FilesLast commitLast update
27 days ago
2 months ago
16 days ago
16 days ago
16 days ago
21 days ago
13 days ago
14 days ago
README

SquareSumAll

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

  • Operator function: computes the sum of squares of all elements of the two input tensors x1 and x2 that have the same shape, and outputs two scalars y1 and y2.

  • Computation formula:

y1=∑i=0N−1x1,i2,y2=∑i=0N−1x2,i2y_1=\sum_{i=0}^{N-1}x_{1,i}^{2},\qquad y_2=\sum_{i=0}^{N-1}x_{2,i}^{2}

Here, NN is the total number of elements in the input tensor.

Parameter Description

Parameter Name Input/Output/Attribute Description Data Type Data Format
x1 Input Denotes the first tensor for which the sum of squares is to be computed, corresponding to x1 in the formula. FLOAT FRACTAL_Z, C1HWNCoC0, NC1HWC0, ND, NCHW, NHWC
x2 Input Denotes the second tensor for which the sum of squares is to be computed, corresponding to x2 in the formula. FLOAT FRACTAL_Z, C1HWNCoC0, NC1HWC0, ND, NCHW, NHWC
y1 Output
  • Denotes the sum of squares of all elements of x1, corresponding to y1 in the formula.
  • The output is a 0-dimension Tensor (rank 0) containing one FLOAT element.
FLOAT FRACTAL_Z, C1HWNCoC0, NC1HWC0, ND
y2 Output
  • Denotes the sum of squares of all elements of x2, corresponding to y2 in the formula.
  • The data type and shape are the same as those of the output y1.
FLOAT FRACTAL_Z, C1HWNCoC0, NC1HWC0, ND

Product Difference Description

The "input format → output format" combinations in the following table apply to both x1→y1 and x2→y2, and the two input paths must use the same data format.

Product Static shape format combinations Dynamic shape format combinations Rank restrictions
Ascend 950PR&950DT products ND→ND, NCHW→ND, NHWC→ND ND→ND, NCHW→ND, NHWC→ND The rank of ND inputs ranges from [0, 8]; the rank of NCHW and NHWC inputs must be 4.
Atlas A3 products
Atlas A2 products
Atlas 200I/500 A2 inference products
Atlas inference products
Atlas training products
FRACTAL_Z→FRACTAL_Z, C1HWNCoC0→C1HWNCoC0, NC1HWC0→NC1HWC0, ND→ND ND→ND The rank of inputs ranges from [0, 8].

Constraint Description

  • During graph compilation, -1 denotes an unknown dimension and [-2] denotes an unknown rank; the two outputs are always rank-0 scalars. During execution, the input shape must be parsed first, and then the following non-empty, same-shape, and format constraints must be met.

  • Empty tensors are not supported for x1 or x2. A rank of 0 indicates a scalar input with one element. If the rank is greater than 0, the length of each dimension must be greater than 0.

  • The execution-time rank and the length of each dimension of x1 and x2 must be exactly the same, and broadcasting is not supported. The shapes [2, 3] and [6] are invalid even if the number of elements is the same.

  • All four parameters of the operator node must use the FLOAT type, and the data formats must meet the valid combinations in the product difference description.

  • NaN, Inf, and overflow of the square of finite values propagate naturally according to FLOAT semantics. The two accumulators and the two outputs do not pollute each other; a non-finite value in x1 does not affect y2, and vice versa.

  • This operator provides only GE graph-mode registration, with the build configuration set to aclnn_exclude, and does not provide aclnn APIs.

Invocation Description

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