Mercurial > mplayer.hg
annotate postproc/swscale.c @ 2575:37da7219ebaf
c optimizations
author | michael |
---|---|
date | Tue, 30 Oct 2001 22:03:25 +0000 |
parents | f2353173d52c |
children | 437ed06579d8 |
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 |
2566 | 36 bicubic scaler |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
37 */ |
2216 | 38 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
39 #define ABS(a) ((a) > 0 ? (a) : (-(a))) |
2469 | 40 #define MIN(a,b) ((a) > (b) ? (b) : (a)) |
41 #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
|
42 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
43 #ifdef HAVE_MMX2 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
44 #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
|
45 #elif defined (HAVE_3DNOW) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
46 #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
|
47 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
48 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
49 #ifdef HAVE_MMX2 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
50 #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
|
51 #else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
52 #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
|
53 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
54 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
55 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
56 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
57 static uint64_t __attribute__((aligned(8))) yCoeff= 0x2568256825682568LL; |
2503 | 58 static uint64_t __attribute__((aligned(8))) vrCoeff= 0x3343334333433343LL; |
59 static uint64_t __attribute__((aligned(8))) ubCoeff= 0x40cf40cf40cf40cfLL; | |
60 static uint64_t __attribute__((aligned(8))) vgCoeff= 0xE5E2E5E2E5E2E5E2LL; | |
61 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
|
62 static uint64_t __attribute__((aligned(8))) w400= 0x0400040004000400LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
63 static uint64_t __attribute__((aligned(8))) w80= 0x0080008000800080LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
64 static uint64_t __attribute__((aligned(8))) w10= 0x0010001000100010LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
65 static uint64_t __attribute__((aligned(8))) bm00001111=0x00000000FFFFFFFFLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
66 static uint64_t __attribute__((aligned(8))) bm00000111=0x0000000000FFFFFFLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
67 static uint64_t __attribute__((aligned(8))) bm11111000=0xFFFFFFFFFF000000LL; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
68 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
69 static uint64_t __attribute__((aligned(8))) b16Dither= 0x0004000400040004LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
70 static uint64_t __attribute__((aligned(8))) b16Dither1=0x0004000400040004LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
71 static uint64_t __attribute__((aligned(8))) b16Dither2=0x0602060206020602LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
72 static uint64_t __attribute__((aligned(8))) g16Dither= 0x0002000200020002LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
73 static uint64_t __attribute__((aligned(8))) g16Dither1=0x0002000200020002LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
74 static uint64_t __attribute__((aligned(8))) g16Dither2=0x0301030103010301LL; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
75 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
76 static uint64_t __attribute__((aligned(8))) b16Mask= 0x001F001F001F001FLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
77 static uint64_t __attribute__((aligned(8))) g16Mask= 0x07E007E007E007E0LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
78 static uint64_t __attribute__((aligned(8))) r16Mask= 0xF800F800F800F800LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
79 static uint64_t __attribute__((aligned(8))) b15Mask= 0x001F001F001F001FLL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
80 static uint64_t __attribute__((aligned(8))) g15Mask= 0x03E003E003E003E0LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
81 static uint64_t __attribute__((aligned(8))) r15Mask= 0x7C007C007C007C00LL; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
82 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
83 static uint64_t __attribute__((aligned(8))) temp0; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
84 static uint64_t __attribute__((aligned(8))) asm_yalpha1; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
85 static uint64_t __attribute__((aligned(8))) asm_uvalpha1; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
86 #endif |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
87 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
88 // temporary storage for 4 yuv lines: |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
89 // 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
|
90 #ifdef HAVE_MMX |
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_y[4][2048]; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
92 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
|
93 #else |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
94 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
|
95 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
|
96 #endif |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
97 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
98 // clipping helper table for C implementations: |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
99 static unsigned char clip_table[768]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
100 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
101 // yuv->rgb conversion tables: |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
102 static int yuvtab_2568[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
103 static int yuvtab_3343[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
104 static int yuvtab_0c92[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
105 static int yuvtab_1a1e[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
106 static int yuvtab_40cf[256]; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
107 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
108 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
109 static uint8_t funnyYCode[10000]; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
110 static uint8_t funnyUVCode[10000]; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
111 |
2469 | 112 static int canMMX2BeUsed=0; |
113 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
114 #define FULL_YSCALEYUV2RGB \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
115 "pxor %%mm7, %%mm7 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
116 "movd %6, %%mm6 \n\t" /*yalpha1*/\ |
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 "punpcklwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
119 "movd %7, %%mm5 \n\t" /*uvalpha1*/\ |
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 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
122 "xorl %%eax, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
123 "1: \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
124 "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
|
125 "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
|
126 "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
|
127 "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
|
128 "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
|
129 "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
|
130 "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
|
131 "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
|
132 "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
|
133 "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
|
134 "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
|
135 "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
|
136 "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
|
137 "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
|
138 "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
|
139 "psubw w80, %%mm1 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
140 "psubw w400, %%mm3 \n\t" /* 8(U-128)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
141 "pmulhw yCoeff, %%mm1 \n\t"\ |
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 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
144 "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
|
145 "movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
146 "pmulhw ubCoeff, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
147 "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
|
148 "pmulhw ugCoeff, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
149 "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
|
150 "psubw w400, %%mm0 \n\t" /* (V-128)8*/\ |
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 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
153 "movq %%mm0, %%mm4 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
154 "pmulhw vrCoeff, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
155 "pmulhw vgCoeff, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
156 "paddw %%mm1, %%mm3 \n\t" /* B*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
157 "paddw %%mm1, %%mm0 \n\t" /* R*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
158 "packuswb %%mm3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
159 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
160 "packuswb %%mm0, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
161 "paddw %%mm4, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
162 "paddw %%mm2, %%mm1 \n\t" /* G*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
163 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
164 "packuswb %%mm1, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
165 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
166 #define YSCALEYUV2RGB \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
167 "movd %6, %%mm6 \n\t" /*yalpha1*/\ |
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 "punpcklwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
170 "movq %%mm6, asm_yalpha1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
171 "movd %7, %%mm5 \n\t" /*uvalpha1*/\ |
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 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
174 "movq %%mm5, asm_uvalpha1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
175 "xorl %%eax, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
176 "1: \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
177 "movq (%2, %%eax), %%mm2 \n\t" /* uvbuf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
178 "movq (%3, %%eax), %%mm3 \n\t" /* uvbuf1[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
179 "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
|
180 "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
|
181 "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
|
182 "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
|
183 "movq asm_uvalpha1, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
184 "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
|
185 "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
|
186 "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
|
187 "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
|
188 "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
|
189 "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
|
190 "psubw w400, %%mm3 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
191 "psubw w400, %%mm4 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
192 "movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
193 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
194 "pmulhw ugCoeff, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
195 "pmulhw vgCoeff, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
196 /* 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
|
197 "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
|
198 "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
|
199 "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
|
200 "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
|
201 "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
|
202 "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
|
203 "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
|
204 "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
|
205 "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
|
206 "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
|
207 "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
|
208 "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
|
209 "pmulhw ubCoeff, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
210 "pmulhw vrCoeff, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
211 "psubw w80, %%mm1 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
212 "psubw w80, %%mm7 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
213 "pmulhw yCoeff, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
214 "pmulhw yCoeff, %%mm7 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
215 /* 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
|
216 "paddw %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
217 "movq %%mm2, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
218 "movq %%mm5, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
219 "movq %%mm4, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
220 "punpcklwd %%mm2, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
221 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
222 "punpcklwd %%mm4, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
223 "paddw %%mm1, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
224 "paddw %%mm1, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
225 "paddw %%mm1, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
226 "punpckhwd %%mm0, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
227 "punpckhwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
228 "punpckhwd %%mm3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
229 "paddw %%mm7, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
230 "paddw %%mm7, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
231 "paddw %%mm7, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
232 /* 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
|
233 "packuswb %%mm0, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
234 "packuswb %%mm6, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
235 "packuswb %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
236 "pxor %%mm7, %%mm7 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
237 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
238 #define YSCALEYUV2RGB1 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
239 "xorl %%eax, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
240 "1: \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
241 "movq (%2, %%eax), %%mm3 \n\t" /* uvbuf0[eax]*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
242 "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
|
243 "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
|
244 "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
|
245 "psubw w400, %%mm3 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
246 "psubw w400, %%mm4 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
247 "movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
248 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
249 "pmulhw ugCoeff, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
250 "pmulhw vgCoeff, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
251 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\ |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
252 "movq (%0, %%eax, 2), %%mm1 \n\t" /*buf0[eax]*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
253 "movq 8(%0, %%eax, 2), %%mm7 \n\t" /*buf0[eax]*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
254 "psraw $4, %%mm1 \n\t" /* buf0[eax] - buf1[eax] >>4*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
255 "psraw $4, %%mm7 \n\t" /* buf0[eax] - buf1[eax] >>4*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
256 "pmulhw ubCoeff, %%mm2 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
257 "pmulhw vrCoeff, %%mm5 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
258 "psubw w80, %%mm1 \n\t" /* 8(Y-16)*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
259 "psubw w80, %%mm7 \n\t" /* 8(Y-16)*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
260 "pmulhw yCoeff, %%mm1 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
261 "pmulhw yCoeff, %%mm7 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
262 /* mm1= Y1, mm2=ub, mm3=ug, mm4=vg mm5=vr, mm7=Y2 */\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
263 "paddw %%mm3, %%mm4 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
264 "movq %%mm2, %%mm0 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
265 "movq %%mm5, %%mm6 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
266 "movq %%mm4, %%mm3 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
267 "punpcklwd %%mm2, %%mm2 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
268 "punpcklwd %%mm5, %%mm5 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
269 "punpcklwd %%mm4, %%mm4 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
270 "paddw %%mm1, %%mm2 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
271 "paddw %%mm1, %%mm5 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
272 "paddw %%mm1, %%mm4 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
273 "punpckhwd %%mm0, %%mm0 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
274 "punpckhwd %%mm6, %%mm6 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
275 "punpckhwd %%mm3, %%mm3 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
276 "paddw %%mm7, %%mm0 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
277 "paddw %%mm7, %%mm6 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
278 "paddw %%mm7, %%mm3 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
279 /* mm0=B1, mm2=B2, mm3=G2, mm4=G1, mm5=R1, mm6=R2 */\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
280 "packuswb %%mm0, %%mm2 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
281 "packuswb %%mm6, %%mm5 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
282 "packuswb %%mm3, %%mm4 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
283 "pxor %%mm7, %%mm7 \n\t" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
284 |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
285 // do vertical chrominance interpolation |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
286 #define YSCALEYUV2RGB1b \ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
287 "xorl %%eax, %%eax \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
288 "1: \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
289 "movq (%2, %%eax), %%mm2 \n\t" /* uvbuf0[eax]*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
290 "movq (%3, %%eax), %%mm3 \n\t" /* uvbuf1[eax]*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
291 "movq 4096(%2, %%eax), %%mm5 \n\t" /* uvbuf0[eax+2048]*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
292 "movq 4096(%3, %%eax), %%mm4 \n\t" /* uvbuf1[eax+2048]*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
293 "paddw %%mm2, %%mm3 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
294 "paddw %%mm5, %%mm4 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
295 "psraw $5, %%mm3 \n\t" /* uvbuf0[eax] - uvbuf1[eax] >>4*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
296 "psraw $5, %%mm4 \n\t" /* uvbuf0[eax+2048] - uvbuf1[eax+2048] >>4*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
297 "psubw w400, %%mm3 \n\t" /* (U-128)8*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
298 "psubw w400, %%mm4 \n\t" /* (V-128)8*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
299 "movq %%mm3, %%mm2 \n\t" /* (U-128)8*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
300 "movq %%mm4, %%mm5 \n\t" /* (V-128)8*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
301 "pmulhw ugCoeff, %%mm3 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
302 "pmulhw vgCoeff, %%mm4 \n\t"\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
303 /* mm2=(U-128)8, mm3=ug, mm4=vg mm5=(V-128)8 */\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
304 "movq (%0, %%eax, 2), %%mm1 \n\t" /*buf0[eax]*/\ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
305 "movq 8(%0, %%eax, 2), %%mm7 \n\t" /*buf0[eax]*/\ |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
306 "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
|
307 "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
|
308 "pmulhw ubCoeff, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
309 "pmulhw vrCoeff, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
310 "psubw w80, %%mm1 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
311 "psubw w80, %%mm7 \n\t" /* 8(Y-16)*/\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
312 "pmulhw yCoeff, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
313 "pmulhw yCoeff, %%mm7 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
314 /* 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
|
315 "paddw %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
316 "movq %%mm2, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
317 "movq %%mm5, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
318 "movq %%mm4, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
319 "punpcklwd %%mm2, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
320 "punpcklwd %%mm5, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
321 "punpcklwd %%mm4, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
322 "paddw %%mm1, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
323 "paddw %%mm1, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
324 "paddw %%mm1, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
325 "punpckhwd %%mm0, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
326 "punpckhwd %%mm6, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
327 "punpckhwd %%mm3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
328 "paddw %%mm7, %%mm0 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
329 "paddw %%mm7, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
330 "paddw %%mm7, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
331 /* 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
|
332 "packuswb %%mm0, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
333 "packuswb %%mm6, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
334 "packuswb %%mm3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
335 "pxor %%mm7, %%mm7 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
336 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
337 #define WRITEBGR32 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
338 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
339 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
340 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
341 "punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
342 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
343 "punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
344 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
345 "movq %%mm2, %%mm0 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
346 "movq %%mm1, %%mm3 \n\t" /* GBGBGBGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
347 "punpcklwd %%mm5, %%mm0 \n\t" /* 0RGB0RGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
348 "punpckhwd %%mm5, %%mm2 \n\t" /* 0RGB0RGB 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
349 "punpcklwd %%mm6, %%mm1 \n\t" /* 0RGB0RGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
350 "punpckhwd %%mm6, %%mm3 \n\t" /* 0RGB0RGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
351 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
352 MOVNTQ(%%mm0, (%4, %%eax, 4))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
353 MOVNTQ(%%mm2, 8(%4, %%eax, 4))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
354 MOVNTQ(%%mm1, 16(%4, %%eax, 4))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
355 MOVNTQ(%%mm3, 24(%4, %%eax, 4))\ |
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 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
358 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
359 " jb 1b \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 #define WRITEBGR16 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
362 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
363 "movq %%mm4, %%mm3 \n\t" /* G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
364 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
365 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
366 "punpcklbw %%mm7, %%mm3 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
367 "punpcklbw %%mm7, %%mm2 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
368 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
369 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
370 "psrlw $3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
371 "psllw $3, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
372 "psllw $8, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
373 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
374 "pand g16Mask, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
375 "pand r16Mask, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
376 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
377 "por %%mm3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
378 "por %%mm5, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
379 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
380 "punpckhbw %%mm7, %%mm4 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
381 "punpckhbw %%mm7, %%mm1 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
382 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R */\ |
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 "psrlw $3, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
385 "psllw $3, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
386 "psllw $8, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
387 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
388 "pand g16Mask, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
389 "pand r16Mask, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
390 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
391 "por %%mm4, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
392 "por %%mm6, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
393 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
394 MOVNTQ(%%mm2, (%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
395 MOVNTQ(%%mm1, 8(%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
396 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
397 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
398 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
399 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
400 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
401 #define WRITEBGR15 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
402 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
403 "movq %%mm4, %%mm3 \n\t" /* G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
404 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
405 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
406 "punpcklbw %%mm7, %%mm3 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
407 "punpcklbw %%mm7, %%mm2 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
408 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
409 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
410 "psrlw $3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
411 "psllw $2, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
412 "psllw $7, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
413 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
414 "pand g15Mask, %%mm3 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
415 "pand r15Mask, %%mm5 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
416 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
417 "por %%mm3, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
418 "por %%mm5, %%mm2 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
419 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
420 "punpckhbw %%mm7, %%mm4 \n\t" /* 0G0G0G0G */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
421 "punpckhbw %%mm7, %%mm1 \n\t" /* 0B0B0B0B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
422 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
423 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
424 "psrlw $3, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
425 "psllw $2, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
426 "psllw $7, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
427 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
428 "pand g15Mask, %%mm4 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
429 "pand r15Mask, %%mm6 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
430 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
431 "por %%mm4, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
432 "por %%mm6, %%mm1 \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
433 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
434 MOVNTQ(%%mm2, (%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
435 MOVNTQ(%%mm1, 8(%4, %%eax, 2))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
436 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
437 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
438 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
439 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
440 // 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
|
441 #define WRITEBGR24 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
442 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
443 "movq %%mm2, %%mm1 \n\t" /* B */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
444 "movq %%mm5, %%mm6 \n\t" /* R */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
445 "punpcklbw %%mm4, %%mm2 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
446 "punpcklbw %%mm7, %%mm5 \n\t" /* 0R0R0R0R 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
447 "punpckhbw %%mm4, %%mm1 \n\t" /* GBGBGBGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
448 "punpckhbw %%mm7, %%mm6 \n\t" /* 0R0R0R0R 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
449 "movq %%mm2, %%mm0 \n\t" /* GBGBGBGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
450 "movq %%mm1, %%mm3 \n\t" /* GBGBGBGB 2 */\ |
2326 | 451 "punpcklwd %%mm5, %%mm0 \n\t" /* 0RGB0RGB 0 */\ |
452 "punpckhwd %%mm5, %%mm2 \n\t" /* 0RGB0RGB 1 */\ | |
453 "punpcklwd %%mm6, %%mm1 \n\t" /* 0RGB0RGB 2 */\ | |
454 "punpckhwd %%mm6, %%mm3 \n\t" /* 0RGB0RGB 3 */\ | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
455 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
456 "movq %%mm0, %%mm4 \n\t" /* 0RGB0RGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
457 "psrlq $8, %%mm0 \n\t" /* 00RGB0RG 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
458 "pand bm00000111, %%mm4 \n\t" /* 00000RGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
459 "pand bm11111000, %%mm0 \n\t" /* 00RGB000 0.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
460 "por %%mm4, %%mm0 \n\t" /* 00RGBRGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
461 "movq %%mm2, %%mm4 \n\t" /* 0RGB0RGB 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
462 "psllq $48, %%mm2 \n\t" /* GB000000 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
463 "por %%mm2, %%mm0 \n\t" /* GBRGBRGB 0 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
464 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
465 "movq %%mm4, %%mm2 \n\t" /* 0RGB0RGB 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
466 "psrld $16, %%mm4 \n\t" /* 000R000R 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
467 "psrlq $24, %%mm2 \n\t" /* 0000RGB0 1.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
468 "por %%mm4, %%mm2 \n\t" /* 000RRGBR 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
469 "pand bm00001111, %%mm2 \n\t" /* 0000RGBR 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
470 "movq %%mm1, %%mm4 \n\t" /* 0RGB0RGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
471 "psrlq $8, %%mm1 \n\t" /* 00RGB0RG 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
472 "pand bm00000111, %%mm4 \n\t" /* 00000RGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
473 "pand bm11111000, %%mm1 \n\t" /* 00RGB000 2.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
474 "por %%mm4, %%mm1 \n\t" /* 00RGBRGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
475 "movq %%mm1, %%mm4 \n\t" /* 00RGBRGB 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
476 "psllq $32, %%mm1 \n\t" /* BRGB0000 2 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
477 "por %%mm1, %%mm2 \n\t" /* BRGBRGBR 1 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
478 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
479 "psrlq $32, %%mm4 \n\t" /* 000000RG 2.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
480 "movq %%mm3, %%mm5 \n\t" /* 0RGB0RGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
481 "psrlq $8, %%mm3 \n\t" /* 00RGB0RG 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
482 "pand bm00000111, %%mm5 \n\t" /* 00000RGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
483 "pand bm11111000, %%mm3 \n\t" /* 00RGB000 3.5 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
484 "por %%mm5, %%mm3 \n\t" /* 00RGBRGB 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
485 "psllq $16, %%mm3 \n\t" /* RGBRGB00 3 */\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
486 "por %%mm4, %%mm3 \n\t" /* RGBRGBRG 2.5 */\ |
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 "leal (%%eax, %%eax, 2), %%ebx \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
489 MOVNTQ(%%mm0, (%4, %%ebx))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
490 MOVNTQ(%%mm2, 8(%4, %%ebx))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
491 MOVNTQ(%%mm3, 16(%4, %%ebx))\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
492 \ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
493 "addl $8, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
494 "cmpl %5, %%eax \n\t"\ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
495 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
496 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
497 |
2519 | 498 static inline void yuv2yuv(uint16_t *buf0, uint16_t *buf1, uint16_t *uvbuf0, uint16_t *uvbuf1, |
499 uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstw, int yalpha, int uvalpha) | |
500 { | |
501 int yalpha1=yalpha^4095; | |
502 int uvalpha1=uvalpha^4095; | |
503 int i; | |
504 | |
2521 | 505 asm volatile ("\n\t"::: "memory"); |
506 | |
2519 | 507 for(i=0;i<dstw;i++) |
508 { | |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
509 ((uint8_t*)dest)[i] = (buf0[i]*yalpha1+buf1[i]*yalpha)>>19; |
2519 | 510 } |
511 | |
512 if(uvalpha != -1) | |
513 { | |
514 for(i=0; i<dstw/2; i++) | |
515 { | |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
516 ((uint8_t*)uDest)[i] = (uvbuf0[i]*uvalpha1+uvbuf1[i]*uvalpha)>>19; |
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
517 ((uint8_t*)vDest)[i] = (uvbuf0[i+2048]*uvalpha1+uvbuf1[i+2048]*uvalpha)>>19; |
2519 | 518 } |
519 } | |
520 } | |
521 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
522 /** |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
523 * vertical scale YV12 to RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
524 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
525 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
|
526 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
|
527 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
528 int yalpha1=yalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
529 int uvalpha1=uvalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
530 int i; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
531 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
532 if(fullUVIpol) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
533 { |
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 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
536 if(dstbpp == 32) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
537 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
538 asm volatile( |
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 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
541 FULL_YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
542 "punpcklbw %%mm1, %%mm3 \n\t" // BGBGBGBG |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
543 "punpcklbw %%mm7, %%mm0 \n\t" // R0R0R0R0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
544 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
545 "movq %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
546 "punpcklwd %%mm0, %%mm3 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
547 "punpckhwd %%mm0, %%mm1 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
548 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
549 MOVNTQ(%%mm3, (%4, %%eax, 4)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
550 MOVNTQ(%%mm1, 8(%4, %%eax, 4)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
551 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
552 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
553 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
554 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
555 |
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), "r" (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" |
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==24) |
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 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
568 // lsb ... msb |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
569 "punpcklbw %%mm1, %%mm3 \n\t" // BGBGBGBG |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
570 "punpcklbw %%mm7, %%mm0 \n\t" // R0R0R0R0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
571 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
572 "movq %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
573 "punpcklwd %%mm0, %%mm3 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
574 "punpckhwd %%mm0, %%mm1 \n\t" // BGR0BGR0 |
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 "movq %%mm3, %%mm2 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
577 "psrlq $8, %%mm3 \n\t" // GR0BGR00 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
578 "pand bm00000111, %%mm2 \n\t" // BGR00000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
579 "pand bm11111000, %%mm3 \n\t" // 000BGR00 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
580 "por %%mm2, %%mm3 \n\t" // BGRBGR00 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
581 "movq %%mm1, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
582 "psllq $48, %%mm1 \n\t" // 000000BG |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
583 "por %%mm1, %%mm3 \n\t" // BGRBGRBG |
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 "movq %%mm2, %%mm1 \n\t" // BGR0BGR0 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
586 "psrld $16, %%mm2 \n\t" // R000R000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
587 "psrlq $24, %%mm1 \n\t" // 0BGR0000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
588 "por %%mm2, %%mm1 \n\t" // RBGRR000 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
589 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
590 "movl %4, %%ebx \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
591 "addl %%eax, %%ebx \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
592 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
593 #ifdef HAVE_MMX2 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
594 //FIXME Alignment |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
595 "movntq %%mm3, (%%ebx, %%eax, 2)\n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
596 "movntq %%mm1, 8(%%ebx, %%eax, 2)\n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
597 #else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
598 "movd %%mm3, (%%ebx, %%eax, 2) \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
599 "psrlq $32, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
600 "movd %%mm3, 4(%%ebx, %%eax, 2) \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
601 "movd %%mm1, 8(%%ebx, %%eax, 2) \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
602 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
603 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
604 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
605 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
606 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
607 :: "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
|
608 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
609 : "%eax", "%ebx" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
610 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
611 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
612 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
613 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
614 asm volatile( |
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 FULL_YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
617 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
618 "paddusb b16Dither, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
619 "paddusb b16Dither, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
620 "paddusb b16Dither, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
621 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
622 "punpcklbw %%mm7, %%mm1 \n\t" // 0G0G0G0G |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
623 "punpcklbw %%mm7, %%mm3 \n\t" // 0B0B0B0B |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
624 "punpcklbw %%mm7, %%mm0 \n\t" // 0R0R0R0R |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
625 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
626 "psrlw $3, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
627 "psllw $2, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
628 "psllw $7, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
629 "pand g15Mask, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
630 "pand r15Mask, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
631 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
632 "por %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
633 "por %%mm1, %%mm0 \n\t" |
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 MOVNTQ(%%mm0, (%4, %%eax, 2)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
636 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
637 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
638 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
639 " jb 1b \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
640 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
641 :: "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
|
642 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
643 : "%eax" |
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 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
649 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
650 FULL_YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
651 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
652 "paddusb g16Dither, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
653 "paddusb b16Dither, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
654 "paddusb b16Dither, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
655 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
656 "punpcklbw %%mm7, %%mm1 \n\t" // 0G0G0G0G |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
657 "punpcklbw %%mm7, %%mm3 \n\t" // 0B0B0B0B |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
658 "punpcklbw %%mm7, %%mm0 \n\t" // 0R0R0R0R |
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 "psrlw $3, %%mm3 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
661 "psllw $3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
662 "psllw $8, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
663 "pand g16Mask, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
664 "pand r16Mask, %%mm0 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
665 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
666 "por %%mm3, %%mm1 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
667 "por %%mm1, %%mm0 \n\t" |
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 MOVNTQ(%%mm0, (%4, %%eax, 2)) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
670 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
671 "addl $4, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
672 "cmpl %5, %%eax \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
673 " jb 1b \n\t" |
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 :: "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
|
676 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
677 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
678 ); |
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 #else |
2521 | 681 asm volatile ("\n\t"::: "memory"); |
682 | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
683 if(dstbpp==32 || dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
684 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
685 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
686 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
687 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
|
688 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
|
689 int V=((uvbuf0[i+2048]*uvalpha1+uvbuf1[i+2048]*uvalpha)>>19); |
2503 | 690 dest[0]=clip_table[((Y + yuvtab_40cf[U]) >>13)]; |
691 dest[1]=clip_table[((Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13)]; | |
692 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
|
693 dest+=dstbpp>>3; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
694 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
695 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
696 else if(dstbpp==16) |
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 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
699 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
700 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
|
701 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
|
702 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
|
703 |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
704 ((uint16_t*)dest)[i] = |
2503 | 705 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
706 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<3)&0x07E0) | | |
707 ((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
|
708 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
709 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
710 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
711 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
712 for(i=0;i<dstw;i++){ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
713 // vertical linear interpolation && yuv2rgb in a single step: |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
714 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
|
715 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
|
716 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
|
717 |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
718 ((uint16_t*)dest)[i] = |
2503 | 719 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
720 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<2)&0x03E0) | | |
721 ((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
|
722 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
723 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
724 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
725 }//FULL_UV_IPOL |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
726 else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
727 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
728 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
729 if(dstbpp == 32) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
730 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
731 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
732 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
733 WRITEBGR32 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
734 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
735 :: "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
|
736 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
737 : "%eax" |
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 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
740 else if(dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
741 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
742 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
743 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
744 WRITEBGR24 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
745 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
746 :: "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
|
747 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
748 : "%eax", "%ebx" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
749 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
750 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
751 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
752 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
753 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
754 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
755 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
756 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
757 "paddusb b16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
758 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
759 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
760 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
761 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
762 WRITEBGR15 |
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 :: "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
|
765 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
766 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
767 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
768 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
769 else if(dstbpp==16) |
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 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
772 YSCALEYUV2RGB |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
773 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
774 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
775 "paddusb g16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
776 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
777 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
778 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
779 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
780 WRITEBGR16 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
781 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
782 :: "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
|
783 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
784 : "%eax" |
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 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
787 #else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
788 //FIXME unroll C loop and dont recalculate UV |
2521 | 789 asm volatile ("\n\t"::: "memory"); |
790 | |
2575 | 791 if(dstbpp==32) |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
792 { |
2575 | 793 for(i=0; i<dstw-1; i+=2){ |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
794 // vertical linear interpolation && yuv2rgb in a single step: |
2575 | 795 int Y1=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
796 int Y2=yuvtab_2568[((buf0[i+1]*yalpha1+buf1[i+1]*yalpha)>>19)]; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
797 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
|
798 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
2575 | 799 |
800 int Cb= yuvtab_40cf[U]; | |
801 int Cg= yuvtab_1a1e[V] + yuvtab_0c92[U]; | |
802 int Cr= yuvtab_3343[V]; | |
803 | |
804 dest[4*i+0]=clip_table[((Y1 + Cb) >>13)]; | |
805 dest[4*i+1]=clip_table[((Y1 + Cg) >>13)]; | |
806 dest[4*i+2]=clip_table[((Y1 + Cr) >>13)]; | |
807 | |
808 dest[4*i+4]=clip_table[((Y2 + Cb) >>13)]; | |
809 dest[4*i+5]=clip_table[((Y2 + Cg) >>13)]; | |
810 dest[4*i+6]=clip_table[((Y2 + Cr) >>13)]; | |
811 } | |
812 } | |
813 if(dstbpp==24) | |
814 { | |
815 for(i=0; i<dstw-1; i+=2){ | |
816 // vertical linear interpolation && yuv2rgb in a single step: | |
817 int Y1=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; | |
818 int Y2=yuvtab_2568[((buf0[i+1]*yalpha1+buf1[i+1]*yalpha)>>19)]; | |
819 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); | |
820 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); | |
821 | |
822 int Cb= yuvtab_40cf[U]; | |
823 int Cg= yuvtab_1a1e[V] + yuvtab_0c92[U]; | |
824 int Cr= yuvtab_3343[V]; | |
825 | |
826 dest[0]=clip_table[((Y1 + Cb) >>13)]; | |
827 dest[1]=clip_table[((Y1 + Cg) >>13)]; | |
828 dest[2]=clip_table[((Y1 + Cr) >>13)]; | |
829 | |
830 dest[3]=clip_table[((Y2 + Cb) >>13)]; | |
831 dest[4]=clip_table[((Y2 + Cg) >>13)]; | |
832 dest[5]=clip_table[((Y2 + Cr) >>13)]; | |
833 dest+=6; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
834 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
835 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
836 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
837 { |
2575 | 838 for(i=0; i<dstw-1; i+=2){ |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
839 // vertical linear interpolation && yuv2rgb in a single step: |
2575 | 840 int Y1=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
841 int Y2=yuvtab_2568[((buf0[i+1]*yalpha1+buf1[i+1]*yalpha)>>19)]; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
842 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
|
843 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
|
844 |
2575 | 845 int Cb= yuvtab_40cf[U]; |
846 int Cg= yuvtab_1a1e[V] + yuvtab_0c92[U]; | |
847 int Cr= yuvtab_3343[V]; | |
848 | |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
849 ((uint16_t*)dest)[i] = |
2575 | 850 (clip_table[(Y1 + Cb) >>13]>>3) | |
851 ((clip_table[(Y1 + Cg) >>13]<<3)&0x07E0) | | |
852 ((clip_table[(Y1 + Cr) >>13]<<8)&0xF800); | |
853 | |
854 ((uint16_t*)dest)[i+1] = | |
855 (clip_table[(Y2 + Cb) >>13]>>3) | | |
856 ((clip_table[(Y2 + Cg) >>13]<<3)&0x07E0) | | |
857 ((clip_table[(Y2 + Cr) >>13]<<8)&0xF800); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
858 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
859 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
860 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
861 { |
2575 | 862 for(i=0; i<dstw-1; i+=2){ |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
863 // vertical linear interpolation && yuv2rgb in a single step: |
2575 | 864 int Y1=yuvtab_2568[((buf0[i]*yalpha1+buf1[i]*yalpha)>>19)]; |
865 int Y2=yuvtab_2568[((buf0[i+1]*yalpha1+buf1[i+1]*yalpha)>>19)]; | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
866 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
|
867 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
|
868 |
2575 | 869 int Cb= yuvtab_40cf[U]; |
870 int Cg= yuvtab_1a1e[V] + yuvtab_0c92[U]; | |
871 int Cr= yuvtab_3343[V]; | |
872 | |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
873 ((uint16_t*)dest)[i] = |
2575 | 874 (clip_table[(Y1 + Cb) >>13]>>3) | |
875 ((clip_table[(Y1 + Cg) >>13]<<2)&0x03E0) | | |
876 ((clip_table[(Y1 + Cr) >>13]<<7)&0x7C00); | |
877 ((uint16_t*)dest)[i+1] = | |
878 (clip_table[(Y2 + Cb) >>13]>>3) | | |
879 ((clip_table[(Y2 + Cg) >>13]<<2)&0x03E0) | | |
880 ((clip_table[(Y2 + Cr) >>13]<<7)&0x7C00); | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
881 } |
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 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
884 } //!FULL_UV_IPOL |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
885 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
886 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
887 /** |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
888 * YV12 to RGB without scaling or interpolating |
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 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
|
891 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
|
892 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
893 int yalpha1=yalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
894 int uvalpha1=uvalpha^4095; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
895 int i; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
896 if(fullUVIpol || allwaysIpol) |
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 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
|
899 return; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
900 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
901 #ifdef HAVE_MMX |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
902 if( yalpha > 2048 ) buf0 = buf1; |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
903 if( uvalpha < 2048 ) // note this is not correct (shifts chrominance by 0.5 pixels) but its a bit faster |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
904 { |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
905 if(dstbpp == 32) |
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 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
908 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
909 WRITEBGR32 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
910 :: "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
|
911 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
912 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
913 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
914 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
915 else if(dstbpp==24) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
916 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
917 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
918 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
919 WRITEBGR24 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
920 :: "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
|
921 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
922 : "%eax", "%ebx" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
923 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
924 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
925 else if(dstbpp==15) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
926 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
927 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
928 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
929 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
930 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
931 "paddusb b16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
932 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
933 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
934 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
935 WRITEBGR15 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
936 :: "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
|
937 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
938 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
939 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
940 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
941 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
942 { |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
943 asm volatile( |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
944 YSCALEYUV2RGB1 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
945 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
946 #ifdef DITHER1XBPP |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
947 "paddusb g16Dither, %%mm2 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
948 "paddusb b16Dither, %%mm4 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
949 "paddusb b16Dither, %%mm5 \n\t" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
950 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
951 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
952 WRITEBGR16 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
953 :: "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
|
954 "m" (yalpha1), "m" (uvalpha1) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
955 : "%eax" |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
956 ); |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
957 } |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
958 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
959 else |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
960 { |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
961 if(dstbpp == 32) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
962 { |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
963 asm volatile( |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
964 YSCALEYUV2RGB1b |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
965 WRITEBGR32 |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
966 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
967 "m" (yalpha1), "m" (uvalpha1) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
968 : "%eax" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
969 ); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
970 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
971 else if(dstbpp==24) |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
972 { |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
973 asm volatile( |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
974 YSCALEYUV2RGB1b |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
975 WRITEBGR24 |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
976 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
977 "m" (yalpha1), "m" (uvalpha1) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
978 : "%eax", "%ebx" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
979 ); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
980 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
981 else if(dstbpp==15) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
982 { |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
983 asm volatile( |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
984 YSCALEYUV2RGB1b |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
985 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
986 #ifdef DITHER1XBPP |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
987 "paddusb b16Dither, %%mm2 \n\t" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
988 "paddusb b16Dither, %%mm4 \n\t" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
989 "paddusb b16Dither, %%mm5 \n\t" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
990 #endif |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
991 WRITEBGR15 |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
992 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
993 "m" (yalpha1), "m" (uvalpha1) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
994 : "%eax" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
995 ); |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
996 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
997 else if(dstbpp==16) |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
998 { |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
999 asm volatile( |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1000 YSCALEYUV2RGB1b |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1001 /* mm2=B, %%mm4=G, %%mm5=R, %%mm7=0 */ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1002 #ifdef DITHER1XBPP |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1003 "paddusb g16Dither, %%mm2 \n\t" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1004 "paddusb b16Dither, %%mm4 \n\t" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1005 "paddusb b16Dither, %%mm5 \n\t" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1006 #endif |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1007 |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1008 WRITEBGR16 |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1009 :: "r" (buf0), "r" (buf1), "r" (uvbuf0), "r" (uvbuf1), "r" (dest), "m" (dstw), |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1010 "m" (yalpha1), "m" (uvalpha1) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1011 : "%eax" |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1012 ); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1013 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1014 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1015 #else |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1016 //FIXME unroll C loop and dont recalculate UV |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1017 asm volatile ("\n\t"::: "memory"); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1018 |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1019 if(dstbpp==32 || dstbpp==24) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1020 { |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1021 for(i=0;i<dstw;i++){ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1022 // vertical linear interpolation && yuv2rgb in a single step: |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1023 int Y=yuvtab_2568[buf0[i]>>7]; |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1024 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1025 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1026 dest[0]=clip_table[((Y + yuvtab_40cf[U]) >>13)]; |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1027 dest[1]=clip_table[((Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13)]; |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1028 dest[2]=clip_table[((Y + yuvtab_3343[V]) >>13)]; |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1029 dest+=dstbpp>>3; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1030 } |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1031 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1032 else if(dstbpp==16) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1033 { |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1034 for(i=0;i<dstw;i++){ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1035 // vertical linear interpolation && yuv2rgb in a single step: |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1036 int Y=yuvtab_2568[buf0[i]>>7]; |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1037 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1038 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1039 |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
1040 ((uint16_t*)dest)[i] = |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1041 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1042 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<3)&0x07E0) | |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1043 ((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
|
1044 } |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1045 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1046 else if(dstbpp==15) |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1047 { |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1048 for(i=0;i<dstw;i++){ |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1049 // vertical linear interpolation && yuv2rgb in a single step: |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1050 int Y=yuvtab_2568[buf0[i]>>7]; |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1051 int U=((uvbuf0[i/2]*uvalpha1+uvbuf1[i/2]*uvalpha)>>19); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1052 int V=((uvbuf0[i/2+2048]*uvalpha1+uvbuf1[i/2+2048]*uvalpha)>>19); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1053 |
2572
f2353173d52c
c optimizations (array is faster than pointer) (16bpp variants tested and 2% faster)
michael
parents:
2569
diff
changeset
|
1054 ((uint16_t*)dest)[i] = |
2569
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1055 (clip_table[(Y + yuvtab_40cf[U]) >>13]>>3) | |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1056 ((clip_table[(Y + yuvtab_1a1e[V] + yuvtab_0c92[U]) >>13]<<2)&0x03E0) | |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1057 ((clip_table[(Y + yuvtab_3343[V]) >>13]<<7)&0x7C00); |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1058 } |
30b736e7feef
interpolate chrominance for every second line in the 1:1 vertical scale function
michael
parents:
2566
diff
changeset
|
1059 } |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1060 #endif |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1061 } |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1062 |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1063 |
2469 | 1064 static inline void hyscale(uint16_t *dst, int dstWidth, uint8_t *src, int srcWidth, int xInc) |
1065 { | |
1066 int i; | |
1067 unsigned int xpos=0; | |
1068 // *** horizontal scale Y line to temp buffer | |
1069 #ifdef ARCH_X86 | |
1070 #ifdef HAVE_MMX2 | |
1071 if(canMMX2BeUsed) | |
1072 { | |
1073 asm volatile( | |
1074 "pxor %%mm7, %%mm7 \n\t" | |
1075 "pxor %%mm2, %%mm2 \n\t" // 2*xalpha | |
1076 "movd %5, %%mm6 \n\t" // xInc&0xFFFF | |
1077 "punpcklwd %%mm6, %%mm6 \n\t" | |
1078 "punpcklwd %%mm6, %%mm6 \n\t" | |
1079 "movq %%mm6, %%mm2 \n\t" | |
1080 "psllq $16, %%mm2 \n\t" | |
1081 "paddw %%mm6, %%mm2 \n\t" | |
1082 "psllq $16, %%mm2 \n\t" | |
1083 "paddw %%mm6, %%mm2 \n\t" | |
1084 "psllq $16, %%mm2 \n\t" //0,t,2t,3t t=xInc&0xFF | |
1085 "movq %%mm2, temp0 \n\t" | |
1086 "movd %4, %%mm6 \n\t" //(xInc*4)&0xFFFF | |
1087 "punpcklwd %%mm6, %%mm6 \n\t" | |
1088 "punpcklwd %%mm6, %%mm6 \n\t" | |
1089 "xorl %%eax, %%eax \n\t" // i | |
1090 "movl %0, %%esi \n\t" // src | |
1091 "movl %1, %%edi \n\t" // buf1 | |
1092 "movl %3, %%edx \n\t" // (xInc*4)>>16 | |
1093 "xorl %%ecx, %%ecx \n\t" | |
1094 "xorl %%ebx, %%ebx \n\t" | |
1095 "movw %4, %%bx \n\t" // (xInc*4)&0xFFFF | |
2520 | 1096 |
2469 | 1097 #define FUNNY_Y_CODE \ |
2520 | 1098 PREFETCH" 1024(%%esi) \n\t"\ |
1099 PREFETCH" 1056(%%esi) \n\t"\ | |
1100 PREFETCH" 1088(%%esi) \n\t"\ | |
2469 | 1101 "call funnyYCode \n\t"\ |
1102 "movq temp0, %%mm2 \n\t"\ | |
1103 "xorl %%ecx, %%ecx \n\t" | |
2520 | 1104 |
2469 | 1105 FUNNY_Y_CODE |
1106 FUNNY_Y_CODE | |
1107 FUNNY_Y_CODE | |
1108 FUNNY_Y_CODE | |
1109 FUNNY_Y_CODE | |
1110 FUNNY_Y_CODE | |
1111 FUNNY_Y_CODE | |
1112 FUNNY_Y_CODE | |
1113 | |
1114 :: "m" (src), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16), | |
1115 "m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF) | |
1116 : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" | |
1117 ); | |
1118 for(i=dstWidth-1; (i*xInc)>>16 >=srcWidth-1; i--) dst[i] = src[srcWidth-1]*128; | |
1119 } | |
1120 else | |
1121 { | |
1122 #endif | |
1123 //NO MMX just normal asm ... | |
1124 asm volatile( | |
1125 "xorl %%eax, %%eax \n\t" // i | |
1126 "xorl %%ebx, %%ebx \n\t" // xx | |
1127 "xorl %%ecx, %%ecx \n\t" // 2*xalpha | |
1128 "1: \n\t" | |
1129 "movzbl (%0, %%ebx), %%edi \n\t" //src[xx] | |
1130 "movzbl 1(%0, %%ebx), %%esi \n\t" //src[xx+1] | |
1131 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
1132 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
1133 "shll $16, %%edi \n\t" | |
1134 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
1135 "movl %1, %%edi \n\t" | |
1136 "shrl $9, %%esi \n\t" | |
1137 "movw %%si, (%%edi, %%eax, 2) \n\t" | |
1138 "addw %4, %%cx \n\t" //2*xalpha += xInc&0xFF | |
1139 "adcl %3, %%ebx \n\t" //xx+= xInc>>8 + carry | |
1140 | |
1141 "movzbl (%0, %%ebx), %%edi \n\t" //src[xx] | |
1142 "movzbl 1(%0, %%ebx), %%esi \n\t" //src[xx+1] | |
1143 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
1144 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
1145 "shll $16, %%edi \n\t" | |
1146 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
1147 "movl %1, %%edi \n\t" | |
1148 "shrl $9, %%esi \n\t" | |
1149 "movw %%si, 2(%%edi, %%eax, 2) \n\t" | |
1150 "addw %4, %%cx \n\t" //2*xalpha += xInc&0xFF | |
1151 "adcl %3, %%ebx \n\t" //xx+= xInc>>8 + carry | |
1152 | |
1153 | |
1154 "addl $2, %%eax \n\t" | |
1155 "cmpl %2, %%eax \n\t" | |
1156 " jb 1b \n\t" | |
1157 | |
1158 | |
1159 :: "r" (src), "m" (dst), "m" (dstWidth), "m" (xInc>>16), "m" (xInc&0xFFFF) | |
1160 : "%eax", "%ebx", "%ecx", "%edi", "%esi" | |
1161 ); | |
1162 #ifdef HAVE_MMX2 | |
1163 } //if MMX2 cant be used | |
1164 #endif | |
1165 #else | |
1166 for(i=0;i<dstWidth;i++){ | |
1167 register unsigned int xx=xpos>>16; | |
1168 register unsigned int xalpha=(xpos&0xFFFF)>>9; | |
2566 | 1169 dst[i]= (src[xx]<<7) + (src[xx+1] - src[xx])*xalpha; |
2469 | 1170 xpos+=xInc; |
1171 } | |
1172 #endif | |
1173 } | |
1174 | |
1175 inline static void hcscale(uint16_t *dst, int dstWidth, | |
1176 uint8_t *src1, uint8_t *src2, int srcWidth, int xInc) | |
1177 { | |
1178 int xpos=0; | |
1179 int i; | |
1180 #ifdef ARCH_X86 | |
1181 #ifdef HAVE_MMX2 | |
1182 if(canMMX2BeUsed) | |
1183 { | |
1184 asm volatile( | |
1185 "pxor %%mm7, %%mm7 \n\t" | |
1186 "pxor %%mm2, %%mm2 \n\t" // 2*xalpha | |
1187 "movd %5, %%mm6 \n\t" // xInc&0xFFFF | |
1188 "punpcklwd %%mm6, %%mm6 \n\t" | |
1189 "punpcklwd %%mm6, %%mm6 \n\t" | |
1190 "movq %%mm6, %%mm2 \n\t" | |
1191 "psllq $16, %%mm2 \n\t" | |
1192 "paddw %%mm6, %%mm2 \n\t" | |
1193 "psllq $16, %%mm2 \n\t" | |
1194 "paddw %%mm6, %%mm2 \n\t" | |
1195 "psllq $16, %%mm2 \n\t" //0,t,2t,3t t=xInc&0xFFFF | |
1196 "movq %%mm2, temp0 \n\t" | |
1197 "movd %4, %%mm6 \n\t" //(xInc*4)&0xFFFF | |
1198 "punpcklwd %%mm6, %%mm6 \n\t" | |
1199 "punpcklwd %%mm6, %%mm6 \n\t" | |
1200 "xorl %%eax, %%eax \n\t" // i | |
1201 "movl %0, %%esi \n\t" // src | |
1202 "movl %1, %%edi \n\t" // buf1 | |
1203 "movl %3, %%edx \n\t" // (xInc*4)>>16 | |
1204 "xorl %%ecx, %%ecx \n\t" | |
1205 "xorl %%ebx, %%ebx \n\t" | |
1206 "movw %4, %%bx \n\t" // (xInc*4)&0xFFFF | |
1207 | |
1208 #define FUNNYUVCODE \ | |
2520 | 1209 PREFETCH" 1024(%%esi) \n\t"\ |
1210 PREFETCH" 1056(%%esi) \n\t"\ | |
1211 PREFETCH" 1088(%%esi) \n\t"\ | |
2469 | 1212 "call funnyUVCode \n\t"\ |
1213 "movq temp0, %%mm2 \n\t"\ | |
1214 "xorl %%ecx, %%ecx \n\t" | |
1215 | |
1216 FUNNYUVCODE | |
1217 FUNNYUVCODE | |
1218 FUNNYUVCODE | |
1219 FUNNYUVCODE | |
1220 | |
1221 FUNNYUVCODE | |
1222 FUNNYUVCODE | |
1223 FUNNYUVCODE | |
1224 FUNNYUVCODE | |
1225 "xorl %%eax, %%eax \n\t" // i | |
1226 "movl %6, %%esi \n\t" // src | |
1227 "movl %1, %%edi \n\t" // buf1 | |
1228 "addl $4096, %%edi \n\t" | |
1229 | |
1230 FUNNYUVCODE | |
1231 FUNNYUVCODE | |
1232 FUNNYUVCODE | |
1233 FUNNYUVCODE | |
1234 | |
1235 FUNNYUVCODE | |
1236 FUNNYUVCODE | |
1237 FUNNYUVCODE | |
1238 FUNNYUVCODE | |
1239 | |
1240 :: "m" (src1), "m" (dst), "m" (dstWidth), "m" ((xInc*4)>>16), | |
1241 "m" ((xInc*4)&0xFFFF), "m" (xInc&0xFFFF), "m" (src2) | |
1242 : "%eax", "%ebx", "%ecx", "%edx", "%esi", "%edi" | |
1243 ); | |
1244 for(i=dstWidth-1; (i*xInc)>>16 >=srcWidth/2-1; i--) | |
1245 { | |
1246 dst[i] = src1[srcWidth/2-1]*128; | |
1247 dst[i+2048] = src2[srcWidth/2-1]*128; | |
1248 } | |
1249 } | |
1250 else | |
1251 { | |
1252 #endif | |
1253 asm volatile( | |
1254 "xorl %%eax, %%eax \n\t" // i | |
1255 "xorl %%ebx, %%ebx \n\t" // xx | |
1256 "xorl %%ecx, %%ecx \n\t" // 2*xalpha | |
1257 "1: \n\t" | |
1258 "movl %0, %%esi \n\t" | |
1259 "movzbl (%%esi, %%ebx), %%edi \n\t" //src[xx] | |
1260 "movzbl 1(%%esi, %%ebx), %%esi \n\t" //src[xx+1] | |
1261 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
1262 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
1263 "shll $16, %%edi \n\t" | |
1264 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
1265 "movl %1, %%edi \n\t" | |
1266 "shrl $9, %%esi \n\t" | |
1267 "movw %%si, (%%edi, %%eax, 2) \n\t" | |
1268 | |
1269 "movzbl (%5, %%ebx), %%edi \n\t" //src[xx] | |
1270 "movzbl 1(%5, %%ebx), %%esi \n\t" //src[xx+1] | |
1271 "subl %%edi, %%esi \n\t" //src[xx+1] - src[xx] | |
1272 "imull %%ecx, %%esi \n\t" //(src[xx+1] - src[xx])*2*xalpha | |
1273 "shll $16, %%edi \n\t" | |
1274 "addl %%edi, %%esi \n\t" //src[xx+1]*2*xalpha + src[xx]*(1-2*xalpha) | |
1275 "movl %1, %%edi \n\t" | |
1276 "shrl $9, %%esi \n\t" | |
1277 "movw %%si, 4096(%%edi, %%eax, 2)\n\t" | |
1278 | |
1279 "addw %4, %%cx \n\t" //2*xalpha += xInc&0xFF | |
1280 "adcl %3, %%ebx \n\t" //xx+= xInc>>8 + carry | |
1281 "addl $1, %%eax \n\t" | |
1282 "cmpl %2, %%eax \n\t" | |
1283 " jb 1b \n\t" | |
1284 | |
1285 :: "m" (src1), "m" (dst), "m" (dstWidth), "m" (xInc>>16), "m" (xInc&0xFFFF), | |
1286 "r" (src2) | |
1287 : "%eax", "%ebx", "%ecx", "%edi", "%esi" | |
1288 ); | |
1289 #ifdef HAVE_MMX2 | |
1290 } //if MMX2 cant be used | |
1291 #endif | |
1292 #else | |
1293 for(i=0;i<dstWidth;i++){ | |
1294 register unsigned int xx=xpos>>16; | |
1295 register unsigned int xalpha=(xpos&0xFFFF)>>9; | |
1296 dst[i]=(src1[xx]*(xalpha^127)+src1[xx+1]*xalpha); | |
1297 dst[i+2048]=(src2[xx]*(xalpha^127)+src2[xx+1]*xalpha); | |
2566 | 1298 /* slower |
1299 dst[i]= (src1[xx]<<7) + (src1[xx+1] - src1[xx])*xalpha; | |
1300 dst[i+2048]=(src2[xx]<<7) + (src2[xx+1] - src2[xx])*xalpha; | |
1301 */ | |
2469 | 1302 xpos+=xInc; |
1303 } | |
1304 #endif | |
1305 } | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1306 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1307 |
2519 | 1308 // *** bilinear scaling and yuv->rgb or yuv->yuv conversion of yv12 slices: |
2216 | 1309 // *** Note: it's called multiple times while decoding a frame, first time y==0 |
1310 // *** Designed to upscale, but may work for downscale too. | |
2274 | 1311 // s_xinc = (src_width << 16) / dst_width |
2216 | 1312 // s_yinc = (src_height << 16) / dst_height |
2519 | 1313 void SwScale_YV12slice(unsigned char* srcptr[],int stride[], int y, int h, |
1314 uint8_t* dstptr[], int dststride, int dstw, int dstbpp, | |
2216 | 1315 unsigned int s_xinc,unsigned int s_yinc){ |
1316 | |
1317 // scaling factors: | |
1318 //static int s_yinc=(vo_dga_src_height<<16)/vo_dga_vp_height; | |
1319 //static int s_xinc=(vo_dga_src_width<<8)/vo_dga_vp_width; | |
1320 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1321 unsigned int s_xinc2; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1322 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1323 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
|
1324 static int s_ypos; |
2216 | 1325 |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1326 // last horzontally interpolated lines, used to avoid unnecessary calculations |
2216 | 1327 static int s_last_ypos; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1328 static int s_last_y1pos; |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1329 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1330 static int static_dstw; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1331 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1332 #ifdef HAVE_MMX2 |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1333 // used to detect a horizontal size change |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1334 static int old_dstw= -1; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1335 static int old_s_xinc= -1; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1336 #endif |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1337 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1338 int srcWidth= (dstw*s_xinc + 0x8000)>>16; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1339 int dstUVw= fullUVIpol ? dstw : dstw/2; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1340 |
2216 | 1341 |
2270 | 1342 #ifdef HAVE_MMX2 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1343 canMMX2BeUsed= (s_xinc <= 0x10000 && (dstw&31)==0 && (srcWidth&15)==0) ? 1 : 0; |
2270 | 1344 #endif |
1345 | |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1346 // 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
|
1347 // n-2 is the last chrominance sample available |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1348 // 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
|
1349 // 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
|
1350 // first and last pixel |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1351 if(canMMX2BeUsed) s_xinc+= 20; |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1352 else s_xinc = ((srcWidth-2)<<16)/(dstw-2) - 20; |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1353 |
2566 | 1354 if(fullUVIpol && !(dstbpp==12)) s_xinc2= s_xinc>>1; |
1355 else s_xinc2= s_xinc; | |
2271 | 1356 // force calculation of the horizontal interpolation of the first line |
1357 | |
2216 | 1358 if(y==0){ |
2469 | 1359 s_last_ypos=-99; |
1360 s_last_y1pos=-99; | |
1361 s_srcypos= s_yinc/2 - 0x8000; | |
1362 s_ypos=0; | |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1363 #ifdef HAVE_MMX2 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1364 // cant downscale !!! |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1365 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
|
1366 { |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1367 uint8_t *fragment; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1368 int imm8OfPShufW1; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1369 int imm8OfPShufW2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1370 int fragmentLength; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1371 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1372 int xpos, xx, xalpha, i; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1373 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1374 old_s_xinc= s_xinc; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1375 old_dstw= dstw; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1376 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1377 static_dstw= dstw; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1378 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1379 // create an optimized horizontal scaling routine |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1380 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1381 //code fragment |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1382 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1383 asm volatile( |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1384 "jmp 9f \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1385 // Begin |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1386 "0: \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1387 "movq (%%esi), %%mm0 \n\t" //FIXME Alignment |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1388 "movq %%mm0, %%mm1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1389 "psrlq $8, %%mm0 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1390 "punpcklbw %%mm7, %%mm1 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1391 "movq %%mm2, %%mm3 \n\t" |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1392 "punpcklbw %%mm7, %%mm0 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1393 "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
|
1394 "pshufw $0xFF, %%mm1, %%mm1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1395 "1: \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1396 "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
|
1397 "pshufw $0xFF, %%mm0, %%mm0 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1398 "2: \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1399 "psrlw $9, %%mm3 \n\t" |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1400 "psubw %%mm1, %%mm0 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1401 "pmullw %%mm3, %%mm0 \n\t" |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1402 "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
|
1403 "psllw $7, %%mm1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1404 "paddw %%mm1, %%mm0 \n\t" |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1405 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1406 "movq %%mm0, (%%edi, %%eax) \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1407 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1408 "addl $8, %%eax \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1409 // End |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1410 "9: \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1411 // "int $3\n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1412 "leal 0b, %0 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1413 "leal 1b, %1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1414 "leal 2b, %2 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1415 "decl %1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1416 "decl %2 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1417 "subl %0, %1 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1418 "subl %0, %2 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1419 "leal 9b, %3 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1420 "subl %0, %3 \n\t" |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1421 :"=r" (fragment), "=r" (imm8OfPShufW1), "=r" (imm8OfPShufW2), |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1422 "=r" (fragmentLength) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1423 ); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1424 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1425 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
|
1426 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1427 /* 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
|
1428 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
|
1429 // 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
|
1430 // 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
|
1431 #ifdef ALT_ERROR |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1432 // 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
|
1433 #endif |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1434 // 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
|
1435 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1436 // s_xinc2+= s_xinc2_diff; |
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1437 // s_xinc+= s_xinc_diff; |
2270 | 1438 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1439 // old_s_xinc= s_xinc; |
2270 | 1440 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1441 for(i=0; i<dstw/8; i++) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1442 { |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1443 int xx=xpos>>16; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1444 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1445 if((i&3) == 0) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1446 { |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1447 int a=0; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1448 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
|
1449 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
|
1450 int d=((xpos+s_xinc*3)>>16) - xx; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1451 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1452 memcpy(funnyYCode + fragmentLength*i/4, fragment, fragmentLength); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1453 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1454 funnyYCode[fragmentLength*i/4 + imm8OfPShufW1]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1455 funnyYCode[fragmentLength*i/4 + imm8OfPShufW2]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1456 a | (b<<2) | (c<<4) | (d<<6); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1457 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1458 // 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
|
1459 // crossing a cache line |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1460 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
|
1461 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1462 funnyYCode[fragmentLength*(i+4)/4]= RET; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1463 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1464 xpos+=s_xinc; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1465 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1466 |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1467 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
|
1468 for(i=0; i<dstUVw/8; i++) |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1469 { |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1470 int xx=xpos>>16; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1471 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1472 if((i&3) == 0) |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1473 { |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1474 int a=0; |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1475 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
|
1476 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
|
1477 int d=((xpos+s_xinc2*3)>>16) - xx; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1478 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1479 memcpy(funnyUVCode + fragmentLength*i/4, fragment, fragmentLength); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1480 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1481 funnyUVCode[fragmentLength*i/4 + imm8OfPShufW1]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1482 funnyUVCode[fragmentLength*i/4 + imm8OfPShufW2]= |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1483 a | (b<<2) | (c<<4) | (d<<6); |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1484 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1485 // 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
|
1486 // crossing a cache line |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1487 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
|
1488 |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1489 funnyUVCode[fragmentLength*(i+4)/4]= RET; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1490 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1491 xpos+=s_xinc2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1492 } |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1493 // funnyCode[0]= RET; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1494 } |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1495 |
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1496 #endif // HAVE_MMX2 |
2216 | 1497 } // reset counters |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1498 |
2216 | 1499 while(1){ |
2519 | 1500 unsigned char *dest =dstptr[0]+dststride*s_ypos; |
1501 unsigned char *uDest=dstptr[1]+(dststride>>1)*(s_ypos>>1); | |
1502 unsigned char *vDest=dstptr[2]+(dststride>>1)*(s_ypos>>1); | |
1503 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1504 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
|
1505 // points to the dst Pixels center in the source (0 is the center of pixel 0,0 in src) |
2519 | 1506 int srcuvpos= dstbpp==12 ? s_srcypos + s_yinc/2 - 0x8000 : |
1507 s_srcypos - 0x8000; | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1508 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
|
1509 int yalpha=((s_srcypos-1)&0xFFFF)>>4; |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1510 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
|
1511 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
|
1512 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
|
1513 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
|
1514 uint16_t *uvbuf1=pix_buf_uv[(y1+1)&1]; // bottom line of the interpolated slice |
2216 | 1515 int i; |
1516 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1517 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
|
1518 |
2519 | 1519 if((y0&1) && dstbpp==12) uvalpha=-1; // there is no alpha if there is no line |
2216 | 1520 |
1521 s_ypos++; s_srcypos+=s_yinc; | |
1522 | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1523 //only interpolate the src line horizontally if we didnt do it allready |
2469 | 1524 if(s_last_ypos!=y0) |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1525 { |
2469 | 1526 unsigned char *src; |
1527 // skip if first line has been horiz scaled alleady | |
1528 if(s_last_ypos != y0-1) | |
1529 { | |
1530 // check if first line is before any available src lines | |
1531 if(y0-1 < y) src=srcptr[0]+(0 )*stride[0]; | |
1532 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
|
1533 |
2469 | 1534 hyscale(buf0, dstw, src, srcWidth, s_xinc); |
1535 } | |
1536 // check if second line is after any available src lines | |
1537 if(y0-y >= h) src=srcptr[0]+(h-1)*stride[0]; | |
1538 else src=srcptr[0]+(y0-y)*stride[0]; | |
1539 | |
1540 // the min() is required to avoid reuseing lines which where not available | |
1541 s_last_ypos= MIN(y0, y+h-1); | |
1542 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
|
1543 } |
2469 | 1544 // printf("%d %d %d %d\n", y, y1, s_last_y1pos, h); |
2216 | 1545 // *** horizontal scale U and V lines to temp buffer |
2469 | 1546 if(s_last_y1pos!=y1) |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1547 { |
2469 | 1548 uint8_t *src1, *src2; |
1549 // skip if first line has been horiz scaled alleady | |
1550 if(s_last_y1pos != y1-1) | |
1551 { | |
1552 // check if first line is before any available src lines | |
1553 if(y1-y/2-1 < 0) | |
1554 { | |
1555 src1= srcptr[1]+(0)*stride[1]; | |
1556 src2= srcptr[2]+(0)*stride[2]; | |
1557 }else{ | |
1558 src1= srcptr[1]+(y1-y/2-1)*stride[1]; | |
1559 src2= srcptr[2]+(y1-y/2-1)*stride[2]; | |
1560 } | |
1561 hcscale(uvbuf0, dstUVw, src1, src2, srcWidth, s_xinc2); | |
1562 } | |
2264
7851375ea156
increased precission of s_xinc s_xinc2 (needed for the mmx2 bugfix)
michael
parents:
2237
diff
changeset
|
1563 |
2469 | 1564 // check if second line is after any available src lines |
1565 if(y1 - y/2 >= h/2) | |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1566 { |
2469 | 1567 src1= srcptr[1]+(h/2-1)*stride[1]; |
1568 src2= srcptr[2]+(h/2-1)*stride[2]; | |
1569 }else{ | |
1570 src1= srcptr[1]+(y1-y/2)*stride[1]; | |
1571 src2= srcptr[2]+(y1-y/2)*stride[2]; | |
2279
9b9c3363abbe
horizontal scaling bugs fixed, should be mostly bugfree now
michael
parents:
2274
diff
changeset
|
1572 } |
2469 | 1573 hcscale(uvbuf1, dstUVw, src1, src2, srcWidth, s_xinc2); |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1574 |
2469 | 1575 // the min() is required to avoid reuseing lines which where not available |
1576 s_last_y1pos= MIN(y1, y/2+h/2-1); | |
1577 } | |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1578 |
2519 | 1579 if(dstbpp==12) //YV12 |
1580 yuv2yuv(buf0, buf1, uvbuf0, uvbuf1, dest, uDest, vDest, dstw, yalpha, uvalpha); | |
1581 else if(ABS(s_yinc - 0x10000) < 10) | |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1582 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
|
1583 else |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1584 yuv2rgbX(buf0, buf1, uvbuf0, uvbuf1, dest, dstw, yalpha, uvalpha, dstbpp); |
2237 | 1585 |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1586 #ifdef HAVE_MMX |
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1587 b16Dither= b16Dither1; |
2232
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1588 b16Dither1= b16Dither2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1589 b16Dither2= b16Dither; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1590 |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1591 g16Dither= g16Dither1; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1592 g16Dither1= g16Dither2; |
65996b3467d7
MMX & MMX2 optimizations (MMX2 is buggy and commented out)
michael
parents:
2230
diff
changeset
|
1593 g16Dither2= g16Dither; |
2316
bcb229557e9b
fixed alignment (static variables where sometimes not 8-byte aligned)
michael
parents:
2297
diff
changeset
|
1594 #endif |
2216 | 1595 } |
2534
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1596 |
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1597 #ifdef HAVE_MMX |
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1598 __asm __volatile(SFENCE:::"memory"); |
2566 | 1599 __asm __volatile(EMMS:::"memory"); |
2534
cc9d3fd626f0
patch from Martin Decky <deckm1am@ss1000.ms.mff.cuni.cz> applied and unnecassery "memory" removed
michael
parents:
2521
diff
changeset
|
1600 #endif |
2216 | 1601 } |
1602 | |
1603 | |
1604 void SwScale_Init(){ | |
1605 // generating tables: | |
1606 int i; | |
1607 for(i=0;i<256;i++){ | |
1608 clip_table[i]=0; | |
1609 clip_table[i+256]=i; | |
1610 clip_table[i+512]=255; | |
1611 yuvtab_2568[i]=(0x2568*(i-16))+(256<<13); | |
1612 yuvtab_3343[i]=0x3343*(i-128); | |
1613 yuvtab_0c92[i]=-0x0c92*(i-128); | |
1614 yuvtab_1a1e[i]=-0x1a1e*(i-128); | |
1615 yuvtab_40cf[i]=0x40cf*(i-128); | |
1616 } | |
1617 | |
1618 } |