Mercurial > libavcodec.hg
annotate ppc/mpegvideo_altivec.c @ 943:0566d1a8426f libavcodec
10l (int i)
author | michael |
---|---|
date | Mon, 30 Dec 2002 12:36:28 +0000 |
parents | ace3ccd18dd2 |
children | 8bec850dc9c7 |
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 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
4 * This library is free software; you can redistribute it and/or |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
5 * 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
|
6 * License as published by the Free Software Foundation; either |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
7 * version 2 of the License, or (at your option) any later version. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
8 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
9 * This library is distributed in the hope that it will be useful, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
10 * 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
|
11 * 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
|
12 * Lesser General Public License for more details. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
13 * |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
14 * You should have received a copy of the GNU Lesser General Public |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
15 * License along with this library; if not, write to the Free Software |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
17 */ |
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 #include <stdlib.h> |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
20 #include <stdio.h> |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
21 #include "../dsputil.h" |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
22 #include "../mpegvideo.h" |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
23 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
24 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
25 // Used when initializing constant vectors |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
26 #define FOUR_INSTANCES(x) x,x,x,x |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
27 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
28 // Swaps two variables (used for altivec registers) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
29 #define SWAP(a,b) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
30 do { \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
31 __typeof__(a) swap_temp=a; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
32 a=b; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
33 b=swap_temp; \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
34 } while (0) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
35 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
36 // 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
|
37 #define TRANSPOSE4(a,b,c,d) \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
38 do { \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
39 __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
|
40 __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
|
41 __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
|
42 __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
|
43 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
44 a = vec_mergeh(_trans_ach, _trans_bdh); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
45 b = vec_mergel(_trans_ach, _trans_bdh); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
46 c = vec_mergeh(_trans_acl, _trans_bdl); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
47 d = vec_mergel(_trans_acl, _trans_bdl); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
48 } while (0) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
49 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
50 #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
|
51 do { \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
52 __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
|
53 __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
|
54 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
55 _A1 = vec_mergeh (a, e); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
56 _B1 = vec_mergel (a, e); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
57 _C1 = vec_mergeh (b, f); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
58 _D1 = vec_mergel (b, f); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
59 _E1 = vec_mergeh (c, g); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
60 _F1 = vec_mergel (c, g); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
61 _G1 = vec_mergeh (d, h); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
62 _H1 = vec_mergel (d, h); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
63 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
64 _A2 = vec_mergeh (_A1, _E1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
65 _B2 = vec_mergel (_A1, _E1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
66 _C2 = vec_mergeh (_B1, _F1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
67 _D2 = vec_mergel (_B1, _F1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
68 _E2 = vec_mergeh (_C1, _G1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
69 _F2 = vec_mergel (_C1, _G1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
70 _G2 = vec_mergeh (_D1, _H1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
71 _H2 = vec_mergel (_D1, _H1); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
72 \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
73 a = vec_mergeh (_A2, _E2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
74 b = vec_mergel (_A2, _E2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
75 c = vec_mergeh (_B2, _F2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
76 d = vec_mergel (_B2, _F2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
77 e = vec_mergeh (_C2, _G2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
78 f = vec_mergel (_C2, _G2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
79 g = vec_mergeh (_D2, _H2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
80 h = vec_mergel (_D2, _H2); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
81 } while (0) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
82 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
83 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
84 // 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
|
85 // 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
|
86 // 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
|
87 #define LOAD4(vec, address) \ |
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 __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
90 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
|
91 vec = vec_ld(0, _load_addr); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
92 vec = vec_perm(vec, vec, _perm_vec); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
93 vec = vec_splat(vec, 0); \ |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
94 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
95 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
96 int dct_quantize_altivec(MpegEncContext* s, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
97 DCTELEM* data, int n, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
98 int qscale, int* overflow) |
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 int lastNonZero; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
101 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
|
102 vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
103 const vector float zero = {FOUR_INSTANCES(0.0f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
104 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
105 // 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
|
106 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
107 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
|
108 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
109 data0 = vec_ld(0, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
110 data1 = vec_ld(16, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
111 data2 = vec_ld(32, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
112 data3 = vec_ld(48, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
113 data4 = vec_ld(64, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
114 data5 = vec_ld(80, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
115 data6 = vec_ld(96, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
116 data7 = vec_ld(112, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
117 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
118 // Transpose the data before we start |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
119 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
|
120 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
121 // 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
|
122 // 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
|
123 // 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
|
124 row0 = vec_ctf(vec_unpackh(data0), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
125 alt0 = vec_ctf(vec_unpackl(data0), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
126 row1 = vec_ctf(vec_unpackh(data1), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
127 alt1 = vec_ctf(vec_unpackl(data1), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
128 row2 = vec_ctf(vec_unpackh(data2), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
129 alt2 = vec_ctf(vec_unpackl(data2), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
130 row3 = vec_ctf(vec_unpackh(data3), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
131 alt3 = vec_ctf(vec_unpackl(data3), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
132 row4 = vec_ctf(vec_unpackh(data4), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
133 alt4 = vec_ctf(vec_unpackl(data4), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
134 row5 = vec_ctf(vec_unpackh(data5), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
135 alt5 = vec_ctf(vec_unpackl(data5), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
136 row6 = vec_ctf(vec_unpackh(data6), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
137 alt6 = vec_ctf(vec_unpackl(data6), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
138 row7 = vec_ctf(vec_unpackh(data7), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
139 alt7 = vec_ctf(vec_unpackl(data7), 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
140 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
141 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
142 // The following block could exist as a separate an altivec dct |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
143 // function. However, if we put it inline, the DCT data can remain |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
144 // in the vector local variables, as floats, which we'll use during the |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
145 // quantize step... |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
146 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
147 const vector float vec_0_298631336 = {FOUR_INSTANCES(0.298631336f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
148 const vector float vec_0_390180644 = {FOUR_INSTANCES(-0.390180644f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
149 const vector float vec_0_541196100 = {FOUR_INSTANCES(0.541196100f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
150 const vector float vec_0_765366865 = {FOUR_INSTANCES(0.765366865f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
151 const vector float vec_0_899976223 = {FOUR_INSTANCES(-0.899976223f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
152 const vector float vec_1_175875602 = {FOUR_INSTANCES(1.175875602f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
153 const vector float vec_1_501321110 = {FOUR_INSTANCES(1.501321110f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
154 const vector float vec_1_847759065 = {FOUR_INSTANCES(-1.847759065f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
155 const vector float vec_1_961570560 = {FOUR_INSTANCES(-1.961570560f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
156 const vector float vec_2_053119869 = {FOUR_INSTANCES(2.053119869f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
157 const vector float vec_2_562915447 = {FOUR_INSTANCES(-2.562915447f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
158 const vector float vec_3_072711026 = {FOUR_INSTANCES(3.072711026f)}; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
159 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
160 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
161 int whichPass, whichHalf; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
162 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
163 for(whichPass = 1; whichPass<=2; whichPass++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
164 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
165 for(whichHalf = 1; whichHalf<=2; whichHalf++) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
166 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
167 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
|
168 vector float tmp10, tmp11, tmp12, tmp13; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
169 vector float z1, z2, z3, z4, z5; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
170 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
171 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
|
172 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
|
173 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
|
174 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
|
175 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
|
176 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
|
177 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
|
178 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
|
179 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
180 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
|
181 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
|
182 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
|
183 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
|
184 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
185 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
186 // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
187 row0 = vec_add(tmp10, tmp11); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
188 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
189 // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
190 row4 = vec_sub(tmp10, tmp11); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
191 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
192 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
193 // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
194 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
|
195 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
196 // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
197 // CONST_BITS-PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
198 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
|
199 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
200 // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
201 // CONST_BITS-PASS1_BITS); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
202 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
|
203 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
204 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
|
205 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
|
206 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
|
207 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
|
208 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
209 // 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
|
210 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
|
211 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
212 // 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
|
213 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
|
214 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
215 // 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
|
216 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
|
217 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
218 // 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
|
219 // z3 = vec_add(z3, z5); // z3 += z5; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
220 // z4 = vec_add(z4, z5); // z4 += z5; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
221 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
222 // 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
|
223 // 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
|
224 // 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
|
225 // 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
|
226 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
227 // 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
|
228 // Same with this one... |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
229 // 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
|
230 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
231 // 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
|
232 // 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
|
233 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
|
234 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
235 // 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
|
236 // 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
|
237 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
|
238 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
239 // 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
|
240 // 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
|
241 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
|
242 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
243 // 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
|
244 // 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
|
245 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
|
246 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
247 // 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
|
248 // 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
|
249 // 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
|
250 // 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
|
251 SWAP(row0, alt0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
252 SWAP(row1, alt1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
253 SWAP(row2, alt2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
254 SWAP(row3, alt3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
255 SWAP(row4, alt4); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
256 SWAP(row5, alt5); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
257 SWAP(row6, alt6); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
258 SWAP(row7, alt7); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
259 } |
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 if (whichPass == 1) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
262 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
263 // transpose the data for the second pass |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
264 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
265 // 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
|
266 SWAP(row4, alt0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
267 SWAP(row5, alt1); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
268 SWAP(row6, alt2); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
269 SWAP(row7, alt3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
270 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
271 // Now, transpose each block of four |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
272 TRANSPOSE4(row0, row1, row2, row3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
273 TRANSPOSE4(row4, row5, row6, row7); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
274 TRANSPOSE4(alt0, alt1, alt2, alt3); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
275 TRANSPOSE4(alt4, alt5, alt6, alt7); |
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 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
278 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
279 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
280 // used after quantise step |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
281 int oldBaseValue = 0; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
282 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
283 // 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
|
284 // still in the row/alt registers |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
285 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
286 const int* biasAddr; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
287 const vector signed int* qmat; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
288 vector float bias, negBias; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
289 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
290 if (s->mb_intra) |
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 vector signed int baseVector; |
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 // 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
|
295 // (it needs special handling). |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
296 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
|
297 vec_ste(baseVector, 0, &oldBaseValue); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
298 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
299 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
|
300 biasAddr = &(s->intra_quant_bias); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
301 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
302 else |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
303 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
304 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
|
305 biasAddr = &(s->inter_quant_bias); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
306 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
307 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
308 // Load the bias vector (We add 0.5 to the bias so that we're |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
309 // rounding when we convert to int, instead of flooring.) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
310 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
311 vector signed int biasInt; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
312 const vector float negOneFloat = (vector float)(FOUR_INSTANCES(-1.0f)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
313 LOAD4(biasInt, biasAddr); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
314 bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
315 negBias = vec_madd(bias, negOneFloat, zero); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
316 } |
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 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
|
320 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
321 q0 = vec_ctf(qmat[0], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
322 q1 = vec_ctf(qmat[2], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
323 q2 = vec_ctf(qmat[4], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
324 q3 = vec_ctf(qmat[6], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
325 q4 = vec_ctf(qmat[8], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
326 q5 = vec_ctf(qmat[10], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
327 q6 = vec_ctf(qmat[12], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
328 q7 = vec_ctf(qmat[14], QMAT_SHIFT); |
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 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
|
331 vec_cmpgt(row0, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
332 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
|
333 vec_cmpgt(row1, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
334 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
|
335 vec_cmpgt(row2, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
336 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
|
337 vec_cmpgt(row3, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
338 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
|
339 vec_cmpgt(row4, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
340 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
|
341 vec_cmpgt(row5, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
342 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
|
343 vec_cmpgt(row6, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
344 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
|
345 vec_cmpgt(row7, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
346 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
347 q0 = vec_ctf(qmat[1], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
348 q1 = vec_ctf(qmat[3], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
349 q2 = vec_ctf(qmat[5], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
350 q3 = vec_ctf(qmat[7], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
351 q4 = vec_ctf(qmat[9], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
352 q5 = vec_ctf(qmat[11], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
353 q6 = vec_ctf(qmat[13], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
354 q7 = vec_ctf(qmat[15], QMAT_SHIFT); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
355 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
356 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
|
357 vec_cmpgt(alt0, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
358 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
|
359 vec_cmpgt(alt1, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
360 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
|
361 vec_cmpgt(alt2, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
362 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
|
363 vec_cmpgt(alt3, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
364 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
|
365 vec_cmpgt(alt4, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
366 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
|
367 vec_cmpgt(alt5, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
368 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
|
369 vec_cmpgt(alt6, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
370 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
|
371 vec_cmpgt(alt7, zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
372 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
373 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
374 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
375 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
376 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
377 // 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
|
378 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
379 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
|
380 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
381 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
|
382 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
|
383 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
|
384 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
|
385 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
|
386 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
|
387 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
|
388 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
|
389 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
390 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
391 // Clamp for overflow |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
392 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
|
393 vector signed short max_q, min_q; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
394 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
395 LOAD4(max_q_int, &(s->max_qcoeff)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
396 LOAD4(min_q_int, &(s->min_qcoeff)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
397 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
398 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
|
399 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
|
400 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
401 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
|
402 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
|
403 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
|
404 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
|
405 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
|
406 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
|
407 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
|
408 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
409 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
410 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
|
411 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
|
412 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
|
413 vector signed char* scanPtr = |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
414 (vector signed char*)(s->intra_scantable.inverse); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
415 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
416 // Determine the largest non-zero index. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
417 zero_01 = vec_pack(vec_cmpeq(data0, (vector short)zero), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
418 vec_cmpeq(data1, (vector short)zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
419 zero_23 = vec_pack(vec_cmpeq(data2, (vector short)zero), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
420 vec_cmpeq(data3, (vector short)zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
421 zero_45 = vec_pack(vec_cmpeq(data4, (vector short)zero), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
422 vec_cmpeq(data5, (vector short)zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
423 zero_67 = vec_pack(vec_cmpeq(data6, (vector short)zero), |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
424 vec_cmpeq(data7, (vector short)zero)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
425 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
426 // 64 biggest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
427 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
|
428 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
|
429 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
|
430 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
|
431 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
432 // 32 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
433 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
|
434 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
|
435 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
436 // 16 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
437 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
|
438 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
439 // 8 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
440 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
|
441 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
442 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
443 // 4 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
444 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
|
445 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
446 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
447 // 2 largest values |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
448 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
|
449 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
450 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
451 // largest value |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
452 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
|
453 vec_mergel(scanIndices_01, negOne)); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
454 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
455 scanIndices_01 = vec_splat(scanIndices_01, 0); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
456 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
457 signed char lastNonZeroChar; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
458 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
459 vec_ste(scanIndices_01, 0, &lastNonZeroChar); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
460 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
461 lastNonZero = lastNonZeroChar; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
462 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
463 // 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
|
464 // 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
|
465 // 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
|
466 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
467 if ((lastNonZero > 0) && (s->idct_permutation_type == FF_TRANSPOSE_IDCT_PERM)) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
468 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
469 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
|
470 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
471 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
472 vec_st(data0, 0, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
473 vec_st(data1, 16, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
474 vec_st(data2, 32, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
475 vec_st(data3, 48, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
476 vec_st(data4, 64, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
477 vec_st(data5, 80, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
478 vec_st(data6, 96, data); |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
479 vec_st(data7, 112, data); |
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 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
482 // special handling of block[0] |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
483 if (s->mb_intra) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
484 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
485 if (!s->h263_aic) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
486 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
487 if (n < 4) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
488 oldBaseValue /= s->y_dc_scale; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
489 else |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
490 oldBaseValue /= s->c_dc_scale; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
491 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
492 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
493 // Divide by 8, rounding the result |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
494 data[0] = (oldBaseValue + 4) >> 3; |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
495 } |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
496 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
497 // 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
|
498 // need to permute the "no" permutation case. |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
499 if ((lastNonZero > 0) && |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
500 (s->idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) && |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
501 (s->idct_permutation_type != FF_NO_IDCT_PERM)) |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
502 { |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
503 ff_block_permute(data, s->idct_permutation, |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
504 s->intra_scantable.scantable, lastNonZero); |
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 |
ace3ccd18dd2
Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
michaelni
parents:
diff
changeset
|
507 return lastNonZero; |
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 |