namespace dnnl::threadpool_interop

Overview

Threadpool interoperability namespace. More…

namespace threadpool_interop {

// structs

struct threadpool_iface;

// global functions

dnnl::stream make_stream(
    const dnnl::engine& aengine,
    threadpool_iface* threadpool
    );

threadpool_iface* get_threadpool(const dnnl::stream& astream);

status sgemm(
    char transa,
    char transb,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    float alpha,
    const float* A,
    dnnl_dim_t lda,
    const float* B,
    dnnl_dim_t ldb,
    float beta,
    float* C,
    dnnl_dim_t ldc,
    threadpool_iface* threadpool
    );

status gemm_u8s8s32(
    char transa,
    char transb,
    char offsetc,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    float alpha,
    const uint8_t* A,
    dnnl_dim_t lda,
    uint8_t ao,
    const int8_t* B,
    dnnl_dim_t ldb,
    int8_t bo,
    float beta,
    int32_t* C,
    dnnl_dim_t ldc,
    const int32_t* co,
    threadpool_iface* threadpool
    );

status gemm_s8s8s32(
    char transa,
    char transb,
    char offsetc,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    float alpha,
    const int8_t* A,
    dnnl_dim_t lda,
    int8_t ao,
    const int8_t* B,
    dnnl_dim_t ldb,
    int8_t bo,
    float beta,
    int32_t* C,
    dnnl_dim_t ldc,
    const int32_t* co,
    threadpool_iface* threadpool
    );

} // namespace threadpool_interop

Detailed Documentation

Threadpool interoperability namespace.

Global Functions

dnnl::stream make_stream(
    const dnnl::engine& aengine,
    threadpool_iface* threadpool
    )

Constructs an execution stream for the specified engine and threadpool.

Parameters:

aengine

Engine to create the stream on.

threadpool

Pointer to an instance of a C++ class that implements dnnl::threapdool_iface interface.

Returns:

An execution stream.

See also:

Using oneDNN with Threadpool-Based Threading

threadpool_iface* get_threadpool(const dnnl::stream& astream)

Returns the pointer to a threadpool that is used by an execution stream.

Parameters:

astream

An execution stream.

Returns:

Output pointer to an instance of a C++ class that implements dnnl::threapdool_iface interface or NULL if the stream was created without threadpool.

See also:

Using oneDNN with Threadpool-Based Threading

status sgemm(
    char transa,
    char transb,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    float alpha,
    const float* A,
    dnnl_dim_t lda,
    const float* B,
    dnnl_dim_t ldb,
    float beta,
    float* C,
    dnnl_dim_t ldc,
    threadpool_iface* threadpool
    )

Performs single-precision matrix-matrix multiply.

The operation is defined as:

C := alpha * op( A ) * op( B ) + beta * C

where

  • op( X ) = X or op( X ) = X**T,

  • alpha and beta are scalars, and

  • A, B, and C are matrices:

    • op( A ) is an MxK matrix,

    • op( B ) is an KxN matrix,

    • C is an MxN matrix.

The matrices are assumed to be stored in row-major order (the elements in each of the matrix rows are contiguous in memory).

Note

This API does not support XERBLA. Instead, unlike the standard BLAS functions, this one returns a dnnl_status_t value to allow error handling.

Parameters:

transa

Transposition flag for matrix A: ‘N’ or ‘n’ means A is not transposed, and ‘T’ or ‘t’ means that A is transposed.

transb

Transposition flag for matrix B: ‘N’ or ‘n’ means B is not transposed, and ‘T’ or ‘t’ means that B is transposed.

M

The M dimension.

N

The N dimension.

K

The K dimension.

alpha

The alpha parameter that is used to scale the product of matrices A and B.

A

A pointer to the A matrix data.

lda

The leading dimension for the matrix A.

B

A pointer to the B matrix data.

ldb

The leading dimension for the matrix B.

beta

The beta parameter that is used to scale the matrix C.

C

A pointer to the C matrix data.

ldc

The leading dimension for the matrix C.

threadpool

A pointer to a threadpool interface (only when built with the THREADPOOL CPU runtime).

Returns:

dnnl_success / dnnl::status::success on success and a status describing the error otherwise.

status gemm_u8s8s32(
    char transa,
    char transb,
    char offsetc,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    float alpha,
    const uint8_t* A,
    dnnl_dim_t lda,
    uint8_t ao,
    const int8_t* B,
    dnnl_dim_t ldb,
    int8_t bo,
    float beta,
    int32_t* C,
    dnnl_dim_t ldc,
    const int32_t* co,
    threadpool_iface* threadpool
    )

Performs integer matrix-matrix multiply on 8-bit unsigned matrix A, 8-bit signed matrix B, and 32-bit signed resulting matrix C.

The operation is defined as:

C := alpha * (op(A) - A_offset) * (op(B) - B_offset) + beta * C + C_offset

where

  • op( X ) = X or op( X ) = X**T,

  • alpha and beta are scalars, and

  • A, B, and C are matrices:

    • op( A ) is an MxK matrix,

    • op( B ) is an KxN matrix,

    • C is an MxN matrix.

  • A_offset is an MxK matrix with every element equal the ao value,

  • B_offset is an KxN matrix with every element equal the bo value,

  • C_offset is an MxN matrix which is defined by the co array of size len :

    • if offsetc = F : the len must be at least 1,

    • if offsetc = C : the len must be at least max(1, m),

    • if offsetc = R : the len must be at least max(1, n),

The matrices are assumed to be stored in row-major order (the elements in each of the matrix rows are contiguous in memory).

Note

This API does not support XERBLA. Instead, unlike the standard BLAS functions, this one returns a dnnl_status_t value to allow error handling.

Warning

On some architectures saturation may happen during intermediate computations, which would lead to unexpected results. For more details, refer to Nuances of int8 Computations.

Parameters:

transa

Transposition flag for matrix A: ‘N’ or ‘n’ means A is not transposed, and ‘T’ or ‘t’ means that A is transposed.

transb

Transposition flag for matrix B: ‘N’ or ‘n’ means B is not transposed, and ‘T’ or ‘t’ means that B is transposed.

offsetc

Flag specifying how offsets should be applied to matrix C:

  • ‘F’ means that the same offset will be applied to each element of the matrix C,

  • ‘C’ means that individual offset will be applied to each element within each column,

  • ‘R’ means that individual offset will be applied to each element within each row.

M

The M dimension.

N

The N dimension.

K

The K dimension.

alpha

The alpha parameter that is used to scale the product of matrices A and B.

A

A pointer to the A matrix data.

lda

The leading dimension for the matrix A.

ao

The offset value for the matrix A.

B

A pointer to the B matrix data.

ldb

The leading dimension for the matrix B.

bo

The offset value for the matrix B.

beta

The beta parameter that is used to scale the matrix C.

C

A pointer to the C matrix data.

ldc

The leading dimension for the matrix C.

co

An array of offset values for the matrix C. The number of elements in the array depends on the value of offsetc.

threadpool

A pointer to a threadpool interface (only when built with the THREADPOOL CPU runtime).

Returns:

dnnl_success / dnnl::status::success on success and a status describing the error otherwise.

status gemm_s8s8s32(
    char transa,
    char transb,
    char offsetc,
    dnnl_dim_t M,
    dnnl_dim_t N,
    dnnl_dim_t K,
    float alpha,
    const int8_t* A,
    dnnl_dim_t lda,
    int8_t ao,
    const int8_t* B,
    dnnl_dim_t ldb,
    int8_t bo,
    float beta,
    int32_t* C,
    dnnl_dim_t ldc,
    const int32_t* co,
    threadpool_iface* threadpool
    )

Performs integer matrix-matrix multiply on 8-bit signed matrix A, 8-bit signed matrix B, and 32-bit signed resulting matrix C.

The operation is defined as:

C := alpha * (op(A) - A_offset) * (op(B) - B_offset) + beta * C + C_offset

where

  • op( X ) = X or op( X ) = X**T,

  • alpha and beta are scalars, and

  • A, B, and C are matrices:

    • op( A ) is an MxK matrix,

    • op( B ) is an KxN matrix,

    • C is an MxN matrix.

  • A_offset is an MxK matrix with every element equal the ao value,

  • B_offset is an KxN matrix with every element equal the bo value,

  • C_offset is an MxN matrix which is defined by the co array of size len :

    • if offsetc = F : the len must be at least 1,

    • if offsetc = C : the len must be at least max(1, m),

    • if offsetc = R : the len must be at least max(1, n),

The matrices are assumed to be stored in row-major order (the elements in each of the matrix rows are contiguous in memory).

Note

This API does not support XERBLA. Instead, unlike the standard BLAS functions, this one returns a dnnl_status_t value to allow error handling.

Warning

On some architectures saturation may happen during intermediate computations, which would lead to unexpected results. For more details, refer to Nuances of int8 Computations.

Parameters:

transa

Transposition flag for matrix A: ‘N’ or ‘n’ means A is not transposed, and ‘T’ or ‘t’ means that A is transposed.

transb

Transposition flag for matrix B: ‘N’ or ‘n’ means B is not transposed, and ‘T’ or ‘t’ means that B is transposed.

offsetc

Flag specifying how offsets should be applied to matrix C:

  • ‘F’ means that the same offset will be applied to each element of the matrix C,

  • ‘C’ means that individual offset will be applied to each element within each column,

  • ‘R’ means that individual offset will be applied to each element within each row.

M

The M dimension.

N

The N dimension.

K

The K dimension.

alpha

The alpha parameter that is used to scale the product of matrices A and B.

A

A pointer to the A matrix data.

lda

The leading dimension for the matrix A.

ao

The offset value for the matrix A.

B

A pointer to the B matrix data.

ldb

The leading dimension for the matrix B.

bo

The offset value for the matrix B.

beta

The beta parameter that is used to scale the matrix C.

C

A pointer to the C matrix data.

ldc

The leading dimension for the matrix C.

co

An array of offset values for the matrix C. The number of elements in the array depends on the value of offsetc.

threadpool

A pointer to a threadpool interface (only when built with the THREADPOOL CPU runtime).

Returns:

dnnl_success / dnnl::status::success on success and a status describing the error otherwise.