inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_MMAD.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_MMAD.h"
  18
  19 namespace kernel_selector {
  20
  21     ParamsKey ConvolutionKernel_MMAD::GetSupportedKey() const
  22     {
  23         ParamsKey k;
  24         k.EnableInputDataType(Datatype::INT8);
  25         k.EnableOutputDataType(Datatype::INT8);
  26         k.EnableInputWeightsType(WeightsType::INT8);
  27         k.EnableInputLayout(DataLayout::byxf_af32);
  28         k.EnableOutputLayout(DataLayout::byxf_af32);
  29         k.EnableTensorOffset();
  30         k.EnableTensorPitches();
  31         k.EnableDilation();
  32         k.EnableBiasPerFeature();
  33         k.EnableBiasPerOutput();
  34         k.EnableNonBiasTerm();
  35         k.EnableBatching();
  36         k.EnableSplitSupport();
  37         k.EnableInt8Quantization();
  38         k.EnableOutputCalibration();
  39         k.DisableTuning();
  40         return k;
  41     }
  42
  43     ConvolutionKernelBase::DispatchData ConvolutionKernel_MMAD::SetDefault(const convolution_params& arg, int) const
  44     {
  45         DispatchData runInfo = ConvolutionKernelBase::SetDefault(arg);
  46
  47         constexpr size_t sub_group_size = 8;
  48
  49         const auto of_maps = arg.output.Feature().v;
  50         const size_t of_threads_per_batch = RoundUp(of_maps, sub_group_size);
  51
  52         runInfo.effiency = FORCE_PRIORITY_4;
  53
  54         runInfo.gws0 = arg.output.X().v;
  55         runInfo.gws1 = arg.output.Y().v;
  56         runInfo.gws2 = of_threads_per_batch * arg.output.Batch().v;
  57
  58         runInfo.lws0 = 1;
  59         runInfo.lws1 = 1;
  60         runInfo.lws2 = sub_group_size;
  61
  62         return runInfo;
  63     }
  64
  65     JitConstants ConvolutionKernel_MMAD::GetJitConstants(const convolution_params& params, const DispatchData& runInfo) const
  66     {
  67         auto jit = Parent::GetJitConstants(params, runInfo);
  68
  69         jit.AddConstant(MakeJitConstant("SUB_GROUP_SIZE", runInfo.lws2));
  70
  71         // pitch for special block format used in this kernel
  72         const size_t ifm_32_aligned = Align(params.weights.IFM().v, 32);
  73         const size_t filter_ofm_block_pitch = (ifm_32_aligned / 32) * params.weights.X().v * params.weights.Y().v * 4 * 8 * 8;
  74         jit.AddConstant(MakeJitConstant("FILTER_OFM_BLOCK_PITCH", filter_ofm_block_pitch));
  75
  76         return jit;
  77     }
  78
  79     KernelsData ConvolutionKernel_MMAD::GetKernelsData(const Params& params, const optional_params& options) const
  80     {
  81         KernelsData kd = GetTunedKernelsDataByIndex(params, options);
  82         if(!kd.empty())
  83             kd[0].estimatedTime = FORCE_PRIORITY_4;
  84         return kd;
  85     }
  86 }