cc68a136 |
1 | @ some color conversion and blitting routines\r |
2 | \r |
3 | @ (c) Copyright 2006, notaz\r |
4 | @ All Rights Reserved\r |
5 | \r |
6 | \r |
7 | @ Convert 0000bbb0 ggg0rrr0 0000bbb0 ggg0rrr0\r |
8 | @ to 00000000 rrr00000 ggg00000 bbb00000 ...\r |
9 | \r |
10 | @ lr = 0x00e000e0, out: r3=lower_pix, r2=higher_pix; trashes rin\r |
11 | @ if sh==2, r8=0x00404040 (sh!=0 destroys flags!)\r |
12 | .macro convRGB32_2 rin sh=0\r |
13 | and r2, lr, \rin, lsr #4 @ blue\r |
14 | and r3, \rin, lr\r |
15 | orr r2, r2, r3, lsl #8 @ g0b0g0b0\r |
16 | \r |
17 | mov r3, r2, lsl #16 @ g0b00000\r |
18 | and \rin,lr, \rin, ror #12 @ 00r000r0 (reversed)\r |
19 | orr r3, r3, \rin, lsr #16 @ g0b000r0\r |
20 | .if \sh == 1\r |
21 | mov r3, r3, ror #17 @ shadow mode\r |
22 | .elseif \sh == 2\r |
23 | adds r3, r3, #0x40000000 @ green\r |
24 | orrcs r3, r3, #0xe0000000\r |
25 | mov r3, r3, ror #8\r |
26 | adds r3, r3, #0x40000000\r |
27 | orrcs r3, r3, #0xe0000000\r |
28 | mov r3, r3, ror #16\r |
29 | adds r3, r3, #0x40000000\r |
30 | orrcs r3, r3, #0xe0000000\r |
31 | mov r3, r3, ror #24\r |
32 | .else\r |
33 | mov r3, r3, ror #16 @ r3=low\r |
34 | .endif\r |
35 | \r |
36 | orr r3, r3, r3, lsr #3\r |
37 | str r3, [r0], #4\r |
38 | \r |
39 | mov r2, r2, lsr #16\r |
40 | orr r2, r2, \rin, lsl #16\r |
41 | .if \sh == 1\r |
42 | mov r2, r2, lsr #1\r |
43 | .elseif \sh == 2\r |
44 | mov r2, r2, ror #8\r |
45 | adds r2, r2, #0x40000000 @ blue\r |
46 | orrcs r2, r2, #0xe0000000\r |
47 | mov r2, r2, ror #8\r |
48 | adds r2, r2, #0x40000000\r |
49 | orrcs r2, r2, #0xe0000000\r |
50 | mov r2, r2, ror #8\r |
51 | adds r2, r2, #0x40000000\r |
52 | orrcs r2, r2, #0xe0000000\r |
53 | mov r2, r2, ror #8\r |
54 | .endif\r |
55 | \r |
56 | orr r2, r2, r2, lsr #3\r |
57 | str r2, [r0], #4\r |
58 | .endm\r |
59 | \r |
60 | \r |
61 | .global vidConvCpyRGB32 @ void *to, void *from, int pixels\r |
62 | \r |
63 | vidConvCpyRGB32:\r |
64 | stmfd sp!, {r4-r7,lr}\r |
65 | \r |
66 | mov r12, r2, lsr #3 @ repeats\r |
67 | mov lr, #0x00e00000\r |
68 | orr lr, lr, #0x00e0\r |
69 | \r |
70 | .loopRGB32:\r |
71 | subs r12, r12, #1\r |
72 | \r |
73 | ldmia r1!, {r4-r7}\r |
74 | convRGB32_2 r4\r |
75 | convRGB32_2 r5\r |
76 | convRGB32_2 r6\r |
77 | convRGB32_2 r7\r |
78 | \r |
79 | bgt .loopRGB32\r |
80 | \r |
81 | ldmfd sp!, {r4-r7,lr}\r |
82 | bx lr\r |
83 | \r |
84 | \r |
85 | .global vidConvCpyRGB32sh @ void *to, void *from, int pixels\r |
86 | \r |
87 | vidConvCpyRGB32sh:\r |
88 | stmfd sp!, {r4-r7,lr}\r |
89 | \r |
90 | mov r12, r2, lsr #3 @ repeats\r |
91 | mov lr, #0x00e00000\r |
92 | orr lr, lr, #0x00e0\r |
93 | \r |
94 | .loopRGB32sh:\r |
95 | subs r12, r12, #1\r |
96 | \r |
97 | ldmia r1!, {r4-r7}\r |
98 | convRGB32_2 r4, 1\r |
99 | convRGB32_2 r5, 1\r |
100 | convRGB32_2 r6, 1\r |
101 | convRGB32_2 r7, 1\r |
102 | \r |
103 | bgt .loopRGB32sh\r |
104 | \r |
105 | ldmfd sp!, {r4-r7,lr}\r |
106 | bx lr\r |
107 | \r |
108 | \r |
109 | .global vidConvCpyRGB32hi @ void *to, void *from, int pixels\r |
110 | \r |
111 | vidConvCpyRGB32hi:\r |
112 | stmfd sp!, {r4-r7,lr}\r |
113 | \r |
114 | mov r12, r2, lsr #3 @ repeats\r |
115 | mov lr, #0x00e00000\r |
116 | orr lr, lr, #0x00e0\r |
117 | \r |
118 | .loopRGB32hi:\r |
119 | ldmia r1!, {r4-r7}\r |
120 | convRGB32_2 r4, 2\r |
121 | convRGB32_2 r5, 2\r |
122 | convRGB32_2 r6, 2\r |
123 | convRGB32_2 r7, 2\r |
124 | \r |
125 | subs r12, r12, #1\r |
126 | bgt .loopRGB32hi\r |
127 | \r |
128 | ldmfd sp!, {r4-r7,lr}\r |
129 | bx lr\r |
130 | \r |
131 | \r |
132 | @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r |
133 | \r |
134 | \r |
135 | @ mode2 blitter for 40 cols\r |
136 | .global vidCpyM2_40col @ void *dest, void *src\r |
137 | \r |
138 | vidCpyM2_40col:\r |
139 | stmfd sp!, {r4-r6,lr}\r |
140 | \r |
141 | mov r12, #224 @ lines\r |
142 | add r1, r1, #8\r |
143 | \r |
144 | vidCpyM2_40_loop_out:\r |
145 | mov r6, #10\r |
146 | vidCpyM2_40_loop:\r |
147 | subs r6, r6, #1\r |
148 | ldmia r1!, {r2-r5}\r |
149 | stmia r0!, {r2-r5}\r |
150 | ldmia r1!, {r2-r5}\r |
151 | stmia r0!, {r2-r5}\r |
152 | bne vidCpyM2_40_loop\r |
153 | subs r12,r12,#1\r |
154 | add r1, r1, #8\r |
155 | bne vidCpyM2_40_loop_out\r |
156 | \r |
157 | ldmfd sp!, {r4-r6,lr}\r |
158 | bx lr\r |
159 | \r |
160 | \r |
161 | @ mode2 blitter for 32 cols\r |
162 | .global vidCpyM2_32col @ void *dest, void *src\r |
163 | \r |
164 | vidCpyM2_32col:\r |
165 | stmfd sp!, {r4-r6,lr}\r |
166 | \r |
167 | mov r12, #224 @ lines\r |
168 | add r1, r1, #8\r |
169 | add r0, r0, #32\r |
170 | \r |
171 | vidCpyM2_32_loop_out:\r |
172 | mov r6, #8\r |
173 | vidCpyM2_32_loop:\r |
174 | subs r6, r6, #1\r |
175 | ldmia r1!, {r2-r5}\r |
176 | stmia r0!, {r2-r5}\r |
177 | ldmia r1!, {r2-r5}\r |
178 | stmia r0!, {r2-r5}\r |
179 | bne vidCpyM2_32_loop\r |
180 | subs r12,r12,#1\r |
181 | add r0, r0, #64\r |
182 | add r1, r1, #8+64\r |
183 | bne vidCpyM2_32_loop_out\r |
184 | \r |
185 | ldmfd sp!, {r4-r6,lr}\r |
186 | bx lr\r |
187 | \r |
188 | \r |
189 | @ mode2 blitter for 32 cols with no borders\r |
190 | .global vidCpyM2_32col_nobord @ void *dest, void *src\r |
191 | \r |
192 | vidCpyM2_32col_nobord:\r |
193 | stmfd sp!, {r4-r6,lr}\r |
194 | \r |
195 | mov r12, #224 @ lines\r |
196 | add r1, r1, #8\r |
197 | b vidCpyM2_32_loop_out\r |
198 | \r |
199 | \r |
200 | @ test\r |
201 | .global spend_cycles @ c\r |
202 | \r |
203 | spend_cycles:\r |
204 | mov r0, r0, lsr #2 @ 4 cycles/iteration\r |
205 | sub r0, r0, #2 @ entry/exit/init\r |
206 | .sc_loop:\r |
207 | subs r0, r0, #1\r |
208 | bpl .sc_loop\r |
209 | \r |
210 | bx lr\r |