Mercurial > libavcodec.hg
annotate ppc/fft_altivec.c @ 2504:f12657081093 libavcodec
INTRA PCM macroblocks support patch by (Loic )lll+ffmpeg m4x org)
This patch adds the support for INTRA PCM macroblocks in CAVLC and CABAC
mode, the deblocking needed a small modification and so did the
intra4x4_pred_mode prediction.
With this patch, the 5 streams of the conformance suite containing INTRA
PCM macroblocks now decode entirely, 4 are completely corrects, 1 is
incorrect since the first B slice because of deblocking in B slice not
yet implemented.
The code is not optimized for speed, it is not necessary IPCM
macroblocks are rare, but it could be optimized for code size, if
someone want to do this, feel free.
author | michael |
---|---|
date | Mon, 07 Feb 2005 00:10:28 +0000 |
parents | dd63cb7e5080 |
children | ef2149182f1c |
rev | line source |
---|---|
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
1 /* |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
2 * FFT/IFFT transforms |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
3 * AltiVec-enabled |
981 | 4 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
5 * Based on code Copyright (c) 2002 Fabrice Bellard. |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
6 * |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
7 * This library is free software; you can redistribute it and/or |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
8 * modify it under the terms of the GNU Lesser General Public |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
9 * License as published by the Free Software Foundation; either |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
10 * version 2 of the License, or (at your option) any later version. |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
11 * |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
12 * This library is distributed in the hope that it will be useful, |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
15 * Lesser General Public License for more details. |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
16 * |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
17 * You should have received a copy of the GNU Lesser General Public |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
18 * License along with this library; if not, write to the Free Software |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
20 */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
21 #include "../dsputil.h" |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
22 |
1277
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1033
diff
changeset
|
23 #include "gcc_fixes.h" |
f3152eb76f1a
altivec gcc-3 fixes by (Magnus Damm <damm at opensource dot se>)
michaelni
parents:
1033
diff
changeset
|
24 |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
25 #include "dsputil_altivec.h" |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
26 |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
27 /* |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
28 those three macros are from libavcodec/fft.c |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
29 and are required for the reference C code |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
30 */ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
31 /* butter fly op */ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
32 #define BF(pre, pim, qre, qim, pre1, pim1, qre1, qim1) \ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
33 {\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
34 FFTSample ax, ay, bx, by;\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
35 bx=pre1;\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
36 by=pim1;\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
37 ax=qre1;\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
38 ay=qim1;\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
39 pre = (bx + ax);\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
40 pim = (by + ay);\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
41 qre = (bx - ax);\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
42 qim = (by - ay);\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
43 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
44 #define MUL16(a,b) ((a) * (b)) |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
45 #define CMUL(pre, pim, are, aim, bre, bim) \ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
46 {\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
47 pre = (MUL16(are, bre) - MUL16(aim, bim));\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
48 pim = (MUL16(are, bim) + MUL16(bre, aim));\ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
49 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
50 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
51 |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
52 /** |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1352
diff
changeset
|
53 * Do a complex FFT with the parameters defined in ff_fft_init(). The |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
54 * input data must be permuted before with s->revtab table. No |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
55 * 1.0/sqrt(n) normalization is done. |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
56 * AltiVec-enabled |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
57 * This code assumes that the 'z' pointer is 16 bytes-aligned |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
58 * It also assumes all FFTComplex are 8 bytes-aligned pair of float |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
59 * The code is exactly the same as the SSE version, except |
981 | 60 * that successive MUL + ADD/SUB have been merged into |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
61 * fused multiply-add ('vec_madd' in altivec) |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
62 */ |
1879
dd63cb7e5080
fft_*() renamed into ff_fft_*() patch by (Gildas Bazin <gbazin at altern dot org>)
michael
parents:
1352
diff
changeset
|
63 void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
64 { |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
65 POWERPC_PERF_DECLARE(altivec_fft_num, s->nbits >= 6); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
66 #ifdef ALTIVEC_USE_REFERENCE_C_CODE |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
67 int ln = s->nbits; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
68 int j, np, np2; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
69 int nblocks, nloops; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
70 register FFTComplex *p, *q; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
71 FFTComplex *exptab = s->exptab; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
72 int l; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
73 FFTSample tmp_re, tmp_im; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
74 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
75 POWERPC_PERF_START_COUNT(altivec_fft_num, s->nbits >= 6); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
76 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
77 np = 1 << ln; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
78 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
79 /* pass 0 */ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
80 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
81 p=&z[0]; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
82 j=(np >> 1); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
83 do { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
84 BF(p[0].re, p[0].im, p[1].re, p[1].im, |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
85 p[0].re, p[0].im, p[1].re, p[1].im); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
86 p+=2; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
87 } while (--j != 0); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
88 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
89 /* pass 1 */ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
90 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
91 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
92 p=&z[0]; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
93 j=np >> 2; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
94 if (s->inverse) { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
95 do { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
96 BF(p[0].re, p[0].im, p[2].re, p[2].im, |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
97 p[0].re, p[0].im, p[2].re, p[2].im); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
98 BF(p[1].re, p[1].im, p[3].re, p[3].im, |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
99 p[1].re, p[1].im, -p[3].im, p[3].re); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
100 p+=4; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
101 } while (--j != 0); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
102 } else { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
103 do { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
104 BF(p[0].re, p[0].im, p[2].re, p[2].im, |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
105 p[0].re, p[0].im, p[2].re, p[2].im); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
106 BF(p[1].re, p[1].im, p[3].re, p[3].im, |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
107 p[1].re, p[1].im, p[3].im, -p[3].re); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
108 p+=4; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
109 } while (--j != 0); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
110 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
111 /* pass 2 .. ln-1 */ |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
112 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
113 nblocks = np >> 3; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
114 nloops = 1 << 2; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
115 np2 = np >> 1; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
116 do { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
117 p = z; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
118 q = z + nloops; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
119 for (j = 0; j < nblocks; ++j) { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
120 BF(p->re, p->im, q->re, q->im, |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
121 p->re, p->im, q->re, q->im); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
122 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
123 p++; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
124 q++; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
125 for(l = nblocks; l < np2; l += nblocks) { |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
126 CMUL(tmp_re, tmp_im, exptab[l].re, exptab[l].im, q->re, q->im); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
127 BF(p->re, p->im, q->re, q->im, |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
128 p->re, p->im, tmp_re, tmp_im); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
129 p++; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
130 q++; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
131 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
132 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
133 p += nloops; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
134 q += nloops; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
135 } |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
136 nblocks = nblocks >> 1; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
137 nloops = nloops << 1; |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
138 } while (nblocks != 0); |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
139 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
140 POWERPC_PERF_STOP_COUNT(altivec_fft_num, s->nbits >= 6); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
141 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
142 #else /* ALTIVEC_USE_REFERENCE_C_CODE */ |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
143 #ifdef CONFIG_DARWIN |
995
edc10966b081
altivec jumbo patch by (Romain Dolbeau <dolbeaur at club-internet dot fr>)
michaelni
parents:
981
diff
changeset
|
144 register const vector float vczero = (const vector float)(0.); |
1033
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
145 #else |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
146 register const vector float vczero = (const vector float){0.,0.,0.,0.}; |
b4172ff70d27
Altivec on non darwin systems patch by Romain Dolbeau
bellard
parents:
1015
diff
changeset
|
147 #endif |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
148 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
149 int ln = s->nbits; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
150 int j, np, np2; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
151 int nblocks, nloops; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
152 register FFTComplex *p, *q; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
153 FFTComplex *cptr, *cptr1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
154 int k; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
155 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
156 POWERPC_PERF_START_COUNT(altivec_fft_num, s->nbits >= 6); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
157 |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
158 np = 1 << ln; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
159 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
160 { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
161 vector float *r, a, b, a1, c1, c2; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
162 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
163 r = (vector float *)&z[0]; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
164 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
165 c1 = vcii(p,p,n,n); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
166 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
167 if (s->inverse) |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
168 { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
169 c2 = vcii(p,p,n,p); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
170 } |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
171 else |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
172 { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
173 c2 = vcii(p,p,p,n); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
174 } |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
175 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
176 j = (np >> 2); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
177 do { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
178 a = vec_ld(0, r); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
179 a1 = vec_ld(sizeof(vector float), r); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
180 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
181 b = vec_perm(a,a,vcprmle(1,0,3,2)); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
182 a = vec_madd(a,c1,b); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
183 /* do the pass 0 butterfly */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
184 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
185 b = vec_perm(a1,a1,vcprmle(1,0,3,2)); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
186 b = vec_madd(a1,c1,b); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
187 /* do the pass 0 butterfly */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
188 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
189 /* multiply third by -i */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
190 b = vec_perm(b,b,vcprmle(2,3,1,0)); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
191 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
192 /* do the pass 1 butterfly */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
193 vec_st(vec_madd(b,c2,a), 0, r); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
194 vec_st(vec_nmsub(b,c2,a), sizeof(vector float), r); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
195 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
196 r += 2; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
197 } while (--j != 0); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
198 } |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
199 /* pass 2 .. ln-1 */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
200 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
201 nblocks = np >> 3; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
202 nloops = 1 << 2; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
203 np2 = np >> 1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
204 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
205 cptr1 = s->exptab1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
206 do { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
207 p = z; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
208 q = z + nloops; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
209 j = nblocks; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
210 do { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
211 cptr = cptr1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
212 k = nloops >> 1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
213 do { |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
214 vector float a,b,c,t1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
215 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
216 a = vec_ld(0, (float*)p); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
217 b = vec_ld(0, (float*)q); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
218 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
219 /* complex mul */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
220 c = vec_ld(0, (float*)cptr); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
221 /* cre*re cim*re */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
222 t1 = vec_madd(c, vec_perm(b,b,vcprmle(2,2,0,0)),vczero); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
223 c = vec_ld(sizeof(vector float), (float*)cptr); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
224 /* -cim*im cre*im */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
225 b = vec_madd(c, vec_perm(b,b,vcprmle(3,3,1,1)),t1); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
226 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
227 /* butterfly */ |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
228 vec_st(vec_add(a,b), 0, (float*)p); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
229 vec_st(vec_sub(a,b), 0, (float*)q); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
230 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
231 p += 2; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
232 q += 2; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
233 cptr += 4; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
234 } while (--k); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
235 |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
236 p += nloops; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
237 q += nloops; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
238 } while (--j); |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
239 cptr1 += nloops * 2; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
240 nblocks = nblocks >> 1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
241 nloops = nloops << 1; |
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
242 } while (nblocks != 0); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
243 |
1352
e8ff4783f188
1) remove TBL support in PPC performance. It's much more useful to use the
michaelni
parents:
1277
diff
changeset
|
244 POWERPC_PERF_STOP_COUNT(altivec_fft_num, s->nbits >= 6); |
1009
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
245 |
3b7cc8e4b83f
AltiVec perf (take 2), plus a couple AltiVec functions by (Romain Dolbeau <dolbeau at irisa dot fr>)
michaelni
parents:
995
diff
changeset
|
246 #endif /* ALTIVEC_USE_REFERENCE_C_CODE */ |
975
e05d525505c5
fft altivec by Romain Dolbeau - simplified patch, test it on PPC with fft-test and wma decoding
bellard
parents:
diff
changeset
|
247 } |