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.

mvn_ops.h 1.7 kB

5 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051
  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_MVN_OPS_H
  17. #define GE_OP_MVN_OPS_H
  18. #include "graph/operator_reg.h"
  19. namespace ge {
  20. /**
  21. *@brief Normalizes the input.
  22. *@par Inputs:
  23. * One input:
  24. *x: An NCHW tensor of type float16 or float32.
  25. *@par Attributes:
  26. *@li normalize_variance: An optional bool specifying whether to normalize the variance, either "true" (default) or "false".
  27. *@li across_channels: An optional bool specifying whether to perform across-channel MVN, either "true" or "false" (default).
  28. *@li eps: An optional float32 epsilon for not dividing by zero. Defaults to "1e-9".
  29. *@par Outputs:
  30. *y: An NCHW tensor of type float16 or float32.
  31. *@attention Constraints:\n
  32. * The input tensor must have the NCHW format, whose shape length must be 4.
  33. */
  34. REG_OP(MVN)
  35. .INPUT(x, TensorType({DT_FLOAT, DT_FLOAT16})) /* "First operand." */
  36. .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT16})) /* "Result, has same element type as inputs" */
  37. .ATTR(normalize_variance, Bool, true)
  38. .ATTR(across_channels, Bool, false)
  39. .ATTR(eps, Float, 1e-9)
  40. .OP_END_FACTORY_REG(MVN)
  41. } // namespace ge
  42. #endif // GE_OP_MVN_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