inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_bfyx_depthwise_weights_lwg.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 "convolution_kernel_bfyx_depthwise_weights_lwg.h"
  18
  19 namespace kernel_selector
  20 {
  21     ParamsKey ConvolutionKernel_bfyx_depthwise_weights_lwg::GetSupportedKey() const
  22     {
  23         ParamsKey k;
  24         k.EnableInputDataType(Datatype::F32);
  25         k.EnableInputDataType(Datatype::F16);
  26         k.EnableInputWeightsType(WeightsType::F16);
  27         k.EnableInputWeightsType(WeightsType::F32);
  28         k.EnableOutputDataType(Datatype::F32);
  29         k.EnableOutputDataType(Datatype::F16);
  30         k.EnableInputLayout(DataLayout::bfyx);
  31         k.EnableOutputLayout(DataLayout::bfyx);
  32         k.EnableTensorOffset();
  33         k.EnableTensorPitches();
  34         k.EnableBiasPerFeature();
  35         k.EnableNonBiasTerm();
  36         k.EnableBatching();
  37         k.EnableSplitSupport();
  38         k.EnableSubGroup();
  39         k.EnableSubGroupShort();
  40         k.EnableDepthwiseSeparableOpt();
  41         k.EnableDilation();
  42         return k;
  43     }
  44
  45     bool ConvolutionKernel_bfyx_depthwise_weights_lwg::Validate(const Params& p, const optional_params& o) const
  46     {
  47         if (!ConvolutionKernelBase::Validate(p, o) ||
  48             !CovolutionCheckInput(p, o))
  49         {
  50             return false;
  51         }
  52
  53        const convolution_params& cp = static_cast<const convolution_params&>(p);
  54        if (!cp.depthwise_separable_opt)
  55            return false;
  56        if ((cp.filterSize.x > 4) ||
  57            (cp.filterSize.y > 4) ||
  58            ((cp.inputs[0].Feature().v != cp.split) && (cp.inputs[0].Feature().v != cp.groups)))
  59        {
  60            return false;
  61        }
  62
  63         return true;
  64     }
  65
  66     ConvolutionKernelBase::DispatchData ConvolutionKernel_bfyx_depthwise_weights_lwg::SetDefault(const convolution_params& params, int) const
  67     {
  68         DispatchData runInfo = Parent::SetDefault(params);
  69         const auto& out = params.output;
  70
  71         std::vector<size_t> global = { out.X().v * out.Y().v, out.Feature().v, out.Batch().v };
  72
  73         runInfo.gws0 = Align(global[0], 16);
  74         runInfo.gws1 = global[1];
  75         runInfo.gws2 = global[2];
  76         runInfo.lws0 = 16;
  77         runInfo.lws1 = 1;
  78         runInfo.lws2 = 1;
  79
  80         runInfo.effiency = FORCE_PRIORITY_6;
  81
  82         return runInfo;
  83     }
  84
  85     JitConstants ConvolutionKernel_bfyx_depthwise_weights_lwg::GetJitConstants(const convolution_params& params, const DispatchData& kd) const
  86     {
  87         auto mem_consts = ConvolutionKernelBase::GetJitConstants(params, kd);
  88
  89         if(params.padding.x != 0 || params.padding.y != 0)
  90             mem_consts.AddConstant(MakeJitConstant("BOUNDARY_CHECK", 1));
  91
  92         return mem_consts;
  93     }
  94
  95     KernelsData ConvolutionKernel_bfyx_depthwise_weights_lwg::GetKernelsData(const Params& params, const optional_params& options) const
  96     {
  97         return GetTunedKernelsDataByIndex(params, options);
  98     }
  99 }