DML_CUMULATIVE_SUMMATION_OPERATOR_DESC structure (directml.h)

Article
01/21/2022

Sums the elements of a tensor along an axis, writing the running tally of the summation into the output tensor.

Syntax

struct DML_CUMULATIVE_SUMMATION_OPERATOR_DESC {
  const DML_TENSOR_DESC *InputTensor;
  const DML_TENSOR_DESC *OutputTensor;
  UINT                  Axis;
  DML_AXIS_DIRECTION    AxisDirection;
  BOOL                  HasExclusiveSum;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

The input tensor containing elements to be summed.

OutputTensor

Type: const DML_TENSOR_DESC*

The output tensor to write the resulting cumulative summations to. This tensor must have the same sizes and data type as the InputTensor.

Axis

Type: UINT

The index of the dimension to sum elements over. This value must be less than the DimensionCount of the InputTensor.

AxisDirection

Type: DML_AXIS_DIRECTION

One of the values of the DML_AXIS_DIRECTION enumeration. If set to DML_AXIS_DIRECTION_INCREASING, then the summation occurs by traversing the tensor along the specified axis by ascending element index. If set to DML_AXIS_DIRECTION_DECREASING, the reverse is true, and the summation occurs by traversing elements by descending index.

HasExclusiveSum

Type: BOOL

If TRUE, then the value of the current element is excluded when writing the running tally to the output tensor. If FALSE, then the value of the current element is included in the running tally.

Examples

The examples in this section all use an input tensor with the following properties.

InputTensor: (Sizes:{1,1,3,4}, DataType:FLOAT32)
[[[[2, 1, 3, 5],
   [3, 8, 7, 3],
   [9, 6, 2, 4]]]]

Example 1. Cumulative summation across horizontal slivers

Axis: 3
AxisDirection: DML_AXIS_DIRECTION_INCREASING
HasExclusiveSum: FALSE

OutputTensor: (Sizes:{1,1,3,4}, DataType:FLOAT32)
[[[[2,  3,  6, 11],     // i.e. [2, 2+1, 2+1+3, 2+1+3+5]
   [3, 11, 18, 21],     //      [...                   ]
   [9, 15, 17, 21]]]]   //      [...                   ]

Example 2. Exclusive sums

Setting HasExclusiveSum to TRUE has the effect of excluding the current element's value from the running tally when writing to the output tensor.

Axis: 3
AxisDirection: DML_AXIS_DIRECTION_INCREASING
HasExclusiveSum: TRUE

OutputTensor: (Sizes:{1,1,3,4}, DataType:FLOAT32)
[[[[0, 2,  3,  6],      // Notice the sum is written before adding the input,
   [0, 3, 11, 18],      // and the final total is not written to any output.
   [0, 9, 15, 17]]]]

Example 3. Axis direction

Setting the AxisDirection to DML_AXIS_DIRECTION_DECREASING has the effect of reversing the traversal order of elements when computing the running tally.

Axis: 3
AxisDirection: DML_AXIS_DIRECTION_DECREASING
HasExclusiveSum: FALSE

OutputTensor: (Sizes:{1,1,3,4}, DataType:FLOAT32)
[[[[11,  9,  8,  5],    // i.e. [2+1+3+5, 1+3+5, 3+5, 5]
   [21, 18, 10,  3],    //      [...                   ]
   [21, 12,  6,  4]]]]  //      [...                   ]

Example 4. Summing along a different axis

In this example, the summation occurs vertically, along the height axis (dimension 2).

Axis: 2
AxisDirection: DML_AXIS_DIRECTION_INCREASING
HasExclusiveSum: FALSE

OutputTensor: (Sizes:{1,1,3,4}, DataType:FLOAT32)
[[[[ 2,  1,  3,  5],   // i.e. [2,    ...]
   [ 5,  9, 10,  8],   //      [2+3,  ...]
   [14, 15, 12, 12]]]] //      [2+3+9 ...]

Remarks

This operator supports in-place execution, meaning that the OutputTensor is permitted to alias the InputTensor during binding.

Availability

This operator was introduced in DML_FEATURE_LEVEL_2_1.

Tensor constraints

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

Tensor support

DML_FEATURE_LEVEL_5_0 and above

Tensor	Kind	Supported dimension counts	Supported data types
InputTensor	Input	1 to 8	FLOAT32, FLOAT16, INT64, INT32, UINT64, UINT32
OutputTensor	Output	1 to 8	FLOAT32, FLOAT16, INT64, INT32, UINT64, UINT32

DML_FEATURE_LEVEL_4_0 and above

Tensor	Kind	Supported dimension counts	Supported data types
InputTensor	Input	1 to 8	FLOAT32, FLOAT16, INT32, UINT32
OutputTensor	Output	1 to 8	FLOAT32, FLOAT16, INT32, UINT32

DML_FEATURE_LEVEL_2_1 and above

Tensor	Kind	Supported dimension counts	Supported data types
InputTensor	Input	4	FLOAT32, FLOAT16, INT32, UINT32
OutputTensor	Output	4	FLOAT32, FLOAT16, INT32, UINT32

Requirements

Requirement	Value
Minimum supported client	Windows 10 Build 20348
Minimum supported server	Windows 10 Build 20348
Header	directml.h

Share via