Mercurial > sdl-ios-xcode
annotate src/video/SDL_yuv_mmx.c @ 1525:23a347cfbed8
Fixed bug #38
I'm using SDL 1.2.9 with Visual C++ 7.0 on Windows 2000.
Here's the setup: my game starts in a window, with
SDL_WM_GrabInput(SDL_GRAB_ON) to constrain the cursor to the game window. The
mouse cursor is outside of the window when the game launches, and when the
window appears the cursor is grabbed and placed at the top left corner of the
inside of the game window. At this point, if I click the mouse without moving
it, the SDL_MOUSEBUTTONDOWN event's mouse coordinates are (65535,65535).
| author | Sam Lantinga <slouken@libsdl.org> |
|---|---|
| date | Tue, 14 Mar 2006 06:00:30 +0000 |
| parents | 40edc79b0926 |
| children | 782fd950bd46 c121d94672cb f420bba13676 |
| 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 */ |
|
1402
d910939febfa
Use consistent identifiers for the various platforms we support.
Sam Lantinga <slouken@libsdl.org>
parents:
1361
diff
changeset
|
22 #include "SDL_config.h" |
| 0 | 23 |
| 1413 | 24 #if 0 /* FIXME: This code needs to be rewritten to reference the static data using relocatable addresses (e.g. http://www.gentoo.org/proj/en/hardened/pic-fix-guide.xml or http://nasm.sourceforge.net/doc/html/nasmdoc8.html#section-8.2) This code currently breaks on systems with readonly text segments (hardened Linux / Intel Mac) */ |
|
1402
d910939febfa
Use consistent identifiers for the various platforms we support.
Sam Lantinga <slouken@libsdl.org>
parents:
1361
diff
changeset
|
25 #if defined(__GNUC__) && defined(__i386__) && SDL_ASSEMBLY_ROUTINES |
| 0 | 26 |
|
1407
0c6941483cc6
Whoops, forgot to check in this fix
Sam Lantinga <slouken@libsdl.org>
parents:
1402
diff
changeset
|
27 #include "SDL_stdinc.h" |
|
0c6941483cc6
Whoops, forgot to check in this fix
Sam Lantinga <slouken@libsdl.org>
parents:
1402
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 #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
|
30 |
|
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
|
31 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
|
32 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
|
33 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
|
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_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
|
36 |
|
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_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
|
38 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
|
39 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
|
40 static unsigned short ASM_ARRAY(MMX_VgrnRGB) = {0xffd2, 0xffd2, 0xffd2, 0xffd2}; |
| 0 | 41 |
|
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
|
42 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
|
43 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
|
44 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
|
45 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
|
46 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
|
47 static unsigned short ASM_ARRAY(MMX_Vgrn565) = {0xffcd, 0xffcd, 0xffcd, 0xffcd}; |
| 0 | 48 |
|
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
|
49 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
|
50 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
|
51 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
|
52 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
|
53 static unsigned short ASM_ARRAY(MMX_blu5x5) = {0x001f, 0x001f, 0x001f, 0x001f}; |
| 0 | 54 |
| 55 /** | |
| 56 This MMX assembler is my first assembler/MMX program ever. | |
| 57 Thus it maybe buggy. | |
| 58 Send patches to: | |
| 59 mvogt@rhrk.uni-kl.de | |
| 60 | |
| 61 After it worked fine I have "obfuscated" the code a bit to have | |
| 62 more parallism in the MMX units. This means I moved | |
| 63 initilisation around and delayed other instruction. | |
| 64 Performance measurement did not show that this brought any advantage | |
| 65 but in theory it _should_ be faster this way. | |
| 66 | |
| 67 The overall performanve gain to the C based dither was 30%-40%. | |
| 68 The MMX routine calculates 256bit=8RGB values in each cycle | |
| 69 (4 for row1 & 4 for row2) | |
| 70 | |
| 71 The red/green/blue.. coefficents are taken from the mpeg_play | |
| 72 player. They look nice, but I dont know if you can have | |
| 73 better values, to avoid integer rounding errors. | |
| 74 | |
| 75 | |
| 76 IMPORTANT: | |
| 77 ========== | |
| 78 | |
| 79 It is a requirement that the cr/cb/lum are 8 byte aligned and | |
| 80 the out are 16byte aligned or you will/may get segfaults | |
| 81 | |
| 82 */ | |
| 83 | |
| 84 void ColorRGBDitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix, | |
| 85 unsigned char *lum, unsigned char *cr, | |
| 86 unsigned char *cb, unsigned char *out, | |
| 87 int rows, int cols, int mod ) | |
| 88 { | |
| 89 Uint32 *row1; | |
| 90 Uint32 *row2; | |
| 91 | |
| 92 unsigned char* y = lum +cols*rows; // Pointer to the end | |
| 93 int x=0; | |
| 94 row1 = (Uint32 *)out; // 32 bit target | |
| 95 row2 = (Uint32 *)out+cols+mod; // start of second row | |
| 96 mod = (mod+cols+mod)*4; // increment for row1 in byte | |
| 97 | |
| 98 __asm__ __volatile__ ( | |
| 99 /* We don't really care about PIC - the code should be rewritten to use | |
| 100 relative addressing for the static tables, so right now we take the | |
| 101 COW hit on the pages this code resides. Big deal. | |
| 102 This spill is just to reduce register pressure in the PIC case. */ | |
| 103 "pushl %%ebx\n" | |
| 104 "movl %0, %%ebx\n" | |
| 105 | |
| 106 ".align 8\n" | |
| 107 "1:\n" | |
| 108 | |
| 109 // create Cr (result in mm1) | |
| 110 "movd (%%ebx), %%mm1\n" // 0 0 0 0 v3 v2 v1 v0 | |
| 111 "pxor %%mm7,%%mm7\n" // 00 00 00 00 00 00 00 00 | |
| 112 "movd (%2), %%mm2\n" // 0 0 0 0 l3 l2 l1 l0 | |
| 113 "punpcklbw %%mm7,%%mm1\n" // 0 v3 0 v2 00 v1 00 v0 | |
| 114 "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
|
115 "psubw _MMX_0080w,%%mm1\n" // mm1-128:r1 r1 r0 r0 r1 r1 r0 r0 |
| 0 | 116 |
| 117 // create Cr_g (result in mm0) | |
| 118 "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
|
119 "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
|
120 "pmullw _MMX_VredRGB,%%mm1\n"// red*89dec=1.4013*64 |
| 0 | 121 "psraw $6, %%mm0\n" // red=red/64 |
| 122 "psraw $6, %%mm1\n" // red=red/64 | |
| 123 | |
| 124 // create L1 L2 (result in mm2,mm4) | |
| 125 // L2=lum+cols | |
| 126 "movq (%2,%4),%%mm3\n" // 0 0 0 0 L3 L2 L1 L0 | |
| 127 "punpckldq %%mm3,%%mm2\n" // L3 L2 L1 L0 l3 l2 l1 l0 | |
| 128 "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
|
129 "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
|
130 "pand _MMX_00FFw,%%mm4\n" // 0 L2 0 L0 0 l2 0 l0 |
| 0 | 131 "psrlw $8,%%mm2\n" // 0 L3 0 L1 0 l3 0 l1 |
| 132 | |
| 133 // create R (result in mm6) | |
| 134 "movq %%mm2,%%mm5\n" // 0 L3 0 L1 0 l3 0 l1 | |
| 135 "movq %%mm4,%%mm6\n" // 0 L2 0 L0 0 l2 0 l0 | |
| 136 "paddsw %%mm1, %%mm5\n" // lum1+red:x R3 x R1 x r3 x r1 | |
| 137 "paddsw %%mm1, %%mm6\n" // lum1+red:x R2 x R0 x r2 x r0 | |
| 138 "packuswb %%mm5,%%mm5\n" // R3 R1 r3 r1 R3 R1 r3 r1 | |
| 139 "packuswb %%mm6,%%mm6\n" // R2 R0 r2 r0 R2 R0 r2 r0 | |
| 140 "pxor %%mm7,%%mm7\n" // 00 00 00 00 00 00 00 00 | |
| 141 "punpcklbw %%mm5,%%mm6\n" // R3 R2 R1 R0 r3 r2 r1 r0 | |
| 142 | |
| 143 // create Cb (result in mm1) | |
| 144 "movd (%1), %%mm1\n" // 0 0 0 0 u3 u2 u1 u0 | |
| 145 "punpcklbw %%mm7,%%mm1\n" // 0 u3 0 u2 00 u1 00 u0 | |
| 146 "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
|
147 "psubw _MMX_0080w,%%mm1\n" // mm1-128:u1 u1 u0 u0 u1 u1 u0 u0 |
| 0 | 148 // create Cb_g (result in mm5) |
| 149 "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
|
150 "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
|
151 "pmullw _MMX_UbluRGB,%%mm1\n" // blue*114dec=1.78125*64 |
| 0 | 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 // process lower lum | |
| 182 "punpcklbw %%mm4,%%mm1\n" // 0 r3 0 r2 0 r1 0 r0 | |
| 183 "punpcklbw %%mm4,%%mm3\n" // 0 b3 0 b2 0 b1 0 b0 | |
| 184 "movq %%mm1,%%mm2\n" // 0 r3 0 r2 0 r1 0 r0 | |
| 185 "movq %%mm3,%%mm0\n" // 0 b3 0 b2 0 b1 0 b0 | |
| 186 "punpcklwd %%mm1,%%mm3\n" // 0 r1 0 b1 0 r0 0 b0 | |
| 187 "punpckhwd %%mm2,%%mm0\n" // 0 r3 0 b3 0 r2 0 b2 | |
| 188 | |
| 189 "pxor %%mm2,%%mm2\n" // 0 0 0 0 0 0 0 0 | |
| 190 "movq %%mm7,%%mm1\n" // G3 G2 G1 G0 g3 g2 g1 g0 | |
| 191 "punpcklbw %%mm1,%%mm2\n" // g3 0 g2 0 g1 0 g0 0 | |
| 192 "punpcklwd %%mm4,%%mm2\n" // 0 0 g1 0 0 0 g0 0 | |
| 193 "por %%mm3, %%mm2\n" // 0 r1 g1 b1 0 r0 g0 b0 | |
| 194 "movq %%mm2,(%3)\n" // wrote out ! row1 | |
| 195 | |
| 196 "pxor %%mm2,%%mm2\n" // 0 0 0 0 0 0 0 0 | |
| 197 "punpcklbw %%mm1,%%mm4\n" // g3 0 g2 0 g1 0 g0 0 | |
| 198 "punpckhwd %%mm2,%%mm4\n" // 0 0 g3 0 0 0 g2 0 | |
| 199 "por %%mm0, %%mm4\n" // 0 r3 g3 b3 0 r2 g2 b2 | |
| 200 "movq %%mm4,8(%3)\n" // wrote out ! row1 | |
| 201 | |
| 202 // fill destination row2 (needed are mm6=Rr,mm7=Gg,mm5=Bb) | |
| 203 // this can be done "destructive" | |
| 204 "pxor %%mm2,%%mm2\n" // 0 0 0 0 0 0 0 0 | |
| 205 "punpckhbw %%mm2,%%mm6\n" // 0 R3 0 R2 0 R1 0 R0 | |
| 206 "punpckhbw %%mm1,%%mm5\n" // G3 B3 G2 B2 G1 B1 G0 B0 | |
| 207 "movq %%mm5,%%mm1\n" // G3 B3 G2 B2 G1 B1 G0 B0 | |
| 208 "punpcklwd %%mm6,%%mm1\n" // 0 R1 G1 B1 0 R0 G0 B0 | |
| 209 "movq %%mm1,(%5)\n" // wrote out ! row2 | |
| 210 "punpckhwd %%mm6,%%mm5\n" // 0 R3 G3 B3 0 R2 G2 B2 | |
| 211 "movq %%mm5,8(%5)\n" // wrote out ! row2 | |
| 212 | |
| 213 "addl $4,%2\n" // lum+4 | |
| 214 "leal 16(%3),%3\n" // row1+16 | |
| 215 "leal 16(%5),%5\n" // row2+16 | |
| 216 "addl $2, %%ebx\n" // cr+2 | |
| 217 "addl $2, %1\n" // cb+2 | |
| 218 | |
| 219 "addl $4,%6\n" // x+4 | |
| 220 "cmpl %4,%6\n" | |
| 221 | |
| 222 "jl 1b\n" | |
| 223 "addl %4, %2\n" // lum += cols | |
| 224 "addl %8, %3\n" // row1+= mod | |
| 225 "addl %8, %5\n" // row2+= mod | |
| 226 "movl $0, %6\n" // x=0 | |
| 227 "cmpl %7, %2\n" | |
| 228 "jl 1b\n" | |
| 229 "emms\n" | |
| 230 "popl %%ebx\n" | |
| 231 : | |
| 232 : "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
|
233 "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)); |
| 0 | 234 } |
| 235 | |
| 236 void Color565DitherYV12MMX1X( int *colortab, Uint32 *rgb_2_pix, | |
| 237 unsigned char *lum, unsigned char *cr, | |
| 238 unsigned char *cb, unsigned char *out, | |
| 239 int rows, int cols, int mod ) | |
| 240 { | |
| 241 Uint16 *row1; | |
| 242 Uint16 *row2; | |
| 243 | |
| 244 unsigned char* y = lum +cols*rows; /* Pointer to the end */ | |
| 245 int x=0; | |
| 246 row1 = (Uint16 *)out; /* 16 bit target */ | |
| 247 row2 = (Uint16 *)out+cols+mod; /* start of second row */ | |
| 248 mod = (mod+cols+mod)*2; /* increment for row1 in byte */ | |
| 249 | |
| 250 | |
| 251 __asm__ __volatile__( | |
| 252 "pushl %%ebx\n" | |
| 253 "movl %0, %%ebx\n" | |
| 254 | |
| 255 ".align 8\n" | |
| 256 "1:\n" | |
| 257 "movd (%1), %%mm0\n" // 4 Cb 0 0 0 0 u3 u2 u1 u0 | |
| 258 "pxor %%mm7, %%mm7\n" | |
| 259 "movd (%%ebx), %%mm1\n" // 4 Cr 0 0 0 0 v3 v2 v1 v0 | |
| 260 "punpcklbw %%mm7, %%mm0\n" // 4 W cb 0 u3 0 u2 0 u1 0 u0 | |
| 261 "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
|
262 "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
|
263 "psubw _MMX_0080w, %%mm1\n" |
| 0 | 264 "movq %%mm0, %%mm2\n" // Cb 0 u3 0 u2 0 u1 0 u0 |
| 265 "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
|
266 "pmullw _MMX_Ugrn565, %%mm2\n" // Cb2green 0 R3 0 R2 0 R1 0 R0 |
| 0 | 267 "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
|
268 "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
|
269 "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
|
270 "pmullw _MMX_Vgrn565, %%mm3\n" // Cr2green |
| 0 | 271 "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
|
272 "pmullw _MMX_Vred5x5, %%mm1\n" // Cr2red |
| 0 | 273 "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
|
274 "pmullw _MMX_Ycoeff, %%mm6\n" // lum1 |
| 0 | 275 "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
|
276 "pmullw _MMX_Ycoeff, %%mm7\n" // lum2 |
| 0 | 277 |
| 278 "movq %%mm6, %%mm4\n" // lum1 | |
| 279 "paddw %%mm0, %%mm6\n" // lum1 +blue 00 B6 00 B4 00 B2 00 B0 | |
| 280 "movq %%mm4, %%mm5\n" // lum1 | |
| 281 "paddw %%mm1, %%mm4\n" // lum1 +red 00 R6 00 R4 00 R2 00 R0 | |
| 282 "paddw %%mm2, %%mm5\n" // lum1 +green 00 G6 00 G4 00 G2 00 G0 | |
| 283 "psraw $6, %%mm4\n" // R1 0 .. 64 | |
| 284 "movq %%mm7, %%mm3\n" // lum2 00 L7 00 L5 00 L3 00 L1 | |
| 285 "psraw $6, %%mm5\n" // G1 - .. + | |
| 286 "paddw %%mm0, %%mm7\n" // Lum2 +blue 00 B7 00 B5 00 B3 00 B1 | |
| 287 "psraw $6, %%mm6\n" // B1 0 .. 64 | |
| 288 "packuswb %%mm4, %%mm4\n" // R1 R1 | |
| 289 "packuswb %%mm5, %%mm5\n" // G1 G1 | |
| 290 "packuswb %%mm6, %%mm6\n" // B1 B1 | |
| 291 "punpcklbw %%mm4, %%mm4\n" | |
| 292 "punpcklbw %%mm5, %%mm5\n" | |
| 293 | |
|
1038
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
294 "pand _MMX_red565, %%mm4\n" |
| 0 | 295 "psllw $3, %%mm5\n" // GREEN 1 |
| 296 "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
|
297 "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
|
298 "pand _MMX_red565, %%mm6\n" |
| 0 | 299 "por %%mm5, %%mm4\n" // |
| 300 "psrlw $11, %%mm6\n" // BLUE 1 | |
| 301 "movq %%mm3, %%mm5\n" // lum2 | |
| 302 "paddw %%mm1, %%mm3\n" // lum2 +red 00 R7 00 R5 00 R3 00 R1 | |
| 303 "paddw %%mm2, %%mm5\n" // lum2 +green 00 G7 00 G5 00 G3 00 G1 | |
| 304 "psraw $6, %%mm3\n" // R2 | |
| 305 "por %%mm6, %%mm4\n" // MM4 | |
| 306 "psraw $6, %%mm5\n" // G2 | |
| 307 "movq (%2, %4), %%mm6\n" // L3 load lum2 | |
| 308 "psraw $6, %%mm7\n" | |
| 309 "packuswb %%mm3, %%mm3\n" | |
| 310 "packuswb %%mm5, %%mm5\n" | |
| 311 "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
|
312 "pand _MMX_00FFw, %%mm6\n" // L3 |
| 0 | 313 "punpcklbw %%mm3, %%mm3\n" |
| 314 "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
|
315 "pmullw _MMX_Ycoeff, %%mm6\n" // lum3 |
| 0 | 316 "punpcklbw %%mm7, %%mm7\n" |
| 317 "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
|
318 "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
|
319 "pand _MMX_red565, %%mm3\n" |
| 0 | 320 "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
|
321 "pand _MMX_grn565, %%mm5\n" |
| 0 | 322 "por %%mm7, %%mm3\n" |
| 323 "movq (%2,%4), %%mm7\n" // L4 load lum2 | |
| 324 "por %%mm5, %%mm3\n" // | |
| 325 "psrlw $8, %%mm7\n" // L4 | |
| 326 "movq %%mm4, %%mm5\n" | |
| 327 "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
|
328 "pmullw _MMX_Ycoeff, %%mm7\n" // lum4 |
| 0 | 329 "punpckhwd %%mm3, %%mm5\n" |
| 330 | |
| 331 "movq %%mm4, (%3)\n" // write row1 | |
| 332 "movq %%mm5, 8(%3)\n" // write row1 | |
| 333 | |
| 334 "movq %%mm6, %%mm4\n" // Lum3 | |
| 335 "paddw %%mm0, %%mm6\n" // Lum3 +blue | |
| 336 | |
| 337 "movq %%mm4, %%mm5\n" // Lum3 | |
| 338 "paddw %%mm1, %%mm4\n" // Lum3 +red | |
| 339 "paddw %%mm2, %%mm5\n" // Lum3 +green | |
| 340 "psraw $6, %%mm4\n" | |
| 341 "movq %%mm7, %%mm3\n" // Lum4 | |
| 342 "psraw $6, %%mm5\n" | |
| 343 "paddw %%mm0, %%mm7\n" // Lum4 +blue | |
| 344 "psraw $6, %%mm6\n" // Lum3 +blue | |
| 345 "movq %%mm3, %%mm0\n" // Lum4 | |
| 346 "packuswb %%mm4, %%mm4\n" | |
| 347 "paddw %%mm1, %%mm3\n" // Lum4 +red | |
| 348 "packuswb %%mm5, %%mm5\n" | |
| 349 "paddw %%mm2, %%mm0\n" // Lum4 +green | |
| 350 "packuswb %%mm6, %%mm6\n" | |
| 351 "punpcklbw %%mm4, %%mm4\n" | |
| 352 "punpcklbw %%mm5, %%mm5\n" | |
| 353 "punpcklbw %%mm6, %%mm6\n" | |
| 354 "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
|
355 "pand _MMX_red565, %%mm4\n" |
| 0 | 356 "psraw $6, %%mm3\n" // psr 6 |
| 357 "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
|
358 "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
|
359 "pand _MMX_grn565, %%mm5\n" |
| 0 | 360 "psrlw $11, %%mm6\n" // BLUE 3 |
| 361 "por %%mm5, %%mm4\n" | |
| 362 "psraw $6, %%mm7\n" | |
| 363 "por %%mm6, %%mm4\n" | |
| 364 "packuswb %%mm3, %%mm3\n" | |
| 365 "packuswb %%mm0, %%mm0\n" | |
| 366 "packuswb %%mm7, %%mm7\n" | |
| 367 "punpcklbw %%mm3, %%mm3\n" | |
| 368 "punpcklbw %%mm0, %%mm0\n" | |
| 369 "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
|
370 "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
|
371 "pand _MMX_red565, %%mm7\n" // BLUE |
| 0 | 372 "psllw $3, %%mm0\n" // GREEN 4 |
| 373 "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
|
374 "pand _MMX_grn565, %%mm0\n" |
| 0 | 375 "por %%mm7, %%mm3\n" |
| 376 "por %%mm0, %%mm3\n" | |
| 377 | |
| 378 "movq %%mm4, %%mm5\n" | |
| 379 | |
| 380 "punpcklwd %%mm3, %%mm4\n" | |
| 381 "punpckhwd %%mm3, %%mm5\n" | |
| 382 | |
| 383 "movq %%mm4, (%5)\n" | |
| 384 "movq %%mm5, 8(%5)\n" | |
| 385 | |
| 386 "addl $8, %6\n" | |
| 387 "addl $8, %2\n" | |
| 388 "addl $4, %%ebx\n" | |
| 389 "addl $4, %1\n" | |
| 390 "cmpl %4, %6\n" | |
| 391 "leal 16(%3), %3\n" | |
| 392 "leal 16(%5),%5\n" // row2+16 | |
| 393 | |
| 394 | |
| 395 "jl 1b\n" | |
| 396 "addl %4, %2\n" // lum += cols | |
| 397 "addl %8, %3\n" // row1+= mod | |
| 398 "addl %8, %5\n" // row2+= mod | |
| 399 "movl $0, %6\n" // x=0 | |
| 400 "cmpl %7, %2\n" | |
| 401 "jl 1b\n" | |
| 402 "emms\n" | |
| 403 "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
|
404 : |
|
29d7db09776e
Ugly hack to make this work with gcc 2.x and 3.x
Sam Lantinga <slouken@libsdl.org>
parents:
949
diff
changeset
|
405 :"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
|
406 "r"(row1),"r"(cols),"r"(row2),"m"(x),"m"(y),"m"(mod)); |
| 0 | 407 } |
| 408 | |
| 409 #endif /* GCC i386 inline assembly */ | |
| 1413 | 410 #endif /* 0 */ |