You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

expanddims_kernel.cc 3.3 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687
  1. /**
  2. * Copyright 2019-2020 Huawei Technologies Co., Ltd
  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. #include "host_kernels/expanddims_kernel.h"
  17. #include <memory>
  18. #include "common/ge_inner_error_codes.h"
  19. #include "common/op/ge_op_utils.h"
  20. #include "common/types.h"
  21. #include "framework/common/debug/ge_log.h"
  22. #include "host_kernels/kernel_utils.h"
  23. #include "inc/kernel_factory.h"
  24. namespace ge {
  25. namespace {
  26. const int kExpandDimsIndexZero = 0;
  27. const size_t kExpandDimsOutputDescNum = 1;
  28. const size_t kExpandDimsInputNum = 2;
  29. } // namespace
  30. Status ExpanddimsKernel::Compute(const NodePtr &node_ptr) {
  31. GELOGI("Expanddims dimension kernel in.");
  32. if (node_ptr == nullptr) {
  33. GELOGE(PARAM_INVALID, "parameter is nullptr");
  34. return PARAM_INVALID;
  35. }
  36. Status ret = KernelUtils::CheckDimensionNodeInfo(node_ptr);
  37. if (ret != SUCCESS) {
  38. GELOGW("GetDimensionNodeInfo failed");
  39. return ret;
  40. }
  41. if (!KernelUtils::CheckFormatSupported(node_ptr)) {
  42. GELOGW("CheckFormatSupported failed");
  43. return NOT_CHANGED;
  44. }
  45. GELOGI("Expanddims dimension kernel success.");
  46. return SUCCESS;
  47. }
  48. Status ExpanddimsKernel::Compute(const ge::OpDescPtr op_desc_ptr, const std::vector<ge::ConstGeTensorPtr> &input,
  49. std::vector<ge::GeTensorPtr> &v_output) {
  50. GELOGI("Expanddims folding kernel in.");
  51. if (op_desc_ptr == nullptr) {
  52. GELOGE(PARAM_INVALID, "Input opdesc is nullptr.");
  53. return PARAM_INVALID;
  54. }
  55. if ((input.size() != kExpandDimsInputNum) || (op_desc_ptr->GetOutputsSize() != kExpandDimsOutputDescNum)) {
  56. GELOGW("Unexpected ExpandDims node, node input size: %zu, node output size: %zu, node name: %s", input.size(),
  57. op_desc_ptr->GetOutputsSize(), op_desc_ptr->GetName().c_str());
  58. return NOT_CHANGED;
  59. }
  60. auto output_tensor_desc = op_desc_ptr->GetOutputDesc(kExpandDimsIndexZero);
  61. GeTensorPtr output_ptr = MakeShared<GeTensor>(output_tensor_desc);
  62. if (output_ptr == nullptr) {
  63. GELOGW("Failed to fold node %s, out of memory", op_desc_ptr->GetName().c_str());
  64. return NOT_CHANGED;
  65. }
  66. // print output tensor information, and will be deleted
  67. GELOGI("Expanddims op %s output tensor data size is %zu", op_desc_ptr->GetName().c_str(),
  68. output_ptr->GetData().size());
  69. size_t data_dim_size = output_ptr->GetTensorDesc().GetShape().GetDims().size();
  70. GELOGI("Expanddims op %s output tensor dim size is %zu", op_desc_ptr->GetName().c_str(), data_dim_size);
  71. if (output_ptr->SetData(input.at(kExpandDimsIndexZero)->GetData()) != GRAPH_SUCCESS) {
  72. GELOGW("Compute: SetData failed");
  73. }
  74. v_output.emplace_back(output_ptr);
  75. GELOGI("Expanddims folding kernel success.");
  76. return SUCCESS;
  77. }
  78. REGISTER_KERNEL(EXPANDDIMS, ExpanddimsKernel);
  79. } // namespace ge

图引擎模块(GE)是MindSpore的一个子模块,其代码由C++实现,位于前端模块ME和底层硬件之间,起到承接作用。图引擎模块以ME下发的图作为输入,然后进行一系列的深度图优化操作,最后输出一张可以在底层硬件上高效运行的图。GE针对昇腾AI处理器的硬件结构特点,做了特定的优化工作,以此来充分发挥出昇腾AI处理器的强大算力。在进行模型训练/推理时,GE会被自动调用而用户并不感知。GE主要由GE API和GE Core两部分组成,详细的架构图如下所示