/github/workspace/src/FilteringFunctions/kernels/plp_conv_i8s_xpulpv2.c
Functions
| Name | |
|---|---|
| void | plp_conv_i8s_xpulpv2(const int8_t * pSrcA, const uint32_t srcALen, const int8_t * pSrcB, const uint32_t srcBLen, int32_t * pRes) Convolution of 8-bit integer vectors kernel for XPULPV2 extension. |
Defines
| Name | |
|---|---|
| shufflemask1 | |
| shufflemask2 | |
| shufflemask3 |
Functions Documentation
function plp_conv_i8s_xpulpv2
void plp_conv_i8s_xpulpv2(
const int8_t * pSrcA,
const uint32_t srcALen,
const int8_t * pSrcB,
const uint32_t srcBLen,
int32_t * pRes
)
Convolution of 8-bit integer vectors kernel for XPULPV2 extension.
Parameters:
- pSrcA points to the first input vector
- srcALen Length of the first input vector
- pSrcB points to the second input vector
- srcBLen Length of the second input vector
- pRes output result returned here
Return: none
Macros Documentation
define shufflemask1
#define shufflemask1 (v4s) { 3, 2, 1, 0 }
define shufflemask2
#define shufflemask2 (v4s) { 1, 2, 3, 5 }
define shufflemask3
#define shufflemask3 (v4s) { 2, 3, 5, 6 }
Source code
/* =====================================================================
* Project: PULP DSP Library
* Title: plp_conv_i8s_xpulpv2.c
* Description: 8-bit integer singlecore convolution for XPULPV2
*
* $Date: 01. July 2019
* $Revision: V0
*
* Target Processor: PULP cores
* ===================================================================== */
/*
* Copyright (C) 2019 ETH Zurich and University of Bologna.
*
* Author: Moritz Scherer, 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"
#define shufflemask1 \
(v4s) { 3, 2, 1, 0 }
#define shufflemask2 \
(v4s) { 1, 2, 3, 5 }
#define shufflemask3 \
(v4s) { 2, 3, 5, 6 }
// Pre-condition: psrcALen >= psrcBLen, established by calling function plp_conv_i8
// Pre-condition: pRes has enough allocated memory, i.e. srcALen + srcBLen-1u
// Pre-condition: srcALen >= 2 and srcBLen >= 2, otherwise use vector dot product
void plp_conv_i8s_xpulpv2(const int8_t *pSrcA,
const uint32_t srcALen,
const int8_t *pSrcB,
const uint32_t srcBLen,
int32_t *pRes) {
const int8_t *pIn1 = pSrcA; /* InputA pointer */
const int8_t *pIn2 = pSrcB; /* InputB pointer */
int32_t *pOut = pRes; /* Output pointer */
const int8_t *px; /* Intermediate inputA pointer */
const int8_t *py; /* Intermediate inputB pointer */
const int8_t *pSrc1, *pSrc2; /* Intermediate pointers */
int32_t sum; /* Accumulators */
uint32_t blockSize1, blockSize2, blockSize3; /* Loop counters */
uint32_t j, k, count, blkCnt; /* Loop counters */
#if defined(PLP_MATH_LOOPUNROLL)
int32_t acc0, acc1, acc2, acc3; /* Accumulators */
int8_t x0, x1, x2, x3, c0; /* Temporary variables to hold state and coefficient values */
#endif
int32_t temp1, temp2;
v4s xmask[] = { (v4s){ 0, 0, 0, 0 }, (v4s){ 0xff, 0, 0, 0 }, (v4s){ 0xff, 0xff, 0, 0 },
(v4s){ 0xff, 0xff, 0xff, 0 } };
v4s ymask[] = { (v4s){ 0, 0, 0, 0 }, (v4s){ 0, 0, 0, 0xff }, (v4s){ 0, 0, 0xff, 0xff },
(v4s){ 0, 0xff, 0xff, 0xff } };
v4s mask;
v4s _x1, _x2, _x3, _x4;
v4s _y1, _y2;
blockSize1 = srcBLen - 1U;
blockSize2 = srcALen - (srcBLen - 1U);
blockSize3 = blockSize1;
/* --------------------------
* Initializations of stage1
* -------------------------*/
/* sum = x[0] * y[0]
* sum = x[0] * y[1] + x[1] * y[0]
* ....
* sum = x[0] * y[srcBlen - 1] + x[1] * y[srcBlen - 2] +...+ x[srcBLen - 1] * y[0]
*/
/* In this stage the MAC operations are increased by 1 for every iteration.
The count variable holds the number of MAC operations performed */
count = 1U;
/* Working pointer of inputA */
px = pIn1;
/* Working pointer of inputB */
py = pIn2;
/* ------------------------
* Stage1 process
* ----------------------*/
/* The first stage starts here */
while (blockSize1 > 0U) {
/* Accumulator is made zero for every iteration */
_y1 = *((v4s *)(py - 3));
_x1 = *((v4s *)(px));
sum = 0;
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
#if defined(PLP_MATH_LOOPUNROLL)
/* Loop unrolling: Compute 4 outputs at a time */
k = count >> 2U;
while (k > 0U) {
sum = __SUMDOTP4(_x1, _y1, sum);
_y1 = *((v4s *)(py - 7));
_x1 = *((v4s *)(px + 4));
px += 4U;
py -= 4U;
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
k--;
}
/* Loop unrolling: Compute remaining outputs */
k = count % 0x4U;
mask = xmask[k];
_x1 = __AND4(_x1, mask);
sum = __SUMDOTP4(_x1, _y1, sum);
#else
/* Initialize k with number of samples */
k = count;
while (k) {
sum = __MAC(sum, *px++, *py--);
k--;
}
#endif /* #if defined (PLP_MATH_LOOPUNROLL) */
/* Store the result in the accumulator in the destination buffer. */
*pOut++ = sum;
/* Update the inputA and inputB pointers for next MAC calculation */
py = pIn2 + count;
px = pIn1;
/* Increment MAC count */
count++;
/* Decrement loop counter */
blockSize1--;
}
/* --------------------------
* Initializations of stage2
* ------------------------*/
/* sum = x[0] * y[srcBLen-1] + x[1] * y[srcBLen-2] +...+ x[srcBLen-1] * y[0]
* sum = x[1] * y[srcBLen-1] + x[2] * y[srcBLen-2] +...+ x[srcBLen] * y[0]
* ....
* sum = x[srcALen-srcBLen-2] * y[srcBLen-1] + x[srcALen] * y[srcBLen-2] +...+ x[srcALen-1] *
* y[0]
*/
/* Working pointer of inputA */
px = pIn1;
/* Working pointer of inputB */
pSrc2 = pIn2 + (srcBLen - 1U);
py = pSrc2;
/* count is index by which the pointer pIn1 to be incremented */
count = 0U;
/* -------------------
* Stage2 process
* ------------------*/
/* Stage2 depends on srcBLen as in this stage srcBLen number of MACS are performed.
* So, to loop unroll over blockSize2,
* srcBLen should be greater than or equal to 4 */
if (srcBLen >= 4U) {
#if defined(PLP_MATH_LOOPUNROLL)
/* Loop unrolling: Compute 4 outputs at a time */
blkCnt = blockSize2 >> 2U;
while (blkCnt > 0U) {
/* Set all accumulators to zero */
acc0 = 0;
acc1 = 0;
acc2 = 0;
acc3 = 0;
/* Apply loop unrolling and compute 4 MACs simultaneously. */
k = srcBLen >> 2U;
/* First part of the processing with loop unrolling. Compute 4 MACs at a time.
** a second loop below computes MACs for the remaining 1 to 3 samples. */
do {
/* Read y[srcBLen - 1] sample */
_x1 = *((v4s *)px); // {x[0],x[1],x[2],x[3]}
_x4 = *((v4s *)(px + 3)); // {x[3],x[4],x[5],x[6]}
_y1 = *((v4s *)(py - 3)); // {y[srcBLen - 4],y[srcBLen - 3],y[srcBLen - 2],y[srcBLen
// - 1]}
px += 4U;
py -= 4U;
_x2 = __builtin_shuffle(_x1, _x4, shufflemask2); // {x[1],x[2],x[3],x[4]}
_x3 = __builtin_shuffle(_x1, _x4, shufflemask3); // {x[2],x[3],x[4],x[5]}
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1); // {y[srcBLen - 1],y[srcBLen -
// 2],y[srcBLen - 3],y[srcBLen - 4]}
acc0 = __SUMDOTP4(_x1, _y1, acc0);
acc1 = __SUMDOTP4(_x2, _y1, acc1);
acc2 = __SUMDOTP4(_x3, _y1, acc2);
acc3 = __SUMDOTP4(_x4, _y1, acc3);
} while (--k);
/* If the srcBLen is not a multiple of 4, compute any remaining MACs here.
** No loop unrolling is used. */
k = srcBLen % 0x4U;
if (k > 0) {
_x1 = *((v4s *)px); // {x[0],x[1],x[2],x[3]}
_x4 = *((v4s *)(px + 3)); // {x[3],x[4],x[5],x[6]}
_y1 = *((v4s *)(py - 3)); // {y[srcBLen - 4],y[srcBLen - 3],y[srcBLen - 2],y[srcBLen
// - 1]}
mask = ymask[k];
_x2 = __builtin_shuffle(_x1, _x4, shufflemask2); // {x[1],x[2],x[3],x[4]}
_x3 = __builtin_shuffle(_x1, _x4, shufflemask3); // {x[2],x[3],x[4],x[5]}
_y1 = __AND4(_y1, mask);
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
/* Perform the multiply-accumulate */
acc0 = __SUMDOTP4(_x1, _y1, acc0);
acc1 = __SUMDOTP4(_x2, _y1, acc1);
acc2 = __SUMDOTP4(_x3, _y1, acc2);
acc3 = __SUMDOTP4(_x4, _y1, acc3);
}
/* Store the result in the accumulator in the destination buffer. */
*pOut++ = acc0;
*pOut++ = acc1;
*pOut++ = acc2;
*pOut++ = acc3;
/* Increment the pointer pIn1 index, count by 4 */
count += 4U;
/* Update the inputA and inputB pointers for next MAC calculation */
px = pIn1 + count;
py = pSrc2;
/* Decrement the loop counter */
blkCnt--;
}
/* If the blockSize2 is not a multiple of 4, compute any remaining output samples here.
** No loop unrolling is used. */
blkCnt = blockSize2 % 0x4U;
#else
/* Initialize blkCnt with number of samples */
blkCnt = blockSize2;
#endif /* #if defined (PLP_MATH_LOOPUNROLL)*/
while (blkCnt > 0U) {
/* Accumulator is made zero for every iteration */
_y1 = *((v4s *)(py - 3));
_x1 = *((v4s *)(px));
sum = 0;
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
#if defined(PLP_MATH_LOOPUNROLL)
/* Loop unrolling: Compute 8 outputs at a time */
k = srcBLen >> 2U;
while (k > 0U) {
sum = __SUMDOTP4(_x1, _y1, sum);
_y1 = *((v4s *)(py - 7));
_x1 = *((v4s *)(px + 4));
px += 4U;
py -= 4U;
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
k--;
}
/* Loop unrolling: Compute remaining outputs */
k = srcBLen % 0x4U;
mask = xmask[k];
_x1 = __AND4(_x1, mask);
sum = __SUMDOTP4(_x1, _y1, sum);
#else
/* Initialize blkCnt with number of samples */
k = srcBLen;
#endif /* #if defined (PLP_MATH_LOOPUNROLL) */
/* Store the result in the accumulator in the destination buffer. */
*pOut++ = sum;
/* Increment the MAC count */
count++;
/* Update the inputA and inputB pointers for next MAC calculation */
px = pIn1 + count;
py = pSrc2;
/* Decrement the loop counter */
blkCnt--;
}
} else {
/* If the srcBLen is not a multiple of 4,
* the blockSize2 loop cannot be unrolled by 4 */
blkCnt = blockSize2;
while (blkCnt > 0U) {
/* Accumulator is made zero for every iteration */
sum = 0;
/* srcBLen number of MACS should be performed */
k = srcBLen;
mask = xmask[k];
_y1 = *((v4s *)(py - 3));
_x1 = *((v4s *)(px));
_x1 = __AND4(_x1, mask);
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
sum = __SUMDOTP4(_x1, _y1, sum);
/* Store the result in the accumulator in the destination buffer. */
*pOut++ = sum;
/* Increment the MAC count */
count++;
/* Update the inputA and inputB pointers for next MAC calculation */
px = pIn1 + count;
py = pSrc2;
/* Decrement the loop counter */
blkCnt--;
}
}
/* --------------------------
* Initializations of stage3
* -------------------------*/
/* sum += x[srcALen-srcBLen+1] * y[srcBLen-1] + x[srcALen-srcBLen+2] * y[srcBLen-2] +...+
* x[srcALen-1] * y[1] sum += x[srcALen-srcBLen+2] * y[srcBLen-1] + x[srcALen-srcBLen+3] *
* y[srcBLen-2] +...+ x[srcALen-1] * y[2]
* ....
* sum += x[srcALen-2] * y[srcBLen-1] + x[srcALen-1] * y[srcBLen-2]
* sum += x[srcALen-1] * y[srcBLen-1]
*/
/* In this stage the MAC operations are decreased by 1 for every iteration.
The blockSize3 variable holds the number of MAC operations performed */
/* Working pointer of inputA */
pSrc1 = pIn1 + (srcALen - (srcBLen - 1U));
px = pSrc1;
/* Working pointer of inputB */
pSrc2 = pIn2 + (srcBLen - 1U);
py = pSrc2;
/* -------------------
* Stage3 process
* ------------------*/
while (blockSize3 > 0U) {
/* Accumulator is made zero for every iteration */
_y1 = *((v4s *)(py - 3));
_x1 = *((v4s *)(px));
sum = 0;
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
#if defined(PLP_MATH_LOOPUNROLL)
/* Loop unrolling: Compute 4 outputs at a time */
k = blockSize3 >> 2U;
while (k > 0U) {
sum = __SUMDOTP4(_x1, _y1, sum);
_y1 = *((v4s *)(py - 7));
_x1 = *((v4s *)(px + 4));
px += 4U;
py -= 4U;
_y1 = __builtin_shuffle(_y1, _y1, shufflemask1);
k--;
}
/* Loop unrolling: Compute remaining outputs */
k = blockSize3 % 0x4U;
mask = xmask[k];
_x1 = __AND4(_x1, mask);
sum = __SUMDOTP4(_x1, _y1, sum);
#else
/* Initialize blkCnt with number of samples */
k = blockSize3;
while (k) {
sum = __MAC(sum, *px++, *py--);
k--;
}
#endif /* defined (PLP_MATH_LOOPUNROLL)*/
/* Store the result in the accumulator in the destination buffer. */
*pOut++ = sum;
/* Update the inputA and inputB pointers for next MAC calculation */
px = ++pSrc1;
py = pSrc2;
/* Decrement the loop counter */
blockSize3--;
}
}
Updated on 2023-03-01 at 16:16:32 +0000