annotate liba52/srfftp_3dnow.h @ 885:cec1562ccf8a

Improvements
author nickols_k
date Sat, 26 May 2001 10:29:41 +0000
parents
children cb432deedb92
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
885
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
1 /*
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
2 * srfftp.h
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
3 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
4 * Copyright (C) Yuqing Deng <Yuqing_Deng@brown.edu> - April 2000
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
5 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
6 * 64 and 128 point split radix fft for ac3dec
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
7 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
8 * The algorithm is desribed in the book:
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
9 * "Computational Frameworks of the Fast Fourier Transform".
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
10 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
11 * The ideas and the the organization of code borrowed from djbfft written by
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
12 * D. J. Bernstein <djb@cr.py.to>. djbff can be found at
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
13 * http://cr.yp.to/djbfft.html.
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
14 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
15 * srfftp.h is free software; you can redistribute it and/or modify
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
16 * it under the terms of the GNU General Public License as published by
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
17 * the Free Software Foundation; either version 2, or (at your option)
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
18 * any later version.
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
19 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
20 * srfftp.h is distributed in the hope that it will be useful,
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
23 * GNU General Public License for more details.
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
24 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
25 * You should have received a copy of the GNU General Public License
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
26 * along with GNU Make; see the file COPYING. If not, write to
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
27 * the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
28 *
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
29 * Modified for using AMD's 3DNow! - 3DNowEx(DSP)! SIMD operations
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
30 * by Nick Kurshev <nickols_k@mail.ru>
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
31 */
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
32
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
33 #ifndef SRFFTP_3DNOW_H__
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
34 #define SRFFTP_3DNOW_H__
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
35
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
36 #ifdef HAVE_3DNOWEX
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
37 #define TRANS_FILL_MM6_MM7_3DNOW()\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
38 asm(\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
39 "movl $-1, %%eax\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
40 "movd %%eax, %%mm7\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
41 "negl %%eax\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
42 "movd %%eax, %%mm6\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
43 "punpckldq %%mm6, %%mm7\n\t" /* -1.0 | 1.0 */\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
44 "pi2fd %%mm7, %%mm7\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
45 "pswapd %%mm7, %%mm6\n\t"/* 1.0 | -1.0 */\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
46 :::"eax","memory");
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
47 #else
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
48 #define TRANS_FILL_MM6_MM7_3DNOW()\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
49 asm(\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
50 "movl $-1, %%eax\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
51 "movd %%eax, %%mm7\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
52 "negl %%eax\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
53 "movd %%eax, %%mm6\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
54 "punpckldq %%mm6, %%mm7\n\t" /* -1.0 | 1.0 */\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
55 "punpckldq %%mm7, %%mm6\n\t" /* 1.0 | -1.0 */\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
56 "pi2fd %%mm7, %%mm7\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
57 "pi2fd %%mm6, %%mm6\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
58 :::"eax","memory");
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
59 #endif
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
60
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
61 #ifdef HAVE_3DNOWEX
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
62 #define PSWAP_MM(mm_base,mm_hlp) "pswapd "##mm_base","##mm_base" \n\t"
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
63 #else
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
64 #define PSWAP_MM(mm_base,mm_hlp)\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
65 "movq "##mm_base","##mm_hlp" \n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
66 "psrlq $32, "##mm_base"\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
67 "punpckldq "##mm_hlp","##mm_base"\n\t"
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
68 #endif
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
69
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
70 #define TRANSZERO_3DNOW(A0,A4,A8,A12) \
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
71 { \
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
72 asm volatile("femms":::"memory");\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
73 TRANS_FILL_MM6_MM7_3DNOW()\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
74 asm(\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
75 "movq %4, %%mm0\n\t" /* mm0 = wTB[0]*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
76 "movq %5, %%mm1\n\t" /* mm1 = wTB[k*2]*/ \
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
77 "movq %%mm0, %%mm5\n\t"/*u.re = wTB[0].re + wTB[k*2].re;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
78 "pfadd %%mm1, %%mm5\n\t"/*u.im = wTB[0].im + wTB[k*2].im; mm5 = u*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
79 "pfmul %%mm6, %%mm0\n\t"/*mm0 = wTB[0].re | -wTB[0].im */\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
80 "pfmul %%mm7, %%mm1\n\t"/*mm1 = -wTB[k*2].re | wTB[k*2].im */\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
81 "pfadd %%mm1, %%mm0\n\t"/*v.im = wTB[0].re - wTB[k*2].re;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
82 "movq %%mm0, %%mm4\n\t"/*v.re =-wTB[0].im + wTB[k*2].im;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
83 PSWAP_MM("%%mm4","%%mm2")/* mm4 = v*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
84 "movq %6, %%mm0\n\t" /* a1 = A0;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
85 "movq %%mm0, %%mm1\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
86 "pfadd %%mm5, %%mm0\n\t" /*A0 = a1 + u;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
87 "pfsub %%mm5, %%mm1\n\t" /*A1 = a1 - u;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
88 "movq %%mm0, %0\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
89 "movq %%mm1, %1\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
90 "movq %7, %%mm2\n\t" /* a1 = A4;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
91 "movq %%mm2, %%mm3\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
92 "pfadd %%mm4, %%mm2\n\t" /*A12 = a1 + v;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
93 "pfsub %%mm4, %%mm3\n\t" /*A4 = a1 - v;*/\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
94 "movq %%mm2, %3\n\t"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
95 "movq %%mm3, %2"\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
96 :"=m"(A0), "=m"(A8), "=m"(A4), "=m"(A12)\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
97 :"m"(wTB[0]), "m"(wTB[k*2]), "0"(A0), "2"(A4)\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
98 :"memory");\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
99 asm volatile("femms":::"memory");\
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
100 }
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
101
cec1562ccf8a Improvements
nickols_k
parents:
diff changeset
102 #endif