1 /*****************************************************************************
2 * ac3_imdct_sse.c: accelerated SSE ac3 DCT
3 *****************************************************************************
4 * Copyright (C) 1999, 2000 VideoLAN
5 * $Id: ac3_imdct_sse.c,v 1.5 2001/07/26 20:00:33 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"
46 # define M_PI 3.14159265358979323846
49 void _M( fft_64p ) ( complex_t *x );
50 void _M( fft_128p ) ( complex_t *a );
52 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse);
53 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse);
54 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt);
55 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt);
58 void _M( imdct_init ) (imdct_t * p_imdct)
61 float scale = 181.019;
63 for (i=0; i < 128; i++)
65 float xcos_i = cos(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
66 float xsin_i = sin(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
67 p_imdct->xcos_sin_sse[i * 4] = xcos_i;
68 p_imdct->xcos_sin_sse[i * 4 + 1] = -xsin_i;
69 p_imdct->xcos_sin_sse[i * 4 + 2] = -xsin_i;
70 p_imdct->xcos_sin_sse[i * 4 + 3] = -xcos_i;
74 void _M( imdct_do_512 ) (imdct_t * p_imdct, float data[], float delay[])
76 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
77 _M( fft_128p ) ( p_imdct->buf );
78 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
79 imdct512_window_delay_sse (p_imdct->buf, data, window, delay);
83 void _M( imdct_do_512_nol ) (imdct_t * p_imdct, float data[], float delay[])
85 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
86 _M( fft_128p ) ( p_imdct->buf );
87 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
88 imdct512_window_delay_nol_sse (p_imdct->buf, data, window, delay);
91 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse)
93 __asm__ __volatile__ (
97 "addl $-4, %%esp\n" /* local variable, loop counter */
106 "movl %%edi, %%ebx\n" /* buf */
107 "movl $64, -4(%%ebp)\n"
111 "movl (%%eax), %%esi\n"
112 "movl 4(%%eax), %%edi\n"
113 "movss (%%ecx, %%esi, 8), %%xmm1\n" /* 2j */
114 "movss (%%ecx, %%edi, 8), %%xmm3\n" /* 2(j+1) */
119 "movaps (%%edx, %%esi, 8), %%xmm0\n" /* -c_j | -s_j | -s_j | c_j */
120 "movaps (%%edx, %%edi, 8), %%xmm2\n" /* -c_j+1 | -s_j+1 | -s_j+1 | c_j+1 */
125 "movss 1020(%%ecx, %%esi, 4), %%xmm4\n" /* 255-2j */
127 "movss 1020(%%ecx, %%edi, 4), %%xmm5\n" /* 255-2(j+1) */
129 "shufps $0, %%xmm1, %%xmm4\n" /* 2j | 2j | 255-2j | 255-2j */
130 "shufps $0, %%xmm3, %%xmm5\n" /* 2(j+1) | 2(j+1) | 255-2(j+1) | 255-2(j+1) */
131 "mulps %%xmm4, %%xmm0\n"
132 "mulps %%xmm5, %%xmm2\n"
133 "movhlps %%xmm0, %%xmm1\n"
134 "movhlps %%xmm2, %%xmm3\n"
136 "addps %%xmm1, %%xmm0\n"
137 "addps %%xmm3, %%xmm2\n"
138 "movlhps %%xmm2, %%xmm0\n"
140 "movaps %%xmm0, -16(%%ebx)\n"
154 : "a" (pmt), "c" (data), "d" (xcos_sin_sse), "D" (buf));
158 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse)
160 __asm__ __volatile__ (
163 "movl $32, %%ebx\n" /* loop counter */
167 "movaps (%%eax), %%xmm0\n" /* im1 | re1 | im0 | re0 */
169 "movaps (%%ecx), %%xmm2\n" /* -c | -s | -s | c */
170 "movhlps %%xmm0, %%xmm1\n" /* im1 | re1 */
171 "movaps 16(%%ecx), %%xmm3\n" /* -c1 | -s1 | -s1 | c1 */
173 "shufps $0x50, %%xmm0, %%xmm0\n" /* im0 | im0 | re0 | re0 */
174 "shufps $0x50, %%xmm1, %%xmm1\n" /* im1 | im1 | re1 | re1 */
176 "movaps 16(%%eax), %%xmm4\n" /* im3 | re3 | im2 | re2 */
178 "shufps $0x27, %%xmm2, %%xmm2\n" /* c | -s | -s | -c */
179 "movhlps %%xmm4, %%xmm5\n" /* im3 | re3 */
180 "shufps $0x27, %%xmm3, %%xmm3\n" /* c1 | -s1 | -s1 | -c1 */
182 "movaps 32(%%ecx), %%xmm6\n" /* -c2 | -s2 | -s2 | c2 */
183 "movaps 48(%%ecx), %%xmm7\n" /* -c3 | -s3 | -s3 | c3 */
185 "shufps $0x50, %%xmm4, %%xmm4\n" /* im2 | im2 | re2 | re2 */
186 "shufps $0x50, %%xmm5, %%xmm5\n" /* im3 | im3 | re3 | re3 */
188 "mulps %%xmm2, %%xmm0\n"
189 "mulps %%xmm3, %%xmm1\n"
191 "shufps $0x27, %%xmm6, %%xmm6\n" /* c2 | -s2 | -s2 | -c2 */
192 "shufps $0x27, %%xmm7, %%xmm7\n" /* c3 | -s3 | -s3 | -c3 */
194 "movhlps %%xmm0, %%xmm2\n"
195 "movhlps %%xmm1, %%xmm3\n"
197 "mulps %%xmm6, %%xmm4\n"
198 "mulps %%xmm7, %%xmm5\n"
200 "addps %%xmm2, %%xmm0\n"
201 "addps %%xmm3, %%xmm1\n"
203 "movhlps %%xmm4, %%xmm6\n"
204 "movhlps %%xmm5, %%xmm7\n"
206 "addps %%xmm6, %%xmm4\n"
207 "addps %%xmm7, %%xmm5\n"
209 "movlhps %%xmm1, %%xmm0\n"
210 "movlhps %%xmm5, %%xmm4\n"
212 "movaps %%xmm0, (%%eax)\n"
213 "movaps %%xmm4, 16(%%eax)\n"
221 : "a" (buf), "c" (xcos_sin_sse) );
224 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt)
226 __asm__ __volatile__ (
237 "movl %%esi, %%ebp\n" /* buf */
238 "movl $16, %%ebx\n" /* loop count */
239 "leal 516(%%ebp), %%esi\n" /* buf[64].im */
240 "leal 504(%%ebp), %%edi\n" /* buf[63].re */
243 ".first_128_samples:\n"
244 "movss (%%esi), %%xmm0\n"
245 "movss 8(%%esi), %%xmm2\n"
246 "movss (%%edi), %%xmm1\n"
247 "movss -8(%%edi), %%xmm3\n"
249 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
250 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
252 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
253 "movaps (%%ecx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
254 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
256 "movss 16(%%esi), %%xmm6\n" /* im2 */
257 "movss 24(%%esi), %%xmm7\n" /* im3 */
258 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
259 "movss -16(%%edi), %%xmm2\n" /* re2 */
260 "movss -24(%%edi), %%xmm3\n" /* re3 */
261 "mulps %%xmm4, %%xmm0\n"
262 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
263 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
264 "addps %%xmm5, %%xmm0\n"
265 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
266 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
267 "movaps 16(%%ecx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
268 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
270 "movaps %%xmm0, (%%eax)\n"
272 "mulps %%xmm4, %%xmm6\n"
275 "addps %%xmm5, %%xmm6\n"
277 "movaps %%xmm6, -16(%%eax)\n"
279 "jnz .first_128_samples\n"
281 "movl %%ebp, %%esi\n" /* buf[0].re */
282 "movl $16, %%ebx\n" /* loop count */
283 "leal 1020(%%ebp), %%edi\n" /* buf[127].im */
286 ".second_128_samples:\n"
287 "movss (%%esi), %%xmm0\n" /* buf[i].re */
288 "movss 8(%%esi), %%xmm2\n" /* re1 */
289 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
290 "movss -8(%%edi), %%xmm3\n" /* im1 */
292 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
293 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
295 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
296 "movaps (%%ecx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
298 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
299 "movss 16(%%esi), %%xmm6\n" /* re2 */
300 "movss 24(%%esi), %%xmm7\n" /* re3 */
301 "movss -16(%%edi), %%xmm2\n" /* im2 */
302 "movss -24(%%edi), %%xmm3\n" /* im3 */
303 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
304 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
305 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
306 "mulps %%xmm4, %%xmm0\n"
307 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
308 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
310 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
311 "addps %%xmm5, %%xmm0\n"
312 "mulps %%xmm4, %%xmm6\n"
314 "movaps 16(%%ecx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
315 "movaps %%xmm0, (%%eax)\n"
316 "addps %%xmm5, %%xmm6\n"
320 "movaps %%xmm6, -16(%%eax)\n"
322 "jnz .second_128_samples\n"
324 "leal 512(%%ebp), %%esi\n" /* buf[64].re */
325 "leal 508(%%ebp), %%edi\n" /* buf[63].im */
326 "movl $16, %%ebx\n" /* loop count */
327 "addl $-1024, %%ecx\n" /* delay */
330 ".first_128_delay:\n"
331 "movss (%%esi), %%xmm0\n"
332 "movss 8(%%esi), %%xmm2\n"
333 "movss (%%edi), %%xmm1\n"
334 "movss -8(%%edi), %%xmm3\n"
336 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
337 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
339 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
340 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
341 "movss 16(%%esi), %%xmm6\n" /* re2 */
342 "movss 24(%%esi), %%xmm7\n" /* re3 */
343 "movss -16(%%edi), %%xmm2\n" /* im2 */
344 "movss -24(%%edi), %%xmm3\n" /* im3 */
345 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
347 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
348 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
349 "mulps %%xmm4, %%xmm0\n"
350 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
351 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
352 "movaps %%xmm0, (%%ecx)\n"
354 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
356 "mulps %%xmm5, %%xmm6\n"
358 "movaps %%xmm6, -16(%%ecx)\n"
360 "jnz .first_128_delay\n"
362 "leal 4(%%ebp), %%esi\n" /* buf[0].im */
363 "leal 1016(%%ebp), %%edi\n" /* buf[127].re */
364 "movl $16, %%ebx\n" /* loop count */
367 ".second_128_delay:\n"
368 "movss (%%esi), %%xmm0\n"
369 "movss 8(%%esi), %%xmm2\n"
370 "movss (%%edi), %%xmm1\n"
371 "movss -8(%%edi), %%xmm3\n"
373 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
374 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
376 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
377 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
378 "movss 16(%%esi), %%xmm6\n" /* im2 */
379 "movss 24(%%esi), %%xmm7\n" /* im3 */
380 "movss -16(%%edi), %%xmm2\n" /* re2 */
381 "movss -24(%%edi), %%xmm3\n" /* re3 */
382 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
384 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
385 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
386 "mulps %%xmm4, %%xmm1\n"
387 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
388 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
389 "movaps %%xmm1, (%%ecx)\n"
391 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
393 "mulps %%xmm5, %%xmm2\n"
395 "movaps %%xmm2, -16(%%ecx)\n"
397 "jnz .second_128_delay\n"
406 : "=S" (buf), "=a" (data_ptr), "=c" (delay_prt), "=d" (window_prt)
407 : "S" (buf), "a" (data_ptr), "c" (delay_prt), "d" (window_prt));
411 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt)
413 __asm__ __volatile__ (
424 "movl %%esi, %%ebp\n" /* buf */
425 "movl $16, %%ebx\n" /* loop count */
426 "leal 516(%%ebp), %%esi\n" /* buf[64].im */
427 "leal 504(%%ebp), %%edi\n" /* buf[63].re */
430 ".first_128_sample:\n"
431 "movss (%%esi), %%xmm0\n"
432 "movss 8(%%esi), %%xmm2\n"
433 "movss (%%edi), %%xmm1\n"
434 "movss -8(%%edi), %%xmm3\n"
436 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
437 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
439 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
440 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
442 "movss 16(%%esi), %%xmm6\n" /* im2 */
443 "movss 24(%%esi), %%xmm7\n" /* im3 */
444 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
445 "movss -16(%%edi), %%xmm2\n" /* re2 */
446 "movss -24(%%edi), %%xmm3\n" /* re3 */
447 "mulps %%xmm4, %%xmm0\n"
448 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
449 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
450 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
451 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
452 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
454 "movaps %%xmm0, (%%eax)\n"
455 "mulps %%xmm4, %%xmm6\n"
459 "movaps %%xmm6, -16(%%eax)\n"
461 "jnz .first_128_sample\n"
463 "movl %%ebp, %%esi\n" /* buf[0].re */
464 "movl $16, %%ebx\n" /* loop count */
465 "leal 1020(%%ebp), %%edi\n" /* buf[127].im */
468 ".second_128_sample:\n"
469 "movss (%%esi), %%xmm0\n" /* buf[i].re */
470 "movss 8(%%esi), %%xmm2\n" /* re1 */
471 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
472 "movss -8(%%edi), %%xmm3\n" /* im1 */
474 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
475 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
477 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
479 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
480 "movss 16(%%esi), %%xmm6\n" /* re2 */
481 "movss 24(%%esi), %%xmm7\n" /* re3 */
482 "movss -16(%%edi), %%xmm2\n" /* im2 */
483 "movss -24(%%edi), %%xmm3\n" /* im3 */
484 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
485 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
486 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
487 "mulps %%xmm4, %%xmm0\n"
488 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
489 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
491 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
492 "mulps %%xmm4, %%xmm6\n"
494 "movaps %%xmm0, (%%eax)\n"
497 "movaps %%xmm6, -16(%%eax)\n"
499 "jnz .second_128_sample\n"
501 "leal 512(%%ebp), %%esi\n" /* buf[64].re */
502 "leal 508(%%ebp), %%edi\n" /* buf[63].im */
503 "movl $16, %%ebx\n" /* loop count */
504 "addl $-1024, %%ecx\n" /* delay */
507 ".first_128_delays:\n"
508 "movss (%%esi), %%xmm0\n"
509 "movss 8(%%esi), %%xmm2\n"
510 "movss (%%edi), %%xmm1\n"
511 "movss -8(%%edi), %%xmm3\n"
513 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
514 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
516 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
517 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
518 "movss 16(%%esi), %%xmm6\n" /* re2 */
519 "movss 24(%%esi), %%xmm7\n" /* re3 */
520 "movss -16(%%edi), %%xmm2\n" /* im2 */
521 "movss -24(%%edi), %%xmm3\n" /* im3 */
522 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
524 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
525 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
526 "mulps %%xmm4, %%xmm0\n"
527 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
528 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
529 "movaps %%xmm0, (%%ecx)\n"
531 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
533 "mulps %%xmm5, %%xmm6\n"
535 "movaps %%xmm6, -16(%%ecx)\n"
537 "jnz .first_128_delays\n"
539 "leal 4(%%ebp), %%esi\n" /* buf[0].im */
540 "leal 1016(%%ebp), %%edi\n" /* buf[127].re */
541 "movl $16, %%ebx\n" /* loop count */
544 ".second_128_delays:\n"
545 "movss (%%esi), %%xmm0\n"
546 "movss 8(%%esi), %%xmm2\n"
547 "movss (%%edi), %%xmm1\n"
548 "movss -8(%%edi), %%xmm3\n"
550 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
551 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
553 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
554 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
555 "movss 16(%%esi), %%xmm6\n" /* im2 */
556 "movss 24(%%esi), %%xmm7\n" /* im3 */
557 "movss -16(%%edi), %%xmm2\n" /* re2 */
558 "movss -24(%%edi), %%xmm3\n" /* re3 */
559 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
561 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
562 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
563 "mulps %%xmm4, %%xmm1\n"
564 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
565 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
566 "movaps %%xmm1, (%%ecx)\n"
568 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
570 "mulps %%xmm5, %%xmm2\n"
572 "movaps %%xmm2, -16(%%ecx)\n"
574 "jnz .second_128_delays\n"
583 : "=S" (buf), "=a" (data_ptr), "=c" (delay_prt), "=d" (window_prt)
584 : "S" (buf), "a" (data_ptr), "c" (delay_prt), "d" (window_prt));