inference-engine/thirdparty/clDNN/src/gpu/batch_norm_grad_gpu.cpp

   1 /*
   2 // Copyright (c) 2016 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 "batch_norm_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 "batch_norm_grad/batch_norm_grad_kernel_base.h"
  23 #include "batch_norm_grad/batch_norm_grad_kernel_selector.h"
  24
  25 namespace cldnn {
  26     namespace gpu {
  27
  28         struct batch_norm_grad_gpu : typed_primitive_gpu_impl<batch_norm_grad>
  29         {
  30             using parent = typed_primitive_gpu_impl<batch_norm_grad>;
  31             using parent::parent;
  32
  33         protected:
  34
  35             virtual kernel::kernel_arguments_data get_arguments(typed_primitive_inst<batch_norm_grad>& instance, int32_t) const override
  36             {
  37                 kernel::kernel_arguments_data args;
  38
  39                 args.inputs = { &instance.input_memory(0), &instance.input_memory(1), &instance.inv_variance_memory() };
  40                 args.output = &instance.output_memory();
  41
  42                 return args;
  43             }
  44
  45         public:
  46
  47             static primitive_impl* create(const batch_norm_grad_node &arg)
  48             {
  49                 auto norm_params = get_default_params<kernel_selector::batch_norm_grad_params>(arg);
  50                 auto norm_optional_params = get_default_optional_params<kernel_selector::batch_norm_grad_optional_params>(arg.get_program());
  51
  52
  53                 auto& kernel_selector = kernel_selector::batch_norm_grad_kernel_selector::Instance();
  54                 auto best_kernels = kernel_selector.GetBestKernels(norm_params, norm_optional_params);
  55
  56                 CLDNN_ERROR_BOOL(arg.id(), "Best_kernel.empty()", best_kernels.empty(), "Cannot find a proper kernel with this arguments");
  57
  58                 auto norm = new batch_norm_grad_gpu(arg, best_kernels[0]);
  59
  60                 return norm;
  61             };
  62         };
  63
  64         namespace {
  65             struct attach {
  66                 attach() {
  67                     auto val_fw = batch_norm_grad_gpu::create;
  68
  69                     implementation_map<batch_norm_grad>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::yxfb), val_fw);
  70                     implementation_map<batch_norm_grad>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::yxfb), val_fw);
  71                     implementation_map<batch_norm_grad>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::bfyx), val_fw);
  72                     implementation_map<batch_norm_grad>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::bfyx), val_fw);
  73                     implementation_map<batch_norm_grad>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::byxf), val_fw);
  74                     implementation_map<batch_norm_grad>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::byxf), val_fw);
  75                 }
  76                 ~attach() {}
  77             };
  78             attach attach_impl;
  79         }
  80     }
  81 }