inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/embed/embed_kernel_ref.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 "embed_kernel_ref.h"
  18 #include "kernel_selector_utils.h"
  19 #include "common_tools.h"
  20
  21 namespace kernel_selector
  22 {
  23
  24         ParamsKey EmbedKernelRef::GetSupportedKey() const
  25         {
  26                 ParamsKey k;
  27                 k.EnableInputDataType(Datatype::F16);
  28                 k.EnableInputDataType(Datatype::F32);
  29                 k.EnableInputDataType(Datatype::INT8);
  30                 k.EnableOutputDataType(Datatype::F16);
  31                 k.EnableOutputDataType(Datatype::F32);
  32                 k.EnableOutputDataType(Datatype::INT8);
  33                 k.EnableInputWeightsType(WeightsType::F16);
  34                 k.EnableInputWeightsType(WeightsType::F32);
  35                 k.EnableInputWeightsType(WeightsType::INT8);
  36         k.EnableAllInputLayout();
  37                 k.EnableOutputLayout(DataLayout::bf);
  38                 k.EnableBiasPerOutput();
  39                 k.EnableBiasPerFeature();
  40                 k.EnableTensorOffset();
  41                 k.EnableTensorPitches();
  42                 k.EnableBatching();
  43         k.EnableNonBiasTerm();
  44                 return k;
  45         }
  46
  47         JitConstants EmbedKernelRef::GetJitConstants(const embed_params& params) const
  48         {
  49         JitConstants jit = WeightBiasKernelBase::GetJitConstants(params);
  50         const auto& input = params.inputs[0];
  51         const auto x_size = input.LogicalSize() / input.Batch().v;
  52         const auto w_size = params.weights.OFM().v;
  53         jit.AddConstant(MakeJitConstant("INPUT0_ELEMENTS_COUNT", x_size));
  54         jit.AddConstant(MakeJitConstant("NUM_OUTPUT_SIZE", w_size));
  55
  56         return jit;
  57         }
  58
  59         EmbedKernelRef::DispatchData EmbedKernelRef::SetDefault(const embed_params& params) const
  60         {
  61                 DispatchData kd;
  62                 std::vector<size_t> global = { params.inputs[0].X().v , params.weights.OFM().v, params.inputs[0].Batch().v };
  63                 std::vector<size_t> local = GetOptimalLocalWorkGroupSizes(global);
  64
  65                 kd.gws0 = global[0];
  66                 kd.gws1 = global[1];
  67                 kd.gws2 = global[2];
  68
  69                 kd.lws0 = local[0];
  70                 kd.lws1 = local[1];
  71                 kd.lws2 = 1;
  72                 return kd;
  73
  74         }
  75
  76         KernelsData EmbedKernelRef::GetKernelsData(const Params& params, const optional_params& options) const
  77         {
  78                 assert(params.GetType() == KernelType::EMBED);
  79
  80                 const embed_params& orgParams = static_cast<const embed_params&>(params);
  81
  82                 const std::vector<WeightsLayout> weightsLayouts = {
  83                         WeightsLayout::oiyx,
  84                 };
  85
  86                 DispatchData runInfo = SetDefault(orgParams);
  87                 KernelData kd = KernelData::Default<embed_params>(params);
  88                 embed_params& newParams = *static_cast<embed_params*>(kd.params.get());
  89
  90                 bool succeed = UpdateWeightsParams(
  91                         newParams,
  92                         options,
  93                         weightsLayouts,
  94                         kd.weightsReorderParams);
  95
  96                 if (!succeed)
  97                 {
  98                         return{};
  99                 }
 100
 101                 auto cldnn_jit = GetJitConstants(newParams);
 102                 auto entry_point = GetEntryPoint(kernelName, newParams.layerID, options);
 103                 auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
 104
 105                 auto& kernel = kd.kernels[0];
 106
 107                 FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point, DEFAULT, true, !newParams.bias.empty());
 108
 109                 kd.estimatedTime = runInfo.effiency;
 110
 111                 return{ kd };
 112         }
 113
 114
 115 }