inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/fully_connected/fully_connected_kernel_bf_io_gemm.cpp

   1 /*
   2 // Copyright (c) 2016 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_kernel_bf_io_gemm.h"
  18
  19 namespace kernel_selector {
  20
  21     ParamsKey FullyConnected_bf_io_GEMM::GetSupportedKey() const
  22     {
  23         ParamsKey k;
  24         k.EnableInputDataType(Datatype::F16);
  25         k.EnableInputDataType(Datatype::F32);
  26         k.EnableOutputDataType(Datatype::F16);
  27         k.EnableOutputDataType(Datatype::F32);
  28         k.EnableInputWeightsType(WeightsType::F16);
  29         k.EnableInputWeightsType(WeightsType::F32);
  30         k.EnableAllInputLayout();
  31         k.EnableOutputLayout(DataLayout::bf);
  32         k.EnableBiasPerOutput();
  33         k.EnableBiasPerFeature();
  34         k.EnableNonBiasTerm();
  35         k.EnableTensorOffset();
  36         k.EnableTensorPitches();
  37         return k;
  38     }
  39
  40     FullyConnected_bf_io_GEMM::DispatchData FullyConnected_bf_io_GEMM::SetDefault(const fully_connected_params& params, int autoTuneIndex) const
  41     {
  42         auto runInfo = Parent::SetDefault(params, autoTuneIndex);
  43
  44         const uint32_t localWorkSizeX = 64;
  45         const uint32_t globalWorkSizeX = localWorkSizeX;
  46
  47         std::vector<size_t> global = { globalWorkSizeX, params.output.Feature().v, params.output.Batch().v };
  48         std::vector<size_t> local = { localWorkSizeX, 1, 1 };
  49
  50         runInfo.gws0 = global[0];
  51         runInfo.gws1 = global[1];
  52         runInfo.gws2 = 1;
  53
  54         runInfo.lws0 = local[0];
  55         runInfo.lws1 = local[1];
  56         runInfo.lws2 = 1;
  57
  58         runInfo.effiency = FORCE_PRIORITY_6;
  59
  60         return runInfo;
  61     }
  62
  63     JitConstants FullyConnected_bf_io_GEMM::GetJitConstants(const fully_connected_params& params, const DispatchData& kd) const
  64     {
  65         auto jit = Parent::GetJitConstants(params, kd);
  66
  67         if (params.inputs[0].GetDType() == Datatype::F16)
  68         {
  69             jit.AddConstant(MakeJitConstant("__fc_f16", ""));
  70         }
  71         else
  72         {
  73             jit.AddConstant(MakeJitConstant("__fc_f32", ""));
  74         }
  75
  76         const uint32_t localWorkSizeX = 64;
  77         const uint32_t globalWorkSizeX = localWorkSizeX;
  78         const uint32_t vecSize = 4;
  79         size_t matrixLineSize = params.inputs[0].Batch().pitch;
  80
  81         jit.AddConstants({
  82             MakeJitConstant("LAST_INPUT_SIZE_REMAINDER", matrixLineSize % (globalWorkSizeX * vecSize)),
  83             MakeJitConstant("LAST_INPUT_SIZE_DIV_4", matrixLineSize % vecSize),
  84         });
  85
  86         return jit;
  87     }
  88
  89     KernelsData FullyConnected_bf_io_GEMM::GetKernelsData(const Params& params, const optional_params& options) const
  90     {
  91         KernelsData res = {};
  92         for (size_t i = 0; i < autoTuneOptions.size(); i++)
  93         {
  94             KernelsData kd = GetTunedKernelsDataByIndex(params, options, DataLayout::bf, { WeightsLayout::oiyx }, FORCE_PRIORITY_6, (int)i);
  95             if (!kd.empty())
  96             {
  97                 res.emplace_back(kd[0]);
  98             }
  99         }
 100
 101         return res;
 102     }
 103 }