Memory

Overview

A container that describes and stores data. More…

// typedefs

typedef struct dnnl_memory_desc* dnnl_memory_desc_t;
typedef const struct dnnl_memory_desc* const_dnnl_memory_desc_t;
typedef struct dnnl_memory* dnnl_memory_t;
typedef const struct dnnl_memory* const_dnnl_memory_t;

// enums

enum dnnl_format_kind_t;
enum dnnl_format_tag_t;
enum dnnl_profiling_data_kind_t;
enum dnnl_sparse_encoding_t;

// structs

struct dnnl_memory;
struct dnnl_memory_desc;
struct dnnl::memory;

// global functions

bool dnnl::operator == (dnnl_data_type_t a, memory::data_type b);
bool dnnl::operator != (dnnl_data_type_t a, memory::data_type b);
bool dnnl::operator == (memory::data_type a, dnnl_data_type_t b);
bool dnnl::operator != (memory::data_type a, dnnl_data_type_t b);
bool dnnl::operator == (dnnl_format_tag_t a, memory::format_tag b);
bool dnnl::operator != (dnnl_format_tag_t a, memory::format_tag b);
bool dnnl::operator == (memory::format_tag a, dnnl_format_tag_t b);
bool dnnl::operator != (memory::format_tag a, dnnl_format_tag_t b);
dnnl_status_t DNNL_API dnnl_memory_desc_destroy(dnnl_memory_desc_t memory_desc);

dnnl_status_t DNNL_API dnnl_memory_desc_clone(
    dnnl_memory_desc_t* memory_desc,
    const_dnnl_memory_desc_t existing_memory_desc
    );

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_strides(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    const dnnl_dims_t strides
    );

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_tag(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    dnnl_format_tag_t tag
    );

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_csr_encoding(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    dnnl_dim_t nnz,
    dnnl_data_type_t indices_dt,
    dnnl_data_type_t pointers_dt
    );

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_packed_encoding(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    dnnl_dim_t nnz
    );

dnnl_status_t DNNL_API dnnl_memory_desc_create_submemory(
    dnnl_memory_desc_t* memory_desc,
    const_dnnl_memory_desc_t parent_memory_desc,
    const dnnl_dims_t dims,
    const dnnl_dims_t offsets
    );

dnnl_status_t DNNL_API dnnl_memory_desc_reshape(
    dnnl_memory_desc_t* out_memory_desc,
    const_dnnl_memory_desc_t in_memory_desc,
    int ndims,
    const dnnl_dims_t dims
    );

dnnl_status_t DNNL_API dnnl_memory_desc_permute_axes(
    dnnl_memory_desc_t* out_memory_desc,
    const_dnnl_memory_desc_t in_memory_desc,
    const int* permutation
    );

dnnl_status_t DNNL_API dnnl_memory_desc_query(
    const_dnnl_memory_desc_t memory_desc,
    dnnl_query_t what,
    void* result
    );

dnnl_status_t DNNL_API dnnl_memory_desc_query_v2(
    const_dnnl_memory_desc_t memory_desc,
    dnnl_query_t what,
    int index,
    void* result
    );

int DNNL_API dnnl_memory_desc_equal(
    const_dnnl_memory_desc_t lhs,
    const_dnnl_memory_desc_t rhs
    );

size_t DNNL_API dnnl_memory_desc_get_size(const_dnnl_memory_desc_t memory_desc);

size_t DNNL_API dnnl_memory_desc_get_size_v2(
    const_dnnl_memory_desc_t memory_desc,
    int index
    );

size_t DNNL_API dnnl_data_type_size(dnnl_data_type_t data_type);

dnnl_status_t DNNL_API dnnl_memory_create(
    dnnl_memory_t* memory,
    const_dnnl_memory_desc_t memory_desc,
    dnnl_engine_t engine,
    void* handle
    );

dnnl_status_t DNNL_API dnnl_memory_create_v2(
    dnnl_memory_t* memory,
    const_dnnl_memory_desc_t memory_desc,
    dnnl_engine_t engine,
    int nhandles,
    void** handles
    );

dnnl_status_t DNNL_API dnnl_memory_get_memory_desc(
    const_dnnl_memory_t memory,
    const_dnnl_memory_desc_t* memory_desc
    );

dnnl_status_t DNNL_API dnnl_memory_get_engine(
    const_dnnl_memory_t memory,
    dnnl_engine_t* engine
    );

dnnl_status_t DNNL_API dnnl_memory_map_data(
    const_dnnl_memory_t memory,
    void** mapped_ptr
    );

dnnl_status_t DNNL_API dnnl_memory_map_data_v2(
    const_dnnl_memory_t memory,
    void** mapped_ptr,
    int index
    );

dnnl_status_t DNNL_API dnnl_memory_unmap_data(
    const_dnnl_memory_t memory,
    void* mapped_ptr
    );

dnnl_status_t DNNL_API dnnl_memory_unmap_data_v2(
    const_dnnl_memory_t memory,
    void* mapped_ptr,
    int index
    );

dnnl_status_t DNNL_API dnnl_memory_get_data_handle(
    const_dnnl_memory_t memory,
    void** handle
    );

dnnl_status_t DNNL_API dnnl_memory_set_data_handle(
    dnnl_memory_t memory,
    void* handle
    );

dnnl_status_t DNNL_API dnnl_memory_get_data_handle_v2(
    const_dnnl_memory_t memory,
    void** handle,
    int index
    );

dnnl_status_t DNNL_API dnnl_memory_set_data_handle_v2(
    dnnl_memory_t memory,
    void* handle,
    int index
    );

dnnl_status_t DNNL_API dnnl_memory_destroy(dnnl_memory_t memory);

// macros

#define DNNL_MEMORY_ALLOCATE
#define DNNL_MEMORY_NONE
#define DNNL_RUNTIME_DIM_VAL
#define DNNL_RUNTIME_F32_VAL
#define DNNL_RUNTIME_S32_VAL
#define DNNL_RUNTIME_SIZE_VAL

Detailed Documentation

A container that describes and stores data.

Memory objects can contain data of various types and formats. There are two levels of abstraction:

  1. Memory descriptor engine-agnostic logical description of data (number of dimensions, dimension sizes, and data type), and, optionally, the information about the physical format of data in memory. If this information is not known yet, a memory descriptor can be created with dnnl::memory::format_tag::any. This allows compute-intensive primitives to choose the best format for computation. The user is responsible for reordering the data into the chosen format when formats do not match.

    A memory descriptor can be initialized either by specifying dimensions and a memory format tag or strides for each of them, or by manipulating the dnnl_memory_desc_t structure directly.

    Warning

    The latter approach requires understanding how the physical data representation is mapped to the structure and is discouraged. This topic is discussed in Understanding Memory Formats.

    The user can query the amount of memory required by a memory descriptor using the dnnl::memory::desc::get_size() function. The size of data in general cannot be computed as the product of dimensions multiplied by the size of the data type. So users are required to use this function for better code portability.

    Two memory descriptors can be compared using the equality and inequality operators. The comparison is especially useful when checking whether it is necessary to reorder data from the user’s data format to a primitive’s format.

  2. Memory object an engine-specific object that handles the memory buffer and its description (a memory descriptor). For the CPU engine or with USM, the memory buffer handle is simply a pointer to void. The memory buffer can be queried using dnnl::memory::get_data_handle() and set using dnnl::memory::set_data_handle(). The underlying SYCL buffer, when used, can be queried using dnnl::sycl_interop::get_buffer and set using dnnl::sycl_interop::set_buffer. A memory object can also be queried for the underlying memory descriptor and for its engine using dnnl::memory::get_desc() and dnnl::memory::get_engine().

Along with ordinary memory descriptors with all dimensions being positive, the library supports zero-volume memory descriptors with one or more dimensions set to zero. This is used to support the NumPy* convention. If a zero-volume memory is passed to a primitive, the primitive typically does not perform any computations with this memory. For example:

  • A concatenation primitive would ignore all memory object with zeroes in the concat dimension / axis.

  • A forward convolution with a source memory object with zero in the minibatch dimension would always produce a destination memory object with a zero in the minibatch dimension and perform no computations.

  • However, a forward convolution with a zero in one of the weights dimensions is ill-defined and is considered to be an error by the library because there is no clear definition of what the output values should be.

Memory buffer of a zero-volume memory is never accessed.

Typedefs

typedef struct dnnl_memory_desc* dnnl_memory_desc_t

A memory descriptor handle.

typedef const struct dnnl_memory_desc* const_dnnl_memory_desc_t

A memory descriptor handle.

typedef struct dnnl_memory* dnnl_memory_t

A memory handle.

typedef const struct dnnl_memory* const_dnnl_memory_t

A constant memory handle.

Global Functions

dnnl_status_t DNNL_API dnnl_memory_desc_destroy(dnnl_memory_desc_t memory_desc)

Destroys a memory descriptor.

Parameters:

memory_desc

Memory descriptor to destroy.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_clone(
    dnnl_memory_desc_t* memory_desc,
    const_dnnl_memory_desc_t existing_memory_desc
    )

Clones a memory descriptor.

The resulting memory descriptor must be destroyed separately.

Parameters:

memory_desc

Output memory descriptor.

existing_memory_desc

Memory descriptor to clone.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_strides(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    const dnnl_dims_t strides
    )

Creates a memory descriptor using dimensions and strides.

Note

As always, the logical order of dimensions corresponds to the abc... format tag, and the physical meaning of the dimensions depends on both the primitive that consumes the memory and the context of that consumption.

Parameters:

memory_desc

Output memory descriptor.

ndims

Number of dimensions

dims

Array of dimensions.

data_type

Elements data type.

strides

Strides in each dimension.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_tag(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    dnnl_format_tag_t tag
    )

Creates a memory descriptor using dimensions and memory format tag.

Note

As always, the logical order of dimensions corresponds to the abc... format tag, and the physical meaning of the dimensions depends on both the primitive that consumes the memory and the context of that consumption.

Parameters:

memory_desc

Output memory descriptor.

ndims

Number of dimensions

dims

Array of dimensions.

data_type

Elements data type.

tag

Memory format tag. Can be dnnl_format_tag_any which would allow a primitive to chose the final memory format. In this case the format_kind field of the memory descriptor would be set to dnnl_format_kind_any.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_csr_encoding(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    dnnl_dim_t nnz,
    dnnl_data_type_t indices_dt,
    dnnl_data_type_t pointers_dt
    )

Creates a memory descriptor for CSR encoding.

Parameters:

memory_desc

Output memory descriptor.

ndims

Number of dimensions

dims

Array of dimensions.

data_type

Elements data type.

nnz

Number of non-zero entries.

indices_dt

Data type of indices.

pointers_dt

Data type of pointers.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_create_with_packed_encoding(
    dnnl_memory_desc_t* memory_desc,
    int ndims,
    const dnnl_dims_t dims,
    dnnl_data_type_t data_type,
    dnnl_dim_t nnz
    )

Creates a memory descriptor for packed sparse encoding.

The created memory descriptor cannot be used to create a memory object. It can only be used to create a primitive descriptor to query the actual memory descriptor (similar to the format tag any).

Warning

The meaning and content of the handles of the memory object that is created using the queried memory descriptor are unspecified therefore using the content is an undefined behavior.

Parameters:

memory_desc

Output memory descriptor.

ndims

Number of dimensions

dims

Array of dimensions.

data_type

Elements data type.

nnz

Number of non-zero entries.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_create_submemory(
    dnnl_memory_desc_t* memory_desc,
    const_dnnl_memory_desc_t parent_memory_desc,
    const dnnl_dims_t dims,
    const dnnl_dims_t offsets
    )

Creates a memory descriptor for a region inside an area described by an existing memory descriptor.

Warning

Some combinations of physical memory layout and/or offsets or dims may result in a failure to create a submemory.

Parameters:

memory_desc

Output memory descriptor.

parent_memory_desc

An existing memory descriptor.

dims

Sizes of the region.

offsets

Offsets to the region from the encompassing memory object in each dimension

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_reshape(
    dnnl_memory_desc_t* out_memory_desc,
    const_dnnl_memory_desc_t in_memory_desc,
    int ndims,
    const dnnl_dims_t dims
    )

Creates a memory descriptor by reshaping an existing one.

The new memory descriptor inherits the data type. This operation is valid only for memory descriptors that have format_kind dnnl_blocked or dnnl_format_kind_any.

The resulting memory descriptor must be destroyed separately.

The operation ensures the transformation of the physical memory format corresponds to the transformation of the logical dimensions. If such transformation is impossible, the function returns dnnl_invalid_arguments.

The reshape operation can be described as a combination of the following basic operations:

  1. Add a dimension of size 1. This is always possible.

  2. Remove a dimension of size 1. This is possible only if the dimension has no padding (i.e. padded_dims[dim] == dims[dim] && dims[dim] == 1).

  3. Split a dimension into multiple ones. This is possible only if the size of the dimension is exactly equal to the product of the split ones and the dimension does not have padding (i.e. padded_dims[dim] = dims[dim]).

  4. Joining multiple consecutive dimensions into a single one. As in the cases above, this requires that the dimensions do not have padding and that the memory format is such that in physical memory these dimensions are dense and have the same order as their logical counterparts. This also assumes that these dimensions are not blocked.

    • Here, dense means: stride for dim[i] == (stride for dim[i + 1]) * dim[i + 1];

    • And same order means: i < j if and only if stride for dim[j] <= stride for dim[i].

Warning

Some combinations of physical memory layout and/or offsets or dimensions may result in a failure to make a reshape.

Parameters:

out_memory_desc

Output memory descriptor.

in_memory_desc

An existing memory descriptor. Must have format_kind set to dnnl_blocked or dnnl_format_kind_any.

ndims

Number of dimensions for the output memory descriptor.

dims

Dimensions for the output memory descriptor.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_permute_axes(
    dnnl_memory_desc_t* out_memory_desc,
    const_dnnl_memory_desc_t in_memory_desc,
    const int* permutation
    )

Creates a memory descriptor by permuting axes in an existing one.

The physical memory layout representation is adjusted accordingly to maintain the consistency between the logical and physical parts of the memory descriptor.

The resulting memory descriptor must be destroyed separately.

The new memory descriptor inherits the data type. This operation is valid only for memory descriptors that have format_kind set to dnnl_blocked or dnnl_format_kind_any.

The logical axes will be permuted in the following manner:

for (i: 0 .. in_memory_desc->ndims)
    out_memory_desc->dims[permutation[i]] = in_memory_desc->dims[i];

Example:

dnnl_memory_desc_t in_md, out_md, expect_out_md;

const int permutation[] = {1, 0}; // swap the first and the second axes

dnnl_dims_t in_dims = {2, 3}, out_dims = {3, 2};
dnnl_format_tag_t in_tag = dnnl_ab, out_tag = dnnl_ba;

dnnl_memory_desc_create_with_tag(
        &in_md, 2, in_dims, data_type, in_tag);
dnnl_memory_desc_create_with_tag(
        &expect_out_md, 2, out_dims, data_type, out_tag);

dnnl_memory_desc_permute_axes(&out_md, in_md, permutation);
assert(dnnl_memory_desc_equal(out_md, expect_out_md));

dnnl_memory_desc_destroy(in_md);
dnnl_memory_desc_destroy(out_md);
dnnl_memory_desc_destroy(expect_out_md);

Parameters:

out_memory_desc

Output memory descriptor.

in_memory_desc

An existing memory descriptor. Must have format_kind set to dnnl_blocked or dnnl_format_kind_any.

permutation

Axes permutation (of size in_memory_desc->ndims).

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_query(
    const_dnnl_memory_desc_t memory_desc,
    dnnl_query_t what,
    void* result
    )

Queries a memory descriptor for various pieces of information.

The following information can be queried:

  • Number of dimensions (dnnl_query_ndims_s32)

  • Dimensions (dnnl_query_dims) in the following order:

    • CNN data tensors: mini-batch, channel, spatial ({N, C, [[D,] H,] W})

    • CNN weight tensors: group (optional), output channel, input channel, spatial ({[G,] O, I, [[D,] H,] W})

    • RNN data tensors: time, mini-batch, channels ({T, N, C}) or layers, directions, states, mini-batch, channels ({L, D, S, N, C})

    • RNN weight tensor: layers, directions, input channel, gates, output channels ({L, D, I, G, O})

  • Data type of the tensor elements (dnnl_query_data_type)

  • Padded dimensions (dnnl_query_padded_dims) - size of the data including padding in each dimension

  • Padded offsets (dnnl_query_padded_offsets) - per-dimension offset from the padding to actual data, the top-level tensor with offsets applied must lie within the padding area.

  • Submemory offset (dnnl_query_submemory_offset_s64) - offset from memory origin to the current block, non-zero only in a description of a memory sub-block.

  • Format kind (dnnl_query_format_kind) - memory format kind

Note

The order of dimensions does not depend on the memory format, so whether the data is laid out in dnnl_nchw or dnnl_nhwc the dims for 4D CN data tensor would be {N, C, H, W}.

The following queries are applicable only to format kind dnnl_blocked.

  • Strides (dnnl_query_strides) between the outermost blocks or in case of plain (non-blocked) formats the strides between dimensions

  • Number of innermost blocks (dnnl_query_inner_nblks_s32), e.g. {4, 16, 4} in case of OIhw_4i16o4i

  • Size of the innermost blocks (dnnl_query_inner_blks), e.g. 3 in case of OIhw_4i16o4i_

  • Logical indices of the blocks (dnnl_query_inner_idxs), e.g. {1, 0, 1} in case of 4i16o4i, because i is the 1st dim and o is the 0st dim

Parameters:

memory_desc

Memory descriptor.

what

Parameter to query.

result

Output result. The type depends on the query. For example, it must be a dnnl_dims_t** if querying for a strides.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_desc_query_v2(
    const_dnnl_memory_desc_t memory_desc,
    dnnl_query_t what,
    int index,
    void* result
    )

Queries a memory descriptor for various pieces of information.

This version support additional queries dnnl_query_sparse_encoding, dnnl_query_nnz_s64 dnnl_query_num_handles_s32 and dnnl_query_data_type for a particular buffer.

The following information can be queried:

  • Number of dimensions (dnnl_query_ndims_s32)

  • Dimensions (dnnl_query_dims) in the following order:

    • CNN data tensors: mini-batch, channel, spatial ({N, C, [[D,] H,] W})

    • CNN weight tensors: group (optional), output channel, input channel, spatial ({[G,] O, I, [[D,] H,] W})

    • RNN data tensors: time, mini-batch, channels ({T, N, C}) or layers, directions, states, mini-batch, channels ({L, D, S, N, C})

    • RNN weight tensor: layers, directions, input channel, gates, output channels ({L, D, I, G, O})

  • Data type of the tensor elements (dnnl_query_data_type)

  • Padded dimensions (dnnl_query_padded_dims) - size of the data including padding in each dimension

  • Padded offsets (dnnl_query_padded_offsets) - per-dimension offset from the padding to actual data, the top-level tensor with offsets applied must lie within the padding area.

  • Submemory offset (dnnl_query_submemory_offset_s64) - offset from memory origin to the current block, non-zero only in a description of a memory sub-block.

  • Format kind (dnnl_query_format_kind) - memory format kind

Note

The order of dimensions does not depend on the memory format, so whether the data is laid out in dnnl_nchw or dnnl_nhwc the dims for 4D CN data tensor would be {N, C, H, W}.

The following queries are applicable only to format kind dnnl_blocked.

  • Strides (dnnl_query_strides) between the outermost blocks or in case of plain (non-blocked) formats the strides between dimensions

  • Number of innermost blocks (dnnl_query_inner_nblks_s32), e.g. {4, 16, 4} in case of OIhw_4i16o4i

  • Size of the innermost blocks (dnnl_query_inner_blks), e.g. 3 in case of OIhw_4i16o4i_

  • Logical indices of the blocks (dnnl_query_inner_idxs), e.g. {1, 0, 1} in case of 4i16o4i, because i is the 1st dim and o is the 0st dim

Parameters:

memory_desc

Memory descriptor.

what

Parameter to query.

index

Index of the parameter to query for. It is mostly used with dnnl_query_data_type to specify which data type is being queried. The main data type (data type of values) has always index 0. For other indices please refer to the API for creating a memory descriptor for sparse encoding.

result

Output result. The type depends on the query. For example, it must be a dnnl_dims_t** if querying for a strides.

Returns:

dnnl_success on success and a status describing the error otherwise.

int DNNL_API dnnl_memory_desc_equal(
    const_dnnl_memory_desc_t lhs,
    const_dnnl_memory_desc_t rhs
    )

Compares two memory descriptors.

Use this function to identify whether a reorder is required between the two memories

Parameters:

lhs

Left-hand side of the comparison.

rhs

Right-hand side of the comparison.

Returns:

1 if the descriptors are the same.

0 if the descriptors are different.

size_t DNNL_API dnnl_memory_desc_get_size(const_dnnl_memory_desc_t memory_desc)

Returns the size of a memory descriptor.

Parameters:

memory_desc

Memory descriptor.

Returns:

The number of bytes required for memory described by a memory descriptor.

size_t DNNL_API dnnl_memory_desc_get_size_v2(
    const_dnnl_memory_desc_t memory_desc,
    int index
    )

Returns the size of the data that corresponds to the given index.

Parameters:

memory_desc

Memory descriptor.

index

Index of the buffer.

Returns:

The number of bytes required for the requested data.

size_t DNNL_API dnnl_data_type_size(dnnl_data_type_t data_type)

Returns the size of data type.

Parameters:

data_type

Data type.

Returns:

The number of bytes occupied by data type.

dnnl_status_t DNNL_API dnnl_memory_create(
    dnnl_memory_t* memory,
    const_dnnl_memory_desc_t memory_desc,
    dnnl_engine_t engine,
    void* handle
    )

Creates a memory object.

Unless handle is equal to DNNL_MEMORY_NONE, the constructed memory object will have the underlying buffer set. In this case, the buffer will be initialized as if dnnl_memory_set_data_handle() had been called.

Parameters:

memory

Output memory object.

memory_desc

Memory descriptor.

engine

Engine to use.

handle

Handle of the memory buffer to use as an underlying storage.

  • A pointer to the user-allocated buffer. In this case the library doesn’t own the buffer.

  • The DNNL_MEMORY_ALLOCATE special value. Instructs the library to allocate the buffer for the memory object. In this case the library owns the buffer.

  • DNNL_MEMORY_NONE to create dnnl_memory without an underlying buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

See also:

dnnl_memory_set_data_handle()

dnnl_status_t DNNL_API dnnl_memory_create_v2(
    dnnl_memory_t* memory,
    const_dnnl_memory_desc_t memory_desc,
    dnnl_engine_t engine,
    int nhandles,
    void** handles
    )

Creates a memory object with multiple handles.

Parameters:

memory

Output memory object.

memory_desc

Memory descriptor.

engine

Engine to use.

nhandles

Number of handles.

handles

Handles of the memory buffers to use as underlying storages. For each element of the handles array the following applies:

  • A pointer to the user-allocated buffer. In this case the library doesn’t own the buffer.

  • The DNNL_MEMORY_ALLOCATE special value. Instructs the library to allocate the buffer for the memory object. In this case the library owns the buffer.

  • DNNL_MEMORY_NONE Instructs the library to skip allocation of the memory buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_get_memory_desc(
    const_dnnl_memory_t memory,
    const_dnnl_memory_desc_t* memory_desc
    )

Returns the memory descriptor for a memory object.

Parameters:

memory

Memory object.

memory_desc

Output memory descriptor (a copy).

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_get_engine(
    const_dnnl_memory_t memory,
    dnnl_engine_t* engine
    )

Returns the engine of a memory object.

Parameters:

memory

Memory object.

engine

Output engine on which the memory is located.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_map_data(
    const_dnnl_memory_t memory,
    void** mapped_ptr
    )

Maps a memory object and returns a host-side pointer to a memory buffer with a copy of its contents.

Mapping enables explicit direct access to memory contents for the engines that do not support it implicitly.

Mapping is an exclusive operation - a memory object cannot be used in other operations until this memory object is unmapped.

Note

Any primitives working with memory should be completed before the memory is mapped. Use dnnl_stream_wait to synchronize the corresponding execution stream.

Note

The dnnl_memory_map_data() and dnnl_memory_unmap_data() functions are mainly provided for debug and testing purposes, and their performance may be suboptimal.

Parameters:

memory

Memory object.

mapped_ptr

Output pointer to the mapped buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_map_data_v2(
    const_dnnl_memory_t memory,
    void** mapped_ptr,
    int index
    )

Maps a memory object and returns a host-side pointer to a memory buffer with a copy of its contents.

The memory buffer corresponds to the given index.

Mapping enables explicit direct access to memory contents for the engines that do not support it implicitly.

Mapping is an exclusive operation - a memory object cannot be used in other operations until this memory object is unmapped.

Note

Any primitives working with memory should be completed before the memory is mapped. Use dnnl_stream_wait to synchronize the corresponding execution stream.

Note

The dnnl_memory_map_data() and dnnl_memory_unmap_data() functions are mainly provided for debug and testing purposes, and their performance may be suboptimal.

Parameters:

memory

Memory object.

mapped_ptr

Output pointer to the mapped buffer.

index

Index of the buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_unmap_data(
    const_dnnl_memory_t memory,
    void* mapped_ptr
    )

Unmaps a memory object and writes back any changes made to the previously mapped memory buffer.

The pointer to the mapped buffer must be obtained via the dnnl_memory_map_data() call.

Note

The dnnl_memory_map_data() and dnnl_memory_unmap_data() functions are mainly provided for debug and testing purposes, and their performance may be suboptimal.

Parameters:

memory

Memory object.

mapped_ptr

Pointer to the mapped buffer that must have been obtained using the dnnl_memory_map_data() function.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_unmap_data_v2(
    const_dnnl_memory_t memory,
    void* mapped_ptr,
    int index
    )

Unmaps a memory object and writes back any changes made to the previously mapped memory buffer.

The pointer to the mapped buffer must be obtained via the dnnl_memory_map_data() call. The buffer corresponds to the given index.

Note

The dnnl_memory_map_data() and dnnl_memory_unmap_data() functions are mainly provided for debug and testing purposes, and their performance may be suboptimal.

Parameters:

memory

Memory object.

mapped_ptr

Pointer to the mapped buffer that must have been obtained using the dnnl_memory_map_data() function.

index

Index of the buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_get_data_handle(
    const_dnnl_memory_t memory,
    void** handle
    )

Returns memory object’s data handle.

Parameters:

memory

Memory object.

handle

Output data handle. For the CPU engine, the data handle is a pointer to the actual data. For OpenCL it is a cl_mem.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_set_data_handle(
    dnnl_memory_t memory,
    void* handle
    )

Sets the underlying memory buffer.

Parameters:

memory

Memory object.

handle

Data handle. For the CPU engine or when USM is used, the memory buffer is a pointer to the actual data. For OpenCL it is a cl_mem.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_get_data_handle_v2(
    const_dnnl_memory_t memory,
    void** handle,
    int index
    )

Returns an underlying memory buffer that corresponds to the given index.

Parameters:

memory

Memory object.

handle

Data handle. For the CPU engine or when USM is used, the memory buffer is a pointer to the actual data. For OpenCL it is a cl_mem.

index

Index of the buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_set_data_handle_v2(
    dnnl_memory_t memory,
    void* handle,
    int index
    )

Sets an underlying memory buffer that corresponds to the given index.

Parameters:

memory

Memory object.

handle

Data handle. For the CPU engine or when USM is used, the memory buffer is a pointer to the actual data. For OpenCL it is a cl_mem.

index

Index of the buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_memory_destroy(dnnl_memory_t memory)

Destroys a memory object.

Parameters:

memory

Memory object to destroy.

Returns:

dnnl_success on success and a status describing the error otherwise.

Macros

#define DNNL_MEMORY_ALLOCATE

Special pointer value that indicates that the library needs to allocate an underlying buffer for a memory object.

#define DNNL_MEMORY_NONE

Special pointer value that indicates that a memory object should not have an underlying buffer.

#define DNNL_RUNTIME_DIM_VAL

A wildcard value for dimensions that are unknown at a primitive creation time.

#define DNNL_RUNTIME_F32_VAL

A wildcard value for floating point values that are unknown at a primitive creation time.

#define DNNL_RUNTIME_S32_VAL

A wildcard value for int32_t values that are unknown at a primitive creation time.

#define DNNL_RUNTIME_SIZE_VAL

A size_t counterpart of the DNNL_RUNTIME_DIM_VAL.

For instance, this value is returned by dnnl_memory_desc_get_size() if either of the dimensions or strides equal to DNNL_RUNTIME_DIM_VAL.