inference-engine/thirdparty/clDNN/src/gpu/activation_grad_gpu.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 "activation_grad_inst.h"
  18 #include "primitive_gpu_base.h"
  19 #include "implementation_map.h"
  20 #include "error_handler.h"
  21 #include "kernel_selector_helper.h"
  22 #include "activation/activation_kernel_selector.h"
  23 #include "activation/activation_kernel_base.h"
  24 #include "api/CPP/activation_grad.hpp"
  25
  26 namespace cldnn { namespace gpu {
  27
  28
  29 struct activation_grad_gpu : typed_primitive_gpu_impl<activation_grad>
  30 {
  31     using parent = typed_primitive_gpu_impl<activation_grad>;
  32     using parent::parent;
  33
  34     virtual kernel::kernel_arguments_data get_arguments(typed_primitive_inst<activation_grad>& instance, int32_t split) const override
  35     {
  36         kernel::kernel_arguments_data args = parent::get_arguments(instance, split);
  37
  38         if (_outer.is_parameterized())
  39         {
  40             args.slope = &instance.slope_memory();
  41         }
  42
  43         return args;
  44     }
  45
  46     static primitive_impl* create(const activation_grad_node& arg)
  47     {
  48         auto activation_grad_params = get_default_params<kernel_selector::activation_params>(arg);
  49         auto activation_grad_optional_params = get_default_optional_params<kernel_selector::activation_optional_params>(arg.get_program());
  50
  51         const auto& primitive = arg.get_primitive();
  52
  53         activation_grad_params.gradient = true;
  54         activation_grad_params.inputs.push_back(convert_data_tensor(arg.get_dependency(1).get_output_layout()));
  55         activation_grad_params.activation.function = get_kernel_selector_activation_grad_param(primitive->activation_grad_func);
  56         activation_grad_params.activation.m = primitive->additional_params.a;
  57         activation_grad_params.activation.n = primitive->additional_params.b;
  58
  59         if (arg.is_parameterized())
  60         {
  61             const auto& slope_layout = arg.slope_input().get_output_layout();
  62             const auto& output_layout = arg.get_output_layout();
  63
  64             const auto params_num = kernel_selector::GetActivationAdditionalParamsNumber(activation_grad_params.activation.function);
  65
  66             CLDNN_ERROR_LESS_THAN(arg.id(), "Slope layout size count", slope_layout.size.count(), "output_layout.size.feature[0] * params_num", static_cast<size_t>(output_layout.size.feature[0] * params_num), "Error - not enough data inside additional params buffer");
  67
  68             activation_grad_params.inputActivationParams.push_back(convert_data_tensor(slope_layout));
  69         }
  70
  71         auto& kernel_selector = kernel_selector::activation_kernel_selector::Instance();
  72         auto best_kernels = kernel_selector.GetBestKernels(activation_grad_params, activation_grad_optional_params);
  73         CLDNN_ERROR_BOOL(arg.id(), "Best_kernel.empty()", best_kernels.empty(), "Cannot find a proper kernel with this arguments");
  74
  75         auto activation_grad = new activation_grad_gpu(arg, best_kernels[0]);
  76
  77         return activation_grad;
  78     };
  79 };
  80
  81
  82 namespace {
  83     struct attach {
  84         attach() {
  85             auto val_fw = activation_grad_gpu::create;
  86
  87             implementation_map<activation_grad>::add({
  88                 { std::make_tuple(engine_types::ocl, data_types::f32, format::yxfb), val_fw},
  89                 { std::make_tuple(engine_types::ocl, data_types::f16, format::yxfb), val_fw},
  90                 { std::make_tuple(engine_types::ocl, data_types::f32, format::bfyx), val_fw },
  91                 { std::make_tuple(engine_types::ocl, data_types::f16, format::bfyx), val_fw },
  92                 { std::make_tuple(engine_types::ocl, data_types::f32, format::byxf), val_fw },
  93                 { std::make_tuple(engine_types::ocl, data_types::f16, format::byxf), val_fw },
  94             });
  95         }
  96         ~attach() {}
  97     };
  98     attach attach_impl;
  99 }
 100 } }