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.

swap_co_ops.h 1.8 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556
  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_SWAP_CO_OPS_H_
  17. #define GE_OP_SWAP_CO_OPS_H_
  18. #include "graph/operator_reg.h"
  19. namespace ge {
  20. /**
  21. *@brief Folds the convolution input weight constant of the preceding layer \n
  22. * of PSROIPooling to convert the N dimension of the weight from \n
  23. * (output_dim, group_size*group_size) to \n
  24. * (group_size*group_size, int((output_dim+15)/C0)*C0).
  25. *@see PSROIPooling
  26. *@par Inputs:
  27. * One input:
  28. *x: An NCHW tensor of type float16 or float32, describing the weight of\n
  29. * convolution. Dim N must equal output_dim*group_size*group_size.
  30. *@par Attributes:
  31. *@li output_dim: A required int32, specifying the number of output channels.\n
  32. * Must be greater than "0".
  33. *@li group_size: A required int32, specifying the number of groups to encode\n
  34. * position-sensitive score maps. Must be within the range (0, 128).
  35. *@par Outputs:
  36. *y: An NCHW tensor of type float16 or float32, describing the result weight\n
  37. * of convolution.
  38. */
  39. REG_OP(SwapCo)
  40. .INPUT(x, TensorType({DT_FLOAT, DT_FLOAT16}))
  41. .ATTR(output_dim, Int, 0)
  42. .ATTR(group_size, Int, 0)
  43. .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT16}))
  44. .OP_END_FACTORY_REG(SwapCo)
  45. } // namespace ge
  46. #endif // GE_OP_SWAP_CO_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