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 2.4 kB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061
  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_RAGGED_ARRAY_OPS_H
  17. #define GE_OP_RAGGED_ARRAY_OPS_H
  18. #include "graph/operator.h"
  19. #include "graph/operator_reg.h"
  20. namespace ge {
  21. /**
  22. *@brief Gather ragged slices from `params` axis `0` according to `indices`.
  23. *@par Inputs:
  24. *@li params_nested_splits: The `nested_row_splits` tensors that define the row-partitioning for the \n
  25. *params` RaggedTensor input.
  26. *@li params_dense_values: The `flat_values` for the `params` RaggedTensor. There was a terminology change \n
  27. *at the python level from dense_values to flat_values, so dense_values is the \n
  28. *deprecated name.
  29. *@li indices: Indices in the outermost dimension of `params` of the values that should be \n
  30. *gathered.
  31. *@li OUTPUT_RAGGED_RANK: The ragged rank of the output RaggedTensor. `output_nested_splits` will contain \n
  32. *this number of `row_splits` tensors. This value should equal \n
  33. *`indices.shape.ndims + params.ragged_rank - 1`.
  34. *@par Outputs:
  35. *y:A Returns The `nested_row_splits` tensors that define the row-partitioning for the \n
  36. *returned RaggedTensor.The `flat_values` for the returned RaggedTensor.
  37. *@par Third-party framework compatibility
  38. * Compatible with tensorflow RaggedGather operator.
  39. */
  40. REG_OP(RaggedGather)
  41. .DYNAMIC_INPUT(params_nested_splits, TensorType({DT_INT32, DT_INT64}))
  42. .INPUT(params_dense_values, TensorType({DT_INT32, DT_INT64}))
  43. .INPUT(indices, TensorType({DT_INT32, DT_INT64}))
  44. .DYNAMIC_OUTPUT(output_nested_splits, TensorType({DT_INT32, DT_INT64}))
  45. .OUTPUT(output_dense_values, TensorType({DT_INT32, DT_INT64}))
  46. .REQUIRED_ATTR(Tsplits, Type)
  47. .ATTR(PARAMS_RAGGED_RANK, Int, 1)
  48. .ATTR(OUTPUT_RAGGED_RANK, Int, 0)
  49. .OP_END_FACTORY_REG(RaggedGather)
  50. } // namespace ge
  51. #endif //GE_OP_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