inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/fully_connected/fully_connected_kernel_image_tutorial.cpp

   1 /*
   2 // Copyright (c) 2016 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 #include "fully_connected_kernel_image_tutorial.h"
  18 #include "kernel_selector_utils.h"
  19
  20 namespace kernel_selector
  21 {
  22     ParamsKey FullyConnected_image_tutorial::GetSupportedKey() const
  23     {
  24         ParamsKey k;
  25         k.EnableInputDataType(Datatype::F16);
  26         k.EnableInputDataType(Datatype::F32);
  27         k.EnableOutputDataType(Datatype::F16);
  28         k.EnableOutputDataType(Datatype::F32);
  29         k.EnableInputWeightsType(WeightsType::F16);
  30         k.EnableInputWeightsType(WeightsType::F32);
  31         k.EnableAllInputLayout();
  32         k.EnableInputLayout(DataLayout::bf);
  33         k.EnableOutputLayout(DataLayout::bf);
  34         k.EnableBiasPerOutput();
  35         k.EnableBiasPerFeature();
  36         k.EnableNonBiasTerm();
  37         k.EnableTensorOffset();
  38         k.EnableTensorPitches();
  39         k.EnableBatching();
  40         return k;
  41     }
  42
  43     FullyConnected_image_tutorial::DispatchData FullyConnected_image_tutorial::SetDefault(const fully_connected_params& params, int ) const
  44     {
  45         auto runInfo = Parent::SetDefault(params);
  46
  47         std::vector<size_t> global = { params.output.Feature().v, params.output.Batch().v };
  48         std::vector<size_t> local  = GetOptimalLocalWorkGroupSizes(global);
  49
  50         runInfo.gws0 = global[0];
  51         runInfo.gws1 = global[1];
  52         runInfo.gws2 = 1;
  53
  54         runInfo.lws0 = local[0];
  55         runInfo.lws1 = local[1];
  56         runInfo.lws2 = 1;
  57
  58         runInfo.effiency = TUTORIAL_PRIORITY;
  59
  60         return runInfo;
  61     }
  62
  63     KernelsData FullyConnected_image_tutorial::GetKernelsData(const Params& params, const optional_params& options) const
  64     {
  65         KernelsData res = {};
  66         for (size_t i = 0; i < autoTuneOptions.size(); i++)
  67         {
  68             KernelsData kd = GetTunedKernelsDataByIndex(params, options, DataLayout::bfyx,
  69                 { WeightsLayout::image_2d_weights_c4_fyx_b }, DONT_USE_IF_HAVE_SOMETHING_ELSE, (int)i);
  70             if (!kd.empty())
  71             {
  72                 res.emplace_back(kd[0]);
  73             }
  74         }
  75         return res;
  76     }
  77 }