1 /*****************************************************************************
2 * ac3_imdct_sse.c: accelerated SSE ac3 DCT
3 *****************************************************************************
4 * Copyright (C) 1999, 2000 VideoLAN
5 * $Id: ac3_imdct_sse.c,v 1.2 2001/05/16 14:51:29 reno 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"
43 #include "ac3_retables.h"
45 void _M( fft_64p ) ( complex_t *x );
46 void _M( fft_128p ) ( complex_t *a );
48 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse);
49 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse);
50 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt);
51 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt);
54 void _M( imdct_init ) (imdct_t * p_imdct)
57 float scale = 181.019;
59 for (i=0; i < 128; i++)
61 float xcos_i = cos(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
62 float xsin_i = sin(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
63 p_imdct->xcos_sin_sse[i * 4] = xcos_i;
64 p_imdct->xcos_sin_sse[i * 4 + 1] = -xsin_i;
65 p_imdct->xcos_sin_sse[i * 4 + 2] = -xsin_i;
66 p_imdct->xcos_sin_sse[i * 4 + 3] = -xcos_i;
70 void _M( imdct_do_512 ) (imdct_t * p_imdct, float data[], float delay[])
72 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
73 _M( fft_128p ) ( p_imdct->buf );
74 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
75 imdct512_window_delay_sse (p_imdct->buf, data, window, delay);
79 void _M( imdct_do_512_nol ) (imdct_t * p_imdct, float data[], float delay[])
81 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
82 _M( fft_128p ) ( p_imdct->buf );
83 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
84 imdct512_window_delay_nol_sse (p_imdct->buf, data, window, delay);
87 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse)
89 __asm__ __volatile__ (
92 "addl $-4, %%esp\n" /* local variable, loop counter */
101 "movl 8(%%ebp), %%eax\n" /* pmt */
102 "movl 12(%%ebp), %%ebx\n" /* buf */
103 "movl 16(%%ebp), %%ecx\n" /* data */
104 "movl 20(%%ebp), %%edx\n" /* xcos_sin_sse */
105 "movl $64, -4(%%ebp)\n"
108 "movl (%%eax), %%esi\n"
109 "movl 4(%%eax), %%edi\n"
110 "movss (%%ecx, %%esi, 8), %%xmm1\n" /* 2j */
111 "movss (%%ecx, %%edi, 8), %%xmm3\n" /* 2(j+1) */
116 "movups (%%edx, %%esi, 8), %%xmm0\n" /* -c_j | -s_j | -s_j | c_j */
117 "movups (%%edx, %%edi, 8), %%xmm2\n" /* -c_j+1 | -s_j+1 | -s_j+1 | c_j+1 */
122 "movss 1020(%%ecx, %%esi, 4), %%xmm4\n" /* 255-2j */
124 "movss 1020(%%ecx, %%edi, 4), %%xmm5\n" /* 255-2(j+1) */
126 "shufps $0, %%xmm1, %%xmm4\n" /* 2j | 2j | 255-2j | 255-2j */
127 "shufps $0, %%xmm3, %%xmm5\n" /* 2(j+1) | 2(j+1) | 255-2(j+1) | 255-2(j+1) */
128 "mulps %%xmm4, %%xmm0\n"
129 "mulps %%xmm5, %%xmm2\n"
130 "movhlps %%xmm0, %%xmm1\n"
131 "movhlps %%xmm2, %%xmm3\n"
133 "addps %%xmm1, %%xmm0\n"
134 "addps %%xmm3, %%xmm2\n"
135 "movlhps %%xmm2, %%xmm0\n"
137 "movups %%xmm0, -16(%%ebx)\n"
153 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse)
155 __asm__ __volatile__ (
157 "movl $32, %%ebx\n" /* loop counter */
160 "movups (%%eax), %%xmm0\n" /* im1 | re1 | im0 | re0 */
162 "movups (%%ecx), %%xmm2\n" /* -c | -s | -s | c */
163 "movhlps %%xmm0, %%xmm1\n" /* im1 | re1 */
164 "movups 16(%%ecx), %%xmm3\n" /* -c1 | -s1 | -s1 | c1 */
166 "shufps $0x50, %%xmm0, %%xmm0\n" /* im0 | im0 | re0 | re0 */
167 "shufps $0x50, %%xmm1, %%xmm1\n" /* im1 | im1 | re1 | re1 */
169 "movups 16(%%eax), %%xmm4\n" /* im3 | re3 | im2 | re2 */
171 "shufps $0x27, %%xmm2, %%xmm2\n" /* c | -s | -s | -c */
172 "movhlps %%xmm4, %%xmm5\n" /* im3 | re3 */
173 "shufps $0x27, %%xmm3, %%xmm3\n" /* c1 | -s1 | -s1 | -c1 */
175 "movups 32(%%ecx), %%xmm6\n" /* -c2 | -s2 | -s2 | c2 */
176 "movups 48(%%ecx), %%xmm7\n" /* -c3 | -s3 | -s3 | c3 */
178 "shufps $0x50, %%xmm4, %%xmm4\n" /* im2 | im2 | re2 | re2 */
179 "shufps $0x50, %%xmm5, %%xmm5\n" /* im3 | im3 | re3 | re3 */
181 "mulps %%xmm2, %%xmm0\n"
182 "mulps %%xmm3, %%xmm1\n"
184 "shufps $0x27, %%xmm6, %%xmm6\n" /* c2 | -s2 | -s2 | -c2 */
185 "shufps $0x27, %%xmm7, %%xmm7\n" /* c3 | -s3 | -s3 | -c3 */
187 "movhlps %%xmm0, %%xmm2\n"
188 "movhlps %%xmm1, %%xmm3\n"
190 "mulps %%xmm6, %%xmm4\n"
191 "mulps %%xmm7, %%xmm5\n"
193 "addps %%xmm2, %%xmm0\n"
194 "addps %%xmm3, %%xmm1\n"
196 "movhlps %%xmm4, %%xmm6\n"
197 "movhlps %%xmm5, %%xmm7\n"
199 "addps %%xmm6, %%xmm4\n"
200 "addps %%xmm7, %%xmm5\n"
202 "movlhps %%xmm1, %%xmm0\n"
203 "movlhps %%xmm5, %%xmm4\n"
205 "movups %%xmm0, (%%eax)\n"
206 "movups %%xmm4, 16(%%eax)\n"
214 : "a" (buf), "c" (xcos_sin_sse) );
217 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt)
219 __asm__ __volatile__ (
221 "movl %%esp, %%ebp\n"
230 "movl 20(%%ebp), %%ebx\n" /* delay */
231 "movl 16(%%ebp), %%edx\n" /* window */
233 "movl 8(%%ebp), %%eax\n" /* buf */
234 "movl $16, %%ecx\n" /* loop count */
235 "leal 516(%%eax), %%esi\n" /* buf[64].im */
236 "leal 504(%%eax), %%edi\n" /* buf[63].re */
237 "movl 12(%%ebp), %%eax\n" /* data */
239 ".first_128_samples:\n"
240 "movss (%%esi), %%xmm0\n"
241 "movss 8(%%esi), %%xmm2\n"
242 "movss (%%edi), %%xmm1\n"
243 "movss -8(%%edi), %%xmm3\n"
245 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
246 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
248 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
249 "movups (%%ebx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
250 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
252 "movss 16(%%esi), %%xmm6\n" /* im2 */
253 "movss 24(%%esi), %%xmm7\n" /* im3 */
254 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
255 "movss -16(%%edi), %%xmm2\n" /* re2 */
256 "movss -24(%%edi), %%xmm3\n" /* re3 */
257 "mulps %%xmm4, %%xmm0\n"
258 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
259 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
260 "addps %%xmm5, %%xmm0\n"
261 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
262 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
263 "movups 16(%%ebx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
264 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
266 "movups %%xmm0, (%%eax)\n"
268 "mulps %%xmm4, %%xmm6\n"
271 "addps %%xmm5, %%xmm6\n"
273 "movups %%xmm6, -16(%%eax)\n"
275 "jnz .first_128_samples\n"
277 "movl 8(%%ebp), %%esi\n" /* buf[0].re */
278 "leal 1020(%%esi), %%edi\n" /* buf[127].im */
279 "movl $16, %%ecx\n" /* loop count */
281 ".second_128_samples:\n"
282 "movss (%%esi), %%xmm0\n" /* buf[i].re */
283 "movss 8(%%esi), %%xmm2\n" /* re1 */
284 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
285 "movss -8(%%edi), %%xmm3\n" /* im1 */
287 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
288 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
290 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
291 "movups (%%ebx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
293 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
294 "movss 16(%%esi), %%xmm6\n" /* re2 */
295 "movss 24(%%esi), %%xmm7\n" /* re3 */
296 "movss -16(%%edi), %%xmm2\n" /* im2 */
297 "movss -24(%%edi), %%xmm3\n" /* im3 */
298 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
299 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
300 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
301 "mulps %%xmm4, %%xmm0\n"
302 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
303 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
305 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
306 "addps %%xmm5, %%xmm0\n"
307 "mulps %%xmm4, %%xmm6\n"
309 "movups 16(%%ebx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
310 "movups %%xmm0, (%%eax)\n"
311 "addps %%xmm5, %%xmm6\n"
315 "movups %%xmm6, -16(%%eax)\n"
317 "jnz .second_128_samples\n"
319 "movl 8(%%ebp), %%eax\n"
320 "leal 512(%%eax), %%esi\n" /* buf[64].re */
321 "leal 508(%%eax), %%edi\n" /* buf[63].im */
322 "movl $16, %%ecx\n" /* loop count */
323 "movl 20(%%ebp), %%eax\n" /* delay */
325 ".first_128_delay:\n"
326 "movss (%%esi), %%xmm0\n"
327 "movss 8(%%esi), %%xmm2\n"
328 "movss (%%edi), %%xmm1\n"
329 "movss -8(%%edi), %%xmm3\n"
331 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
332 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
334 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
335 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
336 "movss 16(%%esi), %%xmm6\n" /* re2 */
337 "movss 24(%%esi), %%xmm7\n" /* re3 */
338 "movss -16(%%edi), %%xmm2\n" /* im2 */
339 "movss -24(%%edi), %%xmm3\n" /* im3 */
340 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
342 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
343 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
344 "mulps %%xmm4, %%xmm0\n"
345 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
346 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
347 "movups %%xmm0, (%%eax)\n"
349 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
351 "mulps %%xmm5, %%xmm6\n"
353 "movups %%xmm6, -16(%%eax)\n"
355 "jnz .first_128_delay\n"
357 "movl 8(%%ebp), %%ebx\n"
358 "leal 4(%%ebx), %%esi\n" /* buf[0].im */
359 "leal 1016(%%ebx), %%edi\n" /* buf[127].re */
360 "movl $16, %%ecx\n" /* loop count */
362 ".second_128_delay:\n"
363 "movss (%%esi), %%xmm0\n"
364 "movss 8(%%esi), %%xmm2\n"
365 "movss (%%edi), %%xmm1\n"
366 "movss -8(%%edi), %%xmm3\n"
368 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
369 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
371 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
372 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
373 "movss 16(%%esi), %%xmm6\n" /* im2 */
374 "movss 24(%%esi), %%xmm7\n" /* im3 */
375 "movss -16(%%edi), %%xmm2\n" /* re2 */
376 "movss -24(%%edi), %%xmm3\n" /* re3 */
377 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
379 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
380 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
381 "mulps %%xmm4, %%xmm1\n"
382 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
383 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
384 "movups %%xmm1, (%%eax)\n"
386 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
388 "mulps %%xmm5, %%xmm2\n"
390 "movups %%xmm2, -16(%%eax)\n"
392 "jnz .second_128_delay\n"
405 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt)
407 __asm__ __volatile__ (
409 "movl %%esp, %%ebp\n"
418 /* movl 20(%%ebp), %%ebx delay */
419 "movl 16(%%ebp), %%edx\n" /* window */
421 "movl 8(%%ebp), %%eax\n" /* buf */
422 "movl $16, %%ecx\n" /* loop count */
423 "leal 516(%%eax), %%esi\n" /* buf[64].im */
424 "leal 504(%%eax), %%edi\n" /* buf[63].re */
425 "movl 12(%%ebp), %%eax\n" /* data */
427 ".first_128_sample:\n"
428 "movss (%%esi), %%xmm0\n"
429 "movss 8(%%esi), %%xmm2\n"
430 "movss (%%edi), %%xmm1\n"
431 "movss -8(%%edi), %%xmm3\n"
433 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
434 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
436 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
437 /* movups (%%ebx), %%xmm5 d3 | d2 | d1 | d0 */
438 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
440 "movss 16(%%esi), %%xmm6\n" /* im2 */
441 "movss 24(%%esi), %%xmm7\n" /* im3 */
442 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
443 "movss -16(%%edi), %%xmm2\n" /* re2 */
444 "movss -24(%%edi), %%xmm3\n" /* re3 */
445 "mulps %%xmm4, %%xmm0\n"
446 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
447 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
448 /* addps %%xmm5, %%xmm0 */
449 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
450 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
451 /* movups 16(%%ebx), %%xmm5 d7 | d6 | d5 | d4 */
452 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
454 "movups %%xmm0, (%%eax)\n"
455 /* addl $32, %%ebx */
456 "mulps %%xmm4, %%xmm6\n"
459 /* addps %%xmm5, %%xmm6 */
461 "movups %%xmm6, -16(%%eax)\n"
463 "jnz .first_128_sample\n"
465 "movl 8(%%ebp), %%esi\n" /* buf[0].re */
466 "leal 1020(%%esi), %%edi\n" /* buf[127].im */
467 "movl $16, %%ecx\n" /* loop count */
469 ".second_128_sample:\n"
470 "movss (%%esi), %%xmm0\n" /* buf[i].re */
471 "movss 8(%%esi), %%xmm2\n" /* re1 */
472 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
473 "movss -8(%%edi), %%xmm3\n" /* im1 */
475 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
476 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
478 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
479 /* movups (%%ebx), %%xmm5 d3 | d2 | d1 | d0 */
481 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
482 "movss 16(%%esi), %%xmm6\n" /* re2 */
483 "movss 24(%%esi), %%xmm7\n" /* re3 */
484 "movss -16(%%edi), %%xmm2\n" /* im2 */
485 "movss -24(%%edi), %%xmm3\n" /* im3 */
486 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
487 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
488 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
489 "mulps %%xmm4, %%xmm0\n"
490 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
491 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
493 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
494 /* addps %%xmm5, %%xmm0 */
495 "mulps %%xmm4, %%xmm6\n"
497 /* movups 16(%%ebx), %%xmm5 d7 | d6 | d5 | d4 */
498 "movups %%xmm0, (%%eax)\n"
499 /* addps %%xmm5, %%xmm6 */
502 /* addl $32, %%ebx */
503 "movups %%xmm6, -16(%%eax)\n"
505 "jnz .second_128_sample\n"
507 "movl 8(%%ebp), %%eax\n"
508 "leal 512(%%eax), %%esi\n" /* buf[64].re */
509 "leal 508(%%eax), %%edi\n" /* buf[63].im */
510 "movl $16, %%ecx\n" /* loop count */
511 "movl 20(%%ebp), %%eax\n" /* delay */
513 ".first_128_delays:\n"
514 "movss (%%esi), %%xmm0\n"
515 "movss 8(%%esi), %%xmm2\n"
516 "movss (%%edi), %%xmm1\n"
517 "movss -8(%%edi), %%xmm3\n"
519 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
520 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
522 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
523 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
524 "movss 16(%%esi), %%xmm6\n" /* re2 */
525 "movss 24(%%esi), %%xmm7\n" /* re3 */
526 "movss -16(%%edi), %%xmm2\n" /* im2 */
527 "movss -24(%%edi), %%xmm3\n" /* im3 */
528 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
530 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
531 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
532 "mulps %%xmm4, %%xmm0\n"
533 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
534 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
535 "movups %%xmm0, (%%eax)\n"
537 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
539 "mulps %%xmm5, %%xmm6\n"
541 "movups %%xmm6, -16(%%eax)\n"
543 "jnz .first_128_delays\n"
545 "movl 8(%%ebp), %%ebx\n"
546 "leal 4(%%ebx), %%esi\n" /* buf[0].im */
547 "leal 1016(%%ebx), %%edi\n" /* buf[127].re */
548 "movl $16, %%ecx\n" /* loop count */
550 ".second_128_delays:\n"
551 "movss (%%esi), %%xmm0\n"
552 "movss 8(%%esi), %%xmm2\n"
553 "movss (%%edi), %%xmm1\n"
554 "movss -8(%%edi), %%xmm3\n"
556 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
557 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
559 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
560 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
561 "movss 16(%%esi), %%xmm6\n" /* im2 */
562 "movss 24(%%esi), %%xmm7\n" /* im3 */
563 "movss -16(%%edi), %%xmm2\n" /* re2 */
564 "movss -24(%%edi), %%xmm3\n" /* re3 */
565 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
567 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
568 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
569 "mulps %%xmm4, %%xmm1\n"
570 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
571 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
572 "movups %%xmm1, (%%eax)\n"
574 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
576 "mulps %%xmm5, %%xmm2\n"
578 "movups %%xmm2, -16(%%eax)\n"
580 "jnz .second_128_delays\n"