inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/fused_conv_eltwise/fused_conv_eltwise_kernel_gemm.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 "fused_conv_eltwise_kernel_gemm.h"
  18 #include "kernel_selector_utils.h"
  19
  20 namespace kernel_selector {
  21
  22     ParamsKey fused_conv_eltwise_kernel_gemm::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.EnableFusedConvEltwSplitSupport();
  41         return k;
  42         }
  43
  44     std::string fused_conv_eltwise_kernel_gemm::GetKernelName(const fused_conv_eltwise_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         bool fused_conv_eltwise_kernel_gemm::Validate(const Params& p, const optional_params& o) const
  57         {
  58                 if (!fused_conv_eltwise_kernel_base::Validate(p, o) ||
  59                         !FusedConvolutionEltwiseCheckInput(p, o))
  60                 {
  61                         return false;
  62                 }
  63
  64                 const convolution_params& cp = static_cast<const convolution_params&>(p);
  65
  66         // make sure it's 1x1 conv
  67         if (cp.filterSize.x != 1 || cp.filterSize.y != 1)
  68             return false;
  69
  70         // make sure stride is 1x1
  71         if (cp.stride.x != 1 || cp.stride.y != 1)
  72             return false;
  73
  74         // input padding not supported
  75         if (cp.inputs[0].X().pad.Total() != 0 ||
  76             cp.inputs[0].Y().pad.Total() != 0 ||
  77             cp.inputs[0].Feature().pad.Total() != 0 ||
  78             cp.inputs[0].Batch().pad.Total() != 0)
  79             return false;
  80
  81         // input and output spatial sizes must match
  82         if (!(cp.output.X().v == cp.inputs[0].X().v) || !(cp.output.Y().v == cp.inputs[0].Y().v))
  83             return false;
  84
  85                 return true;
  86         }
  87
  88     std::vector<WeightsLayout> fused_conv_eltwise_kernel_gemm::GetSupportedWeightLayouts(const fused_conv_eltwise_params& params) const
  89     {
  90         if (params.inputs[0].GetDType() == Datatype::F16)
  91         {
  92             return{ WeightsLayout::iy_xs_os_xsv2_osv16__ao32 };
  93         }
  94         else
  95         {
  96             return{ WeightsLayout::iy_xs_os_xsv2_osv8__ao32 };
  97         }
  98     }
  99
 100     fused_conv_eltwise_kernel_base::DispatchData fused_conv_eltwise_kernel_gemm::SetDefault(const fused_conv_eltwise_params& arg, int) const
 101         {
 102         DispatchData runInfo = Parent::SetDefault(arg);
 103
 104         runInfo.lws0 = 1;
 105         runInfo.lws2 = 1;
 106
 107         if (arg.inputs[0].GetDType() == Datatype::F16)
 108         {
 109             runInfo.gemmStyle = { 1, arg.conv.filterSize.x, 32, 32, 1, 1 };
 110             runInfo.lws1 = 16;
 111             runInfo.effiency = FORCE_PRIORITY_6;
 112         }
 113         else
 114         {
 115             runInfo.gemmStyle = { 2, arg.conv.filterSize.x, 32, 32, 2, 1 };
 116             runInfo.lws1 = 8;
 117             runInfo.effiency = FORCE_PRIORITY_8;
 118         }
 119
 120         size_t sgemm_m = RoundUp(arg.output.X().v * arg.output.Y().v, runInfo.gemmStyle.subBlockDimM);
 121         size_t sgemm_n = RoundUp(arg.output.Feature().v, runInfo.gemmStyle.subBlockDimN);
 122
 123         runInfo.gws0 = RoundUp(CeilDiv(sgemm_n, runInfo.gemmStyle.globalWorkSizeDX), runInfo.lws0);
 124         runInfo.gws1 = RoundUp(CeilDiv(sgemm_m, runInfo.gemmStyle.globalWorkSizeDY), runInfo.lws1);
 125         runInfo.gws2 = arg.output.Batch().v;
 126
 127         return runInfo;
 128         }
 129
 130         JitConstants fused_conv_eltwise_kernel_gemm::GetJitConstants(const fused_conv_eltwise_params& params, const DispatchData& runInfo) const
 131         {
 132                 auto jit = Parent::GetJitConstants(params, runInfo);
 133
 134         jit.AddConstants({
 135             MakeJitConstant("ALIGNED_OFM",                  RoundUp(params.output.Feature().v, runInfo.gemmStyle.subBlockDimN)),
 136             MakeJitConstant("DX",                           runInfo.gemmStyle.globalWorkSizeDX),
 137             MakeJitConstant("DY",                           runInfo.gemmStyle.globalWorkSizeDY),
 138             MakeJitConstant("FILTER_SIZE_X_DIV2",           params.conv.filterSize.x / 2),
 139             MakeJitConstant("INPUT_BUFFER_WIDTH_PADDED",    ""),    // TODO: enable non padding path again
 140             MakeJitConstant("INPUT_BUFFER_HEIGHT_PADDED",   ""),
 141             });
 142
 143         if (CeilDiv(RoundUp(params.output.X().v * params.output.Y().v, runInfo.gemmStyle.subBlockDimM), runInfo.gemmStyle.globalWorkSizeDY) % runInfo.lws1 != 0)
 144             jit.AddConstant(MakeJitConstant("LEFTOVERS", 1));
 145
 146         if (!params.eltw.stride.empty())
 147         {
 148             jit.AddConstant(MakeJitConstant("ELTW_STRIDE_X", params.eltw.stride[0].x));
 149             jit.AddConstant(MakeJitConstant("ELTW_STRIDE_Y", params.eltw.stride[0].y));
 150         }
 151         else
 152         {
 153             jit.AddConstant(MakeJitConstant("ELTW_STRIDE_X", 1));
 154             jit.AddConstant(MakeJitConstant("ELTW_STRIDE_Y", 1));
 155         }
 156
 157                 return jit;
 158         }
 159
 160     KernelsData fused_conv_eltwise_kernel_gemm::GetKernelsData(const Params& params, const optional_params& options) const
 161     {
 162         return GetTunedKernelsDataByIndex(params, options);
 163     }
 164 }