Mercurial > mplayer.hg
annotate libswscale/swscale_altivec_template.c @ 26892:3c360ca1c304
The size of output buffer is stored in 'osize', not 'size'.
This is just for readability, the code behaviour is not changed.
author | eugeni |
---|---|
date | Thu, 29 May 2008 18:35:22 +0000 |
parents | 4ffa92294ee7 |
children | 65b8334df960 |
rev | line source |
---|---|
19200
07769a1d06a9
Fix compile error due to extra "FF" at beginning of file in the FF{MIN,MAX}
pacman
parents:
19181
diff
changeset
|
1 /* |
20094
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
2 * AltiVec-enhanced yuv2yuvX |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
3 * |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
4 * Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org> |
26067 | 5 * based on the equivalent C code in swscale.c |
20094
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
6 * |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
7 * This file is part of FFmpeg. |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
8 * |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
9 * FFmpeg is free software; you can redistribute it and/or modify |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
10 * it under the terms of the GNU General Public License as published by |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
11 * the Free Software Foundation; either version 2 of the License, or |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
12 * (at your option) any later version. |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
13 * |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
14 * FFmpeg is distributed in the hope that it will be useful, |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
17 * GNU General Public License for more details. |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
18 * |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
19 * You should have received a copy of the GNU General Public License |
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
20 * along with FFmpeg; if not, write to the Free Software |
23702 | 21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
20094
aca9e9783f67
Change license headers to say 'FFmpeg' instead of 'this program'.
diego
parents:
20006
diff
changeset
|
22 */ |
18861 | 23 |
24 #define vzero vec_splat_s32(0) | |
25 | |
26 static inline void | |
27 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW) { | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
28 register int i; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
29 vector unsigned int altivec_vectorShiftInt19 = |
25750 | 30 vec_add(vec_splat_u32(10), vec_splat_u32(9)); |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
31 if ((unsigned long)dest % 16) { |
25217 | 32 /* badly aligned store, we force store alignment */ |
33 /* and will handle load misalignment on val w/ vec_perm */ | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
34 vector unsigned char perm1; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
35 vector signed int v1; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
36 for (i = 0 ; (i < dstW) && |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
37 (((unsigned long)dest + i) % 16) ; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
38 int t = val[i] >> 19; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
39 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
40 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
41 perm1 = vec_lvsl(i << 2, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
42 v1 = vec_ld(i << 2, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
43 for ( ; i < (dstW - 15); i+=16) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
44 int offset = i << 2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
45 vector signed int v2 = vec_ld(offset + 16, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
46 vector signed int v3 = vec_ld(offset + 32, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
47 vector signed int v4 = vec_ld(offset + 48, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
48 vector signed int v5 = vec_ld(offset + 64, val); |
25750 | 49 vector signed int v12 = vec_perm(v1, v2, perm1); |
50 vector signed int v23 = vec_perm(v2, v3, perm1); | |
51 vector signed int v34 = vec_perm(v3, v4, perm1); | |
52 vector signed int v45 = vec_perm(v4, v5, perm1); | |
23129 | 53 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
54 vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
55 vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
56 vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
57 vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
58 vector unsigned short vs1 = vec_packsu(vA, vB); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
59 vector unsigned short vs2 = vec_packsu(vC, vD); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
60 vector unsigned char vf = vec_packsu(vs1, vs2); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
61 vec_st(vf, i, dest); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
62 v1 = v5; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
63 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
64 } else { // dest is properly aligned, great |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
65 for (i = 0; i < (dstW - 15); i+=16) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
66 int offset = i << 2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
67 vector signed int v1 = vec_ld(offset, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
68 vector signed int v2 = vec_ld(offset + 16, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
69 vector signed int v3 = vec_ld(offset + 32, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
70 vector signed int v4 = vec_ld(offset + 48, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
71 vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
72 vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
73 vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
74 vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
75 vector unsigned short vs1 = vec_packsu(v5, v6); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
76 vector unsigned short vs2 = vec_packsu(v7, v8); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
77 vector unsigned char vf = vec_packsu(vs1, vs2); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
78 vec_st(vf, i, dest); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
79 } |
18861 | 80 } |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
81 for ( ; i < dstW ; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
82 int t = val[i] >> 19; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
83 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t); |
18861 | 84 } |
85 } | |
86 | |
87 static inline void | |
88 yuv2yuvX_altivec_real(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize, | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
89 int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize, |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
90 uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstW, int chrDstW) |
18861 | 91 { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
92 const vector signed int vini = {(1 << 18), (1 << 18), (1 << 18), (1 << 18)}; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
93 register int i, j; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
94 { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
95 int __attribute__ ((aligned (16))) val[dstW]; |
23129 | 96 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
97 for (i = 0; i < (dstW -7); i+=4) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
98 vec_st(vini, i << 2, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
99 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
100 for (; i < dstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
101 val[i] = (1 << 18); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
102 } |
23129 | 103 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
104 for (j = 0; j < lumFilterSize; j++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
105 vector signed short l1, vLumFilter = vec_ld(j << 1, lumFilter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
106 vector unsigned char perm, perm0 = vec_lvsl(j << 1, lumFilter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
107 vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
108 vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter |
23129 | 109 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
110 perm = vec_lvsl(0, lumSrc[j]); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
111 l1 = vec_ld(0, lumSrc[j]); |
23129 | 112 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
113 for (i = 0; i < (dstW - 7); i+=8) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
114 int offset = i << 2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
115 vector signed short l2 = vec_ld((i << 1) + 16, lumSrc[j]); |
23129 | 116 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
117 vector signed int v1 = vec_ld(offset, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
118 vector signed int v2 = vec_ld(offset + 16, val); |
23129 | 119 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
120 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7] |
23129 | 121 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
122 vector signed int i1 = vec_mule(vLumFilter, ls); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
123 vector signed int i2 = vec_mulo(vLumFilter, ls); |
23129 | 124 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
125 vector signed int vf1 = vec_mergeh(i1, i2); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
126 vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j] |
23129 | 127 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
128 vector signed int vo1 = vec_add(v1, vf1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
129 vector signed int vo2 = vec_add(v2, vf2); |
23129 | 130 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
131 vec_st(vo1, offset, val); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
132 vec_st(vo2, offset + 16, val); |
23129 | 133 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
134 l1 = l2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
135 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
136 for ( ; i < dstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
137 val[i] += lumSrc[j][i] * lumFilter[j]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
138 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
139 } |
25750 | 140 altivec_packIntArrayToCharArray(val, dest, dstW); |
18861 | 141 } |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
142 if (uDest != 0) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
143 int __attribute__ ((aligned (16))) u[chrDstW]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
144 int __attribute__ ((aligned (16))) v[chrDstW]; |
18861 | 145 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
146 for (i = 0; i < (chrDstW -7); i+=4) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
147 vec_st(vini, i << 2, u); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
148 vec_st(vini, i << 2, v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
149 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
150 for (; i < chrDstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
151 u[i] = (1 << 18); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
152 v[i] = (1 << 18); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
153 } |
23129 | 154 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
155 for (j = 0; j < chrFilterSize; j++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
156 vector signed short l1, l1_V, vChrFilter = vec_ld(j << 1, chrFilter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
157 vector unsigned char perm, perm0 = vec_lvsl(j << 1, chrFilter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
158 vChrFilter = vec_perm(vChrFilter, vChrFilter, perm0); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
159 vChrFilter = vec_splat(vChrFilter, 0); // chrFilter[j] is loaded 8 times in vChrFilter |
23129 | 160 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
161 perm = vec_lvsl(0, chrSrc[j]); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
162 l1 = vec_ld(0, chrSrc[j]); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
163 l1_V = vec_ld(2048 << 1, chrSrc[j]); |
23129 | 164 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
165 for (i = 0; i < (chrDstW - 7); i+=8) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
166 int offset = i << 2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
167 vector signed short l2 = vec_ld((i << 1) + 16, chrSrc[j]); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
168 vector signed short l2_V = vec_ld(((i + 2048) << 1) + 16, chrSrc[j]); |
23129 | 169 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
170 vector signed int v1 = vec_ld(offset, u); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
171 vector signed int v2 = vec_ld(offset + 16, u); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
172 vector signed int v1_V = vec_ld(offset, v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
173 vector signed int v2_V = vec_ld(offset + 16, v); |
23129 | 174 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
175 vector signed short ls = vec_perm(l1, l2, perm); // chrSrc[j][i] ... chrSrc[j][i+7] |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
176 vector signed short ls_V = vec_perm(l1_V, l2_V, perm); // chrSrc[j][i+2048] ... chrSrc[j][i+2055] |
23129 | 177 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
178 vector signed int i1 = vec_mule(vChrFilter, ls); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
179 vector signed int i2 = vec_mulo(vChrFilter, ls); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
180 vector signed int i1_V = vec_mule(vChrFilter, ls_V); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
181 vector signed int i2_V = vec_mulo(vChrFilter, ls_V); |
23129 | 182 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
183 vector signed int vf1 = vec_mergeh(i1, i2); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
184 vector signed int vf2 = vec_mergel(i1, i2); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j] |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
185 vector signed int vf1_V = vec_mergeh(i1_V, i2_V); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
186 vector signed int vf2_V = vec_mergel(i1_V, i2_V); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j] |
23129 | 187 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
188 vector signed int vo1 = vec_add(v1, vf1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
189 vector signed int vo2 = vec_add(v2, vf2); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
190 vector signed int vo1_V = vec_add(v1_V, vf1_V); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
191 vector signed int vo2_V = vec_add(v2_V, vf2_V); |
23129 | 192 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
193 vec_st(vo1, offset, u); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
194 vec_st(vo2, offset + 16, u); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
195 vec_st(vo1_V, offset, v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
196 vec_st(vo2_V, offset + 16, v); |
23129 | 197 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
198 l1 = l2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
199 l1_V = l2_V; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
200 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
201 for ( ; i < chrDstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
202 u[i] += chrSrc[j][i] * chrFilter[j]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
203 v[i] += chrSrc[j][i + 2048] * chrFilter[j]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
204 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
205 } |
25750 | 206 altivec_packIntArrayToCharArray(u, uDest, chrDstW); |
207 altivec_packIntArrayToCharArray(v, vDest, chrDstW); | |
18861 | 208 } |
209 } | |
210 | |
211 static inline void hScale_altivec_real(int16_t *dst, int dstW, uint8_t *src, int srcW, int xInc, int16_t *filter, int16_t *filterPos, int filterSize) { | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
212 register int i; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
213 int __attribute__ ((aligned (16))) tempo[4]; |
18861 | 214 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
215 if (filterSize % 4) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
216 for (i=0; i<dstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
217 register int j; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
218 register int srcPos = filterPos[i]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
219 register int val = 0; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
220 for (j=0; j<filterSize; j++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
221 val += ((int)src[srcPos + j])*filter[filterSize*i + j]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
222 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
223 dst[i] = av_clip(val>>7, 0, (1<<15)-1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
224 } |
18861 | 225 } |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
226 else |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
227 switch (filterSize) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
228 case 4: |
18861 | 229 { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
230 for (i=0; i<dstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
231 register int srcPos = filterPos[i]; |
18861 | 232 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
233 vector unsigned char src_v0 = vec_ld(srcPos, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
234 vector unsigned char src_v1, src_vF; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
235 vector signed short src_v, filter_v; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
236 vector signed int val_vEven, val_s; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
237 if ((((int)src + srcPos)% 16) > 12) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
238 src_v1 = vec_ld(srcPos + 16, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
239 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
240 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src)); |
18861 | 241 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
242 src_v = // vec_unpackh sign-extends... |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
243 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
244 // now put our elements in the even slots |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
245 src_v = vec_mergeh(src_v, (vector signed short)vzero); |
18861 | 246 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
247 filter_v = vec_ld(i << 3, filter); |
18861 | 248 // the 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2) |
249 | |
250 // the neat trick : we only care for half the elements, | |
251 // high or low depending on (i<<3)%16 (it's 0 or 8 here), | |
252 // and we're going to use vec_mule, so we chose | |
253 // carefully how to "unpack" the elements into the even slots | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
254 if ((i << 3) % 16) |
25750 | 255 filter_v = vec_mergel(filter_v, (vector signed short)vzero); |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
256 else |
25750 | 257 filter_v = vec_mergeh(filter_v, (vector signed short)vzero); |
18861 | 258 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
259 val_vEven = vec_mule(src_v, filter_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
260 val_s = vec_sums(val_vEven, vzero); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
261 vec_st(val_s, 0, tempo); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
262 dst[i] = av_clip(tempo[3]>>7, 0, (1<<15)-1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
263 } |
18861 | 264 } |
265 break; | |
266 | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
267 case 8: |
18861 | 268 { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
269 for (i=0; i<dstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
270 register int srcPos = filterPos[i]; |
18861 | 271 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
272 vector unsigned char src_v0 = vec_ld(srcPos, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
273 vector unsigned char src_v1, src_vF; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
274 vector signed short src_v, filter_v; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
275 vector signed int val_v, val_s; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
276 if ((((int)src + srcPos)% 16) > 8) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
277 src_v1 = vec_ld(srcPos + 16, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
278 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
279 src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src)); |
18861 | 280 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
281 src_v = // vec_unpackh sign-extends... |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
282 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
283 filter_v = vec_ld(i << 4, filter); |
18861 | 284 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2) |
285 | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
286 val_v = vec_msums(src_v, filter_v, (vector signed int)vzero); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
287 val_s = vec_sums(val_v, vzero); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
288 vec_st(val_s, 0, tempo); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
289 dst[i] = av_clip(tempo[3]>>7, 0, (1<<15)-1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
290 } |
18861 | 291 } |
292 break; | |
293 | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
294 case 16: |
18861 | 295 { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
296 for (i=0; i<dstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
297 register int srcPos = filterPos[i]; |
18861 | 298 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
299 vector unsigned char src_v0 = vec_ld(srcPos, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
300 vector unsigned char src_v1 = vec_ld(srcPos + 16, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
301 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src)); |
18861 | 302 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
303 vector signed short src_vA = // vec_unpackh sign-extends... |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
304 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
305 vector signed short src_vB = // vec_unpackh sign-extends... |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
306 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF)); |
18861 | 307 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
308 vector signed short filter_v0 = vec_ld(i << 5, filter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
309 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
310 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2) |
18861 | 311 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
312 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
313 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc); |
18861 | 314 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
315 vector signed int val_s = vec_sums(val_v, vzero); |
18861 | 316 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
317 vec_st(val_s, 0, tempo); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
318 dst[i] = av_clip(tempo[3]>>7, 0, (1<<15)-1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
319 } |
18861 | 320 } |
321 break; | |
23129 | 322 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
323 default: |
18861 | 324 { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
325 for (i=0; i<dstW; i++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
326 register int j; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
327 register int srcPos = filterPos[i]; |
18861 | 328 |
20584 | 329 vector signed int val_s, val_v = (vector signed int)vzero; |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
330 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter); |
18861 | 331 vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter); |
332 | |
333 vector unsigned char src_v0 = vec_ld(srcPos, src); | |
334 vector unsigned char permS = vec_lvsl(srcPos, src); | |
335 | |
336 for (j = 0 ; j < filterSize - 15; j += 16) { | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
337 vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
338 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS); |
23129 | 339 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
340 vector signed short src_vA = // vec_unpackh sign-extends... |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
341 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
342 vector signed short src_vB = // vec_unpackh sign-extends... |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
343 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF)); |
23129 | 344 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
345 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
346 vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
347 vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
348 vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF); |
23129 | 349 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
350 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
351 val_v = vec_msums(src_vB, filter_v1, val_acc); |
18861 | 352 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
353 filter_v0R = filter_v2R; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
354 src_v0 = src_v1; |
18861 | 355 } |
356 | |
25751 | 357 if (j < filterSize-7) { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
358 // loading src_v0 is useless, it's already done above |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
359 //vector unsigned char src_v0 = vec_ld(srcPos + j, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
360 vector unsigned char src_v1, src_vF; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
361 vector signed short src_v, filter_v1R, filter_v; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
362 if ((((int)src + srcPos)% 16) > 8) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
363 src_v1 = vec_ld(srcPos + j + 16, src); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
364 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
365 src_vF = vec_perm(src_v0, src_v1, permS); |
23129 | 366 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
367 src_v = // vec_unpackh sign-extends... |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
368 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
369 // loading filter_v0R is useless, it's already done above |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
370 //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
371 filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
372 filter_v = vec_perm(filter_v0R, filter_v1R, permF); |
23129 | 373 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
374 val_v = vec_msums(src_v, filter_v, val_v); |
18861 | 375 } |
376 | |
20584 | 377 val_s = vec_sums(val_v, vzero); |
23129 | 378 |
18861 | 379 vec_st(val_s, 0, tempo); |
23129 | 380 dst[i] = av_clip(tempo[3]>>7, 0, (1<<15)-1); |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
381 } |
23129 | 382 |
18861 | 383 } |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
384 } |
18861 | 385 } |
386 | |
387 static inline int yv12toyuy2_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY, | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
388 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
389 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY; |
25750 | 390 // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, srcStride[0], srcStride[1], dstStride[0]); |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
391 uint8_t *ysrc = src[0]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
392 uint8_t *usrc = src[1]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
393 uint8_t *vsrc = src[2]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
394 const int width = c->srcW; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
395 const int height = srcSliceH; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
396 const int lumStride = srcStride[0]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
397 const int chromStride = srcStride[1]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
398 const int dstStride = dstStride_a[0]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
399 const vector unsigned char yperm = vec_lvsl(0, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
400 const int vertLumPerChroma = 2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
401 register unsigned int y; |
23129 | 402 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
403 if (width&15) { |
25750 | 404 yv12toyuy2(ysrc, usrc, vsrc, dst, c->srcW, srcSliceH, lumStride, chromStride, dstStride); |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
405 return srcSliceH; |
18861 | 406 } |
23129 | 407 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
408 /* this code assume: |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
409 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
410 1) dst is 16 bytes-aligned |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
411 2) dstStride is a multiple of 16 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
412 3) width is a multiple of 16 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
413 4) lum&chrom stride are multiple of 8 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
414 */ |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
415 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
416 for (y=0; y<height; y++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
417 int i; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
418 for (i = 0; i < width - 31; i+= 32) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
419 const unsigned int j = i >> 1; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
420 vector unsigned char v_yA = vec_ld(i, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
421 vector unsigned char v_yB = vec_ld(i + 16, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
422 vector unsigned char v_yC = vec_ld(i + 32, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
423 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
424 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
425 vector unsigned char v_uA = vec_ld(j, usrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
426 vector unsigned char v_uB = vec_ld(j + 16, usrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
427 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
428 vector unsigned char v_vA = vec_ld(j, vsrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
429 vector unsigned char v_vB = vec_ld(j + 16, vsrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
430 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
431 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
432 vector unsigned char v_uv_b = vec_mergel(v_u, v_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
433 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
434 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
435 vector unsigned char v_yuy2_2 = vec_mergeh(v_y2, v_uv_b); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
436 vector unsigned char v_yuy2_3 = vec_mergel(v_y2, v_uv_b); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
437 vec_st(v_yuy2_0, (i << 1), dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
438 vec_st(v_yuy2_1, (i << 1) + 16, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
439 vec_st(v_yuy2_2, (i << 1) + 32, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
440 vec_st(v_yuy2_3, (i << 1) + 48, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
441 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
442 if (i < width) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
443 const unsigned int j = i >> 1; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
444 vector unsigned char v_y1 = vec_ld(i, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
445 vector unsigned char v_u = vec_ld(j, usrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
446 vector unsigned char v_v = vec_ld(j, vsrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
447 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
448 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
449 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
450 vec_st(v_yuy2_0, (i << 1), dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
451 vec_st(v_yuy2_1, (i << 1) + 16, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
452 } |
25751 | 453 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
454 usrc += chromStride; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
455 vsrc += chromStride; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
456 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
457 ysrc += lumStride; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
458 dst += dstStride; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
459 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
460 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
461 return srcSliceH; |
18861 | 462 } |
463 | |
464 static inline int yv12touyvy_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY, | |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
465 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
466 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY; |
25750 | 467 // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, srcStride[0], srcStride[1], dstStride[0]); |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
468 uint8_t *ysrc = src[0]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
469 uint8_t *usrc = src[1]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
470 uint8_t *vsrc = src[2]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
471 const int width = c->srcW; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
472 const int height = srcSliceH; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
473 const int lumStride = srcStride[0]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
474 const int chromStride = srcStride[1]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
475 const int dstStride = dstStride_a[0]; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
476 const int vertLumPerChroma = 2; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
477 const vector unsigned char yperm = vec_lvsl(0, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
478 register unsigned int y; |
18861 | 479 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
480 if (width&15) { |
25750 | 481 yv12touyvy(ysrc, usrc, vsrc, dst, c->srcW, srcSliceH, lumStride, chromStride, dstStride); |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
482 return srcSliceH; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
483 } |
18861 | 484 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
485 /* this code assume: |
18861 | 486 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
487 1) dst is 16 bytes-aligned |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
488 2) dstStride is a multiple of 16 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
489 3) width is a multiple of 16 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
490 4) lum&chrom stride are multiple of 8 |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
491 */ |
23129 | 492 |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
493 for (y=0; y<height; y++) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
494 int i; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
495 for (i = 0; i < width - 31; i+= 32) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
496 const unsigned int j = i >> 1; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
497 vector unsigned char v_yA = vec_ld(i, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
498 vector unsigned char v_yB = vec_ld(i + 16, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
499 vector unsigned char v_yC = vec_ld(i + 32, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
500 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
501 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
502 vector unsigned char v_uA = vec_ld(j, usrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
503 vector unsigned char v_uB = vec_ld(j + 16, usrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
504 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
505 vector unsigned char v_vA = vec_ld(j, vsrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
506 vector unsigned char v_vB = vec_ld(j + 16, vsrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
507 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc)); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
508 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
509 vector unsigned char v_uv_b = vec_mergel(v_u, v_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
510 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
511 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
512 vector unsigned char v_uyvy_2 = vec_mergeh(v_uv_b, v_y2); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
513 vector unsigned char v_uyvy_3 = vec_mergel(v_uv_b, v_y2); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
514 vec_st(v_uyvy_0, (i << 1), dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
515 vec_st(v_uyvy_1, (i << 1) + 16, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
516 vec_st(v_uyvy_2, (i << 1) + 32, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
517 vec_st(v_uyvy_3, (i << 1) + 48, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
518 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
519 if (i < width) { |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
520 const unsigned int j = i >> 1; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
521 vector unsigned char v_y1 = vec_ld(i, ysrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
522 vector unsigned char v_u = vec_ld(j, usrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
523 vector unsigned char v_v = vec_ld(j, vsrc); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
524 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
525 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
526 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
527 vec_st(v_uyvy_0, (i << 1), dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
528 vec_st(v_uyvy_1, (i << 1) + 16, dst); |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
529 } |
25751 | 530 if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) { |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
531 usrc += chromStride; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
532 vsrc += chromStride; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
533 } |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
534 ysrc += lumStride; |
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
535 dst += dstStride; |
18861 | 536 } |
23140
4d3870361b73
cosmetics attack, part I: Remove all tabs and prettyprint/reindent the code.
diego
parents:
23129
diff
changeset
|
537 return srcSliceH; |
18861 | 538 } |