FilesLast commitLast update
9 months ago
2 months ago
16 days ago
16 days ago
18 days ago
1 month ago
1 month ago
13 days ago
README

TopKV2

English | 简体中文

Product Support Status

Product Supported
Ascend 950PR&950DT products √
Atlas A3 products √
Atlas A2 products √
Atlas 200I/500 A2 inference products √
Atlas inference products √
Atlas training products √

Function Description

  • API function: returns the k extreme values and their indices of the input tensor on the specified dimension. The TopKV2 operator supports specifying the value of k through an input tensor and supports a configurable output data type of indices.

  • Computation description:

    • When largest=True, the k largest values on the specified dimension and their indices are returned.
    • When largest=False, the k smallest values on the specified dimension and their indices are returned.
    • When sorted=True, the output results are sorted in descending order (largest=True) or ascending order (largest=False).
    • When sorted=False, the output results are not sorted.

Parameter Description

Parameter Name Input/Output/Attribute Description Data Type Data Format
x Input Input tensor for the TopK computation. Dimensions 1-8 are supported. INT8, UINT8, INT16, UINT16, INT32, UINT32, INT64, UINT64, FLOAT16, FLOAT32, BFLOAT16 ND
k Input 0D scalar tensor that indicates the number of extreme values to output on the computation dimension. The value range is [0, x.size(dim)]. INT32, INT64 ND
values Output Output values of the TopK computation. The data type is the same as that of x. The shape is the same as k on the sorting axis and the same as x on non-sorting axes. INT8, UINT8, INT16, UINT16, INT32, UINT32, INT64, UINT64, FLOAT16, FLOAT32, BFLOAT16 ND
indices Output Output indices of the TopK computation. The data type is specified by the indices_dtype attribute. The shape is the same as k on the sorting axis and the same as x on non-sorting axes. INT32, INT64 ND
sorted Attribute Optional Boolean. The default value is True. True indicates that the output results are sorted, and False indicates that the output results are not sorted. Bool -
dim Attribute Optional integer. The default value is -1. It indicates the computation dimension. The value range is [-x.dim(), x.dim()). Int -
largest Attribute Optional Boolean. The default value is True. True indicates that the k largest elements are returned, and False indicates that the k smallest elements are returned. Bool -
indices_dtype Attribute Optional integer. The default value is DT_INT32(3). It indicates the data type of the output indices. DT_INT32(3) or DT_INT64(9) is supported. Int -
sort_policy Attribute Optional integer. The default value is 0. It indicates the sorting policy for TopK candidate results. 0 indicates that the original sorting behavior is used, and 1 indicates that the Bitonic Small TopK path is enabled in the scenario where sorted=True and k is small (2≤k≤32). Int -
  • Atlas 200I/500 A2 inference products, Atlas inference products, and Atlas training products: The BFLOAT16 data type is not supported.

Constraint Description

None

Invocation Description

Invocation Method Invocation Sample Description
Graph-mode invocation test_geir_top_k_v2 Invokes the top_k_v2 operator by building a graph through the operator IR.