Collaboration diagram for Inner product:

Functions
mkldnn_status_t MKLDNN_API	mkldnn_inner_product_forward_desc_init (mkldnn_inner_product_desc_t ip_desc, mkldnn_prop_kind_t prop_kind, const mkldnn_memory_desc_t src_desc, const mkldnn_memory_desc_t weights_desc, const mkldnn_memory_desc_t bias_desc, const mkldnn_memory_desc_t *dst_desc)

mkldnn_status_t MKLDNN_API	mkldnn_inner_product_backward_data_desc_init (mkldnn_inner_product_desc_t ip_desc, const mkldnn_memory_desc_t diff_src_desc, const mkldnn_memory_desc_t weights_desc, const mkldnn_memory_desc_t diff_dst_desc)

mkldnn_status_t MKLDNN_API	mkldnn_inner_product_backward_weights_desc_init (mkldnn_inner_product_desc_t ip_desc, const mkldnn_memory_desc_t src_desc, const mkldnn_memory_desc_t diff_weights_desc, const mkldnn_memory_desc_t diff_bias_desc, const mkldnn_memory_desc_t *diff_dst_desc)

Detailed Description

A primitive to compute an inner product.

Inner product layer is also known as fully connected layer. With spatial dimension:

$dst[n][oc] = \sum\limits_{ic, kh, kw} src[n][ic][kh][kw] \cdot weights[oc][ic][kh][kw] + bias[oc]$

Function Documentation

mkldnn_status_t MKLDNN_API mkldnn_inner_product_backward_data_desc_init	(	mkldnn_inner_product_desc_t *	ip_desc,
		const mkldnn_memory_desc_t *	diff_src_desc,
		const mkldnn_memory_desc_t *	weights_desc,
		const mkldnn_memory_desc_t *	diff_dst_desc
	)

Initializes an inner product descriptor ip_desc for backward propagation with respect to data using memory descriptors.

Note: Memory descriptors are allowed to be initialized with mkldnn_any value of format_kind.

Order of inputs:

diff_dst (mkldnn_query_diff_dst_pd, 0)
weights (mkldnn_query_weights_pd, 0)

Order of outputs:

diff_src (mkldnn_query_diff_src_pd, 0)

mkldnn_status_t MKLDNN_API mkldnn_inner_product_backward_weights_desc_init	(	mkldnn_inner_product_desc_t *	ip_desc,
		const mkldnn_memory_desc_t *	src_desc,
		const mkldnn_memory_desc_t *	diff_weights_desc,
		const mkldnn_memory_desc_t *	diff_bias_desc,
		const mkldnn_memory_desc_t *	diff_dst_desc
	)

Initializes an inner product descriptor ip_desc for backward propagation with respect to weights using memory descriptors.

Note: Memory descriptors are allowed to be initialized with mkldnn_any value of format_kind.

Order of inputs:

src (mkldnn_query_src_pd, 0)
diff_dst (mkldnn_query_diff_dst_pd, 0)

Order of outputs:

diff_weights (mkldnn_query_diff_weights_pd, 0)
diff_bias (mkldnn_query_diff_weights_pd, 1), if created with bias

mkldnn_status_t MKLDNN_API mkldnn_inner_product_forward_desc_init	(	mkldnn_inner_product_desc_t *	ip_desc,
		mkldnn_prop_kind_t	prop_kind,
		const mkldnn_memory_desc_t *	src_desc,
		const mkldnn_memory_desc_t *	weights_desc,
		const mkldnn_memory_desc_t *	bias_desc,
		const mkldnn_memory_desc_t *	dst_desc
	)

Initializes an inner product descriptor ip_desc for forward propagation using prop_kind (possible values are mkldnn_forward_training and mkldnn_forward_inference) and memory descriptors. In order to create an inner product without bias, bias_desc should be either NULL or a pointer to a descriptor with memory format equal to mkldnn_format_undef.

Note: Memory descriptors are allowed to be initialized with mkldnn_any value of format_kind.

Order of inputs:

src (mkldnn_query_src_pd, 0)
weights (mkldnn_query_weights_pd, 0)
bias (mkldnn_query_weights_pd, 1), if created with bias

Order of outputs:

dst (mkldnn_query_dst_pd, 0)

Functions

Detailed Description

Function Documentation