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.

ge_task_info.h 1.7 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667
  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. #ifndef INC_COMMON_OPSKERNEL_GE_TASK_INFO_H_
  17. #define INC_COMMON_OPSKERNEL_GE_TASK_INFO_H_
  18. #include <runtime/rt.h>
  19. #include <stdint.h>
  20. #include <string>
  21. #include <vector>
  22. using std::string;
  23. namespace ge {
  24. // when need to eliminate GETaskKernelHcclInfo, so not need DAVINCI_TRAIN/DAVINCI_CLOUD
  25. struct GETaskKernelHcclInfo {
  26. string input_name;
  27. string hccl_type;
  28. void *inputDataAddr;
  29. void *outputDataAddr;
  30. void *workSpaceAddr;
  31. int32_t count;
  32. int32_t dataType;
  33. int32_t opType;
  34. int64_t rootId;
  35. uint64_t workSpaceMemSize;
  36. std::vector<int64_t> dims;
  37. std::vector<rtStream_t> hcclStreamList;
  38. };
  39. struct GETaskInfo {
  40. uint32_t id;
  41. uint16_t type;
  42. uint32_t streamID;
  43. void *stream; // rtKernelLaunch input argument
  44. void *event;
  45. void *privateDef;
  46. uint32_t privateDefLen;
  47. void *opsKernelStorePtr;
  48. std::vector<GETaskKernelHcclInfo> kernelHcclInfo;
  49. };
  50. struct HcomOpertion {
  51. std::string hcclType;
  52. void *inputPtr;
  53. void *outputPtr;
  54. uint64_t count;
  55. int32_t dataType;
  56. int32_t opType;
  57. int32_t root;
  58. };
  59. } // namespace ge
  60. #endif // INC_COMMON_OPSKERNEL_GE_TASK_INFO_H_

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