Memory Management

Collaboration diagram for Memory Management:

Classes
struct	cldnn_tensor
	N-dimensional vector. Mostly used to represent memory size. More...
struct	cldnn_padding
	Padding information. More...
struct	cldnn_layout
	Memory layout description. More...

Macros
#define	CLDNN_FLOAT_TYPE_MASK   0x80
#define	CLDNN_UINT_TYPE_MASK   0x40
#define	CLDNN_TENSOR_BATCH_DIM_MAX   1
#define	CLDNN_TENSOR_FEATURE_DIM_MAX   1
#define	CLDNN_TENSOR_SPATIAL_DIM_MAX   2
#define	CLDNN_TENSOR_DIM_MAX   8

Typedefs
typedef struct cldnn_memory_impl *	cldnn_memory
	Memory object.

Enumerations
enum	cldnn_format_type {   cldnn_format_yxfb, cldnn_format_byxf, cldnn_format_bfyx, cldnn_format_fyxb,   cldnn_format_os_iyx_osv16, cldnn_format_bs_xs_xsv8_bsv8, cldnn_format_bs_xs_xsv8_bsv16, cldnn_format_bs_x_bsv16,   cldnn_format_bf8_xy16, cldnn_format_image_2d_weights_c4_fyx_b, cldnn_format_image_2d_weights_c1_b_fyx, cldnn_format_format_num,   cldnn_format_any = -1 }
	Represents memory formats (orders). In CNN most of data is describe as 4 dimensional blocks. In Intel(R) clDNN library we describe memory with 4 letters. More...
enum	cldnn_data_type { cldnn_i8 = sizeof(int8_t), cldnn_f16 = sizeof(int16_t) | CLDNN_FLOAT_TYPE_MASK, cldnn_f32 = sizeof(float) | CLDNN_FLOAT_TYPE_MASK, cldnn_u8 = sizeof(uint8_t) | CLDNN_UINT_TYPE_MASK }
	Data type stored in memory.

Functions
CLDNN_API cldnn_memory	cldnn_allocate_memory (cldnn_engine engine, cldnn_layout layout, cldnn_status *status)
	Allocate memory on engine using specified layout.
CLDNN_API cldnn_memory	cldnn_attach_memory (cldnn_layout layout, void *pointer, size_t size, cldnn_status *status)
	Create memory object attached to the buffer allocated by user. More...
CLDNN_API int32_t	cldnn_is_the_same_buffer (cldnn_memory mem1, cldnn_memory mem2, cldnn_status *status)
	Checks if two memory objects refer to the same underlaying buffer.
CLDNN_API void	cldnn_retain_memory (cldnn_memory memory, cldnn_status *status)
	Increment reference counter for the memory object.
CLDNN_API void	cldnn_release_memory (cldnn_memory memory, cldnn_status *status)
	Decrement reference counter for the memory object. Deletes object when counter becomes zero.
CLDNN_API void *	cldnn_lock_memory (cldnn_memory memory, cldnn_status *status)
	Locks memory buffer. Provides direct access to memory data. More...
CLDNN_API void	cldnn_unlock_memory (cldnn_memory memory, cldnn_status *status)
	Unlocks memory locked by cldnn_lock_memory(cldnn_memory memory, cldnn_status* status).
CLDNN_API cldnn_layout	cldnn_get_memory_layout (cldnn_memory memory, cldnn_status *status)
	Returns memory layout. More...
CLDNN_API cldnn_engine	cldnn_get_memory_engine (cldnn_memory memory, cldnn_status *status)
	Returns reference to the engine associated with memory object. More...
CLDNN_API uint16_t	cldnn_float_to_half (float, cldnn_status *)
	converts float(32 bit) to half_t(fp16 bit) More...
CLDNN_API float	cldnn_half_to_float (uint16_t, cldnn_status *)
	converts half_t(f16 bit) to float(32 bit) More...

Detailed Description

---

Class Documentation

cldnn_tensor

struct cldnn_tensor

N-dimensional vector. Mostly used to represent memory size.

Definition at line 266 of file cldnn.h.

Class Members
size_t	batch_num
size_t	feature_num
int32_t	sizes[CLDNN_TENSOR_DIM_MAX]
size_t	spatial_num

cldnn_padding

struct cldnn_padding

Padding information.

Definition at line 275 of file cldnn.h.

Collaboration diagram for cldnn_padding:

Class Members
float	filling_value	Filling value for an element of padding. If data type of elements is different than float it is converted to it using round-towards-nearest-even (for floating-point data types) or round-towards-zero (for integral data types).
cldnn_tensor	lower_size	Lower padding sizes. For spatials, it means size of left (X) and top (Y) padding.
cldnn_tensor	upper_size	Upper padding sizes. For spatials, it means size of right (X) and bottom (Y) padding.

cldnn_layout

struct cldnn_layout

Memory layout description.

Definition at line 295 of file cldnn.h.

Collaboration diagram for cldnn_layout:

Class Members
size_t	data_type	data type (cldnn_data_type) stored in memory.
int32_t	format	Memor format (cldnn_format_type)
cldnn_padding	padding	Explicitly added padding to memory buffer.
cldnn_tensor	size	N-dimensional vector describes size (in elements) of memory (excluding padding).

Enumeration Type Documentation

cldnn_format_type

enum cldnn_format_type

Represents memory formats (orders).
In CNN most of data is describe as 4 dimensional blocks. In Intel(R) clDNN library we describe memory with 4 letters.

• b - number of blocks in batch. For weights formats: output features - conv, neurons - inner product
• f - number of feature maps, features or channels. For weights formats: input features - conv, inputs, inner product
• x - spatial, width
• y - spatial, height /n For explanation how each format type is implemented in memory we will use naming shown bellow (b=2,f=3,y=3,x=3):
  

Enumerator
cldnn_format_yxfb	batch first, feature and than spatials
cldnn_format_byxf	used in bitmaps, input from user i.e b images of RGB format
cldnn_format_bfyx	the most common format for activations in clDNN.
cldnn_format_fyxb	format not used inside clDNN, but supported in reorder as extension for user provided formats.
cldnn_format_os_iyx_osv16	format used only for convolution weights: os - output feature maps slice, i - input feature maps, yx - spatials, sv16 - 16 values of single slice.
cldnn_format_bs_xs_xsv8_bsv8	format used only for fully connected weights: bs - batch slice, xs - x slice, bsv8 - 8 values of single slice.
cldnn_format_bs_xs_xsv8_bsv16	format used only for fully connected weights: bs - batch slice, xs - x slice, bsv16 - 16 values of single slice.
cldnn_format_bs_x_bsv16	format used only for fully connected weights fp16 batch=1 : bs - batch slice (responses slice), bsv16 - 16 values of single batch slice, x - flattened plane of (fyx).
cldnn_format_bf8_xy16	format used only for convolution 1x1 input, xy aligned to 16, f aligned to 8
cldnn_format_image_2d_weights_c4_fyx_b	image format for weights, image 2d, 4-channel, width size is f*y*x/4 (4-channels filled with fyx data), height is b
cldnn_format_image_2d_weights_c1_b_fyx	image format for weights, image 2d, single channel, width size is b, height is f*y*x
cldnn_format_format_num	number of format types

Definition at line 233 of file cldnn.h.

Function Documentation

cldnn_attach_memory()

CLDNN_API cldnn_memory cldnn_attach_memory	(	cldnn_layout	layout,
		void *	pointer,
		size_t	size,
		cldnn_status *	status
	)

Create memory object attached to the buffer allocated by user.

Note

User is responsible for buffer deallocation. Buffer lifetime should be bigger than lifetime of the memory object.

cldnn_float_to_half()

CLDNN_API uint16_t cldnn_float_to_half	(	float	,
		cldnn_status *
	)

converts float(32 bit) to half_t(fp16 bit)

Returns

16bit half_t

cldnn_get_memory_engine()

CLDNN_API cldnn_engine cldnn_get_memory_engine	(	cldnn_memory	memory,
		cldnn_status *	status
	)

Returns reference to the engine associated with memory object.

Returns

The engine associated with memory object. Or NULL if memory was attached to user-allocated buffer.

cldnn_get_memory_layout()

CLDNN_API cldnn_layout cldnn_get_memory_layout	(	cldnn_memory	memory,
		cldnn_status *	status
	)

Returns memory layout.

Returns

cldnn_layout which describes memory.

cldnn_half_to_float()

CLDNN_API float cldnn_half_to_float	(	uint16_t	,
		cldnn_status *
	)

converts half_t(f16 bit) to float(32 bit)

Returns

32bit float

cldnn_lock_memory()

CLDNN_API void* cldnn_lock_memory	(	cldnn_memory	memory,
		cldnn_status *	status
	)

Locks memory buffer. Provides direct access to memory data.

Returns

Direct pointer to the memory data.