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

   1 /*
   2 // Copyright (c) 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 "strided_slice_inst.h"
  18 #include "primitive_gpu_base.h"
  19 #include "implementation_map.h"
  20 #include "kernel_selector_helper.h"
  21 #include "strided_slice/strided_slice_kernel_ref.h"
  22 #include "strided_slice/strided_slice_kernel_selector.h"
  23 #include "error_handler.h"
  24 #include "data_inst.h"
  25
  26 using namespace cldnn;
  27
  28 namespace cldnn
  29 {
  30 namespace gpu
  31 {
  32
  33 struct strided_slice_gpu : typed_primitive_gpu_impl<strided_slice>
  34 {
  35     using parent = typed_primitive_gpu_impl<strided_slice>;
  36     using parent::parent;
  37 public:
  38     static primitive_impl* create(const strided_slice_node& arg)
  39     {
  40         auto strided_slice_params = get_default_params<kernel_selector::strided_slice_params>(arg);
  41         auto strided_slice_optional_params = get_default_optional_params<kernel_selector::strided_slice_optional_params>(arg.get_program());
  42         const int32_t numberOfDims = 4;
  43
  44         auto complete_strided_slice_params = [&](std::vector<int32_t>& param) {
  45             for (size_t i = param.size(); i < numberOfDims; ++i)
  46                 param.push_back(1);
  47         };
  48
  49         auto completeStridedSliceMasks = [&](std::vector<uint8_t>& mask) {
  50             for (size_t i = mask.size(); i < numberOfDims; ++i)
  51                 mask.push_back(0);
  52         };
  53
  54         // Getting data from constant inputs. There are 3 args: Begin, End, Stride
  55         for (size_t i = 1; i < arg.get_dependencies().size(); ++i) {
  56             auto& input = arg.get_dependency(i).as<data>();
  57             auto& mem = input.get_attached_memory();
  58             int32_t* data = static_cast<int32_t*>(mem.lock());
  59             std::vector<int32_t> vData = std::vector<int32_t>(data, data + input.get_output_layout().count());
  60             complete_strided_slice_params(vData);
  61             strided_slice_params.striding_params.push_back(vData);
  62             mem.unlock();
  63         }
  64
  65         strided_slice_params.end_mask = arg.get_primitive()->end_mask;
  66         completeStridedSliceMasks(strided_slice_params.end_mask);
  67         strided_slice_params.begin_mask = arg.get_primitive()->begin_mask;
  68         completeStridedSliceMasks(strided_slice_params.begin_mask);
  69         strided_slice_params.new_axis_mask = arg.get_primitive()->new_axis_mask;
  70         strided_slice_params.shrink_axis_mask = arg.get_primitive()->shrink_axis_mask;
  71         completeStridedSliceMasks(strided_slice_params.shrink_axis_mask);
  72
  73         auto& kernel_selector = kernel_selector::strided_slice_kernel_selector::Instance();
  74         auto best_kernels = kernel_selector.GetBestKernels(strided_slice_params, strided_slice_optional_params);
  75
  76         CLDNN_ERROR_BOOL(arg.id(), "Best_kernel.empty()", best_kernels.empty(), "Cannot find a proper kernel with this arguments");
  77
  78         auto strided_slice = new strided_slice_gpu(arg, best_kernels[0]);
  79
  80         return strided_slice;
  81     }
  82 };
  83
  84 namespace
  85 {
  86     struct attach
  87     {
  88         attach()
  89         {
  90             auto val_fw = strided_slice_gpu::create;
  91             implementation_map<strided_slice>::add(std::make_tuple(engine_types::ocl, data_types::f32, format::bfyx), val_fw);
  92             implementation_map<strided_slice>::add(std::make_tuple(engine_types::ocl, data_types::f16, format::bfyx), val_fw);
  93         }
  94         ~attach() = default;
  95     };
  96     attach attach_impl;
  97 }
  98 } //namespace gpu
  99 } //namespace cldnn