Edit

Share via

DML_UNFOLD_OPERATOR_DESC structure (directml.h)

  • Article

Extracts sliding local blocks from a batched input tensor.

Consider a InputTensor with shape (N,C,*), where * is an arbitrary number of spatial dimensions (<=6). The operation flattens each sliding WindowSize'd block within the InputTensor into a column of an OutputTensor. The OutputTensor is a 3D tensor containing sliding local blocks, e.g., patches of images, of shape (N,C×∏(WindowSizes),BlockCount), where N is batch dimension, C×∏(WindowSizes) is the number of values within a window (a window has ∏(WindowSizes) spatial locations each containing a C-channeled vector), and BlockCount is the total number of blocks. The arguments must satisfy:

BlocksPerDimension[d] = ( ( SpatialSize[d] + StartPadding[d] + EndPadding[d] - Dilations[d] * (WindowSizes[d] - 1) - 1 ) / stride[d] ) + 1

BlockCount = ∏d BlocksPerDimension[d]

Where:

  • SpatialSize is formed by the spatial dimensions of input (*)
  • 0 <= d < DimensionCount

Indexing output at the last dimension (column dimension) gives all values within a certain block.

The StartPadding, EndPadding, Strides, and Dilations arguments specify how the sliding blocks are retrieved.

Important

This API is available as part of the DirectML standalone redistributable package (see Microsoft.AI.DirectML version 1.15.1 and later. Also see DirectML version history.

Syntax

struct DML_UNFOLD_OPERATOR_DESC
{
    const DML_TENSOR_DESC* InputTensor;
    const DML_TENSOR_DESC* OutputTensor;
    UINT DimensionCount;
    _Field_size_(DimensionCount) const UINT* WindowSizes;
    _Field_size_(DimensionCount) const UINT* Strides;
    _Field_size_(DimensionCount) const UINT* Dilations;
    _Field_size_(DimensionCount) const UINT* StartPadding;
    _Field_size_(DimensionCount) const UINT* EndPadding;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

The input tensor to read from.

OutputTensor

Type: const DML_TENSOR_DESC*

The output tensor to write the results to.

DimensionCount

Type: UINT

The spatial dimensions of InputTensor. DimensionCount must be in the range [1, 6].

WindowSizes

Type: _Field_size_(DimensionCount) const UINT*

The size of the sliding window.

Strides

Type: _Field_size_(DimensionCount) const UINT*

The stride of the sliding window (with dimensions WindowSizes) in the input spatial dimensions. They are separate from the tensor strides included in DML_TENSOR_DESC. Step size of the extracted patches.

Dilations

Type: _Field_size_(DimensionCount) const UINT*

The dilations of the sliding window (with dimensions WindowSizes) in the input spatial dimensions, by scaling the space between the kernel points. Dilations of the extracted patch.

StartPadding

Type: _Field_size_(DimensionCount) const UINT*

An array containing the amount of implicit zero-padding to be applied to the beginning of each spatial dimension of InputTensor. Start padding of the source tensor.

EndPadding

Type: _Field_size_(DimensionCount) const UINT*

An array containing the amount of implicit zero-padding to be applied to the end of each spatial dimension of InputTensor. End padding of the source tensor.

Examples

Example 1.

1-channel unfold.

InputTensor: (Sizes:{1, 1, 5, 5}, DataType:FLOAT32)
[[[[ 0.,  1.,  2.,  3.,  4.],
   [ 5.,  6.,  7.,  8.,  9.],
   [10., 11., 12., 13., 14.],
   [15., 16., 17., 18., 19.],
   [20., 21., 22., 23., 24.]]]]
DimensionCount: 2
WindowSizes: {3, 3}
Strides: {1, 1}
Dilations: {1, 1}
StartPadding: {0, 0}
EndPadding: {0, 0}
OutputTensor: (Sizes:{1, 9, 9}, DataType:FLOAT32)
[[[ 0.,  1.,  2.,  5.,  6.,  7., 10., 11., 12.],
  [ 1.,  2.,  3.,  6.,  7.,  8., 11., 12., 13.],
  [ 2.,  3.,  4.,  7.,  8.,  9., 12., 13., 14.],
  [ 5.,  6.,  7., 10., 11., 12., 15., 16., 17.],
  [ 6.,  7.,  8., 11., 12., 13., 16., 17., 18.],
  [ 7.,  8.,  9., 12., 13., 14., 17., 18., 19.],
  [10., 11., 12., 15., 16., 17., 20., 21., 22.],
  [11., 12., 13., 16., 17., 18., 21., 22., 23.],
  [12., 13., 14., 17., 18., 19., 22., 23., 24.]]]

Example 2.

1-channel, padded fold.

InputTensor: (Sizes:{1, 1, 5, 5}, DataType:FLOAT32)
[[[[ 0.,  1.,  2.,  3.,  4.],
   [ 5.,  6.,  7.,  8.,  9.],
   [10., 11., 12., 13., 14.],
   [15., 16., 17., 18., 19.],
   [20., 21., 22., 23., 24.]]]]
DimensionCount: 2
WindowSizes: {3, 3}
Strides: {1, 1}
Dilations: {1, 1}
StartPadding: {1, 0}
EndPadding: {1, 0}
OutputTensor: (Sizes:{1, 9, 15}, DataType:FLOAT32)
[[[ 0.,  0.,  0.,  0.,  1.,  2.,  5.,  6.,  7., 10., 11., 12., 15., 16., 17.],
  [ 0.,  0.,  0.,  1.,  2.,  3.,  6.,  7.,  8., 11., 12., 13., 16., 17., 18.],
  [ 0.,  0.,  0.,  2.,  3.,  4.,  7.,  8.,  9., 12., 13., 14., 17., 18., 19.],
  [ 0.,  1.,  2.,  5.,  6.,  7., 10., 11., 12., 15., 16., 17., 20., 21., 22.],
  [ 1.,  2.,  3.,  6.,  7.,  8., 11., 12., 13., 16., 17., 18., 21., 22., 23.],
  [ 2.,  3.,  4.,  7.,  8.,  9., 12., 13., 14., 17., 18., 19., 22., 23., 24.],
  [ 5.,  6.,  7., 10., 11., 12., 15., 16., 17., 20., 21., 22.,  0.,  0., 0.],
  [ 6.,  7.,  8., 11., 12., 13., 16., 17., 18., 21., 22., 23.,  0.,  0., 0.],
  [ 7.,  8.,  9., 12., 13., 14., 17., 18., 19., 22., 23., 24.,  0.,  0., 0.]]]

Availability

This operator was introduced in DML_FEATURE_LEVEL_6_4.

Tensor constraints

InputTensor and OutputTensor must have the same DimensionCount.

Tensor support

Tensor Kind Supported dimension counts Supported data types
InputTensor Input 3 to 8
OutputTensor Output 3 to 8

Requirements

   
Header directml.h