BRGeMM ukernel¶

Overview¶

BRGeMM ukernel routines. More…

// typedefs

typedef struct dnnl_brgemm* dnnl_brgemm_t;
typedef const struct dnnl_brgemm* const_dnnl_brgemm_t;
typedef struct dnnl_brgemm_pack_b* dnnl_brgemm_pack_b_t;
typedef const struct dnnl_brgemm_pack_b* const_dnnl_brgemm_pack_b_t;

// structs

struct dnnl::ukernel::brgemm;
struct dnnl::ukernel::brgemm_pack_b;
struct dnnl_brgemm;
struct dnnl_brgemm_pack_b;

// global functions

dnnl_status_t DNNL_API dnnl_brgemm_create(
    dnnl_brgemm_t* brgemm,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    dnnl_dim_t batch_size,
    dnnl_dim_t lda,
    dnnl_dim_t ldb,
    dnnl_dim_t ldc,
    dnnl_dim_t ldd,
    dnnl_data_type_t a_dt,
    dnnl_data_type_t b_dt,
    dnnl_data_type_t c_dt,
    dnnl_data_type_t d_dt,
    float alpha,
    float beta,
    const_dnnl_primitive_attr_t attr
    );

dnnl_status_t DNNL_API dnnl_brgemm_get_scratchpad_size(
    const_dnnl_brgemm_t brgemm,
    size_t* size
    );

dnnl_status_t DNNL_API dnnl_brgemm_set_hw_context(const_dnnl_brgemm_t brgemm);
dnnl_status_t DNNL_API dnnl_brgemm_release_hw_context();
dnnl_status_t DNNL_API dnnl_brgemm_generate(dnnl_brgemm_t brgemm);

dnnl_status_t DNNL_API dnnl_brgemm_execute(
    const_dnnl_brgemm_t brgemm,
    const void* A_ptr,
    const void* B_ptr,
    const dnnl_dim_t* A_B_offsets,
    void* C_ptr,
    void* scratchpad_ptr
    );

dnnl_status_t DNNL_API dnnl_brgemm_execute_postops(
    const_dnnl_brgemm_t brgemm,
    const void* A,
    const void* B,
    const dnnl_dim_t* A_B_offsets,
    const void* C_ptr,
    void* D_ptr,
    void* scratchpad_ptr,
    const void* binary_po_ptr
    );

dnnl_status_t DNNL_API dnnl_brgemm_destroy(dnnl_brgemm_t brgemm);

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_create(
    dnnl_brgemm_pack_b_t* brgemm_pack_b,
    dnnl_dim_t K,
    dnnl_dim_t N,
    dnnl_dim_t in_ld,
    dnnl_dim_t out_ld,
    dnnl_data_type_t in_dt,
    dnnl_data_type_t out_dt
    );

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_need_pack(
    const_dnnl_brgemm_pack_b_t brgemm_pack_b,
    int* need_pack
    );

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_generate(dnnl_brgemm_pack_b_t brgemm_pack_b);

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_execute(
    const_dnnl_brgemm_pack_b_t brgemm_pack_b,
    const void* in_ptr,
    void* out_ptr
    );

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_destroy(dnnl_brgemm_pack_b_t brgemm_pack_b);

Detailed Documentation¶

BRGeMM ukernel routines.

Typedefs¶

typedef struct dnnl_brgemm* dnnl_brgemm_t

A brgemm ukernel handle.

typedef const struct dnnl_brgemm* const_dnnl_brgemm_t

A constant brgemm ukernel handle.

typedef struct dnnl_brgemm_pack_b* dnnl_brgemm_pack_b_t

A brgemm ukernel packing B routine handle.

typedef const struct dnnl_brgemm_pack_b* const_dnnl_brgemm_pack_b_t

A constant brgemm ukernel packing B routine handle.

Global Functions¶

dnnl_status_t DNNL_API dnnl_brgemm_create(
    dnnl_brgemm_t* brgemm,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    dnnl_dim_t batch_size,
    dnnl_dim_t lda,
    dnnl_dim_t ldb,
    dnnl_dim_t ldc,
    dnnl_dim_t ldd,
    dnnl_data_type_t a_dt,
    dnnl_data_type_t b_dt,
    dnnl_data_type_t c_dt,
    dnnl_data_type_t d_dt,
    float alpha,
    float beta,
    const_dnnl_primitive_attr_t attr
    )

Creates a BRGeMM ukernel object.

Operates by the following formula: C = alpha * [A x B] + beta * C. D = post-operations(C).

Post-operations applies if one of the following holds:

Non-empty attributes are specified.
Output data type d_dt is different from accumulation data type c_dt.

If any of conditions happens, the final call of the accumulation chain must be dnnl_brgemm_execute_postops, and dnnl_brgemm_execute, otherwise.

Parameters:

brgemm	Output BRGeMM ukernel object.
M	Dimension M of tensor A.
N	Dimension N of tensor B.
K	Dimension K of tensors A and B.
batch_size	Number of batches to process.
lda	Leading dimension of tensor A.
ldb	Leading dimension of tensor B.
ldc	Leading dimension of tensor C.
ldd	Leading dimension of tensor D.
a_dt	Data type of tensor A.
b_dt	Data type of tensor B.
c_dt	Data type of tensor C. Must be dnnl_f32.
d_dt	Data type of tensor D.
alpha	Scale for an accumulation output.
beta	Scale for a tensor C to append on an accumulation output.
attr	Primitive attributes to extend the kernel operations.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_get_scratchpad_size(
    const_dnnl_brgemm_t brgemm,
    size_t* size
    )

Returns the size of a scratchpad memory needed for the BRGeMM ukernel object.

Parameters:

brgemm	BRGeMM ukernel object.
size	Output size of a buffer required for the BRGeMM ukernel object.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_set_hw_context(const_dnnl_brgemm_t brgemm)

Initializes the hardware-specific context.

If no initialization required, returns the success status.

Parameters:

brgemm

BRGeMM ukernel object.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_release_hw_context()

Releases the hardware-specific context.

Must be used after all the execution calls to BRGeMM ukernel objects.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_generate(dnnl_brgemm_t brgemm)

Generates an executable part of BRGeMM ukernel object.

Parameters:

brgemm

BRGeMM ukernel object.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_execute(
    const_dnnl_brgemm_t brgemm,
    const void* A_ptr,
    const void* B_ptr,
    const dnnl_dim_t* A_B_offsets,
    void* C_ptr,
    void* scratchpad_ptr
    )

Executes a BRGeMM ukernel object.

Parameters:

brgemm	BRGeMM ukernel object.
A_ptr	Base pointer to a tensor A.
B_ptr	Base pointer to a tensor B.
A_B_offsets	Pointer to the set of tensor A and tensor B offsets for each batch; the set must be contiguous in memory. Single batch should supply offsets for both tensors A and B simultaneously. The number of batches must coincide with the `batch_size` value passed at the creation stage.
C_ptr	Pointer to a tensor C (accumulation buffer).
scratchpad_ptr	Pointer to a scratchpad buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_execute_postops(
    const_dnnl_brgemm_t brgemm,
    const void* A,
    const void* B,
    const dnnl_dim_t* A_B_offsets,
    const void* C_ptr,
    void* D_ptr,
    void* scratchpad_ptr,
    const void* binary_po_ptr
    )

Executes a BRGeMM ukernel object with post operations.

Parameters:

brgemm	BRGeMM ukernel object.
A	Base pointer to a tensor A.
B	Base pointer to a tensor B.
A_B_offsets	Pointer to a set of tensor A and tensor B offsets for each batch. A set must be contiguous in memory. A single batch should supply offsets for both tensors A and B simultaneously. The number of batches must coincide with the `batch_size` value passed at the creation stage.
C_ptr	Pointer to a tensor C (accumulation buffer).
D_ptr	Pointer to a tensor D (output buffer).
scratchpad_ptr	Pointer to a scratchpad buffer.
binary_po_ptr	Pointer to binary post-op data.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_destroy(dnnl_brgemm_t brgemm)

Destroys a BRGeMM ukernel object.

Parameters:

brgemm

BRGeMM ukernel object to destroy.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_create(
    dnnl_brgemm_pack_b_t* brgemm_pack_b,
    dnnl_dim_t K,
    dnnl_dim_t N,
    dnnl_dim_t in_ld,
    dnnl_dim_t out_ld,
    dnnl_data_type_t in_dt,
    dnnl_data_type_t out_dt
    )

Creates a BRGeMM ukernel packing tensor B object.

Parameters:

brgemm_pack_b	Output BRGeMM ukernel packing B object.
K	Dimension K.
N	Dimension N.
in_ld	Input leading dimension.
out_ld	Output leading dimension. Specifies a block by N dimension during data packing.
in_dt	Input data type.
out_dt	Output data type.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_need_pack(
    const_dnnl_brgemm_pack_b_t brgemm_pack_b,
    int* need_pack
    )

Returns the flag if packing is expected by BRGeMM ukernel kernel.

Parameters:

brgemm_pack_b	BRGeMM ukernel packing B object.
need_pack	Output flag specifying if packing is needed. Possible values are 0 (not needed) and 1 (needed).

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_generate(dnnl_brgemm_pack_b_t brgemm_pack_b)

Generates an executable part of BRGeMM ukernel packing B object.

Parameters:

brgemm_pack_b

BRGeMM ukernel packing B object.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_execute(
    const_dnnl_brgemm_pack_b_t brgemm_pack_b,
    const void* in_ptr,
    void* out_ptr
    )

Executes a BRGeMM ukernel packing tensor B object.

Parameters:

brgemm_pack_b	BRGeMM ukernel packing B object.
in_ptr	Pointer to an input buffer.
out_ptr	Pointer to an output buffer.

Returns:

dnnl_success on success and a status describing the error otherwise.

dnnl_status_t DNNL_API dnnl_brgemm_pack_b_destroy(dnnl_brgemm_pack_b_t brgemm_pack_b)

Destroys a BRGeMM ukernel packing tensor B object.

Parameters:

brgemm_pack_b

BRGeMM ukernel packing B object.

Returns:

dnnl_success on success and a status describing the error otherwise.

Ukernels struct dnnl::ukernel::brgemm

oneDNN v3.6.0 documentation

BRGeMM ukernel¶

Overview¶

Detailed Documentation¶

Typedefs¶

Global Functions¶