inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/index_select/index_select_kernel_base.cpp

   1 // Copyright (c) 2018 Intel Corporation
   2 //
   3 // Licensed under the Apache License, Version 2.0 (the "License");
   4 // you may not use this file except in compliance with the License.
   5 // You may obtain a copy of the License at
   6 //
   7 //      http://www.apache.org/licenses/LICENSE-2.0
   8 //
   9 // Unless required by applicable law or agreed to in writing, software
  10 // distributed under the License is distributed on an "AS IS" BASIS,
  11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12 // See the License for the specific language governing permissions and
  13 // limitations under the License.
  14
  15
  16 #include "index_select_kernel_base.h"
  17
  18 #include "kernel_selector_utils.h"
  19
  20
  21 namespace kernel_selector
  22 {
  23     JitConstants IndexSelectKernelBase::GetJitConstants(const index_select_params& params)
  24     {
  25         JitConstants jit = MakeBaseParamsJitConstants(params);
  26
  27         jit.AddConstant(MakeJitConstant("AXES_NUMBER", params.axes.size()));
  28
  29         if (params.reverse) {
  30             jit.AddConstant(MakeJitConstant("REVERSE", 1));
  31         }
  32
  33         for (size_t i = 0; i < params.axes.size(); i++)
  34         {
  35             std::string size_name = "REVERSE_AXIS_SIZE";
  36             size_t size_value = 0;
  37             if (params.axes.size() > 1) {
  38                 std::stringstream ss;
  39                 ss << "REVERSE_" << toString(params.axes[i]) << "_SIZE";
  40                 size_name = ss.str();
  41             }
  42             jit.AddConstant(MakeJitConstant(toString(params.axes[i]), ""));
  43             if (params.reverse) {
  44                 if (params.axes[i] == IndexSelectAxis::BATCH)
  45                 {
  46                     size_value = params.inputs.at(0).Batch().v;
  47                 }
  48                 else if (params.axes[i] == IndexSelectAxis::X)
  49                 {
  50                     size_value = params.inputs.at(0).X().v;
  51                 }
  52                 else if (params.axes[i] == IndexSelectAxis::Y)
  53                 {
  54                     size_value = params.inputs.at(0).Y().v;
  55                 }
  56                 else if (params.axes[i] == IndexSelectAxis::FEATURE)
  57                 {
  58                     size_value = params.inputs.at(0).Feature().v;
  59                 }
  60             }
  61             jit.AddConstant(MakeJitConstant(size_name, size_value));
  62         }
  63
  64         return jit;
  65     }
  66
  67     IndexSelectKernelBase::DispatchData IndexSelectKernelBase::SetDefault(const index_select_params& params)
  68     {
  69         const auto& output = params.output;
  70         DispatchData kd;
  71
  72         kd.fp16UnitUsed = params.inputs[0].GetDType() == Datatype::F16;
  73
  74         std::vector<size_t> global;
  75
  76         if(params.axes.size() == 1) {
  77             if (params.reverse)
  78             {
  79                 if (params.axes[0] == IndexSelectAxis::BATCH)
  80                 {
  81                     global = { 1, params.inputs.at(0).Batch().v, output.Feature().v };
  82                 }
  83                 else if (params.axes[0] == IndexSelectAxis::X)
  84                 {
  85                     global = { output.Batch().v, params.inputs.at(0).X().v, output.Feature().v };
  86                 }
  87                 else if (params.axes[0] == IndexSelectAxis::Y)
  88                 {
  89                     global = { output.Batch().v, params.inputs.at(0).Y().v, output.Feature().v };
  90                 }
  91                 else if (params.axes[0] == IndexSelectAxis::FEATURE)
  92                 {
  93                     global = { output.Batch().v, params.inputs.at(0).Feature().v, output.Y().v };
  94                 }
  95             }
  96             else
  97             {
  98                 const auto indices = params.inputs.at(1).X().v;
  99
 100                 if (params.axes[0] == IndexSelectAxis::BATCH)
 101                 {
 102                     global = { 1, indices, output.Feature().v };
 103                 }
 104                 else if (params.axes[0] == IndexSelectAxis::X || params.axes[0] == IndexSelectAxis::Y)
 105                 {
 106                     global = { output.Batch().v, indices, output.Feature().v };
 107                 }
 108                 else if (params.axes[0] == IndexSelectAxis::FEATURE)
 109                 {
 110                     global = { output.Batch().v, indices, output.Y().v };
 111                 }
 112             }
 113         }
 114         else
 115         {
 116             if (params.reverse)
 117             {
 118                 global = { output.Batch().v, output.Y().v, output.Feature().v };
 119             }
 120         }
 121
 122         const auto& local = GetOptimalLocalWorkGroupSizes(global);
 123
 124         kd.gws0 = global[0];
 125         kd.gws1 = global[1];
 126         kd.gws2 = global[2];
 127
 128         kd.lws0 = local[0];
 129         kd.lws1 = local[1];
 130         kd.lws2 = local[2];
 131
 132         return kd;
 133     }
 134
 135     KernelsData IndexSelectKernelBase::GetCommonKernelsData(const Params& params, const optional_params& options, float estimated_time) const
 136     {
 137         assert(params.GetType() == KernelType::INDEX_SELECT);
 138
 139         const auto& prim_params = static_cast<const index_select_params&>(params); // NOLINT(cppcoreguidelines-pro-type-static-cast-downcast)
 140
 141         auto run_info     = SetDefault(prim_params);
 142         KernelData k_data = KernelData::Default<index_select_params>(params);
 143
 144         auto cldnn_jit   = GetJitConstants(prim_params);
 145         auto entry_point = GetEntryPoint(kernelName, prim_params.layerID, options);
 146         auto jit         = CreateJit(kernelName, cldnn_jit, entry_point);
 147
 148         auto& kernel = k_data.kernels[0];
 149         FillCLKernelData(kernel, run_info, params.engineInfo, kernelName, jit, entry_point, DEFAULT, false, false, (uint32_t)prim_params.inputs.size());
 150
 151         k_data.estimatedTime = estimated_time;
 152
 153         return {k_data};
 154     }
 155 }