comparison liba52/resample_c.c @ 3653:b11b15df02ed

3F2R sse optimized
author michael
date Sat, 22 Dec 2001 00:33:52 +0000
parents e22ff7ebdc05
children 33c3cff374a1
comparison
equal deleted inserted replaced
3652:65fd971932dc 3653:b11b15df02ed
280 "emms \n\t" 280 "emms \n\t"
281 :: "r" (s16+1024), "r" (f+256) 281 :: "r" (s16+1024), "r" (f+256)
282 :"%esi", "memory" 282 :"%esi", "memory"
283 ); 283 );
284 break; 284 break;
285 case A52_3F2R: //FIXME optimitze 285 case A52_3F2R:
286 for (i = 0; i < 256; i++) { 286 asm volatile(
287 s16[5*i] = convert (f[i]); 287 "movl $-1024, %%esi \n\t"
288 s16[5*i+1] = convert (f[i+512]); 288 "movq magicF2W, %%mm7 \n\t"
289 s16[5*i+2] = convert (f[i+768]); 289 "1: \n\t"
290 s16[5*i+3] = convert (f[i+1024]); 290 "movd (%1, %%esi), %%mm0 \n\t"
291 s16[5*i+4] = convert (f[i+256]); 291 "punpckldq 2048(%1, %%esi), %%mm0\n\t"
292 } 292 "movd 3072(%1, %%esi), %%mm1 \n\t"
293 "punpckldq 4096(%1, %%esi), %%mm1\n\t"
294 "movd 1024(%1, %%esi), %%mm2 \n\t"
295 "punpckldq 4(%1, %%esi), %%mm2 \n\t"
296 "movd 2052(%1, %%esi), %%mm3 \n\t"
297 "punpckldq 3076(%1, %%esi), %%mm3\n\t"
298 "movd 4100(%1, %%esi), %%mm4 \n\t"
299 "punpckldq 1028(%1, %%esi), %%mm4\n\t"
300 "movd 8(%1, %%esi), %%mm5 \n\t"
301 "punpckldq 2056(%1, %%esi), %%mm5\n\t"
302 "leal (%%esi, %%esi, 4), %%edi \n\t"
303 "sarl $1, %%edi \n\t"
304 "psubd %%mm7, %%mm0 \n\t"
305 "psubd %%mm7, %%mm1 \n\t"
306 "psubd %%mm7, %%mm2 \n\t"
307 "psubd %%mm7, %%mm3 \n\t"
308 "psubd %%mm7, %%mm4 \n\t"
309 "psubd %%mm7, %%mm5 \n\t"
310 "packssdw %%mm1, %%mm0 \n\t"
311 "packssdw %%mm3, %%mm2 \n\t"
312 "packssdw %%mm5, %%mm4 \n\t"
313 "movq %%mm0, (%0, %%edi) \n\t"
314 "movq %%mm2, 8(%0, %%edi) \n\t"
315 "movq %%mm4, 16(%0, %%edi) \n\t"
316
317 "movd 3080(%1, %%esi), %%mm0 \n\t"
318 "punpckldq 4104(%1, %%esi), %%mm0\n\t"
319 "movd 1032(%1, %%esi), %%mm1 \n\t"
320 "punpckldq 12(%1, %%esi), %%mm1\n\t"
321 "movd 2060(%1, %%esi), %%mm2 \n\t"
322 "punpckldq 3084(%1, %%esi), %%mm2\n\t"
323 "movd 4108(%1, %%esi), %%mm3 \n\t"
324 "punpckldq 1036(%1, %%esi), %%mm3\n\t"
325 "psubd %%mm7, %%mm0 \n\t"
326 "psubd %%mm7, %%mm1 \n\t"
327 "psubd %%mm7, %%mm2 \n\t"
328 "psubd %%mm7, %%mm3 \n\t"
329 "packssdw %%mm1, %%mm0 \n\t"
330 "packssdw %%mm3, %%mm2 \n\t"
331 "packssdw %%mm5, %%mm4 \n\t"
332 "movq %%mm0, 24(%0, %%edi) \n\t"
333 "movq %%mm2, 32(%0, %%edi) \n\t"
334 "movq %%mm4, 40(%0, %%edi) \n\t"
335
336 "addl $16, %%esi \n\t"
337 " jnz 1b \n\t"
338 "emms \n\t"
339 :: "r" (s16+1280), "r" (f+256)
340 :"%esi", "%edi", "memory"
341 );
293 break; 342 break;
294 case A52_MONO | A52_LFE: 343 case A52_MONO | A52_LFE:
295 asm volatile( 344 asm volatile(
296 "movl $-1024, %%esi \n\t" 345 "movl $-1024, %%esi \n\t"
297 "movq magicF2W, %%mm7 \n\t" 346 "movq magicF2W, %%mm7 \n\t"