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

   1 /*
   2 // Copyright (c) 2016-2019 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 "softmax_inst.h"
  18 #include "primitive_gpu_base.h"
  19 #include "implementation_map.h"
  20 #include "kernel_selector_helper.h"
  21 #include "softmax/softmax_kernel_selector.h"
  22 #include "softmax/softmax_kernel_base.h"
  23 #include "error_handler.h"
  24
  25 namespace cldnn {
  26 namespace gpu {
  27
  28 struct softmax_gpu : typed_primitive_gpu_impl<softmax> {
  29     using parent = typed_primitive_gpu_impl<softmax>;
  30     using parent::parent;
  31
  32     static primitive_impl* create(const softmax_node& arg) {
  33         auto sm_params = get_default_params<kernel_selector::softmax_params>(arg);
  34         auto sm_optional_params =
  35             get_default_optional_params<kernel_selector::softmax_optional_params>(arg.get_program());
  36
  37         auto& input = sm_params.inputs[0];
  38         auto& output = sm_params.output;
  39         const auto primitive = arg.get_primitive();
  40
  41         switch (primitive->dimension) {
  42             case softmax::normalize_x:
  43                 sm_params.dim = kernel_selector::softmax_dim::X;
  44                 break;
  45
  46             case softmax::normalize_y:
  47                 sm_params.dim = kernel_selector::softmax_dim::Y;
  48                 break;
  49
  50             case softmax::normalize_fyx:
  51                 // Flatten fused with softmax
  52                 input = input.FlattenFeatureAndSpatials();
  53                 output = output.FlattenFeatureAndSpatials();
  54
  55                 sm_params.dim = kernel_selector::softmax_dim::FEATURE;
  56                 break;
  57
  58             case softmax::normalize_f:
  59                 sm_params.dim = kernel_selector::softmax_dim::FEATURE;
  60                 break;
  61
  62             case softmax::normalize_z:
  63                 sm_params.dim = kernel_selector::softmax_dim::Z;
  64                 break;
  65
  66             case softmax::normalize_all:
  67                 input = input.FlattenEverything();
  68                 output = output.FlattenEverything();
  69
  70                 sm_params.dim = kernel_selector::softmax_dim::FEATURE;
  71                 break;
  72
  73             default:
  74                 throw std::runtime_error("Wrong API - no such softmax");
  75         }
  76
  77         auto& kernel_selector = kernel_selector::softmax_kernel_selector::Instance();
  78         auto best_kernels = kernel_selector.GetBestKernels(sm_params, sm_optional_params);
  79
  80         CLDNN_ERROR_BOOL(arg.id(),
  81                          "Best_kernel.empty()",
  82                          best_kernels.empty(),
  83                          "Cannot find a proper kernel with this arguments");
  84
  85         auto softmax_node = new softmax_gpu(arg, best_kernels[0]);
  86
  87         return softmax_node;
  88     }
  89 };
  90
  91 namespace {
  92 struct attach {
  93     attach() {
  94         auto val_fw = softmax_gpu::create;
  95         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::yxfb), val_fw);
  96         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::yxfb), val_fw);
  97         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::bfyx), val_fw);
  98         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::bfyx), val_fw);
  99         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::byxf), val_fw);
 100         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::byxf), val_fw);
 101         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::bfzyx), val_fw);
 102         implementation_map<softmax>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::bfzyx), val_fw);
 103     }
 104     ~attach() {}
 105 };
 106
 107 attach attach_impl;
 108 }  // namespace
 109
 110 }  // namespace gpu
 111 }  // namespace cldnn