src/render/SDL_yuv_mmx.c
author Sam Lantinga <slouken@libsdl.org>
Fri, 15 Feb 2013 08:47:44 -0800
changeset 6885 700f1b25f77f
parent 6138 4c64952a58fb
child 7191 75360622e65f
permissions -rw-r--r--
Happy New Year!
     1 /*
     2   Simple DirectMedia Layer
     3   Copyright (C) 1997-2013 Sam Lantinga <slouken@libsdl.org>
     4 
     5   This software is provided 'as-is', without any express or implied
     6   warranty.  In no event will the authors be held liable for any damages
     7   arising from the use of this software.
     8 
     9   Permission is granted to anyone to use this software for any purpose,
    10   including commercial applications, and to alter it and redistribute it
    11   freely, subject to the following restrictions:
    12 
    13   1. The origin of this software must not be misrepresented; you must not
    14      claim that you wrote the original software. If you use this software
    15      in a product, an acknowledgment in the product documentation would be
    16      appreciated but is not required.
    17   2. Altered source versions must be plainly marked as such, and must not be
    18      misrepresented as being the original software.
    19   3. This notice may not be removed or altered from any source distribution.
    20 */
    21 #include "SDL_config.h"
    22 
    23 #if (__GNUC__ > 2) && defined(__i386__) && __OPTIMIZE__ && SDL_ASSEMBLY_ROUTINES
    24 
    25 #include "SDL_stdinc.h"
    26 
    27 #include "mmx.h"
    28 
    29 /* *INDENT-OFF* */
    30 
    31 static mmx_t MMX_0080w    = { .ud = {0x00800080, 0x00800080} };
    32 static mmx_t MMX_00FFw    = { .ud = {0x00ff00ff, 0x00ff00ff} };
    33 static mmx_t MMX_FF00w    = { .ud = {0xff00ff00, 0xff00ff00} };
    34 
    35 static mmx_t MMX_Ycoeff   = { .uw = {0x004a, 0x004a, 0x004a, 0x004a} };
    36 
    37 static mmx_t MMX_UbluRGB  = { .uw = {0x0072, 0x0072, 0x0072, 0x0072} };
    38 static mmx_t MMX_VredRGB  = { .uw = {0x0059, 0x0059, 0x0059, 0x0059} };
    39 static mmx_t MMX_UgrnRGB  = { .uw = {0xffea, 0xffea, 0xffea, 0xffea} };
    40 static mmx_t MMX_VgrnRGB  = { .uw = {0xffd2, 0xffd2, 0xffd2, 0xffd2} };
    41 
    42 static mmx_t MMX_Ublu5x5  = { .uw = {0x0081, 0x0081, 0x0081, 0x0081} };
    43 static mmx_t MMX_Vred5x5  = { .uw = {0x0066, 0x0066, 0x0066, 0x0066} };
    44 static mmx_t MMX_Ugrn565  = { .uw = {0xffe8, 0xffe8, 0xffe8, 0xffe8} };
    45 static mmx_t MMX_Vgrn565  = { .uw = {0xffcd, 0xffcd, 0xffcd, 0xffcd} };
    46 
    47 static mmx_t MMX_red565   = { .uw = {0xf800, 0xf800, 0xf800, 0xf800} };
    48 static mmx_t MMX_grn565   = { .uw = {0x07e0, 0x07e0, 0x07e0, 0x07e0} };
    49 
    50 /**
    51    This MMX assembler is my first assembler/MMX program ever.
    52    Thus it maybe buggy.
    53    Send patches to:
    54    mvogt@rhrk.uni-kl.de
    55 
    56    After it worked fine I have "obfuscated" the code a bit to have
    57    more parallism in the MMX units. This means I moved
    58    initilisation around and delayed other instruction.
    59    Performance measurement did not show that this brought any advantage
    60    but in theory it _should_ be faster this way.
    61 
    62    The overall performanve gain to the C based dither was 30%-40%.
    63    The MMX routine calculates 256bit=8RGB values in each cycle
    64    (4 for row1 & 4 for row2)
    65 
    66    The red/green/blue.. coefficents are taken from the mpeg_play 
    67    player. They look nice, but I dont know if you can have
    68    better values, to avoid integer rounding errors.
    69    
    70 
    71    IMPORTANT:
    72    ==========
    73 
    74    It is a requirement that the cr/cb/lum are 8 byte aligned and
    75    the out are 16byte aligned or you will/may get segfaults
    76 
    77 */
    78 
    79 void ColorRGBDitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix,
    80                               unsigned char *lum, unsigned char *cr,
    81                               unsigned char *cb, unsigned char *out,
    82                               int rows, int cols, int mod )
    83 {
    84     Uint32 *row1;
    85     Uint32 *row2;
    86 
    87     unsigned char* y = lum +cols*rows;    // Pointer to the end
    88     int x = 0;
    89     row1 = (Uint32 *)out;                 // 32 bit target
    90     row2 = (Uint32 *)out+cols+mod;        // start of second row
    91     mod = (mod+cols+mod)*4;               // increment for row1 in byte
    92 
    93     __asm__ __volatile__ (
    94         // tap dance to workaround the inability to use %%ebx at will...
    95         //  move one thing to the stack...
    96         "pushl $0\n"  // save a slot on the stack.
    97         "pushl %%ebx\n"  // save %%ebx.
    98         "movl %0, %%ebx\n"  // put the thing in ebx.
    99         "movl %%ebx,4(%%esp)\n"  // put the thing in the stack slot.
   100         "popl %%ebx\n"  // get back %%ebx (the PIC register).
   101 
   102         ".align 8\n"
   103         "1:\n"
   104 
   105         // create Cr (result in mm1)
   106         "pushl %%ebx\n"
   107         "movl 4(%%esp),%%ebx\n"
   108         "movd (%%ebx),%%mm1\n"   //         0  0  0  0  v3 v2 v1 v0
   109         "popl %%ebx\n"
   110         "pxor %%mm7,%%mm7\n"      //         00 00 00 00 00 00 00 00
   111         "movd (%2), %%mm2\n"           //    0  0  0  0 l3 l2 l1 l0
   112         "punpcklbw %%mm7,%%mm1\n" //         0  v3 0  v2 00 v1 00 v0
   113         "punpckldq %%mm1,%%mm1\n" //         00 v1 00 v0 00 v1 00 v0
   114         "psubw %9,%%mm1\n"        // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0
   115 
   116         // create Cr_g (result in mm0)
   117         "movq %%mm1,%%mm0\n"           // r1 r1 r0 r0 r1 r1 r0 r0
   118         "pmullw %10,%%mm0\n"           // red*-46dec=0.7136*64
   119         "pmullw %11,%%mm1\n"           // red*89dec=1.4013*64
   120         "psraw  $6, %%mm0\n"           // red=red/64
   121         "psraw  $6, %%mm1\n"           // red=red/64
   122 
   123         // create L1 L2 (result in mm2,mm4)
   124         // L2=lum+cols
   125         "movq (%2,%4),%%mm3\n"         //    0  0  0  0 L3 L2 L1 L0
   126         "punpckldq %%mm3,%%mm2\n"      //   L3 L2 L1 L0 l3 l2 l1 l0
   127         "movq %%mm2,%%mm4\n"           //   L3 L2 L1 L0 l3 l2 l1 l0
   128         "pand %12,%%mm2\n"             //   L3 0  L1  0 l3  0 l1  0
   129         "pand %13,%%mm4\n"             //   0  L2  0 L0  0 l2  0 l0
   130         "psrlw $8,%%mm2\n"             //   0  L3  0 L1  0 l3  0 l1
   131 
   132         // create R (result in mm6)
   133         "movq %%mm2,%%mm5\n"           //   0 L3  0 L1  0 l3  0 l1
   134         "movq %%mm4,%%mm6\n"           //   0 L2  0 L0  0 l2  0 l0
   135         "paddsw  %%mm1, %%mm5\n"       // lum1+red:x R3 x R1 x r3 x r1
   136         "paddsw  %%mm1, %%mm6\n"       // lum1+red:x R2 x R0 x r2 x r0
   137         "packuswb %%mm5,%%mm5\n"       //  R3 R1 r3 r1 R3 R1 r3 r1
   138         "packuswb %%mm6,%%mm6\n"       //  R2 R0 r2 r0 R2 R0 r2 r0
   139         "pxor %%mm7,%%mm7\n"      //         00 00 00 00 00 00 00 00
   140         "punpcklbw %%mm5,%%mm6\n"      //  R3 R2 R1 R0 r3 r2 r1 r0
   141 
   142         // create Cb (result in mm1)
   143         "movd (%1), %%mm1\n"      //         0  0  0  0  u3 u2 u1 u0
   144         "punpcklbw %%mm7,%%mm1\n" //         0  u3 0  u2 00 u1 00 u0
   145         "punpckldq %%mm1,%%mm1\n" //         00 u1 00 u0 00 u1 00 u0
   146         "psubw %9,%%mm1\n"        // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0
   147 
   148         // create Cb_g (result in mm5)
   149         "movq %%mm1,%%mm5\n"            // u1 u1 u0 u0 u1 u1 u0 u0
   150         "pmullw %14,%%mm5\n"            // blue*-109dec=1.7129*64
   151         "pmullw %15,%%mm1\n"            // blue*114dec=1.78125*64
   152         "psraw  $6, %%mm5\n"            // blue=red/64
   153         "psraw  $6, %%mm1\n"            // blue=blue/64
   154 
   155         // create G (result in mm7)
   156         "movq %%mm2,%%mm3\n"      //   0  L3  0 L1  0 l3  0 l1
   157         "movq %%mm4,%%mm7\n"      //   0  L2  0 L0  0 l2  0 l1
   158         "paddsw  %%mm5, %%mm3\n"  // lum1+Cb_g:x G3t x G1t x g3t x g1t
   159         "paddsw  %%mm5, %%mm7\n"  // lum1+Cb_g:x G2t x G0t x g2t x g0t
   160         "paddsw  %%mm0, %%mm3\n"  // lum1+Cr_g:x G3  x G1  x g3  x g1
   161         "paddsw  %%mm0, %%mm7\n"  // lum1+blue:x G2  x G0  x g2  x g0
   162         "packuswb %%mm3,%%mm3\n"  // G3 G1 g3 g1 G3 G1 g3 g1
   163         "packuswb %%mm7,%%mm7\n"  // G2 G0 g2 g0 G2 G0 g2 g0
   164         "punpcklbw %%mm3,%%mm7\n" // G3 G2 G1 G0 g3 g2 g1 g0
   165 
   166         // create B (result in mm5)
   167         "movq %%mm2,%%mm3\n"         //   0  L3  0 L1  0 l3  0 l1
   168         "movq %%mm4,%%mm5\n"         //   0  L2  0 L0  0 l2  0 l1
   169         "paddsw  %%mm1, %%mm3\n"     // lum1+blue:x B3 x B1 x b3 x b1
   170         "paddsw  %%mm1, %%mm5\n"     // lum1+blue:x B2 x B0 x b2 x b0
   171         "packuswb %%mm3,%%mm3\n"     // B3 B1 b3 b1 B3 B1 b3 b1
   172         "packuswb %%mm5,%%mm5\n"     // B2 B0 b2 b0 B2 B0 b2 b0
   173         "punpcklbw %%mm3,%%mm5\n"    // B3 B2 B1 B0 b3 b2 b1 b0
   174 
   175         // fill destination row1 (needed are mm6=Rr,mm7=Gg,mm5=Bb)
   176 
   177         "pxor %%mm2,%%mm2\n"           //  0  0  0  0  0  0  0  0
   178         "pxor %%mm4,%%mm4\n"           //  0  0  0  0  0  0  0  0
   179         "movq %%mm6,%%mm1\n"           // R3 R2 R1 R0 r3 r2 r1 r0
   180         "movq %%mm5,%%mm3\n"           // B3 B2 B1 B0 b3 b2 b1 b0
   181 
   182         // process lower lum
   183         "punpcklbw %%mm4,%%mm1\n"      //  0 r3  0 r2  0 r1  0 r0
   184         "punpcklbw %%mm4,%%mm3\n"      //  0 b3  0 b2  0 b1  0 b0
   185         "movq %%mm1,%%mm2\n"           //  0 r3  0 r2  0 r1  0 r0
   186         "movq %%mm3,%%mm0\n"           //  0 b3  0 b2  0 b1  0 b0
   187         "punpcklwd %%mm1,%%mm3\n"      //  0 r1  0 b1  0 r0  0 b0
   188         "punpckhwd %%mm2,%%mm0\n"      //  0 r3  0 b3  0 r2  0 b2
   189 
   190         "pxor %%mm2,%%mm2\n"           //  0  0  0  0  0  0  0  0
   191         "movq %%mm7,%%mm1\n"           // G3 G2 G1 G0 g3 g2 g1 g0
   192         "punpcklbw %%mm1,%%mm2\n"      // g3  0 g2  0 g1  0 g0  0
   193         "punpcklwd %%mm4,%%mm2\n"      //  0  0 g1  0  0  0 g0  0
   194         "por %%mm3, %%mm2\n"          //  0 r1 g1 b1  0 r0 g0 b0
   195         "movq %%mm2,(%3)\n"          // wrote out ! row1
   196 
   197         "pxor %%mm2,%%mm2\n"           //  0  0  0  0  0  0  0  0
   198         "punpcklbw %%mm1,%%mm4\n"      // g3  0 g2  0 g1  0 g0  0
   199         "punpckhwd %%mm2,%%mm4\n"      //  0  0 g3  0  0  0 g2  0
   200         "por %%mm0, %%mm4\n"          //  0 r3 g3 b3  0 r2 g2 b2
   201         "movq %%mm4,8(%3)\n"         // wrote out ! row1
   202 
   203         // fill destination row2 (needed are mm6=Rr,mm7=Gg,mm5=Bb)
   204         // this can be done "destructive"
   205         "pxor %%mm2,%%mm2\n"           //  0  0  0  0  0  0  0  0
   206         "punpckhbw %%mm2,%%mm6\n"      //  0 R3  0 R2  0 R1  0 R0
   207         "punpckhbw %%mm1,%%mm5\n"      // G3 B3 G2 B2 G1 B1 G0 B0
   208         "movq %%mm5,%%mm1\n"           // G3 B3 G2 B2 G1 B1 G0 B0
   209         "punpcklwd %%mm6,%%mm1\n"      //  0 R1 G1 B1  0 R0 G0 B0
   210         "movq %%mm1,(%5)\n"          // wrote out ! row2
   211         "punpckhwd %%mm6,%%mm5\n"      //  0 R3 G3 B3  0 R2 G2 B2
   212         "movq %%mm5,8(%5)\n"         // wrote out ! row2
   213 
   214         "addl $4,%2\n"            // lum+4
   215         "leal 16(%3),%3\n"        // row1+16
   216         "leal 16(%5),%5\n"        // row2+16
   217         "addl $2,(%%esp)\n"        // cr+2
   218         "addl $2,%1\n"           // cb+2
   219 
   220         "addl $4,%6\n"            // x+4
   221         "cmpl %4,%6\n"
   222 
   223         "jl 1b\n"
   224         "addl %4,%2\n" // lum += cols
   225         "addl %8,%3\n" // row1+= mod
   226         "addl %8,%5\n" // row2+= mod
   227         "movl $0,%6\n" // x=0
   228         "cmpl %7,%2\n"
   229         "jl 1b\n"
   230 
   231         "addl $4,%%esp\n"  // get rid of the stack slot we reserved.
   232         "emms\n"  // reset MMX registers.
   233         :
   234         : "m" (cr), "r"(cb),"r"(lum),
   235           "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod),
   236           "m"(MMX_0080w),"m"(MMX_VgrnRGB),"m"(MMX_VredRGB),
   237           "m"(MMX_FF00w),"m"(MMX_00FFw),"m"(MMX_UgrnRGB),
   238           "m"(MMX_UbluRGB)
   239     );
   240 }
   241 
   242 void Color565DitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix,
   243                              unsigned char *lum, unsigned char *cr,
   244                              unsigned char *cb, unsigned char *out,
   245                              int rows, int cols, int mod )
   246 {
   247     Uint16 *row1;
   248     Uint16 *row2;
   249 
   250     unsigned char* y = lum +cols*rows;    /* Pointer to the end */
   251     int x = 0;
   252     row1 = (Uint16 *)out;                 /* 16 bit target */
   253     row2 = (Uint16 *)out+cols+mod;        /* start of second row  */
   254     mod = (mod+cols+mod)*2;               /* increment for row1 in byte */
   255 
   256     __asm__ __volatile__(
   257         // tap dance to workaround the inability to use %%ebx at will...
   258         //  move one thing to the stack...
   259         "pushl $0\n"  // save a slot on the stack.
   260         "pushl %%ebx\n"  // save %%ebx.
   261         "movl %0, %%ebx\n"  // put the thing in ebx.
   262         "movl %%ebx, 4(%%esp)\n"  // put the thing in the stack slot.
   263         "popl %%ebx\n"  // get back %%ebx (the PIC register).
   264 
   265         ".align 8\n"
   266         "1:\n"
   267 
   268         "movd           (%1),                   %%mm0\n" // 4 Cb         0  0  0  0 u3 u2 u1 u0
   269         "pxor           %%mm7,                  %%mm7\n"
   270         "pushl %%ebx\n"
   271         "movl 4(%%esp), %%ebx\n"
   272         "movd (%%ebx), %%mm1\n"   // 4 Cr                0  0  0  0 v3 v2 v1 v0
   273         "popl %%ebx\n"
   274 
   275         "punpcklbw      %%mm7,                  %%mm0\n" // 4 W cb   0 u3  0 u2  0 u1  0 u0
   276         "punpcklbw      %%mm7,                  %%mm1\n" // 4 W cr   0 v3  0 v2  0 v1  0 v0
   277         "psubw          %9,                     %%mm0\n"
   278         "psubw          %9,                     %%mm1\n"
   279         "movq           %%mm0,                  %%mm2\n" // Cb                   0 u3  0 u2  0 u1  0 u0
   280         "movq           %%mm1,                  %%mm3\n" // Cr
   281         "pmullw         %10,                    %%mm2\n" // Cb2green 0 R3  0 R2  0 R1  0 R0
   282         "movq           (%2),                   %%mm6\n" // L1      l7 L6 L5 L4 L3 L2 L1 L0
   283         "pmullw         %11,                    %%mm0\n" // Cb2blue
   284         "pand           %12,                    %%mm6\n" // L1      00 L6 00 L4 00 L2 00 L0
   285         "pmullw         %13,                    %%mm3\n" // Cr2green
   286         "movq           (%2),                   %%mm7\n" // L2
   287         "pmullw         %14,                    %%mm1\n" // Cr2red
   288         "psrlw          $8,                     %%mm7\n"        // L2           00 L7 00 L5 00 L3 00 L1
   289         "pmullw         %15,                    %%mm6\n" // lum1
   290         "paddw          %%mm3,                  %%mm2\n" // Cb2green + Cr2green == green
   291         "pmullw         %15,                    %%mm7\n" // lum2
   292 
   293         "movq           %%mm6,                  %%mm4\n" // lum1
   294         "paddw          %%mm0,                  %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0
   295         "movq           %%mm4,                  %%mm5\n" // lum1
   296         "paddw          %%mm1,                  %%mm4\n" // lum1 +red  00 R6 00 R4 00 R2 00 R0
   297         "paddw          %%mm2,                  %%mm5\n" // lum1 +green 00 G6 00 G4 00 G2 00 G0
   298         "psraw          $6,                     %%mm4\n" // R1 0 .. 64
   299         "movq           %%mm7,                  %%mm3\n" // lum2                       00 L7 00 L5 00 L3 00 L1
   300         "psraw          $6,                     %%mm5\n" // G1  - .. +
   301         "paddw          %%mm0,                  %%mm7\n" // Lum2 +blue 00 B7 00 B5 00 B3 00 B1
   302         "psraw          $6,                     %%mm6\n" // B1         0 .. 64
   303         "packuswb       %%mm4,                  %%mm4\n" // R1 R1
   304         "packuswb       %%mm5,                  %%mm5\n" // G1 G1
   305         "packuswb       %%mm6,                  %%mm6\n" // B1 B1
   306         "punpcklbw      %%mm4,                  %%mm4\n"
   307         "punpcklbw      %%mm5,                  %%mm5\n"
   308 
   309         "pand           %16,                    %%mm4\n"
   310         "psllw          $3,                     %%mm5\n" // GREEN       1
   311         "punpcklbw      %%mm6,                  %%mm6\n"
   312         "pand           %17,                    %%mm5\n"
   313         "pand           %16,                    %%mm6\n"
   314         "por            %%mm5,                  %%mm4\n" //
   315         "psrlw          $11,                    %%mm6\n" // BLUE        1
   316         "movq           %%mm3,                  %%mm5\n" // lum2
   317         "paddw          %%mm1,                  %%mm3\n" // lum2 +red      00 R7 00 R5 00 R3 00 R1
   318         "paddw          %%mm2,                  %%mm5\n" // lum2 +green 00 G7 00 G5 00 G3 00 G1
   319         "psraw          $6,                     %%mm3\n" // R2
   320         "por            %%mm6,                  %%mm4\n" // MM4
   321         "psraw          $6,                     %%mm5\n" // G2
   322         "movq           (%2, %4),               %%mm6\n" // L3 load lum2
   323         "psraw          $6,                     %%mm7\n"
   324         "packuswb       %%mm3,                  %%mm3\n"
   325         "packuswb       %%mm5,                  %%mm5\n"
   326         "packuswb       %%mm7,                  %%mm7\n"
   327         "pand           %12,                    %%mm6\n" // L3
   328         "punpcklbw      %%mm3,                  %%mm3\n"
   329         "punpcklbw      %%mm5,                  %%mm5\n"
   330         "pmullw         %15,                    %%mm6\n" // lum3
   331         "punpcklbw      %%mm7,                  %%mm7\n"
   332         "psllw          $3,                     %%mm5\n" // GREEN 2
   333         "pand           %16,                    %%mm7\n"
   334         "pand           %16,                    %%mm3\n"
   335         "psrlw          $11,                    %%mm7\n" // BLUE  2
   336         "pand           %17,                    %%mm5\n"
   337         "por            %%mm7,                  %%mm3\n"
   338         "movq           (%2,%4),                %%mm7\n" // L4 load lum2
   339         "por            %%mm5,                  %%mm3\n" //
   340         "psrlw          $8,                     %%mm7\n" // L4
   341         "movq           %%mm4,                  %%mm5\n"
   342         "punpcklwd      %%mm3,                  %%mm4\n"
   343         "pmullw         %15,                    %%mm7\n" // lum4
   344         "punpckhwd      %%mm3,                  %%mm5\n"
   345 
   346         "movq           %%mm4,                  (%3)\n"  // write row1
   347         "movq           %%mm5,                  8(%3)\n" // write row1
   348 
   349         "movq           %%mm6,                  %%mm4\n" // Lum3
   350         "paddw          %%mm0,                  %%mm6\n" // Lum3 +blue
   351 
   352         "movq           %%mm4,                  %%mm5\n" // Lum3
   353         "paddw          %%mm1,                  %%mm4\n" // Lum3 +red
   354         "paddw          %%mm2,                  %%mm5\n" // Lum3 +green
   355         "psraw          $6,                     %%mm4\n"
   356         "movq           %%mm7,                  %%mm3\n" // Lum4
   357         "psraw          $6,                     %%mm5\n"
   358         "paddw          %%mm0,                  %%mm7\n" // Lum4 +blue
   359         "psraw          $6,                     %%mm6\n" // Lum3 +blue
   360         "movq           %%mm3,                  %%mm0\n" // Lum4
   361         "packuswb       %%mm4,                  %%mm4\n"
   362         "paddw          %%mm1,                  %%mm3\n" // Lum4 +red
   363         "packuswb       %%mm5,                  %%mm5\n"
   364         "paddw          %%mm2,                  %%mm0\n" // Lum4 +green
   365         "packuswb       %%mm6,                  %%mm6\n"
   366         "punpcklbw      %%mm4,                  %%mm4\n"
   367         "punpcklbw      %%mm5,                  %%mm5\n"
   368         "punpcklbw      %%mm6,                  %%mm6\n"
   369         "psllw          $3,                     %%mm5\n" // GREEN 3
   370         "pand           %16,                    %%mm4\n"
   371         "psraw          $6,                     %%mm3\n" // psr 6
   372         "psraw          $6,                     %%mm0\n"
   373         "pand           %16,                    %%mm6\n" // BLUE
   374         "pand           %17,                    %%mm5\n"
   375         "psrlw          $11,                    %%mm6\n" // BLUE  3
   376         "por            %%mm5,                  %%mm4\n"
   377         "psraw          $6,                     %%mm7\n"
   378         "por            %%mm6,                  %%mm4\n"
   379         "packuswb       %%mm3,                  %%mm3\n"
   380         "packuswb       %%mm0,                  %%mm0\n"
   381         "packuswb       %%mm7,                  %%mm7\n"
   382         "punpcklbw      %%mm3,                  %%mm3\n"
   383         "punpcklbw      %%mm0,                  %%mm0\n"
   384         "punpcklbw      %%mm7,                  %%mm7\n"
   385         "pand           %16,                    %%mm3\n"
   386         "pand           %16,                    %%mm7\n" // BLUE
   387         "psllw          $3,                     %%mm0\n" // GREEN 4
   388         "psrlw          $11,                    %%mm7\n"
   389         "pand           %17,                    %%mm0\n"
   390         "por            %%mm7,                  %%mm3\n"
   391         "por            %%mm0,                  %%mm3\n"
   392 
   393         "movq           %%mm4,                  %%mm5\n"
   394 
   395         "punpcklwd      %%mm3,                  %%mm4\n"
   396         "punpckhwd      %%mm3,                  %%mm5\n"
   397 
   398         "movq           %%mm4,                  (%5)\n"
   399         "movq           %%mm5,                  8(%5)\n"
   400 
   401         "addl           $8,                     %6\n"
   402         "addl           $8,                     %2\n"
   403         "addl           $4,                     (%%esp)\n"
   404         "addl           $4,                     %1\n"
   405         "cmpl           %4,                     %6\n"
   406         "leal           16(%3),                 %3\n"
   407         "leal           16(%5),%5\n" // row2+16
   408 
   409         "jl             1b\n"
   410         "addl           %4,     %2\n" // lum += cols
   411         "addl           %8,     %3\n" // row1+= mod
   412         "addl           %8,     %5\n" // row2+= mod
   413         "movl           $0,     %6\n" // x=0
   414         "cmpl           %7,     %2\n"
   415         "jl             1b\n"
   416         "addl $4, %%esp\n"  // get rid of the stack slot we reserved.
   417         "emms\n"
   418         :
   419         : "m" (cr), "r"(cb),"r"(lum),
   420           "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod),
   421           "m"(MMX_0080w),"m"(MMX_Ugrn565),"m"(MMX_Ublu5x5),
   422           "m"(MMX_00FFw),"m"(MMX_Vgrn565),"m"(MMX_Vred5x5),
   423           "m"(MMX_Ycoeff),"m"(MMX_red565),"m"(MMX_grn565)
   424     );
   425 }
   426 
   427 /* *INDENT-ON* */
   428 
   429 #endif /* GCC3 i386 inline assembly */
   430 
   431 /* vi: set ts=4 sw=4 expandtab: */