast_gpu(arg, best_kernels[0]);
  63     }
  64 };
  65
  66 namespace {
  67     struct attach {
  68         attach() {
  69             auto val_fw = broadcast_gpu::create;
  70
  71             implementation_map<broadcast>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::bfyx), val_fw);
  72             implementation_map<broadcast>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::bfyx), val_fw);
  73             implementation_map<broadcast>::add(std::make_tuple(engine_types::ocl, data_types::i8,  format::bfyx), val_fw);
  74             implementation_map<broadcast>::add(std::make_tuple(engine_types::ocl, data_types::u8,  format::bfyx), val_fw);
  75             implementation_map<broadcast>::add(std::make_tuple(engine_types::ocl, data_types::i32, format::bfyx), val_fw);
  76             implementation_map<broadcast>::add(std::make_tuple(engine_types::ocl, data_types::i64, format::bfyx), val_fw);
  77         }
  78         ~attach() = default;
  79     };
  80
  81     attach attach_impl;
  82
  83 }
  84 } }