Annotation of /trunk/libsdl/patches/libsdl-1.2.9-PIC-yuv-mmx.patch
Parent Directory | Revision Log
Revision 144 -
(hide annotations)
(download)
Tue May 8 20:06:05 2007 UTC (17 years, 4 months ago) by niro
File size: 13260 byte(s)
Tue May 8 20:06:05 2007 UTC (17 years, 4 months ago) by niro
File size: 13260 byte(s)
-import
1 | niro | 144 | --- src/video/SDL_yuv_mmx.c |
2 | +++ src/video/SDL_yuv_mmx.c | ||
3 | @@ -29,32 +29,31 @@ static char rcsid = | ||
4 | #if defined(i386) && defined(__GNUC__) && defined(USE_ASMBLIT) | ||
5 | |||
6 | #include "SDL_types.h" | ||
7 | +#include "mmx.h" | ||
8 | |||
9 | -#define ASM_ARRAY(x) x[] __asm__("_" #x) __attribute__((used)) | ||
10 | - | ||
11 | -static unsigned int ASM_ARRAY(MMX_0080w) = {0x00800080, 0x00800080}; | ||
12 | -static unsigned int ASM_ARRAY(MMX_00FFw) = {0x00ff00ff, 0x00ff00ff}; | ||
13 | -static unsigned int ASM_ARRAY(MMX_FF00w) = {0xff00ff00, 0xff00ff00}; | ||
14 | - | ||
15 | -static unsigned short ASM_ARRAY(MMX_Ycoeff) = {0x004a, 0x004a, 0x004a, 0x004a}; | ||
16 | - | ||
17 | -static unsigned short ASM_ARRAY(MMX_UbluRGB) = {0x0072, 0x0072, 0x0072, 0x0072}; | ||
18 | -static unsigned short ASM_ARRAY(MMX_VredRGB) = {0x0059, 0x0059, 0x0059, 0x0059}; | ||
19 | -static unsigned short ASM_ARRAY(MMX_UgrnRGB) = {0xffea, 0xffea, 0xffea, 0xffea}; | ||
20 | -static unsigned short ASM_ARRAY(MMX_VgrnRGB) = {0xffd2, 0xffd2, 0xffd2, 0xffd2}; | ||
21 | - | ||
22 | -static unsigned short ASM_ARRAY(MMX_Ublu5x5) = {0x0081, 0x0081, 0x0081, 0x0081}; | ||
23 | -static unsigned short ASM_ARRAY(MMX_Vred5x5) = {0x0066, 0x0066, 0x0066, 0x0066}; | ||
24 | -static unsigned short ASM_ARRAY(MMX_Ugrn555) = {0xffe7, 0xffe7, 0xffe7, 0xffe7}; | ||
25 | -static unsigned short ASM_ARRAY(MMX_Vgrn555) = {0xffcc, 0xffcc, 0xffcc, 0xffcc}; | ||
26 | -static unsigned short ASM_ARRAY(MMX_Ugrn565) = {0xffe8, 0xffe8, 0xffe8, 0xffe8}; | ||
27 | -static unsigned short ASM_ARRAY(MMX_Vgrn565) = {0xffcd, 0xffcd, 0xffcd, 0xffcd}; | ||
28 | - | ||
29 | -static unsigned short ASM_ARRAY(MMX_red555) = {0x7c00, 0x7c00, 0x7c00, 0x7c00}; | ||
30 | -static unsigned short ASM_ARRAY(MMX_red565) = {0xf800, 0xf800, 0xf800, 0xf800}; | ||
31 | -static unsigned short ASM_ARRAY(MMX_grn555) = {0x03e0, 0x03e0, 0x03e0, 0x03e0}; | ||
32 | -static unsigned short ASM_ARRAY(MMX_grn565) = {0x07e0, 0x07e0, 0x07e0, 0x07e0}; | ||
33 | -static unsigned short ASM_ARRAY(MMX_blu5x5) = {0x001f, 0x001f, 0x001f, 0x001f}; | ||
34 | +static mmx_t MMX_0080w = { .ud = {0x00800080, 0x00800080} }; | ||
35 | +static mmx_t MMX_00FFw = { .ud = {0x00ff00ff, 0x00ff00ff} }; | ||
36 | +static mmx_t MMX_FF00w = { .ud = {0xff00ff00, 0xff00ff00} }; | ||
37 | + | ||
38 | +static mmx_t MMX_Ycoeff = { .uw = {0x004a, 0x004a, 0x004a, 0x004a} }; | ||
39 | + | ||
40 | +static mmx_t MMX_UbluRGB = { .uw = {0x0072, 0x0072, 0x0072, 0x0072} }; | ||
41 | +static mmx_t MMX_VredRGB = { .uw = {0x0059, 0x0059, 0x0059, 0x0059} }; | ||
42 | +static mmx_t MMX_UgrnRGB = { .uw = {0xffea, 0xffea, 0xffea, 0xffea} }; | ||
43 | +static mmx_t MMX_VgrnRGB = { .uw = {0xffd2, 0xffd2, 0xffd2, 0xffd2} }; | ||
44 | + | ||
45 | +static mmx_t MMX_Ublu5x5 = { .uw = {0x0081, 0x0081, 0x0081, 0x0081} }; | ||
46 | +static mmx_t MMX_Vred5x5 = { .uw = {0x0066, 0x0066, 0x0066, 0x0066} }; | ||
47 | +static mmx_t MMX_Ugrn555 = { .uw = {0xffe7, 0xffe7, 0xffe7, 0xffe7} }; | ||
48 | +static mmx_t MMX_Vgrn555 = { .uw = {0xffcc, 0xffcc, 0xffcc, 0xffcc} }; | ||
49 | +static mmx_t MMX_Ugrn565 = { .uw = {0xffe8, 0xffe8, 0xffe8, 0xffe8} }; | ||
50 | +static mmx_t MMX_Vgrn565 = { .uw = {0xffcd, 0xffcd, 0xffcd, 0xffcd} }; | ||
51 | + | ||
52 | +static mmx_t MMX_red555 = { .uw = {0x7c00, 0x7c00, 0x7c00, 0x7c00} }; | ||
53 | +static mmx_t MMX_red565 = { .uw = {0xf800, 0xf800, 0xf800, 0xf800} }; | ||
54 | +static mmx_t MMX_grn555 = { .uw = {0x03e0, 0x03e0, 0x03e0, 0x03e0} }; | ||
55 | +static mmx_t MMX_grn565 = { .uw = {0x07e0, 0x07e0, 0x07e0, 0x07e0} }; | ||
56 | +static mmx_t MMX_blu5x5 = { .uw = {0x001f, 0x001f, 0x001f, 0x001f} }; | ||
57 | |||
58 | /** | ||
59 | This MMX assembler is my first assembler/MMX program ever. | ||
60 | @@ -116,12 +115,12 @@ void ColorRGBDitherYV12MMX1X( int *color | ||
61 | "movd (%2), %%mm2\n" // 0 0 0 0 l3 l2 l1 l0 | ||
62 | "punpcklbw %%mm7,%%mm1\n" // 0 v3 0 v2 00 v1 00 v0 | ||
63 | "punpckldq %%mm1,%%mm1\n" // 00 v1 00 v0 00 v1 00 v0 | ||
64 | - "psubw _MMX_0080w,%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 | ||
65 | + "psubw %9,%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 | ||
66 | |||
67 | // create Cr_g (result in mm0) | ||
68 | "movq %%mm1,%%mm0\n" // r1 r1 r0 r0 r1 r1 r0 r0 | ||
69 | - "pmullw _MMX_VgrnRGB,%%mm0\n"// red*-46dec=0.7136*64 | ||
70 | - "pmullw _MMX_VredRGB,%%mm1\n"// red*89dec=1.4013*64 | ||
71 | + "pmullw %10,%%mm0\n" // red*-46dec=0.7136*64 | ||
72 | + "pmullw %11,%%mm1\n" // red*89dec=1.4013*64 | ||
73 | "psraw $6, %%mm0\n" // red=red/64 | ||
74 | "psraw $6, %%mm1\n" // red=red/64 | ||
75 | |||
76 | @@ -130,8 +129,8 @@ void ColorRGBDitherYV12MMX1X( int *color | ||
77 | "movq (%2,%4),%%mm3\n" // 0 0 0 0 L3 L2 L1 L0 | ||
78 | "punpckldq %%mm3,%%mm2\n" // L3 L2 L1 L0 l3 l2 l1 l0 | ||
79 | "movq %%mm2,%%mm4\n" // L3 L2 L1 L0 l3 l2 l1 l0 | ||
80 | - "pand _MMX_FF00w,%%mm2\n" // L3 0 L1 0 l3 0 l1 0 | ||
81 | - "pand _MMX_00FFw,%%mm4\n" // 0 L2 0 L0 0 l2 0 l0 | ||
82 | + "pand %12,%%mm2\n" // L3 0 L1 0 l3 0 l1 0 | ||
83 | + "pand %13,%%mm4\n" // 0 L2 0 L0 0 l2 0 l0 | ||
84 | "psrlw $8,%%mm2\n" // 0 L3 0 L1 0 l3 0 l1 | ||
85 | |||
86 | // create R (result in mm6) | ||
87 | @@ -148,11 +147,11 @@ void ColorRGBDitherYV12MMX1X( int *color | ||
88 | "movd (%1), %%mm1\n" // 0 0 0 0 u3 u2 u1 u0 | ||
89 | "punpcklbw %%mm7,%%mm1\n" // 0 u3 0 u2 00 u1 00 u0 | ||
90 | "punpckldq %%mm1,%%mm1\n" // 00 u1 00 u0 00 u1 00 u0 | ||
91 | - "psubw _MMX_0080w,%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 | ||
92 | + "psubw %9,%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 | ||
93 | // create Cb_g (result in mm5) | ||
94 | "movq %%mm1,%%mm5\n" // u1 u1 u0 u0 u1 u1 u0 u0 | ||
95 | - "pmullw _MMX_UgrnRGB,%%mm5\n" // blue*-109dec=1.7129*64 | ||
96 | - "pmullw _MMX_UbluRGB,%%mm1\n" // blue*114dec=1.78125*64 | ||
97 | + "pmullw %14,%%mm5\n" // blue*-109dec=1.7129*64 | ||
98 | + "pmullw %15,%%mm1\n" // blue*114dec=1.78125*64 | ||
99 | "psraw $6, %%mm5\n" // blue=red/64 | ||
100 | "psraw $6, %%mm1\n" // blue=blue/64 | ||
101 | |||
102 | @@ -234,7 +233,10 @@ void ColorRGBDitherYV12MMX1X( int *color | ||
103 | "popl %%ebx\n" | ||
104 | : | ||
105 | : "m" (cr), "r"(cb),"r"(lum), | ||
106 | - "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)); | ||
107 | + "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod), | ||
108 | + "m"(MMX_0080w),"m"(MMX_VgrnRGB),"m"(MMX_VredRGB), | ||
109 | + "m"(MMX_FF00w),"m"(MMX_00FFw),"m"(MMX_UgrnRGB), | ||
110 | + "m"(MMX_UbluRGB)); | ||
111 | } | ||
112 | |||
113 | void Color565DitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix, | ||
114 | @@ -263,21 +265,21 @@ void Color565DitherYV12MMX1X( int *color | ||
115 | "movd (%%ebx), %%mm1\n" // 4 Cr 0 0 0 0 v3 v2 v1 v0 | ||
116 | "punpcklbw %%mm7, %%mm0\n" // 4 W cb 0 u3 0 u2 0 u1 0 u0 | ||
117 | "punpcklbw %%mm7, %%mm1\n" // 4 W cr 0 v3 0 v2 0 v1 0 v0 | ||
118 | - "psubw _MMX_0080w, %%mm0\n" | ||
119 | - "psubw _MMX_0080w, %%mm1\n" | ||
120 | + "psubw %9, %%mm0\n" | ||
121 | + "psubw %9, %%mm1\n" | ||
122 | "movq %%mm0, %%mm2\n" // Cb 0 u3 0 u2 0 u1 0 u0 | ||
123 | "movq %%mm1, %%mm3\n" // Cr | ||
124 | - "pmullw _MMX_Ugrn565, %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0 | ||
125 | + "pmullw %10, %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0 | ||
126 | "movq (%2), %%mm6\n" // L1 l7 L6 L5 L4 L3 L2 L1 L0 | ||
127 | - "pmullw _MMX_Ublu5x5, %%mm0\n" // Cb2blue | ||
128 | - "pand _MMX_00FFw, %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0 | ||
129 | - "pmullw _MMX_Vgrn565, %%mm3\n" // Cr2green | ||
130 | + "pmullw %11, %%mm0\n" // Cb2blue | ||
131 | + "pand %12, %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0 | ||
132 | + "pmullw %13, %%mm3\n" // Cr2green | ||
133 | "movq (%2), %%mm7\n" // L2 | ||
134 | - "pmullw _MMX_Vred5x5, %%mm1\n" // Cr2red | ||
135 | + "pmullw %14, %%mm1\n" // Cr2red | ||
136 | "psrlw $8, %%mm7\n" // L2 00 L7 00 L5 00 L3 00 L1 | ||
137 | - "pmullw _MMX_Ycoeff, %%mm6\n" // lum1 | ||
138 | + "pmullw %15, %%mm6\n" // lum1 | ||
139 | "paddw %%mm3, %%mm2\n" // Cb2green + Cr2green == green | ||
140 | - "pmullw _MMX_Ycoeff, %%mm7\n" // lum2 | ||
141 | + "pmullw %15, %%mm7\n" // lum2 | ||
142 | |||
143 | "movq %%mm6, %%mm4\n" // lum1 | ||
144 | "paddw %%mm0, %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0 | ||
145 | @@ -295,11 +297,11 @@ void Color565DitherYV12MMX1X( int *color | ||
146 | "punpcklbw %%mm4, %%mm4\n" | ||
147 | "punpcklbw %%mm5, %%mm5\n" | ||
148 | |||
149 | - "pand _MMX_red565, %%mm4\n" | ||
150 | + "pand %16, %%mm4\n" | ||
151 | "psllw $3, %%mm5\n" // GREEN 1 | ||
152 | "punpcklbw %%mm6, %%mm6\n" | ||
153 | - "pand _MMX_grn565, %%mm5\n" | ||
154 | - "pand _MMX_red565, %%mm6\n" | ||
155 | + "pand %17, %%mm5\n" | ||
156 | + "pand %16, %%mm6\n" | ||
157 | "por %%mm5, %%mm4\n" // | ||
158 | "psrlw $11, %%mm6\n" // BLUE 1 | ||
159 | "movq %%mm3, %%mm5\n" // lum2 | ||
160 | @@ -313,23 +315,23 @@ void Color565DitherYV12MMX1X( int *color | ||
161 | "packuswb %%mm3, %%mm3\n" | ||
162 | "packuswb %%mm5, %%mm5\n" | ||
163 | "packuswb %%mm7, %%mm7\n" | ||
164 | - "pand _MMX_00FFw, %%mm6\n" // L3 | ||
165 | + "pand %12, %%mm6\n" // L3 | ||
166 | "punpcklbw %%mm3, %%mm3\n" | ||
167 | "punpcklbw %%mm5, %%mm5\n" | ||
168 | - "pmullw _MMX_Ycoeff, %%mm6\n" // lum3 | ||
169 | + "pmullw %15, %%mm6\n" // lum3 | ||
170 | "punpcklbw %%mm7, %%mm7\n" | ||
171 | "psllw $3, %%mm5\n" // GREEN 2 | ||
172 | - "pand _MMX_red565, %%mm7\n" | ||
173 | - "pand _MMX_red565, %%mm3\n" | ||
174 | + "pand %16, %%mm7\n" | ||
175 | + "pand %16, %%mm3\n" | ||
176 | "psrlw $11, %%mm7\n" // BLUE 2 | ||
177 | - "pand _MMX_grn565, %%mm5\n" | ||
178 | + "pand %17, %%mm5\n" | ||
179 | "por %%mm7, %%mm3\n" | ||
180 | "movq (%2,%4), %%mm7\n" // L4 load lum2 | ||
181 | "por %%mm5, %%mm3\n" // | ||
182 | "psrlw $8, %%mm7\n" // L4 | ||
183 | "movq %%mm4, %%mm5\n" | ||
184 | "punpcklwd %%mm3, %%mm4\n" | ||
185 | - "pmullw _MMX_Ycoeff, %%mm7\n" // lum4 | ||
186 | + "pmullw %15, %%mm7\n" // lum4 | ||
187 | "punpckhwd %%mm3, %%mm5\n" | ||
188 | |||
189 | "movq %%mm4, (%3)\n" // write row1 | ||
190 | @@ -356,11 +358,11 @@ void Color565DitherYV12MMX1X( int *color | ||
191 | "punpcklbw %%mm5, %%mm5\n" | ||
192 | "punpcklbw %%mm6, %%mm6\n" | ||
193 | "psllw $3, %%mm5\n" // GREEN 3 | ||
194 | - "pand _MMX_red565, %%mm4\n" | ||
195 | + "pand %16, %%mm4\n" | ||
196 | "psraw $6, %%mm3\n" // psr 6 | ||
197 | "psraw $6, %%mm0\n" | ||
198 | - "pand _MMX_red565, %%mm6\n" // BLUE | ||
199 | - "pand _MMX_grn565, %%mm5\n" | ||
200 | + "pand %16, %%mm6\n" // BLUE | ||
201 | + "pand %17, %%mm5\n" | ||
202 | "psrlw $11, %%mm6\n" // BLUE 3 | ||
203 | "por %%mm5, %%mm4\n" | ||
204 | "psraw $6, %%mm7\n" | ||
205 | @@ -371,11 +373,11 @@ void Color565DitherYV12MMX1X( int *color | ||
206 | "punpcklbw %%mm3, %%mm3\n" | ||
207 | "punpcklbw %%mm0, %%mm0\n" | ||
208 | "punpcklbw %%mm7, %%mm7\n" | ||
209 | - "pand _MMX_red565, %%mm3\n" | ||
210 | - "pand _MMX_red565, %%mm7\n" // BLUE | ||
211 | + "pand %16, %%mm3\n" | ||
212 | + "pand %16, %%mm7\n" // BLUE | ||
213 | "psllw $3, %%mm0\n" // GREEN 4 | ||
214 | "psrlw $11, %%mm7\n" | ||
215 | - "pand _MMX_grn565, %%mm0\n" | ||
216 | + "pand %17, %%mm0\n" | ||
217 | "por %%mm7, %%mm3\n" | ||
218 | "por %%mm0, %%mm3\n" | ||
219 | |||
220 | @@ -407,7 +409,10 @@ void Color565DitherYV12MMX1X( int *color | ||
221 | "popl %%ebx\n" | ||
222 | : | ||
223 | :"m" (cr), "r"(cb),"r"(lum), | ||
224 | - "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)); | ||
225 | + "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod), | ||
226 | + "m"(MMX_0080w),"m"(MMX_Ugrn565),"m"(MMX_Ublu5x5), | ||
227 | + "m"(MMX_00FFw),"m"(MMX_Vgrn565),"m"(MMX_Vred5x5), | ||
228 | + "m"(MMX_Ycoeff),"m"(MMX_red565),"m"(MMX_grn565)); | ||
229 | } | ||
230 | |||
231 | #endif /* GCC i386 inline assembly */ |