Deep Neural Network Library (DNNL)
1.2.0

Performance library for Deep Learning

Shuffle

The shuffle primitive shuffles data along the shuffle axis (here is designated as \(C\)) with the group parameter \(G\). Namely, the shuffle axis is thought to be a 2D tensor of size \((\frac{C}{G} \times G)\) and it is being transposed to \((G \times \frac{C}{G})\).

The formal definition is shown below:

\[ dst(\overline{ou}, c, \overline{in}) = src(\overline{ou}, c', \overline{in}) \]

where

- \(c\) dimension is called a shuffle axis,
- \(G\) is a
`group_size`

, - \(\overline{ou}\) is the outermost indices (to the left from shuffle axis),
- \(\overline{in}\) is the innermost indices (to the right from shuffle axis), and
\(c'\) and \(c\) relate to each other as define by the system:

\[ \begin{cases} c &= u + v\frac{C}{G}, \\ c' &= uG + v, \\ \end{cases} \]

Here, \(u \in [0, \frac{C}{G})\) and \(v \in [0, G)\).

There is no difference between the dnnl_forward_training and dnnl_forward_inference propagation kinds.

The backward propagation computes \(diff\_src(ou, c, in)\), based on \(diff\_dst(ou, c, in)\).

Essentially, backward propagation is the same as forward propagation with \(g\) replaced by \(C / g\).

- The memory format and data type for
`src`

and`dst`

are assumed to be the same, and in the API are typically referred as`data`

(e.g., see`data_desc`

in dnnl::shuffle_forward::desc::desc()). The same holds for`diff_src`

and`diff_dst`

. The corresponding memory descriptors are referred to as`diff_data_desc`

.

The shuffle primitive supports the following combinations of data types:

Propagation | Source / Destination |
---|---|

forward / backward | f32, bf16 |

forward | s32, s8, u8 |

- Warning
- There might be hardware and/or implementation specific restrictions. Check Implementation Limitations section below.

The shuffle primitive works with arbitrary data tensors. There is no special meaning associated with any logical dimensions. However, the shuffle axis is typically referred to as channels (hence in formulas we use \(c\)).

Shuffle operation typically appear in CNN topologies. Hence, in the library the shuffle primitive is optimized for the corresponding memory formats:

Spatial | Logical tensor | Shuffle Axis | Implementations optimized for memory formats |
---|---|---|---|

2D | NCHW | 1 (C) | dnnl_nchw (dnnl_abcd), dnnl_nhwc (dnnl_acdb), optimized^ |

3D | NCDHW | 1 (C) | dnnl_ncdhw (dnnl_abcde), dnnl_ndhwc (dnnl_acdeb), optimized^ |

Here *optimized^* means the format that comes out of any preceding compute-intensive primitive.

The shuffle primitive doesn't support any post-ops or attributes.

- Refer to Data Types for limitations related to data types support.

N/A