开通博客赚积分
发布资源赚积分
分类

源码开发语言/平台
当前位置：首页 > 源码/资料 > 多媒体 > DVD > 查看源码
ditherer_mmx16.cc
资源名称：dvd-munitions.tar.gz　[点击查看]
上传用户：aoeyumen
上传日期：2007-01-06
资源大小：3329k
文件大小：12k
源码类别：

DVD

开发平台：

Unix_Linux

ditherer_mmx16.cc：源码内容
  1. #ifdef HAVE_MMX
  3. #include "ditherer_mmx16.hh"
  5. static long long MMX_0 = 0L;
  6. static unsigned long  MMX_10w[]         = {0x00100010, 0x00100010};                     //dd    00010 0010h, 000100010h
  7. static unsigned long  MMX_80w[]         = {0x00800080, 0x00800080};                     //dd    00080 0080h, 000800080h
  9. static unsigned long  MMX_00FFw[]       = {0x00ff00ff, 0x00ff00ff};                     //dd    000FF 00FFh, 000FF00FFh
  11. static unsigned short MMX_Ublucoeff[]   = {0x81, 0x81, 0x81, 0x81};                     //dd    00081 0081h, 000810081h
  12. static unsigned short MMX_Vredcoeff[]   = {0x66, 0x66, 0x66, 0x66};                     //dd    00066 0066h, 000660066h
  14. static unsigned short MMX_Ugrncoeff[]   = {0xffe8, 0xffe8, 0xffe8, 0xffe8};             //dd    0FFE7 FFE7h, 0FFE7FFE7h
  15. static unsigned short MMX_Vgrncoeff[]   = {0xffcd, 0xffcd, 0xffcd, 0xffcd};             //dd    0FFCC FFCCh, 0FFCCFFCCh
  17. static unsigned short MMX_Ycoeff[]      = {0x4a, 0x4a, 0x4a, 0x4a};                     //dd    0004A 004Ah, 0004A004Ah
  18. static unsigned short MMX_redmask[]     = {0xf800, 0xf800, 0xf800, 0xf800};             //dd    07c00 7c00h, 07c007c00h
  20. static unsigned short MMX_grnmask[]     = {0x7e0, 0x7e0, 0x7e0, 0x7e0};                 //dd    003e0 03e0h, 003e003e0h
  21. // static unsigned short MMX_blumask[]  = {0x1f, 0x1f, 0x1f, 0x1f};                     //dd    0001f 001fh, 0001f001fh
  23. void  Dither_MMX16::ditherBlock(unsigned char *lum, unsigned char *cb, unsigned char *cr,
  24.                      unsigned char *out,
  25.                      int cols, int rows, int screen_width) {
  26.       unsigned short *row1;
  27.       int x;
  28.       unsigned char *y;
  29.       int col1;
  30.       int mod = screen_width-cols;
  32.       row1 = (unsigned short *)out;
  33.       col1 = cols +mod;
  34.       mod += cols +mod;
  35.       mod *=2;
  36.       y = lum +cols*rows;
  37.       x = 0;
  38.       __asm__ __volatile__(
  39.          ".align 8n"
  40.          "1:n"
  41.          "movd           (%1),                   %%mm0n"        // 4 Cb         0  0  0  0 u3 u2 u1 u0
  42.          "pxor           %%mm7,                  %%mm7n"
  43.          "movd           (%0),                   %%mm1n" // 4 Cr                0  0  0  0 v3 v2 v1 v0
  44.          "punpcklbw      %%mm7,                  %%mm0n" // 4 W cb   0 u3  0 u2  0 u1  0 u0
  45.          "punpcklbw      %%mm7,                  %%mm1n" // 4 W cr   0 v3  0 v2  0 v1  0 v0
  46.          "psubw          MMX_80w,                %%mm0n"
  47.          "psubw          MMX_80w,                %%mm1n"
  48.          "movq           %%mm0,                  %%mm2n"        // Cb                   0 u3  0 u2  0 u1  0 u0
  49.          "movq           %%mm1,                  %%mm3n" // Cr
  50.          "pmullw         MMX_Ugrncoeff,          %%mm2n" // Cb2green 0 R3  0 R2  0 R1  0 R0
  51.          "movq           (%2),                   %%mm6n"        // L1      l7 L6 L5 L4 L3 L2 L1 L0
  52.          "pmullw         MMX_Ublucoeff,          %%mm0n" // Cb2blue
  53.          "pand           MMX_00FFw,              %%mm6n" // L1      00 L6 00 L4 00 L2 00 L0
  54.          "pmullw         MMX_Vgrncoeff,          %%mm3n" // Cr2green
  55.          "movq           (%2),                   %%mm7n" // L2
  56.          "pmullw         MMX_Vredcoeff,          %%mm1n" // Cr2red
  57.          //                      "psubw          MMX_10w,                %%mm6n"
  58.          "psrlw          $8,                     %%mm7n"        // L2           00 L7 00 L5 00 L3 00 L1
  59.          "pmullw         MMX_Ycoeff,             %%mm6n" // lum1
  60.          //                      "psubw          MMX_10w,                %%mm7n" // L2
  61.          "paddw          %%mm3,                  %%mm2n" // Cb2green + Cr2green == green
  62.          "pmullw         MMX_Ycoeff,             %%mm7n"  // lum2
  64.          "movq           %%mm6,                  %%mm4n"  // lum1
  65.          "paddw          %%mm0,                  %%mm6n"  // lum1 +blue 00 B6 00 B4 00 B2 00 B0
  66.          "movq           %%mm4,                  %%mm5n"  // lum1
  67.          "paddw          %%mm1,                  %%mm4n"  // lum1 +red  00 R6 00 R4 00 R2 00 R0
  68.          "paddw          %%mm2,                  %%mm5n"  // lum1 +green 00 G6 00 G4 00 G2 00 G0
  69.          "psraw          $6,                     %%mm4n"  // R1 0 .. 64
  70.          "movq           %%mm7,                  %%mm3n"  // lum2                       00 L7 00 L5 00 L3 00 L1
  71.          "psraw          $6,                     %%mm5n"  // G1  - .. +
  72.          "paddw          %%mm0,                  %%mm7n"  // Lum2 +blue 00 B7 00 B5 00 B3 00 B1
  73.          "psraw          $6,                     %%mm6n"  // B1         0 .. 64
  74.          "packuswb       %%mm4,                  %%mm4n"  // R1 R1
  75.          "packuswb       %%mm5,                  %%mm5n"  // G1 G1
  76.          "packuswb       %%mm6,                  %%mm6n"  // B1 B1
  77.          "punpcklbw      %%mm4,                  %%mm4n"
  78.          "punpcklbw      %%mm5,                  %%mm5n"
  80.          "pand           MMX_redmask,            %%mm4n"
  81.          "psllw          $3,                     %%mm5n"  // GREEN       1
  82.          "punpcklbw      %%mm6,                  %%mm6n"
  83.          "pand           MMX_grnmask,            %%mm5n"
  84.          "pand           MMX_redmask,            %%mm6n"
  85.          "por            %%mm5,                  %%mm4n" //
  86.          "psrlw          $11,                    %%mm6n"                // BLUE        1
  87.          "movq           %%mm3,                  %%mm5n" // lum2
  88.          "paddw          %%mm1,                  %%mm3n"        // lum2 +red      00 R7 00 R5 00 R3 00 R1
  89.          "paddw          %%mm2,                  %%mm5n" // lum2 +green 00 G7 00 G5 00 G3 00 G1
  90.          "psraw          $6,                     %%mm3n" // R2
  91.          "por            %%mm6,                  %%mm4n" // MM4
  92.          "psraw          $6,                     %%mm5n" // G2
  93.          "movq           (%2, %3),               %%mm6n"  // L3
  94.          "psraw          $6,                     %%mm7n"
  95.          "packuswb       %%mm3,                  %%mm3n"
  96.          "packuswb       %%mm5,                  %%mm5n"
  97.          "packuswb       %%mm7,                  %%mm7n"
  98.          "pand                   MMX_00FFw,              %%mm6n"  // L3
  99.          "punpcklbw      %%mm3,                  %%mm3n"
  100.          //                              "psubw          MMX_10w,                        %%mm6n"  // L3
  101.          "punpcklbw      %%mm5,                  %%mm5n"
  102.          "pmullw         MMX_Ycoeff,             %%mm6n"  // lum3
  103.          "punpcklbw      %%mm7,                  %%mm7n"
  104.          "psllw          $3,                             %%mm5n"  // GREEN 2
  105.          "pand                   MMX_redmask,    %%mm7n"
  106.          "pand                   MMX_redmask,    %%mm3n"
  107.          "psrlw          $11,                            %%mm7n"  // BLUE  2
  108.          "pand                   MMX_grnmask,    %%mm5n"
  109.          "por                    %%mm7,                  %%mm3n"
  110.          "movq                   (%2,%3),                        %%mm7n"  // L4
  111.          "por                    %%mm5,                  %%mm3n"     //
  112.          "psrlw          $8,                             %%mm7n"    // L4
  113.          "movq                   %%mm4,                  %%mm5n"
  114.          //                              "psubw          MMX_10w,                        %%mm7n"                // L4
  115.          "punpcklwd      %%mm3,                  %%mm4n"
  116.          "pmullw         MMX_Ycoeff,             %%mm7n"    // lum4
  117.          "punpckhwd      %%mm3,                  %%mm5n"
  119.          "movq                   %%mm4,                  (%4)n"
  120.          "movq                   %%mm5,                  8(%4)n"
  122.          "movq                   %%mm6,                  %%mm4n"        // Lum3
  123.          "paddw          %%mm0,                  %%mm6n"                // Lum3 +blue
  125.          "movq                   %%mm4,                  %%mm5n"                        // Lum3
  126.          "paddw          %%mm1,                  %%mm4n"       // Lum3 +red
  127.          "paddw          %%mm2,                  %%mm5n"                        // Lum3 +green
  128.          "psraw          $6,                             %%mm4n"
  129.          "movq                   %%mm7,                  %%mm3n"                        // Lum4
  130.          "psraw          $6,                             %%mm5n"
  131.          "paddw          %%mm0,                  %%mm7n"                   // Lum4 +blue
  132.          "psraw          $6,                             %%mm6n"                        // Lum3 +blue
  133.          "movq                   %%mm3,                  %%mm0n"  // Lum4
  134.          "packuswb       %%mm4,                  %%mm4n"
  135.          "paddw          %%mm1,                  %%mm3n"  // Lum4 +red
  136.          "packuswb       %%mm5,                  %%mm5n"
  137.          "paddw          %%mm2,                  %%mm0n"         // Lum4 +green
  138.          "packuswb       %%mm6,                  %%mm6n"
  139.          "punpcklbw      %%mm4,                  %%mm4n"
  140.          "punpcklbw      %%mm5,                  %%mm5n"
  141.          "punpcklbw      %%mm6,                  %%mm6n"
  142.          "psllw          $3,                             %%mm5n" // GREEN 3
  143.          "pand                   MMX_redmask,    %%mm4n"
  144.          "psraw          $6,                             %%mm3n" // psr 6
  145.          "psraw          $6,                             %%mm0n"
  146.          "pand                   MMX_redmask,    %%mm6n" // BLUE
  147.          "pand                   MMX_grnmask,    %%mm5n"
  148.          "psrlw          $11,                            %%mm6n"  // BLUE  3
  149.          "por                    %%mm5,                  %%mm4n"
  150.          "psraw          $6,                             %%mm7n"
  151.          "por                    %%mm6,                  %%mm4n"
  152.          "packuswb       %%mm3,                  %%mm3n"
  153.          "packuswb       %%mm0,                  %%mm0n"
  154.          "packuswb       %%mm7,                  %%mm7n"
  155.          "punpcklbw      %%mm3,                  %%mm3n"
  156.          "punpcklbw      %%mm0,                  %%mm0n"
  157.          "punpcklbw      %%mm7,                  %%mm7n"
  158.          "pand                   MMX_redmask,    %%mm3n"
  159.          "pand                   MMX_redmask,    %%mm7n" // BLUE
  160.          "psllw          $3,                             %%mm0n" // GREEN 4
  161.          "psrlw          $11,                            %%mm7n"
  162.          "pand                   MMX_grnmask,    %%mm0n"
  163.          "por                    %%mm7,                  %%mm3n"
  164.          "addl                   $8,                             %6n"
  165.          "por                    %%mm0,                  %%mm3n"
  167.          "movq                   %%mm4,                  %%mm5n"
  169.          "punpcklwd      %%mm3,                  %%mm4n"
  170.          "punpckhwd      %%mm3,                  %%mm5n"
  172.          "movq                   %%mm4,                  (%4,%5,2)n"
  173.          "movq                   %%mm5,                  8(%4,%5,2)n"
  175.          "addl                   $8,                             %2n"
  176.          "addl                   $4,                             %0n"
  177.          "addl                   $4,                             %1n"
  178.          "cmpl                   %3,                             %6n"
  179.          "leal                   16(%4),                 %4n"
  180.          "jl             1bn"
  181.          "addl           %3,     %2n"                   /* lum += cols */
  182.          "addl           %7,     %4n"                   /* row1 += mod */
  183.          "movl           $0,     %6n"
  184.          "cmpl           %8,     %2n"
  185.          "jl             1bn"
  186.          :
  187.          : "r" (cr), "r" (cb), "r" (lum), "r" (cols), "r" (row1) ,"r" (col1), "m" (x), "m" (mod)
  188.          , "m" (y)
  189.          );
  190.       __asm__ __volatile__(
  191.          "emmsn"
  192.          );
  194.    }
  196. #endif