Mercurial > hg > peckfft

annotate kf_bfly4.S @ 13:3e85a9101f02 default tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
readme
author Peter Meerwald <p.meerwald@bct-electronic.com>
date Thu, 09 Feb 2012 09:44:39 +0100
parents 655dc5c14169
children
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
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
23
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
24 mov r8, r1, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
25 mov r7, r1, asl #5
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
26 add r1, r1, r1, asl #1
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
27 ldr r2, [r2, #4]
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
28 add r5, r0, r3, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
29 add r4, r0, r4, asl #4
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
30 mov r1, r1, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
31 mov sl, ip
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
32 cmp r2, #0
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
33 beq .forward
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
34 add r2, r0, #0
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
35 add r0, r0, r6, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
36 mov r6, ip
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
37 .inverse_loop:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
38 // C_MUL(scratch[0], Fout[m], *tw1);
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
39 vld1.32 {d18,d19}, [r5,:64]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
40 vld1.32 {d16,d17}, [ip,:64]
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
41 vmul.f32 d20, d18, d16
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
42 vmul.f32 d21, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
43 // load Fout[m2], *tw2
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
44 vld1.32 {d14,d15}, [r4,:64]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
45 vld1.32 {d12,d13}, [r6,:64]
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
46 vmls.f32 d20, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
47 vmla.f32 d21, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
48
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
49 // C_MUL(scratch[3], Fout[m2], *tw2);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
50 vmul.f32 d22, d14, d12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
51 vmul.f32 d23, d12, d15
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
52 // load Fout[m3], *tw3
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
53 vld1.32 {d18,d19}, [r0,:64]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
54 vld1.32 {d16,d17}, [sl,:64]
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
55 vmls.f32 d22, d15, d13
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
56 vmla.f32 d23, d14, d13
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
57
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
58 // C_MUL(scratch[2], Fout[m3], *tw3);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
59 vmul.f32 d24, d18, d16
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
60 vmul.f32 d25, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
61
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
62 // C_SUB(scratch[1], *Fout, scratch[3]);
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
63 vld1.32 {d14,d15}, [r2,:64]
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
64 vsub.f32 q13, q7, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
65
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
66 vmls.f32 d24, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
67 vmla.f32 d25, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
68
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
69 // C_ADDTO(*Fout, scratch[3]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
70 vadd.f32 q7, q7, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
71 // C_ADD(scratch[3], scratch[0], scratch[2]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
72 vadd.f32 q11, q10, q12
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
73 vst1.32 {d16,d17}, [r2,:64]
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
74
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
75 // C_SUB(Fout[m2], *Fout, scratch[3]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
76 vsub.f32 q9, q7, q11
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
77 vst1.32 {d18,d19}, [r4,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
78
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
79 // C_ADDTO(*Fout, scratch[3]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
80 vadd.f32 q7, q7, q11
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
81 vst1.32 {d14,d15}, [r2,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
82
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
83 add ip, ip, r8
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
84 add r6, r6, r7
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
85 add sl, sl, r1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
86
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
87 // C_SUB(scratch[3], scratch[0], scratch[2]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
88 vsub.f32 q11, q10, q12
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
89
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
90 // Fout[m].r = scratch[1].r - scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
91 // Fout[m].i = scratch[1].i + scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
92 vsub.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
93 vadd.f32 d19, d27, d22
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
94 vst1.32 {d18,d19}, [r5,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
95
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
96 // Fout[m3].r = scratch[1].r + scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
97 // Fout[m3].i = scratch[1].i - scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
98 vadd.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
99 vsub.f32 d19, d27, d22
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
100 vst1.32 {d18,d19}, [r0,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
101
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
102 subs r3, r3, #1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
103 bne .inverse_loop
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
104 b .done
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
105 .forward:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
106 add r2, r0, #0
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
107 add r0, r0, r6, asl #4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
108 mov r6, ip
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
109 .forward_loop:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
110 // C_MUL(scratch[0], Fout[m], *tw1);
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
111 vld1.32 {d18,d19}, [r5,:64]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
112 vld1.32 {d16,d17}, [ip,:64]
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
113 vmul.f32 d20, d18, d16
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
114 vmul.f32 d21, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
115 // load Fout[m2], *tw2
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
116 vld1.32 {d14,d15}, [r4,:64]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
117 vld1.32 {d12,d13}, [r6,:64]
10
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 vmla.f32 d21, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
120
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
121 // C_MUL(scratch[3], Fout[m2], *tw2);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
122 vmul.f32 d22, d14, d12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
123 vmul.f32 d23, d12, d15
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
124 // load Fout[m3], *tw3
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
125 vld1.32 {d18,d19}, [r0,:64]
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
126 vld1.32 {d16,d17}, [sl,:64]
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
127 vmls.f32 d22, d15, d13
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
128 vmla.f32 d23, d14, d13
10
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 vmul.f32 d24, d18, d16
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
132 vmul.f32 d25, d16, d19
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
133
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
134 // C_SUB(scratch[1], *Fout, scratch[3]);
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
135 vld1.32 {d14,d15}, [r2,:64]
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
136 vsub.f32 q13, q7, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
137
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
138 vmls.f32 d24, d19, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
139 vmla.f32 d25, d18, d17
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
140
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
141 // C_ADDTO(*Fout, scratch[3]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
142 vadd.f32 q7, q7, q11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
143 // C_ADD(scratch[3], scratch[0], scratch[2]);
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
144 vadd.f32 q11, q10, q12
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
145 vst1.32 {d16,d17}, [r2,:64]
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
146
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
147 // C_SUB(Fout[m2], *Fout, scratch[3]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
148 vsub.f32 q9, q7, q11
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
149 vst1.32 {d18,d19}, [r4,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
150
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
151 // C_ADDTO(*Fout, scratch[3]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
152 vadd.f32 q7, q7, q11
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
153 vst1.32 {d14,d15}, [r2,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
154
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
155 add ip, ip, r8
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
156 add r6, r6, r7
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
157 add sl, sl, r1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
158
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
159 // C_SUB(scratch[3], scratch[0], scratch[2]);
11
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 10
diff changeset
160 vsub.f32 q11, q10, q12
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
161
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
162 // Fout[m].r = scratch[1].r - scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
163 // Fout[m].i = scratch[1].i + scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
164 vadd.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
165 vsub.f32 d19, d27, d22
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
166 vst1.32 {d18,d19}, [r5,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
167
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
168 // Fout[m3].r = scratch[1].r + scratch[3].i;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
169 // Fout[m3].i = scratch[1].i - scratch[3].r;
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
170 vsub.f32 d18, d26, d23
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
171 vadd.f32 d19, d27, d22
12
Peter Meerwald <p.meerwald@bct-electronic.com>
parents: 11
diff changeset
172 vst1.32 {d18,d19}, [r0,:64]!
10
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
173
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
174 subs r3, r3, #1
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
175 bne .forward_loop
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
176 .done:
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
177 ldmfd sp!, {r4, r5, r6, r7, r8, sl}
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
178 bx lr
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
179 .cfi_endproc
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
180 .fnend
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
181 .size kf_bfly4, .-kf_bfly4
Peter Meerwald <p.meerwald@bct-electronic.com>
parents:
diff changeset
182

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