305f44a0ce3fdb3feffd33f1bb63e0efbd16c340
[libav.git] / libavcodec / i386 / fft_sse.c
1 /*
2 * FFT/MDCT transform with SSE optimizations
3 * Copyright (c) 2002 Fabrice Bellard.
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 #include "libavutil/x86_cpu.h"
23 #include "libavcodec/dsputil.h"
24
25 static const int p1p1p1m1[4] __attribute__((aligned(16))) =
26 { 0, 0, 0, 1 << 31 };
27
28 static const int p1p1m1p1[4] __attribute__((aligned(16))) =
29 { 0, 0, 1 << 31, 0 };
30
31 static const int p1p1m1m1[4] __attribute__((aligned(16))) =
32 { 0, 0, 1 << 31, 1 << 31 };
33
34 static const int p1m1p1m1[4] __attribute__((aligned(16))) =
35 { 0, 1 << 31, 0, 1 << 31 };
36
37 static const int m1m1m1m1[4] __attribute__((aligned(16))) =
38 { 1 << 31, 1 << 31, 1 << 31, 1 << 31 };
39
40 #if 0
41 static void print_v4sf(const char *str, __m128 a)
42 {
43 float *p = (float *)&a;
44 printf("%s: %f %f %f %f\n",
45 str, p[0], p[1], p[2], p[3]);
46 }
47 #endif
48
49 /* XXX: handle reverse case */
50 void ff_fft_calc_sse(FFTContext *s, FFTComplex *z)
51 {
52 int ln = s->nbits;
53 x86_reg i;
54 long j;
55 long nblocks, nloops;
56 FFTComplex *p, *cptr;
57
58 asm volatile(
59 "movaps %0, %%xmm4 \n\t"
60 "movaps %1, %%xmm5 \n\t"
61 ::"m"(*p1p1m1m1),
62 "m"(*(s->inverse ? p1p1m1p1 : p1p1p1m1))
63 );
64
65 i = 8 << ln;
66 asm volatile(
67 "1: \n\t"
68 "sub $32, %0 \n\t"
69 /* do the pass 0 butterfly */
70 "movaps (%0,%1), %%xmm0 \n\t"
71 "movaps %%xmm0, %%xmm1 \n\t"
72 "shufps $0x4E, %%xmm0, %%xmm0 \n\t"
73 "xorps %%xmm4, %%xmm1 \n\t"
74 "addps %%xmm1, %%xmm0 \n\t"
75 "movaps 16(%0,%1), %%xmm2 \n\t"
76 "movaps %%xmm2, %%xmm3 \n\t"
77 "shufps $0x4E, %%xmm2, %%xmm2 \n\t"
78 "xorps %%xmm4, %%xmm3 \n\t"
79 "addps %%xmm3, %%xmm2 \n\t"
80 /* multiply third by -i */
81 /* by toggling the sign bit */
82 "shufps $0xB4, %%xmm2, %%xmm2 \n\t"
83 "xorps %%xmm5, %%xmm2 \n\t"
84 /* do the pass 1 butterfly */
85 "movaps %%xmm0, %%xmm1 \n\t"
86 "addps %%xmm2, %%xmm0 \n\t"
87 "subps %%xmm2, %%xmm1 \n\t"
88 "movaps %%xmm0, (%0,%1) \n\t"
89 "movaps %%xmm1, 16(%0,%1) \n\t"
90 "jg 1b \n\t"
91 :"+r"(i)
92 :"r"(z)
93 );
94 /* pass 2 .. ln-1 */
95
96 nblocks = 1 << (ln-3);
97 nloops = 1 << 2;
98 cptr = s->exptab1;
99 do {
100 p = z;
101 j = nblocks;
102 do {
103 i = nloops*8;
104 asm volatile(
105 "1: \n\t"
106 "sub $32, %0 \n\t"
107 "movaps (%2,%0), %%xmm1 \n\t"
108 "movaps (%1,%0), %%xmm0 \n\t"
109 "movaps 16(%2,%0), %%xmm5 \n\t"
110 "movaps 16(%1,%0), %%xmm4 \n\t"
111 "movaps %%xmm1, %%xmm2 \n\t"
112 "movaps %%xmm5, %%xmm6 \n\t"
113 "shufps $0xA0, %%xmm1, %%xmm1 \n\t"
114 "shufps $0xF5, %%xmm2, %%xmm2 \n\t"
115 "shufps $0xA0, %%xmm5, %%xmm5 \n\t"
116 "shufps $0xF5, %%xmm6, %%xmm6 \n\t"
117 "mulps (%3,%0,2), %%xmm1 \n\t" // cre*re cim*re
118 "mulps 16(%3,%0,2), %%xmm2 \n\t" // -cim*im cre*im
119 "mulps 32(%3,%0,2), %%xmm5 \n\t" // cre*re cim*re
120 "mulps 48(%3,%0,2), %%xmm6 \n\t" // -cim*im cre*im
121 "addps %%xmm2, %%xmm1 \n\t"
122 "addps %%xmm6, %%xmm5 \n\t"
123 "movaps %%xmm0, %%xmm3 \n\t"
124 "movaps %%xmm4, %%xmm7 \n\t"
125 "addps %%xmm1, %%xmm0 \n\t"
126 "subps %%xmm1, %%xmm3 \n\t"
127 "addps %%xmm5, %%xmm4 \n\t"
128 "subps %%xmm5, %%xmm7 \n\t"
129 "movaps %%xmm0, (%1,%0) \n\t"
130 "movaps %%xmm3, (%2,%0) \n\t"
131 "movaps %%xmm4, 16(%1,%0) \n\t"
132 "movaps %%xmm7, 16(%2,%0) \n\t"
133 "jg 1b \n\t"
134 :"+r"(i)
135 :"r"(p), "r"(p + nloops), "r"(cptr)
136 );
137 p += nloops*2;
138 } while (--j);
139 cptr += nloops*2;
140 nblocks >>= 1;
141 nloops <<= 1;
142 } while (nblocks != 0);
143 }
144
145 static void imdct_sse(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
146 {
147 x86_reg k;
148 long n4, n2, n;
149 const uint16_t *revtab = s->fft.revtab;
150 const FFTSample *tcos = s->tcos;
151 const FFTSample *tsin = s->tsin;
152 const FFTSample *in1, *in2;
153 FFTComplex *z = (FFTComplex *)tmp;
154
155 n = 1 << s->nbits;
156 n2 = n >> 1;
157 n4 = n >> 2;
158
159 #ifdef ARCH_X86_64
160 asm volatile ("movaps %0, %%xmm8\n\t"::"m"(*p1m1p1m1));
161 #define P1M1P1M1 "%%xmm8"
162 #else
163 #define P1M1P1M1 "%4"
164 #endif
165
166 /* pre rotation */
167 in1 = input;
168 in2 = input + n2 - 4;
169
170 /* Complex multiplication */
171 for (k = 0; k < n4; k += 4) {
172 asm volatile (
173 "movaps %0, %%xmm0 \n\t" // xmm0 = r0 X r1 X : in2
174 "movaps %1, %%xmm3 \n\t" // xmm3 = X i1 X i0: in1
175 "movaps -16+1*%0, %%xmm4 \n\t" // xmm4 = r0 X r1 X : in2
176 "movaps 16+1*%1, %%xmm7 \n\t" // xmm7 = X i1 X i0: in1
177 "movlps %2, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
178 "movlps %3, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
179 "movlps 8+1*%2, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos
180 "movlps 8+1*%3, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin
181 "shufps $95, %%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
182 "shufps $160,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
183 "shufps $95, %%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0
184 "shufps $160,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0
185 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
186 "unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0
187 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
188 "movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0
189 "xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
190 "xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0
191 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
192 "mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR
193 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
194 "shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0
195 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
196 "mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii
197 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
198 "addps %%xmm7, %%xmm4 \n\t" // xmm4 = result
199 ::"m"(in2[-2*k]), "m"(in1[2*k]),
200 "m"(tcos[k]), "m"(tsin[k])
201 #ifndef ARCH_X86_64
202 ,"m"(*p1m1p1m1)
203 #endif
204 );
205 /* Should be in the same block, hack for gcc2.95 & gcc3 */
206 asm (
207 "movlps %%xmm0, %0 \n\t"
208 "movhps %%xmm0, %1 \n\t"
209 "movlps %%xmm4, %2 \n\t"
210 "movhps %%xmm4, %3 \n\t"
211 :"=m"(z[revtab[k]]), "=m"(z[revtab[k + 1]]),
212 "=m"(z[revtab[k + 2]]), "=m"(z[revtab[k + 3]])
213 );
214 }
215
216 ff_fft_calc_sse(&s->fft, z);
217
218 #ifndef ARCH_X86_64
219 #undef P1M1P1M1
220 #define P1M1P1M1 "%3"
221 #endif
222
223 /* post rotation + reordering */
224 for (k = 0; k < n4; k += 4) {
225 asm (
226 "movaps %0, %%xmm0 \n\t" // xmm0 = i1 r1 i0 r0: z
227 "movaps 16+1*%0, %%xmm4 \n\t" // xmm4 = i1 r1 i0 r0: z
228 "movlps %1, %%xmm1 \n\t" // xmm1 = X X R1 R0: tcos
229 "movlps 8+1*%1, %%xmm5 \n\t" // xmm5 = X X R1 R0: tcos
230 "movaps %%xmm0, %%xmm3 \n\t" // xmm3 = i1 r1 i0 r0
231 "movaps %%xmm4, %%xmm7 \n\t" // xmm7 = i1 r1 i0 r0
232 "movlps %2, %%xmm2 \n\t" // xmm2 = X X I1 I0: tsin
233 "movlps 8+1*%2, %%xmm6 \n\t" // xmm6 = X X I1 I0: tsin
234 "shufps $160,%%xmm0, %%xmm0 \n\t" // xmm0 = r1 r1 r0 r0
235 "shufps $245,%%xmm3, %%xmm3 \n\t" // xmm3 = i1 i1 i0 i0
236 "shufps $160,%%xmm4, %%xmm4 \n\t" // xmm4 = r1 r1 r0 r0
237 "shufps $245,%%xmm7, %%xmm7 \n\t" // xmm7 = i1 i1 i0 i0
238 "unpcklps %%xmm2, %%xmm1 \n\t" // xmm1 = I1 R1 I0 R0
239 "unpcklps %%xmm6, %%xmm5 \n\t" // xmm5 = I1 R1 I0 R0
240 "movaps %%xmm1, %%xmm2 \n\t" // xmm2 = I1 R1 I0 R0
241 "movaps %%xmm5, %%xmm6 \n\t" // xmm6 = I1 R1 I0 R0
242 "xorps "P1M1P1M1", %%xmm2 \n\t" // xmm2 = -I1 R1 -I0 R0
243 "mulps %%xmm1, %%xmm0 \n\t" // xmm0 = rI rR rI rR
244 "xorps "P1M1P1M1", %%xmm6 \n\t" // xmm6 = -I1 R1 -I0 R0
245 "mulps %%xmm5, %%xmm4 \n\t" // xmm4 = rI rR rI rR
246 "shufps $177,%%xmm2, %%xmm2 \n\t" // xmm2 = R1 -I1 R0 -I0
247 "shufps $177,%%xmm6, %%xmm6 \n\t" // xmm6 = R1 -I1 R0 -I0
248 "mulps %%xmm2, %%xmm3 \n\t" // xmm3 = Ri -Ii Ri -Ii
249 "mulps %%xmm6, %%xmm7 \n\t" // xmm7 = Ri -Ii Ri -Ii
250 "addps %%xmm3, %%xmm0 \n\t" // xmm0 = result
251 "addps %%xmm7, %%xmm4 \n\t" // xmm4 = result
252 "movaps %%xmm0, %0 \n\t"
253 "movaps %%xmm4, 16+1*%0\n\t"
254 :"+m"(z[k])
255 :"m"(tcos[k]), "m"(tsin[k])
256 #ifndef ARCH_X86_64
257 ,"m"(*p1m1p1m1)
258 #endif
259 );
260 }
261 }
262
263 void ff_imdct_calc_sse(MDCTContext *s, FFTSample *output,
264 const FFTSample *input, FFTSample *tmp)
265 {
266 x86_reg k;
267 long n8, n2, n;
268 FFTComplex *z = (FFTComplex *)tmp;
269
270 n = 1 << s->nbits;
271 n2 = n >> 1;
272 n8 = n >> 3;
273
274 imdct_sse(s, input, tmp);
275
276 /*
277 Mnemonics:
278 0 = z[k].re
279 1 = z[k].im
280 2 = z[k + 1].re
281 3 = z[k + 1].im
282 4 = z[-k - 2].re
283 5 = z[-k - 2].im
284 6 = z[-k - 1].re
285 7 = z[-k - 1].im
286 */
287 k = 16-n;
288 asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1));
289 asm volatile(
290 "1: \n\t"
291 "movaps -16(%4,%0), %%xmm1 \n\t" // xmm1 = 4 5 6 7 = z[-2-k]
292 "neg %0 \n\t"
293 "movaps (%4,%0), %%xmm0 \n\t" // xmm0 = 0 1 2 3 = z[k]
294 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -0 -1 -2 -3
295 "movaps %%xmm0, %%xmm2 \n\t" // xmm2 = -0 -1 -2 -3
296 "shufps $141,%%xmm1, %%xmm0 \n\t" // xmm0 = -1 -3 4 6
297 "shufps $216,%%xmm1, %%xmm2 \n\t" // xmm2 = -0 -2 5 7
298 "shufps $156,%%xmm0, %%xmm0 \n\t" // xmm0 = -1 6 -3 4 !
299 "shufps $156,%%xmm2, %%xmm2 \n\t" // xmm2 = -0 7 -2 5 !
300 "movaps %%xmm0, (%1,%0) \n\t" // output[2*k]
301 "movaps %%xmm2, (%2,%0) \n\t" // output[n2+2*k]
302 "neg %0 \n\t"
303 "shufps $27, %%xmm0, %%xmm0 \n\t" // xmm0 = 4 -3 6 -1
304 "xorps %%xmm7, %%xmm0 \n\t" // xmm0 = -4 3 -6 1 !
305 "shufps $27, %%xmm2, %%xmm2 \n\t" // xmm2 = 5 -2 7 -0 !
306 "movaps %%xmm0, -16(%2,%0) \n\t" // output[n2-4-2*k]
307 "movaps %%xmm2, -16(%3,%0) \n\t" // output[n-4-2*k]
308 "add $16, %0 \n\t"
309 "jle 1b \n\t"
310 :"+r"(k)
311 :"r"(output), "r"(output+n2), "r"(output+n), "r"(z+n8)
312 :"memory"
313 );
314 }
315
316 void ff_imdct_half_sse(MDCTContext *s, FFTSample *output,
317 const FFTSample *input, FFTSample *tmp)
318 {
319 x86_reg j, k;
320 long n8, n4, n;
321 FFTComplex *z = (FFTComplex *)tmp;
322
323 n = 1 << s->nbits;
324 n4 = n >> 2;
325 n8 = n >> 3;
326
327 imdct_sse(s, input, tmp);
328
329 j = -n;
330 k = n-16;
331 asm volatile("movaps %0, %%xmm7 \n\t"::"m"(*m1m1m1m1));
332 asm volatile(
333 "1: \n\t"
334 "movaps (%3,%1), %%xmm0 \n\t"
335 "movaps (%3,%0), %%xmm1 \n\t"
336 "xorps %%xmm7, %%xmm0 \n\t"
337 "movaps %%xmm0, %%xmm2 \n\t"
338 "shufps $141,%%xmm1, %%xmm0 \n\t"
339 "shufps $216,%%xmm1, %%xmm2 \n\t"
340 "shufps $54, %%xmm0, %%xmm0 \n\t"
341 "shufps $156,%%xmm2, %%xmm2 \n\t"
342 "xorps %%xmm7, %%xmm0 \n\t"
343 "movaps %%xmm2, (%2,%1) \n\t"
344 "movaps %%xmm0, (%2,%0) \n\t"
345 "sub $16, %1 \n\t"
346 "add $16, %0 \n\t"
347 "jl 1b \n\t"
348 :"+r"(j), "+r"(k)
349 :"r"(output+n4), "r"(z+n8)
350 :"memory"
351 );
352 }
353