/github/workspace/src/FilteringFunctions/plp_conv_i32_parallel.c

Functions

	Name
void	plp_conv_i32_parallel(const int32_t * pSrcA, const uint32_t srcALen, const int32_t * pSrcB, const uint32_t srcBLen, const uint8_t nPE, int32_t * pRes) Glue code for parallel convolution of 32-bit integer vectors.

Attributes

	Name
HAL_CL_L1 int32_t *	resultsBuffer

Functions Documentation

function plp_conv_i32_parallel

void plp_conv_i32_parallel(
    const int32_t * pSrcA,
    const uint32_t srcALen,
    const int32_t * pSrcB,
    const uint32_t srcBLen,
    const uint8_t nPE,
    int32_t * pRes
)

Glue code for parallel convolution of 32-bit integer vectors.

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
• nPE Number of cores to compute on
• pRes output result returned here

Return: none

Attributes Documentation

variable resultsBuffer

HAL_CL_L1 int32_t * resultsBuffer;

Source code

/* =====================================================================
 * Project:      PULP DSP Library
 * Title:        plp_conv_i32_parallel.c
 * Description:  32-bit paralell integer convolution glue code
 *
 * $Date:        01. July 2019
 * $Revision:    V0
 *
 * Target Processor: PULP cores
 * ===================================================================== */
/*
 * Copyright (C) 2019 ETH Zurich and University of Bologna.
 *
 * Author: Moritz Scherer
 *
 * 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"
#include "rtos_hal.h"

HAL_CL_L1 int32_t *resultsBuffer;

//#define PLP_CONV_SEQUENTIALADDING 1

void plp_conv_i32_parallel(const int32_t *pSrcA,
                           const uint32_t srcALen,
                           const int32_t *pSrcB,
                           const uint32_t srcBLen,
                           const uint8_t nPE,
                           int32_t *pRes) {

    if (hal_cluster_id() == ARCHI_FC_CID) {
        printf("parallel processing supported only for cluster side\n");
        return;
    } else {

        if (nPE == 1) {
            plp_conv_i32(pSrcA, srcALen, pSrcB, srcBLen, pRes);
            return;
        }

        const int32_t *pIn1;
        const int32_t *pIn2;

        uint32_t pIn1Len;
        uint32_t pIn2Len;

        if (srcALen >= srcBLen) {
            pIn2 = pSrcA;
            pIn1 = pSrcB;
            pIn2Len = srcALen;
            pIn1Len = srcBLen;
        } else {
            pIn2 = pSrcB;
            pIn1 = pSrcA;
            pIn2Len = srcBLen;
            pIn1Len = srcALen;
        }

        uint32_t srcAoffset = ((pIn1Len + nPE - 1) / nPE);
        uint32_t resultsoffset = srcAoffset + pIn2Len - 1;
        uint32_t resultsLen =
            resultsoffset * (nPE - 1) + (pIn1Len - (srcAoffset * (nPE - 1))) + pIn2Len - 1;
        int32_t *resBuf;

        if (nPE > 1) {
            resultsBuffer =
                (int32_t *)hal_cl_l1_malloc(sizeof(int32_t) * resultsoffset * nPE);
            resBuf = resultsBuffer;
            for (uint32_t i = resultsLen; i < resultsoffset * nPE; i++) {
                resultsBuffer[i] = 0;
            }
            // printf("Address of resultsBuffer: 0x%x, End: 0x%x\n", resultsBuffer, resultsBuffer +
            // sizeof(int32_t)*resultsLen);
        } else {
            resultsBuffer = pRes;
        }

        plp_conv_instance_i32 S = { .srcALen = pIn1Len,
                                    .srcBLen = pIn2Len,
                                    .pSrcA = pIn1,
                                    .pSrcB = pIn2,
                                    .pRes = resultsBuffer,
                                    .nPE = nPE };

        hal_cl_team_fork(nPE, plp_conv_i32p_xpulpv2, (void *)&S);

        if (nPE > 1) {

            /* Sequential overlap-adding */

#if defined(PLP_CONV_SEQUENTIALADDING)

            for (uint32_t i = 0; i < resultsoffset; i++) {
                pRes[i] = resultsBuffer[i];
            }

            for (uint32_t i = resultsoffset; i < srcALen + srcBLen - 1; i++) {
                pRes[i] = 0;
            }

            for (int32_t i = 1; i < nPE - 1; i++) {
                for (uint32_t j = 0; j < resultsoffset; j++) {
                    pRes[i * srcAoffset + j] += resultsBuffer[j + i * resultsoffset];
                }
            }

            for (uint32_t j = 0; j < resultsLen - resultsoffset * (nPE - 1); j++) {
                pRes[(nPE - 1) * srcAoffset + j] += resultsBuffer[(nPE - 1) * resultsoffset + j];
            }

#else

            /* Parallel overlap-adding */
            plp_conv_parallel_OLA(nPE, pIn1Len, pIn2Len, resultsBuffer);

#if defined(PLP_MATH_LOOPUNROLL)

            uint32_t k = (srcALen + srcBLen - 1) >> 1U;
            int32_t temp1, temp2;

            while (k) {
                temp1 = *resultsBuffer++;
                temp2 = *resultsBuffer++;

                *pRes++ = temp1;
                *pRes++ = temp2;

                k--;
            }

            k = (srcALen + srcBLen - 1) % 0x2U;

            if (k) {
                *pRes++ = *resultsBuffer++;
            }

#else
            for (uint32_t i = 0; i < srcALen + srcBLen - 1; i++) {
                pRes[i] = resultsBuffer[i];
            }
#endif
            hal_cl_l1_free(resBuf, sizeof(int32_t) * resultsoffset * nPE);

#endif
        }
        return;
    }
}

---

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