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.

index_to_addr_ops.h 2.2 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263
  1. /**
  2. * Copyright 2019 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. /*!
  17. * \file index_to_addr_ops.h
  18. * \brief
  19. */
  20. #ifndef OPS_BUILT_IN_OP_PROTO_INC_INDEX_TO_ADDR_OPS_H_
  21. #define OPS_BUILT_IN_OP_PROTO_INC_INDEX_TO_ADDR_OPS_H_
  22. #include "graph/operator_reg.h"
  23. namespace ge {
  24. /**
  25. *@brief get block tensor according to base addr tensor, for hccl remote read to use.
  26. *@par Inputs:
  27. *@li base_addr: A Tensor of type int64/uint64. \n
  28. *@li row:A Tensor of type int64/uint64. \n
  29. *@li col: A Tensor of type int64/uint64.
  30. *@par Outputs:
  31. *addr_table: list of [rank id, host addr, device addr, read size]
  32. *@par Attributes:
  33. *@li ori_shape: An required list int. Shape of base tensor.
  34. *@li block_size: An required list int. Shape of split block tensor.
  35. *@li ori_storage_mode: An optional string from: '"Matrix", "UT"'. Defaults to
  36. "Matrix". Currently only support Matrix storage
  37. *@li block_storage_mode: An optional string from: '"Matrix", "UT"'. Defaults to
  38. "Matrix". Currently only support Matrix storage
  39. *@li rank_id: An optional int of rank id. Defaults is 0
  40. *@li dtype: An optional Type of base tensor. Defaults is DT_FLOAT
  41. */
  42. REG_OP(IndexToAddr)
  43. .INPUT(base_addr, TensorType({DT_INT64, DT_UINT64}))
  44. .INPUT(x, TensorType({DT_INT64, DT_UINT64}))
  45. .OUTPUT(addrs_table, TensorType({DT_INT64, DT_UINT64}))
  46. .REQUIRED_ATTR(ori_shape, ListInt)
  47. .REQUIRED_ATTR(block_size, ListInt)
  48. .ATTR(ori_storage_mode, String, "Matrix")
  49. .ATTR(block_storage_mode, String, "Matrix")
  50. .ATTR(rank_id, Int, 0)
  51. .ATTR(dtype, Type, DT_FLOAT)
  52. .OP_END_FACTORY_REG(IndexToAddr)
  53. } // namespace ge
  54. #endif // OPS_BUILT_IN_OP_PROTO_INC_INDEX_TO_ADDR_OPS_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