Roll#

Versioned name: Roll-7

Category: Data movement

Short description: The Roll operation shifts elements of a tensor along specified axes.

Detailed description: Roll produces a tensor with the same shape as the first input tensor and with elements shifted along dimensions specified in the axes tensor. The shift size is specified in the shift input tensor. Elements that are shifted beyond the last position will be added in the same order starting from the first position.

Example 1. Roll output with shift = 1, axes = 0:

data    = [[ 1,  2,  3],
        [ 4,  5,  6],
        [ 7,  8,  9],
        [10, 11, 12]]
output  = [[10, 11, 12],
        [ 1,  2,  3],
        [ 4,  5,  6],
        [ 7,  8,  9]]

Example 2. Roll output with shift = [-1, 2], axes = [0, 1]:

data    = [[ 1,  2,  3],
        [ 4,  5,  6],
        [ 7,  8,  9],
        [10, 11, 12]]
output  = [[ 5,  6,  4],
        [ 8,  9,  7],
        [11, 12, 10],
        [ 2,  3,  1]]

Example 3. Roll output with shift = [1, 2, 1], axes = [0, 1, 0]:

data    = [[ 1,  2,  3],
        [ 4,  5,  6],
        [ 7,  8,  9],
        [10, 11, 12]]
output  = [[ 8,  9,  7],
        [11, 12, 10],
        [ 2,  3,  1],
        [ 5,  6,  4]]

Attributes

No attributes available.

Inputs:

1: data a tensor of type T. Required.
2: a shift scalar or 1D tensor of type T_IND_1. Specifies the number of places by which the elements of the data tensor are shifted. If shift is a scalar, each dimension specified in the axes tensor are rolled by the same shift value. If shift is a 1D tensor, axes must be a 1D tensor of the same size, and each dimension from axes tensor are rolled by the corresponding value from the shift tensor. If the value of shift is positive, elements are shifted positively (towards larger indices). Otherwise, elements are shifted negatively (towards smaller indices). Required.
3: axes a scalar or 1D tensor of type T_IND_2. Specifies axes along which elements are shifted. If the same axis is referenced more than once, the total shift for that axis will be the sum of all the shifts that belong to that axis. If axes has negative value, axis index will be calculated using the formula: N_dims + axis, where N_dims - total number of dimensions in the data tensor, axis - negative axis index from the axes tensor. Required.

Outputs:

1: output tensor with shape and type equal to the data tensor.

Types

T: any supported type.
T_IND_1: int32 or int64.
T_IND_2: int32 or int64.

Example

Example 1: “shift” and “axes” are 1D tensors.

 <layer ... type="Roll">
     <input>
         <port id="0">
             <dim>3</dim>
             <dim>10</dim>
             <dim>100</dim>
             <dim>200</dim>
         </port>
         <port id="1">
             <dim>2</dim>
         </port>
         <port id="2">
             <dim>2</dim> <!-- shifting along specified axes with the corresponding shift values -->
         </port>
     </input>
     <output>
         <port id="0">
             <dim>3</dim>
             <dim>10</dim>
             <dim>100</dim>
             <dim>200</dim>
         </port>
     </output>
 </layer>

Example 2: “shift” value is a scalar and multiple axes are specified.

 <layer ... type="Roll">
     <input>
         <port id="0">
             <dim>3</dim>
             <dim>10</dim>
             <dim>100</dim>
             <dim>200</dim>
         </port>
         <port id="1">
             <dim>1</dim>
         </port>
         <port id="2">
             <dim>2</dim> <!-- shifting along specified axes with the same shift value -->
         </port>
     </input>
     <output>
         <port id="0">
             <dim>3</dim>
             <dim>10</dim>
             <dim>100</dim>
             <dim>200</dim>
         </port>
     </output>
 </layer>