inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_byx8_f4__fs_bs_yx_bsv4_fsv32.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_byx8_f4__fs_bs_yx_bsv4_fsv32.h"
  18 #include "kernel_selector_utils.h"
  19
  20 namespace kernel_selector {
  21
  22     ParamsKey ConvolutionKernel_byx8_f4__fs_bs_yx_bsv4_fsv32::GetSupportedKey() const
  23     {
  24         ParamsKey k;
  25         k.EnableInputDataType(Datatype::INT8);
  26         k.EnableOutputDataType(Datatype::INT8);
  27         k.EnableInputWeightsType(WeightsType::INT8);
  28         k.EnableInputLayout(DataLayout::byx8_f4);
  29         k.EnableOutputLayout(DataLayout::fs_bs_yx_bsv4_fsv32);
  30         k.EnableTensorOffset();
  31         k.EnableTensorPitches();
  32         k.EnableDilation();
  33         k.EnableBiasPerFeature();
  34         k.EnableBatching();
  35         k.EnableInt8Quantization();
  36         k.EnableOutputCalibration();
  37         k.DisableTuning();
  38         return k;
  39     }
  40
  41     bool ConvolutionKernel_byx8_f4__fs_bs_yx_bsv4_fsv32::Validate(const Params& p, const optional_params& o) const
  42     {
  43         if (!Parent::Validate(p, o))
  44         {
  45             return false;
  46         }
  47
  48         const convolution_params& params = static_cast<const convolution_params&>(p);
  49
  50         // this kernel is designed for quantization use case
  51         if (!params.int8_quantization)
  52             return false;
  53
  54         return true;
  55     }
  56
  57     ConvolutionKernelBase::DispatchData ConvolutionKernel_byx8_f4__fs_bs_yx_bsv4_fsv32::SetDefault(const convolution_params& arg, int) const
  58     {
  59         DispatchData runInfo = ConvolutionKernelBase::SetDefault(arg);
  60
  61         runInfo.effiency = FORCE_PRIORITY_1;
  62
  63         runInfo.gws0 = (arg.output.Batch().v * arg.output.Feature().v) / 4;
  64         runInfo.gws1 = arg.output.X().v / 8;
  65         runInfo.gws2 = arg.output.Y().v / 4;
  66
  67         runInfo.lws0 = 8;
  68         runInfo.lws1 = 1;
  69         runInfo.lws2 = 1;
  70
  71         return runInfo;
  72     }
  73
  74     KernelsData ConvolutionKernel_byx8_f4__fs_bs_yx_bsv4_fsv32::GetKernelsData(const Params& params, const optional_params& options) const
  75     {
  76         KernelsData kd = GetCommonKernelsData(params, options, " -Dcl_intel_subgroups_char");
  77         if (!kd.empty())
  78             kd[0].estimatedTime = FORCE_PRIORITY_3;
  79         return kd;
  80     }
  81 }