display.cc

ac3的解码程序

                        "movq           %%mm4,                  %%mm5\n"  // lum1
                        "paddw          %%mm1,                  %%mm4\n"  // lum1 +red  00 R6 00 R4 00 R2 00 R0
                        "paddw          %%mm2,                  %%mm5\n"  // lum1 +green 00 G6 00 G4 00 G2 00 G0
                        "psraw          $6,                     %%mm4\n"  // R1 0 .. 64
                        "movq           %%mm7,                  %%mm3\n"  // lum2                       00 L7 00 L5 00 L3 00 L1
                        "psraw          $6,                     %%mm5\n"  // G1  - .. +
                        "paddw          %%mm0,                  %%mm7\n"  // Lum2 +blue 00 B7 00 B5 00 B3 00 B1
                        "psraw          $6,                     %%mm6\n"  // B1         0 .. 64
                        "packuswb       %%mm4,                  %%mm4\n"  // R1 R1
                        "packuswb       %%mm5,                  %%mm5\n"  // G1 G1
                        "packuswb       %%mm6,                  %%mm6\n"  // B1 B1
                        "punpcklbw      %%mm4,                  %%mm4\n"
                        "punpcklbw      %%mm5,                  %%mm5\n"

                        "pand           MMX_redmask,            %%mm4\n"
                        "psllw          $3,                     %%mm5\n"  // GREEN       1
                        "punpcklbw      %%mm6,                  %%mm6\n"
                        "pand           MMX_grnmask,            %%mm5\n"
                        "pand           MMX_redmask,            %%mm6\n"
                        "por            %%mm5,                  %%mm4\n" //
                        "psrlw          $11,                    %%mm6\n"                // BLUE        1
                        "movq           %%mm3,                  %%mm5\n" // lum2
                        "paddw          %%mm1,                  %%mm3\n"        // lum2 +red      00 R7 00 R5 00 R3 00 R1
                        "paddw          %%mm2,                  %%mm5\n" // lum2 +green 00 G7 00 G5 00 G3 00 G1
                        "psraw          $6,                     %%mm3\n" // R2
                        "por            %%mm6,                  %%mm4\n" // MM4
                        "psraw          $6,                     %%mm5\n" // G2
                        "movq           (%2, %3),               %%mm6\n"  // L3
                        "psraw          $6,                     %%mm7\n"
                        "packuswb       %%mm3,                  %%mm3\n"
                        "packuswb       %%mm5,                  %%mm5\n"
                        "packuswb       %%mm7,                  %%mm7\n"
                        "pand                   MMX_00FFw,              %%mm6\n"  // L3
                        "punpcklbw      %%mm3,                  %%mm3\n"
//                              "psubw          MMX_10w,                        %%mm6\n"  // L3
                        "punpcklbw      %%mm5,                  %%mm5\n"
                        "pmullw         MMX_Ycoeff,             %%mm6\n"  // lum3
                        "punpcklbw      %%mm7,                  %%mm7\n"
                        "psllw          $3,                             %%mm5\n"  // GREEN 2
                        "pand                   MMX_redmask,    %%mm7\n"
                        "pand                   MMX_redmask,    %%mm3\n"
                        "psrlw          $11,                            %%mm7\n"  // BLUE  2
                        "pand                   MMX_grnmask,    %%mm5\n"
                        "por                    %%mm7,                  %%mm3\n"
                        "movq                   (%2,%3),                        %%mm7\n"  // L4
                        "por                    %%mm5,                  %%mm3\n"     //
                        "psrlw          $8,                             %%mm7\n"    // L4
                        "movq                   %%mm4,                  %%mm5\n"
//                              "psubw          MMX_10w,                        %%mm7\n"                // L4
                        "punpcklwd      %%mm3,                  %%mm4\n"
                        "pmullw         MMX_Ycoeff,             %%mm7\n"    // lum4
                        "punpckhwd      %%mm3,                  %%mm5\n"

                        "movq                   %%mm4,                  (%4)\n"
                        "movq                   %%mm5,                  8(%4)\n"

                        "movq                   %%mm6,                  %%mm4\n"        // Lum3
                        "paddw          %%mm0,                  %%mm6\n"                // Lum3 +blue

                        "movq                   %%mm4,                  %%mm5\n"                        // Lum3
                        "paddw          %%mm1,                  %%mm4\n"       // Lum3 +red
                        "paddw          %%mm2,                  %%mm5\n"                        // Lum3 +green
                        "psraw          $6,                             %%mm4\n"
                        "movq                   %%mm7,                  %%mm3\n"                        // Lum4
                        "psraw          $6,                             %%mm5\n"
                        "paddw          %%mm0,                  %%mm7\n"                   // Lum4 +blue
                        "psraw          $6,                             %%mm6\n"                        // Lum3 +blue
                        "movq                   %%mm3,                  %%mm0\n"  // Lum4
                        "packuswb       %%mm4,                  %%mm4\n"
                        "paddw          %%mm1,                  %%mm3\n"  // Lum4 +red
                        "packuswb       %%mm5,                  %%mm5\n"
                        "paddw          %%mm2,                  %%mm0\n"         // Lum4 +green
                        "packuswb       %%mm6,                  %%mm6\n"
                        "punpcklbw      %%mm4,                  %%mm4\n"
                        "punpcklbw      %%mm5,                  %%mm5\n"
                        "punpcklbw      %%mm6,                  %%mm6\n"
                        "psllw          $3,                             %%mm5\n" // GREEN 3
                        "pand                   MMX_redmask,    %%mm4\n"
                        "psraw          $6,                             %%mm3\n" // psr 6
                        "psraw          $6,                             %%mm0\n"
                        "pand                   MMX_redmask,    %%mm6\n" // BLUE
                        "pand                   MMX_grnmask,    %%mm5\n"
                        "psrlw          $11,                            %%mm6\n"  // BLUE  3
                        "por                    %%mm5,                  %%mm4\n"
                        "psraw          $6,                             %%mm7\n"
                        "por                    %%mm6,                  %%mm4\n"
                        "packuswb       %%mm3,                  %%mm3\n"
                        "packuswb       %%mm0,                  %%mm0\n"
                        "packuswb       %%mm7,                  %%mm7\n"
                        "punpcklbw      %%mm3,                  %%mm3\n"
                        "punpcklbw      %%mm0,                  %%mm0\n"
                        "punpcklbw      %%mm7,                  %%mm7\n"
                        "pand                   MMX_redmask,    %%mm3\n"
                        "pand                   MMX_redmask,    %%mm7\n" // BLUE
                        "psllw          $3,                             %%mm0\n" // GREEN 4
                        "psrlw          $11,                            %%mm7\n"
                        "pand                   MMX_grnmask,    %%mm0\n"
                        "por                    %%mm7,                  %%mm3\n"
                        "addl                   $8,                             %6\n"
                        "por                    %%mm0,                  %%mm3\n"

                        "movq                   %%mm4,                  %%mm5\n"

                        "punpcklwd      %%mm3,                  %%mm4\n"
                        "punpckhwd      %%mm3,                  %%mm5\n"

                        "movq                   %%mm4,                  (%4,%5,2)\n"
                        "movq                   %%mm5,                  8(%4,%5,2)\n"

                        "addl                   $8,                             %2\n"
                        "addl                   $4,                             %0\n"
                        "addl                   $4,                             %1\n"
                        "cmpl                   %3,                             %6\n"
                        "leal                   16(%4),                 %4\n"
                "jl             1b\n"
                "addl           %3,     %2\n"                   /* lum += cols */
                "addl           %7,     %4\n"                   /* row1 += mod */
                "movl           $0,     %6\n"
                "cmpl           %8,     %2\n"
                "jl             1b\n"
                :
                : "r" (cr), "r" (cb), "r" (lum), "r" (cols), "r" (row1) ,"r" (col1), "m" (x), "m" (mod)
                        , "m" (y)
                );
        __asm__ __volatile__(
                "emms\n"
        );

}


int DisplayX11::dither_image_rgb16(unsigned char* src[], unsigned char* dithered_img){
  int height,width,display_width;
  height=coded_picture_height;
  if (shmem_flag == 2)
    display_width = xwidth;
  else
    display_width=coded_picture_width;
  width=coded_picture_width;
  if (rgb_mode == 0)
    Color16DitherImageMod(src[0], src[2], src[1], dithered_img, height, width,display_width -width);
  else
    yuv_2_rgb(src[0],src[1],src[2],width,height,width,width/2,0,dithered_img,0,0,display_width*2,rgb_mode);
  return 0;
}

int DisplayX11::dither_image_rgb24(unsigned char* src[], unsigned char* dithered_img){
  int height,width,display_width;
  height=coded_picture_height;
  if (shmem_flag == 2)
    display_width = xwidth;
  else
    display_width=coded_picture_width;
  width=coded_picture_width;
  if (rgb_mode == 0)
    Color16DitherImageMod(src[0], src[2], src[1], dithered_img, height, width,display_width -width);
  else
    yuv_2_rgb(src[0],src[1],src[2],width,height,width,width/2,0,dithered_img,0,0,display_width*2,rgb_mode);
  return 0;
}


#endif


void DisplayX11::dither_image(unsigned char *src[]){
#ifdef SOLARIS_SDK_XIL
  dithered_image=getImage1Data();
#endif
  if (prog_seq){
    if (chroma_format!=CHROMA444){
#if defined(LINUX) && defined(MMX)
      if (vinfo.c_class==TrueColor && vinfo.depth==24){
        dither_image_rgb24(src, dithered_image);
        return;
      }
      if (bpp == 16 ) {
        dither_image_rgb16(src, dithered_image);
        return;
      }
#endif
      ditherframe(src);
    }
    else
      ditherframe444(src);
  }
  else {
#ifdef SOLARIS_SDK_XIL
    dithered_image2=getImage2Data();
#endif
#if defined(LINUX) && defined(MMX) 
      if (bpp == 16 ){
        dither_image_rgb16(src, dithered_image);
        return;
      }
#endif
    if ((pict_struct==FRAME_PICTURE && topfirst) || pict_struct==BOTTOM_FIELD){
      /* top field first */
      if (chroma_format!=CHROMA444){
        dithertop(src,dithered_image);
        ditherbot(src,dithered_image2);
      }
      else {
        dithertop444(src,dithered_image);
        ditherbot444(src,dithered_image2);
      }
    }
    else {
      /* bottom field first */
      if (chroma_format!=CHROMA444){
        ditherbot(src,dithered_image);
        dithertop(src,dithered_image2);
      }
      else {
        ditherbot444(src,dithered_image);
        dithertop444(src,dithered_image2);
      }
    }
  }
}


/* only for 4:2:0 and 4:2:2! */

void DisplayX11::ditherframe(unsigned char *src[]){
  int i;
  register unsigned int uv;
  register unsigned char *py,*pu,*pv,*dst;
  py = src[0];
  pu = src[1];
  pv = src[2];
  dst = dithered_image;

  for (int j=0; j<coded_picture_height; j+=4){
    /* line j + 0 */
    for (i=0; i<coded_picture_width; i+=8){
      uv = uvtab[(*pu++<<8)|*pv++];
      *dst++ = ytab[((*py++)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +8)<<4)|(uv>>4)];
      uv = uvtab[((*pu++<<8)|*pv++)+1028];
      *dst++ = ytab[((*py++ +2)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +10)<<4)|(uv>>4)];
      uv = uvtab[(*pu++<<8)|*pv++];
      *dst++ = ytab[((*py++)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +8)<<4)|(uv>>4)];
      uv = uvtab[((*pu++<<8)|*pv++)+1028];
      *dst++ = ytab[((*py++ +2)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +10)<<4)|(uv>>4)];
    }

    if (chroma_format==CHROMA420){
      pu -= chrom_width;
      pv -= chrom_width;
    }

    /* line j + 1 */
    for (i=0; i<coded_picture_width; i+=8){
      uv = uvtab[((*pu++<<8)|*pv++)+2056];
      *dst++ = ytab[((*py++ +12)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +4)<<4)|(uv&15)];
      uv = uvtab[((*pu++<<8)|*pv++)+3084];
      *dst++ = ytab[((*py++ +14)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +6)<<4)|(uv&15)];
      uv = uvtab[((*pu++<<8)|*pv++)+2056];
      *dst++ = ytab[((*py++ +12)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +4)<<4)|(uv&15)];
      uv = uvtab[((*pu++<<8)|*pv++)+3084];
      *dst++ = ytab[((*py++ +14)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +6)<<4)|(uv&15)];
    }

    /* line j + 2 */
    for (i=0; i<coded_picture_width; i+=8){
      uv = uvtab[((*pu++<<8)|*pv++)+1542];
      *dst++ = ytab[((*py++ +3)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +11)<<4)|(uv>>4)];
      uv = uvtab[((*pu++<<8)|*pv++)+514];
      *dst++ = ytab[((*py++ +1)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +9)<<4)|(uv>>4)];
      uv = uvtab[((*pu++<<8)|*pv++)+1542];
      *dst++ = ytab[((*py++ +3)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +11)<<4)|(uv>>4)];
      uv = uvtab[((*pu++<<8)|*pv++)+514];
      *dst++ = ytab[((*py++ +1)<<4)|(uv&15)];
      *dst++ = ytab[((*py++ +9)<<4)|(uv>>4)];
    }

    if (chroma_format==CHROMA420){
      pu -= chrom_width;
      pv -= chrom_width;
    }

    /* line j + 3 */
    for (i=0; i<coded_picture_width; i+=8){
      uv = uvtab[((*pu++<<8)|*pv++)+3598];
      *dst++ = ytab[((*py++ +15)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +7)<<4)|(uv&15)];
      uv = uvtab[((*pu++<<8)|*pv++)+2570];
      *dst++ = ytab[((*py++ +13)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +5)<<4)|(uv&15)];
      uv = uvtab[((*pu++<<8)|*pv++)+3598];
      *dst++ = ytab[((*py++ +15)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +7)<<4)|(uv&15)];
      uv = uvtab[((*pu++<<8)|*pv++)+2570];
      *dst++ = ytab[((*py++ +13)<<4)|(uv>>4)];
      *dst++ = ytab[((*py++ +5)<<4)|(uv&15)];
    }
  }
}


void DisplayX11::dithertop(unsigned char *src[], unsigned char *dst){
  int i;
  unsigned int y,uv1,uv2;
  unsigned char *py,*py2,*pu,*pv,*dst2;

  py = src[0];
  py2 = src[0] + (coded_picture_width<<1);
  pu = src[1];
  pv = src[2];
  dst2 = dst + coded_picture_width;

  for (int j=0; j<coded_picture_height; j+=4){
    /* line j + 0, j + 1 */
    for (i=0; i<coded_picture_width; i+=4){
      y = *py++;
      uv2 = (*pu++<<8)|*pv++;
      uv1 = uvtab[uv2];
      uv2 = uvtab[uv2+2056];
      *dst++  = ytab[((y)<<4)|(uv1&15)];
      *dst2++ = ytab[((((y + *py2++)>>1)+12)<<4)|(uv2>>4)];

      y = *py++;
      *dst++  = ytab[((y+8)<<4)|(uv1>>4)];
      *dst2++ = ytab[((((y + *py2++)>>1)+4)<<4)|(uv2&15)];

      y = *py++;
      uv2 = (*pu++<<8)|*pv++;
      uv1 = uvtab[uv2+1028];
      uv2 = uvtab[uv2+3072];
      *dst++  = ytab[((y+2)<<4)|(uv1&15)];
      *dst2++ = ytab[((((y + *py2++)>>1)+14)<<4)|(uv2>>4)];

      y = *py++;
      *dst++  = ytab[((y+10)<<4)|(uv1>>4)];
      *dst2++ = ytab[((((y + *py2++)>>1)+6)<<4)|(uv2&15)];
    }

    py += coded_picture_width;

    if (j!=(coded_picture_height-4)) py2 += coded_picture_width;
    else                             py2 -= coded_picture_width;

    dst += coded_picture_width;
    dst2 += coded_picture_width;

    if (chroma_format==CHROMA420){
      pu -= chrom_width;
      pv -= chrom_width;
    }
    else {
      pu += chrom_width;
      pv += chrom_width;
    }

    /* line j + 2, j + 3 */
    for (i=0; i<coded_picture_width; i+=4){
      y = *py++;
      uv2 = (*pu++<<8)|*pv++;
      uv1 = uvtab[uv2+1542];
      uv2 = uvtab[uv2+3598];
      *dst++  = ytab[((y+3)<<4)|(uv1&15)];
      *dst2++ = ytab[((((y + *py2++)>>1)+15)<<4)|(uv2>>4)];