annotate postproc/swscale_altivec_template.c @ 13809:a9e2fda933be

AMD64 fixes based on patch by Timo Ter¸«£s <timo.teras@iki.fi>
author faust3
date Sat, 30 Oct 2004 12:17:16 +0000
parents 931eee818c52
children 53231c701b29
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
12017
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
1 /*
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
2 AltiVec-enhanced yuv2yuvX
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
3
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
4 Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org>
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
5 based on the equivalent C code in "postproc/swscale.c"
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
6
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
7
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
8 This program is free software; you can redistribute it and/or modify
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
9 it under the terms of the GNU General Public License as published by
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
10 the Free Software Foundation; either version 2 of the License, or
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
11 (at your option) any later version.
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
12
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
13 This program is distributed in the hope that it will be useful,
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
16 GNU General Public License for more details.
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
17
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
18 You should have received a copy of the GNU General Public License
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
19 along with this program; if not, write to the Free Software
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
20 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
21 */
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
22
12130
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
23 #ifdef CONFIG_DARWIN
12532
79a2af950cf7 small linux/altivec compile fix in postproc/ by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12130
diff changeset
24 #define AVV(x...) (x)
12130
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
25 #else
12532
79a2af950cf7 small linux/altivec compile fix in postproc/ by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12130
diff changeset
26 #define AVV(x...) {x}
79a2af950cf7 small linux/altivec compile fix in postproc/ by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12130
diff changeset
27 #endif
79a2af950cf7 small linux/altivec compile fix in postproc/ by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12130
diff changeset
28
12130
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
29 static const vector signed int vzero =
12532
79a2af950cf7 small linux/altivec compile fix in postproc/ by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12130
diff changeset
30 (const vector signed int)AVV(0, 0, 0, 0);
12130
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
31 static const vector unsigned int altivec_vectorShiftInt19 =
12532
79a2af950cf7 small linux/altivec compile fix in postproc/ by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12130
diff changeset
32 (const vector unsigned int)AVV(19, 19, 19, 19);
12130
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
33
12017
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
34 static inline void
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
35 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
36 register int i;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
37 if ((unsigned long)dest % 16) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
38 /* badly aligned store, we force store alignement */
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
39 /* and will handle load misalignement on val w/ vec_perm */
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
40 for (i = 0 ; (i < dstW) &&
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
41 (((unsigned long)dest + i) % 16) ; i++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
42 int t = val[i] >> 19;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
43 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
44 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
45 vector unsigned char perm1 = vec_lvsl(i << 2, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
46 vector signed int v1 = vec_ld(i << 2, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
47 for ( ; i < (dstW - 15); i+=16) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
48 int offset = i << 2;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
49 vector signed int v2 = vec_ld(offset + 16, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
50 vector signed int v3 = vec_ld(offset + 32, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
51 vector signed int v4 = vec_ld(offset + 48, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
52 vector signed int v5 = vec_ld(offset + 64, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
53 vector signed int v12 = vec_perm(v1,v2,perm1);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
54 vector signed int v23 = vec_perm(v2,v3,perm1);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
55 vector signed int v34 = vec_perm(v3,v4,perm1);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
56 vector signed int v45 = vec_perm(v4,v5,perm1);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
57
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
58 vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
59 vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
60 vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
61 vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
62 vector unsigned short vs1 = vec_packsu(vA, vB);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
63 vector unsigned short vs2 = vec_packsu(vC, vD);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
64 vector unsigned char vf = vec_packsu(vs1, vs2);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
65 vec_st(vf, i, dest);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
66 v1 = v5;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
67 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
68 } else { // dest is properly aligned, great
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
69 for (i = 0; i < (dstW - 15); i+=16) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
70 int offset = i << 2;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
71 vector signed int v1 = vec_ld(offset, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
72 vector signed int v2 = vec_ld(offset + 16, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
73 vector signed int v3 = vec_ld(offset + 32, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
74 vector signed int v4 = vec_ld(offset + 48, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
75 vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
76 vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
77 vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
78 vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
79 vector unsigned short vs1 = vec_packsu(v5, v6);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
80 vector unsigned short vs2 = vec_packsu(v7, v8);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
81 vector unsigned char vf = vec_packsu(vs1, vs2);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
82 vec_st(vf, i, dest);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
83 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
84 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
85 for ( ; i < dstW ; i++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
86 int t = val[i] >> 19;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
87 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
88 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
89 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
90
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
91 static inline void
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
92 yuv2yuvX_altivec_real(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize,
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
93 int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize,
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
94 uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstW, int chrDstW)
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
95 {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
96 const vector signed int vini = {(1 << 18), (1 << 18), (1 << 18), (1 << 18)};
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
97 register int i, j;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
98 {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
99 int __attribute__ ((aligned (16))) val[dstW];
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
100
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
101 for (i = 0; i < (dstW -7); i+=4) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
102 vec_st(vini, i << 2, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
103 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
104 for (; i < dstW; i++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
105 val[i] = (1 << 18);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
106 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
107
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
108 for (j = 0; j < lumFilterSize; j++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
109 vector signed short vLumFilter = vec_ld(j << 1, lumFilter);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
110 vector unsigned char perm0 = vec_lvsl(j << 1, lumFilter);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
111 vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
112 vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
113
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
114 vector unsigned char perm = vec_lvsl(0, lumSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
115 vector signed short l1 = vec_ld(0, lumSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
116
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
117 for (i = 0; i < (dstW - 7); i+=8) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
118 int offset = i << 2;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
119 vector signed short l2 = vec_ld((i << 1) + 16, lumSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
120
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
121 vector signed int v1 = vec_ld(offset, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
122 vector signed int v2 = vec_ld(offset + 16, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
123
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
124 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
125
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
126 vector signed int i1 = vec_mule(vLumFilter, ls);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
127 vector signed int i2 = vec_mulo(vLumFilter, ls);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
128
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
129 vector signed int vf1 = vec_mergeh(i1, i2);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
130 vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
131
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
132 vector signed int vo1 = vec_add(v1, vf1);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
133 vector signed int vo2 = vec_add(v2, vf2);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
134
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
135 vec_st(vo1, offset, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
136 vec_st(vo2, offset + 16, val);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
137
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
138 l1 = l2;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
139 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
140 for ( ; i < dstW; i++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
141 val[i] += lumSrc[j][i] * lumFilter[j];
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
142 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
143 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
144 altivec_packIntArrayToCharArray(val,dest,dstW);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
145 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
146 if (uDest != 0) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
147 int __attribute__ ((aligned (16))) u[chrDstW];
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
148 int __attribute__ ((aligned (16))) v[chrDstW];
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
149
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
150 for (i = 0; i < (chrDstW -7); i+=4) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
151 vec_st(vini, i << 2, u);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
152 vec_st(vini, i << 2, v);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
153 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
154 for (; i < chrDstW; i++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
155 u[i] = (1 << 18);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
156 v[i] = (1 << 18);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
157 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
158
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
159 for (j = 0; j < chrFilterSize; j++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
160 vector signed short vChrFilter = vec_ld(j << 1, chrFilter);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
161 vector unsigned char perm0 = vec_lvsl(j << 1, chrFilter);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
162 vChrFilter = vec_perm(vChrFilter, vChrFilter, perm0);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
163 vChrFilter = vec_splat(vChrFilter, 0); // chrFilter[j] is loaded 8 times in vChrFilter
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
164
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
165 vector unsigned char perm = vec_lvsl(0, chrSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
166 vector signed short l1 = vec_ld(0, chrSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
167 vector signed short l1_V = vec_ld(2048 << 1, chrSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
168
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
169 for (i = 0; i < (chrDstW - 7); i+=8) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
170 int offset = i << 2;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
171 vector signed short l2 = vec_ld((i << 1) + 16, chrSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
172 vector signed short l2_V = vec_ld(((i + 2048) << 1) + 16, chrSrc[j]);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
173
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
174 vector signed int v1 = vec_ld(offset, u);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
175 vector signed int v2 = vec_ld(offset + 16, u);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
176 vector signed int v1_V = vec_ld(offset, v);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
177 vector signed int v2_V = vec_ld(offset + 16, v);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
178
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
179 vector signed short ls = vec_perm(l1, l2, perm); // chrSrc[j][i] ... chrSrc[j][i+7]
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
180 vector signed short ls_V = vec_perm(l1_V, l2_V, perm); // chrSrc[j][i+2048] ... chrSrc[j][i+2055]
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
181
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
182 vector signed int i1 = vec_mule(vChrFilter, ls);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
183 vector signed int i2 = vec_mulo(vChrFilter, ls);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
184 vector signed int i1_V = vec_mule(vChrFilter, ls_V);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
185 vector signed int i2_V = vec_mulo(vChrFilter, ls_V);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
186
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
187 vector signed int vf1 = vec_mergeh(i1, i2);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
188 vector signed int vf2 = vec_mergel(i1, i2); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
189 vector signed int vf1_V = vec_mergeh(i1_V, i2_V);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
190 vector signed int vf2_V = vec_mergel(i1_V, i2_V); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
191
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
192 vector signed int vo1 = vec_add(v1, vf1);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
193 vector signed int vo2 = vec_add(v2, vf2);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
194 vector signed int vo1_V = vec_add(v1_V, vf1_V);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
195 vector signed int vo2_V = vec_add(v2_V, vf2_V);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
196
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
197 vec_st(vo1, offset, u);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
198 vec_st(vo2, offset + 16, u);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
199 vec_st(vo1_V, offset, v);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
200 vec_st(vo2_V, offset + 16, v);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
201
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
202 l1 = l2;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
203 l1_V = l2_V;
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
204 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
205 for ( ; i < chrDstW; i++) {
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
206 u[i] += chrSrc[j][i] * chrFilter[j];
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
207 v[i] += chrSrc[j][i + 2048] * chrFilter[j];
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
208 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
209 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
210 altivec_packIntArrayToCharArray(u,uDest,chrDstW);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
211 altivec_packIntArrayToCharArray(v,vDest,chrDstW);
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
212 }
21e5cb258a95 AltiVec support in postproc/ + altivec optimizations for yuv2yuvX patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
diff changeset
213 }
12130
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
214
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
215 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) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
216 register int i;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
217 int __attribute__ ((aligned (16))) tempo[4];
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
218
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
219 if (filterSize % 4) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
220 for(i=0; i<dstW; i++) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
221 register int j;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
222 register int srcPos = filterPos[i];
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
223 register int val = 0;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
224 for(j=0; j<filterSize; j++) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
225 val += ((int)src[srcPos + j])*filter[filterSize*i + j];
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
226 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
227 dst[i] = MIN(MAX(0, val>>7), (1<<15)-1);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
228 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
229 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
230 else
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
231 switch (filterSize) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
232 case 4:
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
233 {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
234 for(i=0; i<dstW; i++) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
235 register int j;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
236 register int srcPos = filterPos[i];
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
237
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
238 vector unsigned char src_v0 = vec_ld(srcPos, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
239 vector unsigned char src_v1;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
240 if ((((int)src + srcPos)% 16) > 12) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
241 src_v1 = vec_ld(srcPos + 16, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
242 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
243 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
244
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
245 vector signed short src_v = // vec_unpackh sign-extends...
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
246 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
247 // now put our elements in the even slots
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
248 src_v = vec_mergeh(src_v, (vector signed short)vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
249
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
250 vector signed short filter_v = vec_ld(i << 3, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
251 // the 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2)
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
252
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
253 // the neat trick : we only care for half the elements,
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
254 // high or low depending on (i<<3)%16 (it's 0 or 8 here),
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
255 // and we're going to use vec_mule, so we chose
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
256 // carefully how to "unpack" the elements into the even slots
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
257 if ((i << 3) % 16)
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
258 filter_v = vec_mergel(filter_v,(vector signed short)vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
259 else
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
260 filter_v = vec_mergeh(filter_v,(vector signed short)vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
261
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
262 vector signed int val_vEven = vec_mule(src_v, filter_v);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
263 vector signed int val_s = vec_sums(val_vEven, vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
264 vec_st(val_s, 0, tempo);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
265 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
266 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
267 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
268 break;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
269
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
270 case 8:
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
271 {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
272 for(i=0; i<dstW; i++) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
273 register int srcPos = filterPos[i];
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
274
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
275 vector unsigned char src_v0 = vec_ld(srcPos, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
276 vector unsigned char src_v1;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
277 if ((((int)src + srcPos)% 16) > 8) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
278 src_v1 = vec_ld(srcPos + 16, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
279 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
280 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
281
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
282 vector signed short src_v = // vec_unpackh sign-extends...
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
283 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
284 vector signed short filter_v = vec_ld(i << 4, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
285 // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
286
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
287 vector signed int val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
288 vector signed int val_s = vec_sums(val_v, vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
289 vec_st(val_s, 0, tempo);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
290 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
291 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
292 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
293 break;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
294
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
295 case 16:
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
296 {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
297 for(i=0; i<dstW; i++) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
298 register int srcPos = filterPos[i];
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
299
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
300 vector unsigned char src_v0 = vec_ld(srcPos, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
301 vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
302 vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
303
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
304 vector signed short src_vA = // vec_unpackh sign-extends...
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
305 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
306 vector signed short src_vB = // vec_unpackh sign-extends...
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
307 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
308
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
309 vector signed short filter_v0 = vec_ld(i << 5, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
310 vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
311 // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
312
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
313 vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
314 vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
315
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
316 vector signed int val_s = vec_sums(val_v, vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
317
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
318 vec_st(val_s, 0, tempo);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
319 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
320 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
321 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
322 break;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
323
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
324 default:
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
325 {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
326 for(i=0; i<dstW; i++) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
327 register int j;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
328 register int srcPos = filterPos[i];
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
329
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
330 vector signed int val_v = (vector signed int)vzero;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
331 vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
332 vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
333
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
334 vector unsigned char src_v0 = vec_ld(srcPos, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
335 vector unsigned char permS = vec_lvsl(srcPos, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
336
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
337 for (j = 0 ; j < filterSize - 15; j += 16) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
338 vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
339 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
340
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
341 vector signed short src_vA = // vec_unpackh sign-extends...
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
342 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
343 vector signed short src_vB = // vec_unpackh sign-extends...
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
344 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
345
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
346 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
347 vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
348 vector signed short filter_v0 = vec_perm(filter_v0R, filter_v1R, permF);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
349 vector signed short filter_v1 = vec_perm(filter_v1R, filter_v2R, permF);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
350
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
351 vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
352 val_v = vec_msums(src_vB, filter_v1, val_acc);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
353
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
354 filter_v0R = filter_v2R;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
355 src_v0 = src_v1;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
356 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
357
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
358 if (j < (filterSize-7)) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
359 // loading src_v0 is useless, it's already done above
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
360 //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
361 vector unsigned char src_v1;
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
362 if ((((int)src + srcPos)% 16) > 8) {
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
363 src_v1 = vec_ld(srcPos + j + 16, src);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
364 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
365 vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
366
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
367 vector signed short src_v = // vec_unpackh sign-extends...
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
368 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
369 // loading filter_v0R is useless, it's already done above
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
370 //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
371 vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
372 vector signed short filter_v = vec_perm(filter_v0R, filter_v1R, permF);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
373
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
374 val_v = vec_msums(src_v, filter_v, val_v);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
375 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
376
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
377 vector signed int val_s = vec_sums(val_v, vzero);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
378
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
379 vec_st(val_s, 0, tempo);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
380 dst[i] = MIN(MAX(0, tempo[3]>>7), (1<<15)-1);
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
381 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
382
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
383 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
384 }
2ef24558b732 AltiVec hScale, all size patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michael
parents: 12017
diff changeset
385 }
12768
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
386
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
387 static inline int yv12toyuy2_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
388 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
389 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
390 // yv12toyuy2( src[0],src[1],src[2],dst,c->srcW,srcSliceH,srcStride[0],srcStride[1],dstStride[0] );
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
391 uint8_t *ysrc = src[0];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
392 uint8_t *usrc = src[1];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
393 uint8_t *vsrc = src[2];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
394 const int width = c->srcW;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
395 const int height = srcSliceH;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
396 const int lumStride = srcStride[0];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
397 const int chromStride = srcStride[1];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
398 const int dstStride = dstStride_a[0];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
399 const vector unsigned char yperm = vec_lvsl(0, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
400 const int vertLumPerChroma = 2;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
401 register unsigned int y;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
402
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
403 /* this code assume:
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
404
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
405 1) dst is 16 bytes-aligned
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
406 2) dstStride is a multiple of 16
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
407 3) width is a multiple of 16
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
408 4) lum&chrom stride are multiple of 8
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
409 */
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
410
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
411 for(y=0; y<height; y++)
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
412 {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
413 int i;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
414 for (i = 0; i < width - 31; i+= 32) {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
415 const unsigned int j = i >> 1;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
416 vector unsigned char v_yA = vec_ld(i, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
417 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
418 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
419 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
420 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
421 vector unsigned char v_uA = vec_ld(j, usrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
422 vector unsigned char v_uB = vec_ld(j + 16, usrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
423 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
424 vector unsigned char v_vA = vec_ld(j, vsrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
425 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
426 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
427 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
428 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
429 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
430 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
431 vector unsigned char v_yuy2_2 = vec_mergeh(v_y2, v_uv_b);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
432 vector unsigned char v_yuy2_3 = vec_mergel(v_y2, v_uv_b);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
433 vec_st(v_yuy2_0, (i << 1), dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
434 vec_st(v_yuy2_1, (i << 1) + 16, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
435 vec_st(v_yuy2_2, (i << 1) + 32, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
436 vec_st(v_yuy2_3, (i << 1) + 48, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
437 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
438 if (i < width) {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
439 const unsigned int j = i >> 1;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
440 vector unsigned char v_y1 = vec_ld(i, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
441 vector unsigned char v_u = vec_ld(j, usrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
442 vector unsigned char v_v = vec_ld(j, vsrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
443 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
444 vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
445 vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
446 vec_st(v_yuy2_0, (i << 1), dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
447 vec_st(v_yuy2_1, (i << 1) + 16, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
448 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
449 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
450 {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
451 usrc += chromStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
452 vsrc += chromStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
453 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
454 ysrc += lumStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
455 dst += dstStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
456 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
457
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
458 return srcSliceH;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
459 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
460
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
461 static inline int yv12touyvy_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
462 int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
463 uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
464 // yv12toyuy2( src[0],src[1],src[2],dst,c->srcW,srcSliceH,srcStride[0],srcStride[1],dstStride[0] );
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
465 uint8_t *ysrc = src[0];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
466 uint8_t *usrc = src[1];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
467 uint8_t *vsrc = src[2];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
468 const int width = c->srcW;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
469 const int height = srcSliceH;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
470 const int lumStride = srcStride[0];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
471 const int chromStride = srcStride[1];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
472 const int dstStride = dstStride_a[0];
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
473 const int vertLumPerChroma = 2;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
474 const vector unsigned char yperm = vec_lvsl(0, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
475 register unsigned int y;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
476
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
477 /* this code assume:
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
478
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
479 1) dst is 16 bytes-aligned
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
480 2) dstStride is a multiple of 16
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
481 3) width is a multiple of 16
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
482 4) lum&chrom stride are multiple of 8
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
483 */
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
484
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
485 for(y=0; y<height; y++)
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
486 {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
487 int i;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
488 for (i = 0; i < width - 31; i+= 32) {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
489 const unsigned int j = i >> 1;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
490 vector unsigned char v_yA = vec_ld(i, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
491 vector unsigned char v_yB = vec_ld(i + 16, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
492 vector unsigned char v_yC = vec_ld(i + 32, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
493 vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
494 vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
495 vector unsigned char v_uA = vec_ld(j, usrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
496 vector unsigned char v_uB = vec_ld(j + 16, usrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
497 vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
498 vector unsigned char v_vA = vec_ld(j, vsrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
499 vector unsigned char v_vB = vec_ld(j + 16, vsrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
500 vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
501 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
502 vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
503 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
504 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
505 vector unsigned char v_uyvy_2 = vec_mergeh(v_uv_b, v_y2);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
506 vector unsigned char v_uyvy_3 = vec_mergel(v_uv_b, v_y2);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
507 vec_st(v_uyvy_0, (i << 1), dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
508 vec_st(v_uyvy_1, (i << 1) + 16, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
509 vec_st(v_uyvy_2, (i << 1) + 32, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
510 vec_st(v_uyvy_3, (i << 1) + 48, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
511 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
512 if (i < width) {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
513 const unsigned int j = i >> 1;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
514 vector unsigned char v_y1 = vec_ld(i, ysrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
515 vector unsigned char v_u = vec_ld(j, usrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
516 vector unsigned char v_v = vec_ld(j, vsrc);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
517 vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
518 vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
519 vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
520 vec_st(v_uyvy_0, (i << 1), dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
521 vec_st(v_uyvy_1, (i << 1) + 16, dst);
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
522 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
523 if((y&(vertLumPerChroma-1))==(vertLumPerChroma-1) )
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
524 {
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
525 usrc += chromStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
526 vsrc += chromStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
527 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
528 ysrc += lumStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
529 dst += dstStride;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
530 }
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
531 return srcSliceH;
931eee818c52 Altivec unscaled YV12 -> packed YUV patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents: 12532
diff changeset
532 }