dsputil_altivec.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

    return s;
}

int pix_norm1_altivec(uint8_t *pix, int line_size)
{
    int i;
    int s __attribute__((aligned(16)));
    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;
    int s __attribute__((aligned(16)));
    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);
        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;
    int s __attribute__((aligned(16)));
    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;
    int s __attribute__((aligned(16)));
    
    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) {
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int i;
    for(i=0; i+7<w; i++){
        dst[i+0] += src[i+0];
        dst[i+1] += src[i+1];
        dst[i+2] += src[i+2];
        dst[i+3] += src[i+3];
        dst[i+4] += src[i+4];
        dst[i+5] += src[i+5];
        dst[i+6] += src[i+6];
        dst[i+7] += src[i+7];
    }
    for(; i<w; i++)
        dst[i+0] += src[i+0];
#else /* ALTIVEC_USE_REFERENCE_C_CODE */
    register int i;
    register vector unsigned char vdst, vsrc;
    
    /* dst and src are 16 bytes-aligned (guaranteed) */
    for(i = 0 ; (i + 15) < w ; i++)
    {
      vdst = vec_ld(i << 4, (unsigned char*)dst);
      vsrc = vec_ld(i << 4, (unsigned char*)src);
      vdst = vec_add(vsrc, vdst);
      vec_st(vdst, i << 4, (unsigned char*)dst);
    }
    /* if w is not a multiple of 16 */
    for (; (i < w) ; i++)
    {
      dst[i] = src[i];
    }
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

/* 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);
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int i;

POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);

    for(i=0; i<h; i++) {
      *((uint32_t*)(block)) = LD32(pixels);
      *((uint32_t*)(block+4)) = LD32(pixels+4);
      *((uint32_t*)(block+8)) = LD32(pixels+8);
      *((uint32_t*)(block+12)) = LD32(pixels+12);
      pixels+=line_size;
      block +=line_size;
    }

POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);

#else /* ALTIVEC_USE_REFERENCE_C_CODE */
    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