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.

// -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(16, (unsigned char*)pixels);
      pixelsv1B = vec_ld(line_size, (unsigned char*)pixels);
      pixelsv2B = vec_ld(16 + line_size, (unsigned char*)pixels);
      pixelsv1C = vec_ld(line_size_2, (unsigned char*)pixels);
      pixelsv2C = vec_ld(16 + line_size_2, (unsigned char*)pixels);
      pixelsv1D = vec_ld(line_size_3, (unsigned char*)pixels);
      pixelsv2D = vec_ld(16 + 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);

#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

/* 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);
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int i;

POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);

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

POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);

#else /* ALTIVEC_USE_REFERENCE_C_CODE */
    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);

#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

/* 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);
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int i;
POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
    for (i = 0; i < h; i++) {
        *((uint32_t *) (block)) =
            (((*((uint32_t *) (block))) |
              ((((const struct unaligned_32 *) (pixels))->l))) -
             ((((*((uint32_t *) (block))) ^
                ((((const struct unaligned_32 *) (pixels))->
                  l))) & 0xFEFEFEFEUL) >> 1));
        *((uint32_t *) (block + 4)) =
            (((*((uint32_t *) (block + 4))) |
              ((((const struct unaligned_32 *) (pixels + 4))->l))) -
             ((((*((uint32_t *) (block + 4))) ^
                ((((const struct unaligned_32 *) (pixels +
                                                  4))->
                  l))) & 0xFEFEFEFEUL) >> 1));
        pixels += line_size;
        block += line_size;
    }
POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);

#else /* ALTIVEC_USE_REFERENCE_C_CODE */
    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);
 
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

/* 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);
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int j;
POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
    for (j = 0; j < 2; j++) {
      int i;
      const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
      const uint32_t b =
        (((const struct unaligned_32 *) (pixels + 1))->l);
      uint32_t l0 =
        (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
      uint32_t h0 =
        ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
      uint32_t l1, h1;
      pixels += line_size;
      for (i = 0; i < h; i += 2) {
        uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
        uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
        h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
        a = (((const struct unaligned_32 *) (pixels))->l);
        b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x02020202UL;
        h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
      } pixels += 4 - line_size * (h + 1);
      block += 4 - line_size * h;
    }

POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);

#else /* ALTIVEC_USE_REFERENCE_C_CODE */
   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);
   pixelsv2 = vec_mergeh(vczero, pixelsv2);
   pixelssum1 = vec_add((vector unsigned short)pixelsv1,
                        (vector unsigned short)pixelsv2);
   pixelssum1 = vec_add(pixelssum1, vctwo);
   
POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1); 
   for (i = 0; i < h ; i++) {
     int rightside = ((unsigned long)block & 0x0000000F);
     blockv = vec_ld(0, block);

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

     pixelsv1 = vec_mergeh(vczero, pixelsv1);
     pixelsv2 = vec_mergeh(vczero, pixelsv2);
     pixelssum2 = vec_add((vector unsigned short)pixelsv1,
                          (vector unsigned short)pixelsv2);
     temp3 = vec_add(pixelssum1, pixelssum2);
     temp3 = vec_sra(temp3, vctwo);
     pixelssum1 = vec_add(pixelssum2, vctwo);
     pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
     
     if (rightside)
     {
       blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
     }
     else
     {
       blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
     }
     
     vec_st(blockv, 0, block);
     
     block += line_size;
     pixels += line_size;
   }
   
POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
#endif /* ALTIVEC_USE_REFERENCE_C_CODE */
}

/* next one assumes that ((line_size % 8) == 0) */
void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
{
POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
#ifdef ALTIVEC_USE_REFERENCE_C_CODE
    int j;
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
    for (j = 0; j < 2; j++) {
      int i;
      const uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
      const uint32_t b =
        (((const struct unaligned_32 *) (pixels + 1))->l);
      uint32_t l0 =
        (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
      uint32_t h0 =
        ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
      uint32_t l1, h1;
      pixels += line_size;
      for (i = 0; i < h; i += 2) {
        uint32_t a = (((const struct unaligned_32 *) (pixels))->l);
        uint32_t b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l1 = (a & 0x03030303UL) + (b & 0x03030303UL);
        h1 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
        a = (((const struct unaligned_32 *) (pixels))->l);
        b = (((const struct unaligned_32 *) (pixels + 1))->l);
        l0 = (a & 0x03030303UL) + (b & 0x03030303UL) + 0x01010101UL;
        h0 = ((a & 0xFCFCFCFCUL) >> 2) + ((b & 0xFCFCFCFCUL) >> 2);
        *((uint32_t *) block) =
          h0 + h1 + (((l0 + l1) >> 2) & 0x0F0F0F0FUL);
        pixels += line_size;
        block += line_size;
      } pixels += 4 - line_size * (h + 1);
      block += 4 - line_size * h;
    }
    
POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);

#else /* ALTIVEC_USE_REFERENCE_C_CODE */
   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 vcone = (const_vector unsigned short)vec_splat_u16(1);
   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);
   pixelsv2 = vec_mergeh(vczero, pixelsv2);
   pixelssum1 = vec_add((vector unsigned short)pixelsv1,
                        (vector unsigned short)pixelsv2);
   pixelssum1 = vec_add(pixelssum1, vcone);
   
POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1); 
   for (i = 0; i < h ; i++) {
     int rightside = ((unsigned long)block & 0x0000000F);
     blockv = vec_ld(0, block);

     temp1 = vec_ld(line_size, pixels);
     temp2 = vec_ld(line_size + 16, pixels);
     pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
     if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F)
     {