annotate alpha/dsputil_alpha_asm.S @ 1706:3ba5c493db6f libavcodec

motion vector vissualization improvements patch by (Wolfgang Hesseler <qv at multimediaware dot com>)
author michael
date Tue, 30 Dec 2003 02:12:12 +0000
parents acd1a808343b
children 503496800167
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
509
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
1 /*
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
2 * Alpha optimized DSP utils
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
3 * Copyright (c) 2002 Falk Hueffner <falk@debian.org>
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
4 *
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
5 * This program is free software; you can redistribute it and/or modify
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
6 * it under the terms of the GNU General Public License as published by
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
7 * the Free Software Foundation; either version 2 of the License, or
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
8 * (at your option) any later version.
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
9 *
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
10 * This program is distributed in the hope that it will be useful,
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
13 * GNU General Public License for more details.
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
14 *
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
15 * You should have received a copy of the GNU General Public License
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
16 * along with this program; if not, write to the Free Software
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
17 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
18 */
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
19
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
20 /*
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
21 * These functions are scheduled for pca56. They should work
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
22 * reasonably on ev6, though.
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
23 */
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
24
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
25 #include "regdef.h"
510
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
26 #ifdef HAVE_AV_CONFIG_H
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
27 #include "config.h"
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
28 #endif
509
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
29
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
30 /* Some nicer register names. */
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
31 #define ta t10
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
32 #define tb t11
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
33 #define tc t12
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
34 #define td AT
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
35 /* Danger: these overlap with the argument list and the return value */
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
36 #define te a5
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
37 #define tf a4
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
38 #define tg a3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
39 #define th v0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
40
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
41 .set noat
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
42 .set noreorder
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
43 .arch pca56
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
44 .text
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
45
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
46 /************************************************************************
511
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
47 * void put_pixels_axp_asm(uint8_t *block, const uint8_t *pixels,
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
48 * int line_size, int h)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
49 */
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
50 .align 6
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
51 .globl put_pixels_axp_asm
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
52 .ent put_pixels_axp_asm
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
53 put_pixels_axp_asm:
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
54 .frame sp, 0, ra
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
55 .prologue 0
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
56
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
57 #ifdef HAVE_GPROF
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
58 lda AT, _mcount
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
59 jsr AT, (AT), _mcount
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
60 #endif
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
61
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
62 and a1, 7, t0
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
63 beq t0, $aligned
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
64
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
65 .align 4
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
66 $unaligned:
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
67 ldq_u t0, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
68 ldq_u t1, 8(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
69 addq a1, a2, a1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
70 nop
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
71
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
72 ldq_u t2, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
73 ldq_u t3, 8(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
74 addq a1, a2, a1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
75 nop
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
76
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
77 ldq_u t4, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
78 ldq_u t5, 8(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
79 addq a1, a2, a1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
80 nop
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
81
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
82 ldq_u t6, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
83 ldq_u t7, 8(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
84 extql t0, a1, t0
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
85 addq a1, a2, a1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
86
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
87 extqh t1, a1, t1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
88 addq a0, a2, t8
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
89 extql t2, a1, t2
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
90 addq t8, a2, t9
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
91
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
92 extqh t3, a1, t3
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
93 addq t9, a2, ta
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
94 extql t4, a1, t4
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
95 or t0, t1, t0
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
96
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
97 extqh t5, a1, t5
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
98 or t2, t3, t2
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
99 extql t6, a1, t6
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
100 or t4, t5, t4
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
101
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
102 extqh t7, a1, t7
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
103 or t6, t7, t6
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
104 stq t0, 0(a0)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
105 stq t2, 0(t8)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
106
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
107 stq t4, 0(t9)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
108 subq a3, 4, a3
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
109 stq t6, 0(ta)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
110 addq ta, a2, a0
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
111
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
112 bne a3, $unaligned
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
113 ret
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
114
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
115 .align 4
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
116 $aligned:
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
117 ldq t0, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
118 addq a1, a2, a1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
119 ldq t1, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
120 addq a1, a2, a1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
121
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
122 ldq t2, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
123 addq a1, a2, a1
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
124 ldq t3, 0(a1)
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
125
668
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
126 addq a0, a2, t4
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
127 addq a1, a2, a1
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
128 addq t4, a2, t5
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
129 subq a3, 4, a3
511
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
130
668
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
131 stq t0, 0(a0)
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
132 addq t5, a2, t6
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
133 stq t1, 0(t4)
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
134 addq t6, a2, a0
511
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
135
668
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
136 stq t2, 0(t5)
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
137 stq t3, 0(t6)
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
138
acd1a808343b Fix for h % 8 == 4 in put_pixels_axp_asm.
mellum
parents: 511
diff changeset
139 bne a3, $aligned
511
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
140 ret
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
141 .end put_pixels_axp_asm
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
142
fa4425cf6b31 Assembly version of put_pixels. This is currently the function that
mellum
parents: 510
diff changeset
143 /************************************************************************
509
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
144 * void put_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels,
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
145 * int line_size)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
146 */
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
147 .align 6
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
148 .globl put_pixels_clamped_mvi_asm
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
149 .ent put_pixels_clamped_mvi_asm
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
150 put_pixels_clamped_mvi_asm:
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
151 .frame sp, 0, ra
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
152 .prologue 0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
153
510
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
154 #ifdef HAVE_GPROF
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
155 lda AT, _mcount
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
156 jsr AT, (AT), _mcount
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
157 #endif
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
158
509
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
159 lda t8, -1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
160 lda t9, 8 # loop counter
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
161 zap t8, 0xaa, t8 # 00ff00ff00ff00ff
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
162
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
163 .align 4
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
164 1: ldq t0, 0(a0)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
165 ldq t1, 8(a0)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
166 ldq t2, 16(a0)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
167 ldq t3, 24(a0)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
168
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
169 maxsw4 t0, zero, t0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
170 subq t9, 2, t9
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
171 maxsw4 t1, zero, t1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
172 lda a0, 32(a0)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
173
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
174 maxsw4 t2, zero, t2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
175 addq a1, a2, ta
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
176 maxsw4 t3, zero, t3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
177 minsw4 t0, t8, t0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
178
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
179 minsw4 t1, t8, t1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
180 minsw4 t2, t8, t2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
181 minsw4 t3, t8, t3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
182 pkwb t0, t0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
183
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
184 pkwb t1, t1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
185 pkwb t2, t2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
186 pkwb t3, t3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
187 stl t0, 0(a1)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
188
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
189 stl t1, 4(a1)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
190 addq ta, a2, a1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
191 stl t2, 0(ta)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
192 stl t3, 4(ta)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
193
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
194 bne t9, 1b
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
195 ret
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
196 .end put_pixels_clamped_mvi_asm
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
197
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
198 /************************************************************************
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
199 * void add_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels,
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
200 * int line_size)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
201 */
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
202 .align 6
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
203 .globl add_pixels_clamped_mvi_asm
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
204 .ent add_pixels_clamped_mvi_asm
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
205 add_pixels_clamped_mvi_asm:
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
206 .frame sp, 0, ra
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
207 .prologue 0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
208
510
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
209 #ifdef HAVE_GPROF
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
210 lda AT, _mcount
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
211 jsr AT, (AT), _mcount
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
212 #endif
ccd90a9cc09b Add profiling support.
mellum
parents: 509
diff changeset
213
509
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
214 lda t1, -1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
215 lda th, 8
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
216 zap t1, 0x33, tg
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
217 nop
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
218
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
219 srl tg, 1, t0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
220 xor tg, t0, tg # 0x8000800080008000
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
221 zap t1, 0xaa, tf # 0x00ff00ff00ff00ff
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
222
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
223 .align 4
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
224 1: ldl t1, 0(a1) # pix0 (try to hit cache line soon)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
225 ldl t4, 4(a1) # pix1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
226 addq a1, a2, te # pixels += line_size
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
227 ldq t0, 0(a0) # shorts0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
228
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
229 ldl t7, 0(te) # pix2 (try to hit cache line soon)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
230 ldl ta, 4(te) # pix3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
231 ldq t3, 8(a0) # shorts1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
232 ldq t6, 16(a0) # shorts2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
233
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
234 ldq t9, 24(a0) # shorts3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
235 unpkbw t1, t1 # 0 0 (quarter/op no.)
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
236 and t0, tg, t2 # 0 1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
237 unpkbw t4, t4 # 1 0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
238
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
239 bic t0, tg, t0 # 0 2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
240 unpkbw t7, t7 # 2 0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
241 and t3, tg, t5 # 1 1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
242 addq t0, t1, t0 # 0 3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
243
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
244 xor t0, t2, t0 # 0 4
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
245 unpkbw ta, ta # 3 0
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
246 and t6, tg, t8 # 2 1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
247 maxsw4 t0, zero, t0 # 0 5
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
248
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
249 bic t3, tg, t3 # 1 2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
250 bic t6, tg, t6 # 2 2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
251 minsw4 t0, tf, t0 # 0 6
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
252 addq t3, t4, t3 # 1 3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
253
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
254 pkwb t0, t0 # 0 7
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
255 xor t3, t5, t3 # 1 4
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
256 maxsw4 t3, zero, t3 # 1 5
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
257 addq t6, t7, t6 # 2 3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
258
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
259 xor t6, t8, t6 # 2 4
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
260 and t9, tg, tb # 3 1
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
261 minsw4 t3, tf, t3 # 1 6
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
262 bic t9, tg, t9 # 3 2
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
263
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
264 maxsw4 t6, zero, t6 # 2 5
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
265 addq t9, ta, t9 # 3 3
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
266 stl t0, 0(a1) # 0 8
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
267 minsw4 t6, tf, t6 # 2 6
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
268
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
269 xor t9, tb, t9 # 3 4
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
270 maxsw4 t9, zero, t9 # 3 5
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
271 lda a0, 32(a0) # block += 16;
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
272 pkwb t3, t3 # 1 7
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
273
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
274 minsw4 t9, tf, t9 # 3 6
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
275 subq th, 2, th
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
276 pkwb t6, t6 # 2 7
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
277 pkwb t9, t9 # 3 7
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
278
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
279 stl t3, 4(a1) # 1 8
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
280 addq te, a2, a1 # pixels += line_size
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
281 stl t6, 0(te) # 2 8
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
282 stl t9, 4(te) # 3 8
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
283
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
284 bne th, 1b
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
285 ret
cab79946302f Implement put_pixels_clamped and add_pixels_clamped in Assembler. This
mellum
parents:
diff changeset
286 .end add_pixels_clamped_mvi_asm