DML_ELEMENT_WISE_DEQUANTIZE_LINEAR_OPERATOR_DESC structure (directml.h)

Performs the following linear dequantization function on every element in InputTensor with respect to its corresponding element in ScaleTensor and ZeroPointTensor, placing the results in the corresponding element of OutputTensor.

f(input, scale, zero_point) = (input - zero_point) * scale

Quantization is a common way to increase performance at the cost of precision. A group of 8-bit int values can be computed faster than a group of 32-bit float values can. Dequantizing converts the encoded data back to its domain.

Syntax

struct DML_ELEMENT_WISE_DEQUANTIZE_LINEAR_OPERATOR_DESC {
  const DML_TENSOR_DESC *InputTensor;
  const DML_TENSOR_DESC *ScaleTensor;
  const DML_TENSOR_DESC *ZeroPointTensor;
  const DML_TENSOR_DESC *OutputTensor;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

The tensor containing the inputs.

ScaleTensor

Type: const DML_TENSOR_DESC*

The tensor containing the scales. A scale value of 0 will result in undefined behavior.

ZeroPointTensor

Type: const DML_TENSOR_DESC*

The tensor containing the zero point that was used for quantization.

OutputTensor

Type: const DML_TENSOR_DESC*

The output tensor to write the results to.

Availability

This operator was introduced in DML_FEATURE_LEVEL_1_0.

Tensor constraints

• InputTensor, OutputTensor, ScaleTensor, and ZeroPointTensor must have the same DimensionCount and Sizes.
• InputTensor and ZeroPointTensor must have the same DataType.
• OutputTensor and ScaleTensor must have the same DataType.

Tensor support

DML_FEATURE_LEVEL_6_2 and above

DML_FEATURE_LEVEL_6_0 and above

DML_FEATURE_LEVEL_3_0 and above

DML_FEATURE_LEVEL_2_1 and above

DML_FEATURE_LEVEL_1_0 and above

Requirements

See also

DML_ELEMENT_WISE_QUANTIZE_LINEAR_OPERATOR_DESC