/github/workspace/src/MatrixFunctionsStride/mat_mult_trans_stride/kernels/plp_mat_mult_trans_stride_q16s_rv32im.c
Functions
Name | |
---|---|
void | plp_mat_mult_trans_stride_q16s_rv32im(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) Matrix multiplication of 16-bit integer matrices kernel for RV32IM extension. |
Defines
Name | |
---|---|
BASIC_VERSION |
Functions Documentation
function plp_mat_mult_trans_stride_q16s_rv32im
void plp_mat_mult_trans_stride_q16s_rv32im(
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
)
Matrix multiplication of 16-bit integer matrices kernel for RV32IM extension.
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).
strided matrix transposed matrix multiplication of a 16-bit fix-point matrices for RV32IM extension.
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.
Macros Documentation
define BASIC_VERSION
#define BASIC_VERSION
Source code
/* =====================================================================
* Project: PULP DSP Library
* Title: plp_mat_mult_trans_stride_q16s_xpulpv2.c
* Description: 16-bit fix-point strided matrix multiplication for RV32IM
*
* $Date: 14. July 2020
* $Revision: V0
*
* Target Processor: PULP cores
* ===================================================================== */
/*
* Copyright (C) 2020 ETH Zurich and University of Bologna.
*
* Author: Tibor Schneider, ETH Zurich
*
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the License); you may
* not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS, WITHOUT
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "plp_math.h"
void plp_mat_mult_trans_stride_q16s_rv32im(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) {
#define BASIC_VERSION // if used don't forget to also use the undefine at end of file
#ifdef BASIC_VERSION
int32_t round = 1 << (shift - 1);
uint32_t m; // loop counter
uint32_t n; // loop counter
uint32_t o; // loop counter
for (m = 0; m < M; m++) {
for (o = 0; o < O; o++) {
int32_t sum = 0;
for (n = 0; n < N; n++) {
int32_t valA = (int32_t)pSrcA[m * strideA + n];
int32_t valB = (int32_t)pSrcB[o * strideB + n];
sum += (valA * valB + round) >> shift;
}
pDstC[m * strideC + o] = (int16_t)sum;
}
}
#else
// TODO hackathon
#endif
#undef BASIC_VERSION
}
Updated on 2023-03-01 at 16:16:33 +0000