DML_QUANTIZED_LINEAR_AVERAGE_POOLING_OPERATOR_DESC structure (directml.h)

Article
01/09/2024

Averages quantized values across the elements within the sliding window over the input tensor. This operator is mathematically equivalent to dequantizing the inputs, then performing average pooling, and then quantizing the output.

Dequantize function

f(Input, Scale, ZeroPoint) = (Input - ZeroPoint) * Scale

Quantize function

f(Input, Scale, ZeroPoint) = clamp(round(Input / Scale) + ZeroPoint, Min, Max)

Important

This API is available as part of the DirectML standalone redistributable package (see Microsoft.AI.DirectML version 1.13 and later. Also see DirectML version history.

Syntax

struct DML_QUANTIZED_LINEAR_AVERAGE_POOLING_OPERATOR_DESC
{
    const DML_TENSOR_DESC* InputTensor;
    const DML_TENSOR_DESC* InputScaleTensor;
    _Maybenull_ const DML_TENSOR_DESC* InputZeroPointTensor;
    const DML_TENSOR_DESC* OutputScaleTensor;
    _Maybenull_ const DML_TENSOR_DESC* OutputZeroPointTensor;
    const DML_TENSOR_DESC* OutputTensor;
    UINT DimensionCount;
    _Field_size_(DimensionCount) const UINT* Strides;
    _Field_size_(DimensionCount) const UINT* WindowSize;
    _Field_size_(DimensionCount) const UINT* StartPadding;
    _Field_size_(DimensionCount) const UINT* EndPadding;
    _Field_size_(DimensionCount) const UINT* Dilations;
    BOOL IncludePadding;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

An input tensor of Sizes { BatchCount, ChannelCount, Height, Width } for 4D, and { BatchCount, ChannelCount, Depth, Height, Weight } for 5D.

InputScaleTensor

Type: const DML_TENSOR_DESC*

A tensor containing the InputTensor scale data. The expected dimensions of InputScaleTensor are { 1, 1, 1, 1 } if per-tensor quantization is required, or { 1, ChannelCount, 1, 1 } if per-channel quantization is required. These scale values are used for dequantizing the InputTensor values.

InputZeroPointTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the InputTensor zero point data. The expected dimensions of InputZeroPointTensor are { 1, 1, 1, 1 } if per-tensor quantization is required, or { 1, ChannelCount, 1, 1 } if per-channel quantization is required. These zero point values are used for dequantizing the InputTensor values.

OutputScaleTensor

Type: const DML_TENSOR_DESC*

A tensor containing the OutputTensor scale data. The expected dimensions of OutputScaleTensor are { 1, 1, 1, 1 } if per-tensor quantization is required, or { 1, ChannelCount, 1, 1 } if per-channel quantization is required. These scale values are used for quantizing the OutputTensor values.

OutputZeroPointTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the OutputTensor zero point data. The expected dimensions of OutputZeroPointTensor are { 1, 1, 1, 1 } if per-tensor quantization is required, or { 1, ChannelCount, 1, 1 } if per-channel quantization is required. This zero point value is used for quantizing the OutputTensor values.

OutputTensor

Type: const DML_TENSOR_DESC*

A description of the output tensor. The sizes of the output tensor can be computed as follows.

OutputTensor->Sizes[0] = InputTensor->Sizes[0];
OutputTensor->Sizes[1] = InputTensor->Sizes[1];

for (UINT i = 0; i < DimensionCount; ++i) {
  UINT PaddedSize = InputTensor->Sizes[i + 2] + StartPadding[i] + EndPadding[i];
  OutputTensor->Sizes[i + 2] = (PaddedSize - WindowSizes[i]) / Strides[i] + 1;
}

DimensionCount

Type: UINT

The number of spatial dimensions of the input tensor InputTensor, which also corresponds to the number of dimensions of the sliding window WindowSize. This value also determines the size of the Strides, StartPadding, and EndPadding arrays. It should be set to 2 when InputTensor is 4D, and 3 when it's a 5D tensor.

Strides