2 // Copyright (c) 2016 Intel Corporation
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
8 // http://www.apache.org/licenses/LICENSE-2.0
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.
17 #include "fully_connected_kernel_MMAD.h"
19 namespace kernel_selector
21 ParamsKey FullyConnectedKernelMMAD::GetSupportedKey() const
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::bf);
29 k.EnableOutputLayout(DataLayout::fb);
30 k.EnableBiasPerOutput();
31 k.EnableBiasPerFeature();
32 k.EnableNonBiasTerm();
33 k.EnableTensorOffset();
34 k.EnableTensorPitches();
36 k.EnableInt8Quantization();
37 k.EnableOutputCalibration();
41 FullyConnectedKernelMMAD::DispatchData FullyConnectedKernelMMAD::SetDefault(const fully_connected_params& params, int) const
43 auto runInfo = Parent::SetDefault(params);
45 constexpr size_t sub_group_size = 8;
46 const auto of_maps = params.output.Feature().v;
47 const size_t of_threads_per_batch = RoundUp(of_maps, sub_group_size);
51 runInfo.gws2 = of_threads_per_batch * params.output.Batch().v;
55 runInfo.lws2 = sub_group_size;
60 JitConstants FullyConnectedKernelMMAD::GetJitConstants(const fully_connected_params& params, const DispatchData& runInfo) const
62 auto jit = Parent::GetJitConstants(params, runInfo);
64 jit.AddConstant(MakeJitConstant("SUB_GROUP_SIZE", runInfo.lws2));
66 // pitch for special block format used in this kernel
67 const size_t ifm_32_aligned = Align(params.weights.IFM().v, 32);
68 const size_t filter_ofm_block_pitch = (ifm_32_aligned / 32) * params.weights.X().v * params.weights.Y().v * 4 * 8 * 8;
69 jit.AddConstant(MakeJitConstant("FILTER_OFM_BLOCK_PITCH", filter_ofm_block_pitch));
74 KernelsData FullyConnectedKernelMMAD::GetKernelsData(const Params& params, const optional_params& options) const
78 for (size_t i = 0; i < autoTuneOptions.size(); i++)
80 KernelsData kd = GetTunedKernelsDataByIndex(params, options, DataLayout::byxf_af32,
81 { WeightsLayout::os_is_yx_isa8_osv8_isv4 }, DONT_USE_IF_HAVE_SOMETHING_ELSE, (int)i);
84 res.emplace_back(kd[0]);