dnnl::primitive_attr Struct Reference

Primitive attributes. More...

#include <dnnl.hpp>

Inheritance diagram for dnnl::primitive_attr:

Collaboration diagram for dnnl::primitive_attr:

Public Member Functions
	primitive_attr ()
	Creates default primitive attributes.
	primitive_attr (dnnl_primitive_attr_t attr)
	Creates primitive attributes from a C dnnl_primitive_attr_t handle. More...
scratchpad_mode	get_scratchpad_mode () const
	Returns the scratchpad mode.
void	set_scratchpad_mode (scratchpad_mode mode)
	Sets scratchpad mode.
void	get_output_scales (int &mask, std::vector< float > &scales) const
	Gets correspondence scale mask and a constant floating point vector of output scales previously set by set_output_scales.
void	set_output_scales (int mask, const std::vector< float > &scales)
	Sets output scales for primitive operations. More...
const post_ops	get_post_ops () const
	Returns post_ops previously set by set_post_ops.
void	set_post_ops (post_ops ops)
	Sets post_ops for future use.
void	set_rnn_data_qparams (float scale, float shift)
	Sets quantization scale and shift for RNN data tensors. More...
void	set_rnn_weights_qparams (int mask, const std::vector< float > &scales)
	Sets quantization scales weights_scales for RNN weights tensors. More...
Public Member Functions inherited from dnnl::handle< dnnl_primitive_attr_t >
	handle ()=default
	Empty constructor. More...
	handle (dnnl_primitive_attr_t t, bool weak=false)
	Constructs a C handle wrapper from a C handle. More...
void	reset (dnnl_primitive_attr_t t, bool weak=false)
	Resets the value of a C handle. More...
dnnl_primitive_attr_t	get (bool allow_emtpy=false) const
	Returns the value of the underlying C handle.

Detailed Description

Primitive attributes.

See also

Primitive Attributes

Examples

cnn_inference_int8.cpp, cpu_rnn_inference_int8.cpp, and performance_profiling.cpp.

Constructor & Destructor Documentation

primitive_attr()

dnnl::primitive_attr::primitive_attr ( dnnl_primitive_attr_t  attr )

inline

Creates primitive attributes from a C dnnl_primitive_attr_t handle.

The resulting handle is never weak and the C handle will be destroyed during the destruction of the C++ object.

Member Function Documentation

set_output_scales()

void dnnl::primitive_attr::set_output_scales ( int  mask, const std::vector< float > &  scales  )

inline

Sets output scales for primitive operations.

The correspondence scale mask is stored for future use.

The mask argument defines the correspondence between the output tensor dimensions and the scales vector. Set the i-th bit of mask to 1 to use a dedicated scaling factor for each slice of the output tensor over the i-th dimension. Set mask to 0 to use a common scaling factor for the whole output tensor.

Note

The dimension order is always native and does not depend on the actual layout used. Examples:

• 2D dimensional data the order of dimensions is always: (n, c)
• 4D dimensional data the order is always: (n, c, h, w)
• 5D dimensional weights the order is always: (g, oc, ic, kh, kw)

Examples

cnn_inference_int8.cpp.

set_rnn_data_qparams()

void dnnl::primitive_attr::set_rnn_data_qparams ( float  scale, float  shift  )

inline

Sets quantization scale and shift for RNN data tensors.

For performance reasons, the low-precision configuration of the RNN primitive expects input activations to have the unsigned int8 data type. Scale and shift used to quantize floating-point data to unsigned integer must be passed to the RNN primitive using attributes.

Note

Quantization scale and shift are common for src_layer, src_iter, dst_iter, and dst_layer.

Examples

cpu_rnn_inference_int8.cpp.

set_rnn_weights_qparams()

void dnnl::primitive_attr::set_rnn_weights_qparams ( int  mask, const std::vector< float > &  scales  )

inline

Sets quantization scales weights_scales for RNN weights tensors.

The low-precision configuration of the RNN primitive expects input weights to have the signed int8 data type. Scales used to quantize floating-point data to signed integer must be passed to the RNN primitive using attributes. The mask argument defines correspondence between output tensor dimensions and the weights_scales array. Set the i-th bit of mask to 1 to use a dedicated scaling factor for each slice of the output tensor over the i-th dimension. Set mask to 0 to use a common scaling factor for the whole output tensor.

Note

The dimension order is always native and does not depend on the actual layout used. For example, five-dimensional weights always have (l, d, i, g, o) logical dimension ordering.

Quantization scales are common for weights_layer and weights_iteration

There is no way to check whether count corresponds to mask until an actual primitive descriptor is created, so it is the user's responsibility to set proper values. The following formula must hold:

\[count = \prod\limits_{d \in mask} output.dims[d]\]

Examples

cpu_rnn_inference_int8.cpp.

---

The documentation for this struct was generated from the following file:

• include/dnnl.hpp