Commit db418ea1 authored by Martin Perdacher's avatar Martin Perdacher
Browse files

changed dsyrk to dgemm

parent c0bf91c7

cmake/blasJoin/blasJoin.cpp

+22 −15
Original line number Diff line number Diff line
@@ -11,31 +11,32 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 


    *joinCounts = 0;



    if ( NUM_THREADS != 0 ){



    omp_set_num_threads(THREADS);

         mkl_set_num_threads(THREADS);

    }

    mkl_set_num_threads(1);



    double *p = (double*) ddr_alloc(sizeof(double) * N );



    // intermediate result

    iresult = (double*) ddr_alloc(BLOCKSIZE * BLOCKSIZE * sizeof(double));



    timer.start();



    const double EPS_SQUARED = (EPS * EPS) / 4.0 ;



    // pre-compute the values P(i) and P(j)

    // #pragma simd

    #pragma omp for

    for ( int i = 0 ; i < N ; i++ ){

        p[i] = EPS_SQUARED - cblas_ddot(D, &x[i*D], 1, &x[i*D], 1) / 2.0;

    }



    // intermediate result

    iresult = (double*) ddr_alloc(BLOCKSIZE * BLOCKSIZE * sizeof(double));

    omp_set_num_threads(THREADS);

    mkl_set_num_threads(THREADS);



    // divide N into NRC rows and NRC cols of size BLOCKSIZE^2

    size_t NRC = (size_t) ceil( (double) N / BLOCKSIZE);

    size_t blockRow = 0, blockCol=0;



    timer.start();



    // #pragma ivdep

    // #pragma vector always

    for ( int i = 1 ; i < NRC ; i++ ){

        blockRow = ( i + 1 == NRC ) ? N - (BLOCKSIZE * (NRC - 1) ) : BLOCKSIZE;

        for ( int j = 0 ; j < i ; j++ ){
@@ -53,7 +54,7 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 
            size_t accum=0;

            #pragma omp parallel for reduction(+:accum)

            for ( int k = 0 ; k < blockRow ; k++ ){

                for ( int l =0 ; l < k ; l++ ){

                for ( int l =0 ; l < blockCol ; l++ ){

                    /* with the 63-rshift we find out whether we are +/- and sum it up to efficiently determine count of join-partners

                     * prevent RAW-dependency with one-liner

                     * count only join-partners (positive ones)
@@ -68,20 +69,26 @@ double blasJoin(const double *x, const size_t N, const size_t D, const double EP 


    // main diagonal

    for ( int i = 0 ; i < NRC ; i++ ){

        blockRow = ( i + 1 == NRC ) ? N - (BLOCKSIZE * i ) : BLOCKSIZE;

        blockRow = ( i + 1 == NRC ) ? N - (BLOCKSIZE * (NRC - 1) ) : BLOCKSIZE;



        // perform symmetric matrix multiplication on lower triangle only

        // C := alpha*A*A' + beta*C



        cblas_dsyrk(CblasRowMajor, CblasLower, CblasNoTrans, blockRow, D, 1.0, x + BLOCKSIZE * D * i, D, 0.0, iresult, blockRow);

        /* note to dsyrk: not sure, why I do not get the correct result in the lower diagonal for very large matrices >8000, however dgemm seems to be faster */

        // cblas_dsyrk(CblasRowMajor, CblasLower, CblasNoTrans, blockRow, D, 1.0, &x[D*i*BLOCKSIZE], D, 0.0, iresult, blockRow);



        // C := alpha*A*B' + beta*C

        cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasTrans, blockRow, blockCol, D, 1.0, &x[i*BLOCKSIZE*D], D, &x[i*BLOCKSIZE*D], D, 0.0, iresult, blockCol);



        size_t accum=0;

        #pragma omp parallel for reduction(+:accum)

        for ( int k = 0 ; k < blockRow ; k++ ){

        for ( int k = 1 ; k < blockRow ; k++ ){

            for ( int l = 0 ; l < k ; l++ ){

                accum += - ( (long long) floor( - ( iresult[k * BLOCKSIZE + l] + p[i*BLOCKSIZE+k] + p[i*BLOCKSIZE+l] ) ) >> 63);

            }

        }

        *joinCounts += accum;



    }



    timer.stop();

cmake/blasJoin/blasJoin.h

+2 −0
Original line number Diff line number Diff line
@@ -12,6 +12,8 @@ 


#include "../util/allocation.h"

#include "../util/timer.h"

#include "../util/dataIo.h"





double blasJoin(const double *x, const size_t N, const size_t D, const double EPS, const unsigned int THREADS, const size_t BLOCKSIZE, size_t *joinCounts);