annotate kf_bfly4.S @ 10:05f6ab0a17c0

backup
author Peter Meerwald <p.meerwald@bct-electronic.com>
date Wed, 21 Sep 2011 15:20:58 +0200
parents
children abdcde012978
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
1 .cpu cortex-a8
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
2 .eabi_attribute 27, 3
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
3 .fpu neon
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
4 .eabi_attribute 23, 1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
5 .eabi_attribute 24, 1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
6 .eabi_attribute 25, 1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
7 .eabi_attribute 26, 2
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
8 .eabi_attribute 30, 1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
9 .eabi_attribute 18, 4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
11 .text
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
12 .align 2
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
13 .global kf_bfly4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
14 .type kf_bfly4, %function
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
15 kf_bfly4:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
16 .fnstart
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
17 .cfi_startproc
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
18 stmfd sp!, {r4, r5, r6, r7, r8, sl}
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
19 .save {r4, r5, r6, r7, r8, sl}
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
20 mov r4, r3, asl #1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
21 add r6, r4, r3
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
22 add ip, r2, #264
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
23 ldr r2, [r2, #4]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
24 cmp r2, #0
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
25 beq .forward
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
26 mov r8, r1, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
27 mov r7, r1, asl #5
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
28 add r1, r1, r1, asl #1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
29 mov r1, r1, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
30 add r2, r0, #0
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
31 add r5, r0, r3, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
32 add r4, r0, r4, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
33 add r0, r0, r6, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
34 mov sl, ip
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
35 mov r6, ip
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
36 .inverse_loop:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
37 // C_MUL(scratch[0], Fout[m], *tw1);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
38 vld1.32 {d18,d19}, [r5]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
39 vld1.32 {d16,d17}, [ip]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
40 vmul.f32 d20, d18, d16
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
41 vmls.f32 d20, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
42 vmul.f32 d21, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
43 vmla.f32 d21, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
44
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
45 // C_MUL(scratch[3], Fout[m2], *tw2);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
46 vld1.32 {d18,d19}, [r4]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
47 vld1.32 {d16,d17}, [r6]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
48 vmul.f32 d22, d18, d16
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
49 vmls.f32 d22, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
50 vmul.f32 d23, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
51 vmla.f32 d23, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
52
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
53 // C_MUL(scratch[2], Fout[m3], *tw3);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
54 vld1.32 {d18,d19}, [r0]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
55 vld1.32 {d16,d17}, [sl]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
56 vmul.f32 d24, d18, d16
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
57 vmls.f32 d24, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
58 vmul.f32 d25, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
59 vmla.f32 d25, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
60
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
61 // C_SUB(scratch[1], *Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
62 vld1.32 {d16,d17}, [r2]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
63 vsubq.f32 q13, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
64
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
65 // C_ADDTO(*Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
66 vaddq.f32 q8, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
67 vst1.32 {d16,d17}, [r2]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
68
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
69 // C_ADD(scratch[3], scratch[0], scratch[2]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
70 vaddq.f32 q11, q10, q12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
71
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
72 // C_SUB(Fout[m2], *Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
73 vsubq.f32 q9, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
74 vst1.32 {d18,d19}, [r4]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
75
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
76 // C_ADDTO(*Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
77 vaddq.f32 q8, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
78 vst1.32 {d16,d17}, [r2]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
79
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
80 add ip, ip, r8
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
81 add r6, r6, r7
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
82 add sl, sl, r1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
83
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
84 // C_SUB(scratch[3], scratch[0], scratch[2]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
85 vsubq.f32 q11, q10, q12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
86
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
87 // Fout[m].r = scratch[1].r - scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
88 // Fout[m].i = scratch[1].i + scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
89 vsub.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
90 vadd.f32 d19, d27, d22
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
91 vst1.32 {d18,d19}, [r5]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
92
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
93 // Fout[m3].r = scratch[1].r + scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
94 // Fout[m3].i = scratch[1].i - scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
95 vadd.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
96 vsub.f32 d19, d27, d22
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
97 vst1.32 {d18,d19}, [r0]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
98
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
99 subs r3, r3, #1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
100 bne .inverse_loop
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
101 b .done
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
102 .forward:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
103 mov r8, r1, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
104 mov r7, r1, asl #5
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
105 add r1, r1, r1, asl #1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
106 mov r1, r1, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
107 add r2, r0, #0
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
108 add r5, r0, r3, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
109 add r4, r0, r4, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
110 add r0, r0, r6, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
111 mov sl, ip
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
112 mov r6, ip
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
113 .forward_loop:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
114 // C_MUL(scratch[0], Fout[m], *tw1);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
115 vld1.32 {d18,d19}, [r5]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
116 vld1.32 {d16,d17}, [ip]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
117 vmul.f32 d20, d18, d16
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
118 vmls.f32 d20, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
119 vmul.f32 d21, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
120 vmla.f32 d21, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
121
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
122 // C_MUL(scratch[3], Fout[m2], *tw2);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
123 vld1.32 {d18,d19}, [r4]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
124 vld1.32 {d16,d17}, [r6]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
125 vmul.f32 d22, d18, d16
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
126 vmls.f32 d22, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
127 vmul.f32 d23, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
128 vmla.f32 d23, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
129
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
130 // C_MUL(scratch[2], Fout[m3], *tw3);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
131 vld1.32 {d18,d19}, [r0]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
132 vld1.32 {d16,d17}, [sl]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
133 vmul.f32 d24, d18, d16
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
134 vmls.f32 d24, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
135 vmul.f32 d25, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
136 vmla.f32 d25, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
137
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
138 // C_SUB(scratch[1], *Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
139 vld1.32 {d16,d17}, [r2]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
140 vsubq.f32 q13, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
141
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
142 // C_ADDTO(*Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
143 vaddq.f32 q8, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
144 vst1.32 {d16,d17}, [r2]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
145
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
146 // C_ADD(scratch[3], scratch[0], scratch[2]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
147 vaddq.f32 q11, q10, q12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
148
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
149 // C_SUB(Fout[m2], *Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
150 vsubq.f32 q9, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
151 vst1.32 {d18,d19}, [r4]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
152
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
153 // C_ADDTO(*Fout, scratch[3]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
154 vaddq.f32 q8, q8, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
155 vst1.32 {d16,d17}, [r2]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
156
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
157 add ip, ip, r8
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
158 add r6, r6, r7
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
159 add sl, sl, r1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
160
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
161 // C_SUB(scratch[3], scratch[0], scratch[2]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
162 vsubq.f32 q11, q10, q12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
163
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
164 // Fout[m].r = scratch[1].r - scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
165 // Fout[m].i = scratch[1].i + scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
166 vadd.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
167 vsub.f32 d19, d27, d22
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
168 vst1.32 {d18,d19}, [r5]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
169
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
170 // Fout[m3].r = scratch[1].r + scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
171 // Fout[m3].i = scratch[1].i - scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
172 vsub.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
173 vadd.f32 d19, d27, d22
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
174 vst1.32 {d18,d19}, [r0]!
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
175
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
176 subs r3, r3, #1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
177 bne .forward_loop
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
178 .done:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
179 ldmfd sp!, {r4, r5, r6, r7, r8, sl}
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
180 bx lr
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
181 .cfi_endproc
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
182 .fnend
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
183 .size kf_bfly4, .-kf_bfly4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
184

Repositories maintained by Peter Meerwald, pmeerw@pmeerw.net.