annotate libmpeg2/motion_comp_arm_s.S @ 34107:4529cc476aa6

Do not use long type, it makes the code more platform dependent for no good reason.
author reimar
date Thu, 13 Oct 2011 16:49:21 +0000
parents da2271c341ee
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
23236
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
1 @ motion_comp_arm_s.S
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
2 @ Copyright (C) 2004 AGAWA Koji <i (AT) atty (DOT) jp>
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
3 @
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
4 @ This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
5 @ See http://libmpeg2.sourceforge.net/ for updates.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
6 @
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
7 @ mpeg2dec is free software; you can redistribute it and/or modify
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
8 @ it under the terms of the GNU General Public License as published by
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
9 @ the Free Software Foundation; either version 2 of the License, or
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
10 @ (at your option) any later version.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
11 @
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
12 @ mpeg2dec is distributed in the hope that it will be useful,
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
13 @ but WITHOUT ANY WARRANTY; without even the implied warranty of
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
14 @ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
15 @ GNU General Public License for more details.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
16 @
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
17 @ You should have received a copy of the GNU General Public License
27572
da2271c341ee Update internal libmpeg2 copy to version 0.5.1.
diego
parents: 26283
diff changeset
18 @ along with mpeg2dec; if not, write to the Free Software
23236
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
19 @ Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
20
27572
da2271c341ee Update internal libmpeg2 copy to version 0.5.1.
diego
parents: 26283
diff changeset
21
23236
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
22 .text
26283
62abac0c8637 cosmetics: Fix some typos and trailing whitespace in local changes.
diego
parents: 23236
diff changeset
23
23236
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
24 @ ----------------------------------------------------------------
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
25 .align
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
26 .global MC_put_o_16_arm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
27 MC_put_o_16_arm:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
28 @@ void func(uint8_t * dest, const uint8_t * ref, int stride, int height)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
29 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
30 stmfd sp!, {r4-r11, lr} @ R14 is also called LR
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
31 and r4, r1, #3
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
32 adr r5, MC_put_o_16_arm_align_jt
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
33 add r5, r5, r4, lsl #2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
34 ldr pc, [r5]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
35
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
36 MC_put_o_16_arm_align0:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
37 ldmia r1, {r4-r7}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
38 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
39 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
40 stmia r0, {r4-r7}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
41 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
42 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
43 bne MC_put_o_16_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
44 ldmfd sp!, {r4-r11, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
45
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
46 .macro PROC shift
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
47 ldmia r1, {r4-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
48 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
49 mov r9, r4, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
50 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
51 mov r10, r5, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
52 orr r9, r9, r5, lsl #(32-\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
53 mov r11, r6, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
54 orr r10, r10, r6, lsl #(32-\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
55 mov r12, r7, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
56 orr r11, r11, r7, lsl #(32-\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
57 orr r12, r12, r8, lsl #(32-\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
58 stmia r0, {r9-r12}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
59 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
60 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
61 .endm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
62
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
63 MC_put_o_16_arm_align1:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
64 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
65 1: PROC(8)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
66 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
67 ldmfd sp!, {r4-r11, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
68 MC_put_o_16_arm_align2:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
69 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
70 1: PROC(16)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
71 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
72 ldmfd sp!, {r4-r11, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
73 MC_put_o_16_arm_align3:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
74 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
75 1: PROC(24)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
76 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
77 ldmfd sp!, {r4-r11, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
78 MC_put_o_16_arm_align_jt:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
79 .word MC_put_o_16_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
80 .word MC_put_o_16_arm_align1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
81 .word MC_put_o_16_arm_align2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
82 .word MC_put_o_16_arm_align3
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
83
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
84 @ ----------------------------------------------------------------
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
85 .align
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
86 .global MC_put_o_8_arm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
87 MC_put_o_8_arm:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
88 @@ void func(uint8_t * dest, const uint8_t * ref, int stride, int height)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
89 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
90 stmfd sp!, {r4-r10, lr} @ R14 is also called LR
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
91 and r4, r1, #3
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
92 adr r5, MC_put_o_8_arm_align_jt
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
93 add r5, r5, r4, lsl #2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
94 ldr pc, [r5]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
95 MC_put_o_8_arm_align0:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
96 ldmia r1, {r4-r5}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
97 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
98 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
99 stmia r0, {r4-r5}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
100 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
101 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
102 bne MC_put_o_8_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
103 ldmfd sp!, {r4-r10, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
104
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
105 .macro PROC8 shift
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
106 ldmia r1, {r4-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
107 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
108 mov r9, r4, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
109 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
110 mov r10, r5, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
111 orr r9, r9, r5, lsl #(32-\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
112 orr r10, r10, r6, lsl #(32-\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
113 stmia r0, {r9-r10}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
114 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
115 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
116 .endm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
117
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
118 MC_put_o_8_arm_align1:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
119 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
120 1: PROC8(8)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
121 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
122 ldmfd sp!, {r4-r10, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
123
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
124 MC_put_o_8_arm_align2:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
125 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
126 1: PROC8(16)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
127 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
128 ldmfd sp!, {r4-r10, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
129
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
130 MC_put_o_8_arm_align3:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
131 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
132 1: PROC8(24)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
133 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
134 ldmfd sp!, {r4-r10, pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
135
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
136 MC_put_o_8_arm_align_jt:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
137 .word MC_put_o_8_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
138 .word MC_put_o_8_arm_align1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
139 .word MC_put_o_8_arm_align2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
140 .word MC_put_o_8_arm_align3
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
141
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
142 @ ----------------------------------------------------------------
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
143 .macro AVG_PW rW1, rW2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
144 mov \rW2, \rW2, lsl #24
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
145 orr \rW2, \rW2, \rW1, lsr #8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
146 eor r9, \rW1, \rW2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
147 and \rW2, \rW1, \rW2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
148 and r10, r9, r12
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
149 add \rW2, \rW2, r10, lsr #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
150 and r10, r9, r11
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
151 add \rW2, \rW2, r10
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
152 .endm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
153
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
154 .align
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
155 .global MC_put_x_16_arm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
156 MC_put_x_16_arm:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
157 @@ void func(uint8_t * dest, const uint8_t * ref, int stride, int height)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
158 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
159 stmfd sp!, {r4-r11,lr} @ R14 is also called LR
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
160 and r4, r1, #3
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
161 adr r5, MC_put_x_16_arm_align_jt
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
162 ldr r11, [r5]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
163 mvn r12, r11
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
164 add r5, r5, r4, lsl #2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
165 ldr pc, [r5, #4]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
166
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
167 .macro ADJ_ALIGN_QW shift, R0, R1, R2, R3, R4
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
168 mov \R0, \R0, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
169 orr \R0, \R0, \R1, lsl #(32 - \shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
170 mov \R1, \R1, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
171 orr \R1, \R1, \R2, lsl #(32 - \shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
172 mov \R2, \R2, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
173 orr \R2, \R2, \R3, lsl #(32 - \shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
174 mov \R3, \R3, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
175 orr \R3, \R3, \R4, lsl #(32 - \shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
176 mov \R4, \R4, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
177 @ and \R4, \R4, #0xFF
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
178 .endm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
179
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
180 MC_put_x_16_arm_align0:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
181 ldmia r1, {r4-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
182 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
183 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
184 AVG_PW r7, r8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
185 AVG_PW r6, r7
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
186 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
187 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
188 stmia r0, {r5-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
189 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
190 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
191 bne MC_put_x_16_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
192 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
193 MC_put_x_16_arm_align1:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
194 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
195 1: ldmia r1, {r4-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
196 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
197 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
198 ADJ_ALIGN_QW 8, r4, r5, r6, r7, r8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
199 AVG_PW r7, r8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
200 AVG_PW r6, r7
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
201 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
202 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
203 stmia r0, {r5-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
204 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
205 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
206 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
207 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
208 MC_put_x_16_arm_align2:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
209 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
210 1: ldmia r1, {r4-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
211 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
212 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
213 ADJ_ALIGN_QW 16, r4, r5, r6, r7, r8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
214 AVG_PW r7, r8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
215 AVG_PW r6, r7
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
216 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
217 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
218 stmia r0, {r5-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
219 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
220 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
221 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
222 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
223 MC_put_x_16_arm_align3:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
224 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
225 1: ldmia r1, {r4-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
226 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
227 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
228 ADJ_ALIGN_QW 24, r4, r5, r6, r7, r8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
229 AVG_PW r7, r8
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
230 AVG_PW r6, r7
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
231 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
232 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
233 stmia r0, {r5-r8}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
234 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
235 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
236 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
237 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
238 MC_put_x_16_arm_align_jt:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
239 .word 0x01010101
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
240 .word MC_put_x_16_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
241 .word MC_put_x_16_arm_align1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
242 .word MC_put_x_16_arm_align2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
243 .word MC_put_x_16_arm_align3
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
244
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
245 @ ----------------------------------------------------------------
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
246 .align
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
247 .global MC_put_x_8_arm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
248 MC_put_x_8_arm:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
249 @@ void func(uint8_t * dest, const uint8_t * ref, int stride, int height)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
250 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
251 stmfd sp!, {r4-r11,lr} @ R14 is also called LR
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
252 and r4, r1, #3
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
253 adr r5, MC_put_x_8_arm_align_jt
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
254 ldr r11, [r5]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
255 mvn r12, r11
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
256 add r5, r5, r4, lsl #2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
257 ldr pc, [r5, #4]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
258
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
259 .macro ADJ_ALIGN_DW shift, R0, R1, R2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
260 mov \R0, \R0, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
261 orr \R0, \R0, \R1, lsl #(32 - \shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
262 mov \R1, \R1, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
263 orr \R1, \R1, \R2, lsl #(32 - \shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
264 mov \R2, \R2, lsr #(\shift)
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
265 @ and \R4, \R4, #0xFF
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
266 .endm
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
267
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
268 MC_put_x_8_arm_align0:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
269 ldmia r1, {r4-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
270 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
271 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
272 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
273 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
274 stmia r0, {r5-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
275 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
276 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
277 bne MC_put_x_8_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
278 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
279 MC_put_x_8_arm_align1:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
280 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
281 1: ldmia r1, {r4-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
282 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
283 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
284 ADJ_ALIGN_DW 8, r4, r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
285 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
286 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
287 stmia r0, {r5-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
288 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
289 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
290 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
291 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
292 MC_put_x_8_arm_align2:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
293 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
294 1: ldmia r1, {r4-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
295 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
296 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
297 ADJ_ALIGN_DW 16, r4, r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
298 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
299 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
300 stmia r0, {r5-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
301 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
302 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
303 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
304 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
305 MC_put_x_8_arm_align3:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
306 and r1, r1, #0xFFFFFFFC
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
307 1: ldmia r1, {r4-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
308 add r1, r1, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
309 pld [r1]
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
310 ADJ_ALIGN_DW 24, r4, r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
311 AVG_PW r5, r6
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
312 AVG_PW r4, r5
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
313 stmia r0, {r5-r6}
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
314 subs r3, r3, #1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
315 add r0, r0, r2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
316 bne 1b
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
317 ldmfd sp!, {r4-r11,pc} @@ update PC with LR content.
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
318 MC_put_x_8_arm_align_jt:
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
319 .word 0x01010101
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
320 .word MC_put_x_8_arm_align0
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
321 .word MC_put_x_8_arm_align1
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
322 .word MC_put_x_8_arm_align2
f0ddd02aec27 iWMMXt-accelerated DCT and motion compensation for ARM processors
gpoirier
parents:
diff changeset
323 .word MC_put_x_8_arm_align3