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.

rank_kernel.cc 2.1 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566
  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/rank_kernel.h"
  17. #include <memory>
  18. #include <vector>
  19. #include "graph/types.h"
  20. #include "common/ge_inner_error_codes.h"
  21. #include "common/op/ge_op_utils.h"
  22. #include "framework/common/debug/ge_log.h"
  23. #include "inc/kernel_factory.h"
  24. #include "omg/omg_inner_types.h"
  25. namespace {
  26. const size_t kRankInputSize = 1;
  27. const uint32_t kRankDataInputIndex = 0;
  28. } // namespace
  29. namespace ge {
  30. Status RankKernel::Compute(const NodePtr &node, std::vector<GeTensorPtr> &v_output) {
  31. if (node == nullptr) {
  32. GELOGE(FAILED, "parameter is null.");
  33. return FAILED;
  34. }
  35. OpDescPtr op_desc = node->GetOpDesc();
  36. GE_CHECK_NOTNULL(op_desc);
  37. size_t input_node_size = op_desc->GetInputsSize();
  38. if (input_node_size != kRankInputSize) {
  39. GELOGW("input node size must be %d", kRankInputSize);
  40. return NOT_CHANGED;
  41. }
  42. const auto &input_shape = op_desc->MutableInputDesc(kRankDataInputIndex);
  43. GE_CHECK_NOTNULL(input_shape);
  44. if (input_shape->GetShape().GetDims() == UNKNOWN_RANK) {
  45. return NOT_CHANGED;
  46. }
  47. auto ndims = input_shape->GetShape().GetDimNum();
  48. GeTensorDesc tensor_desc(op_desc->GetOutputDesc(0));
  49. GeTensorPtr output_ptr;
  50. output_ptr = MakeShared<ge::GeTensor>(tensor_desc, reinterpret_cast<uint8_t *>(&ndims), GetSizeByDataType(DT_INT32));
  51. if (output_ptr == nullptr) {
  52. GELOGE(MEMALLOC_FAILED, "make_shared ge::GeTensor failed");
  53. return MEMALLOC_FAILED;
  54. }
  55. v_output.push_back(output_ptr);
  56. return SUCCESS;
  57. }
  58. REGISTER_KERNEL(RANK, RankKernel);
  59. } // 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