A primitive to perform local response normalization (LRN) across or within channels. More...

Functions
mkldnn_status_t MKLDNN_API	mkldnn_lrn_forward_desc_init (mkldnn_lrn_desc_t lrn_desc, mkldnn_prop_kind_t prop_kind, mkldnn_alg_kind_t alg_kind, const mkldnn_memory_desc_t data_desc, int local_size, float alpha, float beta, float k)
	Initializes an `lrn_desc` for forward propagation using `prop_kind` (possible values are mkldnn_forward_training or mkldnn_forward_inference), `alg_kind`, memory descriptor `data_desc`, and regularization parameters `local_size`, `alpha`, `beta`, and `k`. More...

mkldnn_status_t MKLDNN_API	mkldnn_lrn_backward_desc_init (mkldnn_lrn_desc_t lrn_desc, mkldnn_alg_kind_t alg_kind, const mkldnn_memory_desc_t diff_data_desc, const mkldnn_memory_desc_t *data_desc, int local_size, float alpha, float beta, float k)
	Initializes an `lrn_desc` for backward propagation using `alg_kind`, memory descriptors `data_desc`, and `diff_data_desc`, and regularization parameters `local_size`, `alpha`, `beta`, and `k`. More...

Detailed Description

A primitive to perform local response normalization (LRN) across or within channels.

LRN accross channels:

$dst[n][c][h][w] = \left\{k + \frac{\alpha}{n_{l}} \sum\limits_{i=-(n_{l}-1)/2}^{(n_{l}+1)/2} (src[n][c+i][h][w])^2\right\}^{-\beta} src[n][c][h][w],$

LRN within channels:

$dst[n][c][h][w] = \left\{k + \frac{\alpha}{n_{l}} \sum\limits_{i=-(n_{l}-1)/2}^{(n_{l}+1)/2} (src[n][c][h+i][w+i])^2\right\}^{-\beta} src[n][c][h][w],$

where $n_{l}$ is the local_size.

During training LRN might or might not require workspace on forward (mkldnn_forward_training) and backward (mkldnn_backward) passes. The behavior is implementation specific. Optimized implementations typically require workspace and use it to save some intermediate results from the forward pass that accelerate computations on the backward pass.

To check whether workspace is required one should query the LRN primitive descriptor for the workspace (mkldnn_query_workspace_pd). Success would indicate the workspace is required and its description would be returned.

See also: mkldnn_primitive_desc_query and mkldnn_primitive_desc_query_pd

Function Documentation

◆ mkldnn_lrn_forward_desc_init()

mkldnn_status_t MKLDNN_API mkldnn_lrn_forward_desc_init	(	mkldnn_lrn_desc_t *	lrn_desc,
		mkldnn_prop_kind_t	prop_kind,
		mkldnn_alg_kind_t	alg_kind,
		const mkldnn_memory_desc_t *	data_desc,
		int	local_size,
		float	alpha,
		float	beta,
		float	k
	)

Initializes an lrn_desc for forward propagation using prop_kind (possible values are mkldnn_forward_training or mkldnn_forward_inference), alg_kind, memory descriptor data_desc, and regularization parameters local_size, alpha, beta, and k.

Order of inputs:

src (mkldnn_query_src_pd, 0)

Order of outputs:

dst (mkldnn_query_dst_pd, 0)
workspace (mkldnn_query_workspace_pd, 0), if the underlying implementation requires

◆ mkldnn_lrn_backward_desc_init()

mkldnn_status_t MKLDNN_API mkldnn_lrn_backward_desc_init	(	mkldnn_lrn_desc_t *	lrn_desc,
		mkldnn_alg_kind_t	alg_kind,
		const mkldnn_memory_desc_t *	diff_data_desc,
		const mkldnn_memory_desc_t *	data_desc,
		int	local_size,
		float	alpha,
		float	beta,
		float	k
	)

Initializes an lrn_desc for backward propagation using alg_kind, memory descriptors data_desc, and diff_data_desc, and regularization parameters local_size, alpha, beta, and k.