Mercurial > libavcodec.hg
comparison i386/fft_3dn2.c @ 7542:a8a8205a9081 libavcodec
split-radix FFT
c is 1.9x faster than previous c (on various x86 cpus), sse is 1.6x faster than previous sse.
author | lorenm |
---|---|
date | Tue, 12 Aug 2008 00:26:58 +0000 |
parents | fc843d00867c |
children | ee1cb5ab9f99 |
comparison
equal
deleted
inserted
replaced
7541:570c0c027998 | 7542:a8a8205a9081 |
---|---|
21 */ | 21 */ |
22 | 22 |
23 #include "libavutil/x86_cpu.h" | 23 #include "libavutil/x86_cpu.h" |
24 #include "libavcodec/dsputil.h" | 24 #include "libavcodec/dsputil.h" |
25 | 25 |
26 static const int p1m1[2] __attribute__((aligned(8))) = | 26 #ifdef EMULATE_3DNOWEXT |
27 { 0, 1 << 31 }; | 27 #define ff_fft_calc_3dn2 ff_fft_calc_3dn |
28 #define ff_fft_dispatch_3dn2 ff_fft_dispatch_3dn | |
29 #define ff_fft_dispatch_interleave_3dn2 ff_fft_dispatch_interleave_3dn | |
30 #define ff_imdct_calc_3dn2 ff_imdct_calc_3dn | |
31 #define ff_imdct_half_3dn2 ff_imdct_half_3dn | |
32 #endif | |
28 | 33 |
29 static const int m1p1[2] __attribute__((aligned(8))) = | 34 void ff_fft_dispatch_3dn2(FFTComplex *z, int nbits); |
30 { 1 << 31, 0 }; | 35 void ff_fft_dispatch_interleave_3dn2(FFTComplex *z, int nbits); |
31 | 36 |
32 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z) | 37 void ff_fft_calc_3dn2(FFTContext *s, FFTComplex *z) |
33 { | 38 { |
34 int ln = s->nbits; | 39 int n = 1<<s->nbits; |
35 long j; | 40 int i; |
36 x86_reg i; | 41 ff_fft_dispatch_interleave_3dn2(z, s->nbits); |
37 long nblocks, nloops; | |
38 FFTComplex *p, *cptr; | |
39 | |
40 asm volatile( | |
41 /* FEMMS is not a must here but recommended by AMD */ | |
42 "femms \n\t" | |
43 "movq %0, %%mm7 \n\t" | |
44 ::"m"(*(s->inverse ? m1p1 : p1m1)) | |
45 ); | |
46 | |
47 i = 8 << ln; | |
48 asm volatile( | |
49 "1: \n\t" | |
50 "sub $32, %0 \n\t" | |
51 "movq (%0,%1), %%mm0 \n\t" | |
52 "movq 16(%0,%1), %%mm1 \n\t" | |
53 "movq 8(%0,%1), %%mm2 \n\t" | |
54 "movq 24(%0,%1), %%mm3 \n\t" | |
55 "movq %%mm0, %%mm4 \n\t" | |
56 "movq %%mm1, %%mm5 \n\t" | |
57 "pfadd %%mm2, %%mm0 \n\t" | |
58 "pfadd %%mm3, %%mm1 \n\t" | |
59 "pfsub %%mm2, %%mm4 \n\t" | |
60 "pfsub %%mm3, %%mm5 \n\t" | |
61 "movq %%mm0, %%mm2 \n\t" | |
62 "pswapd %%mm5, %%mm5 \n\t" | |
63 "movq %%mm4, %%mm3 \n\t" | |
64 "pxor %%mm7, %%mm5 \n\t" | |
65 "pfadd %%mm1, %%mm0 \n\t" | |
66 "pfadd %%mm5, %%mm4 \n\t" | |
67 "pfsub %%mm1, %%mm2 \n\t" | |
68 "pfsub %%mm5, %%mm3 \n\t" | |
69 "movq %%mm0, (%0,%1) \n\t" | |
70 "movq %%mm4, 8(%0,%1) \n\t" | |
71 "movq %%mm2, 16(%0,%1) \n\t" | |
72 "movq %%mm3, 24(%0,%1) \n\t" | |
73 "jg 1b \n\t" | |
74 :"+r"(i) | |
75 :"r"(z) | |
76 ); | |
77 /* pass 2 .. ln-1 */ | |
78 | |
79 nblocks = 1 << (ln-3); | |
80 nloops = 1 << 2; | |
81 cptr = s->exptab1; | |
82 do { | |
83 p = z; | |
84 j = nblocks; | |
85 do { | |
86 i = nloops*8; | |
87 asm volatile( | |
88 "1: \n\t" | |
89 "sub $16, %0 \n\t" | |
90 "movq (%1,%0), %%mm0 \n\t" | |
91 "movq 8(%1,%0), %%mm1 \n\t" | |
92 "movq (%2,%0), %%mm2 \n\t" | |
93 "movq 8(%2,%0), %%mm3 \n\t" | |
94 "movq (%3,%0,2), %%mm4 \n\t" | |
95 "movq 8(%3,%0,2), %%mm5 \n\t" | |
96 "pswapd %%mm4, %%mm6 \n\t" // no need for cptr[2] & cptr[3] | |
97 "pswapd %%mm5, %%mm7 \n\t" | |
98 "pfmul %%mm2, %%mm4 \n\t" // cre*re cim*im | |
99 "pfmul %%mm3, %%mm5 \n\t" | |
100 "pfmul %%mm2, %%mm6 \n\t" // cim*re cre*im | |
101 "pfmul %%mm3, %%mm7 \n\t" | |
102 "pfpnacc %%mm6, %%mm4 \n\t" // cre*re-cim*im cim*re+cre*im | |
103 "pfpnacc %%mm7, %%mm5 \n\t" | |
104 "movq %%mm0, %%mm2 \n\t" | |
105 "movq %%mm1, %%mm3 \n\t" | |
106 "pfadd %%mm4, %%mm0 \n\t" | |
107 "pfadd %%mm5, %%mm1 \n\t" | |
108 "pfsub %%mm4, %%mm2 \n\t" | |
109 "pfsub %%mm5, %%mm3 \n\t" | |
110 "movq %%mm0, (%1,%0) \n\t" | |
111 "movq %%mm1, 8(%1,%0) \n\t" | |
112 "movq %%mm2, (%2,%0) \n\t" | |
113 "movq %%mm3, 8(%2,%0) \n\t" | |
114 "jg 1b \n\t" | |
115 :"+r"(i) | |
116 :"r"(p), "r"(p + nloops), "r"(cptr) | |
117 ); | |
118 p += nloops*2; | |
119 } while (--j); | |
120 cptr += nloops*2; | |
121 nblocks >>= 1; | |
122 nloops <<= 1; | |
123 } while (nblocks != 0); | |
124 asm volatile("femms"); | 42 asm volatile("femms"); |
43 if(n <= 8) | |
44 for(i=0; i<n; i+=2) | |
45 FFSWAP(FFTSample, z[i].im, z[i+1].re); | |
125 } | 46 } |
126 | 47 |
127 static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp) | 48 static void imdct_3dn2(MDCTContext *s, const FFTSample *input, FFTSample *tmp) |
128 { | 49 { |
129 long n4, n2, n; | 50 long n4, n2, n; |
160 "movq %%mm0, %0 \n\t" | 81 "movq %%mm0, %0 \n\t" |
161 :"=m"(z[revtab[k]]) | 82 :"=m"(z[revtab[k]]) |
162 ); | 83 ); |
163 } | 84 } |
164 | 85 |
165 ff_fft_calc(&s->fft, z); | 86 ff_fft_calc_3dn2(&s->fft, z); |
166 | 87 |
167 /* post rotation + reordering */ | 88 /* post rotation + reordering */ |
168 for(k = 0; k < n4; k++) { | 89 for(k = 0; k < n4; k++) { |
169 asm volatile( | 90 asm volatile( |
170 "movq %0, %%mm0 \n\t" | 91 "movq %0, %%mm0 \n\t" |