Mercurial > libavcodec.hg
annotate ppc/h264_template_altivec.c @ 6065:180976fd652e libavcodec
10l do not load after the buffer...
author | lu_zero |
---|---|
date | Sun, 23 Dec 2007 15:54:21 +0000 |
parents | 62d040333d51 |
children | a82931915c54 |
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 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
21 //#define DEBUG_ALIGNMENT |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
22 #ifdef DEBUG_ALIGNMENT |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
23 #define ASSERT_ALIGNED(ptr) assert(((unsigned long)ptr&0x0000000F)); |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
24 #else |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
25 #define ASSERT_ALIGNED(ptr) ; |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
26 #endif |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
27 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
28 /* this code assume that stride % 16 == 0 */ |
6059 | 29 |
30 #define CHROMA_MC8_ALTIVEC_CORE \ | |
31 vsrc2ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc2uc);\ | |
32 vsrc3ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc3uc);\ | |
33 \ | |
6061 | 34 psum = vec_mladd(vA, vsrc0ssH, v32ss);\ |
6059 | 35 psum = vec_mladd(vB, vsrc1ssH, psum);\ |
36 psum = vec_mladd(vC, vsrc2ssH, psum);\ | |
37 psum = vec_mladd(vD, vsrc3ssH, psum);\ | |
38 psum = vec_sr(psum, v6us);\ | |
39 \ | |
40 vdst = vec_ld(0, dst);\ | |
41 ppsum = (vec_u8_t)vec_pack(psum, psum);\ | |
42 vfdst = vec_perm(vdst, ppsum, fperm);\ | |
43 \ | |
44 OP_U8_ALTIVEC(fsum, vfdst, vdst);\ | |
45 \ | |
46 vec_st(fsum, 0, dst);\ | |
47 \ | |
48 vsrc0ssH = vsrc2ssH;\ | |
49 vsrc1ssH = vsrc3ssH;\ | |
50 \ | |
51 dst += stride;\ | |
52 src += stride; | |
53 | |
6062
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
54 #define CHROMA_MC8_ALTIVEC_CORE_SIMPLE \ |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
55 \ |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
56 vsrc0ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc0uc);\ |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
57 vsrc1ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc1uc);\ |
6062
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
58 \ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
59 psum = vec_mladd(vA, vsrc0ssH, v32ss);\ |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
60 psum = vec_mladd(vE, vsrc1ssH, psum);\ |
6062
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
61 psum = vec_sr(psum, v6us);\ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
62 \ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
63 vdst = vec_ld(0, dst);\ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
64 ppsum = (vec_u8_t)vec_pack(psum, psum);\ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
65 vfdst = vec_perm(vdst, ppsum, fperm);\ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
66 \ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
67 OP_U8_ALTIVEC(fsum, vfdst, vdst);\ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
68 \ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
69 vec_st(fsum, 0, dst);\ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
70 \ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
71 dst += stride;\ |
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
72 src += stride; |
6059 | 73 |
6063 | 74 void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, uint8_t * src, |
75 int stride, int h, int x, int y) { | |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
76 POWERPC_PERF_DECLARE(PREFIX_h264_chroma_mc8_num, 1); |
5019
41cabe79ba25
use macro Use DECLARE_ALIGNED_16 to align stack-allocated variables
gpoirier
parents:
3947
diff
changeset
|
77 DECLARE_ALIGNED_16(signed int, ABCD[4]) = |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
78 {((8 - x) * (8 - y)), |
6058 | 79 (( x) * (8 - y)), |
80 ((8 - x) * ( y)), | |
81 (( x) * ( y))}; | |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
82 register int i; |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
83 vec_u8_t fperm; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
84 const vec_s32_t vABCD = vec_ld(0, ABCD); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
85 const vec_s16_t vA = vec_splat((vec_s16_t)vABCD, 1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
86 const vec_s16_t vB = vec_splat((vec_s16_t)vABCD, 3); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
87 const vec_s16_t vC = vec_splat((vec_s16_t)vABCD, 5); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
88 const vec_s16_t vD = vec_splat((vec_s16_t)vABCD, 7); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
89 LOAD_ZERO; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
90 const vec_s16_t v32ss = vec_sl(vec_splat_s16(1),vec_splat_u16(5)); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
91 const vec_u16_t v6us = vec_splat_u16(6); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
92 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
|
93 register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0; |
2967 | 94 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
95 vec_u8_t vsrcAuc, vsrcBuc, vsrcperm0, vsrcperm1; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
96 vec_u8_t vsrc0uc, vsrc1uc; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
97 vec_s16_t vsrc0ssH, vsrc1ssH; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
98 vec_u8_t vsrcCuc, vsrc2uc, vsrc3uc; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
99 vec_s16_t vsrc2ssH, vsrc3ssH, psum; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
100 vec_u8_t vdst, ppsum, vfdst, fsum; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
101 |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
102 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
|
103 |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
104 if (((unsigned long)dst) % 16 == 0) { |
6060 | 105 fperm = (vec_u8_t)AVV(0x10, 0x11, 0x12, 0x13, |
106 0x14, 0x15, 0x16, 0x17, | |
107 0x08, 0x09, 0x0A, 0x0B, | |
108 0x0C, 0x0D, 0x0E, 0x0F); | |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
109 } else { |
6060 | 110 fperm = (vec_u8_t)AVV(0x00, 0x01, 0x02, 0x03, |
111 0x04, 0x05, 0x06, 0x07, | |
112 0x18, 0x19, 0x1A, 0x1B, | |
113 0x1C, 0x1D, 0x1E, 0x1F); | |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
114 } |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
115 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
116 vsrcAuc = vec_ld(0, src); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
117 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
118 if (loadSecond) |
6060 | 119 vsrcBuc = vec_ld(16, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
120 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
|
121 vsrcperm1 = vec_lvsl(1, src); |
2967 | 122 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
123 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
|
124 if (reallyBadAlign) |
6060 | 125 vsrc1uc = vsrcBuc; |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
126 else |
6060 | 127 vsrc1uc = vec_perm(vsrcAuc, vsrcBuc, vsrcperm1); |
2967 | 128 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
129 vsrc0ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc0uc); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
130 vsrc1ssH = (vec_s16_t)vec_mergeh(zero_u8v,(vec_u8_t)vsrc1uc); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
131 |
6062
9d1590a4df90
Partially address issue299, no performance change apparently
lu_zero
parents:
6061
diff
changeset
|
132 if (ABCD[3]) { |
6063 | 133 if (!loadSecond) {// -> !reallyBadAlign |
134 for (i = 0 ; i < h ; i++) { | |
135 vsrcCuc = vec_ld(stride + 0, src); | |
136 vsrc2uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0); | |
137 vsrc3uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1); | |
2967 | 138 |
6063 | 139 CHROMA_MC8_ALTIVEC_CORE |
140 } | |
141 } else { | |
142 vec_u8_t vsrcDuc; | |
143 for (i = 0 ; i < h ; i++) { | |
144 vsrcCuc = vec_ld(stride + 0, src); | |
145 vsrcDuc = vec_ld(stride + 16, src); | |
146 vsrc2uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0); | |
147 if (reallyBadAlign) | |
148 vsrc3uc = vsrcDuc; | |
149 else | |
150 vsrc3uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1); | |
151 | |
152 CHROMA_MC8_ALTIVEC_CORE | |
153 } | |
6060 | 154 } |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
155 } else { |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
156 const vec_s16_t vE = vec_add(vB, vC); |
6065 | 157 if (ABCD[2]) { // x == 0 B == 0 |
6063 | 158 if (!loadSecond) {// -> !reallyBadAlign |
159 for (i = 0 ; i < h ; i++) { | |
160 vsrcCuc = vec_ld(stride + 0, src); | |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
161 vsrc1uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
162 CHROMA_MC8_ALTIVEC_CORE_SIMPLE |
2967 | 163 |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
164 vsrc0uc = vsrc1uc; |
6063 | 165 } |
166 } else { | |
167 vec_u8_t vsrcDuc; | |
168 for (i = 0 ; i < h ; i++) { | |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
169 vsrcCuc = vec_ld(0, src); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
170 vsrcDuc = vec_ld(15, src); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
171 vsrc1uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
172 CHROMA_MC8_ALTIVEC_CORE_SIMPLE |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
173 |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
174 vsrc0uc = vsrc1uc; |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
175 } |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
176 } |
6065 | 177 } else { // y == 0 C == 0 |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
178 if (!loadSecond) {// -> !reallyBadAlign |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
179 for (i = 0 ; i < h ; i++) { |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
180 vsrcCuc = vec_ld(0, src); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
181 vsrc0uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm0); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
182 vsrc1uc = vec_perm(vsrcCuc, vsrcCuc, vsrcperm1); |
6065 | 183 |
184 CHROMA_MC8_ALTIVEC_CORE_SIMPLE | |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
185 } |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
186 } else { |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
187 vec_u8_t vsrcDuc; |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
188 for (i = 0 ; i < h ; i++) { |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
189 vsrcCuc = vec_ld(0, src); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
190 vsrcDuc = vec_ld(15, src); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
191 vsrc0uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm0); |
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
diff
changeset
|
192 vsrc1uc = vec_perm(vsrcCuc, vsrcDuc, vsrcperm1); |
6065 | 193 |
194 CHROMA_MC8_ALTIVEC_CORE_SIMPLE | |
6063 | 195 } |
6060 | 196 } |
6064
62d040333d51
Add C/B == 0 cases, 2% slower on CELL but should address Issue299 eventually
lu_zero
parents:
6063
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 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
199 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
|
200 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
201 |
6059 | 202 #undef CHROMA_MC8_ALTIVEC_CORE |
203 | |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
204 /* 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
|
205 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
|
206 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
|
207 register int i; |
2967 | 208 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
209 LOAD_ZERO; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
210 const vec_u8_t permM2 = vec_lvsl(-2, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
211 const vec_u8_t permM1 = vec_lvsl(-1, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
212 const vec_u8_t permP0 = vec_lvsl(+0, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
213 const vec_u8_t permP1 = vec_lvsl(+1, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
214 const vec_u8_t permP2 = vec_lvsl(+2, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
215 const vec_u8_t permP3 = vec_lvsl(+3, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
216 const vec_s16_t v5ss = vec_splat_s16(5); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
217 const vec_u16_t v5us = vec_splat_u16(5); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
218 const vec_s16_t v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2)); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
219 const vec_s16_t v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4)); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
220 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
221 vec_u8_t 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
|
222 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
223 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
|
224 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
225 vec_s16_t srcP0A, srcP0B, srcP1A, srcP1B, |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
226 srcP2A, srcP2B, srcP3A, srcP3B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
227 srcM1A, srcM1B, srcM2A, srcM2B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
228 sum1A, sum1B, sum2A, sum2B, sum3A, sum3B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
229 pp1A, pp1B, pp2A, pp2B, pp3A, pp3B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
230 psumA, psumB, sumA, sumB; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
231 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
232 vec_u8_t sum, vdst, fsum; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
233 |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
234 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
|
235 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
236 for (i = 0 ; i < 16 ; i ++) { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
237 vec_u8_t srcR1 = vec_ld(-2, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
238 vec_u8_t srcR2 = vec_ld(14, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
239 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
240 switch (align) { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
241 default: { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
242 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
|
243 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
|
244 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
|
245 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
|
246 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
|
247 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
|
248 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
249 case 11: { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
250 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
|
251 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
|
252 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
|
253 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
|
254 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
|
255 srcP3 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
256 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
257 case 12: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
258 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
259 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
|
260 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
|
261 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
|
262 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
|
263 srcP2 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
264 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
|
265 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
266 case 13: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
267 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
268 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
|
269 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
|
270 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
|
271 srcP1 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
272 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
|
273 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
|
274 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
275 case 14: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
276 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
277 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
|
278 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
|
279 srcP0 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
280 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
|
281 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
|
282 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
|
283 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
284 case 15: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
285 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
286 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
|
287 srcM1 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
288 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
|
289 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
|
290 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
|
291 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
|
292 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
293 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
294 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
295 srcP0A = (vec_s16_t) vec_mergeh(zero_u8v, srcP0); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
296 srcP0B = (vec_s16_t) vec_mergel(zero_u8v, srcP0); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
297 srcP1A = (vec_s16_t) vec_mergeh(zero_u8v, srcP1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
298 srcP1B = (vec_s16_t) vec_mergel(zero_u8v, srcP1); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
299 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
300 srcP2A = (vec_s16_t) vec_mergeh(zero_u8v, srcP2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
301 srcP2B = (vec_s16_t) vec_mergel(zero_u8v, srcP2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
302 srcP3A = (vec_s16_t) vec_mergeh(zero_u8v, srcP3); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
303 srcP3B = (vec_s16_t) vec_mergel(zero_u8v, srcP3); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
304 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
305 srcM1A = (vec_s16_t) vec_mergeh(zero_u8v, srcM1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
306 srcM1B = (vec_s16_t) vec_mergel(zero_u8v, srcM1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
307 srcM2A = (vec_s16_t) vec_mergeh(zero_u8v, srcM2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
308 srcM2B = (vec_s16_t) vec_mergel(zero_u8v, srcM2); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
309 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
310 sum1A = vec_adds(srcP0A, srcP1A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
311 sum1B = vec_adds(srcP0B, srcP1B); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
312 sum2A = vec_adds(srcM1A, srcP2A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
313 sum2B = vec_adds(srcM1B, srcP2B); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
314 sum3A = vec_adds(srcM2A, srcP3A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
315 sum3B = vec_adds(srcM2B, srcP3B); |
2967 | 316 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
317 pp1A = vec_mladd(sum1A, v20ss, v16ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
318 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
|
319 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
320 pp2A = vec_mladd(sum2A, v5ss, zero_s16v); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
321 pp2B = vec_mladd(sum2B, v5ss, zero_s16v); |
2967 | 322 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
323 pp3A = vec_add(sum3A, pp1A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
324 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
|
325 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
326 psumA = vec_sub(pp3A, pp2A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
327 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
|
328 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
329 sumA = vec_sra(psumA, v5us); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
330 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
|
331 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
332 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
|
333 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
334 ASSERT_ALIGNED(dst); |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
335 vdst = vec_ld(0, dst); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
336 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
337 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
|
338 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
339 vec_st(fsum, 0, dst); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
340 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
341 src += srcStride; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
342 dst += dstStride; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
343 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
344 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
|
345 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
346 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
347 /* 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
|
348 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
|
349 POWERPC_PERF_DECLARE(PREFIX_h264_qpel16_v_lowpass_num, 1); |
2967 | 350 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
351 register int i; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
352 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
353 LOAD_ZERO; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
354 const vec_u8_t perm = vec_lvsl(0, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
355 const vec_s16_t v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2)); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
356 const vec_u16_t v5us = vec_splat_u16(5); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
357 const vec_s16_t v5ss = vec_splat_s16(5); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
358 const vec_s16_t v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4)); |
2967 | 359 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
360 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
|
361 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
362 const vec_u8_t srcM2a = vec_ld(0, srcbis); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
363 const vec_u8_t srcM2b = vec_ld(16, srcbis); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
364 const vec_u8_t srcM2 = vec_perm(srcM2a, srcM2b, perm); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
365 // srcbis += srcStride; |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
366 const vec_u8_t srcM1a = vec_ld(0, srcbis += srcStride); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
367 const vec_u8_t srcM1b = vec_ld(16, srcbis); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
368 const vec_u8_t srcM1 = vec_perm(srcM1a, srcM1b, perm); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
369 // srcbis += srcStride; |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
370 const vec_u8_t srcP0a = vec_ld(0, srcbis += srcStride); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
371 const vec_u8_t srcP0b = vec_ld(16, srcbis); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
372 const vec_u8_t srcP0 = vec_perm(srcP0a, srcP0b, perm); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
373 // srcbis += srcStride; |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
374 const vec_u8_t srcP1a = vec_ld(0, srcbis += srcStride); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
375 const vec_u8_t srcP1b = vec_ld(16, srcbis); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
376 const vec_u8_t srcP1 = vec_perm(srcP1a, srcP1b, perm); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
377 // srcbis += srcStride; |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
378 const vec_u8_t srcP2a = vec_ld(0, srcbis += srcStride); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
379 const vec_u8_t srcP2b = vec_ld(16, srcbis); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
380 const vec_u8_t srcP2 = vec_perm(srcP2a, srcP2b, perm); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
381 // srcbis += srcStride; |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
382 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
383 vec_s16_t srcM2ssA = (vec_s16_t) vec_mergeh(zero_u8v, srcM2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
384 vec_s16_t srcM2ssB = (vec_s16_t) vec_mergel(zero_u8v, srcM2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
385 vec_s16_t srcM1ssA = (vec_s16_t) vec_mergeh(zero_u8v, srcM1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
386 vec_s16_t srcM1ssB = (vec_s16_t) vec_mergel(zero_u8v, srcM1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
387 vec_s16_t srcP0ssA = (vec_s16_t) vec_mergeh(zero_u8v, srcP0); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
388 vec_s16_t srcP0ssB = (vec_s16_t) vec_mergel(zero_u8v, srcP0); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
389 vec_s16_t srcP1ssA = (vec_s16_t) vec_mergeh(zero_u8v, srcP1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
390 vec_s16_t srcP1ssB = (vec_s16_t) vec_mergel(zero_u8v, srcP1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
391 vec_s16_t srcP2ssA = (vec_s16_t) vec_mergeh(zero_u8v, srcP2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
392 vec_s16_t srcP2ssB = (vec_s16_t) vec_mergel(zero_u8v, srcP2); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
393 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
394 vec_s16_t pp1A, pp1B, pp2A, pp2B, pp3A, pp3B, |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
395 psumA, psumB, sumA, sumB, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
396 srcP3ssA, srcP3ssB, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
397 sum1A, sum1B, sum2A, sum2B, sum3A, sum3B; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
398 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
399 vec_u8_t sum, vdst, fsum, srcP3a, srcP3b, srcP3; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
400 |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
401 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
|
402 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
403 for (i = 0 ; i < 16 ; i++) { |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
404 srcP3a = vec_ld(0, srcbis += srcStride); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
405 srcP3b = vec_ld(16, srcbis); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
406 srcP3 = vec_perm(srcP3a, srcP3b, perm); |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
407 srcP3ssA = (vec_s16_t) vec_mergeh(zero_u8v, srcP3); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
408 srcP3ssB = (vec_s16_t) vec_mergel(zero_u8v, srcP3); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
409 // srcbis += srcStride; |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
410 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
411 sum1A = vec_adds(srcP0ssA, srcP1ssA); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
412 sum1B = vec_adds(srcP0ssB, srcP1ssB); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
413 sum2A = vec_adds(srcM1ssA, srcP2ssA); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
414 sum2B = vec_adds(srcM1ssB, srcP2ssB); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
415 sum3A = vec_adds(srcM2ssA, srcP3ssA); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
416 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
|
417 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
418 srcM2ssA = srcM1ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
419 srcM2ssB = srcM1ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
420 srcM1ssA = srcP0ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
421 srcM1ssB = srcP0ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
422 srcP0ssA = srcP1ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
423 srcP0ssB = srcP1ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
424 srcP1ssA = srcP2ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
425 srcP1ssB = srcP2ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
426 srcP2ssA = srcP3ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
427 srcP2ssB = srcP3ssB; |
2967 | 428 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
429 pp1A = vec_mladd(sum1A, v20ss, v16ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
430 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
|
431 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
432 pp2A = vec_mladd(sum2A, v5ss, zero_s16v); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
433 pp2B = vec_mladd(sum2B, v5ss, zero_s16v); |
2967 | 434 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
435 pp3A = vec_add(sum3A, pp1A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
436 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
|
437 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
438 psumA = vec_sub(pp3A, pp2A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
439 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
|
440 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
441 sumA = vec_sra(psumA, v5us); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
442 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
|
443 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
444 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
|
445 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
446 ASSERT_ALIGNED(dst); |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
447 vdst = vec_ld(0, dst); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
448 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
449 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
|
450 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
451 vec_st(fsum, 0, dst); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
452 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
453 dst += dstStride; |
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 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
|
456 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
457 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
458 /* 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
|
459 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
|
460 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
|
461 register int i; |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
462 LOAD_ZERO; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
463 const vec_u8_t permM2 = vec_lvsl(-2, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
464 const vec_u8_t permM1 = vec_lvsl(-1, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
465 const vec_u8_t permP0 = vec_lvsl(+0, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
466 const vec_u8_t permP1 = vec_lvsl(+1, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
467 const vec_u8_t permP2 = vec_lvsl(+2, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
468 const vec_u8_t permP3 = vec_lvsl(+3, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
469 const vec_s16_t v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2)); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
470 const vec_u32_t v10ui = vec_splat_u32(10); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
471 const vec_s16_t v5ss = vec_splat_s16(5); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
472 const vec_s16_t v1ss = vec_splat_s16(1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
473 const vec_s32_t v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9)); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
474 const vec_u32_t 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
|
475 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
476 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
|
477 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
478 vec_s16_t srcP0A, srcP0B, srcP1A, srcP1B, |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
479 srcP2A, srcP2B, srcP3A, srcP3B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
480 srcM1A, srcM1B, srcM2A, srcM2B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
481 sum1A, sum1B, sum2A, sum2B, sum3A, sum3B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
482 pp1A, pp1B, pp2A, pp2B, psumA, psumB; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
483 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
484 const vec_u8_t mperm = (const vec_u8_t) |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
485 AVV(0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
486 0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
487 int16_t *tmpbis = tmp; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
488 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
489 vec_s16_t tmpM1ssA, tmpM1ssB, tmpM2ssA, tmpM2ssB, |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
490 tmpP0ssA, tmpP0ssB, tmpP1ssA, tmpP1ssB, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
491 tmpP2ssA, tmpP2ssB; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
492 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
493 vec_s32_t pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo, |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
494 pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
495 pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo, |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
496 ssumAe, ssumAo, ssumBe, ssumBo; |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
497 vec_u8_t fsum, sumv, sum, vdst; |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
498 vec_s16_t ssume, ssumo; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
499 |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
500 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
|
501 src -= (2 * srcStride); |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
502 for (i = 0 ; i < 21 ; i ++) { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
503 vec_u8_t srcM2, srcM1, srcP0, srcP1, srcP2, srcP3; |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
504 vec_u8_t srcR1 = vec_ld(-2, src); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
505 vec_u8_t srcR2 = vec_ld(14, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
506 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
507 switch (align) { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
508 default: { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
509 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
|
510 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
|
511 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
|
512 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
|
513 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
|
514 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
|
515 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
516 case 11: { |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
517 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
|
518 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
|
519 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
|
520 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
|
521 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
|
522 srcP3 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
523 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
524 case 12: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
525 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
526 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
|
527 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
|
528 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
|
529 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
|
530 srcP2 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
531 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
|
532 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
533 case 13: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
534 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
535 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
|
536 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
|
537 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
|
538 srcP1 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
539 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
|
540 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
|
541 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
542 case 14: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
543 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
544 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
|
545 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
|
546 srcP0 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
547 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
|
548 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
|
549 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
|
550 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
551 case 15: { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
552 vec_u8_t srcR3 = vec_ld(30, src); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
553 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
|
554 srcM1 = srcR2; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
555 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
|
556 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
|
557 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
|
558 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
|
559 } break; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
560 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
561 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
562 srcP0A = (vec_s16_t) vec_mergeh(zero_u8v, srcP0); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
563 srcP0B = (vec_s16_t) vec_mergel(zero_u8v, srcP0); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
564 srcP1A = (vec_s16_t) vec_mergeh(zero_u8v, srcP1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
565 srcP1B = (vec_s16_t) vec_mergel(zero_u8v, srcP1); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
566 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
567 srcP2A = (vec_s16_t) vec_mergeh(zero_u8v, srcP2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
568 srcP2B = (vec_s16_t) vec_mergel(zero_u8v, srcP2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
569 srcP3A = (vec_s16_t) vec_mergeh(zero_u8v, srcP3); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
570 srcP3B = (vec_s16_t) vec_mergel(zero_u8v, srcP3); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
571 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
572 srcM1A = (vec_s16_t) vec_mergeh(zero_u8v, srcM1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
573 srcM1B = (vec_s16_t) vec_mergel(zero_u8v, srcM1); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
574 srcM2A = (vec_s16_t) vec_mergeh(zero_u8v, srcM2); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
575 srcM2B = (vec_s16_t) vec_mergel(zero_u8v, srcM2); |
2967 | 576 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
577 sum1A = vec_adds(srcP0A, srcP1A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
578 sum1B = vec_adds(srcP0B, srcP1B); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
579 sum2A = vec_adds(srcM1A, srcP2A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
580 sum2B = vec_adds(srcM1B, srcP2B); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
581 sum3A = vec_adds(srcM2A, srcP3A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
582 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
|
583 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
584 pp1A = vec_mladd(sum1A, v20ss, sum3A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
585 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
|
586 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
587 pp2A = vec_mladd(sum2A, v5ss, zero_s16v); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
588 pp2B = vec_mladd(sum2B, v5ss, zero_s16v); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
589 |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
590 psumA = vec_sub(pp1A, pp2A); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
591 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
|
592 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
593 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
|
594 vec_st(psumB, 16, tmp); |
2967 | 595 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
596 src += srcStride; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
597 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
|
598 } |
2967 | 599 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
600 tmpM2ssA = vec_ld(0, tmpbis); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
601 tmpM2ssB = vec_ld(16, tmpbis); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
602 tmpbis += tmpStride; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
603 tmpM1ssA = vec_ld(0, tmpbis); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
604 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
|
605 tmpbis += tmpStride; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
606 tmpP0ssA = vec_ld(0, tmpbis); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
607 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
|
608 tmpbis += tmpStride; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
609 tmpP1ssA = vec_ld(0, tmpbis); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
610 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
|
611 tmpbis += tmpStride; |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
612 tmpP2ssA = vec_ld(0, tmpbis); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
613 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
|
614 tmpbis += tmpStride; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
615 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
616 for (i = 0 ; i < 16 ; i++) { |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
617 const vec_s16_t tmpP3ssA = vec_ld(0, tmpbis); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
618 const vec_s16_t tmpP3ssB = 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
|
619 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
620 const vec_s16_t sum1A = vec_adds(tmpP0ssA, tmpP1ssA); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
621 const vec_s16_t sum1B = vec_adds(tmpP0ssB, tmpP1ssB); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
622 const vec_s16_t sum2A = vec_adds(tmpM1ssA, tmpP2ssA); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
623 const vec_s16_t sum2B = vec_adds(tmpM1ssB, tmpP2ssB); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
624 const vec_s16_t sum3A = vec_adds(tmpM2ssA, tmpP3ssA); |
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
625 const vec_s16_t sum3B = vec_adds(tmpM2ssB, tmpP3ssB); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
626 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
627 tmpbis += tmpStride; |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
628 |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
629 tmpM2ssA = tmpM1ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
630 tmpM2ssB = tmpM1ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
631 tmpM1ssA = tmpP0ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
632 tmpM1ssB = tmpP0ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
633 tmpP0ssA = tmpP1ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
634 tmpP0ssB = tmpP1ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
635 tmpP1ssA = tmpP2ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
636 tmpP1ssB = tmpP2ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
637 tmpP2ssA = tmpP3ssA; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
638 tmpP2ssB = tmpP3ssB; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
639 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
640 pp1Ae = vec_mule(sum1A, v20ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
641 pp1Ao = vec_mulo(sum1A, v20ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
642 pp1Be = vec_mule(sum1B, v20ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
643 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
|
644 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
645 pp2Ae = vec_mule(sum2A, v5ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
646 pp2Ao = vec_mulo(sum2A, v5ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
647 pp2Be = vec_mule(sum2B, v5ss); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
648 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
|
649 |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
650 pp3Ae = vec_sra((vec_s32_t)sum3A, v16ui); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
651 pp3Ao = vec_mulo(sum3A, v1ss); |
5530
cd266411b11a
use shorter types vec_"type" instead of the too long vector "type"
gpoirier
parents:
5019
diff
changeset
|
652 pp3Be = vec_sra((vec_s32_t)sum3B, v16ui); |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
653 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
|
654 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
655 pp1cAe = vec_add(pp1Ae, v512si); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
656 pp1cAo = vec_add(pp1Ao, v512si); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
657 pp1cBe = vec_add(pp1Be, v512si); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
658 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
|
659 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
660 pp32Ae = vec_sub(pp3Ae, pp2Ae); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
661 pp32Ao = vec_sub(pp3Ao, pp2Ao); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
662 pp32Be = vec_sub(pp3Be, pp2Be); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
663 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
|
664 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
665 sumAe = vec_add(pp1cAe, pp32Ae); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
666 sumAo = vec_add(pp1cAo, pp32Ao); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
667 sumBe = vec_add(pp1cBe, pp32Be); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
668 sumBo = vec_add(pp1cBo, pp32Bo); |
2967 | 669 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
670 ssumAe = vec_sra(sumAe, v10ui); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
671 ssumAo = vec_sra(sumAo, v10ui); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
672 ssumBe = vec_sra(sumBe, v10ui); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
673 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
|
674 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
675 ssume = vec_packs(ssumAe, ssumBe); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
676 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
|
677 |
3346
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
678 sumv = vec_packsu(ssume, ssumo); |
052765f11f1c
Cosmetics: should not hurt performance, scream if are
lu_zero
parents:
3153
diff
changeset
|
679 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
|
680 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
681 ASSERT_ALIGNED(dst); |
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
682 vdst = vec_ld(0, dst); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
683 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
684 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
|
685 |
5603
861eb234e6ba
remove alignment correction of the destination pointers in luma_16x6
gpoirier
parents:
5530
diff
changeset
|
686 vec_st(fsum, 0, dst); |
2236
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
687 |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
688 dst += dstStride; |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
689 } |
b0102ea621dd
h264 qpel mc, size 16 patch by (Romain Dolbeau <dolbeau at caps-entreprise dot com>)
michael
parents:
diff
changeset
|
690 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
|
691 } |