Contents

Softmax

API Reference

General

The softmax primitive performs forward or backward softmax or logsoftmax operation along a particular axis on data with arbitrary dimensions. All other axes are treated as independent (batch).

Forward

In general form, the operation is defined by the following formulas (the variable names follow the standard Naming Conventions).

Softmax:

\[\dst(\overline{ou}, c, \overline{in}) = \frac {e^{\src(\overline{ou}, c, \overline{in}) - \nu(\overline{ou}, \overline{in})}} { \sum\limits_{ic} e^{\src(\overline{ou}, ic, \overline{in}) - \nu(\overline{ou}, \overline{in})} }\]

Logsoftmax:

\[\dst(\overline{ou}, c, \overline{in}) = \ln\left({\frac { e^{\src(\overline{ou}, c, \overline{in}) - \nu(\overline{ou}, \overline{in})} } { \sum\limits_{ic} e^{\src(\overline{ou}, ic, \overline{in}) - \nu(\overline{ou}, \overline{in})} }}\right) = \left(\src(\overline{ou}, c, \overline{in}) - \nu(\overline{ou}, \overline{in})\right) - \ln\left( \sum\limits_{ic} e^{\src(\overline{ou}, ic, \overline{in}) - \nu(\overline{ou}, \overline{in})} \right)\]

Above

  • \(c\) is the axis over which the operation is computed on,

  • \(\overline{ou}\) is the outermost index (to the left of the axis),

  • \(\overline{in}\) is the innermost index (to the right of the axis), and

  • \(\nu\) is used to produce numerically stable results and defined as:

    \[\nu(\overline{ou}, \overline{in}) = \max\limits_{ic} \src(\overline{ou}, ic, \overline{in})\]

Difference Between Forward Training and Forward Inference

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

Backward

The backward propagation computes \(\diffsrc(ou, c, in)\), based on \(\diffdst(ou, c, in)\) and \(\dst(ou, c, in)\).

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

\(\src\)

DNNL_ARG_SRC

\(\dst\)

DNNL_ARG_DST

\(\diffsrc\)

DNNL_ARG_DIFF_SRC

\(\diffdst\)

DNNL_ARG_DIFF_DST

\(src scale\)

DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC

\(dst scale\)

DNNL_ARG_ATTR_SCALES | DNNL_ARG_DST

\(\text{binary post-op}\)

DNNL_ARG_ATTR_MULTIPLE_POST_OP(binary_post_op_position) | DNNL_ARG_SRC_1

Implementation Details

General Notes

  1. Both forward and backward propagation support in-place operations, meaning that \(\src\) can be used as input and output for forward propagation, and \(\diffdst\) can be used as input and output for backward propagation. In case of in-place operation, the original data will be overwritten. This support is limited to cases when data types of \(\src\) and \(\dst\) or \(\diffsrc\) and \(\diffdst\) are identical.

Post-ops and Attributes

Attributes enable you to modify the behavior of the softmax primitive. The following attributes are supported by the softmax primitive:

Propagation

Type

Operation

Description

Restrictions

forward

attribute

Scales

Scales the corresponding tensor by the given scale factor(s).

Supported only for int8 softmax and one scale per tensor is supported.

forward

post-op

Binary

Applies a Binary operation to the result

General binary post-op restrictions

forward

Post-op

Eltwise

Applies an Eltwise operation to the result.

forward

attribute

Accumulation mode

Defines the implementation’s accumulation arithmetic.

Only the values strict , relaxed , and any are supported.

Accumulation Mode

You can optimize performance of the forward operation when the source and destination floating-point data types of the operation are equal and different from f32. When the destination data type is different from f32, additional memory will be used to accumulate data and store it in the destination memory buffer for a requested data type. Using the additional memory can be opted-out with an accumulation mode setting set to relaxed or any, which will use the precision of destination data type to accumulate intermediate results directly into the destination memory buffer. This performance optimization, however, results in in a minor decrease in accuracy. Depending on the actual data, the difference between strict and relaxed accumulation can reach several units in the last piece (ulps).

Data Type Support

The softmax primitive supports the following combinations of data types:

Propagation

Source

Destination

forward

f32, f64, bf16, f16, u8, s8

f32, f64, bf16, f16, u8, s8

backward

f32, f64, bf16, f16

f32, f64, bf16, f16

Data Representation

Source, Destination, and Their Gradients

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

Implementation Limitations

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

  2. GPU

    • Only tensors of 6 or fewer dimensions are supported.

Performance Tips

  1. Use in-place operations whenever possible.

Example

Softmax Primitive Example

This C++ API example demonstrates how to create and execute a Softmax primitive in forward training propagation mode.

Key optimizations included in this example:

  • In-place primitive execution;

  • Softmax along axis 1 (C) for 2D tensors.

Shuffle Sum