Memory

Classes
struct	dnnl_blocking_desc_t
	Generic description of blocked data layout for most memory formats. More...
struct	dnnl_wino_desc_t
	Description of tensor of weights for winograd 2x3 convolution. More...
struct	dnnl_rnn_packed_desc_t
	Description of tensor of packed weights for rnn. More...
struct	dnnl_memory_extra_desc_t
	Description of extra information stored in memory. More...
struct	dnnl_memory_desc_t
	Memory descriptor. More...
struct	dnnl_memory
	An opaque structure to describe a memory. More...

Macros
#define	DNNL_MAX_NDIMS   12
	Maximum number of dimensions a tensor can have. More...
#define	DNNL_RNN_MAX_N_PARTS   4
	Maximum number of parts of RNN weights tensor that require separate computation. More...

Typedefs
typedef int64_t	dnnl_dim_t
	A type to describe tensor dimension.
typedef dnnl_dim_t	dnnl_dims_t[DNNL_MAX_NDIMS]
	A type to describe tensor dimensions.
typedef struct dnnl_memory *	dnnl_memory_t
	A memory handle.
typedef const struct dnnl_memory *	const_dnnl_memory_t
	A constant memory handle.

Enumerations
enum	dnnl_wino_memory_format_t {   dnnl_wino_undef = 0, dnnl_wino_wei_aaOIoi, dnnl_wino_wei_aaOio, dnnl_wino_wei_aaOBiOo,   dnnl_wino_wei_OBaaIBOIio }
	Winograd-specific formats. More...
enum	dnnl_memory_extra_flags_t { , dnnl_memory_extra_flag_compensation_conv_s8s8 = 0x1U }
	Flags for memory special features. More...

Detailed Description

Macro Definition Documentation

DNNL_MAX_NDIMS

#define DNNL_MAX_NDIMS   12

Maximum number of dimensions a tensor can have.

Only restricts the amount of space used for the tensor description. Individual computational primitives may support only tensors of certain dimensions.

DNNL_RNN_MAX_N_PARTS

#define DNNL_RNN_MAX_N_PARTS   4

Maximum number of parts of RNN weights tensor that require separate computation.

Enumeration Type Documentation

dnnl_wino_memory_format_t

enum dnnl_wino_memory_format_t

Winograd-specific formats.

Enumerator
dnnl_wino_undef	Undefined memory format, used for empty memory descriptors.
dnnl_wino_wei_aaOIoi	Internal weights format for 2x3 Winograd.
dnnl_wino_wei_aaOio	Internal weights format for 2x3 Winograd.
dnnl_wino_wei_aaOBiOo	Internal weights format for 2x3 Winograd.
dnnl_wino_wei_OBaaIBOIio	Internal weights format for 4x3 Winograd.

dnnl_memory_extra_flags_t

enum dnnl_memory_extra_flags_t

Flags for memory special features.

Enumerator
dnnl_memory_extra_flag_compensation_conv_s8s8	Indicates the weights have an additional buffer, that depends on the compensation_mask. For instance, in 4D case with the compensation mask equals (1 << 0) the additional buffer would consist of OC values: O[oc : 0,OC] = -128 * SUM(ic : 0,IC; kh : 0,KH; kw : 0,KW){ weights(oc, ic, kh, kw) }