Mercurial > sdl-ios-xcode
annotate src/video/SDL_yuv_mmx.c @ 1358:c71e05b4dc2e
More header massaging... works great on Windows. ;-)
| author | Sam Lantinga <slouken@libsdl.org> |
|---|---|
| date | Fri, 10 Feb 2006 06:48:43 +0000 |
| parents | c9b51268668f |
| children | 19418e4422cb |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 2 SDL - Simple DirectMedia Layer | |
|
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1148
diff
changeset
|
3 Copyright (C) 1997-2006 Sam Lantinga |
| 0 | 4 |
| 5 This library is free software; you can redistribute it and/or | |
|
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1148
diff
changeset
|
6 modify it under the terms of the GNU Lesser General Public |
| 0 | 7 License as published by the Free Software Foundation; either |
|
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1148
diff
changeset
|
8 version 2.1 of the License, or (at your option) any later version. |
| 0 | 9 |
| 10 This library is distributed in the hope that it will be useful, | |
| 11 but WITHOUT ANY WARRANTY; without even the implied warranty of | |
| 12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
|
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1148
diff
changeset
|
13 Lesser General Public License for more details. |
| 0 | 14 |
|
1312
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1148
diff
changeset
|
15 You should have received a copy of the GNU Lesser General Public |
|
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1148
diff
changeset
|
16 License along with this library; if not, write to the Free Software |
|
c9b51268668f
Updated copyright information and removed rcs id lines (problematic in branch merges)
Sam Lantinga <slouken@libsdl.org>
parents:
1148
diff
changeset
|
17 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA |
| 0 | 18 |
| 19 Sam Lantinga | |
|
252
e8157fcb3114
Updated the source with the correct e-mail address
Sam Lantinga <slouken@libsdl.org>
parents:
0
diff
changeset
|
20 slouken@libsdl.org |
| 0 | 21 */ |
| 22 | |
|
1358
c71e05b4dc2e
More header massaging... works great on Windows. ;-)
Sam Lantinga <slouken@libsdl.org>
parents:
1312
diff
changeset
|
23 #include "SDL_stdinc.h" |
| 0 | 24 |
| 25 #if defined(i386) && defined(__GNUC__) && defined(USE_ASMBLIT) | |
| 26 | |
|
1148
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
27 #define ASM_ARRAY(x) x[] __asm__("_" #x) __attribute__((used)) |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
28 |
|
1148
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
29 static unsigned int ASM_ARRAY(MMX_0080w) = {0x00800080, 0x00800080}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
30 static unsigned int ASM_ARRAY(MMX_00FFw) = {0x00ff00ff, 0x00ff00ff}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
31 static unsigned int ASM_ARRAY(MMX_FF00w) = {0xff00ff00, 0xff00ff00}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
32 |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
33 static unsigned short ASM_ARRAY(MMX_Ycoeff) = {0x004a, 0x004a, 0x004a, 0x004a}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
34 |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
35 static unsigned short ASM_ARRAY(MMX_UbluRGB) = {0x0072, 0x0072, 0x0072, 0x0072}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
36 static unsigned short ASM_ARRAY(MMX_VredRGB) = {0x0059, 0x0059, 0x0059, 0x0059}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
37 static unsigned short ASM_ARRAY(MMX_UgrnRGB) = {0xffea, 0xffea, 0xffea, 0xffea}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
38 static unsigned short ASM_ARRAY(MMX_VgrnRGB) = {0xffd2, 0xffd2, 0xffd2, 0xffd2}; |
| 0 | 39 |
|
1148
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
40 static unsigned short ASM_ARRAY(MMX_Ublu5x5) = {0x0081, 0x0081, 0x0081, 0x0081}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
41 static unsigned short ASM_ARRAY(MMX_Vred5x5) = {0x0066, 0x0066, 0x0066, 0x0066}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
42 static unsigned short ASM_ARRAY(MMX_Ugrn555) = {0xffe7, 0xffe7, 0xffe7, 0xffe7}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
43 static unsigned short ASM_ARRAY(MMX_Vgrn555) = {0xffcc, 0xffcc, 0xffcc, 0xffcc}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
44 static unsigned short ASM_ARRAY(MMX_Ugrn565) = {0xffe8, 0xffe8, 0xffe8, 0xffe8}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
45 static unsigned short ASM_ARRAY(MMX_Vgrn565) = {0xffcd, 0xffcd, 0xffcd, 0xffcd}; |
| 0 | 46 |
|
1148
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
47 static unsigned short ASM_ARRAY(MMX_red555) = {0x7c00, 0x7c00, 0x7c00, 0x7c00}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
48 static unsigned short ASM_ARRAY(MMX_red565) = {0xf800, 0xf800, 0xf800, 0xf800}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
49 static unsigned short ASM_ARRAY(MMX_grn555) = {0x03e0, 0x03e0, 0x03e0, 0x03e0}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
50 static unsigned short ASM_ARRAY(MMX_grn565) = {0x07e0, 0x07e0, 0x07e0, 0x07e0}; |
|
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
51 static unsigned short ASM_ARRAY(MMX_blu5x5) = {0x001f, 0x001f, 0x001f, 0x001f}; |
| 0 | 52 |
| 53 /** | |
| 54 This MMX assembler is my first assembler/MMX program ever. | |
| 55 Thus it maybe buggy. | |
| 56 Send patches to: | |
| 57 mvogt@rhrk.uni-kl.de | |
| 58 | |
| 59 After it worked fine I have "obfuscated" the code a bit to have | |
| 60 more parallism in the MMX units. This means I moved | |
| 61 initilisation around and delayed other instruction. | |
| 62 Performance measurement did not show that this brought any advantage | |
| 63 but in theory it _should_ be faster this way. | |
| 64 | |
| 65 The overall performanve gain to the C based dither was 30%-40%. | |
| 66 The MMX routine calculates 256bit=8RGB values in each cycle | |
| 67 (4 for row1 & 4 for row2) | |
| 68 | |
| 69 The red/green/blue.. coefficents are taken from the mpeg_play | |
| 70 player. They look nice, but I dont know if you can have | |
| 71 better values, to avoid integer rounding errors. | |
| 72 | |
| 73 | |
| 74 IMPORTANT: | |
| 75 ========== | |
| 76 | |
| 77 It is a requirement that the cr/cb/lum are 8 byte aligned and | |
| 78 the out are 16byte aligned or you will/may get segfaults | |
| 79 | |
| 80 */ | |
| 81 | |
| 82 void ColorRGBDitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix, | |
| 83 unsigned char *lum, unsigned char *cr, | |
| 84 unsigned char *cb, unsigned char *out, | |
| 85 int rows, int cols, int mod ) | |
| 86 { | |
| 87 Uint32 *row1; | |
| 88 Uint32 *row2; | |
| 89 | |
| 90 unsigned char* y = lum +cols*rows; // Pointer to the end | |
| 91 int x=0; | |
| 92 row1 = (Uint32 *)out; // 32 bit target | |
| 93 row2 = (Uint32 *)out+cols+mod; // start of second row | |
| 94 mod = (mod+cols+mod)*4; // increment for row1 in byte | |
| 95 | |
| 96 __asm__ __volatile__ ( | |
| 97 /* We don't really care about PIC - the code should be rewritten to use | |
| 98 relative addressing for the static tables, so right now we take the | |
| 99 COW hit on the pages this code resides. Big deal. | |
| 100 This spill is just to reduce register pressure in the PIC case. */ | |
| 101 "pushl %%ebx\n" | |
| 102 "movl %0, %%ebx\n" | |
| 103 | |
| 104 ".align 8\n" | |
| 105 "1:\n" | |
| 106 | |
| 107 // create Cr (result in mm1) | |
| 108 "movd (%%ebx), %%mm1\n" // 0 0 0 0 v3 v2 v1 v0 | |
| 109 "pxor %%mm7,%%mm7\n" // 00 00 00 00 00 00 00 00 | |
| 110 "movd (%2), %%mm2\n" // 0 0 0 0 l3 l2 l1 l0 | |
| 111 "punpcklbw %%mm7,%%mm1\n" // 0 v3 0 v2 00 v1 00 v0 | |
| 112 "punpckldq %%mm1,%%mm1\n" // 00 v1 00 v0 00 v1 00 v0 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
113 "psubw _MMX_0080w,%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 |
| 0 | 114 |
| 115 // create Cr_g (result in mm0) | |
| 116 "movq %%mm1,%%mm0\n" // r1 r1 r0 r0 r1 r1 r0 r0 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
117 "pmullw _MMX_VgrnRGB,%%mm0\n"// red*-46dec=0.7136*64 |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
118 "pmullw _MMX_VredRGB,%%mm1\n"// red*89dec=1.4013*64 |
| 0 | 119 "psraw $6, %%mm0\n" // red=red/64 |
| 120 "psraw $6, %%mm1\n" // red=red/64 | |
| 121 | |
| 122 // create L1 L2 (result in mm2,mm4) | |
| 123 // L2=lum+cols | |
| 124 "movq (%2,%4),%%mm3\n" // 0 0 0 0 L3 L2 L1 L0 | |
| 125 "punpckldq %%mm3,%%mm2\n" // L3 L2 L1 L0 l3 l2 l1 l0 | |
| 126 "movq %%mm2,%%mm4\n" // L3 L2 L1 L0 l3 l2 l1 l0 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
127 "pand _MMX_FF00w,%%mm2\n" // L3 0 L1 0 l3 0 l1 0 |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
128 "pand _MMX_00FFw,%%mm4\n" // 0 L2 0 L0 0 l2 0 l0 |
| 0 | 129 "psrlw $8,%%mm2\n" // 0 L3 0 L1 0 l3 0 l1 |
| 130 | |
| 131 // create R (result in mm6) | |
| 132 "movq %%mm2,%%mm5\n" // 0 L3 0 L1 0 l3 0 l1 | |
| 133 "movq %%mm4,%%mm6\n" // 0 L2 0 L0 0 l2 0 l0 | |
| 134 "paddsw %%mm1, %%mm5\n" // lum1+red:x R3 x R1 x r3 x r1 | |
| 135 "paddsw %%mm1, %%mm6\n" // lum1+red:x R2 x R0 x r2 x r0 | |
| 136 "packuswb %%mm5,%%mm5\n" // R3 R1 r3 r1 R3 R1 r3 r1 | |
| 137 "packuswb %%mm6,%%mm6\n" // R2 R0 r2 r0 R2 R0 r2 r0 | |
| 138 "pxor %%mm7,%%mm7\n" // 00 00 00 00 00 00 00 00 | |
| 139 "punpcklbw %%mm5,%%mm6\n" // R3 R2 R1 R0 r3 r2 r1 r0 | |
| 140 | |
| 141 // create Cb (result in mm1) | |
| 142 "movd (%1), %%mm1\n" // 0 0 0 0 u3 u2 u1 u0 | |
| 143 "punpcklbw %%mm7,%%mm1\n" // 0 u3 0 u2 00 u1 00 u0 | |
| 144 "punpckldq %%mm1,%%mm1\n" // 00 u1 00 u0 00 u1 00 u0 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
145 "psubw _MMX_0080w,%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 |
| 0 | 146 // create Cb_g (result in mm5) |
| 147 "movq %%mm1,%%mm5\n" // u1 u1 u0 u0 u1 u1 u0 u0 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
148 "pmullw _MMX_UgrnRGB,%%mm5\n" // blue*-109dec=1.7129*64 |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
149 "pmullw _MMX_UbluRGB,%%mm1\n" // blue*114dec=1.78125*64 |
| 0 | 150 "psraw $6, %%mm5\n" // blue=red/64 |
| 151 "psraw $6, %%mm1\n" // blue=blue/64 | |
| 152 | |
| 153 // create G (result in mm7) | |
| 154 "movq %%mm2,%%mm3\n" // 0 L3 0 L1 0 l3 0 l1 | |
| 155 "movq %%mm4,%%mm7\n" // 0 L2 0 L0 0 l2 0 l1 | |
| 156 "paddsw %%mm5, %%mm3\n" // lum1+Cb_g:x G3t x G1t x g3t x g1t | |
| 157 "paddsw %%mm5, %%mm7\n" // lum1+Cb_g:x G2t x G0t x g2t x g0t | |
| 158 "paddsw %%mm0, %%mm3\n" // lum1+Cr_g:x G3 x G1 x g3 x g1 | |
| 159 "paddsw %%mm0, %%mm7\n" // lum1+blue:x G2 x G0 x g2 x g0 | |
| 160 "packuswb %%mm3,%%mm3\n" // G3 G1 g3 g1 G3 G1 g3 g1 | |
| 161 "packuswb %%mm7,%%mm7\n" // G2 G0 g2 g0 G2 G0 g2 g0 | |
| 162 "punpcklbw %%mm3,%%mm7\n" // G3 G2 G1 G0 g3 g2 g1 g0 | |
| 163 | |
| 164 // create B (result in mm5) | |
| 165 "movq %%mm2,%%mm3\n" // 0 L3 0 L1 0 l3 0 l1 | |
| 166 "movq %%mm4,%%mm5\n" // 0 L2 0 L0 0 l2 0 l1 | |
| 167 "paddsw %%mm1, %%mm3\n" // lum1+blue:x B3 x B1 x b3 x b1 | |
| 168 "paddsw %%mm1, %%mm5\n" // lum1+blue:x B2 x B0 x b2 x b0 | |
| 169 "packuswb %%mm3,%%mm3\n" // B3 B1 b3 b1 B3 B1 b3 b1 | |
| 170 "packuswb %%mm5,%%mm5\n" // B2 B0 b2 b0 B2 B0 b2 b0 | |
| 171 "punpcklbw %%mm3,%%mm5\n" // B3 B2 B1 B0 b3 b2 b1 b0 | |
| 172 | |
| 173 // fill destination row1 (needed are mm6=Rr,mm7=Gg,mm5=Bb) | |
| 174 | |
| 175 "pxor %%mm2,%%mm2\n" // 0 0 0 0 0 0 0 0 | |
| 176 "pxor %%mm4,%%mm4\n" // 0 0 0 0 0 0 0 0 | |
| 177 "movq %%mm6,%%mm1\n" // R3 R2 R1 R0 r3 r2 r1 r0 | |
| 178 "movq %%mm5,%%mm3\n" // B3 B2 B1 B0 b3 b2 b1 b0 | |
| 179 // process lower lum | |
| 180 "punpcklbw %%mm4,%%mm1\n" // 0 r3 0 r2 0 r1 0 r0 | |
| 181 "punpcklbw %%mm4,%%mm3\n" // 0 b3 0 b2 0 b1 0 b0 | |
| 182 "movq %%mm1,%%mm2\n" // 0 r3 0 r2 0 r1 0 r0 | |
| 183 "movq %%mm3,%%mm0\n" // 0 b3 0 b2 0 b1 0 b0 | |
| 184 "punpcklwd %%mm1,%%mm3\n" // 0 r1 0 b1 0 r0 0 b0 | |
| 185 "punpckhwd %%mm2,%%mm0\n" // 0 r3 0 b3 0 r2 0 b2 | |
| 186 | |
| 187 "pxor %%mm2,%%mm2\n" // 0 0 0 0 0 0 0 0 | |
| 188 "movq %%mm7,%%mm1\n" // G3 G2 G1 G0 g3 g2 g1 g0 | |
| 189 "punpcklbw %%mm1,%%mm2\n" // g3 0 g2 0 g1 0 g0 0 | |
| 190 "punpcklwd %%mm4,%%mm2\n" // 0 0 g1 0 0 0 g0 0 | |
| 191 "por %%mm3, %%mm2\n" // 0 r1 g1 b1 0 r0 g0 b0 | |
| 192 "movq %%mm2,(%3)\n" // wrote out ! row1 | |
| 193 | |
| 194 "pxor %%mm2,%%mm2\n" // 0 0 0 0 0 0 0 0 | |
| 195 "punpcklbw %%mm1,%%mm4\n" // g3 0 g2 0 g1 0 g0 0 | |
| 196 "punpckhwd %%mm2,%%mm4\n" // 0 0 g3 0 0 0 g2 0 | |
| 197 "por %%mm0, %%mm4\n" // 0 r3 g3 b3 0 r2 g2 b2 | |
| 198 "movq %%mm4,8(%3)\n" // wrote out ! row1 | |
| 199 | |
| 200 // fill destination row2 (needed are mm6=Rr,mm7=Gg,mm5=Bb) | |
| 201 // this can be done "destructive" | |
| 202 "pxor %%mm2,%%mm2\n" // 0 0 0 0 0 0 0 0 | |
| 203 "punpckhbw %%mm2,%%mm6\n" // 0 R3 0 R2 0 R1 0 R0 | |
| 204 "punpckhbw %%mm1,%%mm5\n" // G3 B3 G2 B2 G1 B1 G0 B0 | |
| 205 "movq %%mm5,%%mm1\n" // G3 B3 G2 B2 G1 B1 G0 B0 | |
| 206 "punpcklwd %%mm6,%%mm1\n" // 0 R1 G1 B1 0 R0 G0 B0 | |
| 207 "movq %%mm1,(%5)\n" // wrote out ! row2 | |
| 208 "punpckhwd %%mm6,%%mm5\n" // 0 R3 G3 B3 0 R2 G2 B2 | |
| 209 "movq %%mm5,8(%5)\n" // wrote out ! row2 | |
| 210 | |
| 211 "addl $4,%2\n" // lum+4 | |
| 212 "leal 16(%3),%3\n" // row1+16 | |
| 213 "leal 16(%5),%5\n" // row2+16 | |
| 214 "addl $2, %%ebx\n" // cr+2 | |
| 215 "addl $2, %1\n" // cb+2 | |
| 216 | |
| 217 "addl $4,%6\n" // x+4 | |
| 218 "cmpl %4,%6\n" | |
| 219 | |
| 220 "jl 1b\n" | |
| 221 "addl %4, %2\n" // lum += cols | |
| 222 "addl %8, %3\n" // row1+= mod | |
| 223 "addl %8, %5\n" // row2+= mod | |
| 224 "movl $0, %6\n" // x=0 | |
| 225 "cmpl %7, %2\n" | |
| 226 "jl 1b\n" | |
| 227 "emms\n" | |
| 228 "popl %%ebx\n" | |
| 229 : | |
| 230 : "m" (cr), "r"(cb),"r"(lum), | |
|
1148
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
231 "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)); |
| 0 | 232 } |
| 233 | |
| 234 void Color565DitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix, | |
| 235 unsigned char *lum, unsigned char *cr, | |
| 236 unsigned char *cb, unsigned char *out, | |
| 237 int rows, int cols, int mod ) | |
| 238 { | |
| 239 Uint16 *row1; | |
| 240 Uint16 *row2; | |
| 241 | |
| 242 unsigned char* y = lum +cols*rows; /* Pointer to the end */ | |
| 243 int x=0; | |
| 244 row1 = (Uint16 *)out; /* 16 bit target */ | |
| 245 row2 = (Uint16 *)out+cols+mod; /* start of second row */ | |
| 246 mod = (mod+cols+mod)*2; /* increment for row1 in byte */ | |
| 247 | |
| 248 | |
| 249 __asm__ __volatile__( | |
| 250 "pushl %%ebx\n" | |
| 251 "movl %0, %%ebx\n" | |
| 252 | |
| 253 ".align 8\n" | |
| 254 "1:\n" | |
| 255 "movd (%1), %%mm0\n" // 4 Cb 0 0 0 0 u3 u2 u1 u0 | |
| 256 "pxor %%mm7, %%mm7\n" | |
| 257 "movd (%%ebx), %%mm1\n" // 4 Cr 0 0 0 0 v3 v2 v1 v0 | |
| 258 "punpcklbw %%mm7, %%mm0\n" // 4 W cb 0 u3 0 u2 0 u1 0 u0 | |
| 259 "punpcklbw %%mm7, %%mm1\n" // 4 W cr 0 v3 0 v2 0 v1 0 v0 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
260 "psubw _MMX_0080w, %%mm0\n" |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
261 "psubw _MMX_0080w, %%mm1\n" |
| 0 | 262 "movq %%mm0, %%mm2\n" // Cb 0 u3 0 u2 0 u1 0 u0 |
| 263 "movq %%mm1, %%mm3\n" // Cr | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
264 "pmullw _MMX_Ugrn565, %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0 |
| 0 | 265 "movq (%2), %%mm6\n" // L1 l7 L6 L5 L4 L3 L2 L1 L0 |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
266 "pmullw _MMX_Ublu5x5, %%mm0\n" // Cb2blue |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
267 "pand _MMX_00FFw, %%mm6\n" // L1 00 L6 00 L4 00 L2 00 L0 |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
268 "pmullw _MMX_Vgrn565, %%mm3\n" // Cr2green |
| 0 | 269 "movq (%2), %%mm7\n" // L2 |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
270 "pmullw _MMX_Vred5x5, %%mm1\n" // Cr2red |
| 0 | 271 "psrlw $8, %%mm7\n" // L2 00 L7 00 L5 00 L3 00 L1 |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
272 "pmullw _MMX_Ycoeff, %%mm6\n" // lum1 |
| 0 | 273 "paddw %%mm3, %%mm2\n" // Cb2green + Cr2green == green |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
274 "pmullw _MMX_Ycoeff, %%mm7\n" // lum2 |
| 0 | 275 |
| 276 "movq %%mm6, %%mm4\n" // lum1 | |
| 277 "paddw %%mm0, %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0 | |
| 278 "movq %%mm4, %%mm5\n" // lum1 | |
| 279 "paddw %%mm1, %%mm4\n" // lum1 +red 00 R6 00 R4 00 R2 00 R0 | |
| 280 "paddw %%mm2, %%mm5\n" // lum1 +green 00 G6 00 G4 00 G2 00 G0 | |
| 281 "psraw $6, %%mm4\n" // R1 0 .. 64 | |
| 282 "movq %%mm7, %%mm3\n" // lum2 00 L7 00 L5 00 L3 00 L1 | |
| 283 "psraw $6, %%mm5\n" // G1 - .. + | |
| 284 "paddw %%mm0, %%mm7\n" // Lum2 +blue 00 B7 00 B5 00 B3 00 B1 | |
| 285 "psraw $6, %%mm6\n" // B1 0 .. 64 | |
| 286 "packuswb %%mm4, %%mm4\n" // R1 R1 | |
| 287 "packuswb %%mm5, %%mm5\n" // G1 G1 | |
| 288 "packuswb %%mm6, %%mm6\n" // B1 B1 | |
| 289 "punpcklbw %%mm4, %%mm4\n" | |
| 290 "punpcklbw %%mm5, %%mm5\n" | |
| 291 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
292 "pand _MMX_red565, %%mm4\n" |
| 0 | 293 "psllw $3, %%mm5\n" // GREEN 1 |
| 294 "punpcklbw %%mm6, %%mm6\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
295 "pand _MMX_grn565, %%mm5\n" |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
296 "pand _MMX_red565, %%mm6\n" |
| 0 | 297 "por %%mm5, %%mm4\n" // |
| 298 "psrlw $11, %%mm6\n" // BLUE 1 | |
| 299 "movq %%mm3, %%mm5\n" // lum2 | |
| 300 "paddw %%mm1, %%mm3\n" // lum2 +red 00 R7 00 R5 00 R3 00 R1 | |
| 301 "paddw %%mm2, %%mm5\n" // lum2 +green 00 G7 00 G5 00 G3 00 G1 | |
| 302 "psraw $6, %%mm3\n" // R2 | |
| 303 "por %%mm6, %%mm4\n" // MM4 | |
| 304 "psraw $6, %%mm5\n" // G2 | |
| 305 "movq (%2, %4), %%mm6\n" // L3 load lum2 | |
| 306 "psraw $6, %%mm7\n" | |
| 307 "packuswb %%mm3, %%mm3\n" | |
| 308 "packuswb %%mm5, %%mm5\n" | |
| 309 "packuswb %%mm7, %%mm7\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
310 "pand _MMX_00FFw, %%mm6\n" // L3 |
| 0 | 311 "punpcklbw %%mm3, %%mm3\n" |
| 312 "punpcklbw %%mm5, %%mm5\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
313 "pmullw _MMX_Ycoeff, %%mm6\n" // lum3 |
| 0 | 314 "punpcklbw %%mm7, %%mm7\n" |
| 315 "psllw $3, %%mm5\n" // GREEN 2 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
316 "pand _MMX_red565, %%mm7\n" |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
317 "pand _MMX_red565, %%mm3\n" |
| 0 | 318 "psrlw $11, %%mm7\n" // BLUE 2 |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
319 "pand _MMX_grn565, %%mm5\n" |
| 0 | 320 "por %%mm7, %%mm3\n" |
| 321 "movq (%2,%4), %%mm7\n" // L4 load lum2 | |
| 322 "por %%mm5, %%mm3\n" // | |
| 323 "psrlw $8, %%mm7\n" // L4 | |
| 324 "movq %%mm4, %%mm5\n" | |
| 325 "punpcklwd %%mm3, %%mm4\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
326 "pmullw _MMX_Ycoeff, %%mm7\n" // lum4 |
| 0 | 327 "punpckhwd %%mm3, %%mm5\n" |
| 328 | |
| 329 "movq %%mm4, (%3)\n" // write row1 | |
| 330 "movq %%mm5, 8(%3)\n" // write row1 | |
| 331 | |
| 332 "movq %%mm6, %%mm4\n" // Lum3 | |
| 333 "paddw %%mm0, %%mm6\n" // Lum3 +blue | |
| 334 | |
| 335 "movq %%mm4, %%mm5\n" // Lum3 | |
| 336 "paddw %%mm1, %%mm4\n" // Lum3 +red | |
| 337 "paddw %%mm2, %%mm5\n" // Lum3 +green | |
| 338 "psraw $6, %%mm4\n" | |
| 339 "movq %%mm7, %%mm3\n" // Lum4 | |
| 340 "psraw $6, %%mm5\n" | |
| 341 "paddw %%mm0, %%mm7\n" // Lum4 +blue | |
| 342 "psraw $6, %%mm6\n" // Lum3 +blue | |
| 343 "movq %%mm3, %%mm0\n" // Lum4 | |
| 344 "packuswb %%mm4, %%mm4\n" | |
| 345 "paddw %%mm1, %%mm3\n" // Lum4 +red | |
| 346 "packuswb %%mm5, %%mm5\n" | |
| 347 "paddw %%mm2, %%mm0\n" // Lum4 +green | |
| 348 "packuswb %%mm6, %%mm6\n" | |
| 349 "punpcklbw %%mm4, %%mm4\n" | |
| 350 "punpcklbw %%mm5, %%mm5\n" | |
| 351 "punpcklbw %%mm6, %%mm6\n" | |
| 352 "psllw $3, %%mm5\n" // GREEN 3 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
353 "pand _MMX_red565, %%mm4\n" |
| 0 | 354 "psraw $6, %%mm3\n" // psr 6 |
| 355 "psraw $6, %%mm0\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
356 "pand _MMX_red565, %%mm6\n" // BLUE |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
357 "pand _MMX_grn565, %%mm5\n" |
| 0 | 358 "psrlw $11, %%mm6\n" // BLUE 3 |
| 359 "por %%mm5, %%mm4\n" | |
| 360 "psraw $6, %%mm7\n" | |
| 361 "por %%mm6, %%mm4\n" | |
| 362 "packuswb %%mm3, %%mm3\n" | |
| 363 "packuswb %%mm0, %%mm0\n" | |
| 364 "packuswb %%mm7, %%mm7\n" | |
| 365 "punpcklbw %%mm3, %%mm3\n" | |
| 366 "punpcklbw %%mm0, %%mm0\n" | |
| 367 "punpcklbw %%mm7, %%mm7\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
368 "pand _MMX_red565, %%mm3\n" |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
369 "pand _MMX_red565, %%mm7\n" // BLUE |
| 0 | 370 "psllw $3, %%mm0\n" // GREEN 4 |
| 371 "psrlw $11, %%mm7\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
372 "pand _MMX_grn565, %%mm0\n" |
| 0 | 373 "por %%mm7, %%mm3\n" |
| 374 "por %%mm0, %%mm3\n" | |
| 375 | |
| 376 "movq %%mm4, %%mm5\n" | |
| 377 | |
| 378 "punpcklwd %%mm3, %%mm4\n" | |
| 379 "punpckhwd %%mm3, %%mm5\n" | |
| 380 | |
| 381 "movq %%mm4, (%5)\n" | |
| 382 "movq %%mm5, 8(%5)\n" | |
| 383 | |
| 384 "addl $8, %6\n" | |
| 385 "addl $8, %2\n" | |
| 386 "addl $4, %%ebx\n" | |
| 387 "addl $4, %1\n" | |
| 388 "cmpl %4, %6\n" | |
| 389 "leal 16(%3), %3\n" | |
| 390 "leal 16(%5),%5\n" // row2+16 | |
| 391 | |
| 392 | |
| 393 "jl 1b\n" | |
| 394 "addl %4, %2\n" // lum += cols | |
| 395 "addl %8, %3\n" // row1+= mod | |
| 396 "addl %8, %5\n" // row2+= mod | |
| 397 "movl $0, %6\n" // x=0 | |
| 398 "cmpl %7, %2\n" | |
| 399 "jl 1b\n" | |
| 400 "emms\n" | |
| 401 "popl %%ebx\n" | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
402 : |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
403 :"m" (cr), "r"(cb),"r"(lum), |
|
1148
63fb2da89a4b
Patched inline assembly to compile on gcc 4.0.1. Details are here:
Ryan C. Gordon <icculus@icculus.org>
parents:
1038
diff
changeset
|
404 "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)); |
| 0 | 405 } |
| 406 | |
| 407 #endif /* GCC i386 inline assembly */ |