Edit

DML_DIAGONAL_MATRIX_OPERATOR_DESC structure (directml.h)

  • Article

Generates an identity-like matrix with ones (or other explicit value) on the major diagonal, and zeros everywhere else. The diagonal ones may be shifted (via Offset) where OutputTensor[i, i + Offset] = Value, meaning that an argument of Offset greater than zero shifts all values to the right, and less than zero shifts them to the left. This generator operator is useful for models to avoid storing a large constant tensor. Any leading dimensions before the last two are treated as a batch count, meaning that the tensor is treated as stack of 2D matrices.

This operator performs the following pseudocode.

for each coordinate in OutputTensor
    OutputTensor[coordinate] = if (coordinate.y + Offset == coordinate.x) then Value else 0
endfor

Syntax

struct DML_DIAGONAL_MATRIX_OPERATOR_DESC {
  const DML_TENSOR_DESC *OutputTensor;
  INT                   Offset;
  FLOAT                 Value;
};

Members

OutputTensor

Type: const DML_TENSOR_DESC*

The tensor to write the results to. The dimensions are { Batch1, Batch2, OutputHeight, OutputWidth }. The height and width don't need to be square.

Offset

Type: INT

An offset to shift the diagonal lines of Value, with positive offsets shifting the written value to the right/up (viewing the output as a matrix with the top left as 0,0), and negative offsets to the left/down.

Value

Type: FLOAT

A value to fill along the 2D diagonal. The standard value is 1.0. Note that if the DataType of the tensors is not DML_TENSOR_DATA_TYPE_FLOAT16 or DML_TENSOR_DATA_TYPE_FLOAT32, then the value might be truncated (for example, 10.6 will become 10).

Examples

Default identity matrix:

Offset: 0
Value: 1.0
OutputTensor: (Sizes:{1,1,3,3}, DataType:FLOAT32)
    [[[[1, 0, 0],
       [0, 1, 0],
       [0, 0, 1]]]]

Shift ones right/up:

Offset: 1
Value: 1.0
OutputTensor: (Sizes:{1,1,3,3}, DataType:FLOAT32)
    [[[[ 0, 1, 0],
       [ 0, 0, 1],
       [ 0, 0, 0]]]]

Shift ones left/down:

Offset: -1
Value: 1.0
OutputTensor: (Sizes:{1,1,3,2}, DataType:FLOAT32)
    [[[[0, 0],
       [1, 0],
       [0, 1]]]]

Shift the diagonal line of ones so far that all become zeroes:

Offset: -3
Value: 1.0
OutputTensor: (Sizes:{1,1,3,2}, DataType:FLOAT32)
    [[[[0, 0],
       [0, 0],
       [0, 0]]]]

Remarks

Availability

This operator was introduced in DML_FEATURE_LEVEL_2_0.

Tensor support

DML_FEATURE_LEVEL_5_1 and above

Tensor Kind Supported dimension counts Supported data types
OutputTensor Output 2 to 4 FLOAT64, FLOAT32, FLOAT16, INT64, INT32, INT16, INT8, UINT64, UINT32, UINT16, UINT8

DML_FEATURE_LEVEL_4_0 and above

Tensor Kind Supported dimension counts Supported data types
OutputTensor Output 2 to 4 FLOAT32, FLOAT16, INT32, INT16, INT8, UINT32, UINT16, UINT8

DML_FEATURE_LEVEL_2_1 and above

Tensor Kind Supported dimension counts Supported data types
OutputTensor Output 4 FLOAT32, FLOAT16, INT32, INT16, INT8, UINT32, UINT16, UINT8

DML_FEATURE_LEVEL_2_0 and above

Tensor Kind Supported dimension counts Supported data types
OutputTensor Output 4 FLOAT32, FLOAT16

Requirements

Requirement Value
Minimum supported client Windows 10, version 2004 (10.0; Build 19041)
Minimum supported server Windows Server, version 2004 (10.0; Build 19041)
Header directml.h