inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_bfyx_gemm_like.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 "convolution_kernel_bfyx_gemm_like.h"
  18
  19 namespace kernel_selector
  20 {
  21
  22     ParamsKey ConvolutionKernel_bfyx_GEMMLike::GetSupportedKey() const
  23     {
  24         ParamsKey k;
  25         k.EnableInputDataType(Datatype::F16);
  26         k.EnableInputDataType(Datatype::F32);
  27         k.EnableInputWeightsType(WeightsType::F16);
  28         k.EnableInputWeightsType(WeightsType::F32);
  29         k.EnableOutputDataType(Datatype::F16);
  30         k.EnableOutputDataType(Datatype::F32);
  31         k.EnableInputLayout(DataLayout::bfyx);
  32         k.EnableOutputLayout(DataLayout::bfyx);
  33         k.EnableTensorOffset();
  34         k.EnableTensorPitches();
  35         k.EnableSubGroup();
  36         //k.EnableSubGroupShort(); // we need it for FP16 only. we check it on the Validate phase
  37         k.EnableBiasPerFeature();
  38         k.EnableNonBiasTerm();
  39         k.EnableBatching();
  40         k.EnableSplitSupport();
  41         return k;
  42     }
  43
  44     std::string ConvolutionKernel_bfyx_GEMMLike::GetKernelName(const convolution_params& params) const
  45     {
  46         if (params.inputs[0].GetDType() == Datatype::F32)
  47         {
  48             return kernelName + "_fp32";
  49         }
  50         else
  51         {
  52             return kernelName + "_fp16";
  53         }
  54     }
  55
  56     JitConstants ConvolutionKernel_bfyx_GEMMLike::GetJitConstants(const convolution_params& params, const DispatchData& runInfo) const
  57     {
  58         JitConstants jit = Parent::GetJitConstants(params, runInfo);
  59
  60         jit.AddConstants({
  61             MakeJitConstant("ALIGNED_OFM",                  RoundUp(params.output.Feature().v, runInfo.gemmStyle.subBlockDimN)),
  62             MakeJitConstant("DX",                           runInfo.gemmStyle.globalWorkSizeDX),
  63             MakeJitConstant("DY",                           runInfo.gemmStyle.globalWorkSizeDY),
  64             MakeJitConstant("FILTER_SIZE_X_DIV2",           params.filterSize.x / 2),
  65             MakeJitConstant("INPUT_BUFFER_WIDTH_PADDED",    ""),    // TODO: enable non padding path again
  66             MakeJitConstant("INPUT_BUFFER_HEIGHT_PADDED",   ""),
  67         });
  68
  69         if (CeilDiv(RoundUp(params.output.X().v * params.output.Y().v, runInfo.gemmStyle.subBlockDimM), runInfo.gemmStyle.globalWorkSizeDY) % runInfo.lws1 != 0)
  70             jit.AddConstant(MakeJitConstant("LEFTOVERS", 1));
  71
  72         return jit;
  73     }
  74
  75     ConvolutionKernel_bfyx_GEMMLike::Parent::DispatchData ConvolutionKernel_bfyx_GEMMLike::SetDefault(const convolution_params& arg, int autoTuneIndex) const
  76     {
  77         DispatchData runInfo = Parent::SetDefault(arg, autoTuneIndex);
  78
  79         runInfo.lws0 = 1;
  80         runInfo.lws2 = 1;
  81
  82         if (arg.inputs[0].GetDType() == Datatype::F16)
  83         {
  84             runInfo.gemmStyle = { 1, arg.filterSize.x, 32, 32, 1, 1 };
  85             runInfo.lws1 = 16;
  86             runInfo.effiency = FORCE_PRIORITY_6;
  87         }
  88         else
  89         {
  90             runInfo.gemmStyle = { 2, arg.filterSize.x, 32, 32, 2, 1 };
  91             runInfo.lws1 = 8;
  92             runInfo.effiency = FORCE_PRIORITY_8;
  93         }
  94
  95         size_t sgemm_m = RoundUp(arg.output.X().v * arg.output.Y().v, runInfo.gemmStyle.subBlockDimM);
  96         size_t sgemm_n = RoundUp(arg.output.Feature().v, runInfo.gemmStyle.subBlockDimN);
  97
  98         runInfo.gws0 = RoundUp(CeilDiv(sgemm_n, runInfo.gemmStyle.globalWorkSizeDX), runInfo.lws0);
  99         runInfo.gws1 = RoundUp(CeilDiv(sgemm_m, runInfo.gemmStyle.globalWorkSizeDY), runInfo.lws1);
 100         runInfo.gws2 = arg.output.Batch().v;
 101
 102         return runInfo;
 103     }
 104
 105     bool ConvolutionKernel_bfyx_GEMMLike::Validate(const Params& p, const optional_params& o) const
 106     {
 107         if (!Parent::Validate(p, o) ||
 108             !CovolutionCheckInput(p, o))
 109         {
 110             return false;
 111         }
 112
 113         const auto& params = static_cast<const convolution_params&>(p);
 114
 115         if (!params.engineInfo.bSubGroupShortSupport && params.inputs[0].GetDType() == Datatype::F16)
 116         {
 117             return false;
 118         }
 119
 120         return true;
 121     }
 122
 123     std::vector<WeightsLayout> ConvolutionKernel_bfyx_GEMMLike::GetSupportedWeightLayouts(const convolution_params& params) const
 124     {
 125         if (params.inputs[0].GetDType() == Datatype::F16)
 126         {
 127             return{ WeightsLayout::iy_xs_os_xsv2_osv16__ao32 };
 128         }
 129         else
 130         {
 131             return{ WeightsLayout::iy_xs_os_xsv2_osv8__ao32 };
 132         }
 133     }
 134
 135     KernelsData ConvolutionKernel_bfyx_GEMMLike::GetKernelsData(const Params& params, const optional_params& options) const
 136     {
 137         return GetTunedKernelsDataByIndex(params, options);
 138     }
 139 }