struct dnnl_layer_normalization_desc_t


A descriptor of a Layer Normalization operation. More…

#include <dnnl_types.h>

struct dnnl_layer_normalization_desc_t
    // fields

    dnnl_primitive_kind_t primitive_kind;
    dnnl_prop_kind_t prop_kind;
    dnnl_memory_desc_t data_desc;
    dnnl_memory_desc_t diff_data_desc;
    dnnl_memory_desc_t data_scaleshift_desc;
    dnnl_memory_desc_t diff_data_scaleshift_desc;
    dnnl_memory_desc_t stat_desc;
    float layer_norm_epsilon;
    unsigned flags;

Detailed Documentation

A descriptor of a Layer Normalization operation.


dnnl_primitive_kind_t primitive_kind

The kind of primitive.

Used for self-identifying the primitive descriptor. Must be dnnl_layer_normalization.

dnnl_prop_kind_t prop_kind

The kind of propagation.

Possible values: dnnl_forward_training, dnnl_forward_inference, dnnl_backward, and dnnl_backward_data.

dnnl_memory_desc_t data_desc

Source and destination memory descriptor.

dnnl_memory_desc_t diff_data_desc

Source and destination gradient memory descriptor.

dnnl_memory_desc_t data_scaleshift_desc

Scale and shift data and gradient memory descriptors.

Scaleshift memory descriptor uses 2D dnnl_ab format[2, normalized_dim] where 1-st dimension contains gamma parameter, 2-nd dimension contains beta parameter. Normalized_dim is equal to the last logical dimension of the data tensor across which normalization is performed.

dnnl_memory_desc_t stat_desc

Mean and variance data memory descriptors.

Statistics (mean and variance) memory descriptor is the k-dimensional tensor where k is equal to data_tensor_ndims - 1 and may have any plain (stride[last_dim] == 1) user-provided format.

float layer_norm_epsilon

Layer normalization epsilon parameter.