Home | History | Annotate | Download | only in arm64

Lines Matching full:part

187     smull       v20.4s, v2.4h, v0.h[0]      //// y0 * cos4(part of c0 and c1)
189 smull v18.4s, v3.4h, v1.h[2] //// y2 * sin2 (q3 is freed by this time)(part of d1)
192 smull v24.4s, v6.4h, v0.h[1] //// y1 * cos1(part of b0)
194 smull v26.4s, v6.4h, v0.h[3] //// y1 * cos3(part of b1)
196 smull v28.4s, v6.4h, v1.h[1] //// y1 * sin3(part of b2)
198 smull v30.4s, v6.4h, v1.h[3] //// y1 * sin1(part of b3)
200 smlal v24.4s, v7.4h, v0.h[3] //// y1 * cos1 + y3 * cos3(part of b0)
202 smlsl v26.4s, v7.4h, v1.h[3] //// y1 * cos3 - y3 * sin1(part of b1)
204 smlsl v28.4s, v7.4h, v0.h[1] //// y1 * sin3 - y3 * cos1(part of b2)
206 smlsl v30.4s, v7.4h, v1.h[1] //// y1 * sin1 - y3 * sin3(part of b3)
208 smull v22.4s, v10.4h, v0.h[0] //// y4 * cos4(part of c0 and c1)
210 smull v6.4s, v3.4h, v0.h[2] //// y2 * cos2(part of d0)
234 smlal v24.4s, v14.4h, v1.h[1] //// y1 * cos1 + y3 * cos3 + y5 * sin3(part of b0)
235 smlsl v26.4s, v14.4h, v0.h[1] //// y1 * cos3 - y3 * sin1 - y5 * cos1(part of b1)
236 smlal v28.4s, v14.4h, v1.h[3] //// y1 * sin3 - y3 * cos1 + y5 * sin1(part of b2)
237 smlal v30.4s, v14.4h, v0.h[3] //// y1 * sin1 - y3 * sin3 + y5 * cos3(part of b3)
239 smlsl v18.4s, v11.4h, v0.h[2] //// d1 = y2 * sin2 - y6 * cos2(part of a0 and a1)
240 smlal v6.4s, v11.4h, v1.h[2] //// d0 = y2 * cos2 + y6 * sin2(part of a0 and a1)
242 add v10.4s, v20.4s , v22.4s //// c0 = y0 * cos4 + y4 * cos4(part of a0 and a1)
243 sub v20.4s, v20.4s , v22.4s //// c1 = y0 * cos4 - y4 * cos4(part of a0 and a1)
245 smlal v24.4s, v15.4h, v1.h[3] //// b0 = y1 * cos1 + y3 * cos3 + y5 * sin3 + y7 * sin1(part of x0,x7)
246 smlsl v26.4s, v15.4h, v1.h[1] //// b1 = y1 * cos3 - y3 * sin1 - y5 * cos1 - y7 * sin3(part of x1,x6)
247 smlal v28.4s, v15.4h, v0.h[3] //// b2 = y1 * sin3 - y3 * cos1 + y5 * sin1 + y7 * cos3(part of x2,x5)
248 smlsl v30.4s, v15.4h, v0.h[1] //// b3 = y1 * sin1 - y3 * sin3 + y5 * cos3 - y7 * cos1(part of x3,x4)
250 add v14.4s, v10.4s , v6.4s //// a0 = c0 + d0(part of x0,x7)
251 sub v10.4s, v10.4s , v6.4s //// a3 = c0 - d0(part of x3,x4)
252 sub v22.4s, v20.4s , v18.4s //// a2 = c1 - d1(part of x2,x5)
253 add v18.4s, v20.4s , v18.4s //// a1 = c1 + d1(part of x1,x6)
255 add v20.4s, v14.4s , v24.4s //// a0 + b0(part of x0)
256 sub v6.4s, v14.4s , v24.4s //// a0 - b0(part of x7)
258 add v24.4s, v22.4s , v28.4s //// a2 + b2(part of x2)
259 sub v22.4s, v22.4s , v28.4s //// a2 - b2(part of x5)
261 add v28.4s, v18.4s , v26.4s //// a1 + b1(part of x1)
262 sub v18.4s, v18.4s , v26.4s //// a1 - b1(part of x6)
264 add v26.4s, v10.4s , v30.4s //// a3 + b3(part of x3)
265 sub v30.4s, v10.4s , v30.4s //// a3 - b3(part of x4)
304 smull v24.4s, v6.4h, v0.h[1] //// y1 * cos1(part of b0)
305 smull v26.4s, v6.4h, v0.h[3] //// y1 * cos3(part of b1)
306 smull v28.4s, v6.4h, v1.h[1] //// y1 * sin3(part of b2)
307 smull v30.4s, v6.4h, v1.h[3] //// y1 * sin1(part of b3)
309 smlal v24.4s, v7.4h, v0.h[3] //// y1 * cos1 + y3 * cos3(part of b0)
310 smlsl v26.4s, v7.4h, v1.h[3] //// y1 * cos3 - y3 * sin1(part of b1)
311 smlsl v28.4s, v7.4h, v0.h[1] //// y1 * sin3 - y3 * cos1(part of b2)
312 smlsl v30.4s, v7.4h, v1.h[1] //// y1 * sin1 - y3 * sin3(part of b3)
314 smull v18.4s, v3.4h, v1.h[2] //// y2 * sin2 (q3 is freed by this time)(part of d1)
315 smull v6.4s, v3.4h, v0.h[2] //// y2 * cos2(part of d0)
317 smull v20.4s, v2.4h, v0.h[0] //// y0 * cos4(part of c0 and c1)
320 add v14.4s, v20.4s , v6.4s //// a0 = c0 + d0(part of x0,x7)
321 sub v10.4s, v20.4s , v6.4s //// a3 = c0 - d0(part of x3,x4)
322 sub v22.4s, v20.4s , v18.4s //// a2 = c1 - d1(part of x2,x5)
323 add v18.4s, v20.4s , v18.4s //// a1 = c1 + d1(part of x1,x6)
325 add v20.4s, v14.4s , v24.4s //// a0 + b0(part of x0)
326 sub v6.4s, v14.4s , v24.4s //// a0 - b0(part of x7)
328 add v24.4s, v22.4s , v28.4s //// a2 + b2(part of x2)
329 sub v22.4s, v22.4s , v28.4s //// a2 - b2(part of x5)
331 add v28.4s, v18.4s , v26.4s //// a1 + b1(part of x1)
332 sub v18.4s, v18.4s , v26.4s //// a1 - b1(part of x6)
334 add v26.4s, v10.4s , v30.4s //// a3 + b3(part of x3)
335 sub v30.4s, v10.4s , v30.4s //// a3 - b3(part of x4)
353 smull v24.4s, v8.4h, v0.h[1] //// y1 * cos1(part of b0)
354 smull v26.4s, v8.4h, v0.h[3] //// y1 * cos3(part of b1)
355 smull v28.4s, v8.4h, v1.h[1] //// y1 * sin3(part of b2)
356 smull v30.4s, v8.4h, v1.h[3] //// y1 * sin1(part of b3)
358 smlal v24.4s, v9.4h, v0.h[3] //// y1 * cos1 + y3 * cos3(part of b0)
359 smlsl v26.4s, v9.4h, v1.h[3] //// y1 * cos3 - y3 * sin1(part of b1)
360 smlsl v28.4s, v9.4h, v0.h[1] //// y1 * sin3 - y3 * cos1(part of b2)
361 smlsl v30.4s, v9.4h, v1.h[1] //// y1 * sin1 - y3 * sin3(part of b3)
363 smull v18.4s, v5.4h, v1.h[2] //// y2 * sin2 (q4 is freed by this time)(part of d1)
364 smull v8.4s, v5.4h, v0.h[2] //// y2 * cos2(part of d0)
366 smull v20.4s, v4.4h, v0.h[0] //// y0 * cos4(part of c0 and c1)
367 smull v22.4s, v12.4h, v0.h[0] //// y4 * cos4(part of c0 and c1)
369 smlal v24.4s, v16.4h, v1.h[1] //// y1 * cos1 + y3 * cos3 + y5 * sin3(part of b0)
370 smlsl v26.4s, v16.4h, v0.h[1] //// y1 * cos3 - y3 * sin1 - y5 * cos1(part of b1)
371 smlal v28.4s, v16.4h, v1.h[3] //// y1 * sin3 - y3 * cos1 + y5 * sin1(part of b2)
372 smlal v30.4s, v16.4h, v0.h[3] //// y1 * sin1 - y3 * sin3 + y5 * cos3(part of b3)
374 smlsl v18.4s, v13.4h, v0.h[2] //// d1 = y2 * sin2 - y6 * cos2(part of a0 and a1)
375 smlal v8.4s, v13.4h, v1.h[2] //// d0 = y2 * cos2 + y6 * sin2(part of a0 and a1)
377 add v12.4s, v20.4s , v22.4s //// c0 = y0 * cos4 + y4 * cos4(part of a0 and a1)
378 sub v20.4s, v20.4s , v22.4s //// c1 = y0 * cos4 - y4 * cos4(part of a0 and a1)
380 smlal v24.4s, v17.4h, v1.h[3] //// b0 = y1 * cos1 + y3 * cos3 + y5 * sin3 + y7 * sin1(part of e0,e7)
381 smlsl v26.4s, v17.4h, v1.h[1] //// b1 = y1 * cos3 - y3 * sin1 - y5 * cos1 - y7 * sin3(part of e1,e6)
382 smlal v28.4s, v17.4h, v0.h[3] //// b2 = y1 * sin3 - y3 * cos1 + y5 * sin1 + y7 * cos3(part of e2,e5)
383 smlsl v30.4s, v17.4h, v0.h[1] //// b3 = y1 * sin1 - y3 * sin3 + y5 * cos3 - y7 * cos1(part of e3,e4)
385 add v16.4s, v12.4s , v8.4s //// a0 = c0 + d0(part of e0,e7)
386 sub v12.4s, v12.4s , v8.4s //// a3 = c0 - d0(part of e3,e4)
387 sub v22.4s, v20.4s , v18.4s //// a2 = c1 - d1(part of e2,e5)
388 add v18.4s, v20.4s , v18.4s //// a1 = c1 + d1(part of e1,e6)
390 add v20.4s, v16.4s , v24.4s //// a0 + b0(part of e0)
391 sub v8.4s, v16.4s , v24.4s //// a0 - b0(part of e7)
393 add v24.4s, v22.4s , v28.4s //// a2 + b2(part of e2)
394 sub v22.4s, v22.4s , v28.4s //// a2 - b2(part of e5)
396 add v28.4s, v18.4s , v26.4s //// a1 + b1(part of e1)
397 sub v18.4s, v18.4s , v26.4s //// a1 - b1(part of e6)
399 add v26.4s, v12.4s , v30.4s //// a3 + b3(part of e3)
400 part of x4)
443 smull v24.4s, v6.4h, v0.h[1] //// y1 * cos1(part of b0)
444 smull v26.4s, v6.4h, v0.h[3] //// y1 * cos3(part of b1)
445 smull v28.4s, v6.4h, v1.h[1] //// y1 * sin3(part of b2)
446 smull v30.4s, v6.4h, v1.h[3] //// y1 * sin1(part of b3)
448 smlal v24.4s, v7.4h, v0.h[3] //// y1 * cos1 + y3 * cos3(part of b0)
449 smlsl v26.4s, v7.4h, v1.h[3] //// y1 * cos3 - y3 * sin1(part of b1)
450 smlsl v28.4s, v7.4h, v0.h[1] //// y1 * sin3 - y3 * cos1(part of b2)
451 smlsl v30.4s, v7.4h, v1.h[1] //// y1 * sin1 - y3 * sin3(part of b3)
453 smull v20.4s, v2.4h, v0.h[0] //// y0 * cos4(part of c0 and c1)
454 // vmull.s16 q11,d4,d0[0] @// y4 * cos4(part of c0 and c1)
456 smull v18.4s, v3.4h, v1.h[2] //// y2 * sin2 (q3 is freed by this time)(part of d1)
457 smull v6.4s, v3.4h, v0.h[2] //// y2 * cos2(part of d0)
462 sub v22.4s, v20.4s , v6.4s //// a3 = c0 - d0(part of x3,x4)
463 add v4.4s, v20.4s , v6.4s //// a0 = c0 + d0(part of x0,x7)
479 sub v22.4s, v20.4s , v18.4s //// a2 = c1 - d1(part of x2,x5)
480 add v18.4s, v20.4s , v18.4s //// a1 = c1 + d1(part of x1,x6)
523 smull v24.4s, v14.4h, v0.h[1] //// y1 * cos1(part of b0)
525 smull v26.4s, v14.4h, v0.h[3] //// y1 * cos3(part of b1)
526 smull v28.4s, v14.4h, v1.h[1] //// y1 * sin3(part of b2)
527 smull v30.4s, v14.4h, v1.h[3] //// y1 * sin1(part of b3)
529 smlal v24.4s, v15.4h, v0.h[3] //// y1 * cos1 + y3 * cos3(part of b0)
530 smlsl v26.4s, v15.4h, v1.h[3] //// y1 * cos3 - y3 * sin1(part of b1)
531 smlsl v28.4s, v15.4h, v0.h[1] //// y1 * sin3 - y3 * cos1(part of b2)
532 smlsl v30.4s, v15.4h, v1.h[1] //// y1 * sin1 - y3 * sin3(part of b3)
533 smull v20.4s, v10.4h, v0.h[0] //// y0 * cos4(part of c0 and c1)
534 smull v18.4s, v11.4h, v1.h[2] //// y2 * sin2 (q7 is freed by this time)(part of d1)
535 smull v14.4s, v11.4h, v0.h[2] //// y2 * cos2(part of d0)
560 sub v22.4s, v20.4s , v14.4s //// a3 = c0 - d0(part of x3,x4)
561 add v12.4s, v20.4s , v14.4s //// a0 = c0 + d0(part of x0,x7)
580 sub v22.4s, v20.4s , v18.4s //// a2 = c1 - d1(part of x2,x5)
581 add v18.4s, v20.4s , v18.4s //// a1 = c1 + d1(part of x1,x6)
701 smull v24.4s, v6.4h, v0.h[1] //// y1 * cos1(part of b0)
702 smull v26.4s, v6.4h, v0.h[3] //// y1 * cos3(part of b1)
703 smull v28.4s, v6.4h, v1.h[1] //// y1 * sin3(part of b2)
704 smull v30.4s, v6.4h, v1.h[3] //// y1 * sin1(part of b3)
706 smlal v24.4s, v7.4h, v0.h[3] //// y1 * cos1 + y3 * cos3(part of b0)
707 smlsl v26.4s, v7.4h, v1.h[3] //// y1 * cos3 - y3 * sin1(part of b1)
708 smlsl v28.4s, v7.4h, v0.h[1] //// y1 * sin3 - y3 * cos1(part of b2)
709 smlsl v30.4s, v7.4h, v1.h[1] //// y1 * sin1 - y3 * sin3(part of b3)
711 smull v20.4s, v2.4h, v0.h[0] //// y0 * cos4(part of c0 and c1)
712 smull v22.4s, v4.4h, v0.h[0] //// y4 * cos4(part of c0 and c1)
714 smull v18.4s, v3.4h, v1.h[2] //// y2 * sin2 (q3 is freed by this time)(part of d1)
715 smull v6.4s, v3.4h, v0.h[2] //// y2 * cos2(part of d0)
718 smlal v24.4s, v8.4h, v1.h[1] //// y1 * cos1 + y3 * cos3 + y5 * sin3(part of b0)
719 smlsl v26.4s, v8.4h, v0.h[1] //// y1 * cos3 - y3 * sin1 - y5 * cos1(part of b1)
720 smlal v28.4s, v8.4h, v1.h[3] //// y1 * sin3 - y3 * cos1 + y5 * sin1(part of b2)
721 smlal v30.4s, v8.4h, v0.h[3] //// y1 * sin1 - y3 * sin3 + y5 * cos3(part of b3)
723 smlsl v18.4s, v5.4h, v0.h[2] //// d1 = y2 * sin2 - y6 * cos2(part of a0 and a1)
724 smlal v6.4s, v5.4h, v1.h[2] //// d0 = y2 * cos2 + y6 * sin2(part of a0 and a1)
726 add v2.4s, v20.4s , v22.4s //// c0 = y0 * cos4 + y4 * cos4(part of a0 and a1)
727 sub v20.4s, v20.4s , v22.4s //// c1 = y0 * cos4 - y4 * cos4(part of a0 and a1)
729 smlal v24.4s, v9.4h, v1.h[3] //// b0 = y1 * cos1 + y3 * cos3 + y5 * sin3 + y7 * sin1(part of x0,x7)
730 smlsl v26.4s, v9.4h, v1.h[1] //// b1 = y1 * cos3 - y3 * sin1 - y5 * cos1 - y7 * sin3(part of x1,x6)
731 smlal v28.4s, v9.4h, v0.h[3] //// b2 = y1 * sin3 - y3 * cos1 + y5 * sin1 + y7 * cos3(part of x2,x5)
732 smlsl v30.4s, v9.4h, v0.h[1] //// b3 = y1 * sin1 - y3 * sin3 + y5 * cos3 - y7 * cos1(part of x3,x4)
734 sub v22.4s, v2.4s , v6.4s //// a3 = c0 - d0(part of x3,x4)
735 add v4.4s, v2.4s , v6.4s //// a0 = c0 + d0(part of x0,x7)
751 sub v22.4s, v20.4s , v18.4s //// a2 = c1 - d1(part of x2,x5)
752 add v18.4s, v20.4s , v18.4s //// a1 = c1 + d1(part of x1,x6)
797 smull v24.4s, v14.4h, v0.h[1] //// y1 * cos1(part of b0)
798 smull v26.4s, v14.4h, v0.h[3] //// y1 * cos3(part of b1)
799 smull v28.4s, v14.4h, v1.h[1] //// y1 * sin3(part of b2)
800 smull v30.4s, v14.4h, v1.h[3] //// y1 * sin1(part of b3)
801 smlal v24.4s, v15.4h, v0.h[3] //// y1 * cos1 + y3 * cos3(part of b0)
802 smlsl v26.4s, v15.4h, v1.h[3] //// y1 * cos3 - y3 * sin1(part of b1)
803 smlsl v28.4s, v15.4h, v0.h[1] //// y1 * sin3 - y3 * cos1(part of b2)
804 smlsl v30.4s, v15.4h, v1.h[1] //// y1 * sin1 - y3 * sin3(part of b3)
805 smull v20.4s, v10.4h, v0.h[0] //// y0 * cos4(part of c0 and c1)
806 smull v22.4s, v12.4h, v0.h[0] //// y4 * cos4(part of c0 and c1)
807 smull v18.4s, v11.4h, v1.h[2] //// y2 * sin2 (q7 is freed by this time)(part of d1)
808 smull v14.4s, v11.4h, v0.h[2] //// y2 * cos2(part of d0)
809 smlal v24.4s, v16.4h, v1.h[1] //// y1 * cos1 + y3 * cos3 + y5 * sin3(part of b0)
812 part of b1)
815 smlal v28.4s, v16.4h, v1.h[3] //// y1 * sin3 - y3 * cos1 + y5 * sin1(part of b2)
818 smlal v30.4s, v16.4h, v0.h[3] //// y1 * sin1 - y3 * sin3 + y5 * cos3(part of b3)
821 smlsl v18.4s, v13.4h, v0.h[2] //// d1 = y2 * sin2 - y6 * cos2(part of a0 and a1)
824 smlal v14.4s, v13.4h, v1.h[2] //// d0 = y2 * cos2 + y6 * sin2(part of a0 and a1)
826 add v12.4s, v20.4s , v22.4s //// c0 = y0 * cos4 + y4 * cos4(part of a0 and a1)
827 sub v20.4s, v20.4s , v22.4s //// c1 = y0 * cos4 - y4 * cos4(part of a0 and a1)
829 smlal v24.4s, v17.4h, v1.h[3] //// b0 = y1 * cos1 + y3 * cos3 + y5 * sin3 + y7 * sin1(part of x0,x7)
836 smlsl v26.4s, v17.4h, v1.h[1] //// b1 = y1 * cos3 - y3 * sin1 - y5 * cos1 - y7 * sin3(part of x1,x6)
842 smlal v28.4s, v17.4h, v0.h[3] //// b2 = y1 * sin3 - y3 * cos1 + y5 * sin1 + y7 * cos3(part of x2,x5)
843 smlsl v30.4s, v17.4h, v0.h[1] //// b3 = y1 * sin1 - y3 * sin3 + y5 * cos3 - y7 * cos1(part of x3,x4)
845 sub v22.4s, v12.4s , v14.4s //// a3 = c0 - d0(part of x3,x4)
846 add v12.4s, v12.4s , v14.4s //// a0 = c0 + d0(part of x0,x7)
865 sub v22.4s, v20.4s , v18.4s //// a2 = c1 - d1(part of x2,x5)
866 add v18.4s, v20.4s , v18.4s //// a1 = c1 + d1(part of x1,x6)