inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/fully_connected/fully_connected_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 "fully_connected_kernel_MMAD.h"
  18 #include "kernel_selector_utils.h"
  19
  20 namespace kernel_selector
  21 {
  22     ParamsKey FullyConnectedKernelMMAD::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::byxf_af32);
  29         k.EnableOutputLayout(DataLayout::bf);
  30         k.EnableOutputLayout(DataLayout::fb);
  31         k.EnableBiasPerOutput();
  32         k.EnableBiasPerFeature();
  33         k.EnableNonBiasTerm();
  34         k.EnableTensorOffset();
  35         k.EnableTensorPitches();
  36         k.EnableBatching();
  37         k.EnableInt8Quantization();
  38         k.EnableOutputCalibration();
  39         return k;
  40     }
  41
  42     std::unique_ptr<FullyConnectedKernelMMAD::Parent::DispatchData> FullyConnectedKernelMMAD::SetDefault(const fully_connected_params& params) const
  43     {
  44         auto runInfo = Parent::SetDefault(params);
  45
  46         constexpr size_t sub_group_size = 8;
  47         const auto of_maps = params.output.Feature().v;
  48         const size_t of_threads_per_batch = RoundUp(of_maps, sub_group_size);
  49
  50         runInfo->gws0 = 1;
  51         runInfo->gws1 = 1;
  52         runInfo->gws2 = of_threads_per_batch * params.output.Batch().v;
  53
  54         runInfo->lws0 = 1;
  55         runInfo->lws1 = 1;
  56         runInfo->lws2 = sub_group_size;
  57
  58         return std::move(runInfo);
  59     }
  60
  61     JitConstants FullyConnectedKernelMMAD::GetJitConstants(const fully_connected_params& params, const DispatchData& runInfo) const
  62     {
  63         auto jit = Parent::GetJitConstants(params, runInfo);
  64
  65         jit.AddConstant(MakeJitConstant("SUB_GROUP_SIZE", runInfo.lws2));
  66
  67         // pitch for special block format used in this kernel
  68         const size_t ifm_32_aligned = Align(params.weights.IFM().v, 32);
  69         const size_t filter_ofm_block_pitch = (ifm_32_aligned / 32) * params.weights.X().v * params.weights.Y().v * 4 * 8 * 8;
  70         jit.AddConstant(MakeJitConstant("FILTER_OFM_BLOCK_PITCH", filter_ofm_block_pitch));
  71
  72         return jit;
  73     }
  74
  75     KernelsData FullyConnectedKernelMMAD::GetKernelsData(const Params& params, const optional_params& options) const
  76     {
  77         return GetCommonKernelsData(params, options, DataLayout::byxf_af32,
  78         { WeightsLayout::os_is_yx_isa8_osv8_isv4 }
  79         );
  80     }
  81 }