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 \((G \times \frac{C}{G})\). Variable names follow the standard Conventions.
The formal definition is shown below:
Forward¶
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:
Here, \(0 \leq u < \frac{C}{G}\) and \(0 \leq v < G\).
Difference Between Forward Training and Forward Inference¶
There is no difference between the forward_training and forward_inference
propagation kinds.
Backward¶
The backward propagation computes \(\diffsrc(ou, c, in)\), based on \(\diffdst(ou, c, in)\).
Essentially, backward propagation is the same as forward propagation with \(g\) replaced by \(C / g\).
Execution Arguments¶
When executed, the inputs and outputs should be mapped to an execution argument index as specified by the following table.
Primitive input/output |
Execution argument index |
|---|---|
\(\src\) |
|
\(\dst\) |
|
\(\diffsrc\) |
|
\(\diffdst\) |
Operation Details¶
The memory format and data type for
srcanddstare assumed to be the same, and in the API are typically referred asdata(e.g., seedata_descindnnl::shuffle_forward::desc::desc()). The same holds fordiff_srcanddiff_dst. The corresponding memory descriptors are referred to asdiff_data_desc.
Data Types Support¶
The shuffle primitive supports the following combinations of data types:
Note
Here we abbreviate data types names for readability. For example, dnnl::memory::data_type::f32 is
abbreviated to f32.
Propagation |
Source / Destination |
|---|---|
forward / backward |
|
forward |
Data Layouts¶
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) |
|
3D |
NCDHW |
1 (C) |
Here optimized^ means the format that comes out of any preceding compute-intensive primitive.
Post-ops and Attributes¶
The shuffle primitive does not have to support any post-ops or attributes.
API¶
-
struct
dnnl::shuffle_forward: public dnnl::primitive¶ Shuffle forward propagation primitive.
Public Functions
-
shuffle_forward()¶ Default constructor. Produces an empty object.
-
shuffle_forward(const primitive_desc &pd)¶ Constructs a shuffle forward propagation primitive.
- Parameters
pd: Primitive descriptor for a shuffle forward propagation primitive.
-
struct
desc¶ Descriptor for a shuffle forward propagation primitive.
Public Functions
-
desc(prop_kind aprop_kind, const memory::desc &data_desc, int axis, int group_size)¶ Constructs a descriptor for a shuffle forward propagation primitive.
- Parameters
aprop_kind: Propagation kind. Possible values are dnnl::prop_kind::forward_training, and dnnl::prop_kind::forward_inference.data_desc: Source and destination memory descriptor.axis: The axis along which the data is shuffled.group_size: Shuffle group size.
-
-
struct
primitive_desc: public dnnl::primitive_desc¶ Primitive descriptor for a shuffle forward propagation primitive.
Public Functions
-
primitive_desc()¶ Default constructor. Produces an empty object.
-
primitive_desc(const desc &adesc, const engine &aengine, const primitive_attr &attr = primitive_attr(), bool allow_empty = false)¶ Constructs a primitive descriptor for a shuffle forward propagation primitive.
- Parameters
adesc: Descriptor for a shuffle forward propagation primitive.aengine: Engine to use.attr: Primitive attributes to use.allow_empty: A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.
-
-
-
struct
dnnl::shuffle_backward: public dnnl::primitive¶ Shuffle backward propagation primitive.
Public Functions
-
shuffle_backward()¶ Default constructor. Produces an empty object.
-
shuffle_backward(const primitive_desc &pd)¶ Constructs a shuffle backward propagation primitive.
- Parameters
pd: Primitive descriptor for a shuffle backward propagation primitive.
-
struct
desc¶ Descriptor for a shuffle primitive backward propagation primitive.
Public Functions
-
struct
primitive_desc: public dnnl::primitive_desc¶ Primitive descriptor for a shuffle backward propagation primitive.
Public Functions
-
primitive_desc()¶ Default constructor. Produces an empty object.
-
primitive_desc(const desc &adesc, const engine &aengine, const shuffle_forward::primitive_desc &hint_fwd_pd, const primitive_attr &attr = primitive_attr(), bool allow_empty = false)¶ Constructs a primitive descriptor for a shuffle backward propagation primitive.
- Parameters
adesc: Descriptor for a shuffle backward propagation primitive.aengine: Engine to use.attr: Primitive attributes to use.hint_fwd_pd: Primitive descriptor for a shuffle forward propagation primitive. It is used as a hint for deciding which memory format to use.allow_empty: A flag signifying whether construction is allowed to fail without throwing an exception. In this case an empty object will be produced. This flag is optional and defaults to false.
-
-