Mercurial > libavcodec.hg
annotate ppc/h264_altivec.c @ 3804:f472f55953bb libavcodec
prevent warning in vorbis_data.c
author | ods15 |
---|---|
date | Sun, 01 Oct 2006 13:53:08 +0000 |
parents | 56e60411527c |
children | c8c591fe26f8 |
rev | line source |
---|---|
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
1 /* |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
2 * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org> |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
3 * |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
4 * This library is free software; you can redistribute it and/or |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
5 * modify it under the terms of the GNU Lesser General Public |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
6 * License as published by the Free Software Foundation; either |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
7 * version 2 of the License, or (at your option) any later version. |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
8 * |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
9 * This library is distributed in the hope that it will be useful, |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
12 * Lesser General Public License for more details. |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
13 * |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
14 * You should have received a copy of the GNU Lesser General Public |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
15 * License along with this library; if not, write to the Free Software |
3036
0b546eab515d
Update licensing information: The FSF changed postal address.
diego
parents:
2979
diff
changeset
|
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
17 */ |
2967 | 18 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
19 #include "../dsputil.h" |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
20 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
21 #include "gcc_fixes.h" |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
22 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
23 #include "dsputil_altivec.h" |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
24 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
25 #define PUT_OP_U8_ALTIVEC(d, s, dst) d = s |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
26 #define AVG_OP_U8_ALTIVEC(d, s, dst) d = vec_avg(dst, s) |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
27 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
28 #define OP_U8_ALTIVEC PUT_OP_U8_ALTIVEC |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
29 #define PREFIX_h264_chroma_mc8_altivec put_h264_chroma_mc8_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
30 #define PREFIX_h264_chroma_mc8_num altivec_put_h264_chroma_mc8_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
31 #define PREFIX_h264_qpel16_h_lowpass_altivec put_h264_qpel16_h_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
32 #define PREFIX_h264_qpel16_h_lowpass_num altivec_put_h264_qpel16_h_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
33 #define PREFIX_h264_qpel16_v_lowpass_altivec put_h264_qpel16_v_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
34 #define PREFIX_h264_qpel16_v_lowpass_num altivec_put_h264_qpel16_v_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
35 #define PREFIX_h264_qpel16_hv_lowpass_altivec put_h264_qpel16_hv_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
36 #define PREFIX_h264_qpel16_hv_lowpass_num altivec_put_h264_qpel16_hv_lowpass_num |
3577 | 37 #include "h264_template_altivec.c" |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
38 #undef OP_U8_ALTIVEC |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
39 #undef PREFIX_h264_chroma_mc8_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
40 #undef PREFIX_h264_chroma_mc8_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
41 #undef PREFIX_h264_qpel16_h_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
42 #undef PREFIX_h264_qpel16_h_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
43 #undef PREFIX_h264_qpel16_v_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
44 #undef PREFIX_h264_qpel16_v_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
45 #undef PREFIX_h264_qpel16_hv_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
46 #undef PREFIX_h264_qpel16_hv_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
47 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
48 #define OP_U8_ALTIVEC AVG_OP_U8_ALTIVEC |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
49 #define PREFIX_h264_chroma_mc8_altivec avg_h264_chroma_mc8_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
50 #define PREFIX_h264_chroma_mc8_num altivec_avg_h264_chroma_mc8_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
51 #define PREFIX_h264_qpel16_h_lowpass_altivec avg_h264_qpel16_h_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
52 #define PREFIX_h264_qpel16_h_lowpass_num altivec_avg_h264_qpel16_h_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
53 #define PREFIX_h264_qpel16_v_lowpass_altivec avg_h264_qpel16_v_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
54 #define PREFIX_h264_qpel16_v_lowpass_num altivec_avg_h264_qpel16_v_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
55 #define PREFIX_h264_qpel16_hv_lowpass_altivec avg_h264_qpel16_hv_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
56 #define PREFIX_h264_qpel16_hv_lowpass_num altivec_avg_h264_qpel16_hv_lowpass_num |
3577 | 57 #include "h264_template_altivec.c" |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
58 #undef OP_U8_ALTIVEC |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
59 #undef PREFIX_h264_chroma_mc8_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
60 #undef PREFIX_h264_chroma_mc8_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
61 #undef PREFIX_h264_qpel16_h_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
62 #undef PREFIX_h264_qpel16_h_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
63 #undef PREFIX_h264_qpel16_v_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
64 #undef PREFIX_h264_qpel16_v_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
65 #undef PREFIX_h264_qpel16_hv_lowpass_altivec |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
66 #undef PREFIX_h264_qpel16_hv_lowpass_num |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
67 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
68 #define H264_MC(OPNAME, SIZE, CODETYPE) \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
69 static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## CODETYPE (uint8_t *dst, uint8_t *src, int stride){\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
70 OPNAME ## pixels ## SIZE ## _ ## CODETYPE(dst, src, stride, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
71 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
72 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
73 static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){ \ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
74 DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
75 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
76 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src, half, stride, stride, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
77 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
78 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
79 static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
80 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(dst, src, stride, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
81 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
82 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
83 static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
84 DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
85 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
86 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src+1, half, stride, stride, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
87 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
88 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
89 static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
90 DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
91 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
92 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src, half, stride, stride, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
93 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
94 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
95 static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
96 OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(dst, src, stride, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
97 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
98 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
99 static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
100 DECLARE_ALIGNED_16(uint8_t, half[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
101 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(half, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
102 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src+stride, half, stride, stride, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
103 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
104 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
105 static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
106 DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
107 DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
108 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
109 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
110 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
111 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
112 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
113 static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
114 DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
115 DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
116 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
117 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
118 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
119 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
120 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
121 static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
122 DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
123 DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
124 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
125 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
126 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
127 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
128 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
129 static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
130 DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
131 DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
132 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
133 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
134 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
135 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
136 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
137 static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
138 DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
139 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(dst, tmp, src, stride, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
140 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
141 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
142 static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
143 DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
144 DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
145 DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
146 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
147 put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
148 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfHV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
149 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
150 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
151 static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
152 DECLARE_ALIGNED_16(uint8_t, halfH[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
153 DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
154 DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
155 put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
156 put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
157 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfHV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
158 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
159 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
160 static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
161 DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
162 DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
163 DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
164 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
165 put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
166 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfV, halfHV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
167 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
168 \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
169 static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## CODETYPE(uint8_t *dst, uint8_t *src, int stride){\ |
3544
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
170 DECLARE_ALIGNED_16(uint8_t, halfV[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
171 DECLARE_ALIGNED_16(uint8_t, halfHV[SIZE*SIZE]);\ |
8bb61d9a2c40
avoid alignment hacks, luckly gcc does the right thing on arches different from x86
lu_zero
parents:
3337
diff
changeset
|
172 DECLARE_ALIGNED_16(int16_t, tmp[SIZE*(SIZE+8)]);\ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
173 put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
174 put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
175 OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfV, halfHV, stride, SIZE, SIZE);\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
176 }\ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
177 |
3667
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
178 /* this code assume that stride % 16 == 0 */ |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
179 void put_no_rnd_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, int h, int x, int y) { |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
180 signed int ABCD[4] __attribute__((aligned(16))) = |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
181 {((8 - x) * (8 - y)), |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
182 ((x) * (8 - y)), |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
183 ((8 - x) * (y)), |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
184 ((x) * (y))}; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
185 register int i; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
186 vector unsigned char fperm; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
187 const vector signed int vABCD = vec_ld(0, ABCD); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
188 const vector signed short vA = vec_splat((vector signed short)vABCD, 1); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
189 const vector signed short vB = vec_splat((vector signed short)vABCD, 3); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
190 const vector signed short vC = vec_splat((vector signed short)vABCD, 5); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
191 const vector signed short vD = vec_splat((vector signed short)vABCD, 7); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
192 const vector signed int vzero = vec_splat_s32(0); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
193 const vector signed short v28ss = vec_sub(vec_sl(vec_splat_s16(1),vec_splat_u16(5)),vec_splat_s16(4)); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
194 const vector unsigned short v6us = vec_splat_u16(6); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
195 register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
196 register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
197 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
198 vector unsigned char vsrcAuc, vsrcBuc, vsrcperm0, vsrcperm1; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
199 vector unsigned char vsrc0uc, vsrc1uc; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
200 vector signed short vsrc0ssH, vsrc1ssH; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
201 vector unsigned char vsrcCuc, vsrc2uc, vsrc3uc; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
202 vector signed short vsrc2ssH, vsrc3ssH, psum; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
203 vector unsigned char vdst, ppsum, vfdst, fsum; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
204 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
205 if (((unsigned long)dst) % 16 == 0) { |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
206 fperm = (vector unsigned char)AVV(0x10, 0x11, 0x12, 0x13, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
207 0x14, 0x15, 0x16, 0x17, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
208 0x08, 0x09, 0x0A, 0x0B, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
209 0x0C, 0x0D, 0x0E, 0x0F); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
210 } else { |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
211 fperm = (vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
212 0x04, 0x05, 0x06, 0x07, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
213 0x18, 0x19, 0x1A, 0x1B, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
214 0x1C, 0x1D, 0x1E, 0x1F); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
215 } |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
216 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
217 vsrcAuc = vec_ld(0, src); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
218 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
219 if (loadSecond) |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
220 vsrcBuc = vec_ld(16, src); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
221 vsrcperm0 = vec_lvsl(0, src); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
222 vsrcperm1 = vec_lvsl(1, src); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
223 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
224 vsrc0uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm0); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
225 if (reallyBadAlign) |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
226 vsrc1uc = vsrcBuc; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
227 else |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
228 vsrc1uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm1); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
229 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
230 vsrc0ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
231 (vector unsigned char)vsrc0uc); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
232 vsrc1ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
233 (vector unsigned char)vsrc1uc); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
234 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
235 if (!loadSecond) {// -> !reallyBadAlign |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
236 for (i = 0 ; i < h ; i++) { |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
237 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
238 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
239 vsrcCuc = vec_ld(stride + 0, src); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
240 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
241 vsrc2uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
242 vsrc3uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
243 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
244 vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
245 (vector unsigned char)vsrc2uc); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
246 vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
247 (vector unsigned char)vsrc3uc); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
248 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
249 psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0)); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
250 psum = vec_mladd(vB, vsrc1ssH, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
251 psum = vec_mladd(vC, vsrc2ssH, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
252 psum = vec_mladd(vD, vsrc3ssH, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
253 psum = vec_add(v28ss, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
254 psum = vec_sra(psum, v6us); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
255 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
256 vdst = vec_ld(0, dst); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
257 ppsum = (vector unsigned char)vec_packsu(psum, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
258 fsum = vec_perm(vdst, ppsum, fperm); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
259 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
260 vec_st(fsum, 0, dst); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
261 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
262 vsrc0ssH = vsrc2ssH; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
263 vsrc1ssH = vsrc3ssH; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
264 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
265 dst += stride; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
266 src += stride; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
267 } |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
268 } else { |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
269 vector unsigned char vsrcDuc; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
270 for (i = 0 ; i < h ; i++) { |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
271 vsrcCuc = vec_ld(stride + 0, src); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
272 vsrcDuc = vec_ld(stride + 16, src); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
273 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
274 vsrc2uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
275 if (reallyBadAlign) |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
276 vsrc3uc = vsrcDuc; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
277 else |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
278 vsrc3uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
279 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
280 vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
281 (vector unsigned char)vsrc2uc); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
282 vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero, |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
283 (vector unsigned char)vsrc3uc); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
284 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
285 psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0)); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
286 psum = vec_mladd(vB, vsrc1ssH, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
287 psum = vec_mladd(vC, vsrc2ssH, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
288 psum = vec_mladd(vD, vsrc3ssH, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
289 psum = vec_add(v28ss, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
290 psum = vec_sr(psum, v6us); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
291 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
292 vdst = vec_ld(0, dst); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
293 ppsum = (vector unsigned char)vec_pack(psum, psum); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
294 fsum = vec_perm(vdst, ppsum, fperm); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
295 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
296 vec_st(fsum, 0, dst); |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
297 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
298 vsrc0ssH = vsrc2ssH; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
299 vsrc1ssH = vsrc3ssH; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
300 |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
301 dst += stride; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
302 src += stride; |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
303 } |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
304 } |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
305 } |
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
306 |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
307 static inline void put_pixels16_l2_altivec( uint8_t * dst, const uint8_t * src1, |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
308 const uint8_t * src2, int dst_stride, |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
309 int src_stride1, int h) |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
310 { |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
311 int i; |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
312 vector unsigned char a, b, d, tmp1, tmp2, mask, mask_, edges, align; |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
313 |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
314 mask_ = vec_lvsl(0, src2); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
315 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
316 for (i = 0; i < h; i++) { |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
317 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
318 tmp1 = vec_ld(i * src_stride1, src1); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
319 mask = vec_lvsl(i * src_stride1, src1); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
320 tmp2 = vec_ld(i * src_stride1 + 15, src1); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
321 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
322 a = vec_perm(tmp1, tmp2, mask); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
323 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
324 tmp1 = vec_ld(i * 16, src2); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
325 tmp2 = vec_ld(i * 16 + 15, src2); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
326 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
327 b = vec_perm(tmp1, tmp2, mask_); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
328 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
329 tmp1 = vec_ld(0, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
330 mask = vec_lvsl(0, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
331 tmp2 = vec_ld(15, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
332 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
333 d = vec_avg(a, b); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
334 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
335 edges = vec_perm(tmp2, tmp1, mask); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
336 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
337 align = vec_lvsr(0, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
338 |
3659
dd55fb216497
Proper fix for the corner case that would have been corrected before, praise&blame to me and exg in equal shares
lu_zero
parents:
3658
diff
changeset
|
339 tmp2 = vec_perm(d, edges, align); |
3658 | 340 tmp1 = vec_perm(edges, d, align); |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
341 |
3659
dd55fb216497
Proper fix for the corner case that would have been corrected before, praise&blame to me and exg in equal shares
lu_zero
parents:
3658
diff
changeset
|
342 vec_st(tmp2, 15, dst); |
3583
562758eaf7bf
10l, thanks to Emanuele Giaquinta <exg@gentoo.org> for testing and finding the issue
lu_zero
parents:
3577
diff
changeset
|
343 vec_st(tmp1, 0 , dst); |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
344 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
345 dst += dst_stride; |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
346 } |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
347 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
348 |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
349 static inline void avg_pixels16_l2_altivec( uint8_t * dst, const uint8_t * src1, |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
350 const uint8_t * src2, int dst_stride, |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
351 int src_stride1, int h) |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
352 { |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
353 int i; |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
354 vector unsigned char a, b, d, tmp1, tmp2, mask, mask_, edges, align; |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
355 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
356 mask_ = vec_lvsl(0, src2); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
357 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
358 for (i = 0; i < h; i++) { |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
359 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
360 tmp1 = vec_ld(i * src_stride1, src1); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
361 mask = vec_lvsl(i * src_stride1, src1); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
362 tmp2 = vec_ld(i * src_stride1 + 15, src1); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
363 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
364 a = vec_perm(tmp1, tmp2, mask); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
365 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
366 tmp1 = vec_ld(i * 16, src2); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
367 tmp2 = vec_ld(i * 16 + 15, src2); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
368 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
369 b = vec_perm(tmp1, tmp2, mask_); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
370 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
371 tmp1 = vec_ld(0, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
372 mask = vec_lvsl(0, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
373 tmp2 = vec_ld(15, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
374 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
375 d = vec_avg(vec_perm(tmp1, tmp2, mask), vec_avg(a, b)); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
376 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
377 edges = vec_perm(tmp2, tmp1, mask); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
378 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
379 align = vec_lvsr(0, dst); |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
380 |
3659
dd55fb216497
Proper fix for the corner case that would have been corrected before, praise&blame to me and exg in equal shares
lu_zero
parents:
3658
diff
changeset
|
381 tmp2 = vec_perm(d, edges, align); |
3658 | 382 tmp1 = vec_perm(edges, d, align); |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
383 |
3659
dd55fb216497
Proper fix for the corner case that would have been corrected before, praise&blame to me and exg in equal shares
lu_zero
parents:
3658
diff
changeset
|
384 vec_st(tmp2, 15, dst); |
3583
562758eaf7bf
10l, thanks to Emanuele Giaquinta <exg@gentoo.org> for testing and finding the issue
lu_zero
parents:
3577
diff
changeset
|
385 vec_st(tmp1, 0 , dst); |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
386 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
387 dst += dst_stride; |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
388 } |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
389 } |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
390 |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
391 /* Implemented but could be faster |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
392 #define put_pixels16_l2_altivec(d,s1,s2,ds,s1s,h) put_pixels16_l2(d,s1,s2,ds,s1s,16,h) |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
393 #define avg_pixels16_l2_altivec(d,s1,s2,ds,s1s,h) avg_pixels16_l2(d,s1,s2,ds,s1s,16,h) |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
394 */ |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
395 |
3337
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
396 H264_MC(put_, 16, altivec) |
bec1eb6d3746
put_pixels16_l2_altivec and avg_pixels16_l2_altivec
lu_zero
parents:
3089
diff
changeset
|
397 H264_MC(avg_, 16, altivec) |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
398 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
399 void dsputil_h264_init_ppc(DSPContext* c, AVCodecContext *avctx) { |
2967 | 400 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
401 #ifdef HAVE_ALTIVEC |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
402 if (has_altivec()) { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
403 c->put_h264_chroma_pixels_tab[0] = put_h264_chroma_mc8_altivec; |
3667
56e60411527c
AltiVec version of put_no_rnd_h264_chroma_pixels_tab[0] (slightly changed version of put_h264_chroma_pixels_tab[0])
kostya
parents:
3659
diff
changeset
|
404 c->put_no_rnd_h264_chroma_pixels_tab[0] = put_no_rnd_h264_chroma_mc8_altivec; |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
405 c->avg_h264_chroma_pixels_tab[0] = avg_h264_chroma_mc8_altivec; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
406 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
407 #define dspfunc(PFX, IDX, NUM) \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
408 c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
409 c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
410 c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
411 c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
412 c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
413 c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
414 c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
415 c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
416 c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
417 c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
418 c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
419 c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
420 c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
421 c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
422 c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_altivec; \ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
423 c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_altivec |
2967 | 424 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
425 dspfunc(put_h264_qpel, 0, 16); |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
426 dspfunc(avg_h264_qpel, 0, 16); |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
427 #undef dspfunc |
2967 | 428 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
429 } else |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
430 #endif /* HAVE_ALTIVEC */ |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
431 { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
432 // Non-AltiVec PPC optimisations |
2967 | 433 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
434 // ... pending ... |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
435 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
436 } |