Strided Matrix Transposed Matrix Multiplication

Module: Strided Matrix operations

Modules

Name
Strided Matrix Transpose Matrix Multiplication Kernels

Functions

	Name
void	plp_mat_mult_trans_stride_f32(const float restrict pSrcA, const float restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, float *restrict pDstC) Glue code for strided matrix multiplication of 32-bit floating-point matrices.
void	plp_mat_mult_trans_stride_f32_parallel(const float restrict pSrcA, const float restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t nPE, float *restrict pDstC) Glue code for parallel strided matrix multiplication of 32-bit floating-point matrices.
void	plp_mat_mult_trans_stride_i16(const int16_t restrict pSrcA, const int16_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, int32_t *restrict pDstC) Glue code for strided matrix multiplication of 16-bit integer matrices.
void	plp_mat_mult_trans_stride_i16_parallel(const int16_t restrict pSrcA, const int16_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t nPE, int32_t *restrict pDstC) Glue code for parallel strided matrix multiplication of 16-bit integer matrices.
void	plp_mat_mult_trans_stride_i32(const int32_t restrict pSrcA, const int32_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, int32_t *restrict pDstC) Glue code for strided matrix multiplication of 32-bit integer matrices.
void	plp_mat_mult_trans_stride_i32_parallel(const int32_t restrict pSrcA, const int32_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t nPE, int32_t *restrict pDstC) Glue code for parallel strided matrix multiplication of 32-bit integer matrices.
void	plp_mat_mult_trans_stride_i8(const int8_t restrict pSrcA, const int8_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, int32_t *restrict pDstC) Glue code for strided matrix multiplication of 8-bit integer matrices.
void	plp_mat_mult_trans_stride_i8_parallel(const int8_t restrict pSrcA, const int8_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t nPE, int32_t *restrict pDstC) Glue code for parallel strided matrix multiplication of 8-bit integer matrices.
void	plp_mat_mult_trans_stride_q16(const int16_t restrict pSrcA, const int16_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t shift, int16_t *restrict pDstC) Glue code for strided matrix multiplication of 16-bit fix-point matrices.
void	plp_mat_mult_trans_stride_q16_parallel(const int16_t restrict pSrcA, const int16_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t shift, uint32_t nPE, int16_t *restrict pDstC) Glue code for parallel strided matrix multiplication of 16-bit fix-point matrices.
void	plp_mat_mult_trans_stride_q32(const int32_t restrict pSrcA, const int32_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t shift, int32_t *restrict pDstC) Glue code for strided matrix multiplication of 32-bit fix-point matrices.
void	plp_mat_mult_trans_stride_q32_parallel(const int32_t restrict pSrcA, const int32_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t shift, uint32_t nPE, int32_t *restrict pDstC) Glue code for parallel strided matrix multiplication of 32-bit fix-point matrices.
void	plp_mat_mult_trans_stride_q8(const int8_t restrict pSrcA, const int8_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t shift, int8_t *restrict pDstC) Glue code for strided matrix multiplication of 8-bit fix-point matrices.
void	plp_mat_mult_trans_stride_q8_parallel(const int8_t restrict pSrcA, const int8_t restrict pSrcB, uint32_t M, uint32_t N, uint32_t O, uint32_t strideA, uint32_t strideB, uint32_t strideC, uint32_t shift, uint32_t nPE, int8_t *restrict pDstC) Glue code for parallel strided matrix multiplication of 8-bit fix-point matrices.

Detailed Description

This module contains the glue code for Matrix Transposed Matrix Multiplication. The kernel codes (kernels) are in the Moducle Matrix Transposed Matrix Multiplication Kernels.

The Matrix Transpose Matrix Multiplication computes the product of two matrices with dimensions MxN and OxN. Both matrices is accessed row wise, all values form the first are multiplied with the values of the second and then sum of the result gives the value for the result matrix. The resulting matrix has shape MxO.

'pDst[m,o] = pSrcA[m,0]*pSrcB[o,0] + pSrcA[m,1]*pSrcB[o,1] + ... + pSrcA[m,N-1]*pSrcB[o,N-1]`

There are functions for integer 32- 16- and 8-bit data types. For lower precision integers (16- and 8-bit), functions exploiting SIMD instructions are provided.

The naming scheme of the functions follows the following pattern (for example plp_mat_mult_stride_i32):

`plp_<function name>_<data type><precision>[_parallel]`

name	description
function_name	`mat_mult_trans_stride`
data type	{f, i, q} respectively for floats, integers, fixed points
precision	{32, 16, 8} bits

The strideX argument tells how many elements are in between the start of each row of the matrix. In other words, it is the width of the original matrix. Strided Matrix operations

Functions Documentation

function plp_mat_mult_trans_stride_f32

void plp_mat_mult_trans_stride_f32(
    const float *__restrict__ pSrcA,
    const float *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    float *__restrict__ pDstC
)

Glue code for strided matrix multiplication of 32-bit floating-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
pDstC points to the output matrix

Return: none

Glue code for strided matrix transposed matrix multiplication of a 32-bit floating-point matrices.

function plp_mat_mult_trans_stride_f32_parallel

void plp_mat_mult_trans_stride_f32_parallel(
    const float *__restrict__ pSrcA,
    const float *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t nPE,
    float *__restrict__ pDstC
)

Glue code for parallel strided matrix multiplication of 32-bit floating-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
nPE Number of cores to use
pDstC points to the output matrix

Return: none

Glue code for parallel strided matrix transposed matrix multiplication of a 32-bit floating-point matrices.

function plp_mat_mult_trans_stride_i16

void plp_mat_mult_trans_stride_i16(
    const int16_t *__restrict__ pSrcA,
    const int16_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    int32_t *__restrict__ pDstC
)

Glue code for strided matrix multiplication of 16-bit integer matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
pDstC points to the output matrix

Return: none

Glue code for strided matrix transposed matrix multiplication of a 16-bit integer matrices.

function plp_mat_mult_trans_stride_i16_parallel

void plp_mat_mult_trans_stride_i16_parallel(
    const int16_t *__restrict__ pSrcA,
    const int16_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t nPE,
    int32_t *__restrict__ pDstC
)

Glue code for parallel strided matrix multiplication of 16-bit integer matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
nPE Number of cores to use
pDstC points to the output matrix

Return: none

Glue code for parallel strided matrix transposed matrix multiplication of a 16-bit integer matrices.

function plp_mat_mult_trans_stride_i32

void plp_mat_mult_trans_stride_i32(
    const int32_t *__restrict__ pSrcA,
    const int32_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    int32_t *__restrict__ pDstC
)

Glue code for strided matrix multiplication of 32-bit integer matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
pDstC points to the output matrix

Return: none

Glue code for strided matrix transposed matrix multiplication of a 32-bit integer matrices.

function plp_mat_mult_trans_stride_i32_parallel

void plp_mat_mult_trans_stride_i32_parallel(
    const int32_t *__restrict__ pSrcA,
    const int32_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t nPE,
    int32_t *__restrict__ pDstC
)

Glue code for parallel strided matrix multiplication of 32-bit integer matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
nPE Number of cores to use
pDstC points to the output matrix

Return: none

Glue code for parallel strided matrix matrix multiplication of a 32-bit integer matrices.

function plp_mat_mult_trans_stride_i8

void plp_mat_mult_trans_stride_i8(
    const int8_t *__restrict__ pSrcA,
    const int8_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    int32_t *__restrict__ pDstC
)

Glue code for strided matrix multiplication of 8-bit integer matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
pDstC points to the output matrix

Return: none

Glue code for strided matrix transposed matrix multiplication of a 8-bit integer matrices.

function plp_mat_mult_trans_stride_i8_parallel

void plp_mat_mult_trans_stride_i8_parallel(
    const int8_t *__restrict__ pSrcA,
    const int8_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t nPE,
    int32_t *__restrict__ pDstC
)

Glue code for parallel strided matrix multiplication of 8-bit integer matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
nPE Number of cores to use
pDstC points to the output matrix

Return: none

Glue code for parallel strided matrix transposed matrix multiplication of a 8-bit integer matrices.

function plp_mat_mult_trans_stride_q16

void plp_mat_mult_trans_stride_q16(
    const int16_t *__restrict__ pSrcA,
    const int16_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t shift,
    int16_t *__restrict__ pDstC
)

Glue code for strided matrix multiplication of 16-bit fix-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
shift Amount to shift the result of each multiplication.
pDstC points to the output matrix

Return: none

Par: Fix-Point and Shifting

The result will be shifted by the parameter shift to the right (multiplied by 2^-shift). Assume that matrix A is represented as pSrcA * 2^-x, and matrix B as pSrcB * 2^-y (in other words, A has it's x last digits after the binary point). Then, the output is represented as pDstC * 2^-(x + y - shift).

Glue code for strided matrix transposed matrix multiplication of a 16-bit fix-point matrices.

The output of the strided matrix multiplication will also be stored as an 16-bit array. Set the shift parameter such that no overflow ocurrs.

function plp_mat_mult_trans_stride_q16_parallel

void plp_mat_mult_trans_stride_q16_parallel(
    const int16_t *__restrict__ pSrcA,
    const int16_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t shift,
    uint32_t nPE,
    int16_t *__restrict__ pDstC
)

Glue code for parallel strided matrix multiplication of 16-bit fix-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
shift Amount to shift the result of each multiplication.
nPE Number of cores to use
pDstC points to the output matrix

Return: none

Par: Fix-Point and Shifting

The result will be shifted by the parameter shift to the right (multiplied by 2^-shift). Assume that matrix A is represented as pSrcA * 2^-x, and matrix B as pSrcB * 2^-y (in other words, A has it's x last digits after the binary point). Then, the output is represented as pDstC * 2^-(x + y - shift).

Glue code for parallel strided matrix transposed matrix multiplication of a 16-bit fix-point matrices.

The output of the strided matrix multiplication will also be stored as an 16-bit array. Set the shift parameter such that no overflow ocurrs.

function plp_mat_mult_trans_stride_q32

void plp_mat_mult_trans_stride_q32(
    const int32_t *__restrict__ pSrcA,
    const int32_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t shift,
    int32_t *__restrict__ pDstC
)

Glue code for strided matrix multiplication of 32-bit fix-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
shift Amount to shift the result of each multiplication.
pDstC points to the output matrix

Return: none

Par: Fix-Point and Shifting

The result will be shifted by the parameter shift to the right (multiplied by 2^-shift). Assume that matrix A is represented as pSrcA * 2^-x, and matrix B as pSrcB * 2^-y (in other words, A has it's x last digits after the binary point). Then, the output is represented as pDstC * 2^-(x + y - shift).

Glue code for strided matrix transposed matrix multiplication of a 32-bit fix-point matrices.

function plp_mat_mult_trans_stride_q32_parallel

void plp_mat_mult_trans_stride_q32_parallel(
    const int32_t *__restrict__ pSrcA,
    const int32_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t shift,
    uint32_t nPE,
    int32_t *__restrict__ pDstC
)

Glue code for parallel strided matrix multiplication of 32-bit fix-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
shift Amount to shift the result of each multiplication.
nPE Number of cores to use
pDstC points to the output matrix

Return: none

Par: Fix-Point and Shifting

The result will be shifted by the parameter shift to the right (multiplied by 2^-shift). Assume that matrix A is represented as pSrcA * 2^-x, and matrix B as pSrcB * 2^-y (in other words, A has it's x last digits after the binary point). Then, the output is represented as pDstC * 2^-(x + y - shift).

Glue code for parallel strided matrix transposed matrix multiplication of a 32-bit fix-point matrices.

function plp_mat_mult_trans_stride_q8

void plp_mat_mult_trans_stride_q8(
    const int8_t *__restrict__ pSrcA,
    const int8_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t shift,
    int8_t *__restrict__ pDstC
)

Glue code for strided matrix multiplication of 8-bit fix-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
shift Amount to shift the result of each multiplication.
pDstC points to the output matrix

Return: none

Par: Fix-Point and Shifting

The result will be shifted by the parameter shift to the right (multiplied by 2^-shift). Assume that matrix A is represented as pSrcA * 2^-x, and matrix B as pSrcB * 2^-y (in other words, A has it's x last digits after the binary point). Then, the output is represented as pDstC * 2^-(x + y - shift).

Glue code for strided matrix transposed matrix multiplication of a 8-bit fix-point matrices.

The output of the strided matrix multiplication will also be stored as an 8-bit array. Set the shift parameter such that no overflow ocurrs.

function plp_mat_mult_trans_stride_q8_parallel

void plp_mat_mult_trans_stride_q8_parallel(
    const int8_t *__restrict__ pSrcA,
    const int8_t *__restrict__ pSrcB,
    uint32_t M,
    uint32_t N,
    uint32_t O,
    uint32_t strideA,
    uint32_t strideB,
    uint32_t strideC,
    uint32_t shift,
    uint32_t nPE,
    int8_t *__restrict__ pDstC
)

Glue code for parallel strided matrix multiplication of 8-bit fix-point matrices.

Parameters:

pSrcA points to the first input matrix
pSrcB points to the second input matrix
M height of the first input matrix
N width of the first input matrix and hight of the second
O width of the second input matrix
strideA Stride of matrix A (elements between each row)
strideB Stride of matrix B (elements between each row)
strideC Stride of output matrix (elements between each row)
shift Amount to shift the result of each multiplication.
nPE Number of cores to use
pDstC points to the output matrix

Return: none

Par: Fix-Point and Shifting

The result will be shifted by the parameter shift to the right (multiplied by 2^-shift). Assume that matrix A is represented as pSrcA * 2^-x, and matrix B as pSrcB * 2^-y (in other words, A has it's x last digits after the binary point). Then, the output is represented as pDstC * 2^-(x + y - shift).

Glue code for parallel strided matrix transposed matrix multiplication of a 8-bit fix-point matrices.

The output of the strided matrix multiplication will also be stored as an 8-bit array. Set the shift parameter such that no overflow ocurrs.

Updated on 2023-03-01 at 16:16:32 +0000