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

        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);

        /*
          Since we want to use unsigned chars, we can take advantage
          of the fact that abs(a-b)^2 = (a-b)^2.
        */

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

        /* Square the values and add them to our sum */
        sum = vec_msum(t5, t5, sum);

        pix1 += line_size;
        pix2 += line_size;
    }

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

    return s;
}

/**
 * Sum of Squared Errors for a 16x16 block.
 * AltiVec-enhanced.
 * It's the sad16_altivec code above w/ squaring added.
 */
int sse16_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 sum;
    vector signed int sumsqr;

    sum = (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);

        /*
          Since we want to use unsigned chars, we can take advantage
          of the fact that abs(a-b)^2 = (a-b)^2.
        */

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

        /* Square the values and add them to our sum */
        sum = vec_msum(t5, t5, sum);

        pix1 += line_size;
        pix2 += line_size;
    }

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

    return s;
}

int pix_sum_altivec(uint8_t * pix, int line_size)
{
    const vector unsigned int zero = (const vector unsigned int)vec_splat_u32(0);
    vector unsigned char perm, *pixv;
    vector unsigned char t1;
    vector unsigned int sad;
    vector signed int sumdiffs;

    int i;
    DECLARE_ALIGNED_16(int, s);

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

    for (i = 0; i < 16; i++) {
        /* Read the potentially unaligned 16 pixels into t1 */
        perm = vec_lvsl(0, pix);
        pixv = (vector unsigned char *) pix;
        t1 = vec_perm(pixv[0], pixv[1], perm);

        /* Add each 4 pixel group together and put 4 results into sad */
        sad = vec_sum4s(t1, sad);

        pix += 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;
}

void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
{
    int i;
    vector unsigned char perm, bytes, *pixv;
    const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
    vector signed short shorts;

    for(i=0;i<8;i++)
    {
        // Read potentially unaligned pixels.
        // We're reading 16 pixels, and actually only want 8,
        // but we simply ignore the extras.
        perm = vec_lvsl(0, pixels);
        pixv = (vector unsigned char *) pixels;
        bytes = vec_perm(pixv[0], pixv[1], perm);

        // convert the bytes into shorts
        shorts = (vector signed short)vec_mergeh(zero, bytes);

        // save the data to the block, we assume the block is 16-byte aligned
        vec_st(shorts, i*16, (vector signed short*)block);

        pixels += line_size;
    }
}

void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
        const uint8_t *s2, int stride)
{
    int i;
    vector unsigned char perm, bytes, *pixv;
    const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
    vector signed short shorts1, shorts2;

    for(i=0;i<4;i++)
    {
        // Read potentially unaligned pixels
        // We're reading 16 pixels, and actually only want 8,
        // but we simply ignore the extras.
        perm = vec_lvsl(0, s1);
        pixv = (vector unsigned char *) s1;
        bytes = vec_perm(pixv[0], pixv[1], perm);

        // convert the bytes into shorts
        shorts1 = (vector signed short)vec_mergeh(zero, bytes);

        // Do the same for the second block of pixels
        perm = vec_lvsl(0, s2);
        pixv = (vector unsigned char *) s2;
        bytes = vec_perm(pixv[0], pixv[1], perm);

        // convert the bytes into shorts
        shorts2 = (vector signed short)vec_mergeh(zero, bytes);

        // Do the subtraction
        shorts1 = vec_sub(shorts1, shorts2);

        // save the data to the block, we assume the block is 16-byte aligned
        vec_st(shorts1, 0, (vector signed short*)block);

        s1 += stride;
        s2 += stride;
        block += 8;


        // The code below is a copy of the code above... This is a manual
        // unroll.

        // Read potentially unaligned pixels
        // We're reading 16 pixels, and actually only want 8,
        // but we simply ignore the extras.
        perm = vec_lvsl(0, s1);
        pixv = (vector unsigned char *) s1;
        bytes = vec_perm(pixv[0], pixv[1], perm);

        // convert the bytes into shorts
        shorts1 = (vector signed short)vec_mergeh(zero, bytes);

        // Do the same for the second block of pixels
        perm = vec_lvsl(0, s2);
        pixv = (vector unsigned char *) s2;
        bytes = vec_perm(pixv[0], pixv[1], perm);

        // convert the bytes into shorts
        shorts2 = (vector signed short)vec_mergeh(zero, bytes);

        // Do the subtraction
        shorts1 = vec_sub(shorts1, shorts2);

        // save the data to the block, we assume the block is 16-byte aligned
        vec_st(shorts1, 0, (vector signed short*)block);

        s1 += stride;
        s2 += stride;
        block += 8;
    }
}

void add_bytes_altivec(uint8_t *dst, uint8_t *src, int w) {
    register int i;
    register vector unsigned char vdst, vsrc;

    /* dst and src are 16 bytes-aligned (guaranteed) */
    for(i = 0 ; (i + 15) < w ; i+=16)
    {
      vdst = vec_ld(i, (unsigned char*)dst);
      vsrc = vec_ld(i, (unsigned char*)src);
      vdst = vec_add(vsrc, vdst);
      vec_st(vdst, i, (unsigned char*)dst);
    }
    /* if w is not a multiple of 16 */
    for (; (i < w) ; i++)
    {
      dst[i] = src[i];
    }
}

/* next one assumes that ((line_size % 16) == 0) */
void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
POWERPC_PERF_DECLARE(altivec_put_pixels16_num, 1);
    register vector unsigned char pixelsv1, pixelsv2;
    register vector unsigned char pixelsv1B, pixelsv2B;
    register vector unsigned char pixelsv1C, pixelsv2C;
    register vector unsigned char pixelsv1D, pixelsv2D;

    register vector unsigned char perm = vec_lvsl(0, pixels);
    int i;
    register int line_size_2 = line_size << 1;
    register int line_size_3 = line_size + line_size_2;
    register int line_size_4 = line_size << 2;

POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
// hand-unrolling the loop by 4 gains about 15%
// mininum execution time goes from 74 to 60 cycles
// it's faster than -funroll-loops, but using
// -funroll-loops w/ this is bad - 74 cycles again.
// all this is on a 7450, tuning for the 7450
#if 0
    for(i=0; i<h; i++) {
      pixelsv1 = vec_ld(0, (unsigned char*)pixels);
      pixelsv2 = vec_ld(16, (unsigned char*)pixels);
      vec_st(vec_perm(pixelsv1, pixelsv2, perm),
             0, (unsigned char*)block);
      pixels+=line_size;
      block +=line_size;
    }
#else
    for(i=0; i<h; i+=4) {
      pixelsv1 = vec_ld(0, (unsigned char*)pixels);
      pixelsv2 = vec_ld(15, (unsigned char*)pixels);
      pixelsv1B = vec_ld(line_size, (unsigned char*)pixels);
      pixelsv2B = vec_ld(15 + line_size, (unsigned char*)pixels);
      pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels);
      pixelsv2C = vec_ld(15 + line_size_2, (unsigned char*)pixels);
      pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels);
      pixelsv2D = vec_ld(15 + line_size_3, (unsigned char*)pixels);
      vec_st(vec_perm(pixelsv1, pixelsv2, perm),
             0, (unsigned char*)block);
      vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
             line_size, (unsigned char*)block);
      vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
             line_size_2, (unsigned char*)block);
      vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
             line_size_3, (unsigned char*)block);
      pixels+=line_size_4;
      block +=line_size_4;
    }
#endif
POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
}

/* next one assumes that ((line_size % 16) == 0) */
#define op_avg(a,b)  a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
POWERPC_PERF_DECLARE(altivec_avg_pixels16_num, 1);
    register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
    register vector unsigned char perm = vec_lvsl(0, pixels);
    int i;

POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);

    for(i=0; i<h; i++) {
      pixelsv1 = vec_ld(0, (unsigned char*)pixels);
      pixelsv2 = vec_ld(16, (unsigned char*)pixels);
      blockv = vec_ld(0, block);
      pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
      blockv = vec_avg(blockv,pixelsv);
      vec_st(blockv, 0, (unsigned char*)block);
      pixels+=line_size;
      block +=line_size;
    }

POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
}

/* next one assumes that ((line_size % 8) == 0) */
void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
{
POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1);
    register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
    int i;

POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);

   for (i = 0; i < h; i++) {
     /*
       block is 8 bytes-aligned, so we're either in the
       left block (16 bytes-aligned) or in the right block (not)
     */
     int rightside = ((unsigned long)block & 0x0000000F);

     blockv = vec_ld(0, block);
     pixelsv1 = vec_ld(0, (unsigned char*)pixels);
     pixelsv2 = vec_ld(16, (unsigned char*)pixels);
     pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));

     if (rightside)
     {
       pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
     }
     else
     {
       pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
     }

     blockv = vec_avg(blockv, pixelsv);

     vec_st(blockv, 0, block);

     pixels += line_size;
     block += line_size;
   }

POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
}

/* next one assumes that ((line_size % 8) == 0) */
void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
POWERPC_PERF_DECLARE(altivec_put_pixels8_xy2_num, 1);
   register int i;
   register vector unsigned char
     pixelsv1, pixelsv2,
     pixelsavg;
   register vector unsigned char
     blockv, temp1, temp2;
   register vector unsigned short
     pixelssum1, pixelssum2, temp3;
   register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
   register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);

   temp1 = vec_ld(0, pixels);
   temp2 = vec_ld(16, pixels);
   pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F)
   {
     pixelsv2 = temp2;
   }
   else
   {
     pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   }
   pixelsv1 = vec_mergeh(vczero, pixelsv1);