Skip to content

/github/workspace/src/FilteringFunctions/kernels/plp_conv_parallel_OLA_kernel.c

Functions

Name
void plp_conv_parallel_OLA_kernel(void * task_args)
Helper function for parallelized overlap-adding of partial convolution results.

Functions Documentation

function plp_conv_parallel_OLA_kernel

void plp_conv_parallel_OLA_kernel(
    void * task_args
)

Helper function for parallelized overlap-adding of partial convolution results.

Parameters:

Return: none

Source code

/* =====================================================================
 * Project:      PULP DSP Library
 * Title:        plp_conv_parallel_OLA.c
 * Description:  Parallel integer convolution for XPULPV2 using OLA method kernel
 *
 * $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"
#include "rtos_hal.h"

void plp_conv_parallel_OLA_kernel(void *task_args) {

    /* printf("Hello Core %i\n", hal_core_id()); */

    plp_conv_tree_add_instance *S = (plp_conv_tree_add_instance *)task_args;

    const uint8_t coreId = hal_core_id();

    const uint8_t coresPerVector = S->coresPerVector;
    const uint32_t addOffset = S->addOffset;
    const uint32_t addLengthfirst = S->addLengthfirst;
    const uint32_t numVectors = S->numVectors;
    const uint32_t addLengthsecond =
        (coreId >= ((numVectors >> 1) - 1) * coresPerVector && !(numVectors % 2))
            ? S->addLengthsecond
            : S->addLengthfirst;
    uint32_t addLength = (addLengthfirst >= addLengthsecond + addOffset)
                             ? addLengthsecond
                             : addLengthfirst - addOffset;
    uint32_t blockOffset = S->blockOffset; /* (addLengthsecond == addLengthfirst) ? S->blockOffset :
                                              2*(S->blockOffset) - addLengthsecond; */
    int32_t *pRes = (S->pRes + 2 * (coreId / coresPerVector) * blockOffset);

    uint32_t stepSize = (addLength + coresPerVector - 1) / coresPerVector;
    uint8_t shardId = coreId % coresPerVector;
    uint32_t sharedLength = (addLengthsecond - addLength + coresPerVector - 1) / coresPerVector;

#if defined(PLP_MATH_LOOPUNROLL)
    uint32_t k;
    int32_t temp1, temp2, temp3, temp4;
    int32_t *_pRes = (pRes + (shardId * stepSize) + addOffset);
    const int32_t *pIn1 = (pRes + (shardId * stepSize) + addOffset);
    const int32_t *pIn2 = (pRes + (shardId * stepSize) + blockOffset);
    const int32_t *pIn3 = (pRes + (shardId * sharedLength) + blockOffset + addLength);
    int32_t *_pRes2 = (pRes + (shardId * sharedLength) + addLength + addOffset);
#endif

    if (shardId != coresPerVector - 1) {

#if defined(PLP_MATH_LOOPUNROLL)

        k = stepSize >> 1U;

        while (k) {
            temp1 = *pIn1++;
            temp2 = *pIn1++;
            temp3 = *pIn2++;
            temp4 = *pIn2++;

            *_pRes++ = temp1 + temp3;
            *_pRes++ = temp2 + temp4;

            k--;
        }

        k = (stepSize) % 0x2U;

        if (k) {
            *_pRes = *pIn1 + *pIn2;
        }

        if (addLength < addLengthsecond && sharedLength > 0) {

            k = sharedLength >> 1U;

            while (k) {
                temp1 = *pIn3++;
                temp2 = *pIn3++;

                *_pRes2++ = temp1;
                *_pRes2++ = temp2;

                k--;
            }

            k = sharedLength % 0x2U;

            if (k) {
                *_pRes2 = *pIn3;
            }
        }

#else

        for (uint32_t i = shardId * stepSize; i < (shardId + 1) * stepSize; i++) {
            pRes[i + addOffset] += pRes[i + blockOffset];
        }

        if (addLength < addLengthsecond && sharedLength > 0) {
            for (uint32_t i = shardId * sharedLength; i < (shardId + 1) * sharedLength; i++) {
                pRes[i + addLength + addOffset] = pRes[i + blockOffset + addLength];
            }
        }
#endif // if defined(PLP_MATH_LOOPUNROLL)

    } else {

#if defined(PLP_MATH_LOOPUNROLL)

        k = (addLength - shardId * stepSize) >> 1U;

        while (k) {
            temp1 = *pIn1++;
            temp2 = *pIn1++;
            temp3 = *pIn2++;
            temp4 = *pIn2++;

            *_pRes++ = temp1 + temp3;
            *_pRes++ = temp2 + temp4;

            k--;
        }

        k = (addLength - shardId * stepSize) % 0x2U;

        if (k) {
            *_pRes = *pIn1 + *pIn2;
        }

        if (addLength < addLengthsecond) {

            k = ((addLengthsecond - addLength) - shardId * sharedLength) >> 1U;

            while (k) {
                temp1 = *pIn3++;
                temp2 = *pIn3++;

                *_pRes2++ = temp1;
                *_pRes2++ = temp2;

                k--;
            }

            k = ((addLengthsecond - addLength) - shardId * sharedLength) % 0x2U;

            if (k) {
                *_pRes2 = *pIn3;
            }
        }

        if (coreId == (numVectors >> 1) * coresPerVector - 1) {
            if (addLength >= addLengthsecond) {
                S->addLengthsecond = addLengthfirst;
            } else {
                S->addLengthsecond = addLengthfirst + addLengthsecond - addLength;
            }
        }

#else
        for (uint32_t i = shardId * stepSize; i < addLength; i++) {
            pRes[i + addOffset] += pRes[i + blockOffset];
        }

        if (addLength < addLengthsecond) {
            for (uint32_t i = shardId * sharedLength; i < (addLengthsecond - addLength); i++) {
                pRes[i + addLength + addOffset] = pRes[i + blockOffset + addLength];
            }
        }

        if (coreId == (numVectors >> 1) * coresPerVector - 1) {
            if (addLength >= addLengthsecond) {
                S->addLengthsecond = addLengthfirst;
            } else {
                S->addLengthsecond = addLengthfirst + addLengthsecond - addLength;
            }
        }

#endif // if defined(PLP_MATH_LOOPUNROLL)
    }
    hal_team_barrier();
    return;
}

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