inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution_grad_weights/convolution_grad_weights_kernel_base.cpp

   1 /*
   2 // Copyright (c) 2018 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_grad_weights_kernel_base.h"
  18 #include "kernel_selector_utils.h"
  19
  20 namespace kernel_selector
  21 {
  22     std::string convolution_grad_weights_params::to_string() const
  23     {
  24         std::stringstream s;
  25
  26         s << base_params::to_string() << "_";
  27         if (bias.empty())
  28         {
  29             s << "no_bias" << "_";
  30         }
  31         else
  32         {
  33             s << "bias_" << bias[0].PhysicalSize() << "_";
  34         }
  35         s << filterSize.x << "_" << filterSize.y << "_";
  36         s << stride.x << "_" << stride.y << "_";
  37         s << dilation.x << "_" << dilation.y << "_";
  38         s << padding.x << "_" << padding.y << "_";
  39         s << split;
  40
  41         return s.str();
  42     }
  43
  44     JitConstants ConvolutionGradWeightsKernelBase::GetJitConstants(const convolution_grad_weights_params& cp) const
  45     {
  46         JitConstants jit = training_kernel_base::GetJitConstants(cp);
  47         const auto& padding = cp.padding;
  48         const auto& input = cp.inputs[0];
  49
  50         int64_t input_offset_with_padding = (int64_t)input.GetFirstElementOffset() - (cp.filterSize.x - 1 + padding.x)*input.X().pitch - (cp.filterSize.y - 1 + padding.y)*input.Y().pitch;
  51         input_offset_with_padding = std::max(input_offset_with_padding, (int64_t)0);
  52
  53         jit.AddConstants({
  54             MakeJitConstant("STRIDE",                       cp.stride),
  55             MakeJitConstant("PADDING",                      cp.padding),
  56             MakeJitConstant("DILATION",                     cp.dilation),
  57             MakeJitConstant("FILTER_ARRAY_NUM",             cp.split),
  58             MakeJitConstant("INPUT0_OFFSET_WITH_PADDING",   input_offset_with_padding),
  59             MakeJitConstant("DEPTHWISE_SEPARABLE_OPT",      cp.depthwise_separable_opt),
  60             MakeJitConstant("OUTPUT_GRAD_W",                cp.output_grad_w),
  61         });
  62
  63         return jit;
  64     }
  65
  66     ConvolutionGradWeightsKernelBase::DispatchData ConvolutionGradWeightsKernelBase::SetDefault(const convolution_grad_weights_params& params) const
  67     {
  68         auto input_features = params.weights.IFM().v;
  69         auto output_features = params.weights.OFM().v;
  70
  71         DispatchData kd;
  72
  73         kd.fp16UnitUsed = params.inputs[0].GetDType() == Datatype::F16;
  74         size_t gws0 = output_features * input_features;
  75         size_t lws0 = std::min(gws0, static_cast<size_t>(32));
  76         while (gws0 % lws0)
  77         {
  78             lws0--;
  79         }
  80         kd.gws0 = gws0;
  81         kd.gws1 = params.weights.X().v;
  82         kd.gws2 = params.weights.Y().v;
  83         kd.lws0 = lws0;
  84         kd.lws1 = 1;
  85         kd.lws2 = 1;
  86         kd.effiency = DONT_USE_IF_HAVE_SOMETHING_ELSE;
  87         return kd;
  88     }
  89
  90     KernelsData ConvolutionGradWeightsKernelBase::GetKernelsData(const Params& params, const optional_params& options) const
  91     {
  92         assert(params.GetType() == KernelType::CONVOLUTION_GRAD_WEIGHTS);
  93
  94         if (!Validate(params, options))
  95         {
  96             return{};
  97         }
  98
  99         const convolution_grad_weights_params& orgParams = static_cast<const convolution_grad_weights_params&>(params);
 100
 101         const std::vector<WeightsLayout> weightsLayouts = {
 102             WeightsLayout::oiyx,
 103             WeightsLayout::iyxo,
 104             WeightsLayout::yxio,
 105             WeightsLayout::oyxi
 106         };
 107
 108         DispatchData runInfo = SetDefault(orgParams);
 109         KernelData kd = KernelData::Default<convolution_grad_weights_params>(params);
 110         convolution_grad_weights_params& newParams = *static_cast<convolution_grad_weights_params*>(kd.params.get());
 111
 112         bool succeed = UpdateWeightsParams(
 113             newParams,
 114             options,
 115             weightsLayouts,
 116             kd.weightsReorderParams);
 117
 118         if (!succeed)
 119         {
 120             return{};
 121         }
 122
 123         auto cldnn_jit = GetJitConstants(orgParams);
 124         auto entry_point = GetEntryPoint(kernelName, orgParams.layerID, options);
 125         auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
 126
 127         auto& kernel = kd.kernels[0];
 128         FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point, DEFAULT, true, !orgParams.bias.empty());
 129         if (newParams.use_momentum)
 130         {
 131             kernel.arguments.push_back({ ArgumentDescriptor::Types::PREV_WEIGHTS_GRADIENT, 0 });
 132             if (!newParams.bias.empty())
 133                 kernel.arguments.push_back({ ArgumentDescriptor::Types::PREV_BIAS_GRADIENT, 0 });
 134         }
 135         kernel.arguments.push_back({ ArgumentDescriptor::Types::INPUT, 1 });
 136         kernel.arguments.push_back({ ArgumentDescriptor::Types::SPLIT, 0 });
 137         kernel.arguments.push_back({ ArgumentDescriptor::Types::LEARNING_RATE, 0 });
 138
 139         kd.estimatedTime = runInfo.effiency;
 140
 141         return{ kd };
 142     }
 143 }