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.

ragged_array_ops.h 2.5 kB

4 years ago
1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465
  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 ragged_array_ops.h
  18. * \brief
  19. */
  20. #ifndef OPS_BUILT_IN_OP_PROTO_INC_RAGGED_ARRAY_OPS_H_
  21. #define OPS_BUILT_IN_OP_PROTO_INC_RAGGED_ARRAY_OPS_H_
  22. #include "graph/operator.h"
  23. #include "graph/operator_reg.h"
  24. namespace ge {
  25. /**
  26. *@brief Gather ragged slices from `params` axis `0` according to `indices` . \n
  27. *@par Inputs:
  28. *@li params_nested_splits: The `nested_row_splits` tensors that define the row-partitioning for the
  29. *params` RaggedTensor input. It's a dynamic input.
  30. *@li params_dense_values: The `flat_values` for the `params` RaggedTensor. There was a terminology change
  31. *at the python level from dense_values to flat_values, so dense_values is the
  32. *deprecated name.
  33. *@li indices: Indices in the outermost dimension of `params` of the values that should be
  34. *gathered.
  35. *@li OUTPUT_RAGGED_RANK: The ragged rank of the output RaggedTensor. `output_nested_splits` will contain
  36. *this number of `row_splits` tensors. This value should equal
  37. *`indices.shape.ndims + params.ragged_rank - 1` . \n
  38. *@par Outputs:
  39. *y:A Returns The `nested_row_splits` tensors that define the row-partitioning for the
  40. *returned RaggedTensor.The `flat_values` for the returned RaggedTensor . \n
  41. *@par Third-party framework compatibility
  42. * Compatible with tensorflow RaggedGather operator.
  43. */
  44. REG_OP(RaggedGather)
  45. .DYNAMIC_INPUT(params_nested_splits, TensorType({DT_INT32, DT_INT64}))
  46. .INPUT(params_dense_values, TensorType({DT_INT32, DT_INT64}))
  47. .INPUT(indices, TensorType({DT_INT32, DT_INT64}))
  48. .DYNAMIC_OUTPUT(output_nested_splits, TensorType({DT_INT32, DT_INT64}))
  49. .OUTPUT(output_dense_values, TensorType({DT_INT32, DT_INT64}))
  50. .REQUIRED_ATTR(Tsplits, Type)
  51. .ATTR(PARAMS_RAGGED_RANK, Int, 1)
  52. .ATTR(OUTPUT_RAGGED_RANK, Int, 0)
  53. .OP_END_FACTORY_REG(RaggedGather)
  54. } // namespace ge
  55. #endif // OPS_BUILT_IN_OP_PROTO_INC_RAGGED_ARRAY_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