view pca.c @ 728:1fa3820b1a84 libavutil

ARM asm for AV_RN*() ARMv6 and later support unaligned loads and stores for single word/halfword but not double/multiple. GCC is ignorant of this and will always use bytewise accesses for unaligned data. Casting to an int32_t pointer is dangerous since a load/store double or multiple instruction might be used (this happens with some code in FFmpeg). Implementing the AV_[RW]* macros with inline asm using only supported instructions gives fast and safe unaligned accesses. ARM RVCT does the right thing with generic code. This gives an overall speedup of up to 10%.
author mru
date Sat, 18 Apr 2009 00:00:28 +0000
parents 5d344280a1f8
children 0795a743bda1
line wrap: on
line source

/*
 * principal component analysis (PCA)
 * Copyright (c) 2004 Michael Niedermayer <michaelni@gmx.at>
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file libavutil/pca.c
 * principal component analysis (PCA)
 */

#include "common.h"
#include "pca.h"

typedef struct PCA{
    int count;
    int n;
    double *covariance;
    double *mean;
}PCA;

PCA *ff_pca_init(int n){
    PCA *pca;
    if(n<=0)
        return NULL;

    pca= av_mallocz(sizeof(PCA));
    pca->n= n;
    pca->count=0;
    pca->covariance= av_mallocz(sizeof(double)*n*n);
    pca->mean= av_mallocz(sizeof(double)*n);

    return pca;
}

void ff_pca_free(PCA *pca){
    av_freep(&pca->covariance);
    av_freep(&pca->mean);
    av_free(pca);
}

void ff_pca_add(PCA *pca, double *v){
    int i, j;
    const int n= pca->n;

    for(i=0; i<n; i++){
        pca->mean[i] += v[i];
        for(j=i; j<n; j++)
            pca->covariance[j + i*n] += v[i]*v[j];
    }
    pca->count++;
}

int ff_pca(PCA *pca, double *eigenvector, double *eigenvalue){
    int i, j, pass;
    int k=0;
    const int n= pca->n;
    double z[n];

    memset(eigenvector, 0, sizeof(double)*n*n);

    for(j=0; j<n; j++){
        pca->mean[j] /= pca->count;
        eigenvector[j + j*n] = 1.0;
        for(i=0; i<=j; i++){
            pca->covariance[j + i*n] /= pca->count;
            pca->covariance[j + i*n] -= pca->mean[i] * pca->mean[j];
            pca->covariance[i + j*n] = pca->covariance[j + i*n];
        }
        eigenvalue[j]= pca->covariance[j + j*n];
        z[j]= 0;
    }

    for(pass=0; pass < 50; pass++){
        double sum=0;

        for(i=0; i<n; i++)
            for(j=i+1; j<n; j++)
                sum += fabs(pca->covariance[j + i*n]);

        if(sum == 0){
            for(i=0; i<n; i++){
                double maxvalue= -1;
                for(j=i; j<n; j++){
                    if(eigenvalue[j] > maxvalue){
                        maxvalue= eigenvalue[j];
                        k= j;
                    }
                }
                eigenvalue[k]= eigenvalue[i];
                eigenvalue[i]= maxvalue;
                for(j=0; j<n; j++){
                    double tmp= eigenvector[k + j*n];
                    eigenvector[k + j*n]= eigenvector[i + j*n];
                    eigenvector[i + j*n]= tmp;
                }
            }
            return pass;
        }

        for(i=0; i<n; i++){
            for(j=i+1; j<n; j++){
                double covar= pca->covariance[j + i*n];
                double t,c,s,tau,theta, h;

                if(pass < 3 && fabs(covar) < sum / (5*n*n)) //FIXME why pass < 3
                    continue;
                if(fabs(covar) == 0.0) //FIXME should not be needed
                    continue;
                if(pass >=3 && fabs((eigenvalue[j]+z[j])/covar) > (1LL<<32) && fabs((eigenvalue[i]+z[i])/covar) > (1LL<<32)){
                    pca->covariance[j + i*n]=0.0;
                    continue;
                }

                h= (eigenvalue[j]+z[j]) - (eigenvalue[i]+z[i]);
                theta=0.5*h/covar;
                t=1.0/(fabs(theta)+sqrt(1.0+theta*theta));
                if(theta < 0.0) t = -t;

                c=1.0/sqrt(1+t*t);
                s=t*c;
                tau=s/(1.0+c);
                z[i] -= t*covar;
                z[j] += t*covar;

#define ROTATE(a,i,j,k,l) {\
    double g=a[j + i*n];\
    double h=a[l + k*n];\
    a[j + i*n]=g-s*(h+g*tau);\
    a[l + k*n]=h+s*(g-h*tau); }
                for(k=0; k<n; k++) {
                    if(k!=i && k!=j){
                        ROTATE(pca->covariance,FFMIN(k,i),FFMAX(k,i),FFMIN(k,j),FFMAX(k,j))
                    }
                    ROTATE(eigenvector,k,i,k,j)
                }
                pca->covariance[j + i*n]=0.0;
            }
        }
        for (i=0; i<n; i++) {
            eigenvalue[i] += z[i];
            z[i]=0.0;
        }
    }

    return -1;
}

#ifdef TEST

#undef printf
#include <stdio.h>
#include <stdlib.h>
#include "lfg.h"

int main(void){
    PCA *pca;
    int i, j, k;
#define LEN 8
    double eigenvector[LEN*LEN];
    double eigenvalue[LEN];
    AVLFG prng;

    av_lfg_init(&prng, 1);

    pca= ff_pca_init(LEN);

    for(i=0; i<9000000; i++){
        double v[2*LEN+100];
        double sum=0;
        int pos = av_lfg_get(&prng) % LEN;
        int v2  = av_lfg_get(&prng) % 101 - 50;
        v[0]    = av_lfg_get(&prng) % 101 - 50;
        for(j=1; j<8; j++){
            if(j<=pos) v[j]= v[0];
            else       v[j]= v2;
            sum += v[j];
        }
/*        for(j=0; j<LEN; j++){
            v[j] -= v[pos];
        }*/
//        sum += av_lfg_get(&prng) % 10;
/*        for(j=0; j<LEN; j++){
            v[j] -= sum/LEN;
        }*/
//        lbt1(v+100,v+100,LEN);
        ff_pca_add(pca, v);
    }


    ff_pca(pca, eigenvector, eigenvalue);
    for(i=0; i<LEN; i++){
        pca->count= 1;
        pca->mean[i]= 0;

//        (0.5^|x|)^2 = 0.5^2|x| = 0.25^|x|


//        pca.covariance[i + i*LEN]= pow(0.5, fabs
        for(j=i; j<LEN; j++){
            printf("%f ", pca->covariance[i + j*LEN]);
        }
        printf("\n");
    }

#if 1
    for(i=0; i<LEN; i++){
        double v[LEN];
        double error=0;
        memset(v, 0, sizeof(v));
        for(j=0; j<LEN; j++){
            for(k=0; k<LEN; k++){
                v[j] += pca->covariance[FFMIN(k,j) + FFMAX(k,j)*LEN] * eigenvector[i + k*LEN];
            }
            v[j] /= eigenvalue[i];
            error += fabs(v[j] - eigenvector[i + j*LEN]);
        }
        printf("%f ", error);
    }
    printf("\n");
#endif
    for(i=0; i<LEN; i++){
        for(j=0; j<LEN; j++){
            printf("%9.6f ", eigenvector[i + j*LEN]);
        }
        printf("  %9.1f %f\n", eigenvalue[i], eigenvalue[i]/eigenvalue[0]);
    }

    return 0;
}
#endif