inference-engine/thirdparty/clDNN/src/gpu/reorder_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 "reorder_inst.h"
  18 #include "primitive_gpu_base.h"
  19 #include "implementation_map.h"
  20 #include "kernel_selector_helper.h"
  21 #include "reorder/reorder_kernel_selector.h"
  22 #include "reorder/reorder_kernel_base.h"
  23 #include "error_handler.h"
  24
  25 namespace cldnn {
  26 namespace gpu {
  27
  28 struct reorder_gpu : typed_primitive_gpu_impl<reorder> {
  29     using parent = typed_primitive_gpu_impl<reorder>;
  30     using parent::parent;
  31
  32 protected:
  33     bool optimized_out(reorder_inst& instance) const override {
  34         return parent::optimized_out(instance) || _outer.can_be_optimized();
  35     }
  36
  37     kernel::kernel_arguments_data get_arguments(reorder_inst& instance, int32_t split) const override {
  38         kernel::kernel_arguments_data args = parent::get_arguments(instance, split);
  39
  40         if (_outer.has_mean()) {
  41             args.bias = (memory_impl::cptr) &instance.mean_memory();
  42         }
  43
  44         return args;
  45     }
  46
  47 public:
  48     static primitive_impl* create(const reorder_node& arg) {
  49         auto&& input_layout = arg.input().get_output_layout();
  50         auto&& output_layout = arg.get_output_layout();
  51
  52         auto reorder_params = get_default_params<kernel_selector::reorder_params>(arg);
  53         auto reorder_optional_params =
  54             get_default_optional_params<kernel_selector::reorder_optional_params>(arg.get_program());
  55
  56         if (arg.get_output_layout().data_padding) {
  57             reorder_params.has_padded_output = true;
  58         }
  59
  60         if (arg.has_mean()) {
  61             const auto& mean_layout = arg.mean().get_output_layout();
  62             reorder_params.mean = convert_data_tensor(mean_layout);
  63             reorder_params.mode = kernel_selector::mean_subtruct_mode::IN_BUFFER;
  64         } else if (arg.get_primitive()->subtract_per_feature.empty() == false) {
  65             reorder_params.mode = kernel_selector::mean_subtruct_mode::INSIDE_PARAMS;
  66             reorder_params.meanValues = arg.get_primitive()->subtract_per_feature;
  67         } else {
  68             reorder_params.mode = kernel_selector::mean_subtruct_mode::NONE;
  69         }
  70
  71         if (reorder_params.mode != kernel_selector::mean_subtruct_mode::NONE) {
  72             switch (arg.get_primitive()->mean_mode) {
  73                 case cldnn_reorder_mean_mode::mean_none:
  74                     reorder_params.mean_op = kernel_selector::mean_op::NONE;
  75                     break;
  76                 case cldnn_reorder_mean_mode::mean_mul:
  77                     reorder_params.mean_op = kernel_selector::mean_op::MUL;
  78                     break;
  79                 case cldnn_reorder_mean_mode::mean_subtract:
  80                     reorder_params.mean_op = kernel_selector::mean_op::SUB;
  81                     break;
  82                 case cldnn_reorder_mean_mode::mean_div:
  83                     reorder_params.mean_op = kernel_selector::mean_op::DIV;
  84                     break;
  85                 default:
  86                     throw std::out_of_range(arg.id() + ": unsupported mean_mode value.");
  87             }
  88         }
  89
  90         if (output_layout.format == format::winograd_2x3_s1_data) {
  91             reorder_params.winograd_input_offset_x = arg.get_input_offset().spatial[0];
  92             reorder_params.winograd_input_offset_y = arg.get_input_offset().spatial[1];
  93             reorder_params.winograd_nr_tiles_x = ceil_div(output_layout.size.spatial[0], 4);
  94         }
  95
  96         reorder_params.winograd = input_layout.format.is_winograd() || output_layout.format.is_winograd();
  97
  98         auto& kernel_selector = kernel_selector::reorder_kernel_selector::Instance();
  99         auto best_kernels = kernel_selector.GetBestKernels(reorder_params, reorder_optional_params);
 100
 101         CLDNN_ERROR_BOOL(arg.id(),
 102                          "Best_kernel.empty()",
 103                          best_kernels.empty(),
 104                          "Cannot find a proper kernel with this arguments");
 105
 106         auto reorder = new reorder_gpu(arg, best_kernels[0]);
 107
 108         return reorder;
 109     }
 110 };
 111
 112 namespace {
 113 struct attach {
 114     attach() { implementation_map<reorder>::add({{engine_types::ocl, reorder_gpu::create}}); }
 115     ~attach() {}
 116 };
 117
 118 attach attach_impl;
 119
 120 }  // namespace
 121 }  // namespace gpu
 122 }  // namespace cldnn