Attachment #54317 for bug #86481

View | Details | Raw Unified | Return to bug 86481
Collapse All | Expand All





#ifdef SAVE_RCSID
static char rcsid =
 "@(#) $Id: SDL_yuv_mmx.c,v 1.9 2005/02/13 07:10:02 slouken Exp $";
#endif




#include "SDL_types.h"

#if __GNUC__ > 2
#    undef GCC2_HACK
#else
#    define GCC2_HACK
#endif

 
#if defined(GCC2_HACK) && defined (__ELF__)
#define ASM_VAR(X) _##X
#else
#define ASM_VAR(X) X
#endif
 
static volatile unsigned int  ASM_VAR(MMX_0080w)[]    = {0x00800080, 0x00800080};
static volatile unsigned int  ASM_VAR(MMX_00FFw)[]    = {0x00ff00ff, 0x00ff00ff}; 
static volatile unsigned int  ASM_VAR(MMX_FF00w)[]    = {0xff00ff00, 0xff00ff00}; 

static volatile unsigned short ASM_VAR(MMX_Ycoeff)[]  = {0x004a, 0x004a, 0x004a, 0x004a}; 

static volatile unsigned short ASM_VAR(MMX_UbluRGB)[] = {0x0072, 0x0072, 0x0072, 0x0072};    
static volatile unsigned short ASM_VAR(MMX_VredRGB)[] = {0x0059, 0x0059, 0x0059, 0x0059};  
static volatile unsigned short ASM_VAR(MMX_UgrnRGB)[] = {0xffea, 0xffea, 0xffea, 0xffea}; 
static volatile unsigned short ASM_VAR(MMX_VgrnRGB)[] = {0xffd2, 0xffd2, 0xffd2, 0xffd2};  

static volatile unsigned short ASM_VAR(MMX_Ublu5x5)[] = {0x0081, 0x0081, 0x0081, 0x0081};
static volatile unsigned short ASM_VAR(MMX_Vred5x5)[] = {0x0066, 0x0066, 0x0066, 0x0066};
static volatile unsigned short ASM_VAR(MMX_Ugrn555)[] = {0xffe7, 0xffe7, 0xffe7, 0xffe7};
static volatile unsigned short ASM_VAR(MMX_Vgrn555)[] = {0xffcc, 0xffcc, 0xffcc, 0xffcc};
static volatile unsigned short ASM_VAR(MMX_Ugrn565)[] = {0xffe8, 0xffe8, 0xffe8, 0xffe8};
static volatile unsigned short ASM_VAR(MMX_Vgrn565)[] = {0xffcd, 0xffcd, 0xffcd, 0xffcd};

static volatile unsigned short ASM_VAR(MMX_red555)[]  = {0x7c00, 0x7c00, 0x7c00, 0x7c00};
static volatile unsigned short ASM_VAR(MMX_red565)[]  = {0xf800, 0xf800, 0xf800, 0xf800};
static volatile unsigned short ASM_VAR(MMX_grn555)[]  = {0x03e0, 0x03e0, 0x03e0, 0x03e0};
static volatile unsigned short ASM_VAR(MMX_grn565)[]  = {0x07e0, 0x07e0, 0x07e0, 0x07e0};
static volatile unsigned short ASM_VAR(MMX_blu5x5)[]  = {0x001f, 0x001f, 0x001f, 0x001f};

/**
   This MMX assembler is my first assembler/MMX program ever.

		 "movd (%2), %%mm2\n"           //    0  0  0  0 l3 l2 l1 l0
		 "punpcklbw %%mm7,%%mm1\n" //         0  v3 0  v2 00 v1 00 v0
		 "punpckldq %%mm1,%%mm1\n" //         00 v1 00 v0 00 v1 00 v0
#ifdef GCC2_HACK
		 "psubw _MMX_0080w,%%mm1\n"  // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 
#else
		 "psubw %[_MMX_0080w],%%mm1\n"  // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 
#endif

		 // create Cr_g (result in mm0)
		 "movq %%mm1,%%mm0\n"           // r1 r1 r0 r0 r1 r1 r0 r0
#ifdef GCC2_HACK
		 "pmullw _MMX_VgrnRGB,%%mm0\n"// red*-46dec=0.7136*64
		 "pmullw _MMX_VredRGB,%%mm1\n"// red*89dec=1.4013*64
#else
		 "pmullw %[_MMX_VgrnRGB],%%mm0\n"// red*-46dec=0.7136*64
		 "pmullw %[_MMX_VredRGB],%%mm1\n"// red*89dec=1.4013*64
#endif
		 "psraw  $6, %%mm0\n"           // red=red/64
		 "psraw  $6, %%mm1\n"           // red=red/64
		 

		 "movq (%2,%4),%%mm3\n"         //    0  0  0  0 L3 L2 L1 L0
		 "punpckldq %%mm3,%%mm2\n"      //   L3 L2 L1 L0 l3 l2 l1 l0
		 "movq %%mm2,%%mm4\n"           //   L3 L2 L1 L0 l3 l2 l1 l0
#ifdef GCC2_HACK
		 "pand _MMX_FF00w,%%mm2\n"      //   L3 0  L1  0 l3  0 l1  0
		 "pand _MMX_00FFw,%%mm4\n"      //   0  L2  0 L0  0 l2  0 l0
#else
		 "pand %[_MMX_FF00w],%%mm2\n"      //   L3 0  L1  0 l3  0 l1  0
		 "pand %[_MMX_00FFw],%%mm4\n"      //   0  L2  0 L0  0 l2  0 l0
#endif
		 "psrlw $8,%%mm2\n"             //   0  L3  0 L1  0 l3  0 l1

		 // create R (result in mm6)

		 "movd (%1), %%mm1\n"      //         0  0  0  0  u3 u2 u1 u0
		 "punpcklbw %%mm7,%%mm1\n" //         0  u3 0  u2 00 u1 00 u0
		 "punpckldq %%mm1,%%mm1\n" //         00 u1 00 u0 00 u1 00 u0
#ifdef GCC2_HACK
		 "psubw _MMX_0080w,%%mm1\n"  // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 
#else
		 "psubw %[_MMX_0080w],%%mm1\n"  // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 
#endif
		 // create Cb_g (result in mm5)
		 "movq %%mm1,%%mm5\n"            // u1 u1 u0 u0 u1 u1 u0 u0
#ifdef GCC2_HACK
		 "pmullw _MMX_UgrnRGB,%%mm5\n"    // blue*-109dec=1.7129*64
		 "pmullw _MMX_UbluRGB,%%mm1\n"    // blue*114dec=1.78125*64
#else
		 "pmullw %[_MMX_UgrnRGB],%%mm5\n"    // blue*-109dec=1.7129*64
		 "pmullw %[_MMX_UbluRGB],%%mm1\n"    // blue*114dec=1.78125*64
#endif
		 "psraw  $6, %%mm5\n"            // blue=red/64
		 "psraw  $6, %%mm1\n"            // blue=blue/64


		 "popl %%ebx\n"
		 :
		 : "m" (cr), "r"(cb),"r"(lum),
		 "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)
#ifndef GCC2_HACK
		 ,[_MMX_0080w] "m" (*MMX_0080w),
		 [_MMX_00FFw] "m" (*MMX_00FFw),
		 [_MMX_FF00w] "m" (*MMX_FF00w),
		 [_MMX_VgrnRGB] "m" (*MMX_VgrnRGB),
		 [_MMX_VredRGB] "m" (*MMX_VredRGB),
		 [_MMX_UgrnRGB] "m" (*MMX_UgrnRGB),
		 [_MMX_UbluRGB] "m" (*MMX_UbluRGB)
#endif
			 );
}

void Color565DitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix,

         "movd           (%%ebx),                %%mm1\n" // 4 Cr                0  0  0  0 v3 v2 v1 v0
         "punpcklbw      %%mm7,                  %%mm0\n" // 4 W cb   0 u3  0 u2  0 u1  0 u0
         "punpcklbw      %%mm7,                  %%mm1\n" // 4 W cr   0 v3  0 v2  0 v1  0 v0
#ifdef GCC2_HACK
         "psubw          _MMX_0080w,             %%mm0\n"
         "psubw          _MMX_0080w,             %%mm1\n"
#else
         "psubw          %[_MMX_0080w],             %%mm0\n"
         "psubw          %[_MMX_0080w],             %%mm1\n"
#endif
         "movq           %%mm0,                  %%mm2\n" // Cb                   0 u3  0 u2  0 u1  0 u0
         "movq           %%mm1,                  %%mm3\n" // Cr
#ifdef GCC2_HACK
         "pmullw         _MMX_Ugrn565,           %%mm2\n" // Cb2green 0 R3  0 R2  0 R1  0 R0
#else
         "pmullw         %[_MMX_Ugrn565],           %%mm2\n" // Cb2green 0 R3  0 R2  0 R1  0 R0
#endif
         "movq           (%2),                   %%mm6\n" // L1      l7 L6 L5 L4 L3 L2 L1 L0
#ifdef GCC2_HACK
         "pmullw         _MMX_Ublu5x5,           %%mm0\n" // Cb2blue
         "pand           _MMX_00FFw,             %%mm6\n" // L1      00 L6 00 L4 00 L2 00 L0
         "pmullw         _MMX_Vgrn565,           %%mm3\n" // Cr2green
#else
         "pmullw         %[_MMX_Ublu5x5],           %%mm0\n" // Cb2blue
         "pand           %[_MMX_00FFw],             %%mm6\n" // L1      00 L6 00 L4 00 L2 00 L0
         "pmullw         %[_MMX_Vgrn565],           %%mm3\n" // Cr2green
#endif
         "movq           (%2),                   %%mm7\n" // L2
#ifdef GCC2_HACK
         "pmullw         _MMX_Vred5x5,           %%mm1\n" // Cr2red
#else
         "pmullw         %[_MMX_Vred5x5],           %%mm1\n" // Cr2red
#endif
         "psrlw          $8,                     %%mm7\n"        // L2           00 L7 00 L5 00 L3 00 L1
#ifdef GCC2_HACK
         "pmullw         _MMX_Ycoeff,            %%mm6\n" // lum1
#else
         "pmullw         %[_MMX_Ycoeff],            %%mm6\n" // lum1
#endif
         "paddw          %%mm3,                  %%mm2\n" // Cb2green + Cr2green == green
#ifdef GCC2_HACK
         "pmullw         _MMX_Ycoeff,            %%mm7\n" // lum2
#else
         "pmullw         %[_MMX_Ycoeff],            %%mm7\n" // lum2
#endif

         "movq           %%mm6,                  %%mm4\n" // lum1
         "paddw          %%mm0,                  %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0

         "punpcklbw      %%mm4,                  %%mm4\n"
         "punpcklbw      %%mm5,                  %%mm5\n"

#ifdef GCC2_HACK
         "pand           _MMX_red565,            %%mm4\n"
#else
         "pand           %[_MMX_red565],            %%mm4\n"
#endif
         "psllw          $3,                     %%mm5\n" // GREEN       1
         "punpcklbw      %%mm6,                  %%mm6\n"
#ifdef GCC2_HACK
         "pand           _MMX_grn565,            %%mm5\n"
         "pand           _MMX_red565,            %%mm6\n"
#else
         "pand           %[_MMX_grn565],            %%mm5\n"
         "pand           %[_MMX_red565],            %%mm6\n"
#endif
         "por            %%mm5,                  %%mm4\n" //
         "psrlw          $11,                    %%mm6\n" // BLUE        1
         "movq           %%mm3,                  %%mm5\n" // lum2

         "packuswb       %%mm3,                  %%mm3\n"
         "packuswb       %%mm5,                  %%mm5\n"
         "packuswb       %%mm7,                  %%mm7\n"
#ifdef GCC2_HACK
         "pand           _MMX_00FFw,             %%mm6\n" // L3
#else
         "pand           %[_MMX_00FFw],             %%mm6\n" // L3
#endif
         "punpcklbw      %%mm3,                  %%mm3\n"
         "punpcklbw      %%mm5,                  %%mm5\n"
#ifdef GCC2_HACK
         "pmullw         _MMX_Ycoeff,            %%mm6\n" // lum3
#else
         "pmullw         %[_MMX_Ycoeff],            %%mm6\n" // lum3
#endif
         "punpcklbw      %%mm7,                  %%mm7\n"
         "psllw          $3,                     %%mm5\n" // GREEN 2
#ifdef GCC2_HACK
         "pand           _MMX_red565,            %%mm7\n"
         "pand           _MMX_red565,            %%mm3\n"
#else
         "pand           %[_MMX_red565],            %%mm7\n"
         "pand           %[_MMX_red565],            %%mm3\n"
#endif
         "psrlw          $11,                    %%mm7\n" // BLUE  2
#ifdef GCC2_HACK
         "pand           _MMX_grn565,            %%mm5\n"
#else
         "pand           %[_MMX_grn565],            %%mm5\n"
#endif
         "por            %%mm7,                  %%mm3\n"
         "movq           (%2,%4),                %%mm7\n" // L4 load lum2
         "por            %%mm5,                  %%mm3\n" //
         "psrlw          $8,                     %%mm7\n" // L4
         "movq           %%mm4,                  %%mm5\n"
         "punpcklwd      %%mm3,                  %%mm4\n"
#ifdef GCC2_HACK
         "pmullw         _MMX_Ycoeff,            %%mm7\n" // lum4
#else
         "pmullw         %[_MMX_Ycoeff],            %%mm7\n" // lum4
#endif
         "punpckhwd      %%mm3,                  %%mm5\n"

         "movq           %%mm4,                  (%3)\n"  // write row1

         "punpcklbw      %%mm5,                  %%mm5\n"
         "punpcklbw      %%mm6,                  %%mm6\n"
         "psllw          $3,                     %%mm5\n" // GREEN 3
#ifdef GCC2_HACK
         "pand           _MMX_red565,            %%mm4\n"
#else
         "pand           %[_MMX_red565],            %%mm4\n"
#endif
         "psraw          $6,                     %%mm3\n" // psr 6
         "psraw          $6,                     %%mm0\n"
#ifdef GCC2_HACK
         "pand           _MMX_red565,            %%mm6\n" // BLUE
         "pand           _MMX_grn565,            %%mm5\n"
#else
         "pand           %[_MMX_red565],            %%mm6\n" // BLUE
         "pand           %[_MMX_grn565],            %%mm5\n"
#endif
         "psrlw          $11,                    %%mm6\n" // BLUE  3
         "por            %%mm5,                  %%mm4\n"
         "psraw          $6,                     %%mm7\n"

         "punpcklbw      %%mm3,                  %%mm3\n"
         "punpcklbw      %%mm0,                  %%mm0\n"
         "punpcklbw      %%mm7,                  %%mm7\n"
#ifdef GCC2_HACK
         "pand           _MMX_red565,            %%mm3\n"
         "pand           _MMX_red565,            %%mm7\n" // BLUE
#else
         "pand           %[_MMX_red565],            %%mm3\n"
         "pand           %[_MMX_red565],            %%mm7\n" // BLUE
#endif
         "psllw          $3,                     %%mm0\n" // GREEN 4
         "psrlw          $11,                    %%mm7\n"
#ifdef GCC2_HACK
         "pand           _MMX_grn565,            %%mm0\n"
#else
         "pand           %[_MMX_grn565],            %%mm0\n"
#endif
         "por            %%mm7,                  %%mm3\n"
         "por            %%mm0,                  %%mm3\n"


	 "jl             1b\n"
         "emms\n"
	 "popl %%ebx\n"
         :
         :"m" (cr), "r"(cb),"r"(lum),
	 "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)
#ifndef GCC2_HACK
	 ,[_MMX_0080w] "m" (*MMX_0080w),
	 [_MMX_Ugrn565] "m" (*MMX_Ugrn565),
	 [_MMX_Ublu5x5] "m" (*MMX_Ublu5x5),
	 [_MMX_00FFw] "m" (*MMX_00FFw),

	 [_MMX_Ycoeff] "m" (*MMX_Ycoeff),
	 [_MMX_red565] "m" (*MMX_red565),
	 [_MMX_grn565] "m" (*MMX_grn565)
#endif
		 );
}

#undef GCC2_HACK

#endif /* GCC i386 inline assembly */

Return to bug 86481

Lines 22-28 Link Here

(-)SDL12/src/video/SDL_yuv_mmx.c:1.8 (-38 / +155 lines)
22		22
23	#ifdef SAVE_RCSID	23	#ifdef SAVE_RCSID
24	static char rcsid =	24	static char rcsid =
25	"@(#) $Id: SDL_yuv_mmx.c,v 1.8 2004/08/24 07:40:14 slouken Exp $";	25	"@(#) $Id: SDL_yuv_mmx.c,v 1.9 2005/02/13 07:10:02 slouken Exp $";
26	#endif	26	#endif
27		27
28		28
Lines 30-58 Link Here
30		30
31	#include "SDL_types.h"	31	#include "SDL_types.h"
32		32
33	static unsigned int MMX_0080w[] = {0x00800080, 0x00800080};	33	#if __GNUC__ > 2
34	static unsigned int MMX_00FFw[] = {0x00ff00ff, 0x00ff00ff};	34	# undef GCC2_HACK
35	static unsigned int MMX_FF00w[] = {0xff00ff00, 0xff00ff00};	35	#else
36		36	# define GCC2_HACK
37	static unsigned short MMX_Ycoeff[] = {0x004a, 0x004a, 0x004a, 0x004a};	37	#endif
38		38
39	static unsigned short MMX_UbluRGB[] = {0x0072, 0x0072, 0x0072, 0x0072};	39
40	static unsigned short MMX_VredRGB[] = {0x0059, 0x0059, 0x0059, 0x0059};	40	#if defined(GCC2_HACK) && defined (__ELF__)
41	static unsigned short MMX_UgrnRGB[] = {0xffea, 0xffea, 0xffea, 0xffea};	41	#define ASM_VAR(X) _##X
42	static unsigned short MMX_VgrnRGB[] = {0xffd2, 0xffd2, 0xffd2, 0xffd2};	42	#else
43		43	#define ASM_VAR(X) X
44	static unsigned short MMX_Ublu5x5[] = {0x0081, 0x0081, 0x0081, 0x0081};	44	#endif
45	static unsigned short MMX_Vred5x5[] = {0x0066, 0x0066, 0x0066, 0x0066};	45
46	static unsigned short MMX_Ugrn555[] = {0xffe7, 0xffe7, 0xffe7, 0xffe7};	46	static volatile unsigned int ASM_VAR(MMX_0080w)[] = {0x00800080, 0x00800080};
47	static unsigned short MMX_Vgrn555[] = {0xffcc, 0xffcc, 0xffcc, 0xffcc};	47	static volatile unsigned int ASM_VAR(MMX_00FFw)[] = {0x00ff00ff, 0x00ff00ff};
48	static unsigned short MMX_Ugrn565[] = {0xffe8, 0xffe8, 0xffe8, 0xffe8};	48	static volatile unsigned int ASM_VAR(MMX_FF00w)[] = {0xff00ff00, 0xff00ff00};
49	static unsigned short MMX_Vgrn565[] = {0xffcd, 0xffcd, 0xffcd, 0xffcd};	49
50		50	static volatile unsigned short ASM_VAR(MMX_Ycoeff)[] = {0x004a, 0x004a, 0x004a, 0x004a};
51	static unsigned short MMX_red555[] = {0x7c00, 0x7c00, 0x7c00, 0x7c00};	51
52	static unsigned short MMX_red565[] = {0xf800, 0xf800, 0xf800, 0xf800};	52	static volatile unsigned short ASM_VAR(MMX_UbluRGB)[] = {0x0072, 0x0072, 0x0072, 0x0072};
53	static unsigned short MMX_grn555[] = {0x03e0, 0x03e0, 0x03e0, 0x03e0};	53	static volatile unsigned short ASM_VAR(MMX_VredRGB)[] = {0x0059, 0x0059, 0x0059, 0x0059};
54	static unsigned short MMX_grn565[] = {0x07e0, 0x07e0, 0x07e0, 0x07e0};	54	static volatile unsigned short ASM_VAR(MMX_UgrnRGB)[] = {0xffea, 0xffea, 0xffea, 0xffea};
55	static unsigned short MMX_blu5x5[] = {0x001f, 0x001f, 0x001f, 0x001f};	55	static volatile unsigned short ASM_VAR(MMX_VgrnRGB)[] = {0xffd2, 0xffd2, 0xffd2, 0xffd2};
		56
		57	static volatile unsigned short ASM_VAR(MMX_Ublu5x5)[] = {0x0081, 0x0081, 0x0081, 0x0081};
		58	static volatile unsigned short ASM_VAR(MMX_Vred5x5)[] = {0x0066, 0x0066, 0x0066, 0x0066};
		59	static volatile unsigned short ASM_VAR(MMX_Ugrn555)[] = {0xffe7, 0xffe7, 0xffe7, 0xffe7};
		60	static volatile unsigned short ASM_VAR(MMX_Vgrn555)[] = {0xffcc, 0xffcc, 0xffcc, 0xffcc};
		61	static volatile unsigned short ASM_VAR(MMX_Ugrn565)[] = {0xffe8, 0xffe8, 0xffe8, 0xffe8};
		62	static volatile unsigned short ASM_VAR(MMX_Vgrn565)[] = {0xffcd, 0xffcd, 0xffcd, 0xffcd};
		63
		64	static volatile unsigned short ASM_VAR(MMX_red555)[] = {0x7c00, 0x7c00, 0x7c00, 0x7c00};
		65	static volatile unsigned short ASM_VAR(MMX_red565)[] = {0xf800, 0xf800, 0xf800, 0xf800};
		66	static volatile unsigned short ASM_VAR(MMX_grn555)[] = {0x03e0, 0x03e0, 0x03e0, 0x03e0};
		67	static volatile unsigned short ASM_VAR(MMX_grn565)[] = {0x07e0, 0x07e0, 0x07e0, 0x07e0};
		68	static volatile unsigned short ASM_VAR(MMX_blu5x5)[] = {0x001f, 0x001f, 0x001f, 0x001f};
56		69
57	/**	70	/**
58	This MMX assembler is my first assembler/MMX program ever.	71	This MMX assembler is my first assembler/MMX program ever.
Lines 114-125 Link Here
114	"movd (%2), %%mm2\n" // 0 0 0 0 l3 l2 l1 l0	127	"movd (%2), %%mm2\n" // 0 0 0 0 l3 l2 l1 l0
115	"punpcklbw %%mm7,%%mm1\n" // 0 v3 0 v2 00 v1 00 v0	128	"punpcklbw %%mm7,%%mm1\n" // 0 v3 0 v2 00 v1 00 v0
116	"punpckldq %%mm1,%%mm1\n" // 00 v1 00 v0 00 v1 00 v0	129	"punpckldq %%mm1,%%mm1\n" // 00 v1 00 v0 00 v1 00 v0
		130	#ifdef GCC2_HACK
		131	"psubw _MMX_0080w,%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0
		132	#else
117	"psubw %[_MMX_0080w],%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0	133	"psubw %[_MMX_0080w],%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0
		134	#endif
118		135
119	// create Cr_g (result in mm0)	136	// create Cr_g (result in mm0)
120	"movq %%mm1,%%mm0\n" // r1 r1 r0 r0 r1 r1 r0 r0	137	"movq %%mm1,%%mm0\n" // r1 r1 r0 r0 r1 r1 r0 r0
		138	#ifdef GCC2_HACK
		139	"pmullw _MMX_VgrnRGB,%%mm0\n"// red-46dec=0.713664
		140	"pmullw _MMX_VredRGB,%%mm1\n"// red89dec=1.401364
		141	#else
121	"pmullw %[_MMX_VgrnRGB],%%mm0\n"// red-46dec=0.713664	142	"pmullw %[_MMX_VgrnRGB],%%mm0\n"// red-46dec=0.713664
122	"pmullw %[_MMX_VredRGB],%%mm1\n"// red89dec=1.401364	143	"pmullw %[_MMX_VredRGB],%%mm1\n"// red89dec=1.401364
		144	#endif
123	"psraw $6, %%mm0\n" // red=red/64	145	"psraw $6, %%mm0\n" // red=red/64
124	"psraw $6, %%mm1\n" // red=red/64	146	"psraw $6, %%mm1\n" // red=red/64
125		147
Lines 128-135 Link Here
128	"movq (%2,%4),%%mm3\n" // 0 0 0 0 L3 L2 L1 L0	150	"movq (%2,%4),%%mm3\n" // 0 0 0 0 L3 L2 L1 L0
129	"punpckldq %%mm3,%%mm2\n" // L3 L2 L1 L0 l3 l2 l1 l0	151	"punpckldq %%mm3,%%mm2\n" // L3 L2 L1 L0 l3 l2 l1 l0
130	"movq %%mm2,%%mm4\n" // L3 L2 L1 L0 l3 l2 l1 l0	152	"movq %%mm2,%%mm4\n" // L3 L2 L1 L0 l3 l2 l1 l0
		153	#ifdef GCC2_HACK
		154	"pand _MMX_FF00w,%%mm2\n" // L3 0 L1 0 l3 0 l1 0
		155	"pand _MMX_00FFw,%%mm4\n" // 0 L2 0 L0 0 l2 0 l0
		156	#else
131	"pand %[_MMX_FF00w],%%mm2\n" // L3 0 L1 0 l3 0 l1 0	157	"pand %[_MMX_FF00w],%%mm2\n" // L3 0 L1 0 l3 0 l1 0
132	"pand %[_MMX_00FFw],%%mm4\n" // 0 L2 0 L0 0 l2 0 l0	158	"pand %[_MMX_00FFw],%%mm4\n" // 0 L2 0 L0 0 l2 0 l0
		159	#endif
133	"psrlw $8,%%mm2\n" // 0 L3 0 L1 0 l3 0 l1	160	"psrlw $8,%%mm2\n" // 0 L3 0 L1 0 l3 0 l1
134		161
135	// create R (result in mm6)	162	// create R (result in mm6)
Lines 146-156 Link Here
146	"movd (%1), %%mm1\n" // 0 0 0 0 u3 u2 u1 u0	173	"movd (%1), %%mm1\n" // 0 0 0 0 u3 u2 u1 u0
147	"punpcklbw %%mm7,%%mm1\n" // 0 u3 0 u2 00 u1 00 u0	174	"punpcklbw %%mm7,%%mm1\n" // 0 u3 0 u2 00 u1 00 u0
148	"punpckldq %%mm1,%%mm1\n" // 00 u1 00 u0 00 u1 00 u0	175	"punpckldq %%mm1,%%mm1\n" // 00 u1 00 u0 00 u1 00 u0
		176	#ifdef GCC2_HACK
		177	"psubw _MMX_0080w,%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0
		178	#else
149	"psubw %[_MMX_0080w],%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0	179	"psubw %[_MMX_0080w],%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0
		180	#endif
150	// create Cb_g (result in mm5)	181	// create Cb_g (result in mm5)
151	"movq %%mm1,%%mm5\n" // u1 u1 u0 u0 u1 u1 u0 u0	182	"movq %%mm1,%%mm5\n" // u1 u1 u0 u0 u1 u1 u0 u0
		183	#ifdef GCC2_HACK
		184	"pmullw _MMX_UgrnRGB,%%mm5\n" // blue-109dec=1.712964
		185	"pmullw _MMX_UbluRGB,%%mm1\n" // blue114dec=1.7812564
		186	#else
152	"pmullw %[_MMX_UgrnRGB],%%mm5\n" // blue-109dec=1.712964	187	"pmullw %[_MMX_UgrnRGB],%%mm5\n" // blue-109dec=1.712964
153	"pmullw %[_MMX_UbluRGB],%%mm1\n" // blue114dec=1.7812564	188	"pmullw %[_MMX_UbluRGB],%%mm1\n" // blue114dec=1.7812564
		189	#endif
154	"psraw $6, %%mm5\n" // blue=red/64	190	"psraw $6, %%mm5\n" // blue=red/64
155	"psraw $6, %%mm1\n" // blue=blue/64	191	"psraw $6, %%mm1\n" // blue=blue/64
156		192
Lines 232-246 Link Here
232	"popl %%ebx\n"	268	"popl %%ebx\n"
233	:	269	:
234	: "m" (cr), "r"(cb),"r"(lum),	270	: "m" (cr), "r"(cb),"r"(lum),
235	"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod),	271	"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)
236	[_MMX_0080w] "m" (*MMX_0080w),	272	#ifndef GCC2_HACK
237	[_MMX_00FFw] "m" (*MMX_00FFw),	273	,[_MMX_0080w] "m" (*MMX_0080w),
238	[_MMX_FF00w] "m" (*MMX_FF00w),	274	[_MMX_00FFw] "m" (*MMX_00FFw),
239	[_MMX_VgrnRGB] "m" (*MMX_VgrnRGB),	275	[_MMX_FF00w] "m" (*MMX_FF00w),
240	[_MMX_VredRGB] "m" (*MMX_VredRGB),	276	[_MMX_VgrnRGB] "m" (*MMX_VgrnRGB),
241	[_MMX_UgrnRGB] "m" (*MMX_UgrnRGB),	277	[_MMX_VredRGB] "m" (*MMX_VredRGB),
242	[_MMX_UbluRGB] "m" (*MMX_UbluRGB)	278	[_MMX_UgrnRGB] "m" (*MMX_UgrnRGB),
243	);	279	[_MMX_UbluRGB] "m" (*MMX_UbluRGB)
		280	#endif
		281	);
244	}	282	}
245		283
246	void Color565DitherYV12MMX1X( int colortab, Uint32 rgb_2_pix,	284	void Color565DitherYV12MMX1X( int colortab, Uint32 rgb_2_pix,
Lines 269-289 Link Here
269	"movd (%%ebx), %%mm1\n" // 4 Cr 0 0 0 0 v3 v2 v1 v0	307	"movd (%%ebx), %%mm1\n" // 4 Cr 0 0 0 0 v3 v2 v1 v0
270	"punpcklbw %%mm7, %%mm0\n" // 4 W cb 0 u3 0 u2 0 u1 0 u0	308	"punpcklbw %%mm7, %%mm0\n" // 4 W cb 0 u3 0 u2 0 u1 0 u0
271	"punpcklbw %%mm7, %%mm1\n" // 4 W cr 0 v3 0 v2 0 v1 0 v0	309	"punpcklbw %%mm7, %%mm1\n" // 4 W cr 0 v3 0 v2 0 v1 0 v0
		310	#ifdef GCC2_HACK
		311	"psubw _MMX_0080w, %%mm0\n"
		312	"psubw _MMX_0080w, %%mm1\n"
		313	#else
272	"psubw %[_MMX_0080w], %%mm0\n"	314	"psubw %[_MMX_0080w], %%mm0\n"
273	"psubw %[_MMX_0080w], %%mm1\n"	315	"psubw %[_MMX_0080w], %%mm1\n"
		316	#endif
274	"movq %%mm0, %%mm2\n" // Cb 0 u3 0 u2 0 u1 0 u0	317	"movq %%mm0, %%mm2\n" // Cb 0 u3 0 u2 0 u1 0 u0
275	"movq %%mm1, %%mm3\n" // Cr	318	"movq %%mm1, %%mm3\n" // Cr
		319	#ifdef GCC2_HACK
		320	"pmullw _MMX_Ugrn565, %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0
		321	#else
276	"pmullw %[_MMX_Ugrn565], %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0	322	"pmullw %[_MMX_Ugrn565], %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0
		323	#endif
277	"movq (%2), %%mm6\n" // L1 l7 L6 L5 L4 L3 L2 L1 L0	324	"movq (%2), %%mm6\n" // L1 l7 L6 L5 L4 L3 L2 L1 L0
		325	#ifdef GCC2_HACK
		326	"pmullw _MMX_Ublu5x5, %%mm0\n" // Cb2blue
		327	"pand _MMX_00FFw, %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0
		328	"pmullw _MMX_Vgrn565, %%mm3\n" // Cr2green
		329	#else
278	"pmullw %[_MMX_Ublu5x5], %%mm0\n" // Cb2blue	330	"pmullw %[_MMX_Ublu5x5], %%mm0\n" // Cb2blue
279	"pand %[_MMX_00FFw], %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0	331	"pand %[_MMX_00FFw], %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0
280	"pmullw %[_MMX_Vgrn565], %%mm3\n" // Cr2green	332	"pmullw %[_MMX_Vgrn565], %%mm3\n" // Cr2green
		333	#endif
281	"movq (%2), %%mm7\n" // L2	334	"movq (%2), %%mm7\n" // L2
		335	#ifdef GCC2_HACK
		336	"pmullw _MMX_Vred5x5, %%mm1\n" // Cr2red
		337	#else
282	"pmullw %[_MMX_Vred5x5], %%mm1\n" // Cr2red	338	"pmullw %[_MMX_Vred5x5], %%mm1\n" // Cr2red
		339	#endif
283	"psrlw $8, %%mm7\n" // L2 00 L7 00 L5 00 L3 00 L1	340	"psrlw $8, %%mm7\n" // L2 00 L7 00 L5 00 L3 00 L1
		341	#ifdef GCC2_HACK
		342	"pmullw _MMX_Ycoeff, %%mm6\n" // lum1
		343	#else
284	"pmullw %[_MMX_Ycoeff], %%mm6\n" // lum1	344	"pmullw %[_MMX_Ycoeff], %%mm6\n" // lum1
		345	#endif
285	"paddw %%mm3, %%mm2\n" // Cb2green + Cr2green == green	346	"paddw %%mm3, %%mm2\n" // Cb2green + Cr2green == green
		347	#ifdef GCC2_HACK
		348	"pmullw _MMX_Ycoeff, %%mm7\n" // lum2
		349	#else
286	"pmullw %[_MMX_Ycoeff], %%mm7\n" // lum2	350	"pmullw %[_MMX_Ycoeff], %%mm7\n" // lum2
		351	#endif
287		352
288	"movq %%mm6, %%mm4\n" // lum1	353	"movq %%mm6, %%mm4\n" // lum1
289	"paddw %%mm0, %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0	354	"paddw %%mm0, %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0
Lines 301-311 Link Here
301	"punpcklbw %%mm4, %%mm4\n"	366	"punpcklbw %%mm4, %%mm4\n"
302	"punpcklbw %%mm5, %%mm5\n"	367	"punpcklbw %%mm5, %%mm5\n"
303		368
		369	#ifdef GCC2_HACK
		370	"pand _MMX_red565, %%mm4\n"
		371	#else
304	"pand %[_MMX_red565], %%mm4\n"	372	"pand %[_MMX_red565], %%mm4\n"
		373	#endif
305	"psllw $3, %%mm5\n" // GREEN 1	374	"psllw $3, %%mm5\n" // GREEN 1
306	"punpcklbw %%mm6, %%mm6\n"	375	"punpcklbw %%mm6, %%mm6\n"
		376	#ifdef GCC2_HACK
		377	"pand _MMX_grn565, %%mm5\n"
		378	"pand _MMX_red565, %%mm6\n"
		379	#else
307	"pand %[_MMX_grn565], %%mm5\n"	380	"pand %[_MMX_grn565], %%mm5\n"
308	"pand %[_MMX_red565], %%mm6\n"	381	"pand %[_MMX_red565], %%mm6\n"
		382	#endif
309	"por %%mm5, %%mm4\n" //	383	"por %%mm5, %%mm4\n" //
310	"psrlw $11, %%mm6\n" // BLUE 1	384	"psrlw $11, %%mm6\n" // BLUE 1
311	"movq %%mm3, %%mm5\n" // lum2	385	"movq %%mm3, %%mm5\n" // lum2
Lines 319-341 Link Here
319	"packuswb %%mm3, %%mm3\n"	393	"packuswb %%mm3, %%mm3\n"
320	"packuswb %%mm5, %%mm5\n"	394	"packuswb %%mm5, %%mm5\n"
321	"packuswb %%mm7, %%mm7\n"	395	"packuswb %%mm7, %%mm7\n"
		396	#ifdef GCC2_HACK
		397	"pand _MMX_00FFw, %%mm6\n" // L3
		398	#else
322	"pand %[_MMX_00FFw], %%mm6\n" // L3	399	"pand %[_MMX_00FFw], %%mm6\n" // L3
		400	#endif
323	"punpcklbw %%mm3, %%mm3\n"	401	"punpcklbw %%mm3, %%mm3\n"
324	"punpcklbw %%mm5, %%mm5\n"	402	"punpcklbw %%mm5, %%mm5\n"
		403	#ifdef GCC2_HACK
		404	"pmullw _MMX_Ycoeff, %%mm6\n" // lum3
		405	#else
325	"pmullw %[_MMX_Ycoeff], %%mm6\n" // lum3	406	"pmullw %[_MMX_Ycoeff], %%mm6\n" // lum3
		407	#endif
326	"punpcklbw %%mm7, %%mm7\n"	408	"punpcklbw %%mm7, %%mm7\n"
327	"psllw $3, %%mm5\n" // GREEN 2	409	"psllw $3, %%mm5\n" // GREEN 2
		410	#ifdef GCC2_HACK
		411	"pand _MMX_red565, %%mm7\n"
		412	"pand _MMX_red565, %%mm3\n"
		413	#else
328	"pand %[_MMX_red565], %%mm7\n"	414	"pand %[_MMX_red565], %%mm7\n"
329	"pand %[_MMX_red565], %%mm3\n"	415	"pand %[_MMX_red565], %%mm3\n"
		416	#endif
330	"psrlw $11, %%mm7\n" // BLUE 2	417	"psrlw $11, %%mm7\n" // BLUE 2
		418	#ifdef GCC2_HACK
		419	"pand _MMX_grn565, %%mm5\n"
		420	#else
331	"pand %[_MMX_grn565], %%mm5\n"	421	"pand %[_MMX_grn565], %%mm5\n"
		422	#endif
332	"por %%mm7, %%mm3\n"	423	"por %%mm7, %%mm3\n"
333	"movq (%2,%4), %%mm7\n" // L4 load lum2	424	"movq (%2,%4), %%mm7\n" // L4 load lum2
334	"por %%mm5, %%mm3\n" //	425	"por %%mm5, %%mm3\n" //
335	"psrlw $8, %%mm7\n" // L4	426	"psrlw $8, %%mm7\n" // L4
336	"movq %%mm4, %%mm5\n"	427	"movq %%mm4, %%mm5\n"
337	"punpcklwd %%mm3, %%mm4\n"	428	"punpcklwd %%mm3, %%mm4\n"
		429	#ifdef GCC2_HACK
		430	"pmullw _MMX_Ycoeff, %%mm7\n" // lum4
		431	#else
338	"pmullw %[_MMX_Ycoeff], %%mm7\n" // lum4	432	"pmullw %[_MMX_Ycoeff], %%mm7\n" // lum4
		433	#endif
339	"punpckhwd %%mm3, %%mm5\n"	434	"punpckhwd %%mm3, %%mm5\n"
340		435
341	"movq %%mm4, (%3)\n" // write row1	436	"movq %%mm4, (%3)\n" // write row1
Lines 362-372 Link Here
362	"punpcklbw %%mm5, %%mm5\n"	457	"punpcklbw %%mm5, %%mm5\n"
363	"punpcklbw %%mm6, %%mm6\n"	458	"punpcklbw %%mm6, %%mm6\n"
364	"psllw $3, %%mm5\n" // GREEN 3	459	"psllw $3, %%mm5\n" // GREEN 3
		460	#ifdef GCC2_HACK
		461	"pand _MMX_red565, %%mm4\n"
		462	#else
365	"pand %[_MMX_red565], %%mm4\n"	463	"pand %[_MMX_red565], %%mm4\n"
		464	#endif
366	"psraw $6, %%mm3\n" // psr 6	465	"psraw $6, %%mm3\n" // psr 6
367	"psraw $6, %%mm0\n"	466	"psraw $6, %%mm0\n"
		467	#ifdef GCC2_HACK
		468	"pand _MMX_red565, %%mm6\n" // BLUE
		469	"pand _MMX_grn565, %%mm5\n"
		470	#else
368	"pand %[_MMX_red565], %%mm6\n" // BLUE	471	"pand %[_MMX_red565], %%mm6\n" // BLUE
369	"pand %[_MMX_grn565], %%mm5\n"	472	"pand %[_MMX_grn565], %%mm5\n"
		473	#endif
370	"psrlw $11, %%mm6\n" // BLUE 3	474	"psrlw $11, %%mm6\n" // BLUE 3
371	"por %%mm5, %%mm4\n"	475	"por %%mm5, %%mm4\n"
372	"psraw $6, %%mm7\n"	476	"psraw $6, %%mm7\n"
Lines 377-387 Link Here
377	"punpcklbw %%mm3, %%mm3\n"	481	"punpcklbw %%mm3, %%mm3\n"
378	"punpcklbw %%mm0, %%mm0\n"	482	"punpcklbw %%mm0, %%mm0\n"
379	"punpcklbw %%mm7, %%mm7\n"	483	"punpcklbw %%mm7, %%mm7\n"
		484	#ifdef GCC2_HACK
		485	"pand _MMX_red565, %%mm3\n"
		486	"pand _MMX_red565, %%mm7\n" // BLUE
		487	#else
380	"pand %[_MMX_red565], %%mm3\n"	488	"pand %[_MMX_red565], %%mm3\n"
381	"pand %[_MMX_red565], %%mm7\n" // BLUE	489	"pand %[_MMX_red565], %%mm7\n" // BLUE
		490	#endif
382	"psllw $3, %%mm0\n" // GREEN 4	491	"psllw $3, %%mm0\n" // GREEN 4
383	"psrlw $11, %%mm7\n"	492	"psrlw $11, %%mm7\n"
		493	#ifdef GCC2_HACK
		494	"pand _MMX_grn565, %%mm0\n"
		495	#else
384	"pand %[_MMX_grn565], %%mm0\n"	496	"pand %[_MMX_grn565], %%mm0\n"
		497	#endif
385	"por %%mm7, %%mm3\n"	498	"por %%mm7, %%mm3\n"
386	"por %%mm0, %%mm3\n"	499	"por %%mm0, %%mm3\n"
387		500
Lines 411-420 Link Here
411	"jl 1b\n"	524	"jl 1b\n"
412	"emms\n"	525	"emms\n"
413	"popl %%ebx\n"	526	"popl %%ebx\n"
414	:	527	:
415	:"m" (cr), "r"(cb),"r"(lum),	528	:"m" (cr), "r"(cb),"r"(lum),
416	"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod),	529	"r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)
417	[_MMX_0080w] "m" (*MMX_0080w),	530	#ifndef GCC2_HACK
		531	,[_MMX_0080w] "m" (*MMX_0080w),
418	[_MMX_Ugrn565] "m" (*MMX_Ugrn565),	532	[_MMX_Ugrn565] "m" (*MMX_Ugrn565),
419	[_MMX_Ublu5x5] "m" (*MMX_Ublu5x5),	533	[_MMX_Ublu5x5] "m" (*MMX_Ublu5x5),
420	[_MMX_00FFw] "m" (*MMX_00FFw),	534	[_MMX_00FFw] "m" (*MMX_00FFw),
Lines 423-429 Link Here
423	[_MMX_Ycoeff] "m" (*MMX_Ycoeff),	537	[_MMX_Ycoeff] "m" (*MMX_Ycoeff),
424	[_MMX_red565] "m" (*MMX_red565),	538	[_MMX_red565] "m" (*MMX_red565),
425	[_MMX_grn565] "m" (*MMX_grn565)	539	[_MMX_grn565] "m" (*MMX_grn565)
426	);	540	#endif
		541	);
427	}	542	}
428		543
		544	#undef GCC2_HACK
		545
429	#endif /* GCC i386 inline assembly */	546	#endif /* GCC i386 inline assembly */