annotate ppc/fft_altivec_s.S @ 12518:67e7e49058c2 libavcodec

Split and then simplify address generation macro. Allows nasm to work for this code.
author reimar
date Sun, 26 Sep 2010 09:08:11 +0000
parents 6f064ab48463
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
1 /*
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
2 * FFT transform with Altivec optimizations
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
3 * Copyright (c) 2009 Loren Merritt
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
4 *
12188
6f064ab48463 more credits to D. J. Bernstein for fft
lorenm
parents: 12089
diff changeset
5 * This algorithm (though not any of the implementation details) is
6f064ab48463 more credits to D. J. Bernstein for fft
lorenm
parents: 12089
diff changeset
6 * based on libdjbfft by D. J. Bernstein.
6f064ab48463 more credits to D. J. Bernstein for fft
lorenm
parents: 12089
diff changeset
7 *
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
8 * This file is part of FFmpeg.
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
9 *
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
10 * FFmpeg is free software; you can redistribute it and/or
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
11 * modify it under the terms of the GNU Lesser General Public
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
12 * License as published by the Free Software Foundation; either
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
13 * version 2.1 of the License, or (at your option) any later version.
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
14 *
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
15 * FFmpeg is distributed in the hope that it will be useful,
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
18 * Lesser General Public License for more details.
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
19 *
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
20 * You should have received a copy of the GNU Lesser General Public
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
21 * License along with FFmpeg; if not, write to the Free Software
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
22 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
23 */
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
24
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
25 /*
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
26 * These functions are not individually interchangeable with the C versions.
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
27 * While C takes arrays of FFTComplex, Altivec leaves intermediate results
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
28 * in blocks as convenient to the vector size.
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
29 * i.e. {4x real, 4x imaginary, 4x real, ...}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
30 *
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
31 * I ignore standard calling convention.
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
32 * Instead, the following registers are treated as global constants:
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
33 * v14: zero
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
34 * v15..v18: cosines
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
35 * v19..v29: permutations
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
36 * r9: 16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
37 * r12: ff_cos_tabs
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
38 * and the rest are free for local use.
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
39 */
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
40
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
41 #include "config.h"
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
42 #include "asm.S"
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
43
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
44 .text
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
45
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
46 .macro addi2 ra, imm // add 32-bit immediate
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
47 .if \imm & 0xffff
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
48 addi \ra, \ra, \imm@l
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
49 .endif
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
50 .if (\imm+0x8000)>>16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
51 addis \ra, \ra, \imm@ha
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
52 .endif
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
53 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
54
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
55 .macro FFT4 a0, a1, a2, a3 // in:0-1 out:2-3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
56 vperm \a2,\a0,\a1,v20 // vcprm(0,1,s2,s1) // {r0,i0,r3,i2}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
57 vperm \a3,\a0,\a1,v21 // vcprm(2,3,s0,s3) // {r1,i1,r2,i3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
58 vaddfp \a0,\a2,\a3 // {t1,t2,t6,t5}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
59 vsubfp \a1,\a2,\a3 // {t3,t4,t8,t7}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
60 vmrghw \a2,\a0,\a1 // vcprm(0,s0,1,s1) // {t1,t3,t2,t4}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
61 vperm \a3,\a0,\a1,v22 // vcprm(2,s3,3,s2) // {t6,t7,t5,t8}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
62 vaddfp \a0,\a2,\a3 // {r0,r1,i0,i1}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
63 vsubfp \a1,\a2,\a3 // {r2,r3,i2,i3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
64 vperm \a2,\a0,\a1,v23 // vcprm(0,1,s0,s1) // {r0,r1,r2,r3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
65 vperm \a3,\a0,\a1,v24 // vcprm(2,3,s2,s3) // {i0,i1,i2,i3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
66 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
67
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
68 .macro FFT4x2 a0, a1, b0, b1, a2, a3, b2, b3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
69 vperm \a2,\a0,\a1,v20 // vcprm(0,1,s2,s1) // {r0,i0,r3,i2}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
70 vperm \a3,\a0,\a1,v21 // vcprm(2,3,s0,s3) // {r1,i1,r2,i3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
71 vperm \b2,\b0,\b1,v20
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
72 vperm \b3,\b0,\b1,v21
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
73 vaddfp \a0,\a2,\a3 // {t1,t2,t6,t5}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
74 vsubfp \a1,\a2,\a3 // {t3,t4,t8,t7}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
75 vaddfp \b0,\b2,\b3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
76 vsubfp \b1,\b2,\b3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
77 vmrghw \a2,\a0,\a1 // vcprm(0,s0,1,s1) // {t1,t3,t2,t4}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
78 vperm \a3,\a0,\a1,v22 // vcprm(2,s3,3,s2) // {t6,t7,t5,t8}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
79 vmrghw \b2,\b0,\b1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
80 vperm \b3,\b0,\b1,v22
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
81 vaddfp \a0,\a2,\a3 // {r0,r1,i0,i1}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
82 vsubfp \a1,\a2,\a3 // {r2,r3,i2,i3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
83 vaddfp \b0,\b2,\b3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
84 vsubfp \b1,\b2,\b3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
85 vperm \a2,\a0,\a1,v23 // vcprm(0,1,s0,s1) // {r0,r1,r2,r3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
86 vperm \a3,\a0,\a1,v24 // vcprm(2,3,s2,s3) // {i0,i1,i2,i3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
87 vperm \b2,\b0,\b1,v23
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
88 vperm \b3,\b0,\b1,v24
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
89 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
90
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
91 .macro FFT8 a0, a1, b0, b1, a2, a3, b2, b3, b4 // in,out:a0-b1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
92 vmrghw \b2,\b0,\b1 // vcprm(0,s0,1,s1) // {r4,r6,i4,i6}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
93 vmrglw \b3,\b0,\b1 // vcprm(2,s2,3,s3) // {r5,r7,i5,i7}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
94 vperm \a2,\a0,\a1,v20 // FFT4 ...
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
95 vperm \a3,\a0,\a1,v21
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
96 vaddfp \b0,\b2,\b3 // {t1,t3,t2,t4}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
97 vsubfp \b1,\b2,\b3 // {r5,r7,i5,i7}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
98 vperm \b4,\b1,\b1,v25 // vcprm(2,3,0,1) // {i5,i7,r5,r7}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
99 vaddfp \a0,\a2,\a3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
100 vsubfp \a1,\a2,\a3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
101 vmaddfp \b1,\b1,v17,v14 // * {-1,1,1,-1}/sqrt(2)
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
102 vmaddfp \b1,\b4,v18,\b1 // * { 1,1,1,1 }/sqrt(2) // {t8,ta,t7,t9}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
103 vmrghw \a2,\a0,\a1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
104 vperm \a3,\a0,\a1,v22
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
105 vperm \b2,\b0,\b1,v26 // vcprm(1,2,s3,s0) // {t3,t2,t9,t8}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
106 vperm \b3,\b0,\b1,v27 // vcprm(0,3,s2,s1) // {t1,t4,t7,ta}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
107 vaddfp \a0,\a2,\a3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
108 vsubfp \a1,\a2,\a3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
109 vaddfp \b0,\b2,\b3 // {t1,t2,t9,ta}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
110 vsubfp \b1,\b2,\b3 // {t6,t5,tc,tb}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
111 vperm \a2,\a0,\a1,v23
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
112 vperm \a3,\a0,\a1,v24
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
113 vperm \b2,\b0,\b1,v28 // vcprm(0,2,s1,s3) // {t1,t9,t5,tb}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
114 vperm \b3,\b0,\b1,v29 // vcprm(1,3,s0,s2) // {t2,ta,t6,tc}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
115 vsubfp \b0,\a2,\b2 // {r4,r5,r6,r7}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
116 vsubfp \b1,\a3,\b3 // {i4,i5,i6,i7}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
117 vaddfp \a0,\a2,\b2 // {r0,r1,r2,r3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
118 vaddfp \a1,\a3,\b3 // {i0,i1,i2,i3}
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
119 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
120
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
121 .macro BF d0,d1,s0,s1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
122 vsubfp \d1,\s0,\s1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
123 vaddfp \d0,\s0,\s1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
124 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
125
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
126 .macro zip d0,d1,s0,s1
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
127 vmrghw \d0,\s0,\s1
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
128 vmrglw \d1,\s0,\s1
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
129 .endm
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
130
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
131 .macro def_fft4 interleave
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
132 fft4\interleave\()_altivec:
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
133 lvx v0, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
134 lvx v1,r9,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
135 FFT4 v0,v1,v2,v3
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
136 .ifnb \interleave
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
137 zip v0,v1,v2,v3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
138 stvx v0, 0,r3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
139 stvx v1,r9,r3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
140 .else
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
141 stvx v2, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
142 stvx v3,r9,r3
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
143 .endif
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
144 blr
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
145 .endm
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
146
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
147 .macro def_fft8 interleave
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
148 fft8\interleave\()_altivec:
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
149 addi r4,r3,32
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
150 lvx v0, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
151 lvx v1,r9,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
152 lvx v2, 0,r4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
153 lvx v3,r9,r4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
154 FFT8 v0,v1,v2,v3,v4,v5,v6,v7,v8
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
155 .ifnb \interleave
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
156 zip v4,v5,v0,v1
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
157 zip v6,v7,v2,v3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
158 stvx v4, 0,r3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
159 stvx v5,r9,r3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
160 stvx v6, 0,r4
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
161 stvx v7,r9,r4
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
162 .else
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
163 stvx v0, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
164 stvx v1,r9,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
165 stvx v2, 0,r4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
166 stvx v3,r9,r4
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
167 .endif
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
168 blr
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
169 .endm
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
170
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
171 .macro def_fft16 interleave
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
172 fft16\interleave\()_altivec:
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
173 addi r5,r3,64
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
174 addi r6,r3,96
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
175 addi r4,r3,32
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
176 lvx v0, 0,r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
177 lvx v1,r9,r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
178 lvx v2, 0,r6
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
179 lvx v3,r9,r6
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
180 FFT4x2 v0,v1,v2,v3,v4,v5,v6,v7
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
181 lvx v0, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
182 lvx v1,r9,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
183 lvx v2, 0,r4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
184 lvx v3,r9,r4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
185 FFT8 v0,v1,v2,v3,v8,v9,v10,v11,v12
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
186 vmaddfp v8,v4,v15,v14 // r2*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
187 vmaddfp v9,v5,v15,v14 // i2*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
188 vmaddfp v10,v6,v15,v14 // r3*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
189 vmaddfp v11,v7,v15,v14 // i3*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
190 vmaddfp v8,v5,v16,v8 // i2*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
191 vnmsubfp v9,v4,v16,v9 // r2*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
192 vnmsubfp v10,v7,v16,v10 // i3*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
193 vmaddfp v11,v6,v16,v11 // r3*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
194 BF v10,v12,v10,v8
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
195 BF v11,v13,v9,v11
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
196 BF v0,v4,v0,v10
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
197 BF v3,v7,v3,v12
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
198 BF v1,v5,v1,v11
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
199 BF v2,v6,v2,v13
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
200 .ifnb \interleave
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
201 zip v8, v9,v0,v1
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
202 zip v10,v11,v2,v3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
203 zip v12,v13,v4,v5
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
204 zip v14,v15,v6,v7
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
205 stvx v8, 0,r3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
206 stvx v9,r9,r3
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
207 stvx v10, 0,r4
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
208 stvx v11,r9,r4
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
209 stvx v12, 0,r5
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
210 stvx v13,r9,r5
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
211 stvx v14, 0,r6
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
212 stvx v15,r9,r6
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
213 .else
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
214 stvx v0, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
215 stvx v4, 0,r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
216 stvx v3,r9,r4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
217 stvx v7,r9,r6
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
218 stvx v1,r9,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
219 stvx v5,r9,r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
220 stvx v2, 0,r4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
221 stvx v6, 0,r6
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
222 .endif
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
223 blr
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
224 .endm
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
225
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
226 // void pass(float *z, float *wre, int n)
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
227 .macro PASS interleave, suffix
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
228 fft_pass\suffix\()_altivec:
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
229 mtctr r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
230 slwi r0,r5,4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
231 slwi r7,r5,6 // o2
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
232 slwi r5,r5,5 // o1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
233 add r10,r5,r7 // o3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
234 add r0,r4,r0 // wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
235 addi r6,r5,16 // o1+16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
236 addi r8,r7,16 // o2+16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
237 addi r11,r10,16 // o3+16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
238 1:
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
239 lvx v8, 0,r4 // wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
240 lvx v10, 0,r0 // wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
241 sub r0,r0,r9
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
242 lvx v9, 0,r0
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
243 vperm v9,v9,v10,v19 // vcprm(s0,3,2,1) => wim[0 .. -3]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
244 lvx v4,r3,r7 // r2 = z[o2]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
245 lvx v5,r3,r8 // i2 = z[o2+16]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
246 lvx v6,r3,r10 // r3 = z[o3]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
247 lvx v7,r3,r11 // i3 = z[o3+16]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
248 vmaddfp v10,v4,v8,v14 // r2*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
249 vmaddfp v11,v5,v8,v14 // i2*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
250 vmaddfp v12,v6,v8,v14 // r3*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
251 vmaddfp v13,v7,v8,v14 // i3*wre
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
252 lvx v0, 0,r3 // r0 = z[0]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
253 lvx v3,r3,r6 // i1 = z[o1+16]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
254 vmaddfp v10,v5,v9,v10 // i2*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
255 vnmsubfp v11,v4,v9,v11 // r2*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
256 vnmsubfp v12,v7,v9,v12 // i3*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
257 vmaddfp v13,v6,v9,v13 // r3*wim
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
258 lvx v1,r3,r9 // i0 = z[16]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
259 lvx v2,r3,r5 // r1 = z[o1]
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
260 BF v12,v8,v12,v10
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
261 BF v13,v9,v11,v13
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
262 BF v0,v4,v0,v12
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
263 BF v3,v7,v3,v8
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
264 .if !\interleave
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
265 stvx v0, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
266 stvx v4,r3,r7
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
267 stvx v3,r3,r6
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
268 stvx v7,r3,r11
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
269 .endif
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
270 BF v1,v5,v1,v13
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
271 BF v2,v6,v2,v9
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
272 .if !\interleave
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
273 stvx v1,r3,r9
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
274 stvx v2,r3,r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
275 stvx v5,r3,r8
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
276 stvx v6,r3,r10
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
277 .else
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
278 vmrghw v8,v0,v1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
279 vmrglw v9,v0,v1
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
280 stvx v8, 0,r3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
281 stvx v9,r3,r9
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
282 vmrghw v8,v2,v3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
283 vmrglw v9,v2,v3
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
284 stvx v8,r3,r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
285 stvx v9,r3,r6
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
286 vmrghw v8,v4,v5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
287 vmrglw v9,v4,v5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
288 stvx v8,r3,r7
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
289 stvx v9,r3,r8
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
290 vmrghw v8,v6,v7
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
291 vmrglw v9,v6,v7
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
292 stvx v8,r3,r10
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
293 stvx v9,r3,r11
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
294 .endif
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
295 addi r3,r3,32
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
296 addi r4,r4,16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
297 bdnz 1b
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
298 sub r3,r3,r5
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
299 blr
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
300 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
301
12089
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
302 #define M_SQRT1_2 0.70710678118654752440 /* 1/sqrt(2) */
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
303
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
304 #define WORD_0 0x00,0x01,0x02,0x03
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
305 #define WORD_1 0x04,0x05,0x06,0x07
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
306 #define WORD_2 0x08,0x09,0x0a,0x0b
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
307 #define WORD_3 0x0c,0x0d,0x0e,0x0f
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
308 #define WORD_s0 0x10,0x11,0x12,0x13
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
309 #define WORD_s1 0x14,0x15,0x16,0x17
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
310 #define WORD_s2 0x18,0x19,0x1a,0x1b
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
311 #define WORD_s3 0x1c,0x1d,0x1e,0x1f
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
312
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
313 #define vcprm(a, b, c, d) .byte WORD_##a, WORD_##b, WORD_##c, WORD_##d
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
314
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
315 .rodata
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
316 .align 4
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
317 fft_data:
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
318 .float 0, 0, 0, 0
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
319 .float 1, 0.92387953, M_SQRT1_2, 0.38268343
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
320 .float 0, 0.38268343, M_SQRT1_2, 0.92387953
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
321 .float -M_SQRT1_2, M_SQRT1_2, M_SQRT1_2,-M_SQRT1_2
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
322 .float M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, M_SQRT1_2
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
323 vcprm(s0,3,2,1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
324 vcprm(0,1,s2,s1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
325 vcprm(2,3,s0,s3)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
326 vcprm(2,s3,3,s2)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
327 vcprm(0,1,s0,s1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
328 vcprm(2,3,s2,s3)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
329 vcprm(2,3,0,1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
330 vcprm(1,2,s3,s0)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
331 vcprm(0,3,s2,s1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
332 vcprm(0,2,s1,s3)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
333 vcprm(1,3,s0,s2)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
334
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
335 .macro lvm b, r, regs:vararg
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
336 lvx \r, 0, \b
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
337 addi \b, \b, 16
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
338 .ifnb \regs
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
339 lvm \b, \regs
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
340 .endif
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
341 .endm
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
342
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
343 .macro stvm b, r, regs:vararg
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
344 stvx \r, 0, \b
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
345 addi \b, \b, 16
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
346 .ifnb \regs
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
347 stvm \b, \regs
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
348 .endif
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
349 .endm
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
350
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
351 .macro fft_calc interleave
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
352 extfunc ff_fft_calc\interleave\()_altivec
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
353 mflr r0
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
354 stp r0, 2*PS(r1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
355 stpu r1, -(160+16*PS)(r1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
356 addi r6, r1, 16*PS
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
357 stvm r6, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
358 mfvrsave r0
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
359 stw r0, 15*PS(r1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
360 li r6, 0xfffffffc
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
361 mtvrsave r6
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
362
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
363 movrel r6, fft_data
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
364 lvm r6, v14, v15, v16, v17, v18, v19, v20, v21
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
365 lvm r6, v22, v23, v24, v25, v26, v27, v28, v29
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
366
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
367 li r9, 16
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
368 movrel r12, X(ff_cos_tabs)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
369
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
370 movrel r6, fft_dispatch_tab\interleave\()_altivec
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
371 lwz r3, 0(r3)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
372 subi r3, r3, 2
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
373 slwi r3, r3, 2+ARCH_PPC64
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
374 lpx r3, r3, r6
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
375 mtctr r3
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
376 mr r3, r4
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
377 bctrl
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
378
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
379 addi r6, r1, 16*PS
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
380 lvm r6, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
381 lwz r6, 15*PS(r1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
382 mtvrsave r6
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
383 lp r1, 0(r1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
384 lp r0, 2*PS(r1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
385 mtlr r0
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
386 blr
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
387 .endm
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
388
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
389 .macro DECL_FFT suffix, bits, n, n2, n4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
390 fft\n\suffix\()_altivec:
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
391 mflr r0
12089
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
392 stp r0,PS*(\bits-3)(r1)
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
393 bl fft\n2\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
394 addi2 r3,\n*4
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
395 bl fft\n4\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
396 addi2 r3,\n*2
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
397 bl fft\n4\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
398 addi2 r3,\n*-6
12089
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
399 lp r0,PS*(\bits-3)(r1)
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
400 lp r4,\bits*PS(r12)
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
401 mtlr r0
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
402 li r5,\n/16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
403 b fft_pass\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
404 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
405
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
406 .macro DECL_FFTS interleave, suffix
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
407 .text
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
408 def_fft4 \suffix
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
409 def_fft8 \suffix
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
410 def_fft16 \suffix
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
411 PASS \interleave, \suffix
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
412 DECL_FFT \suffix, 5, 32, 16, 8
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
413 DECL_FFT \suffix, 6, 64, 32, 16
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
414 DECL_FFT \suffix, 7, 128, 64, 32
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
415 DECL_FFT \suffix, 8, 256, 128, 64
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
416 DECL_FFT \suffix, 9, 512, 256, 128
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
417 DECL_FFT \suffix,10, 1024, 512, 256
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
418 DECL_FFT \suffix,11, 2048, 1024, 512
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
419 DECL_FFT \suffix,12, 4096, 2048, 1024
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
420 DECL_FFT \suffix,13, 8192, 4096, 2048
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
421 DECL_FFT \suffix,14,16384, 8192, 4096
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
422 DECL_FFT \suffix,15,32768,16384, 8192
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
423 DECL_FFT \suffix,16,65536,32768,16384
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
424
12089
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
425 fft_calc \suffix
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
426
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
427 .rodata
12089
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
428 .align 3
5638941ec8ef PPC: convert Altivec FFT to pure assembler
mru
parents: 12088
diff changeset
429 fft_dispatch_tab\suffix\()_altivec:
12085
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
430 PTR fft4\suffix\()_altivec
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
431 PTR fft8\suffix\()_altivec
8454bb880008 PPC: add _interleave versions of fft{4,6,16}_altivec
mru
parents: 12046
diff changeset
432 PTR fft16\suffix\()_altivec
12046
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
433 PTR fft32\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
434 PTR fft64\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
435 PTR fft128\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
436 PTR fft256\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
437 PTR fft512\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
438 PTR fft1024\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
439 PTR fft2048\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
440 PTR fft4096\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
441 PTR fft8192\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
442 PTR fft16384\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
443 PTR fft32768\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
444 PTR fft65536\suffix\()_altivec
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
445 .endm
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
446
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
447 DECL_FFTS 0
ae57be2ef58c PPC: Altivec split-radix FFT
mru
parents:
diff changeset
448 DECL_FFTS 1, _interleave