Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
1330 | killagreg | 1 | ;-----------------------------------------------------------------------------; |
2 | ; Fast integer squareroot routines for avr-gcc project (C)ChaN, 2008 |
||
3 | ; http://elm-chan.org/docs/avrlib/sqrt32.S |
||
4 | ;-----------------------------------------------------------------------------; |
||
5 | ; uint16_t isqrt32 (uint32_t n); |
||
6 | ; uint8_t isqrt16 (uint16_t n); |
||
7 | ; uint16_t ihypot (int16_t x, int16_t y); |
||
8 | |||
9 | ;-----------------------------------------------------------------------------: |
||
10 | ; 32bit integer squareroot |
||
11 | ;-----------------------------------------------------------------------------; |
||
12 | ; uint16_t isqrt32 ( |
||
13 | ; uint32_t n |
||
14 | ; ); |
||
15 | ; |
||
16 | ; Return Value: |
||
17 | ; Squareroot of n. |
||
18 | ; |
||
19 | ; Size = 53 words |
||
20 | ; Clock = 532..548 cycles |
||
21 | ; Stack = 0 byte |
||
22 | |||
23 | .global isqrt32 |
||
24 | .func isqrt32 |
||
25 | |||
26 | isqrt32: |
||
27 | clr r0 |
||
28 | clr r18 |
||
29 | clr r19 |
||
30 | clr r20 |
||
31 | ldi r21, 1 |
||
32 | clr r27 |
||
33 | clr r30 |
||
34 | clr r31 |
||
35 | ldi r26, 16 |
||
36 | 1: lsl r22 |
||
37 | rol r23 |
||
38 | rol r24 |
||
39 | rol r25 |
||
40 | rol r0 |
||
41 | rol r18 |
||
42 | rol r19 |
||
43 | rol r20 |
||
44 | lsl r22 |
||
45 | rol r23 |
||
46 | rol r24 |
||
47 | rol r25 |
||
48 | rol r0 |
||
49 | rol r18 |
||
50 | rol r19 |
||
51 | rol r20 |
||
52 | brpl 2f |
||
53 | add r0, r21 |
||
54 | adc r18, r27 |
||
55 | adc r19, r30 |
||
56 | adc r20, r31 |
||
57 | rjmp 3f |
||
58 | 2: sub r0, r21 |
||
59 | sbc r18, r27 |
||
60 | sbc r19, r30 |
||
61 | sbc r20, r31 |
||
62 | 3: lsl r21 |
||
63 | rol r27 |
||
64 | rol r30 |
||
65 | andi r21, 0b11111000 |
||
66 | ori r21, 0b00000101 |
||
67 | sbrc r20, 7 |
||
68 | subi r21, 2 |
||
69 | dec r26 |
||
70 | brne 1b |
||
71 | lsr r30 |
||
72 | ror r27 |
||
73 | ror r21 |
||
74 | lsr r30 |
||
75 | ror r27 |
||
76 | ror r21 |
||
77 | mov r24, r21 |
||
78 | mov r25, r27 |
||
79 | ret |
||
80 | .endfunc |
||
81 | |||
82 | |||
83 | |||
84 | ;-----------------------------------------------------------------------------: |
||
85 | ; 16bit integer squareroot |
||
86 | ;-----------------------------------------------------------------------------; |
||
87 | ; uint8_t isqrt16 ( |
||
88 | ; uint16_t n |
||
89 | ; ); |
||
90 | ; |
||
91 | ; Return Value: |
||
92 | ; Squareroot of n. |
||
93 | ; |
||
94 | ; Size = 33 words |
||
95 | ; Clock = 181..189 cycles |
||
96 | ; Stack = 0 byte |
||
97 | |||
98 | .global isqrt16 |
||
99 | .func isqrt16 |
||
100 | |||
101 | isqrt16: |
||
102 | clr r18 |
||
103 | clr r19 |
||
104 | ldi r20, 1 |
||
105 | clr r21 |
||
106 | ldi r22, 8 |
||
107 | 1: lsl r24 |
||
108 | rol r25 |
||
109 | rol r18 |
||
110 | rol r19 |
||
111 | lsl r24 |
||
112 | rol r25 |
||
113 | rol r18 |
||
114 | rol r19 |
||
115 | brpl 2f |
||
116 | add r18, r20 |
||
117 | adc r19, r21 |
||
118 | rjmp 3f |
||
119 | 2: sub r18, r20 |
||
120 | sbc r19, r21 |
||
121 | 3: lsl r20 |
||
122 | rol r21 |
||
123 | andi r20, 0b11111000 |
||
124 | ori r20, 0b00000101 |
||
125 | sbrc r19, 7 |
||
126 | subi r20, 2 |
||
127 | dec r22 |
||
128 | brne 1b |
||
129 | lsr r21 |
||
130 | ror r20 |
||
131 | lsr r21 |
||
132 | ror r20 |
||
133 | mov r24, r20 |
||
134 | ret |
||
135 | .endfunc |
||
136 | |||
137 | |||
138 | |||
139 | ;-----------------------------------------------------------------------------: |
||
140 | ; 16bit integer hypot (megaAVR is required) |
||
141 | ;-----------------------------------------------------------------------------; |
||
142 | ; uint16_t ihypot ( |
||
143 | ; int16_t x, |
||
144 | ; int16_t y |
||
145 | ; ); |
||
146 | ; |
||
147 | ; Return Value: |
||
148 | ; Squareroot of (x*x + y*y) |
||
149 | ; |
||
150 | ; Size = 42 words |
||
151 | ; Clock = 581..597 cycles |
||
152 | ; Stack = 0 byte |
||
153 | |||
154 | .global ihypot |
||
155 | .func ihypot |
||
156 | |||
157 | ihypot: |
||
158 | clr r26 |
||
159 | sbrs r25, 7 |
||
160 | rjmp 1f |
||
161 | com r24 |
||
162 | com r25 |
||
163 | adc r24, r26 |
||
164 | adc r25, r26 |
||
165 | 1: sbrs r23, 7 |
||
166 | rjmp 2f |
||
167 | com r22 |
||
168 | com r23 |
||
169 | adc r22, r26 |
||
170 | adc r23, r26 |
||
171 | 2: mul r22, r22 |
||
172 | movw r18, r0 |
||
173 | mul r23, r23 |
||
174 | movw r20, r0 |
||
175 | mul r22, r23 |
||
176 | add r19, r0 |
||
177 | adc r20, r1 |
||
178 | adc r21, r26 |
||
179 | add r19, r0 |
||
180 | adc r20, r1 |
||
181 | adc r21, r26 |
||
182 | mul r24, r24 |
||
183 | movw r30, r0 |
||
184 | mul r25, r25 |
||
185 | add r18, r30 |
||
186 | adc r19, r31 |
||
187 | adc r20, r0 |
||
188 | adc r21, r1 |
||
189 | mul r24, r25 |
||
190 | add r19, r0 |
||
191 | adc r20, r1 |
||
192 | adc r21, r26 |
||
193 | add r19, r0 |
||
194 | adc r20, r1 |
||
195 | adc r21, r26 |
||
196 | movw r24, r20 |
||
197 | movw r22, r18 |
||
198 | clr r1 |
||
199 | rjmp isqrt32 |
||
200 | .endfunc |
||
201 | |||
202 | |||
203 |