inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/fully_connected_grad_input/fully_connected_grad_input_kernel_base.cpp

   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 #include "fully_connected_grad_input_kernel_base.h"
  18 #include "kernel_selector_utils.h"
  19
  20 namespace kernel_selector
  21 {
  22     JitConstants FullyConnectedGradInputKernelBase::GetJitConstants(const fully_connected_grad_input_params& params) const
  23     {
  24         return WeightBiasKernelBase::GetJitConstants(params);
  25     }
  26
  27     FullyConnectedGradInputKernelBase::DispatchData FullyConnectedGradInputKernelBase::SetDefault(const fully_connected_grad_input_params& params) const
  28     {
  29         DispatchData kd;
  30
  31         kd.fp16UnitUsed = params.inputs[0].GetDType() == Datatype::F16;
  32         size_t gws0 = params.output.Batch().v * params.weights.IFM().v;
  33         size_t lws0 = std::min(gws0, static_cast<size_t>(32));
  34         while (gws0 % lws0)
  35         {
  36             lws0--;
  37         }
  38         kd.gws0 = gws0;
  39         kd.gws1 = params.weights.X().v;
  40         kd.gws2 = params.weights.Y().v;
  41         kd.lws0 = lws0;
  42         kd.lws1 = 1;
  43         kd.lws2 = 1;
  44         kd.effiency = DONT_USE_IF_HAVE_SOMETHING_ELSE;
  45         return kd;
  46     }
  47
  48     KernelsData FullyConnectedGradInputKernelBase::GetKernelsData(const Params& params, const optional_params& options) const
  49     {
  50         assert(params.GetType() == KernelType::FULLY_CONNECTED_GRAD_INPUT);
  51
  52         const fully_connected_grad_input_params& orgParams = static_cast<const fully_connected_grad_input_params&>(params);
  53
  54         const std::vector<WeightsLayout> weightsLayouts = {
  55             WeightsLayout::oi,
  56             WeightsLayout::io,
  57             WeightsLayout::oiyx,
  58             WeightsLayout::iyxo,
  59             WeightsLayout::yxio,
  60             WeightsLayout::oyxi
  61         };
  62
  63         DispatchData runInfo = SetDefault(orgParams);
  64         KernelData kd = KernelData::Default<fully_connected_grad_input_params>(params);
  65         fully_connected_grad_input_params& newParams = *static_cast<fully_connected_grad_input_params*>(kd.params.get());
  66
  67         bool succeed = UpdateWeightsParams(
  68             newParams,
  69             options,
  70             weightsLayouts,
  71             kd.weightsReorderParams);
  72
  73         if (!succeed)
  74         {
  75             return{};
  76         }
  77
  78         auto cldnn_jit = GetJitConstants(orgParams);
  79         auto entry_point = GetEntryPoint(kernelName, orgParams.layerID, options);
  80         auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
  81
  82         auto& kernel = kd.kernels[0];
  83         FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point, DEFAULT, true, !orgParams.bias.empty());
  84         kernel.arguments.push_back({ ArgumentDescriptor::Types::INPUT, 1 });
  85
  86         kd.estimatedTime = runInfo.effiency;
  87
  88         return{ kd };
  89     }
  90 }