Rate This Document

Findability

Accuracy

Completeness

Readability

Function Description

Table 1 lists the operators supported by KuDNN.

**Table 1** Supported operators
Operator	Description
GEMM (single-threaded) & Linear (multi-threaded)	Matrix multiplication operator.
Linear+postops	Performs ReLU, SiLU, GeLU, ReSEXT, ResMul, and Residential operations following the Linear operator computation.
Softmax	Softmax normalization operator.
Conv2d/Conv3d	2D/3D convolution operator.
Group Normalization	Group normalization operator.
Layer Normalization	Performs layer normalization.
Embedding	Mapping operator.
Root Mean Square Normalization	Root mean square normalization operator.
Rotary Position Embedding	Rotary positional embedding (RoPE) operator.

Table 2 describes the common objects.

**Table 2** Parameters to be passed for initializing KuDNN::TensorInfo
Parameter	Data Type	Description
shape	struct KuDNN::Shape	{d1, d2, d3, m, n} (up to 5D) must correspond to the layout.
type	struct KuDNN::Element::Type	The following four enumeration types are supported: KuDNN::Element::TypeT::F16 KuDNN::Element::TypeT::F32 KuDNN::Element::TypeT::BF16 KuDNN::Element::TypeT::S8
layout	enum class KuDNN::Layout	Enumeration type. For details about the supported layouts, see the table of layout types supported during the initialization of the TensorInfo object of each operator.

Parent topic: KuDNN Functions