Mercurial > mplayer.hg
annotate postproc/swscale_template.c @ 2564:3d04a0991dce
cosmetic
author | nick |
---|---|
date | Tue, 30 Oct 2001 17:46:07 +0000 |
parents | f2e70944d02a |
children | a350d8bed636 |
rev | line source |
---|---|
2216 | 1 |
2 // Software scaling and colorspace conversion routines for MPlayer | |
3 | |
2269 | 4 // Orginal C implementation by A'rpi/ESP-team <arpi@thot.banki.hu> |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
5 // current version mostly by Michael Niedermayer (michaelni@gmx.at) |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
6 // the parts written by michael are under GNU GPL |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
7 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
8 #include <inttypes.h> |
2476 | 9 #include <string.h> |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
10 #include "../config.h" |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
11 #include "swscale.h" |
2520 | 12 #include "../mmx_defs.h" |
2540 | 13 #undef MOVNTQ |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
14 |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
15 //#undef HAVE_MMX2 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
16 //#undef HAVE_MMX |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
17 //#undef ARCH_X86 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
18 #define DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
19 int fullUVIpol=0; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
20 //disables the unscaled height version |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
21 int allwaysIpol=0; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
22 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
23 #define RET 0xC3 //near return opcode |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
24 /* |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
25 NOTES |
2216 | 26 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
27 known BUGS with known cause (no bugreports please!, but patches are welcome :) ) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
28 horizontal MMX2 scaler reads 1-7 samples too much (might cause a sig11) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
29 |
2326 | 30 Supported output formats BGR15 BGR16 BGR24 BGR32 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
31 BGR15 & BGR16 MMX verions support dithering |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
32 Special versions: fast Y 1:1 scaling (no interpolation in y direction) |
2216 | 33 |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
34 TODO |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
35 more intelligent missalignment avoidance for the horizontal scaler |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
36 */ |
2216 | 37 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
38 #define ABS(a) ((a) > 0 ? (a) : (-(a))) |
2469 | 39 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
40 #define MAX(a,b) ((a) < (b) ? (b) : (a)) | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
41 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
42 #ifdef HAVE_MMX2 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
43 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
44 #elif defined (HAVE_3DNOW) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
45 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
46 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
47 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
48 #ifdef HAVE_MMX2 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
49 #define MOVNTQ(a,b) "movntq " #a ", " #b " \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
50 #else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
51 #define MOVNTQ(a,b) "movq " #a ", " #b " \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
52 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
53 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
54 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
55 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
56 static uint64_t __attribute__((aligned(8))) yCoeff= 0x2568256825682568LL; |
2503 | 57 static uint64_t __attribute__((aligned(8))) vrCoeff= 0x3343334333433343LL; |
58 static uint64_t __attribute__((aligned(8))) ubCoeff= 0x40cf40cf40cf40cfLL; | |
59 static uint64_t __attribute__((aligned(8))) vgCoeff= 0xE5E2E5E2E5E2E5E2LL; | |
60 static uint64_t __attribute__((aligned(8))) ugCoeff= 0xF36EF36EF36EF36ELL; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
61 static uint64_t __attribute__((aligned(8))) w400= 0x0400040004000400LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
62 static uint64_t __attribute__((aligned(8))) w80= 0x0080008000800080LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
63 static uint64_t __attribute__((aligned(8))) w10= 0x0010001000100010LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
64 static uint64_t __attribute__((aligned(8))) bm00001111=0x00000000FFFFFFFFLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
65 static uint64_t __attribute__((aligned(8))) bm00000111=0x0000000000FFFFFFLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
66 static uint64_t __attribute__((aligned(8))) bm11111000=0xFFFFFFFFFF000000LL; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
67 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
68 static uint64_t __attribute__((aligned(8))) b16Dither= 0x0004000400040004LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
69 static uint64_t __attribute__((aligned(8))) b16Dither1=0x0004000400040004LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
70 static uint64_t __attribute__((aligned(8))) b16Dither2=0x0602060206020602LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
71 static uint64_t __attribute__((aligned(8))) g16Dither= 0x0002000200020002LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
72 static uint64_t __attribute__((aligned(8))) g16Dither1=0x0002000200020002LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
73 static uint64_t __attribute__((aligned(8))) g16Dither2=0x0301030103010301LL; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
74 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
75 static uint64_t __attribute__((aligned(8))) b16Mask= 0x001F001F001F001FLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
76 static uint64_t __attribute__((aligned(8))) g16Mask= 0x07E007E007E007E0LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
77 static uint64_t __attribute__((aligned(8))) r16Mask= 0xF800F800F800F800LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
78 static uint64_t __attribute__((aligned(8))) b15Mask= 0x001F001F001F001FLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
79 static uint64_t __attribute__((aligned(8))) g15Mask= 0x03E003E003E003E0LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
80 static uint64_t __attribute__((aligned(8))) r15Mask= 0x7C007C007C007C00LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
81 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
82 static uint64_t __attribute__((aligned(8))) temp0; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
83 static uint64_t __attribute__((aligned(8))) asm_yalpha1; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
84 static uint64_t __attribute__((aligned(8))) asm_uvalpha1; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
85 #endif |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
86 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
87 // temporary storage for 4 yuv lines: |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
88 // 16bit for now (mmx likes it more compact) |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
89 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
90 static uint16_t __attribute__((aligned(8))) pix_buf_y[4][2048]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
91 static uint16_t __attribute__((aligned(8))) pix_buf_uv[2][2048*2]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
92 #else |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
93 static uint16_t pix_buf_y[4][2048]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
94 static uint16_t pix_buf_uv[2][2048*2]; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
95 #endif |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
96 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
97 // clipping helper table for C implementations: |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
98 static unsigned char clip_table[768]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
99 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
100 // yuv->rgb conversion tables: |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
101 static int yuvtab_2568[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
102 static int yuvtab_3343[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
103 static int yuvtab_0c92[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
104 static int yuvtab_1a1e[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
105 static int yuvtab_40cf[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
106 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
107 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
108 static uint8_t funnyYCode[10000]; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
109 static uint8_t funnyUVCode[10000]; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
110 |
2469 | 111 static int canMMX2BeUsed=0; |
112 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
113 #define FULL_YSCALEYUV2RGB \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
114 "pxor %%mm7, %%mm7 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
115 "movd %6, %%mm6 \n\t" /*yalpha1*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
116 "punpcklwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
117 "punpcklwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
118 "movd %7, %%mm5 \n\t" /*uvalpha1*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
119 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
120 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
121 "xorl %%eax, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
122 "1: \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
123 "movq (%0, %%eax, 2), %%mm0 \n\t" /*buf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
124 "movq (%1, %%eax, 2), %%mm1 \n\t" /*buf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
125 "movq (%2, %%eax,2), %%mm2 \n\t" /* uvbuf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
126 "movq (%3, %%eax,2), %%mm3 \n\t" /* uvbuf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
127 "psubw %%mm1, %%mm0 \n\t" /* buf0[eax] - buf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
128 "psubw %%mm3, %%mm2 \n\t" /* uvbuf0[eax] - uvbuf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
129 "pmulhw %%mm6, %%mm0 \n\t" /* (buf0[eax] - buf1[eax])yalpha1>>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
130 "pmulhw %%mm5, %%mm2 \n\t" /* (uvbuf0[eax] - uvbuf1[eax])uvalpha1>>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
131 "psraw $4, %%mm1 \n\t" /* buf0[eax] - buf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
132 "movq 4096(%2, %%eax,2), %%mm4 \n\t" /* uvbuf0[eax+2048]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
133 "psraw $4, %%mm3 \n\t" /* uvbuf0[eax] - uvbuf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
134 "paddw %%mm0, %%mm1 \n\t" /* buf0[eax]yalpha1 + buf1[eax](1-yalpha1) >>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
135 "movq 4096(%3, %%eax,2), %%mm0 \n\t" /* uvbuf1[eax+2048]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
136 "paddw %%mm2, %%mm3 \n\t" /* uvbuf0[eax]uvalpha1 - uvbuf1[eax](1-uvalpha1)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
137 "psubw %%mm0, %%mm4 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
138 "psubw w80, %%mm1 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
139 "psubw w400, %%mm3 \n\t" /* 8(U-128)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
140 "pmulhw yCoeff, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
141 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
142 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
143 "pmulhw %%mm5, %%mm4 \n\t" /* (uvbuf0[eax+2048] - uvbuf1[eax+2048])uvalpha1>>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
144 "movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
145 "pmulhw ubCoeff, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
146 "psraw $4, %%mm0 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
147 "pmulhw ugCoeff, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
148 "paddw %%mm4, %%mm0 \n\t" /* uvbuf0[eax+2048]uvalpha1 - uvbuf1[eax+2048](1-uvalpha1)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
149 "psubw w400, %%mm0 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
150 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
151 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
152 "movq %%mm0, %%mm4 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
153 "pmulhw vrCoeff, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
154 "pmulhw vgCoeff, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
155 "paddw %%mm1, %%mm3 \n\t" /* B*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
156 "paddw %%mm1, %%mm0 \n\t" /* R*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
157 "packuswb %%mm3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
158 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
159 "packuswb %%mm0, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
160 "paddw %%mm4, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
161 "paddw %%mm2, %%mm1 \n\t" /* G*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
162 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
163 "packuswb %%mm1, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
164 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
165 #define YSCALEYUV2RGB \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
166 "movd %6, %%mm6 \n\t" /*yalpha1*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
167 "punpcklwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
168 "punpcklwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
169 "movq %%mm6, asm_yalpha1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
170 "movd %7, %%mm5 \n\t" /*uvalpha1*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
171 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
172 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
173 "movq %%mm5, asm_uvalpha1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
174 "xorl %%eax, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
175 "1: \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
176 "movq (%2, %%eax), %%mm2 \n\t" /* uvbuf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
177 "movq (%3, %%eax), %%mm3 \n\t" /* uvbuf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
178 "movq 4096(%2, %%eax), %%mm5 \n\t" /* uvbuf0[eax+2048]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
179 "movq 4096(%3, %%eax), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
180 "psubw %%mm3, %%mm2 \n\t" /* uvbuf0[eax] - uvbuf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
181 "psubw %%mm4, %%mm5 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
182 "movq asm_uvalpha1, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
183 "pmulhw %%mm0, %%mm2 \n\t" /* (uvbuf0[eax] - uvbuf1[eax])uvalpha1>>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
184 "pmulhw %%mm0, %%mm5 \n\t" /* (uvbuf0[eax+2048] - uvbuf1[eax+2048])uvalpha1>>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
185 "psraw $4, %%mm3 \n\t" /* uvbuf0[eax] - uvbuf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
186 "psraw $4, %%mm4 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
187 "paddw %%mm2, %%mm3 \n\t" /* uvbuf0[eax]uvalpha1 - uvbuf1[eax](1-uvalpha1)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
188 "paddw %%mm5, %%mm4 \n\t" /* uvbuf0[eax+2048]uvalpha1 - uvbuf1[eax+2048](1-uvalpha1)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
189 "psubw w400, %%mm3 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
190 "psubw w400, %%mm4 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
191 "movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
192 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
193 "pmulhw ugCoeff, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
194 "pmulhw vgCoeff, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
195 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
196 "movq (%0, %%eax, 2), %%mm0 \n\t" /*buf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
197 "movq (%1, %%eax, 2), %%mm1 \n\t" /*buf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
198 "movq 8(%0, %%eax, 2), %%mm6 \n\t" /*buf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
199 "movq 8(%1, %%eax, 2), %%mm7 \n\t" /*buf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
200 "psubw %%mm1, %%mm0 \n\t" /* buf0[eax] - buf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
201 "psubw %%mm7, %%mm6 \n\t" /* buf0[eax] - buf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
202 "pmulhw asm_yalpha1, %%mm0 \n\t" /* (buf0[eax] - buf1[eax])yalpha1>>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
203 "pmulhw asm_yalpha1, %%mm6 \n\t" /* (buf0[eax] - buf1[eax])yalpha1>>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
204 "psraw $4, %%mm1 \n\t" /* buf0[eax] - buf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
205 "psraw $4, %%mm7 \n\t" /* buf0[eax] - buf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
206 "paddw %%mm0, %%mm1 \n\t" /* buf0[eax]yalpha1 + buf1[eax](1-yalpha1) >>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
207 "paddw %%mm6, %%mm7 \n\t" /* buf0[eax]yalpha1 + buf1[eax](1-yalpha1) >>16*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
208 "pmulhw ubCoeff, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
209 "pmulhw vrCoeff, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
210 "psubw w80, %%mm1 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
211 "psubw w80, %%mm7 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
212 "pmulhw yCoeff, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
213 "pmulhw yCoeff, %%mm7 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
214 /* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
215 "paddw %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
216 "movq %%mm2, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
217 "movq %%mm5, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
218 "movq %%mm4, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
219 "punpcklwd %%mm2, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
220 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
221 "punpcklwd %%mm4, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
222 "paddw %%mm1, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
223 "paddw %%mm1, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
224 "paddw %%mm1, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
225 "punpckhwd %%mm0, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
226 "punpckhwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
227 "punpckhwd %%mm3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
228 "paddw %%mm7, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
229 "paddw %%mm7, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
230 "paddw %%mm7, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
231 /* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
232 "packuswb %%mm0, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
233 "packuswb %%mm6, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
234 "packuswb %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
235 "pxor %%mm7, %%mm7 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
236 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
237 #define YSCALEYUV2RGB1 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
238 "xorl %%eax, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
239 "1: \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
240 "movq (%2, %%eax), %%mm3 \n\t" /* uvbuf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
241 "movq 4096(%2, %%eax), %%mm4 \n\t" /* uvbuf0[eax+2048]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
242 "psraw $4, %%mm3 \n\t" /* uvbuf0[eax] - uvbuf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
243 "psraw $4, %%mm4 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
244 "psubw w400, %%mm3 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
245 "psubw w400, %%mm4 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
246 "movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
247 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
248 "pmulhw ugCoeff, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
249 "pmulhw vgCoeff, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
250 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
251 "movq (%1, %%eax, 2), %%mm1 \n\t" /*buf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
252 "movq 8(%1, %%eax, 2), %%mm7 \n\t" /*buf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
253 "psraw $4, %%mm1 \n\t" /* buf0[eax] - buf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
254 "psraw $4, %%mm7 \n\t" /* buf0[eax] - buf1[eax] >>4*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
255 "pmulhw ubCoeff, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
256 "pmulhw vrCoeff, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
257 "psubw w80, %%mm1 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
258 "psubw w80, %%mm7 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
259 "pmulhw yCoeff, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
260 "pmulhw yCoeff, %%mm7 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
261 /* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
262 "paddw %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
263 "movq %%mm2, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
264 "movq %%mm5, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
265 "movq %%mm4, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
266 "punpcklwd %%mm2, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
267 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
268 "punpcklwd %%mm4, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
269 "paddw %%mm1, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
270 "paddw %%mm1, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
271 "paddw %%mm1, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
272 "punpckhwd %%mm0, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
273 "punpckhwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
274 "punpckhwd %%mm3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
275 "paddw %%mm7, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
276 "paddw %%mm7, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
277 "paddw %%mm7, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
278 /* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
279 "packuswb %%mm0, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
280 "packuswb %%mm6, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
281 "packuswb %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
282 "pxor %%mm7, %%mm7 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
283 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
284 #define WRITEBGR32 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
285 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
286 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
287 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
288 "punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
289 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
290 "punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
291 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
292 "movq %%mm2, %%mm0 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
293 "movq %%mm1, %%mm3 \n\t" /* GBGBGBGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
294 "punpcklwd %%mm5, %%mm0 \n\t" /* 0RGB0RGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
295 "punpckhwd %%mm5, %%mm2 \n\t" /* 0RGB0RGB 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
296 "punpcklwd %%mm6, %%mm1 \n\t" /* 0RGB0RGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
297 "punpckhwd %%mm6, %%mm3 \n\t" /* 0RGB0RGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
298 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
299 MOVNTQ(%%mm0, (%4, %%eax, 4))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
300 MOVNTQ(%%mm2, 8(%4, %%eax, 4))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
301 MOVNTQ(%%mm1, 16(%4, %%eax, 4))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
302 MOVNTQ(%%mm3, 24(%4, %%eax, 4))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
303 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
304 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
305 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
306 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
307 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
308 #define WRITEBGR16 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
309 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
310 "movq %%mm4, %%mm3 \n\t" /* G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
311 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
312 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
313 "punpcklbw %%mm7, %%mm3 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
314 "punpcklbw %%mm7, %%mm2 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
315 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
316 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
317 "psrlw $3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
318 "psllw $3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
319 "psllw $8, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
320 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
321 "pand g16Mask, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
322 "pand r16Mask, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
323 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
324 "por %%mm3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
325 "por %%mm5, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
326 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
327 "punpckhbw %%mm7, %%mm4 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
328 "punpckhbw %%mm7, %%mm1 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
329 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
330 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
331 "psrlw $3, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
332 "psllw $3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
333 "psllw $8, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
334 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
335 "pand g16Mask, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
336 "pand r16Mask, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
337 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
338 "por %%mm4, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
339 "por %%mm6, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
340 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
341 MOVNTQ(%%mm2, (%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
342 MOVNTQ(%%mm1, 8(%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
343 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
344 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
345 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
346 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
347 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
348 #define WRITEBGR15 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
349 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
350 "movq %%mm4, %%mm3 \n\t" /* G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
351 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
352 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
353 "punpcklbw %%mm7, %%mm3 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
354 "punpcklbw %%mm7, %%mm2 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
355 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
356 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
357 "psrlw $3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
358 "psllw $2, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
359 "psllw $7, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
360 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
361 "pand g15Mask, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
362 "pand r15Mask, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
363 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
364 "por %%mm3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
365 "por %%mm5, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
366 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
367 "punpckhbw %%mm7, %%mm4 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
368 "punpckhbw %%mm7, %%mm1 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
369 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
370 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
371 "psrlw $3, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
372 "psllw $2, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
373 "psllw $7, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
374 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
375 "pand g15Mask, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
376 "pand r15Mask, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
377 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
378 "por %%mm4, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
379 "por %%mm6, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
380 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
381 MOVNTQ(%%mm2, (%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
382 MOVNTQ(%%mm1, 8(%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
383 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
384 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
385 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
386 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
387 // FIXME find a faster way to shuffle it to BGR24 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
388 #define WRITEBGR24 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
389 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
390 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
391 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
392 "punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
393 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
394 "punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
395 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
396 "movq %%mm2, %%mm0 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
397 "movq %%mm1, %%mm3 \n\t" /* GBGBGBGB 2 */\ |
2326 | 398 "punpcklwd %%mm5, %%mm0 \n\t" /* 0RGB0RGB 0 */\ |
399 "punpckhwd %%mm5, %%mm2 \n\t" /* 0RGB0RGB 1 */\ | |
400 "punpcklwd %%mm6, %%mm1 \n\t" /* 0RGB0RGB 2 */\ | |
401 "punpckhwd %%mm6, %%mm3 \n\t" /* 0RGB0RGB 3 */\ | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
402 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
403 "movq %%mm0, %%mm4 \n\t" /* 0RGB0RGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
404 "psrlq $8, %%mm0 \n\t" /* 00RGB0RG 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
405 "pand bm00000111, %%mm4 \n\t" /* 00000RGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
406 "pand bm11111000, %%mm0 \n\t" /* 00RGB000 0.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
407 "por %%mm4, %%mm0 \n\t" /* 00RGBRGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
408 "movq %%mm2, %%mm4 \n\t" /* 0RGB0RGB 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
409 "psllq $48, %%mm2 \n\t" /* GB000000 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
410 "por %%mm2, %%mm0 \n\t" /* GBRGBRGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
411 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
412 "movq %%mm4, %%mm2 \n\t" /* 0RGB0RGB 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
413 "psrld $16, %%mm4 \n\t" /* 000R000R 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
414 "psrlq $24, %%mm2 \n\t" /* 0000RGB0 1.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
415 "por %%mm4, %%mm2 \n\t" /* 000RRGBR 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
416 "pand bm00001111, %%mm2 \n\t" /* 0000RGBR 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
417 "movq %%mm1, %%mm4 \n\t" /* 0RGB0RGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
418 "psrlq $8, %%mm1 \n\t" /* 00RGB0RG 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
419 "pand bm00000111, %%mm4 \n\t" /* 00000RGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
420 "pand bm11111000, %%mm1 \n\t" /* 00RGB000 2.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
421 "por %%mm4, %%mm1 \n\t" /* 00RGBRGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
422 "movq %%mm1, %%mm4 \n\t" /* 00RGBRGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
423 "psllq $32, %%mm1 \n\t" /* BRGB0000 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
424 "por %%mm1, %%mm2 \n\t" /* BRGBRGBR 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
425 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
426 "psrlq $32, %%mm4 \n\t" /* 000000RG 2.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
427 "movq %%mm3, %%mm5 \n\t" /* 0RGB0RGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
428 "psrlq $8, %%mm3 \n\t" /* 00RGB0RG 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
429 "pand bm00000111, %%mm5 \n\t" /* 00000RGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
430 "pand bm11111000, %%mm3 \n\t" /* 00RGB000 3.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
431 "por %%mm5, %%mm3 \n\t" /* 00RGBRGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
432 "psllq $16, %%mm3 \n\t" /* RGBRGB00 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
433 "por %%mm4, %%mm3 \n\t" /* RGBRGBRG 2.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
434 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
435 "leal (%%eax, %%eax, 2), %%ebx \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
436 MOVNTQ(%%mm0, (%4, %%ebx))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
437 MOVNTQ(%%mm2, 8(%4, %%ebx))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
438 MOVNTQ(%%mm3, 16(%4, %%ebx))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
439 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
440 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
441 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
442 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
443 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
444 |
2519 | 445 static inline void yuv2yuv(uint16_t *buf0, uint16_t *buf1, uint16_t *uvbuf0, uint16_t *uvbuf1, |
446 uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstw, int yalpha, int uvalpha) | |
447 { | |
448 int yalpha1=yalpha^4095; | |
449 int uvalpha1=uvalpha^4095; | |
450 int i; | |
451 | |
2521 | 452 asm volatile ("\n\t"::: "memory"); |
453 | |
2519 | 454 for(i=0;i<dstw;i++) |
455 { | |
456 ((uint8_t*)dest)[0] = (buf0[i]*yalpha1+buf1[i]*yalpha)>>19; | |
457 dest++; | |
458 } | |
459 | |
460 if(uvalpha != -1) | |
461 { | |
462 for(i=0; i<dstw/2; i++) | |
463 { | |
464 ((uint8_t*)uDest)[0] = (uvbuf0[i]*uvalpha1+uvbuf1[i]*uvalpha)>>19; | |
465 ((uint8_t*)vDest)[0] = (uvbuf0[i+2048]*uvalpha1+uvbuf1[i+2048]*uvalpha)>>19; | |
466 uDest++; | |
467 vDest++; | |
468 } | |
469 } | |
470 } | |
471 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
472 /** |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
473 * vertical scale YV12 to RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
474 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
475 static inline void yuv2rgbX(uint16_t *buf0, uint16_t *buf1, uint16_t *uvbuf0, uint16_t *uvbuf1, |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
476 uint8_t *dest, int dstw, int yalpha, int uvalpha, int dstbpp) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
477 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
478 int yalpha1=yalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
479 int uvalpha1=uvalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
480 int i; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
481 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
482 if(fullUVIpol) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
483 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
484 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
485 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
486 if(dstbpp == 32) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
487 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
488 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
489 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
490 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
491 FULL_YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
492 "punpcklbw %%mm1, %%mm3 \n\t" // BGBGBGBG |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
493 "punpcklbw %%mm7, %%mm0 \n\t" // R0R0R0R0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
494 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
495 "movq %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
496 "punpcklwd %%mm0, %%mm3 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
497 "punpckhwd %%mm0, %%mm1 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
498 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
499 MOVNTQ(%%mm3, (%4, %%eax, 4)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
500 MOVNTQ(%%mm1, 8(%4, %%eax, 4)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
501 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
502 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
503 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
504 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
505 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
506 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
507 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
508 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
509 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
510 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
511 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
512 else if(dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
513 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
514 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
515 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
516 FULL_YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
517 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
518 // lsb ... msb |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
519 "punpcklbw %%mm1, %%mm3 \n\t" // BGBGBGBG |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
520 "punpcklbw %%mm7, %%mm0 \n\t" // R0R0R0R0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
521 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
522 "movq %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
523 "punpcklwd %%mm0, %%mm3 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
524 "punpckhwd %%mm0, %%mm1 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
525 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
526 "movq %%mm3, %%mm2 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
527 "psrlq $8, %%mm3 \n\t" // GR0BGR00 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
528 "pand bm00000111, %%mm2 \n\t" // BGR00000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
529 "pand bm11111000, %%mm3 \n\t" // 000BGR00 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
530 "por %%mm2, %%mm3 \n\t" // BGRBGR00 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
531 "movq %%mm1, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
532 "psllq $48, %%mm1 \n\t" // 000000BG |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
533 "por %%mm1, %%mm3 \n\t" // BGRBGRBG |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
534 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
535 "movq %%mm2, %%mm1 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
536 "psrld $16, %%mm2 \n\t" // R000R000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
537 "psrlq $24, %%mm1 \n\t" // 0BGR0000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
538 "por %%mm2, %%mm1 \n\t" // RBGRR000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
539 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
540 "movl %4, %%ebx \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
541 "addl %%eax, %%ebx \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
542 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
543 #ifdef HAVE_MMX2 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
544 //FIXME Alignment |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
545 "movntq %%mm3, (%%ebx, %%eax, 2)\n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
546 "movntq %%mm1, 8(%%ebx, %%eax, 2)\n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
547 #else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
548 "movd %%mm3, (%%ebx, %%eax, 2) \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
549 "psrlq $32, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
550 "movd %%mm3, 4(%%ebx, %%eax, 2) \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
551 "movd %%mm1, 8(%%ebx, %%eax, 2) \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
552 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
553 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
554 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
555 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
556 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
557 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "m" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
558 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
559 : "%eax", "%ebx" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
560 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
561 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
562 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
563 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
564 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
565 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
566 FULL_YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
567 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
568 "paddusb b16Dither, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
569 "paddusb b16Dither, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
570 "paddusb b16Dither, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
571 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
572 "punpcklbw %%mm7, %%mm1 \n\t" // 0G0G0G0G |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
573 "punpcklbw %%mm7, %%mm3 \n\t" // 0B0B0B0B |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
574 "punpcklbw %%mm7, %%mm0 \n\t" // 0R0R0R0R |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
575 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
576 "psrlw $3, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
577 "psllw $2, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
578 "psllw $7, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
579 "pand g15Mask, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
580 "pand r15Mask, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
581 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
582 "por %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
583 "por %%mm1, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
584 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
585 MOVNTQ(%%mm0, (%4, %%eax, 2)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
586 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
587 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
588 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
589 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
590 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
591 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
592 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
593 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
594 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
595 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
596 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
597 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
598 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
599 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
600 FULL_YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
601 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
602 "paddusb g16Dither, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
603 "paddusb b16Dither, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
604 "paddusb b16Dither, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
605 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
606 "punpcklbw %%mm7, %%mm1 \n\t" // 0G0G0G0G |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
607 "punpcklbw %%mm7, %%mm3 \n\t" // 0B0B0B0B |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
608 "punpcklbw %%mm7, %%mm0 \n\t" // 0R0R0R0R |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
609 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
610 "psrlw $3, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
611 "psllw $3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
612 "psllw $8, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
613 "pand g16Mask, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
614 "pand r16Mask, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
615 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
616 "por %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
617 "por %%mm1, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
618 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
619 MOVNTQ(%%mm0, (%4, %%eax, 2)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
620 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
621 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
622 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
623 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
624 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
625 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
626 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
627 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
628 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
629 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
630 #else |
2521 | 631 asm volatile ("\n\t"::: "memory"); |
632 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
633 if(dstbpp==32 || dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
634 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
635 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
636 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
637 int Y=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
638 int U=((uvbuf0[i]*uvalpha1+uvbuf1[i]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
639 int V=((uvbuf0[i+2048]*uvalpha1+uvbuf1[i+2048]*uvalpha)>>19); |
2503 | 640 dest[0]=clip_table[((Y + yuvtab_40cf[U]) >>13)]; |
641 dest[1]=clip_table[((Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13)]; | |
642 dest[2]=clip_table[((Y + yuvtab_3343[V]) >>13)]; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
643 dest+=dstbpp>>3; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
644 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
645 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
646 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
647 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
648 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
649 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
650 int Y=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
651 int U=((uvbuf0[i]*uvalpha1+uvbuf1[i]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
652 int V=((uvbuf0[i+2048]*uvalpha1+uvbuf1[i+2048]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
653 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
654 ((uint16_t*)dest)[0] = |
2503 | 655 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
656 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<3)&0x07E0) | | |
657 ((clip_table[(Y + yuvtab_3343[V]) >>13]<<8)&0xF800); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
658 dest+=2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
659 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
660 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
661 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
662 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
663 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
664 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
665 int Y=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
666 int U=((uvbuf0[i]*uvalpha1+uvbuf1[i]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
667 int V=((uvbuf0[i+2048]*uvalpha1+uvbuf1[i+2048]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
668 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
669 ((uint16_t*)dest)[0] = |
2503 | 670 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
671 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<2)&0x03E0) | | |
672 ((clip_table[(Y + yuvtab_3343[V]) >>13]<<7)&0x7C00); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
673 dest+=2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
674 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
675 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
676 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
677 }//FULL_UV_IPOL |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
678 else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
679 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
680 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
681 if(dstbpp == 32) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
682 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
683 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
684 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
685 WRITEBGR32 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
686 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
687 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
688 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
689 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
690 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
691 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
692 else if(dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
693 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
694 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
695 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
696 WRITEBGR24 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
697 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
698 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
699 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
700 : "%eax", "%ebx" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
701 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
702 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
703 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
704 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
705 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
706 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
707 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
708 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
709 "paddusb b16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
710 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
711 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
712 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
713 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
714 WRITEBGR15 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
715 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
716 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
717 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
718 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
719 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
720 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
721 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
722 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
723 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
724 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
725 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
726 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
727 "paddusb g16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
728 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
729 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
730 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
731 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
732 WRITEBGR16 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
733 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
734 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
735 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
736 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
737 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
738 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
739 #else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
740 //FIXME unroll C loop and dont recalculate UV |
2521 | 741 asm volatile ("\n\t"::: "memory"); |
742 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
743 if(dstbpp==32 || dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
744 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
745 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
746 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
747 int Y=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
748 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
749 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
2503 | 750 dest[0]=clip_table[((Y + yuvtab_40cf[U]) >>13)]; |
751 dest[1]=clip_table[((Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13)]; | |
752 dest[2]=clip_table[((Y + yuvtab_3343[V]) >>13)]; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
753 dest+=dstbpp>>3; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
754 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
755 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
756 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
757 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
758 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
759 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
760 int Y=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
761 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
762 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
763 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
764 ((uint16_t*)dest)[0] = |
2503 | 765 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
766 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<3)&0x07E0) | | |
767 ((clip_table[(Y + yuvtab_3343[V]) >>13]<<8)&0xF800); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
768 dest+=2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
769 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
770 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
771 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
772 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
773 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
774 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
775 int Y=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
776 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
777 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
778 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
779 ((uint16_t*)dest)[0] = |
2503 | 780 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
781 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<2)&0x03E0) | | |
782 ((clip_table[(Y + yuvtab_3343[V]) >>13]<<7)&0x7C00); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
783 dest+=2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
784 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
785 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
786 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
787 } //!FULL_UV_IPOL |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
788 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
789 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
790 /** |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
791 * YV12 to RGB without scaling or interpolating |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
792 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
793 static inline void yuv2rgb1(uint16_t *buf0, uint16_t *buf1, uint16_t *uvbuf0, uint16_t *uvbuf1, |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
794 uint8_t *dest, int dstw, int yalpha, int uvalpha, int dstbpp) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
795 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
796 int yalpha1=yalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
797 int uvalpha1=uvalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
798 int i; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
799 if(fullUVIpol || allwaysIpol) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
800 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
801 yuv2rgbX(buf0, buf1, uvbuf0, uvbuf1, dest, dstw, yalpha, uvalpha, dstbpp); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
802 return; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
803 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
804 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
805 if(dstbpp == 32) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
806 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
807 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
808 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
809 WRITEBGR32 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
810 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
811 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
812 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
813 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
814 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
815 else if(dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
816 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
817 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
818 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
819 WRITEBGR24 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
820 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
821 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
822 : "%eax", "%ebx" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
823 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
824 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
825 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
826 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
827 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
828 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
829 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
830 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
831 "paddusb b16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
832 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
833 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
834 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
835 WRITEBGR15 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
836 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
837 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
838 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
839 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
840 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
841 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
842 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
843 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
844 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
845 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
846 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
847 "paddusb g16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
848 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
849 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
850 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
851 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
852 WRITEBGR16 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
853 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
854 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
855 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
856 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
857 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
858 #else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
859 //FIXME unroll C loop and dont recalculate UV |
2521 | 860 asm volatile ("\n\t"::: "memory"); |
861 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
862 if(dstbpp==32 || dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
863 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
864 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
865 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
866 int Y=yuvtab_2568[buf0[i]>>7]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
867 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
868 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
2503 | 869 dest[0]=clip_table[((Y + yuvtab_40cf[U]) >>13)]; |
870 dest[1]=clip_table[((Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13)]; | |
871 dest[2]=clip_table[((Y + yuvtab_3343[V]) >>13)]; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
872 dest+=dstbpp>>3; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
873 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
874 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
875 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
876 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
877 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
878 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
879 int Y=yuvtab_2568[buf0[i]>>7]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
880 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
881 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
882 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
883 ((uint16_t*)dest)[0] = |
2503 | 884 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
885 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<3)&0x07E0) | | |
886 ((clip_table[(Y + yuvtab_3343[V]) >>13]<<8)&0xF800); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
887 dest+=2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
888 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
889 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
890 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
891 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
892 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
893 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
894 int Y=yuvtab_2568[buf0[i]>>7]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
895 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
896 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
897 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
898 ((uint16_t*)dest)[0] = |
2503 | 899 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
900 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<2)&0x03E0) | | |
901 ((clip_table[(Y + yuvtab_3343[V]) >>13]<<7)&0x7C00); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
902 dest+=2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
903 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
904 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
905 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
906 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
907 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
908 |
2469 | 909 static inline void hyscale(uint16_t *dst, int dstWidth, uint8_t *src, int srcWidth, int xInc) |
910 { | |
911 int i; | |
912 unsigned int xpos=0; | |
913 // *** horizontal scale Y line to temp buffer | |
914 #ifdef ARCH_X86 | |
915 | |
916 #ifdef HAVE_MMX2 | |
917 if(canMMX2BeUsed) | |
918 { | |
919 asm volatile( | |
920 "pxor %%mm7, %%mm7 \n\t" | |
921 "pxor %%mm2, %%mm2 \n\t" // 2*xalpha | |
922 "movd %5, %%mm6 \n\t" // xInc&0xFFFF | |
923 "punpcklwd %%mm6, %%mm6 \n\t" | |
924 "punpcklwd %%mm6, %%mm6 \n\t" | |
925 "movq %%mm6, %%mm2 \n\t" | |
926 "psllq $16, %%mm2 \n\t" | |
927 "paddw %%mm6, %%mm2 \n\t" | |
928 "psllq $16, %%mm2 \n\t" | |
929 "paddw %%mm6, %%mm2 \n\t" | |
930 "psllq $16, %%mm2 \n\t" //0,t,2t,3t t=xInc&0xFF | |
931 "movq %%mm2, temp0 \n\t" | |
932 "movd %4, %%mm6 \n\t" //(xInc*4)&0xFFFF | |
933 "punpcklwd %%mm6, %%mm6 \n\t" | |
934 "punpcklwd %%mm6, %%mm6 \n\t" | |
935 "xorl %%eax, %%eax \n\t" // i | |
936 "movl %0, %%esi \n\t" // src | |
937 "movl %1, %%edi \n\t" // buf1 | |
938 "movl %3, %%edx \n\t" // (xInc*4)>>16 | |
939 "xorl %%ecx, %%ecx \n\t" | |
940 "xorl %%ebx, %%ebx \n\t" | |
941 "movw %4, %%bx \n\t" // (xInc*4)&0xFFFF | |
2520 | 942 |
2469 | 943 #define FUNNY_Y_CODE \ |
2520 | 944 PREFETCH" 1024(%%esi) \n\t"\ |
945 PREFETCH" 1056(%%esi) \n\t"\ | |
946 PREFETCH" 1088(%%esi) \n\t"\ | |
2469 | 947 "call funnyYCode \n\t"\ |
948 "movq temp0, %%mm2 \n\t"\ | |
949 "xorl %%ecx, %%ecx \n\t" | |
2520 | 950 |
2469 | 951 FUNNY_Y_CODE |
952 FUNNY_Y_CODE | |
953 FUNNY_Y_CODE | |
954 FUNNY_Y_CODE | |
955 FUNNY_Y_CODE | |
956 FUNNY_Y_CODE | |
957 FUNNY_Y_CODE | |
958 FUNNY_Y_CODE | |
959 | |
960 :: "m" (src), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16), | |
961 "m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF) | |
962 : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" | |
963 ); | |
964 for(i=dstWidth-1; (i*xInc)>>16 >=srcWidth-1; i--) dst[i] = src[srcWidth-1]*128; | |
965 } | |
966 else | |
967 { | |
968 #endif | |
969 //NO MMX just normal asm ... | |
970 asm volatile( | |
971 "xorl %%eax, %%eax \n\t" // i | |
972 "xorl %%ebx, %%ebx \n\t" // xx | |
973 "xorl %%ecx, %%ecx \n\t" // 2*xalpha | |
974 "1: \n\t" | |
975 "movzbl (%0, %%ebx), %%edi \n\t" //src[xx] | |
976 "movzbl 1(%0, %%ebx), %%esi \n\t" //src[xx+1] | |
977 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
978 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
979 "shll $16, %%edi \n\t" | |
980 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
981 "movl %1, %%edi \n\t" | |
982 "shrl $9, %%esi \n\t" | |
983 "movw %%si, (%%edi, %%eax, 2) \n\t" | |
984 "addw %4, %%cx \n\t" //2*xalpha += xInc&0xFF | |
985 "adcl %3, %%ebx \n\t" //xx+= xInc>>8 + carry | |
986 | |
987 "movzbl (%0, %%ebx), %%edi \n\t" //src[xx] | |
988 "movzbl 1(%0, %%ebx), %%esi \n\t" //src[xx+1] | |
989 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
990 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
991 "shll $16, %%edi \n\t" | |
992 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
993 "movl %1, %%edi \n\t" | |
994 "shrl $9, %%esi \n\t" | |
995 "movw %%si, 2(%%edi, %%eax, 2) \n\t" | |
996 "addw %4, %%cx \n\t" //2*xalpha += xInc&0xFF | |
997 "adcl %3, %%ebx \n\t" //xx+= xInc>>8 + carry | |
998 | |
999 | |
1000 "addl $2, %%eax \n\t" | |
1001 "cmpl %2, %%eax \n\t" | |
1002 " jb 1b \n\t" | |
1003 | |
1004 | |
1005 :: "r" (src), "m" (dst), "m" (dstWidth), "m" (xInc>>16), "m" (xInc&0xFFFF) | |
1006 : "%eax", "%ebx", "%ecx", "%edi", "%esi" | |
1007 ); | |
1008 #ifdef HAVE_MMX2 | |
1009 } //if MMX2 cant be used | |
1010 #endif | |
1011 #else | |
1012 for(i=0;i<dstWidth;i++){ | |
1013 register unsigned int xx=xpos>>16; | |
1014 register unsigned int xalpha=(xpos&0xFFFF)>>9; | |
1015 dst[i]=(src[xx]*(xalpha^127)+src[xx+1]*xalpha); | |
1016 xpos+=xInc; | |
1017 } | |
1018 #endif | |
1019 } | |
1020 | |
1021 inline static void hcscale(uint16_t *dst, int dstWidth, | |
1022 uint8_t *src1, uint8_t *src2, int srcWidth, int xInc) | |
1023 { | |
1024 int xpos=0; | |
1025 int i; | |
1026 #ifdef ARCH_X86 | |
1027 #ifdef HAVE_MMX2 | |
1028 if(canMMX2BeUsed) | |
1029 { | |
1030 asm volatile( | |
1031 "pxor %%mm7, %%mm7 \n\t" | |
1032 "pxor %%mm2, %%mm2 \n\t" // 2*xalpha | |
1033 "movd %5, %%mm6 \n\t" // xInc&0xFFFF | |
1034 "punpcklwd %%mm6, %%mm6 \n\t" | |
1035 "punpcklwd %%mm6, %%mm6 \n\t" | |
1036 "movq %%mm6, %%mm2 \n\t" | |
1037 "psllq $16, %%mm2 \n\t" | |
1038 "paddw %%mm6, %%mm2 \n\t" | |
1039 "psllq $16, %%mm2 \n\t" | |
1040 "paddw %%mm6, %%mm2 \n\t" | |
1041 "psllq $16, %%mm2 \n\t" //0,t,2t,3t t=xInc&0xFFFF | |
1042 "movq %%mm2, temp0 \n\t" | |
1043 "movd %4, %%mm6 \n\t" //(xInc*4)&0xFFFF | |
1044 "punpcklwd %%mm6, %%mm6 \n\t" | |
1045 "punpcklwd %%mm6, %%mm6 \n\t" | |
1046 "xorl %%eax, %%eax \n\t" // i | |
1047 "movl %0, %%esi \n\t" // src | |
1048 "movl %1, %%edi \n\t" // buf1 | |
1049 "movl %3, %%edx \n\t" // (xInc*4)>>16 | |
1050 "xorl %%ecx, %%ecx \n\t" | |
1051 "xorl %%ebx, %%ebx \n\t" | |
1052 "movw %4, %%bx \n\t" // (xInc*4)&0xFFFF | |
1053 | |
1054 #define FUNNYUVCODE \ | |
2520 | 1055 PREFETCH" 1024(%%esi) \n\t"\ |
1056 PREFETCH" 1056(%%esi) \n\t"\ | |
1057 PREFETCH" 1088(%%esi) \n\t"\ | |
2469 | 1058 "call funnyUVCode \n\t"\ |
1059 "movq temp0, %%mm2 \n\t"\ | |
1060 "xorl %%ecx, %%ecx \n\t" | |
1061 | |
1062 FUNNYUVCODE | |
1063 FUNNYUVCODE | |
1064 FUNNYUVCODE | |
1065 FUNNYUVCODE | |
1066 | |
1067 FUNNYUVCODE | |
1068 FUNNYUVCODE | |
1069 FUNNYUVCODE | |
1070 FUNNYUVCODE | |
1071 | |
1072 | |
1073 "xorl %%eax, %%eax \n\t" // i | |
1074 "movl %6, %%esi \n\t" // src | |
1075 "movl %1, %%edi \n\t" // buf1 | |
1076 "addl $4096, %%edi \n\t" | |
1077 | |
1078 FUNNYUVCODE | |
1079 FUNNYUVCODE | |
1080 FUNNYUVCODE | |
1081 FUNNYUVCODE | |
1082 | |
1083 FUNNYUVCODE | |
1084 FUNNYUVCODE | |
1085 FUNNYUVCODE | |
1086 FUNNYUVCODE | |
1087 | |
1088 :: "m" (src1), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16), | |
1089 "m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF), "m" (src2) | |
1090 : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" | |
1091 ); | |
1092 for(i=dstWidth-1; (i*xInc)>>16 >=srcWidth/2-1; i--) | |
1093 { | |
1094 dst[i] = src1[srcWidth/2-1]*128; | |
1095 dst[i+2048] = src2[srcWidth/2-1]*128; | |
1096 } | |
1097 } | |
1098 else | |
1099 { | |
1100 #endif | |
1101 asm volatile( | |
1102 "xorl %%eax, %%eax \n\t" // i | |
1103 "xorl %%ebx, %%ebx \n\t" // xx | |
1104 "xorl %%ecx, %%ecx \n\t" // 2*xalpha | |
1105 "1: \n\t" | |
1106 "movl %0, %%esi \n\t" | |
1107 "movzbl (%%esi, %%ebx), %%edi \n\t" //src[xx] | |
1108 "movzbl 1(%%esi, %%ebx), %%esi \n\t" //src[xx+1] | |
1109 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
1110 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
1111 "shll $16, %%edi \n\t" | |
1112 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
1113 "movl %1, %%edi \n\t" | |
1114 "shrl $9, %%esi \n\t" | |
1115 "movw %%si, (%%edi, %%eax, 2) \n\t" | |
1116 | |
1117 "movzbl (%5, %%ebx), %%edi \n\t" //src[xx] | |
1118 "movzbl 1(%5, %%ebx), %%esi \n\t" //src[xx+1] | |
1119 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
1120 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
1121 "shll $16, %%edi \n\t" | |
1122 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
1123 "movl %1, %%edi \n\t" | |
1124 "shrl $9, %%esi \n\t" | |
1125 "movw %%si, 4096(%%edi, %%eax, 2)\n\t" | |
1126 | |
1127 "addw %4, %%cx \n\t" //2*xalpha += xInc&0xFF | |
1128 "adcl %3, %%ebx \n\t" //xx+= xInc>>8 + carry | |
1129 "addl $1, %%eax \n\t" | |
1130 "cmpl %2, %%eax \n\t" | |
1131 " jb 1b \n\t" | |
1132 | |
1133 :: "m" (src1), "m" (dst), "m" (dstWidth), "m" (xInc>>16), "m" (xInc&0xFFFF), | |
1134 "r" (src2) | |
1135 : "%eax", "%ebx", "%ecx", "%edi", "%esi" | |
1136 ); | |
1137 #ifdef HAVE_MMX2 | |
1138 } //if MMX2 cant be used | |
1139 #endif | |
1140 #else | |
1141 for(i=0;i<dstWidth;i++){ | |
1142 register unsigned int xx=xpos>>16; | |
1143 register unsigned int xalpha=(xpos&0xFFFF)>>9; | |
1144 dst[i]=(src1[xx]*(xalpha^127)+src1[xx+1]*xalpha); | |
1145 dst[i+2048]=(src2[xx]*(xalpha^127)+src2[xx+1]*xalpha); | |
1146 xpos+=xInc; | |
1147 } | |
1148 #endif | |
1149 } | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1150 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1151 |
2519 | 1152 // *** bilinear scaling and yuv->rgb or yuv->yuv conversion of yv12 slices: |
2216 | 1153 // *** Note: it's called multiple times while decoding a frame, first time y==0 |
1154 // *** Designed to upscale, but may work for downscale too. | |
2274 | 1155 // s_xinc = (src_width << 16) / dst_width |
2216 | 1156 // s_yinc = (src_height << 16) / dst_height |
2519 | 1157 void SwScale_YV12slice(unsigned char* srcptr[],int stride[], int y, int h, |
1158 uint8_t* dstptr[], int dststride, int dstw, int dstbpp, | |
2216 | 1159 unsigned int s_xinc,unsigned int s_yinc){ |
1160 | |
1161 // scaling factors: | |
1162 //static int s_yinc=(vo_dga_src_height<<16)/vo_dga_vp_height; | |
1163 //static int s_xinc=(vo_dga_src_width<<8)/vo_dga_vp_width; | |
1164 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1165 unsigned int s_xinc2; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1166 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1167 static int s_srcypos; // points to the dst Pixels center in the source (0 is the center of pixel 0,0 in src) |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1168 static int s_ypos; |
2216 | 1169 |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1170 // last horzontally interpolated lines, used to avoid unnecessary calculations |
2216 | 1171 static int s_last_ypos; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1172 static int s_last_y1pos; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1173 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1174 static int static_dstw; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1175 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1176 #ifdef HAVE_MMX2 |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1177 // used to detect a horizontal size change |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1178 static int old_dstw= -1; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1179 static int old_s_xinc= -1; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1180 #endif |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1181 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1182 int srcWidth= (dstw*s_xinc + 0x8000)>>16; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1183 int dstUVw= fullUVIpol ? dstw : dstw/2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1184 |
2216 | 1185 |
2270 | 1186 #ifdef HAVE_MMX2 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1187 canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0 && (srcWidth&15)==0) ? 1 : 0; |
2270 | 1188 #endif |
1189 | |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1190 // match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src to pixel n-2 of dst |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1191 // n-2 is the last chrominance sample available |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1192 // FIXME this is not perfect, but noone shuld notice the difference, the more correct variant |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1193 // would be like the vertical one, but that would require some special code for the |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1194 // first and last pixel |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1195 if(canMMX2BeUsed) s_xinc+= 20; |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1196 else s_xinc = ((srcWidth-2)<<16)/(dstw-2) - 20; |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1197 |
2519 | 1198 if(fullUVIpol && !dstbpp==12) s_xinc2= s_xinc>>1; |
1199 else s_xinc2= s_xinc; | |
2271 | 1200 // force calculation of the horizontal interpolation of the first line |
1201 | |
2216 | 1202 if(y==0){ |
2469 | 1203 s_last_ypos=-99; |
1204 s_last_y1pos=-99; | |
1205 s_srcypos= s_yinc/2 - 0x8000; | |
1206 s_ypos=0; | |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1207 #ifdef HAVE_MMX2 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1208 // cant downscale !!! |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1209 if((old_s_xinc != s_xinc || old_dstw!=dstw) && canMMX2BeUsed) |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1210 { |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1211 uint8_t *fragment; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1212 int imm8OfPShufW1; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1213 int imm8OfPShufW2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1214 int fragmentLength; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1215 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1216 int xpos, xx, xalpha, i; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1217 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1218 old_s_xinc= s_xinc; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1219 old_dstw= dstw; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1220 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1221 static_dstw= dstw; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1222 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1223 // create an optimized horizontal scaling routine |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1224 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1225 //code fragment |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1226 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1227 asm volatile( |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1228 "jmp 9f \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1229 // Begin |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1230 "0: \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1231 "movq (%%esi), %%mm0 \n\t" //FIXME Alignment |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1232 "movq %%mm0, %%mm1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1233 "psrlq $8, %%mm0 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1234 "punpcklbw %%mm7, %%mm1 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1235 "movq %%mm2, %%mm3 \n\t" |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1236 "punpcklbw %%mm7, %%mm0 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1237 "addw %%bx, %%cx \n\t" //2*xalpha += (4*s_xinc)&0xFFFF |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1238 "pshufw $0xFF, %%mm1, %%mm1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1239 "1: \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1240 "adcl %%edx, %%esi \n\t" //xx+= (4*s_xinc)>>16 + carry |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1241 "pshufw $0xFF, %%mm0, %%mm0 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1242 "2: \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1243 "psrlw $9, %%mm3 \n\t" |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1244 "psubw %%mm1, %%mm0 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1245 "pmullw %%mm3, %%mm0 \n\t" |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1246 "paddw %%mm6, %%mm2 \n\t" // 2*alpha += xpos&0xFFFF |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1247 "psllw $7, %%mm1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1248 "paddw %%mm1, %%mm0 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1249 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1250 "movq %%mm0, (%%edi, %%eax) \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1251 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1252 "addl $8, %%eax \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1253 // End |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1254 "9: \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1255 // "int $3\n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1256 "leal 0b, %0 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1257 "leal 1b, %1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1258 "leal 2b, %2 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1259 "decl %1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1260 "decl %2 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1261 "subl %0, %1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1262 "subl %0, %2 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1263 "leal 9b, %3 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1264 "subl %0, %3 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1265 :"=r" (fragment), "=r" (imm8OfPShufW1), "=r" (imm8OfPShufW2), |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1266 "=r" (fragmentLength) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1267 ); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1268 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1269 xpos= 0; //s_xinc/2 - 0x8000; // difference between pixel centers |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1270 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1271 /* choose xinc so that all 8 parts fit exactly |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1272 Note: we cannot use just 1 part because it would not fit in the code cache */ |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1273 // s_xinc2_diff= -((((s_xinc2*(dstw/8))&0xFFFF))/(dstw/8))-10; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1274 // s_xinc_diff= -((((s_xinc*(dstw/8))&0xFFFF))/(dstw/8)); |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1275 #ifdef ALT_ERROR |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1276 // s_xinc2_diff+= ((0x10000/(dstw/8))); |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1277 #endif |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1278 // s_xinc_diff= s_xinc2_diff*2; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1279 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1280 // s_xinc2+= s_xinc2_diff; |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1281 // s_xinc+= s_xinc_diff; |
2270 | 1282 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1283 // old_s_xinc= s_xinc; |
2270 | 1284 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1285 for(i=0; i<dstw/8; i++) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1286 { |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1287 int xx=xpos>>16; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1288 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1289 if((i&3) == 0) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1290 { |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1291 int a=0; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1292 int b=((xpos+s_xinc)>>16) - xx; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1293 int c=((xpos+s_xinc*2)>>16) - xx; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1294 int d=((xpos+s_xinc*3)>>16) - xx; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1295 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1296 memcpy(funnyYCode + fragmentLength*i/4, fragment, fragmentLength); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1297 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1298 funnyYCode[fragmentLength*i/4 + imm8OfPShufW1]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1299 funnyYCode[fragmentLength*i/4 + imm8OfPShufW2]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1300 a | (b<<2) | (c<<4) | (d<<6); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1301 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1302 // if we dont need to read 8 bytes than dont :), reduces the chance of |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1303 // crossing a cache line |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1304 if(d<3) funnyYCode[fragmentLength*i/4 + 1]= 0x6E; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1305 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1306 funnyYCode[fragmentLength*(i+4)/4]= RET; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1307 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1308 xpos+=s_xinc; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1309 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1310 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1311 xpos= 0; //s_xinc2/2 - 0x10000; // difference between centers of chrom samples |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1312 for(i=0; i<dstUVw/8; i++) |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1313 { |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1314 int xx=xpos>>16; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1315 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1316 if((i&3) == 0) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1317 { |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1318 int a=0; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1319 int b=((xpos+s_xinc2)>>16) - xx; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1320 int c=((xpos+s_xinc2*2)>>16) - xx; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1321 int d=((xpos+s_xinc2*3)>>16) - xx; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1322 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1323 memcpy(funnyUVCode + fragmentLength*i/4, fragment, fragmentLength); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1324 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1325 funnyUVCode[fragmentLength*i/4 + imm8OfPShufW1]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1326 funnyUVCode[fragmentLength*i/4 + imm8OfPShufW2]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1327 a | (b<<2) | (c<<4) | (d<<6); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1328 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1329 // if we dont need to read 8 bytes than dont :), reduces the chance of |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1330 // crossing a cache line |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1331 if(d<3) funnyUVCode[fragmentLength*i/4 + 1]= 0x6E; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1332 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1333 funnyUVCode[fragmentLength*(i+4)/4]= RET; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1334 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1335 xpos+=s_xinc2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1336 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1337 // funnyCode[0]= RET; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1338 } |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1339 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1340 #endif // HAVE_MMX2 |
2216 | 1341 } // reset counters |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1342 |
2216 | 1343 while(1){ |
2519 | 1344 unsigned char *dest =dstptr[0]+dststride*s_ypos; |
1345 unsigned char *uDest=dstptr[1]+(dststride>>1)*(s_ypos>>1); | |
1346 unsigned char *vDest=dstptr[2]+(dststride>>1)*(s_ypos>>1); | |
1347 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1348 int y0=(s_srcypos + 0xFFFF)>>16; // first luminance source line number below the dst line |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1349 // points to the dst Pixels center in the source (0 is the center of pixel 0,0 in src) |
2519 | 1350 int srcuvpos= dstbpp==12 ? s_srcypos + s_yinc/2 - 0x8000 : |
1351 s_srcypos - 0x8000; | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1352 int y1=(srcuvpos + 0x1FFFF)>>17; // first chrominance source line number below the dst line |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1353 int yalpha=((s_srcypos-1)&0xFFFF)>>4; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1354 int uvalpha=((srcuvpos-1)&0x1FFFF)>>5; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1355 uint16_t *buf0=pix_buf_y[y0&1]; // top line of the interpolated slice |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1356 uint16_t *buf1=pix_buf_y[((y0+1)&1)]; // bottom line of the interpolated slice |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1357 uint16_t *uvbuf0=pix_buf_uv[y1&1]; // top line of the interpolated slice |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1358 uint16_t *uvbuf1=pix_buf_uv[(y1+1)&1]; // bottom line of the interpolated slice |
2216 | 1359 int i; |
1360 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1361 if(y0>=y+h) break; // FIXME wrong, skips last lines, but they are dupliactes anyway |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1362 |
2519 | 1363 if((y0&1) && dstbpp==12) uvalpha=-1; // there is no alpha if there is no line |
2216 | 1364 |
1365 s_ypos++; s_srcypos+=s_yinc; | |
1366 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1367 //only interpolate the src line horizontally if we didnt do it allready |
2469 | 1368 if(s_last_ypos!=y0) |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1369 { |
2469 | 1370 unsigned char *src; |
1371 // skip if first line has been horiz scaled alleady | |
1372 if(s_last_ypos != y0-1) | |
1373 { | |
1374 // check if first line is before any available src lines | |
1375 if(y0-1 < y) src=srcptr[0]+(0 )*stride[0]; | |
1376 else src=srcptr[0]+(y0-y-1)*stride[0]; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1377 |
2469 | 1378 hyscale(buf0, dstw, src, srcWidth, s_xinc); |
1379 } | |
1380 // check if second line is after any available src lines | |
1381 if(y0-y >= h) src=srcptr[0]+(h-1)*stride[0]; | |
1382 else src=srcptr[0]+(y0-y)*stride[0]; | |
1383 | |
1384 // the min() is required to avoid reuseing lines which where not available | |
1385 s_last_ypos= MIN(y0, y+h-1); | |
1386 hyscale(buf1, dstw, src, srcWidth, s_xinc); | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1387 } |
2469 | 1388 // printf("%d %d %d %d\n", y, y1, s_last_y1pos, h); |
2216 | 1389 // *** horizontal scale U and V lines to temp buffer |
2469 | 1390 if(s_last_y1pos!=y1) |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1391 { |
2469 | 1392 uint8_t *src1, *src2; |
1393 // skip if first line has been horiz scaled alleady | |
1394 if(s_last_y1pos != y1-1) | |
1395 { | |
1396 // check if first line is before any available src lines | |
1397 if(y1-y/2-1 < 0) | |
1398 { | |
1399 src1= srcptr[1]+(0)*stride[1]; | |
1400 src2= srcptr[2]+(0)*stride[2]; | |
1401 }else{ | |
1402 src1= srcptr[1]+(y1-y/2-1)*stride[1]; | |
1403 src2= srcptr[2]+(y1-y/2-1)*stride[2]; | |
1404 } | |
1405 hcscale(uvbuf0, dstUVw, src1, src2, srcWidth, s_xinc2); | |
1406 } | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1407 |
2469 | 1408 // check if second line is after any available src lines |
1409 if(y1 - y/2 >= h/2) | |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1410 { |
2469 | 1411 src1= srcptr[1]+(h/2-1)*stride[1]; |
1412 src2= srcptr[2]+(h/2-1)*stride[2]; | |
1413 }else{ | |
1414 src1= srcptr[1]+(y1-y/2)*stride[1]; | |
1415 src2= srcptr[2]+(y1-y/2)*stride[2]; | |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1416 } |
2469 | 1417 hcscale(uvbuf1, dstUVw, src1, src2, srcWidth, s_xinc2); |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1418 |
2469 | 1419 // the min() is required to avoid reuseing lines which where not available |
1420 s_last_y1pos= MIN(y1, y/2+h/2-1); | |
1421 } | |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1422 |
2519 | 1423 if(dstbpp==12) //YV12 |
1424 yuv2yuv(buf0, buf1, uvbuf0, uvbuf1, dest, uDest, vDest, dstw, yalpha, uvalpha); | |
1425 else if(ABS(s_yinc - 0x10000) < 10) | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1426 yuv2rgb1(buf0, buf1, uvbuf0, uvbuf1, dest, dstw, yalpha, uvalpha, dstbpp); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1427 else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1428 yuv2rgbX(buf0, buf1, uvbuf0, uvbuf1, dest, dstw, yalpha, uvalpha, dstbpp); |
2237 | 1429 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1430 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1431 b16Dither= b16Dither1; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1432 b16Dither1= b16Dither2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1433 b16Dither2= b16Dither; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1434 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1435 g16Dither= g16Dither1; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1436 g16Dither1= g16Dither2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1437 g16Dither2= g16Dither; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1438 #endif |
2216 | 1439 } |
2534
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1440 |
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1441 #ifdef HAVE_MMX |
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1442 __asm __volatile(SFENCE:::"memory"); |
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1443 __asm __volatile(EMMS); |
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1444 #endif |
2216 | 1445 } |
1446 | |
1447 | |
1448 void SwScale_Init(){ | |
1449 // generating tables: | |
1450 int i; | |
1451 for(i=0;i<256;i++){ | |
1452 clip_table[i]=0; | |
1453 clip_table[i+256]=i; | |
1454 clip_table[i+512]=255; | |
1455 yuvtab_2568[i]=(0x2568*(i-16))+(256<<13); | |
1456 yuvtab_3343[i]=0x3343*(i-128); | |
1457 yuvtab_0c92[i]=-0x0c92*(i-128); | |
1458 yuvtab_1a1e[i]=-0x1a1e*(i-128); | |
1459 yuvtab_40cf[i]=0x40cf*(i-128); | |
1460 } | |
1461 | |
1462 } |