A primitive to perform softmax. More...

Functions
mkldnn_status_t MKLDNN_API	mkldnn_softmax_forward_desc_init (mkldnn_softmax_desc_t softmax_desc, mkldnn_prop_kind_t prop_kind, const mkldnn_memory_desc_t data_desc, int softmax_axis)
	Initializes a `softmax_desc` for forward propagation using `prop_kind` (possible values are mkldnn_forward_training and mkldnn_forward_inference) and memory descriptor `data_desc`. More...

mkldnn_status_t MKLDNN_API	mkldnn_softmax_backward_desc_init (mkldnn_softmax_desc_t softmax_desc, const mkldnn_memory_desc_t diff_desc, const mkldnn_memory_desc_t *data_desc, int softmax_axis)
	Initializes a `softmax_desc` for backward propagation using memory descriptors `diff_desc` and `data_desc`. More...

Detailed Description

A primitive to perform softmax.

$dst[u][c][in] = \frac{\exp(src[ou][c][in]) - \max\limits_{c}(src[ou][c][in])} {\sum\limits_{c}\{\exp(src[ou][c][in]) - \max\limits_{c}(src[ou][c][in])\}},$

where $ou, iu$ are outer and inner sizes repectively, defined by data_desc.dims and softmax_axis.

Function Documentation

◆ mkldnn_softmax_forward_desc_init()

mkldnn_status_t MKLDNN_API mkldnn_softmax_forward_desc_init	(	mkldnn_softmax_desc_t *	softmax_desc,
		mkldnn_prop_kind_t	prop_kind,
		const mkldnn_memory_desc_t *	data_desc,
		int	softmax_axis
	)

Initializes a softmax_desc for forward propagation using prop_kind (possible values are mkldnn_forward_training and mkldnn_forward_inference) and memory descriptor data_desc.

Order of inputs:

src (mkldnn_query_src_pd, 0)

Order of outputs:

dst (mkldnn_query_dst_pd, 0)

◆ mkldnn_softmax_backward_desc_init()

mkldnn_status_t MKLDNN_API mkldnn_softmax_backward_desc_init	(	mkldnn_softmax_desc_t *	softmax_desc,
		const mkldnn_memory_desc_t *	diff_desc,
		const mkldnn_memory_desc_t *	data_desc,
		int	softmax_axis
	)

Initializes a softmax_desc for backward propagation using memory descriptors diff_desc and data_desc.