annotate ppc/h264_template_altivec.c @ 4468:5455cb5a759c libavcodec

fix decoding of takethat.mp3
author michael
date Sat, 03 Feb 2007 12:42:12 +0000
parents c8c591fe26f8
children 41cabe79ba25
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
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 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3577
diff changeset
4 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3577
diff changeset
5 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3577
diff changeset
6 * FFmpeg is free software; you can redistribute it and/or
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
7 * 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
8 * License as published by the Free Software Foundation; either
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3577
diff changeset
9 * version 2.1 of the License, or (at your option) any later version.
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
10 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3577
diff changeset
11 * FFmpeg is distributed in the hope that it will be useful,
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
12 * 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
13 * 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
14 * 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
15 *
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3577
diff changeset
17 * License along with FFmpeg; if not, write to the Free Software
3036
0b546eab515d Update licensing information: The FSF changed postal address.
diego
parents: 2967
diff changeset
18 * 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
19 */
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 /* this code assume that stride % 16 == 0 */
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
22 void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, int stride, int h, int x, int y) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
23 POWERPC_PERF_DECLARE(PREFIX_h264_chroma_mc8_num, 1);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
24 signed int ABCD[4] __attribute__((aligned(16))) =
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
25 {((8 - x) * (8 - y)),
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
26 ((x) * (8 - y)),
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
27 ((8 - x) * (y)),
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
28 ((x) * (y))};
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
29 register int i;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
30 vector unsigned char fperm;
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
31 const vector signed int vABCD = vec_ld(0, ABCD);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
32 const vector signed short vA = vec_splat((vector signed short)vABCD, 1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
33 const vector signed short vB = vec_splat((vector signed short)vABCD, 3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
34 const vector signed short vC = vec_splat((vector signed short)vABCD, 5);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
35 const vector signed short vD = vec_splat((vector signed short)vABCD, 7);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
36 const vector signed int vzero = vec_splat_s32(0);
3137
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
37 const vector signed short v32ss = vec_sl(vec_splat_s16(1),vec_splat_u16(5));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
38 const vector unsigned short v6us = vec_splat_u16(6);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
39 register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
40 register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
41
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
42 vector unsigned char vsrcAuc, vsrcBuc, vsrcperm0, vsrcperm1;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
43 vector unsigned char vsrc0uc, vsrc1uc;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
44 vector signed short vsrc0ssH, vsrc1ssH;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
45 vector unsigned char vsrcCuc, vsrc2uc, vsrc3uc;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
46 vector signed short vsrc2ssH, vsrc3ssH, psum;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
47 vector unsigned char vdst, ppsum, vfdst, fsum;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
48
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
49 POWERPC_PERF_START_COUNT(PREFIX_h264_chroma_mc8_num, 1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
50
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
51 if (((unsigned long)dst) % 16 == 0) {
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
52 fperm = (vector unsigned char)AVV(0x10, 0x11, 0x12, 0x13,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
53 0x14, 0x15, 0x16, 0x17,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
54 0x08, 0x09, 0x0A, 0x0B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
55 0x0C, 0x0D, 0x0E, 0x0F);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
56 } else {
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
57 fperm = (vector unsigned char)AVV(0x00, 0x01, 0x02, 0x03,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
58 0x04, 0x05, 0x06, 0x07,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
59 0x18, 0x19, 0x1A, 0x1B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
60 0x1C, 0x1D, 0x1E, 0x1F);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
61 }
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
62
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
63 vsrcAuc = vec_ld(0, src);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
64
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
65 if (loadSecond)
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
66 vsrcBuc = vec_ld(16, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
67 vsrcperm0 = vec_lvsl(0, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
68 vsrcperm1 = vec_lvsl(1, src);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
69
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
70 vsrc0uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
71 if (reallyBadAlign)
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
72 vsrc1uc = vsrcBuc;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
73 else
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
74 vsrc1uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
75
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
76 vsrc0ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
77 (vector unsigned char)vsrc0uc);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
78 vsrc1ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
79 (vector unsigned char)vsrc1uc);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
80
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
81 if (!loadSecond) {// -> !reallyBadAlign
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
82 for (i = 0 ; i < h ; i++) {
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
83
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
84
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
85 vsrcCuc = vec_ld(stride + 0, src);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
86
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
87 vsrc2uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
88 vsrc3uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
89
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
90 vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
91 (vector unsigned char)vsrc2uc);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
92 vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
93 (vector unsigned char)vsrc3uc);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
94
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
95 psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0));
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
96 psum = vec_mladd(vB, vsrc1ssH, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
97 psum = vec_mladd(vC, vsrc2ssH, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
98 psum = vec_mladd(vD, vsrc3ssH, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
99 psum = vec_add(v32ss, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
100 psum = vec_sra(psum, v6us);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
101
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
102 vdst = vec_ld(0, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
103 ppsum = (vector unsigned char)vec_packsu(psum, psum);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
104 vfdst = vec_perm(vdst, ppsum, fperm);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
105
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
106 OP_U8_ALTIVEC(fsum, vfdst, vdst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
107
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
108 vec_st(fsum, 0, dst);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
109
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
110 vsrc0ssH = vsrc2ssH;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
111 vsrc1ssH = vsrc3ssH;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
112
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
113 dst += stride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
114 src += stride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
115 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
116 } else {
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
117 vector unsigned char vsrcDuc;
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
118 for (i = 0 ; i < h ; i++) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
119 vsrcCuc = vec_ld(stride + 0, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
120 vsrcDuc = vec_ld(stride + 16, src);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
121
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
122 vsrc2uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
123 if (reallyBadAlign)
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
124 vsrc3uc = vsrcDuc;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
125 else
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
126 vsrc3uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
127
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
128 vsrc2ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
129 (vector unsigned char)vsrc2uc);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
130 vsrc3ssH = (vector signed short)vec_mergeh((vector unsigned char)vzero,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
131 (vector unsigned char)vsrc3uc);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
132
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
133 psum = vec_mladd(vA, vsrc0ssH, vec_splat_s16(0));
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
134 psum = vec_mladd(vB, vsrc1ssH, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
135 psum = vec_mladd(vC, vsrc2ssH, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
136 psum = vec_mladd(vD, vsrc3ssH, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
137 psum = vec_add(v32ss, psum);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
138 psum = vec_sr(psum, v6us);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
139
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
140 vdst = vec_ld(0, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
141 ppsum = (vector unsigned char)vec_pack(psum, psum);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
142 vfdst = vec_perm(vdst, ppsum, fperm);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
143
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
144 OP_U8_ALTIVEC(fsum, vfdst, vdst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
145
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
146 vec_st(fsum, 0, dst);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
147
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
148 vsrc0ssH = vsrc2ssH;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
149 vsrc1ssH = vsrc3ssH;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
150
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
151 dst += stride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
152 src += stride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
153 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
154 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
155 POWERPC_PERF_STOP_COUNT(PREFIX_h264_chroma_mc8_num, 1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
156 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
157
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
158 /* this code assume stride % 16 == 0 */
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
159 static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
160 POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_h_lowpass_num, 1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
161 register int i;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
162
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
163 const vector signed int vzero = vec_splat_s32(0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
164 const vector unsigned char permM2 = vec_lvsl(-2, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
165 const vector unsigned char permM1 = vec_lvsl(-1, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
166 const vector unsigned char permP0 = vec_lvsl(+0, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
167 const vector unsigned char permP1 = vec_lvsl(+1, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
168 const vector unsigned char permP2 = vec_lvsl(+2, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
169 const vector unsigned char permP3 = vec_lvsl(+3, src);
3137
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
170 const vector signed short v5ss = vec_splat_s16(5);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
171 const vector unsigned short v5us = vec_splat_u16(5);
3137
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
172 const vector signed short v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
173 const vector signed short v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
174 const vector unsigned char dstperm = vec_lvsr(0, dst);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
175 const vector unsigned char neg1 =
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
176 (const vector unsigned char) vec_splat_s8(-1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
177
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
178 const vector unsigned char dstmask =
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
179 vec_perm((const vector unsigned char)vzero,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
180 neg1, dstperm);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
181
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
182 vector unsigned char srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
183
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
184 register int align = ((((unsigned long)src) - 2) % 16);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
185
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
186 vector signed short srcP0A, srcP0B, srcP1A, srcP1B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
187 srcP2A, srcP2B, srcP3A, srcP3B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
188 srcM1A, srcM1B, srcM2A, srcM2B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
189 sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
190 pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
191 psumA, psumB, sumA, sumB;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
192
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
193 vector unsigned char sum, dst1, dst2, vdst, fsum,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
194 rsum, fdst1, fdst2;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
195
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
196 POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
197
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
198 for (i = 0 ; i < 16 ; i ++) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
199 vector unsigned char srcR1 = vec_ld(-2, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
200 vector unsigned char srcR2 = vec_ld(14, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
201
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
202 switch (align) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
203 default: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
204 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
205 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
206 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
207 srcP1 = vec_perm(srcR1, srcR2, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
208 srcP2 = vec_perm(srcR1, srcR2, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
209 srcP3 = vec_perm(srcR1, srcR2, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
210 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
211 case 11: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
212 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
213 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
214 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
215 srcP1 = vec_perm(srcR1, srcR2, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
216 srcP2 = vec_perm(srcR1, srcR2, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
217 srcP3 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
218 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
219 case 12: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
220 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
221 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
222 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
223 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
224 srcP1 = vec_perm(srcR1, srcR2, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
225 srcP2 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
226 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
227 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
228 case 13: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
229 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
230 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
231 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
232 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
233 srcP1 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
234 srcP2 = vec_perm(srcR2, srcR3, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
235 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
236 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
237 case 14: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
238 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
239 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
240 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
241 srcP0 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
242 srcP1 = vec_perm(srcR2, srcR3, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
243 srcP2 = vec_perm(srcR2, srcR3, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
244 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
245 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
246 case 15: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
247 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
248 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
249 srcM1 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
250 srcP0 = vec_perm(srcR2, srcR3, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
251 srcP1 = vec_perm(srcR2, srcR3, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
252 srcP2 = vec_perm(srcR2, srcR3, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
253 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
254 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
255 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
256
3350
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
257 srcP0A = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
258 vec_mergeh((vector unsigned char)vzero, srcP0);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
259 srcP0B = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
260 vec_mergel((vector unsigned char)vzero, srcP0);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
261 srcP1A = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
262 vec_mergeh((vector unsigned char)vzero, srcP1);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
263 srcP1B = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
264 vec_mergel((vector unsigned char)vzero, srcP1);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
265
3350
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
266 srcP2A = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
267 vec_mergeh((vector unsigned char)vzero, srcP2);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
268 srcP2B = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
269 vec_mergel((vector unsigned char)vzero, srcP2);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
270 srcP3A = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
271 vec_mergeh((vector unsigned char)vzero, srcP3);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
272 srcP3B = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
273 vec_mergel((vector unsigned char)vzero, srcP3);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
274
3350
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
275 srcM1A = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
276 vec_mergeh((vector unsigned char)vzero, srcM1);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
277 srcM1B = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
278 vec_mergel((vector unsigned char)vzero, srcM1);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
279 srcM2A = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
280 vec_mergeh((vector unsigned char)vzero, srcM2);
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
281 srcM2B = (vector signed short)
eff63ac2b545 Make gcc-3 happy again, thanks to Olivier Castan <castan.o@free.fr> for pointing the issue
lu_zero
parents: 3346
diff changeset
282 vec_mergel((vector unsigned char)vzero, srcM2);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
283
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
284 sum1A = vec_adds(srcP0A, srcP1A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
285 sum1B = vec_adds(srcP0B, srcP1B);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
286 sum2A = vec_adds(srcM1A, srcP2A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
287 sum2B = vec_adds(srcM1B, srcP2B);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
288 sum3A = vec_adds(srcM2A, srcP3A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
289 sum3B = vec_adds(srcM2B, srcP3B);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
290
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
291 pp1A = vec_mladd(sum1A, v20ss, v16ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
292 pp1B = vec_mladd(sum1B, v20ss, v16ss);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
293
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
294 pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
295 pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
296
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
297 pp3A = vec_add(sum3A, pp1A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
298 pp3B = vec_add(sum3B, pp1B);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
299
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
300 psumA = vec_sub(pp3A, pp2A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
301 psumB = vec_sub(pp3B, pp2B);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
302
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
303 sumA = vec_sra(psumA, v5us);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
304 sumB = vec_sra(psumB, v5us);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
305
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
306 sum = vec_packsu(sumA, sumB);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
307
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
308 dst1 = vec_ld(0, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
309 dst2 = vec_ld(16, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
310 vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
311
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
312 OP_U8_ALTIVEC(fsum, sum, vdst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
313
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
314 rsum = vec_perm(fsum, fsum, dstperm);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
315 fdst1 = vec_sel(dst1, rsum, dstmask);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
316 fdst2 = vec_sel(rsum, dst2, dstmask);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
317
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
318 vec_st(fdst1, 0, dst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
319 vec_st(fdst2, 16, dst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
320
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
321 src += srcStride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
322 dst += dstStride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
323 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
324 POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_h_lowpass_num, 1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
325 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
326
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
327 /* this code assume stride % 16 == 0 */
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
328 static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t * dst, uint8_t * src, int dstStride, int srcStride) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
329 POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_v_lowpass_num, 1);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
330
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
331 register int i;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
332
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
333 const vector signed int vzero = vec_splat_s32(0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
334 const vector unsigned char perm = vec_lvsl(0, src);
3137
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
335 const vector signed short v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
336 const vector unsigned short v5us = vec_splat_u16(5);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
337 const vector signed short v5ss = vec_splat_s16(5);
3137
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
338 const vector signed short v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
339 const vector unsigned char dstperm = vec_lvsr(0, dst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
340 const vector unsigned char neg1 = (const vector unsigned char)vec_splat_s8(-1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
341 const vector unsigned char dstmask = vec_perm((const vector unsigned char)vzero, neg1, dstperm);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
342
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
343 uint8_t *srcbis = src - (srcStride * 2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
344
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
345 const vector unsigned char srcM2a = vec_ld(0, srcbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
346 const vector unsigned char srcM2b = vec_ld(16, srcbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
347 const vector unsigned char srcM2 = vec_perm(srcM2a, srcM2b, perm);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
348 // srcbis += srcStride;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
349 const vector unsigned char srcM1a = vec_ld(0, srcbis += srcStride);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
350 const vector unsigned char srcM1b = vec_ld(16, srcbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
351 const vector unsigned char srcM1 = vec_perm(srcM1a, srcM1b, perm);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
352 // srcbis += srcStride;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
353 const vector unsigned char srcP0a = vec_ld(0, srcbis += srcStride);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
354 const vector unsigned char srcP0b = vec_ld(16, srcbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
355 const vector unsigned char srcP0 = vec_perm(srcP0a, srcP0b, perm);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
356 // srcbis += srcStride;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
357 const vector unsigned char srcP1a = vec_ld(0, srcbis += srcStride);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
358 const vector unsigned char srcP1b = vec_ld(16, srcbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
359 const vector unsigned char srcP1 = vec_perm(srcP1a, srcP1b, perm);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
360 // srcbis += srcStride;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
361 const vector unsigned char srcP2a = vec_ld(0, srcbis += srcStride);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
362 const vector unsigned char srcP2b = vec_ld(16, srcbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
363 const vector unsigned char srcP2 = vec_perm(srcP2a, srcP2b, perm);
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
364 // srcbis += srcStride;
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
365
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
366 vector signed short srcM2ssA = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
367 vec_mergeh((vector unsigned char)vzero, srcM2);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
368 vector signed short srcM2ssB = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
369 vec_mergel((vector unsigned char)vzero, srcM2);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
370 vector signed short srcM1ssA = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
371 vec_mergeh((vector unsigned char)vzero, srcM1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
372 vector signed short srcM1ssB = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
373 vec_mergel((vector unsigned char)vzero, srcM1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
374 vector signed short srcP0ssA = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
375 vec_mergeh((vector unsigned char)vzero, srcP0);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
376 vector signed short srcP0ssB = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
377 vec_mergel((vector unsigned char)vzero, srcP0);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
378 vector signed short srcP1ssA = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
379 vec_mergeh((vector unsigned char)vzero, srcP1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
380 vector signed short srcP1ssB = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
381 vec_mergel((vector unsigned char)vzero, srcP1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
382 vector signed short srcP2ssA = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
383 vec_mergeh((vector unsigned char)vzero, srcP2);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
384 vector signed short srcP2ssB = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
385 vec_mergel((vector unsigned char)vzero, srcP2);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
386
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
387 vector signed short pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
388 psumA, psumB, sumA, sumB,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
389 srcP3ssA, srcP3ssB,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
390 sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
391
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
392 vector unsigned char sum, dst1, dst2, vdst, fsum, rsum, fdst1, fdst2,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
393 srcP3a, srcP3b, srcP3;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
394
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
395 POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
396
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
397 for (i = 0 ; i < 16 ; i++) {
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
398 srcP3a = vec_ld(0, srcbis += srcStride);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
399 srcP3b = vec_ld(16, srcbis);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
400 srcP3 = vec_perm(srcP3a, srcP3b, perm);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
401 srcP3ssA = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
402 vec_mergeh((vector unsigned char)vzero, srcP3);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
403 srcP3ssB = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
404 vec_mergel((vector unsigned char)vzero, srcP3);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
405 // srcbis += srcStride;
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
406
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
407 sum1A = vec_adds(srcP0ssA, srcP1ssA);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
408 sum1B = vec_adds(srcP0ssB, srcP1ssB);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
409 sum2A = vec_adds(srcM1ssA, srcP2ssA);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
410 sum2B = vec_adds(srcM1ssB, srcP2ssB);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
411 sum3A = vec_adds(srcM2ssA, srcP3ssA);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
412 sum3B = vec_adds(srcM2ssB, srcP3ssB);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
413
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
414 srcM2ssA = srcM1ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
415 srcM2ssB = srcM1ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
416 srcM1ssA = srcP0ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
417 srcM1ssB = srcP0ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
418 srcP0ssA = srcP1ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
419 srcP0ssB = srcP1ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
420 srcP1ssA = srcP2ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
421 srcP1ssB = srcP2ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
422 srcP2ssA = srcP3ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
423 srcP2ssB = srcP3ssB;
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
424
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
425 pp1A = vec_mladd(sum1A, v20ss, v16ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
426 pp1B = vec_mladd(sum1B, v20ss, v16ss);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
427
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
428 pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
429 pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
430
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
431 pp3A = vec_add(sum3A, pp1A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
432 pp3B = vec_add(sum3B, pp1B);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
433
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
434 psumA = vec_sub(pp3A, pp2A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
435 psumB = vec_sub(pp3B, pp2B);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
436
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
437 sumA = vec_sra(psumA, v5us);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
438 sumB = vec_sra(psumB, v5us);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
439
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
440 sum = vec_packsu(sumA, sumB);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
441
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
442 dst1 = vec_ld(0, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
443 dst2 = vec_ld(16, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
444 vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
445
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
446 OP_U8_ALTIVEC(fsum, sum, vdst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
447
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
448 rsum = vec_perm(fsum, fsum, dstperm);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
449 fdst1 = vec_sel(dst1, rsum, dstmask);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
450 fdst2 = vec_sel(rsum, dst2, dstmask);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
451
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
452 vec_st(fdst1, 0, dst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
453 vec_st(fdst2, 16, dst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
454
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
455 dst += dstStride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
456 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
457 POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_v_lowpass_num, 1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
458 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
459
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
460 /* this code assume stride % 16 == 0 *and* tmp is properly aligned */
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
461 static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t * dst, int16_t * tmp, uint8_t * src, int dstStride, int tmpStride, int srcStride) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
462 POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_hv_lowpass_num, 1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
463 register int i;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
464 const vector signed int vzero = vec_splat_s32(0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
465 const vector unsigned char permM2 = vec_lvsl(-2, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
466 const vector unsigned char permM1 = vec_lvsl(-1, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
467 const vector unsigned char permP0 = vec_lvsl(+0, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
468 const vector unsigned char permP1 = vec_lvsl(+1, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
469 const vector unsigned char permP2 = vec_lvsl(+2, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
470 const vector unsigned char permP3 = vec_lvsl(+3, src);
3137
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
471 const vector signed short v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
472 const vector unsigned int v10ui = vec_splat_u32(10);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
473 const vector signed short v5ss = vec_splat_s16(5);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
474 const vector signed short v1ss = vec_splat_s16(1);
3137
876d6280faf7 Fix AltiVec H.264 decoding on FSF gcc.
diego
parents: 3036
diff changeset
475 const vector signed int v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9));
3153
7aa01243b4d3 use vec_splat_u32 to initialise a v16ui (patch by Likai Liu < liulk _at_ cs.bu.edu >)
aurel
parents: 3137
diff changeset
476 const vector unsigned int v16ui = vec_sl(vec_splat_u32(1),vec_splat_u32(4));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
477
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
478 register int align = ((((unsigned long)src) - 2) % 16);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
479
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
480 const vector unsigned char neg1 = (const vector unsigned char)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
481 vec_splat_s8(-1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
482
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
483 vector signed short srcP0A, srcP0B, srcP1A, srcP1B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
484 srcP2A, srcP2B, srcP3A, srcP3B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
485 srcM1A, srcM1B, srcM2A, srcM2B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
486 sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
487 pp1A, pp1B, pp2A, pp2B, psumA, psumB;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
488
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
489 const vector unsigned char dstperm = vec_lvsr(0, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
490
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
491 const vector unsigned char dstmask = vec_perm((const vector unsigned char)vzero, neg1, dstperm);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
492
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
493 const vector unsigned char mperm = (const vector unsigned char)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
494 AVV(0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
495 0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
496 int16_t *tmpbis = tmp;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
497
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
498 vector signed short tmpM1ssA, tmpM1ssB, tmpM2ssA, tmpM2ssB,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
499 tmpP0ssA, tmpP0ssB, tmpP1ssA, tmpP1ssB,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
500 tmpP2ssA, tmpP2ssB;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
501
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
502 vector signed int pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
503 pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
504 pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
505 ssumAe, ssumAo, ssumBe, ssumBo;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
506 vector unsigned char fsum, sumv, sum, dst1, dst2, vdst,
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
507 rsum, fdst1, fdst2;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
508 vector signed short ssume, ssumo;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
509
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
510 POWERPC_PERF_START_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
511 src -= (2 * srcStride);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
512 for (i = 0 ; i < 21 ; i ++) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
513 vector unsigned char srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
514 vector unsigned char srcR1 = vec_ld(-2, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
515 vector unsigned char srcR2 = vec_ld(14, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
516
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
517 switch (align) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
518 default: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
519 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
520 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
521 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
522 srcP1 = vec_perm(srcR1, srcR2, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
523 srcP2 = vec_perm(srcR1, srcR2, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
524 srcP3 = vec_perm(srcR1, srcR2, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
525 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
526 case 11: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
527 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
528 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
529 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
530 srcP1 = vec_perm(srcR1, srcR2, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
531 srcP2 = vec_perm(srcR1, srcR2, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
532 srcP3 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
533 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
534 case 12: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
535 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
536 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
537 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
538 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
539 srcP1 = vec_perm(srcR1, srcR2, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
540 srcP2 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
541 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
542 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
543 case 13: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
544 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
545 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
546 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
547 srcP0 = vec_perm(srcR1, srcR2, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
548 srcP1 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
549 srcP2 = vec_perm(srcR2, srcR3, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
550 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
551 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
552 case 14: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
553 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
554 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
555 srcM1 = vec_perm(srcR1, srcR2, permM1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
556 srcP0 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
557 srcP1 = vec_perm(srcR2, srcR3, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
558 srcP2 = vec_perm(srcR2, srcR3, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
559 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
560 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
561 case 15: {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
562 vector unsigned char srcR3 = vec_ld(30, src);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
563 srcM2 = vec_perm(srcR1, srcR2, permM2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
564 srcM1 = srcR2;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
565 srcP0 = vec_perm(srcR2, srcR3, permP0);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
566 srcP1 = vec_perm(srcR2, srcR3, permP1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
567 srcP2 = vec_perm(srcR2, srcR3, permP2);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
568 srcP3 = vec_perm(srcR2, srcR3, permP3);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
569 } break;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
570 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
571
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
572 srcP0A = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
573 vec_mergeh((vector unsigned char)vzero, srcP0);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
574 srcP0B = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
575 vec_mergel((vector unsigned char)vzero, srcP0);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
576 srcP1A = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
577 vec_mergeh((vector unsigned char)vzero, srcP1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
578 srcP1B = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
579 vec_mergel((vector unsigned char)vzero, srcP1);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
580
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
581 srcP2A = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
582 vec_mergeh((vector unsigned char)vzero, srcP2);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
583 srcP2B = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
584 vec_mergel((vector unsigned char)vzero, srcP2);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
585 srcP3A = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
586 vec_mergeh((vector unsigned char)vzero, srcP3);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
587 srcP3B = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
588 vec_mergel((vector unsigned char)vzero, srcP3);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
589
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
590 srcM1A = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
591 vec_mergeh((vector unsigned char)vzero, srcM1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
592 srcM1B = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
593 vec_mergel((vector unsigned char)vzero, srcM1);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
594 srcM2A = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
595 vec_mergeh((vector unsigned char)vzero, srcM2);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
596 srcM2B = (vector signed short)
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
597 vec_mergel((vector unsigned char)vzero, srcM2);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
598
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
599 sum1A = vec_adds(srcP0A, srcP1A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
600 sum1B = vec_adds(srcP0B, srcP1B);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
601 sum2A = vec_adds(srcM1A, srcP2A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
602 sum2B = vec_adds(srcM1B, srcP2B);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
603 sum3A = vec_adds(srcM2A, srcP3A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
604 sum3B = vec_adds(srcM2B, srcP3B);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
605
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
606 pp1A = vec_mladd(sum1A, v20ss, sum3A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
607 pp1B = vec_mladd(sum1B, v20ss, sum3B);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
608
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
609 pp2A = vec_mladd(sum2A, v5ss, (vector signed short)vzero);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
610 pp2B = vec_mladd(sum2B, v5ss, (vector signed short)vzero);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
611
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
612 psumA = vec_sub(pp1A, pp2A);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
613 psumB = vec_sub(pp1B, pp2B);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
614
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
615 vec_st(psumA, 0, tmp);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
616 vec_st(psumB, 16, tmp);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
617
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
618 src += srcStride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
619 tmp += tmpStride; /* int16_t*, and stride is 16, so it's OK here */
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
620 }
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
621
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
622 tmpM2ssA = vec_ld(0, tmpbis);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
623 tmpM2ssB = vec_ld(16, tmpbis);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
624 tmpbis += tmpStride;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
625 tmpM1ssA = vec_ld(0, tmpbis);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
626 tmpM1ssB = vec_ld(16, tmpbis);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
627 tmpbis += tmpStride;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
628 tmpP0ssA = vec_ld(0, tmpbis);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
629 tmpP0ssB = vec_ld(16, tmpbis);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
630 tmpbis += tmpStride;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
631 tmpP1ssA = vec_ld(0, tmpbis);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
632 tmpP1ssB = vec_ld(16, tmpbis);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
633 tmpbis += tmpStride;
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
634 tmpP2ssA = vec_ld(0, tmpbis);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
635 tmpP2ssB = vec_ld(16, tmpbis);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
636 tmpbis += tmpStride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
637
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
638 for (i = 0 ; i < 16 ; i++) {
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
639 const vector signed short tmpP3ssA = vec_ld(0, tmpbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
640 const vector signed short tmpP3ssB = vec_ld(16, tmpbis);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
641
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
642 const vector signed short sum1A = vec_adds(tmpP0ssA, tmpP1ssA);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
643 const vector signed short sum1B = vec_adds(tmpP0ssB, tmpP1ssB);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
644 const vector signed short sum2A = vec_adds(tmpM1ssA, tmpP2ssA);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
645 const vector signed short sum2B = vec_adds(tmpM1ssB, tmpP2ssB);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
646 const vector signed short sum3A = vec_adds(tmpM2ssA, tmpP3ssA);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
647 const vector signed short sum3B = vec_adds(tmpM2ssB, tmpP3ssB);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
648
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
649 tmpbis += tmpStride;
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
650
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
651 tmpM2ssA = tmpM1ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
652 tmpM2ssB = tmpM1ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
653 tmpM1ssA = tmpP0ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
654 tmpM1ssB = tmpP0ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
655 tmpP0ssA = tmpP1ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
656 tmpP0ssB = tmpP1ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
657 tmpP1ssA = tmpP2ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
658 tmpP1ssB = tmpP2ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
659 tmpP2ssA = tmpP3ssA;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
660 tmpP2ssB = tmpP3ssB;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
661
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
662 pp1Ae = vec_mule(sum1A, v20ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
663 pp1Ao = vec_mulo(sum1A, v20ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
664 pp1Be = vec_mule(sum1B, v20ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
665 pp1Bo = vec_mulo(sum1B, v20ss);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
666
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
667 pp2Ae = vec_mule(sum2A, v5ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
668 pp2Ao = vec_mulo(sum2A, v5ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
669 pp2Be = vec_mule(sum2B, v5ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
670 pp2Bo = vec_mulo(sum2B, v5ss);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
671
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
672 pp3Ae = vec_sra((vector signed int)sum3A, v16ui);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
673 pp3Ao = vec_mulo(sum3A, v1ss);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
674 pp3Be = vec_sra((vector signed int)sum3B, v16ui);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
675 pp3Bo = vec_mulo(sum3B, v1ss);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
676
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
677 pp1cAe = vec_add(pp1Ae, v512si);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
678 pp1cAo = vec_add(pp1Ao, v512si);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
679 pp1cBe = vec_add(pp1Be, v512si);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
680 pp1cBo = vec_add(pp1Bo, v512si);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
681
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
682 pp32Ae = vec_sub(pp3Ae, pp2Ae);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
683 pp32Ao = vec_sub(pp3Ao, pp2Ao);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
684 pp32Be = vec_sub(pp3Be, pp2Be);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
685 pp32Bo = vec_sub(pp3Bo, pp2Bo);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
686
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
687 sumAe = vec_add(pp1cAe, pp32Ae);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
688 sumAo = vec_add(pp1cAo, pp32Ao);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
689 sumBe = vec_add(pp1cBe, pp32Be);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
690 sumBo = vec_add(pp1cBo, pp32Bo);
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2236
diff changeset
691
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
692 ssumAe = vec_sra(sumAe, v10ui);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
693 ssumAo = vec_sra(sumAo, v10ui);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
694 ssumBe = vec_sra(sumBe, v10ui);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
695 ssumBo = vec_sra(sumBo, v10ui);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
696
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
697 ssume = vec_packs(ssumAe, ssumBe);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
698 ssumo = vec_packs(ssumAo, ssumBo);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
699
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
700 sumv = vec_packsu(ssume, ssumo);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
701 sum = vec_perm(sumv, sumv, mperm);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
702
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
703 dst1 = vec_ld(0, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
704 dst2 = vec_ld(16, dst);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
705 vdst = vec_perm(dst1, dst2, vec_lvsl(0, dst));
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
706
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
707 OP_U8_ALTIVEC(fsum, sum, vdst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
708
3346
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
709 rsum = vec_perm(fsum, fsum, dstperm);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
710 fdst1 = vec_sel(dst1, rsum, dstmask);
052765f11f1c Cosmetics: should not hurt performance, scream if are
lu_zero
parents: 3153
diff changeset
711 fdst2 = vec_sel(rsum, dst2, dstmask);
2236
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
712
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
713 vec_st(fdst1, 0, dst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
714 vec_st(fdst2, 16, dst);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
715
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
716 dst += dstStride;
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
717 }
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
718 POWERPC_PERF_STOP_COUNT(PREFIX_h264_qpel16_hv_lowpass_num, 1);
b0102ea621dd h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff changeset
719 }