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.

npu_loss_scale_ops.h 2.1 kB

5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475
  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 GE_OP_NN_LOSS_SCALE_OPS_H
  17. #define GE_OP_NN_LOSS_SCALE_OPS_H
  18. #include "graph/operator_reg.h"
  19. namespace ge {
  20. REG_OP(NPUAllocFloatStatusOperator)
  21. .OUTPUT(data, TensorType({DT_FLOAT}))
  22. .OP_END_FACTORY_REG(NPUAllocFloatStatusOperator)
  23. REG_OP(NPUClearFloatStatusOperator)
  24. .INPUT(addr, TensorType{DT_FLOAT})
  25. .OUTPUT(data, TensorType({DT_FLOAT}))
  26. .OP_END_FACTORY_REG(NPUClearFloatStatusOperator)
  27. REG_OP(NPUGetFloatStatusOperator)
  28. .INPUT(addr, TensorType{DT_FLOAT})
  29. .OUTPUT(data, TensorType({DT_FLOAT}))
  30. .OP_END_FACTORY_REG(NPUGetFloatStatusOperator)
  31. /**
  32. *@brief Produces a variable with 0 in memory.
  33. *@par Outputs:
  34. *y: A Tensor of type int32, output eight numbers with a value of zero.
  35. */
  36. REG_OP(NPUAllocFloatStatus)
  37. .OUTPUT(data, TensorType({DT_FLOAT}))
  38. .OP_END_FACTORY_REG(NPUAllocFloatStatus)
  39. /**
  40. *@brief Set the value of address 0x40000 to 0 in each core.
  41. *@par Inputs:
  42. *@li addr: A tensor of type float32.
  43. *@par Outputs:
  44. *data: A Tensor of type float32.
  45. */
  46. REG_OP(NPUClearFloatStatus)
  47. .INPUT(addr, TensorType{DT_FLOAT})
  48. .OUTPUT(data, TensorType({DT_FLOAT}))
  49. .OP_END_FACTORY_REG(NPUClearFloatStatus)
  50. /**
  51. *@brief Get the value of address 0x40000.
  52. *@par Inputs:
  53. *@li addr: A tensor of type float32.
  54. *@par Outputs:
  55. *data: A Tensor of type float32.
  56. */
  57. REG_OP(NPUGetFloatStatus)
  58. .INPUT(addr, TensorType{DT_FLOAT})
  59. .OUTPUT(data, TensorType({DT_FLOAT}))
  60. .OP_END_FACTORY_REG(NPUGetFloatStatus)
  61. } // namespace ge
  62. #endif // GE_OP_NN_LOSS_SCALE_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