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.

op_imp.cc 2.6 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879
  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 <stdint.h>
  17. #include <functional>
  18. #include <vector>
  19. #include "debug/ge_log.h"
  20. #include "debug/ge_util.h"
  21. using namespace std;
  22. namespace ge {
  23. GE_FUNC_DEV_VISIBILITY GE_FUNC_HOST_VISIBILITY graphStatus
  24. BroadCastInfer(const function<vector<int64_t>()>& get_in1_shape, const function<vector<int64_t>()>& get_in2_shape,
  25. const function<void(const vector<int64_t>& outShape)>& set_out_shape) {
  26. auto x1_shape = get_in1_shape();
  27. auto x2_shape = get_in2_shape();
  28. vector<int64_t> y_shape;
  29. if (x1_shape.empty()) {
  30. y_shape = x2_shape;
  31. set_out_shape(y_shape);
  32. return GRAPH_SUCCESS;
  33. }
  34. if (x2_shape.empty()) {
  35. y_shape = x1_shape;
  36. set_out_shape(y_shape);
  37. return GRAPH_SUCCESS;
  38. }
  39. int len_diff = static_cast<int>(x1_shape.size() - x2_shape.size());
  40. if (len_diff >= 0) {
  41. for (int i = 0; i < len_diff; i++) {
  42. y_shape.push_back(x1_shape[i]);
  43. }
  44. int x2_shape_size = static_cast<int>(x2_shape.size());
  45. for (int i = 0; i < x2_shape_size; i++) {
  46. bool shapeFlag =
  47. ((x1_shape[i + len_diff] != x2_shape[i]) && (std::min(x1_shape[i + len_diff], x2_shape[i]) != 1));
  48. if (shapeFlag) {
  49. GE_LOGE("operands could not be broadcast together");
  50. return GRAPH_FAILED;
  51. }
  52. y_shape.push_back(std::max(x1_shape[i + len_diff], x2_shape[i]));
  53. }
  54. } else {
  55. for (int i = 0; i < -len_diff; i++) {
  56. y_shape.push_back(x2_shape[i]);
  57. }
  58. int x1_shape_size = static_cast<int>(x1_shape.size());
  59. for (int i = 0; i < x1_shape_size; i++) {
  60. bool shapeFlag =
  61. ((x1_shape[i] != x2_shape[i - len_diff]) && (std::min(x1_shape[i], x2_shape[i - len_diff]) != 1));
  62. if (shapeFlag) {
  63. GE_LOGE("operands could not be broadcast together");
  64. return GRAPH_FAILED;
  65. }
  66. y_shape.push_back(std::max(x1_shape[i], x2_shape[i - len_diff]));
  67. }
  68. }
  69. set_out_shape(y_shape);
  70. return GRAPH_SUCCESS;
  71. }
  72. } // 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