The DNNL documentation relies on a set of standard naming conventions for variables. This section describes these conventions.
Neural network models consist of operations of the following form:
\[ \dst = f(\src, \weights), \]
where \(\dst\) and \(\src\) are activation tensors, and \(\weights\) are learnable tensors.
The backward propagation consists then in computing the gradients with respect to the \(\src\) and \(\weights\) respectively:
\[ \diffsrc = df_{\src}(\diffdst, \src, \weights, \dst), \]
and
\[ \diffweights = df_{\weights}(\diffdst, \src, \weights, \dst). \]
While DNNL uses src, dst, and weights as generic names for the activations and learnable tensors, for a specific operation there might be commonly used and widely known specific names for these tensors. For instance, the convolution operation has a learnable tensor called bias. For usability reasons, DNNL primitives use such names in initialization or other functions to simplify the coding.
To summarize, DNNL uses the following commonly used notations for tensors:
| Name | Me |
|---|---|
src | Source tensor |
dst | Destination tensor |
weights | Weights tensor |
bias | Bias tensor (used in Convolution, Inner Product and other primitives) |
scale_shift | Scale and shift tensors (used in Batch Normalization and Layer Normalization) |
workspace | Workspace tensor that carries additional information from the forward propagation to the backward propagation |
scratchpad | Temporary tensor that is required to store the intermediate results |
diff_src | Gradient tensor with respect to the source |
diff_dst | Gradient tensor with respect to the destination |
diff_weights | Gradient tensor with respect to the weights |
diff_bias | Gradient tensor with respect to the bias |
diff_scale_shift | Gradient tensor with respect to the scale and shift |
*_layer | RNN layer data or weights tensors |
*_iter | RNN recurrent data or weights tensors |
DNNL uses the following notations in the documentation formulas and verbose output. Here, lower-case letters are used to denote indices in a particular spatial dimension, the sizes of which are denoted by corresponding upper-case letters.
| Name | Semantics |
|---|---|
n (or mb) | batch |
g | groups |
oc, od, oh, ow | output channels, depth, height, and width |
ic, id, ih, iw | input channels, depth, height, and width |
kd, kh, kw | kernel (filter) depth, height, and width |
sd, sh, sw | stride by depth, height, and width |
dd, dh, dw | dilation by depth, height, and width |
pd, ph, pw | padding by depth, height, and width |
The following notations are used when describing RNN primitives.
| Name | Semantics |
|---|---|
| \(\cdot\) | matrix multiply operator |
| \( * \) | element-wise multiplication operator |
| W | input weights |
| U | recurrent weights |
| \(^T\) | transposition |
| B | bias |
| h | hidden state |
| a | intermediate value |
| x | input |
| \( _t {}_{}\) | timestamp |
| \( l \) | layer index |
| activation | tanh, relu, logistic |
| c | cell state |
| \(\tilde{c}\) | candidate state |
| i | input gate |
| f | forget gate |
| o | output gate |
| u | update gate |
| r | reset gate |
When describing tensor memory formats, which is the DNNL term for the way that the data is laid out in memory, documentation uses letters of the English alphabet to describe an order of dimensions and their semantics.
The canonical sequence of letters is a, b, c, ..., z. In this notation, the ab tag denotes a two-dimensional tensor with a denoting the outermost dimension and b denoting the innermost dimension, where the latter is dense in memory. Further, the ba tag denotes a two-dimensional tensor but with last two dimensions transposed: instead of the naturally dense b dimension, now a is the dense dimension. If we suppose that the two-dimensional tensor is a matrix and the a and b dimensions represent the number of columns and rows, then ab would denote the row-major (C) format and ba would denote the column-major (Fortran) format.
Upper-case letters are used to indicate that the data is laid out in blocks for a particular dimension. In such cases, the format name contains both upper- and lower-case letters for that dimension with a lower-case letter preceded by the block size. For example, the Ab16a tag denotes a format similar to row-major but with columns split into contiguous blocks of 16 elements each. Moreover, the implicit assumption is that if the number of columns is not divisible by 16, the last block in the in-memory representation will contain padding zeroes.
Since there are many widely used names for specific deep learning domains like convolutional neural networks (CNNs), DNNL also supports memory format tags in which dimensions have specifically assigned meaning like 'image width', 'image height', etc. The following table summarizes notations used in such memory format tags.
| Letter | Dimension |
|---|---|
| n | batch |
| g | groups |
| c | channels |
| o | output channels |
| i | input channels |
| h | height |
| w | width |
| d | depth |
| t | timestamp |
| l | layer |
| d | direction |
| g | gate |
| s | state |
The canonical sequence of dimensions for four-dimensional data tensors in CNNs is (batch, channels, spatial dimensions). Spatial dimensions are ordered for tensors with three spatial dimensions as (depth, height, width), for tensors with two spatial dimensions as (height, width), and as just (width) for tensors with only one spatial dimension.
In this notation, nchw is a memory format tag for a four-dimensional tensor, with the first dimension corresponding to batch, the second to channel, and the remaining two to spatial dimensions. Due to the canonical order of dimensions for CNNs, this tag is the same as abcd. As another example, nhwc is the same as acdb.