annotate alpha/motion_est_mvi_asm.S @ 11089:68fa9b353337 libavcodec

Check direct_8x8_inference_flag.
author michael
date Sun, 07 Feb 2010 15:12:37 +0000
parents 6f1b210e58d1
children 58c2da0a371b
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
1 /*
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
2 * Alpha optimized DSP utils
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
3 * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
4 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3130
diff changeset
5 * This file is part of FFmpeg.
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3130
diff changeset
6 *
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3130
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
3130
9361d38a3af0 Change license header to LGPL for consistency.
mellum
parents: 2967
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
9361d38a3af0 Change license header to LGPL for consistency.
mellum
parents: 2967
diff changeset
9 * 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: 3130
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
11 *
3947
c8c591fe26f8 Change license headers to say 'FFmpeg' instead of 'this program/this library'
diego
parents: 3130
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
3130
9361d38a3af0 Change license header to LGPL for consistency.
mellum
parents: 2967
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
9361d38a3af0 Change license header to LGPL for consistency.
mellum
parents: 2967
diff changeset
15 * Lesser General Public License for more details.
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
16 *
3130
9361d38a3af0 Change license header to LGPL for consistency.
mellum
parents: 2967
diff changeset
17 * 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: 3130
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
3130
9361d38a3af0 Change license header to LGPL for consistency.
mellum
parents: 2967
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
20 */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
21
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
22 #include "regdef.h"
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
23
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
24 /* Some nicer register names. */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
25 #define ta t10
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
26 #define tb t11
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
27 #define tc t12
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
28 #define td AT
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
29 /* Danger: these overlap with the argument list and the return value */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
30 #define te a5
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
31 #define tf a4
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
32 #define tg a3
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
33 #define th v0
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2122
diff changeset
34
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
35 .set noat
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
36 .set noreorder
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
37 .arch pca56
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
38 .text
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
39
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
40 /*****************************************************************************
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
41 * int pix_abs16x16_mvi_asm(uint8_t *pix1, uint8_t *pix2, int line_size)
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
42 *
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
43 * This code is written with a pca56 in mind. For ev6, one should
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
44 * really take the increased latency of 3 cycles for MVI instructions
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
45 * into account.
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
46 *
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
47 * It is important to keep the loading and first use of a register as
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
48 * far apart as possible, because if a register is accessed before it
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
49 * has been fetched from memory, the CPU will stall.
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
50 */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
51 .align 4
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
52 .globl pix_abs16x16_mvi_asm
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
53 .ent pix_abs16x16_mvi_asm
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
54 pix_abs16x16_mvi_asm:
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
55 .frame sp, 0, ra, 0
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
56 .prologue 0
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
57
8590
7a463923ecd1 Change semantic of CONFIG_*, HAVE_* and ARCH_*.
aurel
parents: 4512
diff changeset
58 #if CONFIG_GPROF
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
59 lda AT, _mcount
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
60 jsr AT, (AT), _mcount
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
61 #endif
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
62
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
63 and a2, 7, t0
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
64 clr v0
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
65 beq t0, $aligned
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
66 .align 4
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
67 $unaligned:
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
68 /* Registers:
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
69 line 0:
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
70 t0: left_u -> left lo -> left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
71 t1: mid
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
72 t2: right_u -> right hi -> right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
73 t3: ref left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
74 t4: ref right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
75 line 1:
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
76 t5: left_u -> left lo -> left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
77 t6: mid
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
78 t7: right_u -> right hi -> right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
79 t8: ref left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
80 t9: ref right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
81 temp:
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
82 ta: left hi
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
83 tb: right lo
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
84 tc: error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
85 td: error right */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
86
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
87 /* load line 0 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
88 ldq_u t0, 0(a2) # left_u
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
89 ldq_u t1, 8(a2) # mid
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
90 ldq_u t2, 16(a2) # right_u
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
91 ldq t3, 0(a1) # ref left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
92 ldq t4, 8(a1) # ref right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
93 addq a1, a3, a1 # pix1
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
94 addq a2, a3, a2 # pix2
2967
ef2149182f1c COSMETICS: Remove all trailing whitespace.
diego
parents: 2122
diff changeset
95 /* load line 1 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
96 ldq_u t5, 0(a2) # left_u
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
97 ldq_u t6, 8(a2) # mid
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
98 ldq_u t7, 16(a2) # right_u
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
99 ldq t8, 0(a1) # ref left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
100 ldq t9, 8(a1) # ref right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
101 addq a1, a3, a1 # pix1
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
102 addq a2, a3, a2 # pix2
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
103 /* calc line 0 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
104 extql t0, a2, t0 # left lo
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
105 extqh t1, a2, ta # left hi
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
106 extql t1, a2, tb # right lo
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
107 or t0, ta, t0 # left
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
108 extqh t2, a2, t2 # right hi
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
109 perr t3, t0, tc # error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
110 or t2, tb, t2 # right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
111 perr t4, t2, td # error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
112 addq v0, tc, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
113 addq v0, td, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
114 /* calc line 1 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
115 extql t5, a2, t5 # left lo
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
116 extqh t6, a2, ta # left hi
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
117 extql t6, a2, tb # right lo
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
118 or t5, ta, t5 # left
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
119 extqh t7, a2, t7 # right hi
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
120 perr t8, t5, tc # error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
121 or t7, tb, t7 # right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
122 perr t9, t7, td # error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
123 addq v0, tc, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
124 addq v0, td, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
125 /* loop */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
126 subq a4, 2, a4 # h -= 2
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
127 bne a4, $unaligned
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
128 ret
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
129
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
130 .align 4
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
131 $aligned:
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
132 /* load line 0 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
133 ldq t0, 0(a2) # left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
134 ldq t1, 8(a2) # right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
135 addq a2, a3, a2 # pix2
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
136 ldq t2, 0(a1) # ref left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
137 ldq t3, 8(a1) # ref right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
138 addq a1, a3, a1 # pix1
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
139 /* load line 1 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
140 ldq t4, 0(a2) # left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
141 ldq t5, 8(a2) # right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
142 addq a2, a3, a2 # pix2
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
143 ldq t6, 0(a1) # ref left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
144 ldq t7, 8(a1) # ref right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
145 addq a1, a3, a1 # pix1
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
146 /* load line 2 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
147 ldq t8, 0(a2) # left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
148 ldq t9, 8(a2) # right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
149 addq a2, a3, a2 # pix2
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
150 ldq ta, 0(a1) # ref left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
151 ldq tb, 8(a1) # ref right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
152 addq a1, a3, a1 # pix1
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
153 /* load line 3 */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
154 ldq tc, 0(a2) # left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
155 ldq td, 8(a2) # right
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
156 addq a2, a3, a2 # pix2
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
157 ldq te, 0(a1) # ref left
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
158 ldq a0, 8(a1) # ref right
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
159 /* calc line 0 */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
160 perr t0, t2, t0 # error left
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
161 addq a1, a3, a1 # pix1
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
162 perr t1, t3, t1 # error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
163 addq v0, t0, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
164 /* calc line 1 */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
165 perr t4, t6, t0 # error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
166 addq v0, t1, v0 # add error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
167 perr t5, t7, t1 # error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
168 addq v0, t0, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
169 /* calc line 2 */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
170 perr t8, ta, t0 # error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
171 addq v0, t1, v0 # add error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
172 perr t9, tb, t1 # error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
173 addq v0, t0, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
174 /* calc line 3 */
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
175 perr tc, te, t0 # error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
176 addq v0, t1, v0 # add error right
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
177 perr td, a0, t1 # error right
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
178 addq v0, t0, v0 # add error left
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
179 addq v0, t1, v0 # add error right
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
180 /* loop */
8625
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
181 subq a4, 4, a4 # h -= 4
6f1b210e58d1 Alpha: fix pix_abs16
mru
parents: 8590
diff changeset
182 bne a4, $aligned
705
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
183 ret
107a56aa74f5 Add Alpha assembly for pix_abs16x16. Optimized for pca56, no large win
mellum
parents:
diff changeset
184 .end pix_abs16x16_mvi_asm