DML_CONVOLUTION_OPERATOR_DESC structure (directml.h)

Article
12/02/2022

Performs a convolution of the FilterTensor with the InputTensor. This operator supports a number of standard convolution configurations. These standard configurations include forward and backward (transposed) convolution by setting the Direction and Mode fields, as well as depth-wise convolution by setting the GroupCount field.

A summary of the steps involved: perform the convolution into the output tensor; reshape the bias to the same dimension sizes as the output tensor; add the reshaped bias tensor to the output tensor.

Syntax

struct DML_CONVOLUTION_OPERATOR_DESC {
  const DML_TENSOR_DESC     *InputTensor;
  const DML_TENSOR_DESC     *FilterTensor;
  const DML_TENSOR_DESC     *BiasTensor;
  const DML_TENSOR_DESC     *OutputTensor;
  DML_CONVOLUTION_MODE      Mode;
  DML_CONVOLUTION_DIRECTION Direction;
  UINT                      DimensionCount;
  const UINT                *Strides;
  const UINT                *Dilations;
  const UINT                *StartPadding;
  const UINT                *EndPadding;
  const UINT                *OutputPadding;
  UINT                      GroupCount;
  const DML_OPERATOR_DESC   *FusedActivation;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

A tensor containing the input data. The expected dimensions of the InputTensor are:

{ BatchCount, InputChannelCount, InputWidth } for 3D,
{ BatchCount, InputChannelCount, InputHeight, InputWidth } for 4D, and
{ BatchCount, InputChannelCount, InputDepth, InputHeight, InputWidth } for 5D.

FilterTensor

Type: const DML_TENSOR_DESC*

A tensor containing the filter data. The expected dimensions of the FilterTensor are:

{ FilterBatchCount, FilterChannelCount, FilterWidth } for 3D,
{ FilterBatchCount, FilterChannelCount, FilterHeight, FilterWidth } for 4D, and
{ FilterBatchCount, FilterChannelCount, FilterDepth, FilterHeight, FilterWidth } for 5D.

BiasTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the bias data. The bias tensor is a tensor containing data which is broadcasted across the output tensor at the end of the convolution which is added to the result. The expected dimensions of the BiasTensor are:

{ 1, OutputChannelCount, 1 } for 3D,
{ 1, OutputChannelCount, 1, 1 } for 4D, and
{ 1, OutputChannelCount, 1, 1, 1 } for 5D.

For each output channel, the single bias value for that channel is added to every element in that channel of the OutputTensor. That is, the BiasTensor is broadcasted to the size of the OutputTensor, and what the operator returns is the summation of this broadcasted BiasTensor with the result from convolution.

OutputTensor

Type: const DML_TENSOR_DESC*

A tensor to write the results to. The expected dimensions of the OutputTensor are:

{ BatchCount, OutputChannelCount, OutputWidth } for 3D,
{ BatchCount, OutputChannelCount, OutputHeight, OutputWidth } for 4D, and
{ BatchCount, OutputChannelCount, OutputDepth, OutputHeight, OutputWidth } for 5D.

Mode

Type: DML_CONVOLUTION_MODE

The mode to use for the convolution operation. DML_CONVOLUTION_MODE_CROSS_CORRELATION is the behavior for required for typical inference scenarios. In contrast, DML_CONVOLUTION_MODE_CONVOLUTION flips the order of elements in each filter kernel along each spatial dimension.

Direction

Type: DML_CONVOLUTION_DIRECTION

The direction of the convolution operation. DML_CONVOLUTION_DIRECTION_FORWARD is the primary form of convolution used for inference where a combination of DML_CONVOLUTION_DIRECTION_FORWARD and DML_CONVOLUTION_DIRECTION_BACKWARD are used during training.

DimensionCount

Type: UINT

The number of spatial dimensions for the convolution operation. Spatial dimensions are the lower dimensions of the convolution FilterTensor. For example, the width and height dimension are spatial dimensions of a 4D convolution filter tensor. This value also determines the size of the Strides, Dilations, StartPadding, EndPadding, and OutputPadding arrays. It should be set to 2 when InputTensor.DimensionCount is 4, and 3 when InputTensor.DimensionCount is 5.

Strides