inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/fully_connected_grad_weights/fully_connected_grad_weights_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_weights_kernel_base.h"
  18 #include "kernel_selector_utils.h"
  19
  20 namespace kernel_selector
  21 {
  22     JitConstants FullyConnectedGradWeightsKernelBase::GetJitConstants(const fully_connected_grad_weights_params& params) const
  23     {
  24         JitConstants jit = training_kernel_base::GetJitConstants(params);
  25
  26         return jit;
  27     }
  28
  29     FullyConnectedGradWeightsKernelBase::DispatchData FullyConnectedGradWeightsKernelBase::SetDefault(const fully_connected_grad_weights_params& params) const
  30     {
  31         DispatchData kd;
  32
  33         kd.fp16UnitUsed = params.inputs[0].GetDType() == Datatype::F16;
  34         size_t gws0 = params.weights.OFM().v * params.weights.IFM().v;
  35         size_t lws0 = std::min(gws0, static_cast<size_t>(32));
  36         while (gws0 % lws0)
  37         {
  38             lws0--;
  39         }
  40         kd.gws0 = gws0;
  41         kd.gws1 = params.weights.X().v;
  42         kd.gws2 = params.weights.Y().v;
  43         kd.lws0 = lws0;
  44         kd.lws1 = 1;
  45         kd.lws2 = 1;
  46         kd.effiency = DONT_USE_IF_HAVE_SOMETHING_ELSE;
  47         return kd;
  48     }
  49
  50     KernelsData FullyConnectedGradWeightsKernelBase::GetKernelsData(const Params& params, const optional_params& options) const
  51     {
  52         assert(params.GetType() == KernelType::FULLY_CONNECTED_GRAD_WEIGHTS);
  53
  54         const fully_connected_grad_weights_params& orgParams = static_cast<const fully_connected_grad_weights_params&>(params);
  55
  56         const std::vector<WeightsLayout> weightsLayouts = {
  57             WeightsLayout::oi,
  58             WeightsLayout::io,
  59             WeightsLayout::oiyx,
  60             WeightsLayout::iyxo,
  61             WeightsLayout::yxio,
  62             WeightsLayout::oyxi
  63         };
  64
  65         DispatchData runInfo = SetDefault(orgParams);
  66         KernelData kd = KernelData::Default<fully_connected_grad_weights_params>(params);
  67         fully_connected_grad_weights_params& newParams = *static_cast<fully_connected_grad_weights_params*>(kd.params.get());
  68
  69         bool succeed = UpdateWeightsParams(
  70             newParams,
  71             options,
  72             weightsLayouts,
  73             kd.weightsReorderParams);
  74
  75         if (!succeed)
  76         {
  77             return{};
  78         }
  79
  80         auto cldnn_jit = GetJitConstants(orgParams);
  81         auto entry_point = GetEntryPoint(kernelName, orgParams.layerID, options);
  82         auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
  83
  84         auto& kernel = kd.kernels[0];
  85         FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point, DEFAULT, true, !orgParams.bias.empty());
  86         if (orgParams.use_momentum)
  87         {
  88             kernel.arguments.push_back({ ArgumentDescriptor::Types::PREV_WEIGHTS_GRADIENT, 0 });
  89             if (!orgParams.bias.empty())
  90                 kernel.arguments.push_back({ ArgumentDescriptor::Types::PREV_BIAS_GRADIENT, 0 });
  91         }
  92         kernel.arguments.push_back({ ArgumentDescriptor::Types::INPUT, 1 });
  93         kernel.arguments.push_back({ ArgumentDescriptor::Types::LEARNING_RATE, 0 });
  94
  95         kd.estimatedTime = runInfo.effiency;
  96
  97         return{ kd };
  98     }
  99 }