MatrixDiagV3

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: generates matrices from the input values of one or more diagonals, and positions outside the diagonal band are filled with $\mathrm{p}\mathrm{a}\mathrm{d}\mathrm{d}\mathrm{i}\mathrm{n}\mathrm{g}\_\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}$.
• Assume that the sizes of the last two dimensions of the output tensor are $\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{r}\mathrm{o}\mathrm{w}\mathrm{s}$ and $\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{c}\mathrm{o}\mathrm{l}\mathrm{s}$ respectively, and $k=[k_l,k_u]$ indicates the range of diagonals to write. In the single-diagonal scenario, $k_l=k_u$. Let $d=j-i$ indicate the diagonal number of position $(i,j)$. Then the maximum diagonal length is:

$$\mathrm{m}\mathrm{a}\mathrm{x}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\_\mathrm{l}\mathrm{e}\mathrm{n}=\min (\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{r}\mathrm{o}\mathrm{w}\mathrm{s}+\min (k_u,0),\mathrm{n}\mathrm{u}\mathrm{m}\_\mathrm{c}\mathrm{o}\mathrm{l}\mathrm{s}-\max (k_l,0))$$

Output elements satisfy:

$$y_{...,i,j}=\{\begin{array}{ll}{x}_{...,k_u-d,p(i,j)}& k_l\le d\le k_u\\ \text{padding\_value}& \text{otherwise}\end{array}$$

Here, $p(i,j)$ indicates the position of the diagonal element in the last dimension of input $x$, and the left/right alignment mode is controlled by the align attribute. When $k$ is a single integer or $k_l=k_u$, $x$ indicates a single diagonal; when $k_l<k_u$, $x$ indicates a diagonal band.

Parameter Description

Parameter Name	Input/Output/Attribute	Description	Data Type	Data Format
x	Input	`x` in the formula. When `k` indicates a single diagonal, the last dimension of `x` stores the data of this diagonal; when `k` indicates a diagonal band, the second-to-last dimension of `x` stores the number of diagonals, and the last dimension stores the data of each diagonal padded according to `align`.	DOUBLE, FLOAT, FLOAT16, INT8, INT16, INT32, INT64, UINT8, UINT16, UINT32, UINT64, COMPLEX64, COMPLEX128, BOOL	ND
k	Input	kl and ku in the formula. It can be a scalar, indicating a single diagonal; it can also be a vector of length 2, indicating the lower bound and upper bound of a diagonal band.	INT32	ND
num_rows	Input	Number of rows of the output matrix, which is `num_rows` in the formula. A value of `-1` indicates that it is automatically inferred from `k` and `x`.	INT32	ND
num_cols	Input	Number of columns of the output matrix, which is `num_cols` in the formula. A value of `-1` indicates that it is automatically inferred from `k` and `x`.	INT32	ND
padding_value	Input	`padding_value` in the formula. It is used to fill positions outside the specified diagonal band, and its data type is the same as that of `x`.	DOUBLE, FLOAT, FLOAT16, INT8, INT16, INT32, INT64, UINT8, UINT16, UINT32, UINT64, COMPLEX64, COMPLEX128, BOOL	ND
align	Optional attribute	Specifies the alignment mode of superdiagonals and subdiagonals. `RIGHT_LEFT`, `LEFT_RIGHT`, `LEFT_LEFT`, and `RIGHT_RIGHT` are supported, and the default value is `RIGHT_LEFT`.	STRING	-
y	Output	`y` in the formula, indicating the generated matrix tensor. Its data type is the same as that of `x`. When `k` is a single element or `k[0] == k[1]`, the rank of `y` is the rank of `x` plus 1; otherwise, the rank of `y` is the same as that of `x`.	DOUBLE, FLOAT, FLOAT16, INT8, INT16, INT32, INT64, UINT8, UINT16, UINT32, UINT64, COMPLEX64, COMPLEX128, BOOL	ND

Constraint Description

• The rank of x is at least 1.
• The number of elements in k can only be 1 or 2; when k has 2 elements, k[0] <= k[1] must be satisfied.
• When k indicates a diagonal band, the length of the second-to-last dimension of x must be equal to $k_u-k_l+1$, and the length of the last dimension must be equal to $\mathrm{m}\mathrm{a}\mathrm{x}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\_\mathrm{l}\mathrm{e}\mathrm{n}$.

Invocation Description

Invocation Method	Invocation Sample	Description
Graph-mode invocation	test_geir_matrix_diag_v3	Invokes the MatrixDiagV3 operator by building a graph through the operator IR.