1 /*****************************************************************************
2 * ac3_imdct_sse.c: ac3 DCT
3 *****************************************************************************
4 * Copyright (C) 1999, 2000 VideoLAN
5 * $Id: ac3_imdct_sse.c,v 1.1 2001/05/14 15:58:04 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 *****************************************************************************/
35 #include "intf_msg.h" /* intf_DbgMsg(), intf_ErrMsg() */
37 #include "stream_control.h"
38 #include "input_ext-dec.h"
40 #include "ac3_decoder.h"
42 #include "ac3_imdct_sse.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 fft_64p_sse (complex_t *x);
92 void fft_128p_sse(complex_t *a);
93 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse);
94 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse);
95 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt);
96 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt);
99 int imdct_init_sse (imdct_t * p_imdct)
102 float scale = 181.019;
104 intf_WarnMsg (1, "ac3dec: using MMX_SSE for imdct");
105 p_imdct->imdct_do_512 = imdct_do_512_sse;
106 p_imdct->imdct_do_512_nol = imdct_do_512_nol_sse;
107 p_imdct->fft_64p = fft_64p_sse;
109 for (i=0; i < 128; i++)
111 float xcos_i = cos(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
112 float xsin_i = sin(2.0f * M_PI * (8*i+1)/(8*N)) * scale;
113 p_imdct->xcos_sin_sse[i * 4] = xcos_i;
114 p_imdct->xcos_sin_sse[i * 4 + 1] = -xsin_i;
115 p_imdct->xcos_sin_sse[i * 4 + 2] = -xsin_i;
116 p_imdct->xcos_sin_sse[i * 4 + 3] = -xcos_i;
121 void imdct_do_512_sse (imdct_t * p_imdct, float data[], float delay[])
123 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
124 fft_128p_sse (p_imdct->buf);
125 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
126 imdct512_window_delay_sse (p_imdct->buf, data, window, delay);
130 void imdct_do_512_nol_sse (imdct_t * p_imdct, float data[], float delay[])
132 imdct512_pre_ifft_twiddle_sse (pm128, p_imdct->buf, data, p_imdct->xcos_sin_sse);
133 fft_128p_sse (p_imdct->buf);
134 imdct512_post_ifft_twiddle_sse (p_imdct->buf, p_imdct->xcos_sin_sse);
135 imdct512_window_delay_nol_sse (p_imdct->buf, data, window, delay);
138 static void imdct512_pre_ifft_twiddle_sse (const int *pmt, complex_t *buf, float *data, float *xcos_sin_sse)
140 __asm__ __volatile__ (
142 "movl %%esp, %%ebp\n"
143 "addl $-4, %%esp\n" /* local variable, loop counter */
152 "movl 8(%%ebp), %%eax\n" /* pmt */
153 "movl 12(%%ebp), %%ebx\n" /* buf */
154 "movl 16(%%ebp), %%ecx\n" /* data */
155 "movl 20(%%ebp), %%edx\n" /* xcos_sin_sse */
156 "movl $64, -4(%%ebp)\n"
159 "movl (%%eax), %%esi\n"
160 "movl 4(%%eax), %%edi\n"
161 "movss (%%ecx, %%esi, 8), %%xmm1\n" /* 2j */
162 "movss (%%ecx, %%edi, 8), %%xmm3\n" /* 2(j+1) */
167 "movups (%%edx, %%esi, 8), %%xmm0\n" /* -c_j | -s_j | -s_j | c_j */
168 "movups (%%edx, %%edi, 8), %%xmm2\n" /* -c_j+1 | -s_j+1 | -s_j+1 | c_j+1 */
173 "movss 1020(%%ecx, %%esi, 4), %%xmm4\n" /* 255-2j */
175 "movss 1020(%%ecx, %%edi, 4), %%xmm5\n" /* 255-2(j+1) */
177 "shufps $0, %%xmm1, %%xmm4\n" /* 2j | 2j | 255-2j | 255-2j */
178 "shufps $0, %%xmm3, %%xmm5\n" /* 2(j+1) | 2(j+1) | 255-2(j+1) | 255-2(j+1) */
179 "mulps %%xmm4, %%xmm0\n"
180 "mulps %%xmm5, %%xmm2\n"
181 "movhlps %%xmm0, %%xmm1\n"
182 "movhlps %%xmm2, %%xmm3\n"
184 "addps %%xmm1, %%xmm0\n"
185 "addps %%xmm3, %%xmm2\n"
186 "movlhps %%xmm2, %%xmm0\n"
188 "movups %%xmm0, -16(%%ebx)\n"
204 static void imdct512_post_ifft_twiddle_sse (complex_t *buf, float *xcos_sin_sse)
206 __asm__ __volatile__ (
208 "movl $32, %%ecx\n" /* loop counter */
211 "movups (%%eax), %%xmm0\n" /* im1 | re1 | im0 | re0 */
213 "movups (%%ebx), %%xmm2\n" /* -c | -s | -s | c */
214 "movhlps %%xmm0, %%xmm1\n" /* im1 | re1 */
215 "movups 16(%%ebx), %%xmm3\n" /* -c1 | -s1 | -s1 | c1 */
217 "shufps $0x50, %%xmm0, %%xmm0\n" /* im0 | im0 | re0 | re0 */
218 "shufps $0x50, %%xmm1, %%xmm1\n" /* im1 | im1 | re1 | re1 */
220 "movups 16(%%eax), %%xmm4\n" /* im3 | re3 | im2 | re2 */
222 "shufps $0x27, %%xmm2, %%xmm2\n" /* c | -s | -s | -c */
223 "movhlps %%xmm4, %%xmm5\n" /* im3 | re3 */
224 "shufps $0x27, %%xmm3, %%xmm3\n" /* c1 | -s1 | -s1 | -c1 */
226 "movups 32(%%ebx), %%xmm6\n" /* -c2 | -s2 | -s2 | c2 */
227 "movups 48(%%ebx), %%xmm7\n" /* -c3 | -s3 | -s3 | c3 */
229 "shufps $0x50, %%xmm4, %%xmm4\n" /* im2 | im2 | re2 | re2 */
230 "shufps $0x50, %%xmm5, %%xmm5\n" /* im3 | im3 | re3 | re3 */
232 "mulps %%xmm2, %%xmm0\n"
233 "mulps %%xmm3, %%xmm1\n"
235 "shufps $0x27, %%xmm6, %%xmm6\n" /* c2 | -s2 | -s2 | -c2 */
236 "shufps $0x27, %%xmm7, %%xmm7\n" /* c3 | -s3 | -s3 | -c3 */
238 "movhlps %%xmm0, %%xmm2\n"
239 "movhlps %%xmm1, %%xmm3\n"
241 "mulps %%xmm6, %%xmm4\n"
242 "mulps %%xmm7, %%xmm5\n"
244 "addps %%xmm2, %%xmm0\n"
245 "addps %%xmm3, %%xmm1\n"
247 "movhlps %%xmm4, %%xmm6\n"
248 "movhlps %%xmm5, %%xmm7\n"
250 "addps %%xmm6, %%xmm4\n"
251 "addps %%xmm7, %%xmm5\n"
253 "movlhps %%xmm1, %%xmm0\n"
254 "movlhps %%xmm5, %%xmm4\n"
256 "movups %%xmm0, (%%eax)\n"
257 "movups %%xmm4, 16(%%eax)\n"
265 : "a" (buf), "b" (xcos_sin_sse) );
268 static void imdct512_window_delay_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt)
270 __asm__ __volatile__ (
272 "movl %%esp, %%ebp\n"
281 "movl 20(%%ebp), %%ebx\n" /* delay */
282 "movl 16(%%ebp), %%edx\n" /* window */
284 "movl 8(%%ebp), %%eax\n" /* buf */
285 "movl $16, %%ecx\n" /* loop count */
286 "leal 516(%%eax), %%esi\n" /* buf[64].im */
287 "leal 504(%%eax), %%edi\n" /* buf[63].re */
288 "movl 12(%%ebp), %%eax\n" /* data */
290 ".first_128_samples:\n"
291 "movss (%%esi), %%xmm0\n"
292 "movss 8(%%esi), %%xmm2\n"
293 "movss (%%edi), %%xmm1\n"
294 "movss -8(%%edi), %%xmm3\n"
296 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
297 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
299 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
300 "movups (%%ebx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
301 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
303 "movss 16(%%esi), %%xmm6\n" /* im2 */
304 "movss 24(%%esi), %%xmm7\n" /* im3 */
305 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
306 "movss -16(%%edi), %%xmm2\n" /* re2 */
307 "movss -24(%%edi), %%xmm3\n" /* re3 */
308 "mulps %%xmm4, %%xmm0\n"
309 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
310 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
311 "addps %%xmm5, %%xmm0\n"
312 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
313 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
314 "movups 16(%%ebx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
315 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
317 "movups %%xmm0, (%%eax)\n"
319 "mulps %%xmm4, %%xmm6\n"
322 "addps %%xmm5, %%xmm6\n"
324 "movups %%xmm6, -16(%%eax)\n"
326 "jnz .first_128_samples\n"
328 "movl 8(%%ebp), %%esi\n" /* buf[0].re */
329 "leal 1020(%%esi), %%edi\n" /* buf[127].im */
330 "movl $16, %%ecx\n" /* loop count */
332 ".second_128_samples:\n"
333 "movss (%%esi), %%xmm0\n" /* buf[i].re */
334 "movss 8(%%esi), %%xmm2\n" /* re1 */
335 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
336 "movss -8(%%edi), %%xmm3\n" /* im1 */
338 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
339 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
341 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
342 "movups (%%ebx), %%xmm5\n" /* d3 | d2 | d1 | d0 */
344 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
345 "movss 16(%%esi), %%xmm6\n" /* re2 */
346 "movss 24(%%esi), %%xmm7\n" /* re3 */
347 "movss -16(%%edi), %%xmm2\n" /* im2 */
348 "movss -24(%%edi), %%xmm3\n" /* im3 */
349 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
350 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
351 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
352 "mulps %%xmm4, %%xmm0\n"
353 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
354 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
356 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
357 "addps %%xmm5, %%xmm0\n"
358 "mulps %%xmm4, %%xmm6\n"
360 "movups 16(%%ebx), %%xmm5\n" /* d7 | d6 | d5 | d4 */
361 "movups %%xmm0, (%%eax)\n"
362 "addps %%xmm5, %%xmm6\n"
366 "movups %%xmm6, -16(%%eax)\n"
368 "jnz .second_128_samples\n"
370 "movl 8(%%ebp), %%eax\n"
371 "leal 512(%%eax), %%esi\n" /* buf[64].re */
372 "leal 508(%%eax), %%edi\n" /* buf[63].im */
373 "movl $16, %%ecx\n" /* loop count */
374 "movl 20(%%ebp), %%eax\n" /* delay */
376 ".first_128_delay:\n"
377 "movss (%%esi), %%xmm0\n"
378 "movss 8(%%esi), %%xmm2\n"
379 "movss (%%edi), %%xmm1\n"
380 "movss -8(%%edi), %%xmm3\n"
382 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
383 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
385 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
386 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
387 "movss 16(%%esi), %%xmm6\n" /* re2 */
388 "movss 24(%%esi), %%xmm7\n" /* re3 */
389 "movss -16(%%edi), %%xmm2\n" /* im2 */
390 "movss -24(%%edi), %%xmm3\n" /* im3 */
391 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
393 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
394 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
395 "mulps %%xmm4, %%xmm0\n"
396 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
397 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
398 "movups %%xmm0, (%%eax)\n"
400 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
402 "mulps %%xmm5, %%xmm6\n"
404 "movups %%xmm6, -16(%%eax)\n"
406 "jnz .first_128_delay\n"
408 "movl 8(%%ebp), %%ebx\n"
409 "leal 4(%%ebx), %%esi\n" /* buf[0].im */
410 "leal 1016(%%ebx), %%edi\n" /* buf[127].re */
411 "movl $16, %%ecx\n" /* loop count */
413 ".second_128_delay:\n"
414 "movss (%%esi), %%xmm0\n"
415 "movss 8(%%esi), %%xmm2\n"
416 "movss (%%edi), %%xmm1\n"
417 "movss -8(%%edi), %%xmm3\n"
419 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
420 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
422 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
423 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
424 "movss 16(%%esi), %%xmm6\n" /* im2 */
425 "movss 24(%%esi), %%xmm7\n" /* im3 */
426 "movss -16(%%edi), %%xmm2\n" /* re2 */
427 "movss -24(%%edi), %%xmm3\n" /* re3 */
428 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
430 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
431 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
432 "mulps %%xmm4, %%xmm1\n"
433 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
434 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
435 "movups %%xmm1, (%%eax)\n"
437 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
439 "mulps %%xmm5, %%xmm2\n"
441 "movups %%xmm2, -16(%%eax)\n"
443 "jnz .second_128_delay\n"
456 static void imdct512_window_delay_nol_sse (complex_t *buf, float *data_ptr, float *window_prt, float *delay_prt)
458 __asm__ __volatile__ (
460 "movl %%esp, %%ebp\n"
469 /* movl 20(%%ebp), %%ebx delay */
470 "movl 16(%%ebp), %%edx\n" /* window */
472 "movl 8(%%ebp), %%eax\n" /* buf */
473 "movl $16, %%ecx\n" /* loop count */
474 "leal 516(%%eax), %%esi\n" /* buf[64].im */
475 "leal 504(%%eax), %%edi\n" /* buf[63].re */
476 "movl 12(%%ebp), %%eax\n" /* data */
478 ".first_128_sample:\n"
479 "movss (%%esi), %%xmm0\n"
480 "movss 8(%%esi), %%xmm2\n"
481 "movss (%%edi), %%xmm1\n"
482 "movss -8(%%edi), %%xmm3\n"
484 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
485 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
487 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
488 /* movups (%%ebx), %%xmm5 d3 | d2 | d1 | d0 */
489 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
491 "movss 16(%%esi), %%xmm6\n" /* im2 */
492 "movss 24(%%esi), %%xmm7\n" /* im3 */
493 "subps %%xmm1, %%xmm0\n" /* -re1 | im1 | -re0 | im0 */
494 "movss -16(%%edi), %%xmm2\n" /* re2 */
495 "movss -24(%%edi), %%xmm3\n" /* re3 */
496 "mulps %%xmm4, %%xmm0\n"
497 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
498 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
499 /* addps %%xmm5, %%xmm0 */
500 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
501 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
502 /* movups 16(%%ebx), %%xmm5 d7 | d6 | d5 | d4 */
503 "subps %%xmm2, %%xmm6\n" /* -re3 | im3 | -re2 | im2 */
505 "movups %%xmm0, (%%eax)\n"
506 /* addl $32, %%ebx */
507 "mulps %%xmm4, %%xmm6\n"
510 /* addps %%xmm5, %%xmm6 */
512 "movups %%xmm6, -16(%%eax)\n"
514 "jnz .first_128_sample\n"
516 "movl 8(%%ebp), %%esi\n" /* buf[0].re */
517 "leal 1020(%%esi), %%edi\n" /* buf[127].im */
518 "movl $16, %%ecx\n" /* loop count */
520 ".second_128_sample:\n"
521 "movss (%%esi), %%xmm0\n" /* buf[i].re */
522 "movss 8(%%esi), %%xmm2\n" /* re1 */
523 "movss (%%edi), %%xmm1\n" /* buf[127-i].im */
524 "movss -8(%%edi), %%xmm3\n" /* im1 */
526 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
527 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im1 */
529 "movups (%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
530 /* movups (%%ebx), %%xmm5 d3 | d2 | d1 | d0 */
532 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
533 "movss 16(%%esi), %%xmm6\n" /* re2 */
534 "movss 24(%%esi), %%xmm7\n" /* re3 */
535 "movss -16(%%edi), %%xmm2\n" /* im2 */
536 "movss -24(%%edi), %%xmm3\n" /* im3 */
537 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
538 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
539 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
540 "mulps %%xmm4, %%xmm0\n"
541 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
542 "movups 16(%%edx), %%xmm4\n" /* w7 | w6 | w5 | w4 */
544 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
545 /* addps %%xmm5, %%xmm0 */
546 "mulps %%xmm4, %%xmm6\n"
548 /* movups 16(%%ebx), %%xmm5 d7 | d6 | d5 | d4 */
549 "movups %%xmm0, (%%eax)\n"
550 /* addps %%xmm5, %%xmm6 */
553 /* addl $32, %%ebx */
554 "movups %%xmm6, -16(%%eax)\n"
556 "jnz .second_128_sample\n"
558 "movl 8(%%ebp), %%eax\n"
559 "leal 512(%%eax), %%esi\n" /* buf[64].re */
560 "leal 508(%%eax), %%edi\n" /* buf[63].im */
561 "movl $16, %%ecx\n" /* loop count */
562 "movl 20(%%ebp), %%eax\n" /* delay */
564 ".first_128_delays:\n"
565 "movss (%%esi), %%xmm0\n"
566 "movss 8(%%esi), %%xmm2\n"
567 "movss (%%edi), %%xmm1\n"
568 "movss -8(%%edi), %%xmm3\n"
570 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | re1 | 0.0 | re0 */
571 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | im1 | 0.0 | im0 */
573 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
574 "shufps $0xb1, %%xmm1, %%xmm1\n"/* im1 | 0.0 | im0 | 0.0 */
575 "movss 16(%%esi), %%xmm6\n" /* re2 */
576 "movss 24(%%esi), %%xmm7\n" /* re3 */
577 "movss -16(%%edi), %%xmm2\n" /* im2 */
578 "movss -24(%%edi), %%xmm3\n" /* im3 */
579 "subps %%xmm1, %%xmm0\n" /* -im1 | re1 | -im0 | re0 */
581 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | re3 | 0.0 | re2 */
582 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | im3 | 0.0 | im2 */
583 "mulps %%xmm4, %%xmm0\n"
584 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
585 "shufps $0xb1, %%xmm2, %%xmm2\n"/* im3 | 0.0 | im2 | 0.0 */
586 "movups %%xmm0, (%%eax)\n"
588 "subps %%xmm2, %%xmm6\n" /* -im3 | re3 | -im2 | re2 */
590 "mulps %%xmm5, %%xmm6\n"
592 "movups %%xmm6, -16(%%eax)\n"
594 "jnz .first_128_delays\n"
596 "movl 8(%%ebp), %%ebx\n"
597 "leal 4(%%ebx), %%esi\n" /* buf[0].im */
598 "leal 1016(%%ebx), %%edi\n" /* buf[127].re */
599 "movl $16, %%ecx\n" /* loop count */
601 ".second_128_delays:\n"
602 "movss (%%esi), %%xmm0\n"
603 "movss 8(%%esi), %%xmm2\n"
604 "movss (%%edi), %%xmm1\n"
605 "movss -8(%%edi), %%xmm3\n"
607 "movlhps %%xmm2, %%xmm0\n" /* 0.0 | im1 | 0.0 | im0 */
608 "movlhps %%xmm3, %%xmm1\n" /* 0.0 | re1 | 0.0 | re0 */
610 "movups -16(%%edx), %%xmm4\n" /* w3 | w2 | w1 | w0 */
611 "shufps $0xb1, %%xmm1, %%xmm1\n"/* re1 | 0.0 | re0 | 0.0 */
612 "movss 16(%%esi), %%xmm6\n" /* im2 */
613 "movss 24(%%esi), %%xmm7\n" /* im3 */
614 "movss -16(%%edi), %%xmm2\n" /* re2 */
615 "movss -24(%%edi), %%xmm3\n" /* re3 */
616 "subps %%xmm0, %%xmm1\n" /* re1 | -im1 | re0 | -im0 */
618 "movlhps %%xmm7, %%xmm6\n" /* 0.0 | im3 | 0.0 | im2 */
619 "movlhps %%xmm3, %%xmm2\n" /* 0.0 | re3 | 0.0 | re2 */
620 "mulps %%xmm4, %%xmm1\n"
621 "movups (%%edx), %%xmm5\n" /* w7 | w6 | w5 | w4 */
622 "shufps $0xb1, %%xmm2, %%xmm2\n"/* re3 | 0.0 | re2 | 0.0 */
623 "movups %%xmm1, (%%eax)\n"
625 "subps %%xmm6, %%xmm2\n" /* re | -im3 | re | -im2 */
627 "mulps %%xmm5, %%xmm2\n"
629 "movups %%xmm2, -16(%%eax)\n"
631 "jnz .second_128_delays\n"