dsputil_altivec.c

ffmpeg的完整源代码和作者自己写的文档。不但有在Linux的工程哦

/*
 * Copyright (c) 2002 Brian Foley
 * Copyright (c) 2002 Dieter Shirley
 * Copyright (c) 2003-2004 Romain Dolbeau <[e-mail]>
 *
 * 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
 */

#include "dsputil.h"

#include "gcc_fixes.h"

#include "dsputil_altivec.h"
#include "util_altivec.h"

int sad16_x2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
    int i;
    DECLARE_ALIGNED_16(int, s);
    const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
    vector unsigned char *tv;
    vector unsigned char pix1v, pix2v, pix2iv, avgv, t5;
    vector unsigned int sad;
    vector signed int sumdiffs;

    s = 0;
    sad = (vector unsigned int)vec_splat_u32(0);
    for(i=0;i<h;i++) {
        /*
           Read unaligned pixels into our vectors. The vectors are as follows:
           pix1v: pix1[0]-pix1[15]
           pix2v: pix2[0]-pix2[15]      pix2iv: pix2[1]-pix2[16]
        */
        tv = (vector unsigned char *) pix1;
        pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));

        tv = (vector unsigned char *) &pix2[0];
        pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));

        tv = (vector unsigned char *) &pix2[1];
        pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));

        /* Calculate the average vector */
        avgv = vec_avg(pix2v, pix2iv);

        /* Calculate a sum of abs differences vector */
        t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));

        /* Add each 4 pixel group together and put 4 results into sad */
        sad = vec_sum4s(t5, sad);

        pix1 += line_size;
        pix2 += line_size;
    }
    /* Sum up the four partial sums, and put the result into s */
    sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
    sumdiffs = vec_splat(sumdiffs, 3);
    vec_ste(sumdiffs, 0, &s);

    return s;
}

int sad16_y2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
    int i;
    DECLARE_ALIGNED_16(int, s);
    const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
    vector unsigned char *tv;
    vector unsigned char pix1v, pix2v, pix3v, avgv, t5;
    vector unsigned int sad;
    vector signed int sumdiffs;
    uint8_t *pix3 = pix2 + line_size;

    s = 0;
    sad = (vector unsigned int)vec_splat_u32(0);

    /*
       Due to the fact that pix3 = pix2 + line_size, the pix3 of one
       iteration becomes pix2 in the next iteration. We can use this
       fact to avoid a potentially expensive unaligned read, each
       time around the loop.
       Read unaligned pixels into our vectors. The vectors are as follows:
       pix2v: pix2[0]-pix2[15]
       Split the pixel vectors into shorts
    */
    tv = (vector unsigned char *) &pix2[0];
    pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));

    for(i=0;i<h;i++) {
        /*
           Read unaligned pixels into our vectors. The vectors are as follows:
           pix1v: pix1[0]-pix1[15]
           pix3v: pix3[0]-pix3[15]
        */
        tv = (vector unsigned char *) pix1;
        pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));

        tv = (vector unsigned char *) &pix3[0];
        pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));

        /* Calculate the average vector */
        avgv = vec_avg(pix2v, pix3v);

        /* Calculate a sum of abs differences vector */
        t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));

        /* Add each 4 pixel group together and put 4 results into sad */
        sad = vec_sum4s(t5, sad);

        pix1 += line_size;
        pix2v = pix3v;
        pix3 += line_size;

    }

    /* Sum up the four partial sums, and put the result into s */
    sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
    sumdiffs = vec_splat(sumdiffs, 3);
    vec_ste(sumdiffs, 0, &s);
    return s;
}

int sad16_xy2_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
    int i;
    DECLARE_ALIGNED_16(int, s);
    uint8_t *pix3 = pix2 + line_size;
    const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
    const vector unsigned short two = (const vector unsigned short)vec_splat_u16(2);
    vector unsigned char *tv, avgv, t5;
    vector unsigned char pix1v, pix2v, pix3v, pix2iv, pix3iv;
    vector unsigned short pix2lv, pix2hv, pix2ilv, pix2ihv;
    vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
    vector unsigned short avghv, avglv;
    vector unsigned short t1, t2, t3, t4;
    vector unsigned int sad;
    vector signed int sumdiffs;

    sad = (vector unsigned int)vec_splat_u32(0);

    s = 0;

    /*
       Due to the fact that pix3 = pix2 + line_size, the pix3 of one
       iteration becomes pix2 in the next iteration. We can use this
       fact to avoid a potentially expensive unaligned read, as well
       as some splitting, and vector addition each time around the loop.
       Read unaligned pixels into our vectors. The vectors are as follows:
       pix2v: pix2[0]-pix2[15]  pix2iv: pix2[1]-pix2[16]
       Split the pixel vectors into shorts
    */
    tv = (vector unsigned char *) &pix2[0];
    pix2v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[0]));

    tv = (vector unsigned char *) &pix2[1];
    pix2iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix2[1]));

    pix2hv = (vector unsigned short) vec_mergeh(zero, pix2v);
    pix2lv = (vector unsigned short) vec_mergel(zero, pix2v);
    pix2ihv = (vector unsigned short) vec_mergeh(zero, pix2iv);
    pix2ilv = (vector unsigned short) vec_mergel(zero, pix2iv);
    t1 = vec_add(pix2hv, pix2ihv);
    t2 = vec_add(pix2lv, pix2ilv);

    for(i=0;i<h;i++) {
        /*
           Read unaligned pixels into our vectors. The vectors are as follows:
           pix1v: pix1[0]-pix1[15]
           pix3v: pix3[0]-pix3[15]      pix3iv: pix3[1]-pix3[16]
        */
        tv = (vector unsigned char *) pix1;
        pix1v = vec_perm(tv[0], tv[1], vec_lvsl(0, pix1));

        tv = (vector unsigned char *) &pix3[0];
        pix3v = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[0]));

        tv = (vector unsigned char *) &pix3[1];
        pix3iv = vec_perm(tv[0], tv[1], vec_lvsl(0, &pix3[1]));

        /*
          Note that AltiVec does have vec_avg, but this works on vector pairs
          and rounds up. We could do avg(avg(a,b),avg(c,d)), but the rounding
          would mean that, for example, avg(3,0,0,1) = 2, when it should be 1.
          Instead, we have to split the pixel vectors into vectors of shorts,
          and do the averaging by hand.
        */

        /* Split the pixel vectors into shorts */
        pix3hv = (vector unsigned short) vec_mergeh(zero, pix3v);
        pix3lv = (vector unsigned short) vec_mergel(zero, pix3v);
        pix3ihv = (vector unsigned short) vec_mergeh(zero, pix3iv);
        pix3ilv = (vector unsigned short) vec_mergel(zero, pix3iv);

        /* Do the averaging on them */
        t3 = vec_add(pix3hv, pix3ihv);
        t4 = vec_add(pix3lv, pix3ilv);

        avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
        avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);

        /* Pack the shorts back into a result */
        avgv = vec_pack(avghv, avglv);

        /* Calculate a sum of abs differences vector */
        t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));

        /* Add each 4 pixel group together and put 4 results into sad */
        sad = vec_sum4s(t5, sad);

        pix1 += line_size;
        pix3 += line_size;
        /* Transfer the calculated values for pix3 into pix2 */
        t1 = t3;
        t2 = t4;
    }
    /* Sum up the four partial sums, and put the result into s */
    sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
    sumdiffs = vec_splat(sumdiffs, 3);
    vec_ste(sumdiffs, 0, &s);

    return s;
}

int sad16_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
    int i;
    DECLARE_ALIGNED_16(int, s);
    const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
    vector unsigned char perm1, perm2, *pix1v, *pix2v;
    vector unsigned char t1, t2, t3,t4, t5;
    vector unsigned int sad;
    vector signed int sumdiffs;

    sad = (vector unsigned int)vec_splat_u32(0);


    for(i=0;i<h;i++) {
        /* Read potentially unaligned pixels into t1 and t2 */
        perm1 = vec_lvsl(0, pix1);
        pix1v = (vector unsigned char *) pix1;
        perm2 = vec_lvsl(0, pix2);
        pix2v = (vector unsigned char *) pix2;
        t1 = vec_perm(pix1v[0], pix1v[1], perm1);
        t2 = vec_perm(pix2v[0], pix2v[1], perm2);

        /* Calculate a sum of abs differences vector */
        t3 = vec_max(t1, t2);
        t4 = vec_min(t1, t2);
        t5 = vec_sub(t3, t4);

        /* Add each 4 pixel group together and put 4 results into sad */
        sad = vec_sum4s(t5, sad);

        pix1 += line_size;
        pix2 += line_size;
    }

    /* Sum up the four partial sums, and put the result into s */
    sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
    sumdiffs = vec_splat(sumdiffs, 3);
    vec_ste(sumdiffs, 0, &s);

    return s;
}

int sad8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
    int i;
    DECLARE_ALIGNED_16(int, s);
    const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
    vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
    vector unsigned char t1, t2, t3,t4, t5;
    vector unsigned int sad;
    vector signed int sumdiffs;

    sad = (vector unsigned int)vec_splat_u32(0);

    permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);

    for(i=0;i<h;i++) {
        /* Read potentially unaligned pixels into t1 and t2
           Since we're reading 16 pixels, and actually only want 8,
           mask out the last 8 pixels. The 0s don't change the sum. */
        perm1 = vec_lvsl(0, pix1);
        pix1v = (vector unsigned char *) pix1;
        perm2 = vec_lvsl(0, pix2);
        pix2v = (vector unsigned char *) pix2;
        t1 = vec_and(vec_perm(pix1v[0], pix1v[1], perm1), permclear);
        t2 = vec_and(vec_perm(pix2v[0], pix2v[1], perm2), permclear);

        /* Calculate a sum of abs differences vector */
        t3 = vec_max(t1, t2);
        t4 = vec_min(t1, t2);
        t5 = vec_sub(t3, t4);

        /* Add each 4 pixel group together and put 4 results into sad */
        sad = vec_sum4s(t5, sad);

        pix1 += line_size;
        pix2 += line_size;
    }

    /* Sum up the four partial sums, and put the result into s */
    sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
    sumdiffs = vec_splat(sumdiffs, 3);
    vec_ste(sumdiffs, 0, &s);

    return s;
}

int pix_norm1_altivec(uint8_t *pix, int line_size)
{
    int i;
    DECLARE_ALIGNED_16(int, s);
    const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
    vector unsigned char *tv;
    vector unsigned char pixv;
    vector unsigned int sv;
    vector signed int sum;

    sv = (vector unsigned int)vec_splat_u32(0);

    s = 0;
    for (i = 0; i < 16; i++) {
        /* Read in the potentially unaligned pixels */
        tv = (vector unsigned char *) pix;
        pixv = vec_perm(tv[0], tv[1], vec_lvsl(0, pix));

        /* Square the values, and add them to our sum */
        sv = vec_msum(pixv, pixv, sv);

        pix += line_size;
    }
    /* Sum up the four partial sums, and put the result into s */
    sum = vec_sums((vector signed int) sv, (vector signed int) zero);
    sum = vec_splat(sum, 3);
    vec_ste(sum, 0, &s);

    return s;
}

/**
 * Sum of Squared Errors for a 8x8 block.
 * AltiVec-enhanced.
 * It's the sad8_altivec code above w/ squaring added.
 */
int sse8_altivec(void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)
{
    int i;
    DECLARE_ALIGNED_16(int, s);
    const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
    vector unsigned char perm1, perm2, permclear, *pix1v, *pix2v;
    vector unsigned char t1, t2, t3,t4, t5;
    vector unsigned int sum;
    vector signed int sumsqr;

    sum = (vector unsigned int)vec_splat_u32(0);

    permclear = (vector unsigned char)AVV(255,255,255,255,255,255,255,255,0,0,0,0,0,0,0,0);


    for(i=0;i<h;i++) {
        /* Read potentially unaligned pixels into t1 and t2
           Since we're reading 16 pixels, and actually only want 8,
           mask out the last 8 pixels. The 0s don't change the sum. */
        perm1 = vec_lvsl(0, pix1);