inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/fully_connected/fully_connected_kernel_bs_f_bsv16_af8.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_bs_f_bsv16_af8.h"
  18
  19 namespace kernel_selector
  20 {
  21     ParamsKey FullyConnected_bs_f_bsv16_af8::GetSupportedKey() const
  22     {
  23         ParamsKey k;
  24         k.EnableInputDataType(Datatype::F16);
  25         k.EnableOutputDataType(Datatype::F16);
  26         k.EnableInputWeightsType(WeightsType::F16);
  27         k.EnableInputWeightsType(WeightsType::F32);
  28         k.EnableAllInputLayout();
  29         k.EnableOutputLayout(DataLayout::fb);
  30         k.EnableOutputLayout(DataLayout::bs_f_bsv16__af8);
  31         k.EnableBatching();
  32         k.EnableBiasPerFeature();
  33         k.EnableNonBiasTerm();
  34         k.EnableSubGroup();
  35         return k;
  36     }
  37
  38     FullyConnected_bs_f_bsv16_af8::DispatchData FullyConnected_bs_f_bsv16_af8::SetDefault(const fully_connected_params& arg, int ) const
  39     {
  40         auto kd = FullyConnectedBlockKernelBase::SetDefault(arg);
  41
  42         size_t groups_per_batches = GetLocalGroupsSize(arg);
  43         kd.gws0 = Align(arg.output.LogicalSize() / (GetBatchesPerWorkItem(arg) * groups_per_batches), 16);
  44         kd.gws1 = groups_per_batches;
  45         kd.lws0 = 16;
  46         kd.lws1 = 1;
  47
  48         return kd;
  49     }
  50
  51     static bool check_input_layout(const DataTensor& t)
  52     {
  53         bool b16_layout = false;
  54         b16_layout |= t.GetLayout() == DataLayout::bs_f_bsv16__af8;
  55         b16_layout |= DataTensor::Channelndex(t.GetLayout(), Tensor::DataChannelName::BATCH) == 0 && t.Batch().v == 16;
  56         return b16_layout;
  57     }
  58
  59     bool FullyConnected_bs_f_bsv16_af8::Validate(const Params& p, const optional_params& o) const
  60     {
  61         if (!FullyConnectedBlockKernelBase::Validate(p, o))
  62         {
  63             return false;
  64         }
  65
  66         const auto& params = static_cast<const fully_connected_params&>(p);
  67         const auto& optParams = static_cast<const fully_connected_optional_params&>(o);
  68
  69         if (!params.engineInfo.bSubGroupShortSupport && params.inputs[0].GetDType() == Datatype::F16)
  70         {
  71             return false;
  72         }
  73
  74         const bool bProperBatch = params.inputs[0].Batch().v == 16;
  75         const bool bProperInput = check_input_layout(params.inputs[0]);
  76         const bool bSupportedLayout = optParams.allowInputReordering || bProperInput;
  77
  78         if (!bProperBatch || !bSupportedLayout)
  79         {
  80             return false;
  81         }
  82
  83         return true;
  84     }
  85
  86     KernelsData FullyConnected_bs_f_bsv16_af8::GetKernelsData(const Params& params, const optional_params& optParams) const
  87     {
  88         KernelsData res = {};
  89         for (size_t i = 0; i < autoTuneOptions.size(); i++)
  90         {
  91             KernelsData kd = GetTunedKernelsDataByIndex(params, optParams, DataLayout::bs_f_bsv16__af8, { WeightsLayout::os_i_osv16__ai8 }, FORCE_PRIORITY_2, (int)i);
  92             if (!kd.empty())
  93             {
  94                 res.emplace_back(kd[0]);
  95             }
  96         }
  97
  98         return res;
  99     }
 100 }