Mercurial > mplayer.hg
comparison liba52/imdct_3dnow.h @ 8230:330086b89d8f
minor optimization & gcc-CVS fix/workaround patch by (Glen Nakamura <glen at imodulo dot com>)
author | michael |
---|---|
date | Mon, 18 Nov 2002 10:12:03 +0000 |
parents | d3aedd7db02c |
children | 772d6d27fd66 |
comparison
equal
deleted
inserted
replaced
8229:0f7774d3c5d9 | 8230:330086b89d8f |
---|---|
37 "pfadd %%mm1, %%mm3\n\t" /* vi.im = x[3].re - x[1].re; */ | 37 "pfadd %%mm1, %%mm3\n\t" /* vi.im = x[3].re - x[1].re; */ |
38 "movq %%mm3, %%mm4\n\t" /* vi.re =-x[3].im + x[1].im; mm4 = vi */ | 38 "movq %%mm3, %%mm4\n\t" /* vi.re =-x[3].im + x[1].im; mm4 = vi */ |
39 #ifdef HAVE_3DNOWEX | 39 #ifdef HAVE_3DNOWEX |
40 "pswapd %%mm4, %%mm4\n\t" | 40 "pswapd %%mm4, %%mm4\n\t" |
41 #else | 41 #else |
42 "movq %%mm4, %%mm5\n\t" | 42 "punpckldq %%mm4, %%mm5\n\t" |
43 "psrlq $32, %%mm4\n\t" | 43 "punpckhdq %%mm5, %%mm4\n\t" |
44 "punpckldq %%mm5, %%mm4\n\t" | |
45 #endif | 44 #endif |
46 "movq (%1), %%mm5\n\t" /* yb.re = x[0].re - x[2].re; */ | 45 "movq (%1), %%mm5\n\t" /* yb.re = x[0].re - x[2].re; */ |
47 "movq (%1), %%mm6\n\t" /* yt.re = x[0].re + x[2].re; */ | 46 "movq (%1), %%mm6\n\t" /* yt.re = x[0].re + x[2].re; */ |
48 "movq 24(%1), %%mm7\n\t" /* u.re = x[3].re + x[1].re; */ | 47 "movq 24(%1), %%mm7\n\t" /* u.re = x[3].re + x[1].re; */ |
49 "pfsub 16(%1), %%mm5\n\t" /* yb.im = x[0].im - x[2].im; mm5 = yb */ | 48 "pfsub 16(%1), %%mm5\n\t" /* yb.im = x[0].im - x[2].im; mm5 = yb */ |
115 "movq 16(%1), %%mm4\n\t" | 114 "movq 16(%1), %%mm4\n\t" |
116 "movq %%mm1, %%mm2\n\t" | 115 "movq %%mm1, %%mm2\n\t" |
117 #ifdef HAVE_3DNOWEX | 116 #ifdef HAVE_3DNOWEX |
118 "pswapd %%mm3, %%mm3\n\t" | 117 "pswapd %%mm3, %%mm3\n\t" |
119 #else | 118 #else |
120 "movq %%mm3, %%mm6\n\t" | 119 "punpckldq %%mm3, %%mm6\n\t" |
121 "psrlq $32, %%mm3\n\t" | 120 "punpckhdq %%mm6, %%mm3\n\t" |
122 "punpckldq %%mm6, %%mm3\n\t" | |
123 #endif | 121 #endif |
124 "pfadd %%mm0, %%mm1\n\t" | 122 "pfadd %%mm0, %%mm1\n\t" |
125 "movq %%mm4, %%mm5\n\t" | 123 "movq %%mm4, %%mm5\n\t" |
126 "pfsub %%mm0, %%mm2\n\t" | 124 "pfsub %%mm0, %%mm2\n\t" |
127 "pfadd %%mm3, %%mm4\n\t" | 125 "pfadd %%mm3, %%mm4\n\t" |
147 "movq 56(%3), %%mm3\n\t" | 145 "movq 56(%3), %%mm3\n\t" |
148 "pfsub 40(%3), %%mm0\n\t" | 146 "pfsub 40(%3), %%mm0\n\t" |
149 #ifdef HAVE_3DNOWEX | 147 #ifdef HAVE_3DNOWEX |
150 "pswapd %%mm1, %%mm1\n\t" | 148 "pswapd %%mm1, %%mm1\n\t" |
151 #else | 149 #else |
152 "movq %%mm1, %%mm2\n\t" | 150 "punpckldq %%mm1, %%mm2\n\t" |
153 "psrlq $32, %%mm1\n\t" | 151 "punpckhdq %%mm2, %%mm1\n\t" |
154 "punpckldq %%mm2,%%mm1\n\t" | |
155 #endif | 152 #endif |
156 "pxor %%mm7, %%mm1\n\t" | 153 "pxor %%mm7, %%mm1\n\t" |
157 "pfadd %%mm1, %%mm0\n\t" | 154 "pfadd %%mm1, %%mm0\n\t" |
158 #ifdef HAVE_3DNOWEX | 155 #ifdef HAVE_3DNOWEX |
159 "pswapd %%mm3, %%mm3\n\t" | 156 "pswapd %%mm3, %%mm3\n\t" |
160 #else | 157 #else |
161 "movq %%mm3, %%mm2\n\t" | 158 "punpckldq %%mm3, %%mm2\n\t" |
162 "psrlq $32, %%mm3\n\t" | 159 "punpckhdq %%mm2, %%mm3\n\t" |
163 "punpckldq %%mm2,%%mm3\n\t" | |
164 #endif | 160 #endif |
165 "pxor %%mm6, %%mm3\n\t" | 161 "pxor %%mm6, %%mm3\n\t" |
166 "pfadd %%mm3, %%mm0\n\t" | 162 "pfadd %%mm3, %%mm0\n\t" |
167 "movq %%mm0, %%mm1\n\t" | 163 "movq %%mm0, %%mm1\n\t" |
168 "pxor %%mm6, %%mm1\n\t" | 164 "pxor %%mm6, %%mm1\n\t" |
171 | 167 |
172 "movq 40(%3), %%mm5\n\t" | 168 "movq 40(%3), %%mm5\n\t" |
173 #ifdef HAVE_3DNOWEX | 169 #ifdef HAVE_3DNOWEX |
174 "pswapd %%mm5, %%mm5\n\t" | 170 "pswapd %%mm5, %%mm5\n\t" |
175 #else | 171 #else |
176 "movq %%mm5, %%mm1\n\t" | 172 "punpckldq %%mm5, %%mm1\n\t" |
177 "psrlq $32, %%mm5\n\t" | 173 "punpckhdq %%mm1, %%mm5\n\t" |
178 "punpckldq %%mm1,%%mm5\n\t" | |
179 #endif | 174 #endif |
180 "movq %%mm5, %0\n\t" | 175 "movq %%mm5, %0\n\t" |
181 | 176 |
182 "movq 8(%3), %%mm1\n\t" | 177 "movq 8(%3), %%mm1\n\t" |
183 "movq %%mm1, %%mm2\n\t" | 178 "movq %%mm1, %%mm2\n\t" |
196 "movq %1, %%mm0\n\t" | 191 "movq %1, %%mm0\n\t" |
197 #ifdef HAVE_3DNOWEX | 192 #ifdef HAVE_3DNOWEX |
198 "pswapd %3, %%mm1\n\t" | 193 "pswapd %3, %%mm1\n\t" |
199 #else | 194 #else |
200 "movq %3, %%mm1\n\t" | 195 "movq %3, %%mm1\n\t" |
201 "psrlq $32, %%mm1\n\t" | 196 "punpckldq %%mm1, %%mm2\n\t" |
202 "punpckldq %3, %%mm1\n\t" | 197 "punpckhdq %%mm2, %%mm1\n\t" |
203 #endif | 198 #endif |
204 "pxor %%mm6, %%mm1\n\t" | 199 "pxor %%mm6, %%mm1\n\t" |
205 "pfadd %%mm1, %%mm0\n\t" | 200 "pfadd %%mm1, %%mm0\n\t" |
206 "movq %2, %%mm2\n\t" | 201 "movq %2, %%mm2\n\t" |
207 "movq 56(%4), %%mm3\n\t" | 202 "movq 56(%4), %%mm3\n\t" |
208 "pxor %%mm7, %%mm3\n\t" | 203 "pxor %%mm7, %%mm3\n\t" |
209 "pfadd %%mm3, %%mm2\n\t" | 204 "pfadd %%mm3, %%mm2\n\t" |
210 #ifdef HAVE_3DNOWEX | 205 #ifdef HAVE_3DNOWEX |
211 "pswapd %%mm2, %%mm2\n\t" | 206 "pswapd %%mm2, %%mm2\n\t" |
212 #else | 207 #else |
213 "movq %%mm2, %%mm5\n\t" | 208 "punpckldq %%mm2, %%mm5\n\t" |
214 "psrlq $32, %%mm2\n\t" | 209 "punpckhdq %%mm5, %%mm2\n\t" |
215 "punpckldq %%mm5,%%mm2\n\t" | |
216 #endif | 210 #endif |
217 "movq 24(%4), %%mm3\n\t" | 211 "movq 24(%4), %%mm3\n\t" |
218 "pfsub %%mm2, %%mm0\n\t" | 212 "pfsub %%mm2, %%mm0\n\t" |
219 "movq %%mm3, %%mm4\n\t" | 213 "movq %%mm3, %%mm4\n\t" |
220 "movq %%mm0, %%mm1\n\t" | 214 "movq %%mm0, %%mm1\n\t" |
361 "movq %%mm0, %%mm2\n\t" | 355 "movq %%mm0, %%mm2\n\t" |
362 "pfmul %%mm1, %%mm0\n\t" | 356 "pfmul %%mm1, %%mm0\n\t" |
363 #ifdef HAVE_3DNOWEX | 357 #ifdef HAVE_3DNOWEX |
364 "pswapd %%mm1, %%mm1\n\t" | 358 "pswapd %%mm1, %%mm1\n\t" |
365 #else | 359 #else |
366 "movq %%mm1, %%mm5\n\t" | 360 "punpckldq %%mm1, %%mm5\n\t" |
367 "psrlq $32, %%mm1\n\t" | 361 "punpckhdq %%mm5, %%mm1\n\t" |
368 "punpckldq %%mm5, %%mm1\n\t" | |
369 #endif | 362 #endif |
370 "pfmul %%mm1, %%mm2\n\t" | 363 "pfmul %%mm1, %%mm2\n\t" |
371 #ifdef HAVE_3DNOWEX | 364 #ifdef HAVE_3DNOWEX |
372 "pfpnacc %%mm2, %%mm0\n\t" | 365 "pfpnacc %%mm2, %%mm0\n\t" |
373 #else | 366 #else |
434 for (i=0; i < 128; i++) { | 427 for (i=0; i < 128; i++) { |
435 __asm__ __volatile__ ( | 428 __asm__ __volatile__ ( |
436 "movq %1, %%mm0\n\t" /* ac3_buf[i].re | ac3_buf[i].im */ | 429 "movq %1, %%mm0\n\t" /* ac3_buf[i].re | ac3_buf[i].im */ |
437 "movq %%mm0, %%mm1\n\t" /* ac3_buf[i].re | ac3_buf[i].im */ | 430 "movq %%mm0, %%mm1\n\t" /* ac3_buf[i].re | ac3_buf[i].im */ |
438 #ifndef HAVE_3DNOWEX | 431 #ifndef HAVE_3DNOWEX |
439 "movq %%mm1, %%mm2\n\t" | 432 "punpckldq %%mm1, %%mm2\n\t" |
440 "psrlq $32, %%mm1\n\t" | 433 "punpckhdq %%mm2, %%mm1\n\t" |
441 "punpckldq %%mm2, %%mm1\n\t" | |
442 #else | 434 #else |
443 "pswapd %%mm1, %%mm1\n\t" /* ac3_buf[i].re | ac3_buf[i].im */ | 435 "pswapd %%mm1, %%mm1\n\t" /* ac3_buf[i].re | ac3_buf[i].im */ |
444 #endif | 436 #endif |
445 "movd %3, %%mm3\n\t" /* ac3_xsin[i] */ | 437 "movd %3, %%mm3\n\t" /* ac3_xsin[i] */ |
446 "punpckldq %2, %%mm3\n\t" /* ac3_xsin[i] | ac3_xcos[i] */ | 438 "punpckldq %2, %%mm3\n\t" /* ac3_xsin[i] | ac3_xcos[i] */ |
447 "pfmul %%mm3, %%mm0\n\t" | 439 "pfmul %%mm3, %%mm0\n\t" |
448 "pfmul %%mm3, %%mm1\n\t" | 440 "pfmul %%mm3, %%mm1\n\t" |
449 #ifndef HAVE_3DNOWEX | 441 #ifndef HAVE_3DNOWEX |
450 "pxor %%mm7, %%mm0\n\t" | 442 "pxor %%mm7, %%mm0\n\t" |
451 "pfacc %%mm1, %%mm0\n\t" | 443 "pfacc %%mm1, %%mm0\n\t" |
452 "movd %%mm0, 4%0\n\t" | 444 "punpckldq %%mm0, %%mm1\n\t" |
453 "psrlq $32, %%mm0\n\t" | 445 "punpckhdq %%mm1, %%mm0\n\t" |
454 "movd %%mm0, %0\n\t" | 446 "movq %%mm0, %0\n\t" |
455 #else | 447 #else |
456 "pfpnacc %%mm1, %%mm0\n\t" /* mm0 = mm0[0] - mm0[1] | mm1[0] + mm1[1] */ | 448 "pfpnacc %%mm1, %%mm0\n\t" /* mm0 = mm0[0] - mm0[1] | mm1[0] + mm1[1] */ |
457 "pswapd %%mm0, %%mm0\n\t" | 449 "pswapd %%mm0, %%mm0\n\t" |
458 "movq %%mm0, %0" | 450 "movq %%mm0, %0" |
459 #endif | 451 #endif |
536 "punpckldq 508(%2), %%mm1\n\t" | 528 "punpckldq 508(%2), %%mm1\n\t" |
537 #ifdef HAVE_3DNOWEX | 529 #ifdef HAVE_3DNOWEX |
538 "pswapd (%3), %%mm3\n\t" | 530 "pswapd (%3), %%mm3\n\t" |
539 "pswapd -512(%3), %%mm4\n\t" | 531 "pswapd -512(%3), %%mm4\n\t" |
540 #else | 532 #else |
541 "movq (%3), %%mm3\n\t"/**delay_ptr++=-buf[64+i].re**--window_ptr;*/ | 533 "movq (%3), %%mm3\n\t" |
534 "punpckldq %%mm3, %%mm2\n\t" | |
535 "punpckhdq %%mm2, %%mm3\n\t" | |
542 "movq -512(%3), %%mm4\n\t" | 536 "movq -512(%3), %%mm4\n\t" |
543 "psrlq $32, %%mm3\n\t"/*delay_ptr[128]=buf[i].im**window_ptr[-512];*/ | 537 "punpckldq %%mm4, %%mm2\n\t" |
544 "psrlq $32, %%mm4\n\t"/**delay_ptr++=buf[64-i-1].im**--window_ptr;*/ | 538 "punpckhdq %%mm2, %%mm4\n\t" |
545 "punpckldq (%3), %%mm3\n\t"/*delay_ptr[128]=-buf[128-i-1].re**window_ptr[-512];*/ | |
546 "punpckldq -512(%3), %%mm4\n\t" | |
547 #endif | 539 #endif |
548 "pfmul %%mm3, %%mm0\n\t" | 540 "pfmul %%mm3, %%mm0\n\t" |
549 "pfmul %%mm4, %%mm1\n\t" | 541 "pfmul %%mm4, %%mm1\n\t" |
550 "pxor %%mm6, %%mm0\n\t" | 542 "pxor %%mm6, %%mm0\n\t" |
551 "pxor %%mm7, %%mm1\n\t" | 543 "pxor %%mm7, %%mm1\n\t" |