Mercurial > libavcodec.hg
comparison x86/fft_sse.c @ 12399:020540442072 libavcodec
Convert ff_imdct_half_sse() to yasm.
This is to avoid split asm sections that attempt to preserve some
registers between sections.
author | alexc |
---|---|
date | Sun, 22 Aug 2010 14:39:58 +0000 |
parents | db9ef48dc0e4 |
children | b64b8e5a2d3a |
comparison
equal
deleted
inserted
replaced
12398:31736ceab3aa | 12399:020540442072 |
---|---|
69 ); | 69 ); |
70 } | 70 } |
71 memcpy(z, s->tmp_buf, n*sizeof(FFTComplex)); | 71 memcpy(z, s->tmp_buf, n*sizeof(FFTComplex)); |
72 } | 72 } |
73 | 73 |
74 void ff_imdct_half_sse(FFTContext *s, FFTSample *output, const FFTSample *input) | |
75 { | |
76 av_unused x86_reg i, j, k, l; | |
77 long n = 1 << s->mdct_bits; | |
78 long n2 = n >> 1; | |
79 long n4 = n >> 2; | |
80 long n8 = n >> 3; | |
81 const uint16_t *revtab = s->revtab + n8; | |
82 const FFTSample *tcos = s->tcos; | |
83 const FFTSample *tsin = s->tsin; | |
84 FFTComplex *z = (FFTComplex *)output; | |
85 | |
86 /* pre rotation */ | |
87 for(k=n8-2; k>=0; k-=2) { | |
88 __asm__ volatile( | |
89 "movaps (%2,%1,2), %%xmm0 \n" // { z[k].re, z[k].im, z[k+1].re, z[k+1].im } | |
90 "movaps -16(%2,%0,2), %%xmm1 \n" // { z[-k-2].re, z[-k-2].im, z[-k-1].re, z[-k-1].im } | |
91 "movaps %%xmm0, %%xmm2 \n" | |
92 "shufps $0x88, %%xmm1, %%xmm0 \n" // { z[k].re, z[k+1].re, z[-k-2].re, z[-k-1].re } | |
93 "shufps $0x77, %%xmm2, %%xmm1 \n" // { z[-k-1].im, z[-k-2].im, z[k+1].im, z[k].im } | |
94 "movlps (%3,%1), %%xmm4 \n" | |
95 "movlps (%4,%1), %%xmm5 \n" | |
96 "movhps -8(%3,%0), %%xmm4 \n" // { cos[k], cos[k+1], cos[-k-2], cos[-k-1] } | |
97 "movhps -8(%4,%0), %%xmm5 \n" // { sin[k], sin[k+1], sin[-k-2], sin[-k-1] } | |
98 "movaps %%xmm0, %%xmm2 \n" | |
99 "movaps %%xmm1, %%xmm3 \n" | |
100 "mulps %%xmm5, %%xmm0 \n" // re*sin | |
101 "mulps %%xmm4, %%xmm1 \n" // im*cos | |
102 "mulps %%xmm4, %%xmm2 \n" // re*cos | |
103 "mulps %%xmm5, %%xmm3 \n" // im*sin | |
104 "subps %%xmm0, %%xmm1 \n" // -> re | |
105 "addps %%xmm3, %%xmm2 \n" // -> im | |
106 "movaps %%xmm1, %%xmm0 \n" | |
107 "unpcklps %%xmm2, %%xmm1 \n" // { z[k], z[k+1] } | |
108 "unpckhps %%xmm2, %%xmm0 \n" // { z[-k-2], z[-k-1] } | |
109 ::"r"(-4*k), "r"(4*k), | |
110 "r"(input+n4), "r"(tcos+n8), "r"(tsin+n8) | |
111 ); | |
112 #if ARCH_X86_64 | |
113 // if we have enough regs, don't let gcc make the luts latency-bound | |
114 // but if not, latency is faster than spilling | |
115 __asm__("movlps %%xmm0, %0 \n" | |
116 "movhps %%xmm0, %1 \n" | |
117 "movlps %%xmm1, %2 \n" | |
118 "movhps %%xmm1, %3 \n" | |
119 :"=m"(z[revtab[-k-2]]), | |
120 "=m"(z[revtab[-k-1]]), | |
121 "=m"(z[revtab[ k ]]), | |
122 "=m"(z[revtab[ k+1]]) | |
123 ); | |
124 #else | |
125 __asm__("movlps %%xmm0, %0" :"=m"(z[revtab[-k-2]])); | |
126 __asm__("movhps %%xmm0, %0" :"=m"(z[revtab[-k-1]])); | |
127 __asm__("movlps %%xmm1, %0" :"=m"(z[revtab[ k ]])); | |
128 __asm__("movhps %%xmm1, %0" :"=m"(z[revtab[ k+1]])); | |
129 #endif | |
130 } | |
131 | |
132 ff_fft_dispatch_sse(z, s->nbits); | |
133 | |
134 /* post rotation + reinterleave + reorder */ | |
135 | |
136 #define CMUL(j,xmm0,xmm1)\ | |
137 "movaps (%2,"#j",2), %%xmm6 \n"\ | |
138 "movaps 16(%2,"#j",2), "#xmm0"\n"\ | |
139 "movaps %%xmm6, "#xmm1"\n"\ | |
140 "movaps "#xmm0",%%xmm7 \n"\ | |
141 "mulps (%3,"#j"), %%xmm6 \n"\ | |
142 "mulps (%4,"#j"), "#xmm0"\n"\ | |
143 "mulps (%4,"#j"), "#xmm1"\n"\ | |
144 "mulps (%3,"#j"), %%xmm7 \n"\ | |
145 "subps %%xmm6, "#xmm0"\n"\ | |
146 "addps %%xmm7, "#xmm1"\n" | |
147 | |
148 j = -n2; | |
149 k = n2-16; | |
150 __asm__ volatile( | |
151 "1: \n" | |
152 CMUL(%0, %%xmm0, %%xmm1) | |
153 CMUL(%1, %%xmm4, %%xmm5) | |
154 "shufps $0x1b, %%xmm1, %%xmm1 \n" | |
155 "shufps $0x1b, %%xmm5, %%xmm5 \n" | |
156 "movaps %%xmm4, %%xmm6 \n" | |
157 "unpckhps %%xmm1, %%xmm4 \n" | |
158 "unpcklps %%xmm1, %%xmm6 \n" | |
159 "movaps %%xmm0, %%xmm2 \n" | |
160 "unpcklps %%xmm5, %%xmm0 \n" | |
161 "unpckhps %%xmm5, %%xmm2 \n" | |
162 "movaps %%xmm6, (%2,%1,2) \n" | |
163 "movaps %%xmm4, 16(%2,%1,2) \n" | |
164 "movaps %%xmm0, (%2,%0,2) \n" | |
165 "movaps %%xmm2, 16(%2,%0,2) \n" | |
166 "sub $16, %1 \n" | |
167 "add $16, %0 \n" | |
168 "jl 1b \n" | |
169 :"+&r"(j), "+&r"(k) | |
170 :"r"(z+n8), "r"(tcos+n8), "r"(tsin+n8) | |
171 :"memory" | |
172 ); | |
173 } | |
174 | |
175 void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input) | 74 void ff_imdct_calc_sse(FFTContext *s, FFTSample *output, const FFTSample *input) |
176 { | 75 { |
177 x86_reg j, k; | 76 x86_reg j, k; |
178 long n = 1 << s->mdct_bits; | 77 long n = 1 << s->mdct_bits; |
179 long n4 = n >> 2; | 78 long n4 = n >> 2; |