1 /*****************************************************************************
2 * ac3_imdct_sse.c: accelerated SSE ac3 DCT
3 *****************************************************************************
4 * Copyright (C) 1999, 2000 VideoLAN
5 * $Id: ac3_imdct_sse.c,v 1.1 2001/05/15 16:19:42 sam Exp $
7 * Authors: Renaud Dartus <reno@videolan.org>
8 * Aaron Holtzman <aholtzma@engr.uvic.ca>
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License as published by
12 * the Free Software Foundation; either version 2 of the License, or
13 * (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 * GNU General Public License for more details.
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
23 *****************************************************************************/
25 #define MODULE_NAME imdctsse
26 #include "modules_inner.h"
28 /*****************************************************************************
30 *****************************************************************************/
41 #include "ac3_imdct.h"
42 #include "ac3_imdct_common.h"
44 static const float window[] = {
45 0.00014, 0.00024, 0.00037, 0.00051, 0.00067, 0.00086, 0.00107, 0.00130,
46 0.00157, 0.00187, 0.00220, 0.00256, 0.00297, 0.00341, 0.00390, 0.00443,
47 0.00501, 0.00564, 0.00632, 0.00706, 0.00785, 0.00871, 0.00962, 0.01061,
48 0.01166, 0.01279, 0.01399, 0.01526, 0.01662, 0.01806, 0.01959, 0.02121,
49 0.02292, 0.02472, 0.02662, 0.02863, 0.03073, 0.03294, 0.03527, 0.03770,
50 0.04025, 0.04292, 0.04571, 0.04862, 0.05165, 0.05481, 0.05810, 0.06153,
51 0.06508, 0.06878, 0.07261, 0.07658, 0.08069, 0.08495, 0.08935, 0.09389,
52 0.09859, 0.10343, 0.10842, 0.11356, 0.11885, 0.12429, 0.12988, 0.13563,
53 0.14152, 0.14757, 0.15376, 0.16011, 0.16661, 0.17325, 0.18005, 0.18699,
54 0.19407, 0.20130, 0.20867, 0.21618, 0.22382, 0.23161, 0.23952, 0.24757,
55 0.25574, 0.26404, 0.27246, 0.28100, 0.28965, 0.29841, 0.30729, 0.31626,
56 0.32533, 0.33450, 0.34376, 0.35311, 0.36253, 0.37204, 0.38161, 0.39126,
57 0.40096, 0.41072, 0.42054, 0.43040, 0.44030, 0.45023, 0.46020, 0.47019,
58 0.48020, 0.49022, 0.50025, 0.51028, 0.52031, 0.53033, 0.54033, 0.55031,
59 0.56026, 0.57019, 0.58007, 0.58991, 0.59970, 0.60944, 0.61912, 0.62873,
60 0.63827, 0.64774, 0.65713, 0.66643, 0.67564, 0.68476, 0.69377, 0.70269,
61 0.71150, 0.72019, 0.72877, 0.73723, 0.74557, 0.75378, 0.76186, 0.76981,
62 0.77762, 0.78530, 0.79283, 0.80022, 0.80747, 0.81457, 0.82151, 0.82831,
63 0.83496, 0.84145, 0.84779, 0.85398, 0.86001, 0.86588, 0.87160, 0.87716,
64 0.88257, 0.88782, 0.89291, 0.89785, 0.90264, 0.90728, 0.91176, 0.91610,
65 0.92028, 0.92432, 0.92822, 0.93197, 0.93558, 0.93906, 0.94240, 0.94560,
66 0.94867, 0.95162, 0.95444, 0.95713, 0.95971, 0.96217, 0.96451, 0.96674,
67 0.96887, 0.97089, 0.97281, 0.97463, 0.97635, 0.97799, 0.97953, 0.98099,
68 0.98236, 0.98366, 0.98488, 0.98602, 0.98710, 0.98811, 0.98905, 0.98994,
69 0.99076, 0.99153, 0.99225, 0.99291, 0.99353, 0.99411, 0.99464, 0.99513,
70 0.99558, 0.99600, 0.99639, 0.99674, 0.99706, 0.99736, 0.99763, 0.99788,
71 0.99811, 0.99831, 0.99850, 0.99867, 0.99882, 0.99895, 0.99908, 0.99919,
72 0.99929, 0.99938, 0.99946, 0.99953, 0.99959, 0.99965, 0.99969, 0.99974,
73 0.99978, 0.99981, 0.99984, 0.99986, 0.99988, 0.99990, 0.99992, 0.99993,
74 0.99994, 0.99995, 0.99996, 0.99997, 0.99998, 0.99998, 0.99998, 0.99999,
75 0.99999, 0.99999, 0.99999, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000,
76 1.00000, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000, 1.00000
79 static const int pm128[128] =
81 0, 16, 32, 48, 64, 80, 96, 112, 8, 40, 72, 104, 24, 56, 88, 120,
82 4, 20, 36, 52, 68, 84, 100, 116, 12, 28, 44, 60, 76, 92, 108, 124,
83 2, 18, 34, 50, 66, 82, 98, 114, 10, 42, 74, 106, 26, 58, 90, 122,
84 6, 22, 38, 54, 70, 86, 102, 118, 14, 46, 78, 110, 30, 62, 94, 126,
85 1, 17, 33, 49, 65, 81, 97, 113, 9, 41, 73, 105, 25, 57, 89, 121,
86 5, 21, 37, 53, 69, 85, 101, 117, 13, 29, 45, 61, 77, 93, 109, 125,
87 3, 19, 35, 51, 67, 83, 99, 115, 11, 43, 75, 107, 27, 59, 91, 123,
88 7, 23, 39, 55, 71, 87, 103, 119, 15, 31, 47, 63, 79, 95, 111, 127
91 void _M( fft_64p ) ( complex_t *x );
92 void _M( fft_128p ) ( complex_t *a );
94 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse);
95 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse);
96 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, const float *window_prt, float *delay_prt);
97 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, const float *window_prt, float *delay_prt);
100 void _M( imdct_init ) (imdct_t * p_imdct)
103 float scale = 181.019;
105 for (i=0; i < 128; i++)
107 float xcos_i = cos(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
108 float xsin_i = sin(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
109 p_imdct->xcos_sin_sse[i * 4] = xcos_i;
110 p_imdct->xcos_sin_sse[i * 4 + 1] = -xsin_i;
111 p_imdct->xcos_sin_sse[i * 4 + 2] = -xsin_i;
112 p_imdct->xcos_sin_sse[i * 4 + 3] = -xcos_i;
116 void _M( imdct_do_512 ) (imdct_t * p_imdct, float data[], float delay[])
118 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
119 _M( fft_128p ) ( p_imdct->buf );
120 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
121 imdct512_window_delay_sse (p_imdct->buf, data, window, delay);
125 void _M( imdct_do_512_nol ) (imdct_t * p_imdct, float data[], float delay[])
127 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
128 _M( fft_128p ) ( p_imdct->buf );
129 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
130 imdct512_window_delay_nol_sse (p_imdct->buf, data, window, delay);
133 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse)
135 __asm__ __volatile__ (
137 "movl %%esp, %%ebp\n"
138 "addl $-4, %%esp\n" /* local variable, loop counter */
147 "movl 8(%%ebp), %%eax\n" /* pmt */
148 "movl 12(%%ebp), %%ebx\n" /* buf */
149 "movl 16(%%ebp), %%ecx\n" /* data */
150 "movl 20(%%ebp), %%edx\n" /* xcos_sin_sse */
151 "movl $64, -4(%%ebp)\n"
154 "movl (%%eax), %%esi\n"
155 "movl 4(%%eax), %%edi\n"
156 "movss (%%ecx, %%esi, 8), %%xmm1\n" /* 2j */
157 "movss (%%ecx, %%edi, 8), %%xmm3\n" /* 2(j+1) */
162 "movups (%%edx, %%esi, 8), %%xmm0\n" /* -c_j | -s_j | -s_j | c_j */
163 "movups (%%edx, %%edi, 8), %%xmm2\n" /* -c_j+1 | -s_j+1 | -s_j+1 | c_j+1 */
168 "movss 1020(%%ecx, %%esi, 4), %%xmm4\n" /* 255-2j */
170 "movss 1020(%%ecx, %%edi, 4), %%xmm5\n" /* 255-2(j+1) */
172 "shufps $0, %%xmm1, %%xmm4\n" /* 2j | 2j | 255-2j | 255-2j */
173 "shufps $0, %%xmm3, %%xmm5\n" /* 2(j+1) | 2(j+1) | 255-2(j+1) | 255-2(j+1) */
174 "mulps %%xmm4, %%xmm0\n"
175 "mulps %%xmm5, %%xmm2\n"
176 "movhlps %%xmm0, %%xmm1\n"
177 "movhlps %%xmm2, %%xmm3\n"
179 "addps %%xmm1, %%xmm0\n"
180 "addps %%xmm3, %%xmm2\n"
181 "movlhps %%xmm2, %%xmm0\n"
183 "movups %%xmm0, -16(%%ebx)\n"
199 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse)
201 __asm__ __volatile__ (
203 "movl $32, %%ebx\n" /* loop counter */
206 "movups (%%eax), %%xmm0\n" /* im1 | re1 | im0 | re0 */
208 "movups (%%ecx), %%xmm2\n" /* -c | -s | -s | c */
209 "movhlps %%xmm0, %%xmm1\n" /* im1 | re1 */
210 "movups 16(%%ecx), %%xmm3\n" /* -c1 | -s1 | -s1 | c1 */
212 "shufps $0x50, %%xmm0, %%xmm0\n" /* im0 | im0 | re0 | re0 */
213 "shufps $0x50, %%xmm1, %%xmm1\n" /* im1 | im1 | re1 | re1 */
215 "movups 16(%%eax), %%xmm4\n" /* im3 | re3 | im2 | re2 */
217 "shufps $0x27, %%xmm2, %%xmm2\n" /* c | -s | -s | -c */
218 "movhlps %%xmm4, %%xmm5\n" /* im3 | re3 */
219 "shufps $0x27, %%xmm3, %%xmm3\n" /* c1 | -s1 | -s1 | -c1 */
221 "movups 32(%%ecx), %%xmm6\n" /* -c2 | -s2 | -s2 | c2 */
222 "movups 48(%%ecx), %%xmm7\n" /* -c3 | -s3 | -s3 | c3 */
224 "shufps $0x50, %%xmm4, %%xmm4\n" /* im2 | im2 | re2 | re2 */
225 "shufps $0x50, %%xmm5, %%xmm5\n" /* im3 | im3 | re3 | re3 */
227 "mulps %%xmm2, %%xmm0\n"
228 "mulps %%xmm3, %%xmm1\n"
230 "shufps $0x27, %%xmm6, %%xmm6\n" /* c2 | -s2 | -s2 | -c2 */
231 "shufps $0x27, %%xmm7, %%xmm7\n" /* c3 | -s3 | -s3 | -c3 */
233 "movhlps %%xmm0, %%xmm2\n"
234 "movhlps %%xmm1, %%xmm3\n"
236 "mulps %%xmm6, %%xmm4\n"
237 "mulps %%xmm7, %%xmm5\n"
239 "addps %%xmm2, %%xmm0\n"
240 "addps %%xmm3, %%xmm1\n"
242 "movhlps %%xmm4, %%xmm6\n"
243 "movhlps %%xmm5, %%xmm7\n"
245 "addps %%xmm6, %%xmm4\n"
246 "addps %%xmm7, %%xmm5\n"
248 "movlhps %%xmm1, %%xmm0\n"
249 "movlhps %%xmm5, %%xmm4\n"
251 "movups %%xmm0, (%%eax)\n"
252 "movups %%xmm4, 16(%%eax)\n"
260 : "a" (buf), "c" (xcos_sin_sse) );
263 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, const float *window_prt, float *delay_prt)
265 __asm__ __volatile__ (
267 "movl %%esp, %%ebp\n"
276 "movl 20(%%ebp), %%ebx\n" /* delay */
277 "movl 16(%%ebp), %%edx\n" /* window */
279 "movl 8(%%ebp), %%eax\n" /* buf */
280 "movl $16, %%ecx\n" /* loop count */
281 "leal 516(%%eax), %%esi\n" /* buf[64].im */
282 "leal 504(%%eax), %%edi\n" /* buf[63].re */
283 "movl 12(%%ebp), %%eax\n" /* data */
285 ".first_128_samples:\n"
286 "movss (%%esi), %%xmm0\n"
287 "movss 8(%%esi), %%xmm2\n"
288 "movss (%%edi), %%xmm1\n"
289 "movss -8(%%edi), %%xmm3\n"
291 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
292 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
294 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
295 "movups (%%ebx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
296 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
298 "movss 16(%%esi), %%xmm6\n" /* im2 */
299 "movss 24(%%esi), %%xmm7\n" /* im3 */
300 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
301 "movss -16(%%edi), %%xmm2\n" /* re2 */
302 "movss -24(%%edi), %%xmm3\n" /* re3 */
303 "mulps %%xmm4, %%xmm0\n"
304 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
305 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
306 "addps %%xmm5, %%xmm0\n"
307 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
308 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
309 "movups 16(%%ebx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
310 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
312 "movups %%xmm0, (%%eax)\n"
314 "mulps %%xmm4, %%xmm6\n"
317 "addps %%xmm5, %%xmm6\n"
319 "movups %%xmm6, -16(%%eax)\n"
321 "jnz .first_128_samples\n"
323 "movl 8(%%ebp), %%esi\n" /* buf[0].re */
324 "leal 1020(%%esi), %%edi\n" /* buf[127].im */
325 "movl $16, %%ecx\n" /* loop count */
327 ".second_128_samples:\n"
328 "movss (%%esi), %%xmm0\n" /* buf[i].re */
329 "movss 8(%%esi), %%xmm2\n" /* re1 */
330 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
331 "movss -8(%%edi), %%xmm3\n" /* im1 */
333 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
334 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
336 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
337 "movups (%%ebx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
339 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
340 "movss 16(%%esi), %%xmm6\n" /* re2 */
341 "movss 24(%%esi), %%xmm7\n" /* re3 */
342 "movss -16(%%edi), %%xmm2\n" /* im2 */
343 "movss -24(%%edi), %%xmm3\n" /* im3 */
344 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
345 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
346 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
347 "mulps %%xmm4, %%xmm0\n"
348 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
349 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
351 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
352 "addps %%xmm5, %%xmm0\n"
353 "mulps %%xmm4, %%xmm6\n"
355 "movups 16(%%ebx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
356 "movups %%xmm0, (%%eax)\n"
357 "addps %%xmm5, %%xmm6\n"
361 "movups %%xmm6, -16(%%eax)\n"
363 "jnz .second_128_samples\n"
365 "movl 8(%%ebp), %%eax\n"
366 "leal 512(%%eax), %%esi\n" /* buf[64].re */
367 "leal 508(%%eax), %%edi\n" /* buf[63].im */
368 "movl $16, %%ecx\n" /* loop count */
369 "movl 20(%%ebp), %%eax\n" /* delay */
371 ".first_128_delay:\n"
372 "movss (%%esi), %%xmm0\n"
373 "movss 8(%%esi), %%xmm2\n"
374 "movss (%%edi), %%xmm1\n"
375 "movss -8(%%edi), %%xmm3\n"
377 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
378 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
380 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
381 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
382 "movss 16(%%esi), %%xmm6\n" /* re2 */
383 "movss 24(%%esi), %%xmm7\n" /* re3 */
384 "movss -16(%%edi), %%xmm2\n" /* im2 */
385 "movss -24(%%edi), %%xmm3\n" /* im3 */
386 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
388 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
389 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
390 "mulps %%xmm4, %%xmm0\n"
391 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
392 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
393 "movups %%xmm0, (%%eax)\n"
395 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
397 "mulps %%xmm5, %%xmm6\n"
399 "movups %%xmm6, -16(%%eax)\n"
401 "jnz .first_128_delay\n"
403 "movl 8(%%ebp), %%ebx\n"
404 "leal 4(%%ebx), %%esi\n" /* buf[0].im */
405 "leal 1016(%%ebx), %%edi\n" /* buf[127].re */
406 "movl $16, %%ecx\n" /* loop count */
408 ".second_128_delay:\n"
409 "movss (%%esi), %%xmm0\n"
410 "movss 8(%%esi), %%xmm2\n"
411 "movss (%%edi), %%xmm1\n"
412 "movss -8(%%edi), %%xmm3\n"
414 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
415 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
417 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
418 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
419 "movss 16(%%esi), %%xmm6\n" /* im2 */
420 "movss 24(%%esi), %%xmm7\n" /* im3 */
421 "movss -16(%%edi), %%xmm2\n" /* re2 */
422 "movss -24(%%edi), %%xmm3\n" /* re3 */
423 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
425 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
426 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
427 "mulps %%xmm4, %%xmm1\n"
428 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
429 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
430 "movups %%xmm1, (%%eax)\n"
432 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
434 "mulps %%xmm5, %%xmm2\n"
436 "movups %%xmm2, -16(%%eax)\n"
438 "jnz .second_128_delay\n"
451 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, const float *window_prt, float *delay_prt)
453 __asm__ __volatile__ (
455 "movl %%esp, %%ebp\n"
464 /* movl 20(%%ebp), %%ebx delay */
465 "movl 16(%%ebp), %%edx\n" /* window */
467 "movl 8(%%ebp), %%eax\n" /* buf */
468 "movl $16, %%ecx\n" /* loop count */
469 "leal 516(%%eax), %%esi\n" /* buf[64].im */
470 "leal 504(%%eax), %%edi\n" /* buf[63].re */
471 "movl 12(%%ebp), %%eax\n" /* data */
473 ".first_128_sample:\n"
474 "movss (%%esi), %%xmm0\n"
475 "movss 8(%%esi), %%xmm2\n"
476 "movss (%%edi), %%xmm1\n"
477 "movss -8(%%edi), %%xmm3\n"
479 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
480 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
482 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
483 /* movups (%%ebx), %%xmm5 d3 | d2 | d1 | d0 */
484 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
486 "movss 16(%%esi), %%xmm6\n" /* im2 */
487 "movss 24(%%esi), %%xmm7\n" /* im3 */
488 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
489 "movss -16(%%edi), %%xmm2\n" /* re2 */
490 "movss -24(%%edi), %%xmm3\n" /* re3 */
491 "mulps %%xmm4, %%xmm0\n"
492 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
493 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
494 /* addps %%xmm5, %%xmm0 */
495 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
496 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
497 /* movups 16(%%ebx), %%xmm5 d7 | d6 | d5 | d4 */
498 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
500 "movups %%xmm0, (%%eax)\n"
501 /* addl $32, %%ebx */
502 "mulps %%xmm4, %%xmm6\n"
505 /* addps %%xmm5, %%xmm6 */
507 "movups %%xmm6, -16(%%eax)\n"
509 "jnz .first_128_sample\n"
511 "movl 8(%%ebp), %%esi\n" /* buf[0].re */
512 "leal 1020(%%esi), %%edi\n" /* buf[127].im */
513 "movl $16, %%ecx\n" /* loop count */
515 ".second_128_sample:\n"
516 "movss (%%esi), %%xmm0\n" /* buf[i].re */
517 "movss 8(%%esi), %%xmm2\n" /* re1 */
518 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
519 "movss -8(%%edi), %%xmm3\n" /* im1 */
521 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
522 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
524 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
525 /* movups (%%ebx), %%xmm5 d3 | d2 | d1 | d0 */
527 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
528 "movss 16(%%esi), %%xmm6\n" /* re2 */
529 "movss 24(%%esi), %%xmm7\n" /* re3 */
530 "movss -16(%%edi), %%xmm2\n" /* im2 */
531 "movss -24(%%edi), %%xmm3\n" /* im3 */
532 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
533 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
534 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
535 "mulps %%xmm4, %%xmm0\n"
536 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
537 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
539 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
540 /* addps %%xmm5, %%xmm0 */
541 "mulps %%xmm4, %%xmm6\n"
543 /* movups 16(%%ebx), %%xmm5 d7 | d6 | d5 | d4 */
544 "movups %%xmm0, (%%eax)\n"
545 /* addps %%xmm5, %%xmm6 */
548 /* addl $32, %%ebx */
549 "movups %%xmm6, -16(%%eax)\n"
551 "jnz .second_128_sample\n"
553 "movl 8(%%ebp), %%eax\n"
554 "leal 512(%%eax), %%esi\n" /* buf[64].re */
555 "leal 508(%%eax), %%edi\n" /* buf[63].im */
556 "movl $16, %%ecx\n" /* loop count */
557 "movl 20(%%ebp), %%eax\n" /* delay */
559 ".first_128_delays:\n"
560 "movss (%%esi), %%xmm0\n"
561 "movss 8(%%esi), %%xmm2\n"
562 "movss (%%edi), %%xmm1\n"
563 "movss -8(%%edi), %%xmm3\n"
565 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
566 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
568 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
569 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
570 "movss 16(%%esi), %%xmm6\n" /* re2 */
571 "movss 24(%%esi), %%xmm7\n" /* re3 */
572 "movss -16(%%edi), %%xmm2\n" /* im2 */
573 "movss -24(%%edi), %%xmm3\n" /* im3 */
574 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
576 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
577 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
578 "mulps %%xmm4, %%xmm0\n"
579 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
580 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
581 "movups %%xmm0, (%%eax)\n"
583 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
585 "mulps %%xmm5, %%xmm6\n"
587 "movups %%xmm6, -16(%%eax)\n"
589 "jnz .first_128_delays\n"
591 "movl 8(%%ebp), %%ebx\n"
592 "leal 4(%%ebx), %%esi\n" /* buf[0].im */
593 "leal 1016(%%ebx), %%edi\n" /* buf[127].re */
594 "movl $16, %%ecx\n" /* loop count */
596 ".second_128_delays:\n"
597 "movss (%%esi), %%xmm0\n"
598 "movss 8(%%esi), %%xmm2\n"
599 "movss (%%edi), %%xmm1\n"
600 "movss -8(%%edi), %%xmm3\n"
602 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
603 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
605 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
606 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
607 "movss 16(%%esi), %%xmm6\n" /* im2 */
608 "movss 24(%%esi), %%xmm7\n" /* im3 */
609 "movss -16(%%edi), %%xmm2\n" /* re2 */
610 "movss -24(%%edi), %%xmm3\n" /* re3 */
611 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
613 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
614 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
615 "mulps %%xmm4, %%xmm1\n"
616 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
617 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
618 "movups %%xmm1, (%%eax)\n"
620 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
622 "mulps %%xmm5, %%xmm2\n"
624 "movups %%xmm2, -16(%%eax)\n"
626 "jnz .second_128_delays\n"