Mercurial > libavcodec.hg
comparison ppc/fft_altivec.c @ 12046:ae57be2ef58c libavcodec
PPC: Altivec split-radix FFT
1.8x faster than altivec radix-2 on a G4
8% faster vorbis decoding
Patch (mostly) by Loren Merritt
author | mru |
---|---|
date | Thu, 01 Jul 2010 23:21:39 +0000 |
parents | 3cd4cd0509cd |
children | c80c7a717156 |
comparison
equal
deleted
inserted
replaced
12045:f399ea693f46 | 12046:ae57be2ef58c |
---|---|
1 /* | 1 /* |
2 * FFT/IFFT transforms | 2 * FFT/IFFT transforms |
3 * AltiVec-enabled | 3 * AltiVec-enabled |
4 * Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org> | 4 * Copyright (c) 2009 Loren Merritt |
5 * Based on code Copyright (c) 2002 Fabrice Bellard | |
6 * | 5 * |
7 * This file is part of FFmpeg. | 6 * This file is part of FFmpeg. |
8 * | 7 * |
9 * FFmpeg is free software; you can redistribute it and/or | 8 * FFmpeg is free software; you can redistribute it and/or |
10 * modify it under the terms of the GNU Lesser General Public | 9 * modify it under the terms of the GNU Lesser General Public |
20 * License along with FFmpeg; if not, write to the Free Software | 19 * License along with FFmpeg; if not, write to the Free Software |
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | 20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
22 */ | 21 */ |
23 #include "libavcodec/fft.h" | 22 #include "libavcodec/fft.h" |
24 #include "util_altivec.h" | 23 #include "util_altivec.h" |
25 #include "dsputil_altivec.h" | 24 #include "types_altivec.h" |
25 #include "regs.h" | |
26 | 26 |
27 /** | 27 /** |
28 * Do a complex FFT with the parameters defined in ff_fft_init(). The | 28 * Do a complex FFT with the parameters defined in ff_fft_init(). The |
29 * input data must be permuted before with s->revtab table. No | 29 * input data must be permuted before with s->revtab table. No |
30 * 1.0/sqrt(n) normalization is done. | 30 * 1.0/sqrt(n) normalization is done. |
31 * AltiVec-enabled | 31 * AltiVec-enabled |
32 * This code assumes that the 'z' pointer is 16 bytes-aligned | 32 * This code assumes that the 'z' pointer is 16 bytes-aligned |
33 * It also assumes all FFTComplex are 8 bytes-aligned pair of float | 33 * It also assumes all FFTComplex are 8 bytes-aligned pair of float |
34 * The code is exactly the same as the SSE version, except | |
35 * that successive MUL + ADD/SUB have been merged into | |
36 * fused multiply-add ('vec_madd' in altivec) | |
37 */ | 34 */ |
35 | |
36 // Pointers to functions. Not using function pointer syntax, because | |
37 // that involves an extra level of indirection on some PPC ABIs. | |
38 extern void *ff_fft_dispatch_altivec[2][15]; | |
39 | |
40 // Convert from simd order to C order. | |
41 static void swizzle(vec_f *z, int n) | |
42 { | |
43 int i; | |
44 n >>= 1; | |
45 for (i = 0; i < n; i += 2) { | |
46 vec_f re = z[i]; | |
47 vec_f im = z[i+1]; | |
48 z[i] = vec_mergeh(re, im); | |
49 z[i+1] = vec_mergel(re, im); | |
50 } | |
51 } | |
52 | |
38 static void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) | 53 static void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) |
39 { | 54 { |
40 register const vector float vczero = (const vector float)vec_splat_u32(0.); | 55 register vec_f v14 __asm__("v14") = {0,0,0,0}; |
41 | 56 register vec_f v15 __asm__("v15") = *(const vec_f*)ff_cos_16; |
42 int ln = s->nbits; | 57 register vec_f v16 __asm__("v16") = {0, 0.38268343, M_SQRT1_2, 0.92387953}; |
43 int j, np, np2; | 58 register vec_f v17 __asm__("v17") = {-M_SQRT1_2, M_SQRT1_2, M_SQRT1_2,-M_SQRT1_2}; |
44 int nblocks, nloops; | 59 register vec_f v18 __asm__("v18") = { M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, M_SQRT1_2}; |
45 register FFTComplex *p, *q; | 60 register vec_u8 v19 __asm__("v19") = vcprm(s0,3,2,1); |
46 FFTComplex *cptr, *cptr1; | 61 register vec_u8 v20 __asm__("v20") = vcprm(0,1,s2,s1); |
47 int k; | 62 register vec_u8 v21 __asm__("v21") = vcprm(2,3,s0,s3); |
48 | 63 register vec_u8 v22 __asm__("v22") = vcprm(2,s3,3,s2); |
49 np = 1 << ln; | 64 register vec_u8 v23 __asm__("v23") = vcprm(0,1,s0,s1); |
50 | 65 register vec_u8 v24 __asm__("v24") = vcprm(2,3,s2,s3); |
51 { | 66 register vec_u8 v25 __asm__("v25") = vcprm(2,3,0,1); |
52 vector float *r, a, b, a1, c1, c2; | 67 register vec_u8 v26 __asm__("v26") = vcprm(1,2,s3,s0); |
53 | 68 register vec_u8 v27 __asm__("v27") = vcprm(0,3,s2,s1); |
54 r = (vector float *)&z[0]; | 69 register vec_u8 v28 __asm__("v28") = vcprm(0,2,s1,s3); |
55 | 70 register vec_u8 v29 __asm__("v29") = vcprm(1,3,s0,s2); |
56 c1 = vcii(p,p,n,n); | 71 register FFTSample *const*cos_tabs __asm__("r12") = ff_cos_tabs; |
57 | 72 register FFTComplex *zarg __asm__("r3") = z; |
58 if (s->inverse) { | 73 __asm__( |
59 c2 = vcii(p,p,n,p); | 74 "mtctr %0 \n" |
60 } else { | 75 "li "r(9)", 16 \n" |
61 c2 = vcii(p,p,p,n); | 76 "subi "r(1)","r(1) ",%1 \n" |
62 } | 77 "bctrl \n" |
63 | 78 "addi "r(1)","r(1) ",%1 \n" |
64 j = (np >> 2); | 79 ::"r"(ff_fft_dispatch_altivec[1][s->nbits-2]), "i"(12*sizeof(void*)), |
65 do { | 80 "r"(zarg), "r"(cos_tabs), |
66 a = vec_ld(0, r); | 81 "v"(v14),"v"(v15),"v"(v16),"v"(v17),"v"(v18),"v"(v19),"v"(v20),"v"(v21), |
67 a1 = vec_ld(sizeof(vector float), r); | 82 "v"(v22),"v"(v23),"v"(v24),"v"(v25),"v"(v26),"v"(v27),"v"(v28),"v"(v29) |
68 | 83 : "lr","ctr","r0","r4","r5","r6","r7","r8","r9","r10","r11", |
69 b = vec_perm(a,a,vcprmle(1,0,3,2)); | 84 "v0","v1","v2","v3","v4","v5","v6","v7","v8","v9","v10","v11","v12","v13" |
70 a = vec_madd(a,c1,b); | 85 ); |
71 /* do the pass 0 butterfly */ | 86 if (s->nbits <= 4) |
72 | 87 swizzle((vec_f*)z, 1<<s->nbits); |
73 b = vec_perm(a1,a1,vcprmle(1,0,3,2)); | |
74 b = vec_madd(a1,c1,b); | |
75 /* do the pass 0 butterfly */ | |
76 | |
77 /* multiply third by -i */ | |
78 b = vec_perm(b,b,vcprmle(2,3,1,0)); | |
79 | |
80 /* do the pass 1 butterfly */ | |
81 vec_st(vec_madd(b,c2,a), 0, r); | |
82 vec_st(vec_nmsub(b,c2,a), sizeof(vector float), r); | |
83 | |
84 r += 2; | |
85 } while (--j != 0); | |
86 } | |
87 /* pass 2 .. ln-1 */ | |
88 | |
89 nblocks = np >> 3; | |
90 nloops = 1 << 2; | |
91 np2 = np >> 1; | |
92 | |
93 cptr1 = s->exptab1; | |
94 do { | |
95 p = z; | |
96 q = z + nloops; | |
97 j = nblocks; | |
98 do { | |
99 cptr = cptr1; | |
100 k = nloops >> 1; | |
101 do { | |
102 vector float a,b,c,t1; | |
103 | |
104 a = vec_ld(0, (float*)p); | |
105 b = vec_ld(0, (float*)q); | |
106 | |
107 /* complex mul */ | |
108 c = vec_ld(0, (float*)cptr); | |
109 /* cre*re cim*re */ | |
110 t1 = vec_madd(c, vec_perm(b,b,vcprmle(2,2,0,0)),vczero); | |
111 c = vec_ld(sizeof(vector float), (float*)cptr); | |
112 /* -cim*im cre*im */ | |
113 b = vec_madd(c, vec_perm(b,b,vcprmle(3,3,1,1)),t1); | |
114 | |
115 /* butterfly */ | |
116 vec_st(vec_add(a,b), 0, (float*)p); | |
117 vec_st(vec_sub(a,b), 0, (float*)q); | |
118 | |
119 p += 2; | |
120 q += 2; | |
121 cptr += 4; | |
122 } while (--k); | |
123 | |
124 p += nloops; | |
125 q += nloops; | |
126 } while (--j); | |
127 cptr1 += nloops * 2; | |
128 nblocks = nblocks >> 1; | |
129 nloops = nloops << 1; | |
130 } while (nblocks != 0); | |
131 } | 88 } |
132 | 89 |
133 av_cold void ff_fft_init_altivec(FFTContext *s) | 90 av_cold void ff_fft_init_altivec(FFTContext *s) |
134 { | 91 { |
135 s->fft_calc = ff_fft_calc_altivec; | 92 if (HAVE_GNU_AS) |
136 s->split_radix = 0; | 93 s->fft_calc = ff_fft_calc_altivec; |
94 s->split_radix = 1; | |
137 } | 95 } |