Mercurial > libpostproc.hg
annotate postprocess_altivec_template.c @ 107:7cbdc99cc753 libpostproc
Do not ignore .depend files, they are no longer generated.
author | diego |
---|---|
date | Thu, 24 Apr 2008 10:17:55 +0000 |
parents | 883d75fc0712 |
children | 83d51d1fb580 |
rev | line source |
---|---|
0 | 1 /* |
22
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
2 * AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org> |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
3 * |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
4 * based on code by Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at) |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
5 * |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
6 * This file is part of FFmpeg. |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
7 * |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
8 * FFmpeg is free software; you can redistribute it and/or modify |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
9 * it under the terms of the GNU General Public License as published by |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
10 * the Free Software Foundation; either version 2 of the License, or |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
11 * (at your option) any later version. |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
12 * |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
13 * FFmpeg is distributed in the hope that it will be useful, |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
16 * GNU General Public License for more details. |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
17 * |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
18 * You should have received a copy of the GNU General Public License |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
19 * along with FFmpeg; if not, write to the Free Software |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
da3bfee1fa67
Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents:
0
diff
changeset
|
21 */ |
0 | 22 |
42 | 23 #include "avutil.h" |
0 | 24 |
25 #define ALTIVEC_TRANSPOSE_8x8_SHORT(src_a,src_b,src_c,src_d,src_e,src_f,src_g,src_h) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
26 do { \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
27 __typeof__(src_a) tempA1, tempB1, tempC1, tempD1; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
28 __typeof__(src_a) tempE1, tempF1, tempG1, tempH1; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
29 __typeof__(src_a) tempA2, tempB2, tempC2, tempD2; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
30 __typeof__(src_a) tempE2, tempF2, tempG2, tempH2; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
31 tempA1 = vec_mergeh (src_a, src_e); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
32 tempB1 = vec_mergel (src_a, src_e); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
33 tempC1 = vec_mergeh (src_b, src_f); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
34 tempD1 = vec_mergel (src_b, src_f); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
35 tempE1 = vec_mergeh (src_c, src_g); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
36 tempF1 = vec_mergel (src_c, src_g); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
37 tempG1 = vec_mergeh (src_d, src_h); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
38 tempH1 = vec_mergel (src_d, src_h); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
39 tempA2 = vec_mergeh (tempA1, tempE1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
40 tempB2 = vec_mergel (tempA1, tempE1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
41 tempC2 = vec_mergeh (tempB1, tempF1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
42 tempD2 = vec_mergel (tempB1, tempF1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
43 tempE2 = vec_mergeh (tempC1, tempG1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
44 tempF2 = vec_mergel (tempC1, tempG1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
45 tempG2 = vec_mergeh (tempD1, tempH1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
46 tempH2 = vec_mergel (tempD1, tempH1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
47 src_a = vec_mergeh (tempA2, tempE2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
48 src_b = vec_mergel (tempA2, tempE2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
49 src_c = vec_mergeh (tempB2, tempF2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
50 src_d = vec_mergel (tempB2, tempF2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
51 src_e = vec_mergeh (tempC2, tempG2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
52 src_f = vec_mergel (tempC2, tempG2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
53 src_g = vec_mergeh (tempD2, tempH2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
54 src_h = vec_mergel (tempD2, tempH2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
55 } while (0) |
0 | 56 |
57 | |
58 static inline int vertClassify_altivec(uint8_t src[], int stride, PPContext *c) { | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
59 /* |
0 | 60 this code makes no assumption on src or stride. |
61 One could remove the recomputation of the perm | |
62 vector by assuming (stride % 16) == 0, unfortunately | |
63 this is not always true. | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
64 */ |
99 | 65 DECLARE_ALIGNED(16, short, data[8]) = |
66 { | |
67 ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1, | |
68 data[0] * 2 + 1, | |
69 c->QP * 2, | |
70 c->QP * 4 | |
71 }; | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
72 int numEq; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
73 uint8_t *src2 = src; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
74 vector signed short v_dcOffset; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
75 vector signed short v2QP; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
76 vector unsigned short v4QP; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
77 vector unsigned short v_dcThreshold; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
78 const int properStride = (stride % 16); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
79 const int srcAlign = ((unsigned long)src2 % 16); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
80 const int two_vectors = ((srcAlign > 8) || properStride) ? 1 : 0; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
81 const vector signed int zero = vec_splat_s32(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
82 const vector signed short mask = vec_splat_s16(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
83 vector signed int v_numEq = vec_splat_s32(0); |
99 | 84 vector signed short v_data = vec_ld(0, data); |
85 vector signed short v_srcAss0, v_srcAss1, v_srcAss2, v_srcAss3, | |
86 v_srcAss4, v_srcAss5, v_srcAss6, v_srcAss7; | |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
87 //FIXME avoid this mess if possible |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
88 register int j0 = 0, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
89 j1 = stride, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
90 j2 = 2 * stride, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
91 j3 = 3 * stride, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
92 j4 = 4 * stride, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
93 j5 = 5 * stride, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
94 j6 = 6 * stride, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
95 j7 = 7 * stride; |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
96 vector unsigned char v_srcA0, v_srcA1, v_srcA2, v_srcA3, |
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
97 v_srcA4, v_srcA5, v_srcA6, v_srcA7; |
0 | 98 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
99 v_dcOffset = vec_splat(v_data, 0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
100 v_dcThreshold = (vector unsigned short)vec_splat(v_data, 1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
101 v2QP = vec_splat(v_data, 2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
102 v4QP = (vector unsigned short)vec_splat(v_data, 3); |
0 | 103 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
104 src2 += stride * 4; |
0 | 105 |
106 #define LOAD_LINE(i) \ | |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
107 { \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
108 vector unsigned char perm##i = vec_lvsl(j##i, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
109 vector unsigned char v_srcA2##i; \ |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
110 vector unsigned char v_srcA1##i = vec_ld(j##i, src2); \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
111 if (two_vectors) \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
112 v_srcA2##i = vec_ld(j##i + 16, src2); \ |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
113 v_srcA##i = \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
114 vec_perm(v_srcA1##i, v_srcA2##i, perm##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
115 v_srcAss##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
116 (vector signed short)vec_mergeh((vector signed char)zero, \ |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
117 (vector signed char)v_srcA##i); } |
0 | 118 |
119 #define LOAD_LINE_ALIGNED(i) \ | |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
120 v_srcA##i = vec_ld(j##i, src2); \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
121 v_srcAss##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
122 (vector signed short)vec_mergeh((vector signed char)zero, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
123 (vector signed char)v_srcA##i) |
0 | 124 |
88 | 125 /* Special-casing the aligned case is worthwhile, as all calls from |
126 * the (transposed) horizontable deblocks will be aligned, in addition | |
127 * to the naturally aligned vertical deblocks. */ | |
0 | 128 if (properStride && srcAlign) { |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
129 LOAD_LINE_ALIGNED(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
130 LOAD_LINE_ALIGNED(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
131 LOAD_LINE_ALIGNED(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
132 LOAD_LINE_ALIGNED(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
133 LOAD_LINE_ALIGNED(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
134 LOAD_LINE_ALIGNED(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
135 LOAD_LINE_ALIGNED(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
136 LOAD_LINE_ALIGNED(7); |
0 | 137 } else { |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
138 LOAD_LINE(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
139 LOAD_LINE(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
140 LOAD_LINE(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
141 LOAD_LINE(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
142 LOAD_LINE(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
143 LOAD_LINE(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
144 LOAD_LINE(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
145 LOAD_LINE(7); |
0 | 146 } |
147 #undef LOAD_LINE | |
148 #undef LOAD_LINE_ALIGNED | |
149 | |
150 #define ITER(i, j) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
151 const vector signed short v_diff##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
152 vec_sub(v_srcAss##i, v_srcAss##j); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
153 const vector signed short v_sum##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
154 vec_add(v_diff##i, v_dcOffset); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
155 const vector signed short v_comp##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
156 (vector signed short)vec_cmplt((vector unsigned short)v_sum##i, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
157 v_dcThreshold); \ |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
158 const vector signed short v_part##i = vec_and(mask, v_comp##i); |
0 | 159 |
103 | 160 { |
161 ITER(0, 1) | |
162 ITER(1, 2) | |
163 ITER(2, 3) | |
164 ITER(3, 4) | |
165 ITER(4, 5) | |
166 ITER(5, 6) | |
167 ITER(6, 7) | |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
168 |
103 | 169 v_numEq = vec_sum4s(v_part0, v_numEq); |
170 v_numEq = vec_sum4s(v_part1, v_numEq); | |
171 v_numEq = vec_sum4s(v_part2, v_numEq); | |
172 v_numEq = vec_sum4s(v_part3, v_numEq); | |
173 v_numEq = vec_sum4s(v_part4, v_numEq); | |
174 v_numEq = vec_sum4s(v_part5, v_numEq); | |
175 v_numEq = vec_sum4s(v_part6, v_numEq); | |
176 } | |
100
b944f0b99b23
Remove declarations after statements from vertClassify_altivec
lu_zero
parents:
99
diff
changeset
|
177 |
0 | 178 #undef ITER |
179 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
180 v_numEq = vec_sums(v_numEq, zero); |
0 | 181 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
182 v_numEq = vec_splat(v_numEq, 3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
183 vec_ste(v_numEq, 0, &numEq); |
0 | 184 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
185 if (numEq > c->ppMode.flatnessThreshold){ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
186 const vector unsigned char mmoP1 = (const vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
187 AVV(0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
188 0x00, 0x01, 0x12, 0x13, 0x08, 0x09, 0x1A, 0x1B); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
189 const vector unsigned char mmoP2 = (const vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
190 AVV(0x04, 0x05, 0x16, 0x17, 0x0C, 0x0D, 0x1E, 0x1F, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
191 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f, 0x1f); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
192 const vector unsigned char mmoP = (const vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
193 vec_lvsl(8, (unsigned char*)0); |
0 | 194 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
195 vector signed short mmoL1 = vec_perm(v_srcAss0, v_srcAss2, mmoP1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
196 vector signed short mmoL2 = vec_perm(v_srcAss4, v_srcAss6, mmoP2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
197 vector signed short mmoL = vec_perm(mmoL1, mmoL2, mmoP); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
198 vector signed short mmoR1 = vec_perm(v_srcAss5, v_srcAss7, mmoP1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
199 vector signed short mmoR2 = vec_perm(v_srcAss1, v_srcAss3, mmoP2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
200 vector signed short mmoR = vec_perm(mmoR1, mmoR2, mmoP); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
201 vector signed short mmoDiff = vec_sub(mmoL, mmoR); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
202 vector unsigned short mmoSum = (vector unsigned short)vec_add(mmoDiff, v2QP); |
0 | 203 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
204 if (vec_any_gt(mmoSum, v4QP)) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
205 return 0; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
206 else |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
207 return 1; |
0 | 208 } |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
209 else return 2; |
0 | 210 } |
211 | |
212 static inline void doVertLowPass_altivec(uint8_t *src, int stride, PPContext *c) { | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
213 /* |
0 | 214 this code makes no assumption on src or stride. |
215 One could remove the recomputation of the perm | |
216 vector by assuming (stride % 16) == 0, unfortunately | |
217 this is not always true. Quite a lot of load/stores | |
63 | 218 can be removed by assuming proper alignment of |
0 | 219 src & stride :-( |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
220 */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
221 uint8_t *src2 = src; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
222 const vector signed int zero = vec_splat_s32(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
223 const int properStride = (stride % 16); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
224 const int srcAlign = ((unsigned long)src2 % 16); |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
225 DECLARE_ALIGNED(16, short, qp[8]) = {c->QP}; |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
226 vector signed short vqp = vec_ld(0, qp); |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
227 vector signed short vb0, vb1, vb2, vb3, vb4, vb5, vb6, vb7, vb8, vb9; |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
228 vector unsigned char vbA0, vbA1, vbA2, vbA3, vbA4, vbA5, vbA6, vbA7, vbA8, vbA9; |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
229 vector unsigned char vbB0, vbB1, vbB2, vbB3, vbB4, vbB5, vbB6, vbB7, vbB8, vbB9; |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
230 vector unsigned char vbT0, vbT1, vbT2, vbT3, vbT4, vbT5, vbT6, vbT7, vbT8, vbT9; |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
231 vector unsigned char perml0, perml1, perml2, perml3, perml4, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
232 perml5, perml6, perml7, perml8, perml9; |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
233 register int j0 = 0, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
234 j1 = stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
235 j2 = 2 * stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
236 j3 = 3 * stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
237 j4 = 4 * stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
238 j5 = 5 * stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
239 j6 = 6 * stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
240 j7 = 7 * stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
241 j8 = 8 * stride, |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
242 j9 = 9 * stride; |
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
243 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
244 vqp = vec_splat(vqp, 0); |
0 | 245 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
246 src2 += stride*3; |
0 | 247 |
248 #define LOAD_LINE(i) \ | |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
249 perml##i = vec_lvsl(i * stride, src2); \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
250 vbA##i = vec_ld(i * stride, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
251 vbB##i = vec_ld(i * stride + 16, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
252 vbT##i = vec_perm(vbA##i, vbB##i, perml##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
253 vb##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
254 (vector signed short)vec_mergeh((vector unsigned char)zero, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
255 (vector unsigned char)vbT##i) |
0 | 256 |
257 #define LOAD_LINE_ALIGNED(i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
258 vbT##i = vec_ld(j##i, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
259 vb##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
260 (vector signed short)vec_mergeh((vector signed char)zero, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
261 (vector signed char)vbT##i) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
262 |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
263 /* Special-casing the aligned case is worthwhile, as all calls from |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
264 * the (transposed) horizontable deblocks will be aligned, in addition |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
265 * to the naturally aligned vertical deblocks. */ |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
266 if (properStride && srcAlign) { |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
267 LOAD_LINE_ALIGNED(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
268 LOAD_LINE_ALIGNED(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
269 LOAD_LINE_ALIGNED(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
270 LOAD_LINE_ALIGNED(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
271 LOAD_LINE_ALIGNED(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
272 LOAD_LINE_ALIGNED(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
273 LOAD_LINE_ALIGNED(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
274 LOAD_LINE_ALIGNED(7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
275 LOAD_LINE_ALIGNED(8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
276 LOAD_LINE_ALIGNED(9); |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
277 } else { |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
278 LOAD_LINE(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
279 LOAD_LINE(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
280 LOAD_LINE(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
281 LOAD_LINE(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
282 LOAD_LINE(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
283 LOAD_LINE(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
284 LOAD_LINE(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
285 LOAD_LINE(7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
286 LOAD_LINE(8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
287 LOAD_LINE(9); |
103 | 288 } |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
289 #undef LOAD_LINE |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
290 #undef LOAD_LINE_ALIGNED |
103 | 291 { |
292 const vector unsigned short v_2 = vec_splat_u16(2); | |
293 const vector unsigned short v_4 = vec_splat_u16(4); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
294 |
103 | 295 const vector signed short v_diff01 = vec_sub(vb0, vb1); |
296 const vector unsigned short v_cmp01 = | |
297 (const vector unsigned short) vec_cmplt(vec_abs(v_diff01), vqp); | |
298 const vector signed short v_first = vec_sel(vb1, vb0, v_cmp01); | |
299 const vector signed short v_diff89 = vec_sub(vb8, vb9); | |
300 const vector unsigned short v_cmp89 = | |
301 (const vector unsigned short) vec_cmplt(vec_abs(v_diff89), vqp); | |
302 const vector signed short v_last = vec_sel(vb8, vb9, v_cmp89); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
303 |
103 | 304 const vector signed short temp01 = vec_mladd(v_first, (vector signed short)v_4, vb1); |
305 const vector signed short temp02 = vec_add(vb2, vb3); | |
306 const vector signed short temp03 = vec_add(temp01, (vector signed short)v_4); | |
307 const vector signed short v_sumsB0 = vec_add(temp02, temp03); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
308 |
103 | 309 const vector signed short temp11 = vec_sub(v_sumsB0, v_first); |
310 const vector signed short v_sumsB1 = vec_add(temp11, vb4); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
311 |
103 | 312 const vector signed short temp21 = vec_sub(v_sumsB1, v_first); |
313 const vector signed short v_sumsB2 = vec_add(temp21, vb5); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
314 |
103 | 315 const vector signed short temp31 = vec_sub(v_sumsB2, v_first); |
316 const vector signed short v_sumsB3 = vec_add(temp31, vb6); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
317 |
103 | 318 const vector signed short temp41 = vec_sub(v_sumsB3, v_first); |
319 const vector signed short v_sumsB4 = vec_add(temp41, vb7); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
320 |
103 | 321 const vector signed short temp51 = vec_sub(v_sumsB4, vb1); |
322 const vector signed short v_sumsB5 = vec_add(temp51, vb8); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
323 |
103 | 324 const vector signed short temp61 = vec_sub(v_sumsB5, vb2); |
325 const vector signed short v_sumsB6 = vec_add(temp61, v_last); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
326 |
103 | 327 const vector signed short temp71 = vec_sub(v_sumsB6, vb3); |
328 const vector signed short v_sumsB7 = vec_add(temp71, v_last); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
329 |
103 | 330 const vector signed short temp81 = vec_sub(v_sumsB7, vb4); |
331 const vector signed short v_sumsB8 = vec_add(temp81, v_last); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
332 |
103 | 333 const vector signed short temp91 = vec_sub(v_sumsB8, vb5); |
334 const vector signed short v_sumsB9 = vec_add(temp91, v_last); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
335 |
103 | 336 #define COMPUTE_VR(i, j, k) \ |
337 const vector signed short temps1##i = \ | |
338 vec_add(v_sumsB##i, v_sumsB##k); \ | |
339 const vector signed short temps2##i = \ | |
340 vec_mladd(vb##j, (vector signed short)v_2, temps1##i); \ | |
341 const vector signed short vr##j = vec_sra(temps2##i, v_4) | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
342 |
103 | 343 COMPUTE_VR(0, 1, 2); |
344 COMPUTE_VR(1, 2, 3); | |
345 COMPUTE_VR(2, 3, 4); | |
346 COMPUTE_VR(3, 4, 5); | |
347 COMPUTE_VR(4, 5, 6); | |
348 COMPUTE_VR(5, 6, 7); | |
349 COMPUTE_VR(6, 7, 8); | |
350 COMPUTE_VR(7, 8, 9); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
351 |
103 | 352 const vector signed char neg1 = vec_splat_s8(-1); |
353 const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, | |
354 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
355 |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
356 #define PACK_AND_STORE(i) \ |
104 | 357 { const vector unsigned char perms##i = \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
358 vec_lvsr(i * stride, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
359 const vector unsigned char vf##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
360 vec_packsu(vr##i, (vector signed short)zero); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
361 const vector unsigned char vg##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
362 vec_perm(vf##i, vbT##i, permHH); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
363 const vector unsigned char mask##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
364 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
365 const vector unsigned char vg2##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
366 vec_perm(vg##i, vg##i, perms##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
367 const vector unsigned char svA##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
368 vec_sel(vbA##i, vg2##i, mask##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
369 const vector unsigned char svB##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
370 vec_sel(vg2##i, vbB##i, mask##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
371 vec_st(svA##i, i * stride, src2); \ |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
372 vec_st(svB##i, i * stride + 16, src2);} |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
373 |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
374 #define PACK_AND_STORE_ALIGNED(i) \ |
104 | 375 { const vector unsigned char vf##i = \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
376 vec_packsu(vr##i, (vector signed short)zero); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
377 const vector unsigned char vg##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
378 vec_perm(vf##i, vbT##i, permHH); \ |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
379 vec_st(vg##i, i * stride, src2);} |
0 | 380 |
103 | 381 /* Special-casing the aligned case is worthwhile, as all calls from |
382 * the (transposed) horizontable deblocks will be aligned, in addition | |
383 * to the naturally aligned vertical deblocks. */ | |
384 if (properStride && srcAlign) { | |
385 PACK_AND_STORE_ALIGNED(1) | |
386 PACK_AND_STORE_ALIGNED(2) | |
387 PACK_AND_STORE_ALIGNED(3) | |
388 PACK_AND_STORE_ALIGNED(4) | |
389 PACK_AND_STORE_ALIGNED(5) | |
390 PACK_AND_STORE_ALIGNED(6) | |
391 PACK_AND_STORE_ALIGNED(7) | |
392 PACK_AND_STORE_ALIGNED(8) | |
393 } else { | |
394 PACK_AND_STORE(1) | |
395 PACK_AND_STORE(2) | |
396 PACK_AND_STORE(3) | |
397 PACK_AND_STORE(4) | |
398 PACK_AND_STORE(5) | |
399 PACK_AND_STORE(6) | |
400 PACK_AND_STORE(7) | |
401 PACK_AND_STORE(8) | |
402 } | |
403 #undef PACK_AND_STORE | |
404 #undef PACK_AND_STORE_ALIGNED | |
0 | 405 } |
101
db57626d7d76
Remove declarations after statements from doVertLowPass_altivec
lu_zero
parents:
100
diff
changeset
|
406 } |
0 | 407 |
408 | |
409 | |
410 static inline void doVertDefFilter_altivec(uint8_t src[], int stride, PPContext *c) { | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
411 /* |
0 | 412 this code makes no assumption on src or stride. |
413 One could remove the recomputation of the perm | |
414 vector by assuming (stride % 16) == 0, unfortunately | |
415 this is not always true. Quite a lot of load/stores | |
63 | 416 can be removed by assuming proper alignment of |
0 | 417 src & stride :-( |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
418 */ |
102
eeb4558c950a
Remove declarations after statements from doVertDefFilter_altivec
lu_zero
parents:
101
diff
changeset
|
419 uint8_t *src2 = src + stride*3; |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
420 const vector signed int zero = vec_splat_s32(0); |
102
eeb4558c950a
Remove declarations after statements from doVertDefFilter_altivec
lu_zero
parents:
101
diff
changeset
|
421 DECLARE_ALIGNED(16, short, qp[8]) = {8*c->QP}; |
eeb4558c950a
Remove declarations after statements from doVertDefFilter_altivec
lu_zero
parents:
101
diff
changeset
|
422 vector signed short vqp = vec_splat( |
eeb4558c950a
Remove declarations after statements from doVertDefFilter_altivec
lu_zero
parents:
101
diff
changeset
|
423 (vector signed short)vec_ld(0, qp), 0); |
0 | 424 |
425 #define LOAD_LINE(i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
426 const vector unsigned char perm##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
427 vec_lvsl(i * stride, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
428 const vector unsigned char vbA##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
429 vec_ld(i * stride, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
430 const vector unsigned char vbB##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
431 vec_ld(i * stride + 16, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
432 const vector unsigned char vbT##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
433 vec_perm(vbA##i, vbB##i, perm##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
434 const vector signed short vb##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
435 (vector signed short)vec_mergeh((vector unsigned char)zero, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
436 (vector unsigned char)vbT##i) |
0 | 437 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
438 LOAD_LINE(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
439 LOAD_LINE(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
440 LOAD_LINE(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
441 LOAD_LINE(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
442 LOAD_LINE(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
443 LOAD_LINE(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
444 LOAD_LINE(7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
445 LOAD_LINE(8); |
0 | 446 #undef LOAD_LINE |
447 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
448 const vector signed short v_1 = vec_splat_s16(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
449 const vector signed short v_2 = vec_splat_s16(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
450 const vector signed short v_5 = vec_splat_s16(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
451 const vector signed short v_32 = vec_sl(v_1, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
452 (vector unsigned short)v_5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
453 /* middle energy */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
454 const vector signed short l3minusl6 = vec_sub(vb3, vb6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
455 const vector signed short l5minusl4 = vec_sub(vb5, vb4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
456 const vector signed short twotimes_l3minusl6 = vec_mladd(v_2, l3minusl6, (vector signed short)zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
457 const vector signed short mE = vec_mladd(v_5, l5minusl4, twotimes_l3minusl6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
458 const vector signed short absmE = vec_abs(mE); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
459 /* left & right energy */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
460 const vector signed short l1minusl4 = vec_sub(vb1, vb4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
461 const vector signed short l3minusl2 = vec_sub(vb3, vb2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
462 const vector signed short l5minusl8 = vec_sub(vb5, vb8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
463 const vector signed short l7minusl6 = vec_sub(vb7, vb6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
464 const vector signed short twotimes_l1minusl4 = vec_mladd(v_2, l1minusl4, (vector signed short)zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
465 const vector signed short twotimes_l5minusl8 = vec_mladd(v_2, l5minusl8, (vector signed short)zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
466 const vector signed short lE = vec_mladd(v_5, l3minusl2, twotimes_l1minusl4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
467 const vector signed short rE = vec_mladd(v_5, l7minusl6, twotimes_l5minusl8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
468 /* d */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
469 const vector signed short ddiff = vec_sub(absmE, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
470 vec_min(vec_abs(lE), |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
471 vec_abs(rE))); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
472 const vector signed short ddiffclamp = vec_max(ddiff, (vector signed short)zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
473 const vector signed short dtimes64 = vec_mladd(v_5, ddiffclamp, v_32); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
474 const vector signed short d = vec_sra(dtimes64, vec_splat_u16(6)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
475 const vector signed short minusd = vec_sub((vector signed short)zero, d); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
476 const vector signed short finald = vec_sel(minusd, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
477 d, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
478 vec_cmpgt(vec_sub((vector signed short)zero, mE), |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
479 (vector signed short)zero)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
480 /* q */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
481 const vector signed short qtimes2 = vec_sub(vb4, vb5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
482 /* for a shift right to behave like /2, we need to add one |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
483 to all negative integer */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
484 const vector signed short rounddown = vec_sel((vector signed short)zero, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
485 v_1, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
486 vec_cmplt(qtimes2, (vector signed short)zero)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
487 const vector signed short q = vec_sra(vec_add(qtimes2, rounddown), vec_splat_u16(1)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
488 /* clamp */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
489 const vector signed short dclamp_P1 = vec_max((vector signed short)zero, finald); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
490 const vector signed short dclamp_P = vec_min(dclamp_P1, q); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
491 const vector signed short dclamp_N1 = vec_min((vector signed short)zero, finald); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
492 const vector signed short dclamp_N = vec_max(dclamp_N1, q); |
0 | 493 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
494 const vector signed short dclampedfinal = vec_sel(dclamp_N, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
495 dclamp_P, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
496 vec_cmpgt(q, (vector signed short)zero)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
497 const vector signed short dornotd = vec_sel((vector signed short)zero, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
498 dclampedfinal, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
499 vec_cmplt(absmE, vqp)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
500 /* add/subtract to l4 and l5 */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
501 const vector signed short vb4minusd = vec_sub(vb4, dornotd); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
502 const vector signed short vb5plusd = vec_add(vb5, dornotd); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
503 /* finally, stores */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
504 const vector unsigned char st4 = vec_packsu(vb4minusd, (vector signed short)zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
505 const vector unsigned char st5 = vec_packsu(vb5plusd, (vector signed short)zero); |
0 | 506 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
507 const vector signed char neg1 = vec_splat_s8(-1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
508 const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
509 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F); |
0 | 510 |
511 #define STORE(i) \ | |
104 | 512 { const vector unsigned char perms##i = \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
513 vec_lvsr(i * stride, src2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
514 const vector unsigned char vg##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
515 vec_perm(st##i, vbT##i, permHH); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
516 const vector unsigned char mask##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
517 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
518 const vector unsigned char vg2##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
519 vec_perm(vg##i, vg##i, perms##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
520 const vector unsigned char svA##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
521 vec_sel(vbA##i, vg2##i, mask##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
522 const vector unsigned char svB##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
523 vec_sel(vg2##i, vbB##i, mask##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
524 vec_st(svA##i, i * stride, src2); \ |
102
eeb4558c950a
Remove declarations after statements from doVertDefFilter_altivec
lu_zero
parents:
101
diff
changeset
|
525 vec_st(svB##i, i * stride + 16, src2);} |
0 | 526 |
102
eeb4558c950a
Remove declarations after statements from doVertDefFilter_altivec
lu_zero
parents:
101
diff
changeset
|
527 STORE(4) |
eeb4558c950a
Remove declarations after statements from doVertDefFilter_altivec
lu_zero
parents:
101
diff
changeset
|
528 STORE(5) |
0 | 529 } |
530 | |
531 static inline void dering_altivec(uint8_t src[], int stride, PPContext *c) { | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
532 /* |
0 | 533 this code makes no assumption on src or stride. |
534 One could remove the recomputation of the perm | |
535 vector by assuming (stride % 16) == 0, unfortunately | |
536 this is not always true. Quite a lot of load/stores | |
63 | 537 can be removed by assuming proper alignment of |
0 | 538 src & stride :-( |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
539 */ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
540 uint8_t *srcCopy = src; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
541 DECLARE_ALIGNED(16, uint8_t, dt[16]); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
542 const vector signed int zero = vec_splat_s32(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
543 vector unsigned char v_dt; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
544 dt[0] = deringThreshold; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
545 v_dt = vec_splat(vec_ld(0, dt), 0); |
0 | 546 |
104 | 547 #define LOAD_LINE(i) \ |
548 const vector unsigned char perm##i = \ | |
549 vec_lvsl(i * stride, srcCopy); \ | |
550 vector unsigned char sA##i = vec_ld(i * stride, srcCopy); \ | |
551 vector unsigned char sB##i = vec_ld(i * stride + 16, srcCopy); \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
552 vector unsigned char src##i = vec_perm(sA##i, sB##i, perm##i) |
0 | 553 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
554 LOAD_LINE(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
555 LOAD_LINE(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
556 LOAD_LINE(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
557 LOAD_LINE(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
558 LOAD_LINE(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
559 LOAD_LINE(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
560 LOAD_LINE(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
561 LOAD_LINE(7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
562 LOAD_LINE(8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
563 LOAD_LINE(9); |
0 | 564 #undef LOAD_LINE |
565 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
566 vector unsigned char v_avg; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
567 { |
0 | 568 const vector unsigned char trunc_perm = (vector unsigned char) |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
569 AVV(0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
570 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18); |
0 | 571 const vector unsigned char trunc_src12 = vec_perm(src1, src2, trunc_perm); |
572 const vector unsigned char trunc_src34 = vec_perm(src3, src4, trunc_perm); | |
573 const vector unsigned char trunc_src56 = vec_perm(src5, src6, trunc_perm); | |
574 const vector unsigned char trunc_src78 = vec_perm(src7, src8, trunc_perm); | |
575 | |
576 #define EXTRACT(op) do { \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
577 const vector unsigned char s##op##_1 = vec_##op(trunc_src12, trunc_src34); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
578 const vector unsigned char s##op##_2 = vec_##op(trunc_src56, trunc_src78); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
579 const vector unsigned char s##op##_6 = vec_##op(s##op##_1, s##op##_2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
580 const vector unsigned char s##op##_8h = vec_mergeh(s##op##_6, s##op##_6); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
581 const vector unsigned char s##op##_8l = vec_mergel(s##op##_6, s##op##_6); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
582 const vector unsigned char s##op##_9 = vec_##op(s##op##_8h, s##op##_8l); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
583 const vector unsigned char s##op##_9h = vec_mergeh(s##op##_9, s##op##_9); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
584 const vector unsigned char s##op##_9l = vec_mergel(s##op##_9, s##op##_9); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
585 const vector unsigned char s##op##_10 = vec_##op(s##op##_9h, s##op##_9l); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
586 const vector unsigned char s##op##_10h = vec_mergeh(s##op##_10, s##op##_10); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
587 const vector unsigned char s##op##_10l = vec_mergel(s##op##_10, s##op##_10); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
588 const vector unsigned char s##op##_11 = vec_##op(s##op##_10h, s##op##_10l); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
589 const vector unsigned char s##op##_11h = vec_mergeh(s##op##_11, s##op##_11); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
590 const vector unsigned char s##op##_11l = vec_mergel(s##op##_11, s##op##_11); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
591 v_##op = vec_##op(s##op##_11h, s##op##_11l); } while (0) |
0 | 592 |
593 vector unsigned char v_min; | |
594 vector unsigned char v_max; | |
595 EXTRACT(min); | |
596 EXTRACT(max); | |
597 #undef EXTRACT | |
598 | |
599 if (vec_all_lt(vec_sub(v_max, v_min), v_dt)) | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
600 return; |
0 | 601 |
602 v_avg = vec_avg(v_min, v_max); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
603 } |
0 | 604 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
605 DECLARE_ALIGNED(16, signed int, S[8]); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
606 { |
0 | 607 const vector unsigned short mask1 = (vector unsigned short) |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
608 AVV(0x0001, 0x0002, 0x0004, 0x0008, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
609 0x0010, 0x0020, 0x0040, 0x0080); |
0 | 610 const vector unsigned short mask2 = (vector unsigned short) |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
611 AVV(0x0100, 0x0200, 0x0000, 0x0000, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
612 0x0000, 0x0000, 0x0000, 0x0000); |
0 | 613 |
614 const vector unsigned int vuint32_16 = vec_sl(vec_splat_u32(1), vec_splat_u32(4)); | |
615 const vector unsigned int vuint32_1 = vec_splat_u32(1); | |
616 | |
617 #define COMPARE(i) \ | |
618 vector signed int sum##i; \ | |
619 do { \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
620 const vector unsigned char cmp##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
621 (vector unsigned char)vec_cmpgt(src##i, v_avg); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
622 const vector unsigned short cmpHi##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
623 (vector unsigned short)vec_mergeh(cmp##i, cmp##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
624 const vector unsigned short cmpLi##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
625 (vector unsigned short)vec_mergel(cmp##i, cmp##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
626 const vector signed short cmpHf##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
627 (vector signed short)vec_and(cmpHi##i, mask1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
628 const vector signed short cmpLf##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
629 (vector signed short)vec_and(cmpLi##i, mask2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
630 const vector signed int sump##i = vec_sum4s(cmpHf##i, zero); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
631 const vector signed int sumq##i = vec_sum4s(cmpLf##i, sump##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
632 sum##i = vec_sums(sumq##i, zero); } while (0) |
0 | 633 |
634 COMPARE(0); | |
635 COMPARE(1); | |
636 COMPARE(2); | |
637 COMPARE(3); | |
638 COMPARE(4); | |
639 COMPARE(5); | |
640 COMPARE(6); | |
641 COMPARE(7); | |
642 COMPARE(8); | |
643 COMPARE(9); | |
644 #undef COMPARE | |
645 | |
646 vector signed int sumA2; | |
647 vector signed int sumB2; | |
648 { | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
649 const vector signed int sump02 = vec_mergel(sum0, sum2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
650 const vector signed int sump13 = vec_mergel(sum1, sum3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
651 const vector signed int sumA = vec_mergel(sump02, sump13); |
0 | 652 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
653 const vector signed int sump46 = vec_mergel(sum4, sum6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
654 const vector signed int sump57 = vec_mergel(sum5, sum7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
655 const vector signed int sumB = vec_mergel(sump46, sump57); |
0 | 656 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
657 const vector signed int sump8A = vec_mergel(sum8, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
658 const vector signed int sump9B = vec_mergel(sum9, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
659 const vector signed int sumC = vec_mergel(sump8A, sump9B); |
0 | 660 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
661 const vector signed int tA = vec_sl(vec_nor(zero, sumA), vuint32_16); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
662 const vector signed int tB = vec_sl(vec_nor(zero, sumB), vuint32_16); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
663 const vector signed int tC = vec_sl(vec_nor(zero, sumC), vuint32_16); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
664 const vector signed int t2A = vec_or(sumA, tA); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
665 const vector signed int t2B = vec_or(sumB, tB); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
666 const vector signed int t2C = vec_or(sumC, tC); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
667 const vector signed int t3A = vec_and(vec_sra(t2A, vuint32_1), |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
668 vec_sl(t2A, vuint32_1)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
669 const vector signed int t3B = vec_and(vec_sra(t2B, vuint32_1), |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
670 vec_sl(t2B, vuint32_1)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
671 const vector signed int t3C = vec_and(vec_sra(t2C, vuint32_1), |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
672 vec_sl(t2C, vuint32_1)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
673 const vector signed int yA = vec_and(t2A, t3A); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
674 const vector signed int yB = vec_and(t2B, t3B); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
675 const vector signed int yC = vec_and(t2C, t3C); |
0 | 676 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
677 const vector unsigned char strangeperm1 = vec_lvsl(4, (unsigned char*)0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
678 const vector unsigned char strangeperm2 = vec_lvsl(8, (unsigned char*)0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
679 const vector signed int sumAd4 = vec_perm(yA, yB, strangeperm1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
680 const vector signed int sumAd8 = vec_perm(yA, yB, strangeperm2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
681 const vector signed int sumBd4 = vec_perm(yB, yC, strangeperm1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
682 const vector signed int sumBd8 = vec_perm(yB, yC, strangeperm2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
683 const vector signed int sumAp = vec_and(yA, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
684 vec_and(sumAd4,sumAd8)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
685 const vector signed int sumBp = vec_and(yB, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
686 vec_and(sumBd4,sumBd8)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
687 sumA2 = vec_or(sumAp, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
688 vec_sra(sumAp, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
689 vuint32_16)); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
690 sumB2 = vec_or(sumBp, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
691 vec_sra(sumBp, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
692 vuint32_16)); |
0 | 693 } |
694 vec_st(sumA2, 0, S); | |
695 vec_st(sumB2, 16, S); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
696 } |
0 | 697 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
698 /* I'm not sure the following is actually faster |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
699 than straight, unvectorized C code :-( */ |
0 | 700 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
701 DECLARE_ALIGNED(16, int, tQP2[4]); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
702 tQP2[0]= c->QP/2 + 1; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
703 vector signed int vQP2 = vec_ld(0, tQP2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
704 vQP2 = vec_splat(vQP2, 0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
705 const vector signed int vsint32_8 = vec_splat_s32(8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
706 const vector unsigned int vuint32_4 = vec_splat_u32(4); |
0 | 707 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
708 const vector unsigned char permA1 = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
709 AVV(0x00, 0x01, 0x02, 0x10, 0x11, 0x12, 0x1F, 0x1F, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
710 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
711 const vector unsigned char permA2 = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
712 AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x10, 0x11, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
713 0x12, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F, 0x1F); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
714 const vector unsigned char permA1inc = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
715 AVV(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
716 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
717 const vector unsigned char permA2inc = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
718 AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
719 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
720 const vector unsigned char magic = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
721 AVV(0x01, 0x02, 0x01, 0x02, 0x04, 0x02, 0x01, 0x02, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
722 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
723 const vector unsigned char extractPerm = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
724 AVV(0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
725 0x10, 0x10, 0x10, 0x01, 0x10, 0x10, 0x10, 0x01); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
726 const vector unsigned char extractPermInc = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
727 AVV(0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
728 0x00, 0x00, 0x00, 0x01, 0x00, 0x00, 0x00, 0x01); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
729 const vector unsigned char identity = vec_lvsl(0,(unsigned char *)0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
730 const vector unsigned char tenRight = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
731 AVV(0x00, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
732 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
733 const vector unsigned char eightLeft = (vector unsigned char) |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
734 AVV(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
735 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x08); |
0 | 736 |
737 | |
738 #define F_INIT(i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
739 vector unsigned char tenRightM##i = tenRight; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
740 vector unsigned char permA1M##i = permA1; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
741 vector unsigned char permA2M##i = permA2; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
742 vector unsigned char extractPermM##i = extractPerm |
0 | 743 |
744 #define F2(i, j, k, l) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
745 if (S[i] & (1 << (l+1))) { \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
746 const vector unsigned char a_##j##_A##l = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
747 vec_perm(src##i, src##j, permA1M##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
748 const vector unsigned char a_##j##_B##l = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
749 vec_perm(a_##j##_A##l, src##k, permA2M##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
750 const vector signed int a_##j##_sump##l = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
751 (vector signed int)vec_msum(a_##j##_B##l, magic, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
752 (vector unsigned int)zero); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
753 vector signed int F_##j##_##l = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
754 vec_sr(vec_sums(a_##j##_sump##l, vsint32_8), vuint32_4); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
755 F_##j##_##l = vec_splat(F_##j##_##l, 3); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
756 const vector signed int p_##j##_##l = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
757 (vector signed int)vec_perm(src##j, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
758 (vector unsigned char)zero, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
759 extractPermM##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
760 const vector signed int sum_##j##_##l = vec_add( p_##j##_##l, vQP2);\ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
761 const vector signed int diff_##j##_##l = vec_sub( p_##j##_##l, vQP2);\ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
762 vector signed int newpm_##j##_##l; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
763 if (vec_all_lt(sum_##j##_##l, F_##j##_##l)) \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
764 newpm_##j##_##l = sum_##j##_##l; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
765 else if (vec_all_gt(diff_##j##_##l, F_##j##_##l)) \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
766 newpm_##j##_##l = diff_##j##_##l; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
767 else newpm_##j##_##l = F_##j##_##l; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
768 const vector unsigned char newpm2_##j##_##l = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
769 vec_splat((vector unsigned char)newpm_##j##_##l, 15); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
770 const vector unsigned char mask##j##l = vec_add(identity, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
771 tenRightM##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
772 src##j = vec_perm(src##j, newpm2_##j##_##l, mask##j##l); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
773 } \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
774 permA1M##i = vec_add(permA1M##i, permA1inc); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
775 permA2M##i = vec_add(permA2M##i, permA2inc); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
776 tenRightM##i = vec_sro(tenRightM##i, eightLeft); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
777 extractPermM##i = vec_add(extractPermM##i, extractPermInc) |
0 | 778 |
779 #define ITER(i, j, k) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
780 F_INIT(i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
781 F2(i, j, k, 0); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
782 F2(i, j, k, 1); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
783 F2(i, j, k, 2); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
784 F2(i, j, k, 3); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
785 F2(i, j, k, 4); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
786 F2(i, j, k, 5); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
787 F2(i, j, k, 6); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
788 F2(i, j, k, 7) |
0 | 789 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
790 ITER(0, 1, 2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
791 ITER(1, 2, 3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
792 ITER(2, 3, 4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
793 ITER(3, 4, 5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
794 ITER(4, 5, 6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
795 ITER(5, 6, 7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
796 ITER(6, 7, 8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
797 ITER(7, 8, 9); |
0 | 798 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
799 const vector signed char neg1 = vec_splat_s8(-1); |
0 | 800 |
801 #define STORE_LINE(i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
802 const vector unsigned char permST##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
803 vec_lvsr(i * stride, srcCopy); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
804 const vector unsigned char maskST##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
805 vec_perm((vector unsigned char)zero, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
806 (vector unsigned char)neg1, permST##i);\ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
807 src##i = vec_perm(src##i ,src##i, permST##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
808 sA##i= vec_sel(sA##i, src##i, maskST##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
809 sB##i= vec_sel(src##i, sB##i, maskST##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
810 vec_st(sA##i, i * stride, srcCopy); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
811 vec_st(sB##i, i * stride + 16, srcCopy) |
0 | 812 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
813 STORE_LINE(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
814 STORE_LINE(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
815 STORE_LINE(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
816 STORE_LINE(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
817 STORE_LINE(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
818 STORE_LINE(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
819 STORE_LINE(7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
820 STORE_LINE(8); |
0 | 821 |
822 #undef STORE_LINE | |
823 #undef ITER | |
824 #undef F2 | |
825 } | |
826 | |
827 #define doHorizLowPass_altivec(a...) doHorizLowPass_C(a) | |
828 #define doHorizDefFilter_altivec(a...) doHorizDefFilter_C(a) | |
829 #define do_a_deblock_altivec(a...) do_a_deblock_C(a) | |
830 | |
831 static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride, | |
98 | 832 uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise) |
0 | 833 { |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
834 const vector signed int zero = vec_splat_s32(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
835 const vector signed short vsint16_1 = vec_splat_s16(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
836 vector signed int v_dp = zero; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
837 vector signed int v_sysdp = zero; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
838 int d, sysd, i; |
0 | 839 |
98 | 840 tempBlurredPast[127]= maxNoise[0]; |
841 tempBlurredPast[128]= maxNoise[1]; | |
842 tempBlurredPast[129]= maxNoise[2]; | |
0 | 843 |
844 #define LOAD_LINE(src, i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
845 register int j##src##i = i * stride; \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
846 vector unsigned char perm##src##i = vec_lvsl(j##src##i, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
847 const vector unsigned char v_##src##A1##i = vec_ld(j##src##i, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
848 const vector unsigned char v_##src##A2##i = vec_ld(j##src##i + 16, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
849 const vector unsigned char v_##src##A##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
850 vec_perm(v_##src##A1##i, v_##src##A2##i, perm##src##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
851 vector signed short v_##src##Ass##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
852 (vector signed short)vec_mergeh((vector signed char)zero, \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
853 (vector signed char)v_##src##A##i) |
0 | 854 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
855 LOAD_LINE(src, 0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
856 LOAD_LINE(src, 1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
857 LOAD_LINE(src, 2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
858 LOAD_LINE(src, 3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
859 LOAD_LINE(src, 4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
860 LOAD_LINE(src, 5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
861 LOAD_LINE(src, 6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
862 LOAD_LINE(src, 7); |
0 | 863 |
98 | 864 LOAD_LINE(tempBlurred, 0); |
865 LOAD_LINE(tempBlurred, 1); | |
866 LOAD_LINE(tempBlurred, 2); | |
867 LOAD_LINE(tempBlurred, 3); | |
868 LOAD_LINE(tempBlurred, 4); | |
869 LOAD_LINE(tempBlurred, 5); | |
870 LOAD_LINE(tempBlurred, 6); | |
871 LOAD_LINE(tempBlurred, 7); | |
0 | 872 #undef LOAD_LINE |
873 | |
874 #define ACCUMULATE_DIFFS(i) \ | |
104 | 875 vector signed short v_d##i = vec_sub(v_tempBlurredAss##i, \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
876 v_srcAss##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
877 v_dp = vec_msums(v_d##i, v_d##i, v_dp); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
878 v_sysdp = vec_msums(v_d##i, vsint16_1, v_sysdp) |
0 | 879 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
880 ACCUMULATE_DIFFS(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
881 ACCUMULATE_DIFFS(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
882 ACCUMULATE_DIFFS(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
883 ACCUMULATE_DIFFS(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
884 ACCUMULATE_DIFFS(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
885 ACCUMULATE_DIFFS(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
886 ACCUMULATE_DIFFS(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
887 ACCUMULATE_DIFFS(7); |
0 | 888 #undef ACCUMULATE_DIFFS |
889 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
890 v_dp = vec_sums(v_dp, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
891 v_sysdp = vec_sums(v_sysdp, zero); |
0 | 892 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
893 v_dp = vec_splat(v_dp, 3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
894 v_sysdp = vec_splat(v_sysdp, 3); |
0 | 895 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
896 vec_ste(v_dp, 0, &d); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
897 vec_ste(v_sysdp, 0, &sysd); |
0 | 898 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
899 i = d; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
900 d = (4*d |
98 | 901 +(*(tempBlurredPast-256)) |
902 +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1)) | |
903 +(*(tempBlurredPast+256)) | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
904 +4)>>3; |
0 | 905 |
98 | 906 *tempBlurredPast=i; |
0 | 907 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
908 if (d > maxNoise[1]) { |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
909 if (d < maxNoise[2]) { |
98 | 910 #define OP(i) v_tempBlurredAss##i = vec_avg(v_tempBlurredAss##i, v_srcAss##i); |
0 | 911 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
912 OP(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
913 OP(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
914 OP(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
915 OP(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
916 OP(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
917 OP(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
918 OP(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
919 OP(7); |
0 | 920 #undef OP |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
921 } else { |
98 | 922 #define OP(i) v_tempBlurredAss##i = v_srcAss##i; |
0 | 923 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
924 OP(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
925 OP(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
926 OP(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
927 OP(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
928 OP(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
929 OP(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
930 OP(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
931 OP(7); |
0 | 932 #undef OP |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
933 } |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
934 } else { |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
935 if (d < maxNoise[0]) { |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
936 const vector signed short vsint16_7 = vec_splat_s16(7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
937 const vector signed short vsint16_4 = vec_splat_s16(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
938 const vector unsigned short vuint16_3 = vec_splat_u16(3); |
0 | 939 |
940 #define OP(i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
941 const vector signed short v_temp##i = \ |
104 | 942 vec_mladd(v_tempBlurredAss##i, \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
943 vsint16_7, v_srcAss##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
944 const vector signed short v_temp2##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
945 vec_add(v_temp##i, vsint16_4); \ |
98 | 946 v_tempBlurredAss##i = vec_sr(v_temp2##i, vuint16_3) |
0 | 947 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
948 OP(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
949 OP(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
950 OP(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
951 OP(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
952 OP(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
953 OP(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
954 OP(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
955 OP(7); |
0 | 956 #undef OP |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
957 } else { |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
958 const vector signed short vsint16_3 = vec_splat_s16(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
959 const vector signed short vsint16_2 = vec_splat_s16(2); |
0 | 960 |
961 #define OP(i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
962 const vector signed short v_temp##i = \ |
104 | 963 vec_mladd(v_tempBlurredAss##i, \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
964 vsint16_3, v_srcAss##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
965 const vector signed short v_temp2##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
966 vec_add(v_temp##i, vsint16_2); \ |
98 | 967 v_tempBlurredAss##i = vec_sr(v_temp2##i, (vector unsigned short)vsint16_2) |
0 | 968 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
969 OP(0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
970 OP(1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
971 OP(2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
972 OP(3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
973 OP(4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
974 OP(5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
975 OP(6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
976 OP(7); |
0 | 977 #undef OP |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
978 } |
0 | 979 } |
980 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
981 const vector signed char neg1 = vec_splat_s8(-1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
982 const vector unsigned char permHH = (const vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
983 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, 0x1F); |
0 | 984 |
985 #define PACK_AND_STORE(src, i) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
986 const vector unsigned char perms##src##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
987 vec_lvsr(i * stride, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
988 const vector unsigned char vf##src##i = \ |
98 | 989 vec_packsu(v_tempBlurredAss##i, (vector signed short)zero); \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
990 const vector unsigned char vg##src##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
991 vec_perm(vf##src##i, v_##src##A##i, permHH); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
992 const vector unsigned char mask##src##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
993 vec_perm((vector unsigned char)zero, (vector unsigned char)neg1, perms##src##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
994 const vector unsigned char vg2##src##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
995 vec_perm(vg##src##i, vg##src##i, perms##src##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
996 const vector unsigned char svA##src##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
997 vec_sel(v_##src##A1##i, vg2##src##i, mask##src##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
998 const vector unsigned char svB##src##i = \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
999 vec_sel(vg2##src##i, v_##src##A2##i, mask##src##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1000 vec_st(svA##src##i, i * stride, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1001 vec_st(svB##src##i, i * stride + 16, src) |
0 | 1002 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1003 PACK_AND_STORE(src, 0); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1004 PACK_AND_STORE(src, 1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1005 PACK_AND_STORE(src, 2); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1006 PACK_AND_STORE(src, 3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1007 PACK_AND_STORE(src, 4); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1008 PACK_AND_STORE(src, 5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1009 PACK_AND_STORE(src, 6); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1010 PACK_AND_STORE(src, 7); |
98 | 1011 PACK_AND_STORE(tempBlurred, 0); |
1012 PACK_AND_STORE(tempBlurred, 1); | |
1013 PACK_AND_STORE(tempBlurred, 2); | |
1014 PACK_AND_STORE(tempBlurred, 3); | |
1015 PACK_AND_STORE(tempBlurred, 4); | |
1016 PACK_AND_STORE(tempBlurred, 5); | |
1017 PACK_AND_STORE(tempBlurred, 6); | |
1018 PACK_AND_STORE(tempBlurred, 7); | |
0 | 1019 #undef PACK_AND_STORE |
1020 } | |
1021 | |
1022 static inline void transpose_16x8_char_toPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) { | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1023 const vector unsigned char zero = vec_splat_u8(0); |
0 | 1024 |
1025 #define LOAD_DOUBLE_LINE(i, j) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1026 vector unsigned char perm1##i = vec_lvsl(i * stride, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1027 vector unsigned char perm2##i = vec_lvsl(j * stride, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1028 vector unsigned char srcA##i = vec_ld(i * stride, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1029 vector unsigned char srcB##i = vec_ld(i * stride + 16, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1030 vector unsigned char srcC##i = vec_ld(j * stride, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1031 vector unsigned char srcD##i = vec_ld(j * stride+ 16, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1032 vector unsigned char src##i = vec_perm(srcA##i, srcB##i, perm1##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1033 vector unsigned char src##j = vec_perm(srcC##i, srcD##i, perm2##i) |
0 | 1034 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1035 LOAD_DOUBLE_LINE(0, 1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1036 LOAD_DOUBLE_LINE(2, 3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1037 LOAD_DOUBLE_LINE(4, 5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1038 LOAD_DOUBLE_LINE(6, 7); |
0 | 1039 #undef LOAD_DOUBLE_LINE |
1040 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1041 vector unsigned char tempA = vec_mergeh(src0, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1042 vector unsigned char tempB = vec_mergel(src0, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1043 vector unsigned char tempC = vec_mergeh(src1, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1044 vector unsigned char tempD = vec_mergel(src1, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1045 vector unsigned char tempE = vec_mergeh(src2, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1046 vector unsigned char tempF = vec_mergel(src2, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1047 vector unsigned char tempG = vec_mergeh(src3, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1048 vector unsigned char tempH = vec_mergel(src3, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1049 vector unsigned char tempI = vec_mergeh(src4, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1050 vector unsigned char tempJ = vec_mergel(src4, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1051 vector unsigned char tempK = vec_mergeh(src5, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1052 vector unsigned char tempL = vec_mergel(src5, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1053 vector unsigned char tempM = vec_mergeh(src6, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1054 vector unsigned char tempN = vec_mergel(src6, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1055 vector unsigned char tempO = vec_mergeh(src7, zero); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1056 vector unsigned char tempP = vec_mergel(src7, zero); |
0 | 1057 |
104 | 1058 vector unsigned char temp0 = vec_mergeh(tempA, tempI); |
1059 vector unsigned char temp1 = vec_mergel(tempA, tempI); | |
1060 vector unsigned char temp2 = vec_mergeh(tempB, tempJ); | |
1061 vector unsigned char temp3 = vec_mergel(tempB, tempJ); | |
1062 vector unsigned char temp4 = vec_mergeh(tempC, tempK); | |
1063 vector unsigned char temp5 = vec_mergel(tempC, tempK); | |
1064 vector unsigned char temp6 = vec_mergeh(tempD, tempL); | |
1065 vector unsigned char temp7 = vec_mergel(tempD, tempL); | |
1066 vector unsigned char temp8 = vec_mergeh(tempE, tempM); | |
1067 vector unsigned char temp9 = vec_mergel(tempE, tempM); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1068 vector unsigned char temp10 = vec_mergeh(tempF, tempN); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1069 vector unsigned char temp11 = vec_mergel(tempF, tempN); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1070 vector unsigned char temp12 = vec_mergeh(tempG, tempO); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1071 vector unsigned char temp13 = vec_mergel(tempG, tempO); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1072 vector unsigned char temp14 = vec_mergeh(tempH, tempP); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1073 vector unsigned char temp15 = vec_mergel(tempH, tempP); |
0 | 1074 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1075 tempA = vec_mergeh(temp0, temp8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1076 tempB = vec_mergel(temp0, temp8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1077 tempC = vec_mergeh(temp1, temp9); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1078 tempD = vec_mergel(temp1, temp9); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1079 tempE = vec_mergeh(temp2, temp10); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1080 tempF = vec_mergel(temp2, temp10); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1081 tempG = vec_mergeh(temp3, temp11); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1082 tempH = vec_mergel(temp3, temp11); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1083 tempI = vec_mergeh(temp4, temp12); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1084 tempJ = vec_mergel(temp4, temp12); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1085 tempK = vec_mergeh(temp5, temp13); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1086 tempL = vec_mergel(temp5, temp13); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1087 tempM = vec_mergeh(temp6, temp14); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1088 tempN = vec_mergel(temp6, temp14); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1089 tempO = vec_mergeh(temp7, temp15); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1090 tempP = vec_mergel(temp7, temp15); |
0 | 1091 |
104 | 1092 temp0 = vec_mergeh(tempA, tempI); |
1093 temp1 = vec_mergel(tempA, tempI); | |
1094 temp2 = vec_mergeh(tempB, tempJ); | |
1095 temp3 = vec_mergel(tempB, tempJ); | |
1096 temp4 = vec_mergeh(tempC, tempK); | |
1097 temp5 = vec_mergel(tempC, tempK); | |
1098 temp6 = vec_mergeh(tempD, tempL); | |
1099 temp7 = vec_mergel(tempD, tempL); | |
1100 temp8 = vec_mergeh(tempE, tempM); | |
1101 temp9 = vec_mergel(tempE, tempM); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1102 temp10 = vec_mergeh(tempF, tempN); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1103 temp11 = vec_mergel(tempF, tempN); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1104 temp12 = vec_mergeh(tempG, tempO); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1105 temp13 = vec_mergel(tempG, tempO); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1106 temp14 = vec_mergeh(tempH, tempP); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1107 temp15 = vec_mergel(tempH, tempP); |
0 | 1108 |
104 | 1109 vec_st(temp0, 0, dst); |
1110 vec_st(temp1, 16, dst); | |
1111 vec_st(temp2, 32, dst); | |
1112 vec_st(temp3, 48, dst); | |
1113 vec_st(temp4, 64, dst); | |
1114 vec_st(temp5, 80, dst); | |
1115 vec_st(temp6, 96, dst); | |
1116 vec_st(temp7, 112, dst); | |
1117 vec_st(temp8, 128, dst); | |
1118 vec_st(temp9, 144, dst); | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1119 vec_st(temp10, 160, dst); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1120 vec_st(temp11, 176, dst); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1121 vec_st(temp12, 192, dst); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1122 vec_st(temp13, 208, dst); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1123 vec_st(temp14, 224, dst); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1124 vec_st(temp15, 240, dst); |
0 | 1125 } |
1126 | |
1127 static inline void transpose_8x16_char_fromPackedAlign_altivec(unsigned char* dst, unsigned char* src, int stride) { | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1128 const vector unsigned char zero = vec_splat_u8(0); |
0 | 1129 |
1130 #define LOAD_DOUBLE_LINE(i, j) \ | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1131 vector unsigned char src##i = vec_ld(i * 16, src); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1132 vector unsigned char src##j = vec_ld(j * 16, src) |
0 | 1133 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1134 LOAD_DOUBLE_LINE(0, 1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1135 LOAD_DOUBLE_LINE(2, 3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1136 LOAD_DOUBLE_LINE(4, 5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1137 LOAD_DOUBLE_LINE(6, 7); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1138 LOAD_DOUBLE_LINE(8, 9); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1139 LOAD_DOUBLE_LINE(10, 11); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1140 LOAD_DOUBLE_LINE(12, 13); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1141 LOAD_DOUBLE_LINE(14, 15); |
0 | 1142 #undef LOAD_DOUBLE_LINE |
1143 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1144 vector unsigned char tempA = vec_mergeh(src0, src8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1145 vector unsigned char tempB; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1146 vector unsigned char tempC = vec_mergeh(src1, src9); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1147 vector unsigned char tempD; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1148 vector unsigned char tempE = vec_mergeh(src2, src10); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1149 vector unsigned char tempG = vec_mergeh(src3, src11); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1150 vector unsigned char tempI = vec_mergeh(src4, src12); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1151 vector unsigned char tempJ; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1152 vector unsigned char tempK = vec_mergeh(src5, src13); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1153 vector unsigned char tempL; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1154 vector unsigned char tempM = vec_mergeh(src6, src14); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1155 vector unsigned char tempO = vec_mergeh(src7, src15); |
0 | 1156 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1157 vector unsigned char temp0 = vec_mergeh(tempA, tempI); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1158 vector unsigned char temp1 = vec_mergel(tempA, tempI); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1159 vector unsigned char temp2; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1160 vector unsigned char temp3; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1161 vector unsigned char temp4 = vec_mergeh(tempC, tempK); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1162 vector unsigned char temp5 = vec_mergel(tempC, tempK); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1163 vector unsigned char temp6; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1164 vector unsigned char temp7; |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1165 vector unsigned char temp8 = vec_mergeh(tempE, tempM); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1166 vector unsigned char temp9 = vec_mergel(tempE, tempM); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1167 vector unsigned char temp12 = vec_mergeh(tempG, tempO); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1168 vector unsigned char temp13 = vec_mergel(tempG, tempO); |
0 | 1169 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1170 tempA = vec_mergeh(temp0, temp8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1171 tempB = vec_mergel(temp0, temp8); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1172 tempC = vec_mergeh(temp1, temp9); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1173 tempD = vec_mergel(temp1, temp9); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1174 tempI = vec_mergeh(temp4, temp12); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1175 tempJ = vec_mergel(temp4, temp12); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1176 tempK = vec_mergeh(temp5, temp13); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1177 tempL = vec_mergel(temp5, temp13); |
0 | 1178 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1179 temp0 = vec_mergeh(tempA, tempI); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1180 temp1 = vec_mergel(tempA, tempI); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1181 temp2 = vec_mergeh(tempB, tempJ); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1182 temp3 = vec_mergel(tempB, tempJ); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1183 temp4 = vec_mergeh(tempC, tempK); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1184 temp5 = vec_mergel(tempC, tempK); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1185 temp6 = vec_mergeh(tempD, tempL); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1186 temp7 = vec_mergel(tempD, tempL); |
0 | 1187 |
1188 | |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1189 const vector signed char neg1 = vec_splat_s8(-1); |
0 | 1190 #define STORE_DOUBLE_LINE(i, j) \ |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1191 vector unsigned char dstA##i = vec_ld(i * stride, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1192 vector unsigned char dstB##i = vec_ld(i * stride + 16, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1193 vector unsigned char dstA##j = vec_ld(j * stride, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1194 vector unsigned char dstB##j = vec_ld(j * stride+ 16, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1195 vector unsigned char align##i = vec_lvsr(i * stride, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1196 vector unsigned char align##j = vec_lvsr(j * stride, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1197 vector unsigned char mask##i = vec_perm(zero, (vector unsigned char)neg1, align##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1198 vector unsigned char mask##j = vec_perm(zero, (vector unsigned char)neg1, align##j); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1199 vector unsigned char dstR##i = vec_perm(temp##i, temp##i, align##i);\ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1200 vector unsigned char dstR##j = vec_perm(temp##j, temp##j, align##j);\ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1201 vector unsigned char dstAF##i = vec_sel(dstA##i, dstR##i, mask##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1202 vector unsigned char dstBF##i = vec_sel(dstR##i, dstB##i, mask##i); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1203 vector unsigned char dstAF##j = vec_sel(dstA##j, dstR##j, mask##j); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1204 vector unsigned char dstBF##j = vec_sel(dstR##j, dstB##j, mask##j); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1205 vec_st(dstAF##i, i * stride, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1206 vec_st(dstBF##i, i * stride + 16, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1207 vec_st(dstAF##j, j * stride, dst); \ |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1208 vec_st(dstBF##j, j * stride + 16, dst) |
0 | 1209 |
95
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1210 STORE_DOUBLE_LINE(0,1); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1211 STORE_DOUBLE_LINE(2,3); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1212 STORE_DOUBLE_LINE(4,5); |
c24dab9bca80
cosmetics: Fix indentation to be 4 spaces and consistently place {}.
diego
parents:
88
diff
changeset
|
1213 STORE_DOUBLE_LINE(6,7); |
0 | 1214 } |