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 "convolution_kernel_MMAD.h"
19 namespace kernel_selector {
21 ParamsKey ConvolutionKernel_MMAD::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::byxf_af32);
29 k.EnableTensorOffset();
30 k.EnableTensorPitches();
32 k.EnableBiasPerFeature();
33 k.EnableBiasPerOutput();
34 k.EnableNonBiasTerm();
36 k.EnableSplitSupport();
37 k.EnableInt8Quantization();
38 k.EnableOutputCalibration();
43 ConvolutionKernelBase::DispatchData ConvolutionKernel_MMAD::SetDefault(const convolution_params& arg, int) const
45 DispatchData runInfo = ConvolutionKernelBase::SetDefault(arg);
47 constexpr size_t sub_group_size = 8;
49 const auto of_maps = arg.output.Feature().v;
50 const size_t of_threads_per_batch = RoundUp(of_maps, sub_group_size);
52 runInfo.effiency = FORCE_PRIORITY_4;
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;
60 runInfo.lws2 = sub_group_size;
65 JitConstants ConvolutionKernel_MMAD::GetJitConstants(const convolution_params& params, const DispatchData& runInfo) const
67 auto jit = Parent::GetJitConstants(params, runInfo);
69 jit.AddConstant(MakeJitConstant("SUB_GROUP_SIZE", runInfo.lws2));
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));
79 KernelsData ConvolutionKernel_MMAD::GetKernelsData(const Params& params, const optional_params& options) const
81 KernelsData kd = GetTunedKernelsDataByIndex(params, options);
83 kd[0].estimatedTime = FORCE_PRIORITY_4;