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_1x1_gemm_MMAD.h"
19 namespace kernel_selector {
21 ParamsKey ConvolutionKernel_1x1_gemm_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.EnableDepthwiseSeparableOpt();
38 k.EnableInt8Quantization();
39 k.EnableOutputCalibration();
44 bool ConvolutionKernel_1x1_gemm_MMAD::Validate(const Params& p, const optional_params& o) const
46 if (!ConvolutionKernelBase::Validate(p, o))
51 const auto& params = static_cast<const convolution_params&>(p);
53 if (params.filterSize.x != 1 || params.filterSize.y != 1)
56 if (params.stride.x != 1 || params.stride.y != 1)
59 if (params.padding.x != 0 || params.padding.y != 0)
62 const auto& input = params.inputs[0];
64 // we do not support padded input
65 if (input.X().pad.Total() != 0 || input.Y().pad.Total() != 0)
68 if (params.split != 1)
74 ConvolutionKernelBase::DispatchData ConvolutionKernel_1x1_gemm_MMAD::SetDefault(const convolution_params& arg, int) const
76 DispatchData runInfo = ConvolutionKernelBase::SetDefault(arg);
78 // Sub-group size used by "convolution_1x1_gemm_MMAD" kernel.
79 constexpr size_t sub_group_size = 8;
81 const auto of_maps = arg.output.Feature().v;
82 const size_t of_threads_per_batch = RoundUp(of_maps, sub_group_size);
84 runInfo.effiency = FORCE_PRIORITY_2;
86 runInfo.gws0 = RoundUp(arg.output.X().v * arg.output.Y().v, 8) / 8;
87 runInfo.gws1 = of_threads_per_batch * arg.output.Batch().v;
91 runInfo.lws1 = sub_group_size;
97 JitConstants ConvolutionKernel_1x1_gemm_MMAD::GetJitConstants(const convolution_params& params, const DispatchData& runInfo) const
99 auto jit = Parent::GetJitConstants(params, runInfo);
101 jit.AddConstant(MakeJitConstant("SUB_GROUP_SIZE", runInfo.lws1));
103 // pitch for special block format used in this kernel
104 const size_t ifm_32_aligned = Align(params.weights.IFM().v, 32);
105 const size_t filter_ofm_block_pitch = (ifm_32_aligned / 32) * params.weights.X().v * params.weights.Y().v * 4 * 8 * 8;
106 jit.AddConstant(MakeJitConstant("FILTER_OFM_BLOCK_PITCH", filter_ofm_block_pitch));
111 KernelsData ConvolutionKernel_1x1_gemm_MMAD::GetKernelsData(const Params& params, const optional_params& options) const
113 return GetTunedKernelsDataByIndex(params, options);