Mercurial > libavcodec.hg
annotate ppc/mpegvideo_altivec.c @ 3963:075b3e6bdc05 libavcodec
Fixes:
rv10.c:494: warning: implicit declaration of function ¡Æff_mpeg4_init_direct_mv¡Ç
Approved by Michael.
author | rathann |
---|---|
date | Sun, 08 Oct 2006 21:26:19 +0000 |
parents | c8c591fe26f8 |
children | b28edd190fc0 |
rev | line source |
---|---|
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
1 /* |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
2 * Copyright (c) 2002 Dieter Shirley |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
3 * |
1942
9241d99f7443
forgotten copyright & dct_quantize_altivec disabled patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
4 * dct_unquantize_h263_altivec: |
9241d99f7443
forgotten copyright & dct_quantize_altivec disabled patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
5 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> |
9241d99f7443
forgotten copyright & dct_quantize_altivec disabled patch by (Romain Dolbeau <dolbeau at irisa dot fr>)
michael
parents:
1839
diff
changeset
|
6 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
7 * This file is part of FFmpeg. |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
8 * |
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
9 * FFmpeg is free software; you can redistribute it and/or |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
10 * modify it under the terms of the GNU Lesser General Public |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
11 * License as published by the Free Software Foundation; either |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
12 * version 2.1 of the License, or (at your option) any later version. |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
13 * |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
14 * FFmpeg is distributed in the hope that it will be useful, |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
17 * Lesser General Public License for more details. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
18 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
19 * You should have received a copy of the GNU Lesser General Public |
3947
c8c591fe26f8
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
3036
diff
changeset
|
20 * License along with FFmpeg; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
22 */ |
1277
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1107
diff
changeset
|
23 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
24 #include <stdlib.h> |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
25 #include <stdio.h> |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
26 #include "../dsputil.h" |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
27 #include "../mpegvideo.h" |
1277
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1107
diff
changeset
|
28 |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1107
diff
changeset
|
29 #include "gcc_fixes.h" |
2967 | 30 |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
31 #include "dsputil_altivec.h" |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
32 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
33 // Swaps two variables (used for altivec registers) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
34 #define SWAP(a,b) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
35 do { \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
36 __typeof__(a) swap_temp=a; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
37 a=b; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
38 b=swap_temp; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
39 } while (0) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
40 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
41 // transposes a matrix consisting of four vectors with four elements each |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
42 #define TRANSPOSE4(a,b,c,d) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
43 do { \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
44 __typeof__(a) _trans_ach = vec_mergeh(a, c); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
45 __typeof__(a) _trans_acl = vec_mergel(a, c); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
46 __typeof__(a) _trans_bdh = vec_mergeh(b, d); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
47 __typeof__(a) _trans_bdl = vec_mergel(b, d); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
48 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
49 a = vec_mergeh(_trans_ach, _trans_bdh); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
50 b = vec_mergel(_trans_ach, _trans_bdh); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
51 c = vec_mergeh(_trans_acl, _trans_bdl); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
52 d = vec_mergel(_trans_acl, _trans_bdl); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
53 } while (0) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
54 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
55 #define TRANSPOSE8(a,b,c,d,e,f,g,h) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
56 do { \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
57 __typeof__(a) _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
58 __typeof__(a) _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
59 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
60 _A1 = vec_mergeh (a, e); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
61 _B1 = vec_mergel (a, e); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
62 _C1 = vec_mergeh (b, f); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
63 _D1 = vec_mergel (b, f); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
64 _E1 = vec_mergeh (c, g); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
65 _F1 = vec_mergel (c, g); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
66 _G1 = vec_mergeh (d, h); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
67 _H1 = vec_mergel (d, h); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
68 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
69 _A2 = vec_mergeh (_A1, _E1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
70 _B2 = vec_mergel (_A1, _E1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
71 _C2 = vec_mergeh (_B1, _F1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
72 _D2 = vec_mergel (_B1, _F1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
73 _E2 = vec_mergeh (_C1, _G1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
74 _F2 = vec_mergel (_C1, _G1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
75 _G2 = vec_mergeh (_D1, _H1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
76 _H2 = vec_mergel (_D1, _H1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
77 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
78 a = vec_mergeh (_A2, _E2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
79 b = vec_mergel (_A2, _E2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
80 c = vec_mergeh (_B2, _F2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
81 d = vec_mergel (_B2, _F2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
82 e = vec_mergeh (_C2, _G2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
83 f = vec_mergel (_C2, _G2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
84 g = vec_mergeh (_D2, _H2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
85 h = vec_mergel (_D2, _H2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
86 } while (0) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
87 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
88 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
89 // Loads a four-byte value (int or float) from the target address |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
90 // into every element in the target vector. Only works if the |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
91 // target address is four-byte aligned (which should be always). |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
92 #define LOAD4(vec, address) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
93 { \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
94 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
95 vector unsigned char _perm_vec = vec_lvsl(0,(address)); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
96 vec = vec_ld(0, _load_addr); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
97 vec = vec_perm(vec, vec, _perm_vec); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
98 vec = vec_splat(vec, 0); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
99 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
100 |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
101 |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
102 #ifdef CONFIG_DARWIN |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
103 #define FOUROF(a) (a) |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
104 #else |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
105 // slower, for dumb non-apple GCC |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
106 #define FOUROF(a) {a,a,a,a} |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
107 #endif |
2967 | 108 int dct_quantize_altivec(MpegEncContext* s, |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
109 DCTELEM* data, int n, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
110 int qscale, int* overflow) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
111 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
112 int lastNonZero; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
113 vector float row0, row1, row2, row3, row4, row5, row6, row7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
114 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
115 const_vector float zero = (const_vector float)FOUROF(0.); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
116 // used after quantise step |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
117 int oldBaseValue = 0; |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
118 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
119 // Load the data into the row/alt vectors |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
120 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
121 vector signed short data0, data1, data2, data3, data4, data5, data6, data7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
122 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
123 data0 = vec_ld(0, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
124 data1 = vec_ld(16, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
125 data2 = vec_ld(32, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
126 data3 = vec_ld(48, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
127 data4 = vec_ld(64, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
128 data5 = vec_ld(80, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
129 data6 = vec_ld(96, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
130 data7 = vec_ld(112, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
131 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
132 // Transpose the data before we start |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
133 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
134 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
135 // load the data into floating point vectors. We load |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
136 // the high half of each row into the main row vectors |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
137 // and the low half into the alt vectors. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
138 row0 = vec_ctf(vec_unpackh(data0), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
139 alt0 = vec_ctf(vec_unpackl(data0), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
140 row1 = vec_ctf(vec_unpackh(data1), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
141 alt1 = vec_ctf(vec_unpackl(data1), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
142 row2 = vec_ctf(vec_unpackh(data2), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
143 alt2 = vec_ctf(vec_unpackl(data2), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
144 row3 = vec_ctf(vec_unpackh(data3), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
145 alt3 = vec_ctf(vec_unpackl(data3), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
146 row4 = vec_ctf(vec_unpackh(data4), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
147 alt4 = vec_ctf(vec_unpackl(data4), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
148 row5 = vec_ctf(vec_unpackh(data5), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
149 alt5 = vec_ctf(vec_unpackl(data5), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
150 row6 = vec_ctf(vec_unpackh(data6), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
151 alt6 = vec_ctf(vec_unpackl(data6), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
152 row7 = vec_ctf(vec_unpackh(data7), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
153 alt7 = vec_ctf(vec_unpackl(data7), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
154 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
155 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
156 // The following block could exist as a separate an altivec dct |
2979 | 157 // function. However, if we put it inline, the DCT data can remain |
158 // in the vector local variables, as floats, which we'll use during the | |
159 // quantize step... | |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
160 { |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
161 const vector float vec_0_298631336 = (vector float)FOUROF(0.298631336f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
162 const vector float vec_0_390180644 = (vector float)FOUROF(-0.390180644f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
163 const vector float vec_0_541196100 = (vector float)FOUROF(0.541196100f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
164 const vector float vec_0_765366865 = (vector float)FOUROF(0.765366865f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
165 const vector float vec_0_899976223 = (vector float)FOUROF(-0.899976223f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
166 const vector float vec_1_175875602 = (vector float)FOUROF(1.175875602f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
167 const vector float vec_1_501321110 = (vector float)FOUROF(1.501321110f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
168 const vector float vec_1_847759065 = (vector float)FOUROF(-1.847759065f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
169 const vector float vec_1_961570560 = (vector float)FOUROF(-1.961570560f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
170 const vector float vec_2_053119869 = (vector float)FOUROF(2.053119869f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
171 const vector float vec_2_562915447 = (vector float)FOUROF(-2.562915447f); |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
172 const vector float vec_3_072711026 = (vector float)FOUROF(3.072711026f); |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
173 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
174 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
175 int whichPass, whichHalf; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
176 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
177 for(whichPass = 1; whichPass<=2; whichPass++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
178 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
179 for(whichHalf = 1; whichHalf<=2; whichHalf++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
180 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
181 vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
182 vector float tmp10, tmp11, tmp12, tmp13; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
183 vector float z1, z2, z3, z4, z5; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
184 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
185 tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
186 tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
187 tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
188 tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
189 tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
190 tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
191 tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
192 tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
193 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
194 tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
195 tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
196 tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
197 tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
198 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
199 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
200 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
201 row0 = vec_add(tmp10, tmp11); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
202 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
203 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
204 row4 = vec_sub(tmp10, tmp11); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
205 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
206 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
207 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
208 z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
209 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
210 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), |
2979 | 211 // CONST_BITS-PASS1_BITS); |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
212 row2 = vec_madd(tmp13, vec_0_765366865, z1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
213 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
214 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), |
2979 | 215 // CONST_BITS-PASS1_BITS); |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
216 row6 = vec_madd(tmp12, vec_1_847759065, z1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
217 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
218 z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
219 z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
220 z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
221 z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
222 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
223 // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
224 z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
225 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
226 // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
227 z3 = vec_madd(z3, vec_1_961570560, z5); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
228 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
229 // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
230 z4 = vec_madd(z4, vec_0_390180644, z5); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
231 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
232 // The following adds are rolled into the multiplies above |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
233 // z3 = vec_add(z3, z5); // z3 += z5; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
234 // z4 = vec_add(z4, z5); // z4 += z5; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
235 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
236 // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
237 // Wow! It's actually more effecient to roll this multiply |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
238 // into the adds below, even thought the multiply gets done twice! |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
239 // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
240 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
241 // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
242 // Same with this one... |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
243 // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
244 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
245 // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
246 // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
247 row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
248 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
249 // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
250 // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
251 row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
252 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
253 // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
254 // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
255 row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
256 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
257 // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
258 // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
259 row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
260 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
261 // Swap the row values with the alts. If this is the first half, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
262 // this sets up the low values to be acted on in the second half. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
263 // If this is the second half, it puts the high values back in |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
264 // the row values where they are expected to be when we're done. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
265 SWAP(row0, alt0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
266 SWAP(row1, alt1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
267 SWAP(row2, alt2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
268 SWAP(row3, alt3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
269 SWAP(row4, alt4); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
270 SWAP(row5, alt5); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
271 SWAP(row6, alt6); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
272 SWAP(row7, alt7); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
273 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
274 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
275 if (whichPass == 1) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
276 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
277 // transpose the data for the second pass |
2967 | 278 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
279 // First, block transpose the upper right with lower left. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
280 SWAP(row4, alt0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
281 SWAP(row5, alt1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
282 SWAP(row6, alt2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
283 SWAP(row7, alt3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
284 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
285 // Now, transpose each block of four |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
286 TRANSPOSE4(row0, row1, row2, row3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
287 TRANSPOSE4(row4, row5, row6, row7); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
288 TRANSPOSE4(alt0, alt1, alt2, alt3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
289 TRANSPOSE4(alt4, alt5, alt6, alt7); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
290 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
291 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
292 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
293 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
294 // perform the quantise step, using the floating point data |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
295 // still in the row/alt registers |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
296 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
297 const int* biasAddr; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
298 const vector signed int* qmat; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
299 vector float bias, negBias; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
300 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
301 if (s->mb_intra) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
302 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
303 vector signed int baseVector; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
304 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
305 // We must cache element 0 in the intra case |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
306 // (it needs special handling). |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
307 baseVector = vec_cts(vec_splat(row0, 0), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
308 vec_ste(baseVector, 0, &oldBaseValue); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
309 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
310 qmat = (vector signed int*)s->q_intra_matrix[qscale]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
311 biasAddr = &(s->intra_quant_bias); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
312 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
313 else |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
314 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
315 qmat = (vector signed int*)s->q_inter_matrix[qscale]; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
316 biasAddr = &(s->inter_quant_bias); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
317 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
318 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
319 // Load the bias vector (We add 0.5 to the bias so that we're |
2979 | 320 // rounding when we convert to int, instead of flooring.) |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
321 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
322 vector signed int biasInt; |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
323 const vector float negOneFloat = (vector float)FOUROF(-1.0f); |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
324 LOAD4(biasInt, biasAddr); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
325 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
326 negBias = vec_madd(bias, negOneFloat, zero); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
327 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
328 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
329 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
330 vector float q0, q1, q2, q3, q4, q5, q6, q7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
331 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
332 q0 = vec_ctf(qmat[0], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
333 q1 = vec_ctf(qmat[2], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
334 q2 = vec_ctf(qmat[4], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
335 q3 = vec_ctf(qmat[6], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
336 q4 = vec_ctf(qmat[8], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
337 q5 = vec_ctf(qmat[10], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
338 q6 = vec_ctf(qmat[12], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
339 q7 = vec_ctf(qmat[14], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
340 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
341 row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
342 vec_cmpgt(row0, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
343 row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
344 vec_cmpgt(row1, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
345 row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
346 vec_cmpgt(row2, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
347 row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
348 vec_cmpgt(row3, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
349 row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
350 vec_cmpgt(row4, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
351 row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
352 vec_cmpgt(row5, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
353 row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
354 vec_cmpgt(row6, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
355 row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
356 vec_cmpgt(row7, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
357 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
358 q0 = vec_ctf(qmat[1], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
359 q1 = vec_ctf(qmat[3], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
360 q2 = vec_ctf(qmat[5], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
361 q3 = vec_ctf(qmat[7], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
362 q4 = vec_ctf(qmat[9], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
363 q5 = vec_ctf(qmat[11], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
364 q6 = vec_ctf(qmat[13], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
365 q7 = vec_ctf(qmat[15], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
366 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
367 alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
368 vec_cmpgt(alt0, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
369 alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
370 vec_cmpgt(alt1, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
371 alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
372 vec_cmpgt(alt2, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
373 alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
374 vec_cmpgt(alt3, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
375 alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
376 vec_cmpgt(alt4, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
377 alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
378 vec_cmpgt(alt5, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
379 alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
380 vec_cmpgt(alt6, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
381 alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
382 vec_cmpgt(alt7, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
383 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
384 |
2967 | 385 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
386 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
387 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
388 // Store the data back into the original block |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
389 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
390 vector signed short data0, data1, data2, data3, data4, data5, data6, data7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
391 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
392 data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
393 data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
394 data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
395 data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
396 data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
397 data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
398 data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
399 data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
400 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
401 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
402 // Clamp for overflow |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
403 vector signed int max_q_int, min_q_int; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
404 vector signed short max_q, min_q; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
405 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
406 LOAD4(max_q_int, &(s->max_qcoeff)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
407 LOAD4(min_q_int, &(s->min_qcoeff)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
408 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
409 max_q = vec_pack(max_q_int, max_q_int); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
410 min_q = vec_pack(min_q_int, min_q_int); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
411 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
412 data0 = vec_max(vec_min(data0, max_q), min_q); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
413 data1 = vec_max(vec_min(data1, max_q), min_q); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
414 data2 = vec_max(vec_min(data2, max_q), min_q); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
415 data4 = vec_max(vec_min(data4, max_q), min_q); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
416 data5 = vec_max(vec_min(data5, max_q), min_q); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
417 data6 = vec_max(vec_min(data6, max_q), min_q); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
418 data7 = vec_max(vec_min(data7, max_q), min_q); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
419 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
420 |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
421 { |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
422 vector bool char zero_01, zero_23, zero_45, zero_67; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
423 vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
424 vector signed char negOne = vec_splat_s8(-1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
425 vector signed char* scanPtr = |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
426 (vector signed char*)(s->intra_scantable.inverse); |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
427 signed char lastNonZeroChar; |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
428 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
429 // Determine the largest non-zero index. |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
430 zero_01 = vec_pack(vec_cmpeq(data0, (vector signed short)zero), |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
431 vec_cmpeq(data1, (vector signed short)zero)); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
432 zero_23 = vec_pack(vec_cmpeq(data2, (vector signed short)zero), |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
433 vec_cmpeq(data3, (vector signed short)zero)); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
434 zero_45 = vec_pack(vec_cmpeq(data4, (vector signed short)zero), |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
435 vec_cmpeq(data5, (vector signed short)zero)); |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
436 zero_67 = vec_pack(vec_cmpeq(data6, (vector signed short)zero), |
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
437 vec_cmpeq(data7, (vector signed short)zero)); |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
438 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
439 // 64 biggest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
440 scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
441 scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
442 scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
443 scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
444 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
445 // 32 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
446 scanIndices_01 = vec_max(scanIndices_01, scanIndices_23); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
447 scanIndices_45 = vec_max(scanIndices_45, scanIndices_67); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
448 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
449 // 16 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
450 scanIndices_01 = vec_max(scanIndices_01, scanIndices_45); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
451 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
452 // 8 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
453 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
454 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
455 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
456 // 4 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
457 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
458 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
459 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
460 // 2 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
461 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
462 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
463 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
464 // largest value |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
465 scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
466 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
467 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
468 scanIndices_01 = vec_splat(scanIndices_01, 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
469 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
470 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
471 vec_ste(scanIndices_01, 0, &lastNonZeroChar); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
472 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
473 lastNonZero = lastNonZeroChar; |
2967 | 474 |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
475 // While the data is still in vectors we check for the transpose IDCT permute |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
476 // and handle it using the vector unit if we can. This is the permute used |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
477 // by the altivec idct, so it is common when using the altivec dct. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
478 |
1092 | 479 if ((lastNonZero > 0) && (s->dsp.idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
480 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
481 TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
482 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
483 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
484 vec_st(data0, 0, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
485 vec_st(data1, 16, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
486 vec_st(data2, 32, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
487 vec_st(data3, 48, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
488 vec_st(data4, 64, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
489 vec_st(data5, 80, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
490 vec_st(data6, 96, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
491 vec_st(data7, 112, data); |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
492 } |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
493 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
494 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
495 // special handling of block[0] |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
496 if (s->mb_intra) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
497 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
498 if (!s->h263_aic) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
499 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
500 if (n < 4) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
501 oldBaseValue /= s->y_dc_scale; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
502 else |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
503 oldBaseValue /= s->c_dc_scale; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
504 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
505 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
506 // Divide by 8, rounding the result |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
507 data[0] = (oldBaseValue + 4) >> 3; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
508 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
509 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
510 // We handled the tranpose permutation above and we don't |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
511 // need to permute the "no" permutation case. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
512 if ((lastNonZero > 0) && |
1092 | 513 (s->dsp.idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) && |
514 (s->dsp.idct_permutation_type != FF_NO_IDCT_PERM)) | |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
515 { |
1107
28f0e53706a7
altivec compilation fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1092
diff
changeset
|
516 ff_block_permute(data, s->dsp.idct_permutation, |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
517 s->intra_scantable.scantable, lastNonZero); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
518 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
519 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
520 return lastNonZero; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
521 } |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
522 #undef FOUROF |
828
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
523 |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
524 /* |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
525 AltiVec version of dct_unquantize_h263 |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
526 this code assumes `block' is 16 bytes-aligned |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
527 */ |
2967 | 528 void dct_unquantize_h263_altivec(MpegEncContext *s, |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
529 DCTELEM *block, int n, int qscale) |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
530 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
531 POWERPC_PERF_DECLARE(altivec_dct_unquantize_h263_num, 1); |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
532 int i, level, qmul, qadd; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
533 int nCoeffs; |
2967 | 534 |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
535 assert(s->block_last_index[n]>=0); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
536 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
537 POWERPC_PERF_START_COUNT(altivec_dct_unquantize_h263_num, 1); |
2967 | 538 |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
539 qadd = (qscale - 1) | 1; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
540 qmul = qscale << 1; |
2967 | 541 |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
542 if (s->mb_intra) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
543 if (!s->h263_aic) { |
2967 | 544 if (n < 4) |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
545 block[0] = block[0] * s->y_dc_scale; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
546 else |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
547 block[0] = block[0] * s->c_dc_scale; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
548 }else |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
549 qadd = 0; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
550 i = 1; |
2967 | 551 nCoeffs= 63; //does not allways use zigzag table |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
552 } else { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
553 i = 0; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
554 nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ]; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
555 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
556 |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
557 #ifdef ALTIVEC_USE_REFERENCE_C_CODE |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
558 for(;i<=nCoeffs;i++) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
559 level = block[i]; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
560 if (level) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
561 if (level < 0) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
562 level = level * qmul - qadd; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
563 } else { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
564 level = level * qmul + qadd; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
565 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
566 block[i] = level; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
567 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
568 } |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
569 #else /* ALTIVEC_USE_REFERENCE_C_CODE */ |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
570 { |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
571 register const_vector signed short vczero = (const_vector signed short)vec_splat_s16(0); |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
572 short __attribute__ ((aligned(16))) qmul8[] = |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
573 { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
574 qmul, qmul, qmul, qmul, |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
575 qmul, qmul, qmul, qmul |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
576 }; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
577 short __attribute__ ((aligned(16))) qadd8[] = |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
578 { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
579 qadd, qadd, qadd, qadd, |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
580 qadd, qadd, qadd, qadd |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
581 }; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
582 short __attribute__ ((aligned(16))) nqadd8[] = |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
583 { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
584 -qadd, -qadd, -qadd, -qadd, |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
585 -qadd, -qadd, -qadd, -qadd |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
586 }; |
1839
b370288f004d
Metrowerks CodeWarrior patches by (John Dalgliesh <johnd at defyne dot org>)
michael
parents:
1352
diff
changeset
|
587 register vector signed short blockv, qmulv, qaddv, nqaddv, temp1; |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
588 register vector bool short blockv_null, blockv_neg; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
589 register short backup_0 = block[0]; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
590 register int j = 0; |
2967 | 591 |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
592 qmulv = vec_ld(0, qmul8); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
593 qaddv = vec_ld(0, qadd8); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
594 nqaddv = vec_ld(0, nqadd8); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
595 |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
596 #if 0 // block *is* 16 bytes-aligned, it seems. |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
597 // first make sure block[j] is 16 bytes-aligned |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
598 for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
599 level = block[j]; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
600 if (level) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
601 if (level < 0) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
602 level = level * qmul - qadd; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
603 } else { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
604 level = level * qmul + qadd; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
605 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
606 block[j] = level; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
607 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
608 } |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
609 #endif |
2967 | 610 |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
611 // vectorize all the 16 bytes-aligned blocks |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
612 // of 8 elements |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
613 for(; (j + 7) <= nCoeffs ; j+=8) |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
614 { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
615 blockv = vec_ld(j << 1, block); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
616 blockv_neg = vec_cmplt(blockv, vczero); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
617 blockv_null = vec_cmpeq(blockv, vczero); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
618 // choose between +qadd or -qadd as the third operand |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
619 temp1 = vec_sel(qaddv, nqaddv, blockv_neg); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
620 // multiply & add (block{i,i+7} * qmul [+-] qadd) |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
621 temp1 = vec_mladd(blockv, qmulv, temp1); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
622 // put 0 where block[{i,i+7} used to have 0 |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
623 blockv = vec_sel(temp1, blockv, blockv_null); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
624 vec_st(blockv, j << 1, block); |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
625 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
626 |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
627 // if nCoeffs isn't a multiple of 8, finish the job |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
628 // using good old scalar units. |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
629 // (we could do it using a truncated vector, |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
630 // but I'm not sure it's worth the hassle) |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
631 for(; j <= nCoeffs ; j++) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
632 level = block[j]; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
633 if (level) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
634 if (level < 0) { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
635 level = level * qmul - qadd; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
636 } else { |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
637 level = level * qmul + qadd; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
638 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
639 block[j] = level; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
640 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
641 } |
2967 | 642 |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
643 if (i == 1) |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
644 { // cheat. this avoid special-casing the first iteration |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
645 block[0] = backup_0; |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
646 } |
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
647 } |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
648 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
1001
diff
changeset
|
649 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
650 POWERPC_PERF_STOP_COUNT(altivec_dct_unquantize_h263_num, nCoeffs == 63); |
1001
95cbffdc98a9
dct_unquantize_h263_altivec by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
995
diff
changeset
|
651 } |