Lines Matching refs:s1

39   int32x4x2_t in[8], s1[32], s2[32], s3[32];
69 s1[16] = multiply_shift_and_narrow_s32_dual(in[1], cospi_31_64);
71 s1[31] = multiply_shift_and_narrow_s32_dual(in[1], cospi_1_64);
73 s1[20] = multiply_shift_and_narrow_s32_dual(in[5], cospi_27_64);
74 s1[27] = multiply_shift_and_narrow_s32_dual(in[5], cospi_5_64);
76 s1[23] = multiply_shift_and_narrow_s32_dual(in[3], -cospi_29_64);
77 s1[24] = multiply_shift_and_narrow_s32_dual(in[3], cospi_3_64);
84 s1[4] = multiply_shift_and_narrow_s32_dual(in[4], cospi_28_64);
85 s1[7] = multiply_shift_and_narrow_s32_dual(in[4], cospi_4_64);
87 s1[17] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], -cospi_4_64,
88 s1[31], cospi_28_64);
89 s1[30] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], cospi_28_64,
90 s1[31], cospi_4_64);
92 s1[21] = multiply_accumulate_shift_and_narrow_s32_dual(s1[20], -cospi_20_64,
93 s1[27], cospi_12_64);
94 s1[26] = multiply_accumulate_shift_and_narrow_s32_dual(s1[20], cospi_12_64,
95 s1[27], cospi_20_64);
97 s1[22] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_12_64,
98 s1[24], -cospi_20_64);
99 s1[25] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_20_64,
100 s1[24], cospi_12_64);
103 s1[0] = multiply_shift_and_narrow_s32_dual(in[0], cospi_16_64);
110 s2[20] = highbd_idct_sub_dual(s1[23], s1[20]);
111 s2[21] = highbd_idct_sub_dual(s1[22], s1[21]);
112 s2[22] = highbd_idct_add_dual(s1[21], s1[22]);
113 s2[23] = highbd_idct_add_dual(s1[20], s1[23]);
114 s2[24] = highbd_idct_add_dual(s1[24], s1[27]);
115 s2[25] = highbd_idct_add_dual(s1[25], s1[26]);
116 s2[26] = highbd_idct_sub_dual(s1[25], s1[26]);
117 s2[27] = highbd_idct_sub_dual(s1[24], s1[27]);
120 s1[5] = sub_multiply_shift_and_narrow_s32_dual(s1[7], s1[4], cospi_16_64);
121 s1[6] = add_multiply_shift_and_narrow_s32_dual(s1[4], s1[7], cospi_16_64);
123 s1[18] = multiply_accumulate_shift_and_narrow_s32_dual(s1[17], -cospi_8_64,
124 s1[30], cospi_24_64);
125 s1[29] = multiply_accumulate_shift_and_narrow_s32_dual(s1[17], cospi_24_64,
126 s1[30], cospi_8_64);
128 s1[19] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], -cospi_8_64,
129 s1[31], cospi_24_64);
130 s1[28] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], cospi_24_64,
131 s1[31], cospi_8_64);
133 s1[20] = multiply_accumulate_shift_and_narrow_s32_dual(s2[20], -cospi_24_64,
135 s1[27] = multiply_accumulate_shift_and_narrow_s32_dual(s2[20], -cospi_8_64,
138 s1[21] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_24_64,
140 s1[26] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_8_64,
144 s2[0] = highbd_idct_add_dual(s1[0], s1[7]);
145 s2[1] = highbd_idct_add_dual(s1[0], s1[6]);
146 s2[2] = highbd_idct_add_dual(s1[0], s1[5]);
147 s2[3] = highbd_idct_add_dual(s1[0], s1[4]);
148 s2[4] = highbd_idct_sub_dual(s1[0], s1[4]);
149 s2[5] = highbd_idct_sub_dual(s1[0], s1[5]);
150 s2[6] = highbd_idct_sub_dual(s1[0], s1[6]);
151 s2[7] = highbd_idct_sub_dual(s1[0], s1[7]);
159 s2[16] = highbd_idct_add_dual(s1[16], s2[23]);
160 s2[17] = highbd_idct_add_dual(s1[17], s2[22]);
161 s2[18] = highbd_idct_add_dual(s1[18], s1[21]);
162 s2[19] = highbd_idct_add_dual(s1[19], s1[20]);
163 s2[20] = highbd_idct_sub_dual(s1[19], s1[20]);
164 s2[21] = highbd_idct_sub_dual(s1[18], s1[21]);
165 s2[22] = highbd_idct_sub_dual(s1[17], s2[22]);
166 s2[23] = highbd_idct_sub_dual(s1[16], s2[23]);
168 s3[24] = highbd_idct_sub_dual(s1[31], s2[24]);
169 s3[25] = highbd_idct_sub_dual(s1[30], s2[25]);
170 s3[26] = highbd_idct_sub_dual(s1[29], s1[26]);
171 s3[27] = highbd_idct_sub_dual(s1[28], s1[27]);
172 s2[28] = highbd_idct_add_dual(s1[27], s1[28]);
173 s2[29] = highbd_idct_add_dual(s1[26], s1[29]);
174 s2[30] = highbd_idct_add_dual(s2[25], s1[30]);
175 s2[31] = highbd_idct_add_dual(s2[24], s1[31]);
178 s1[0] = highbd_idct_add_dual(s2[0], s2[15]);
179 s1[1] = highbd_idct_add_dual(s2[1], s2[14]);
180 s1[2] = highbd_idct_add_dual(s2[2], s2[13]);
181 s1[3] = highbd_idct_add_dual(s2[3], s2[12]);
182 s1[4] = highbd_idct_add_dual(s2[4], s2[11]);
183 s1[5] = highbd_idct_add_dual(s2[5], s2[10]);
184 s1[6] = highbd_idct_add_dual(s2[6], s2[9]);
185 s1[7] = highbd_idct_add_dual(s2[7], s2[8]);
186 s1[8] = highbd_idct_sub_dual(s2[7], s2[8]);
187 s1[9] = highbd_idct_sub_dual(s2[6], s2[9]);
188 s1[10] = highbd_idct_sub_dual(s2[5], s2[10]);
189 s1[11] = highbd_idct_sub_dual(s2[4], s2[11]);
190 s1[12] = highbd_idct_sub_dual(s2[3], s2[12]);
191 s1[13] = highbd_idct_sub_dual(s2[2], s2[13]);
192 s1[14] = highbd_idct_sub_dual(s2[1], s2[14]);
193 s1[15] = highbd_idct_sub_dual(s2[0], s2[15]);
195 s1[20] = sub_multiply_shift_and_narrow_s32_dual(s3[27], s2[20], cospi_16_64);
196 s1[27] = add_multiply_shift_and_narrow_s32_dual(s2[20], s3[27], cospi_16_64);
198 s1[21] = sub_multiply_shift_and_narrow_s32_dual(s3[26], s2[21], cospi_16_64);
199 s1[26] = add_multiply_shift_and_narrow_s32_dual(s2[21], s3[26], cospi_16_64);
201 s1[22] = sub_multiply_shift_and_narrow_s32_dual(s3[25], s2[22], cospi_16_64);
202 s1[25] = add_multiply_shift_and_narrow_s32_dual(s2[22], s3[25], cospi_16_64);
204 s1[23] = sub_multiply_shift_and_narrow_s32_dual(s3[24], s2[23], cospi_16_64);
205 s1[24] = add_multiply_shift_and_narrow_s32_dual(s2[23], s3[24], cospi_16_64);
208 s3[0] = highbd_idct_add_dual(s1[0], s2[31]);
209 s3[1] = highbd_idct_add_dual(s1[1], s2[30]);
210 s3[2] = highbd_idct_add_dual(s1[2], s2[29]);
211 s3[3] = highbd_idct_add_dual(s1[3], s2[28]);
212 s3[4] = highbd_idct_add_dual(s1[4], s1[27]);
213 s3[5] = highbd_idct_add_dual(s1[5], s1[26]);
214 s3[6] = highbd_idct_add_dual(s1[6], s1[25]);
215 s3[7] = highbd_idct_add_dual(s1[7], s1[24]);
216 s3[8] = highbd_idct_add_dual(s1[8], s1[23]);
217 s3[9] = highbd_idct_add_dual(s1[9], s1[22]);
218 s3[10] = highbd_idct_add_dual(s1[10], s1[21]);
219 s3[11] = highbd_idct_add_dual(s1[11], s1[20]);
220 s3[12] = highbd_idct_add_dual(s1[12], s2[19]);
221 s3[13] = highbd_idct_add_dual(s1[13], s2[18]);
222 s3[14] = highbd_idct_add_dual(s1[14], s2[17]);
223 s3[15] = highbd_idct_add_dual(s1[15], s2[16]);
224 s3[16] = highbd_idct_sub_dual(s1[15], s2[16]);
225 s3[17] = highbd_idct_sub_dual(s1[14], s2[17]);
226 s3[18] = highbd_idct_sub_dual(s1[13], s2[18]);
227 s3[19] = highbd_idct_sub_dual(s1[12], s2[19]);
228 s3[20] = highbd_idct_sub_dual(s1[11], s1[20]);
229 s3[21] = highbd_idct_sub_dual(s1[10], s1[21]);
230 s3[22] = highbd_idct_sub_dual(s1[9], s1[22]);
231 s3[23] = highbd_idct_sub_dual(s1[8], s1[23]);
232 s3[24] = highbd_idct_sub_dual(s1[7], s1[24]);
233 s3[25] = highbd_idct_sub_dual(s1[6], s1[25]);
234 s3[26] = highbd_idct_sub_dual(s1[5], s1[26]);
235 s3[27] = highbd_idct_sub_dual(s1[4], s1[27]);
236 s3[28] = highbd_idct_sub_dual(s1[3], s2[28]);
237 s3[29] = highbd_idct_sub_dual(s1[2], s2[29]);
238 s3[30] = highbd_idct_sub_dual(s1[1], s2[30]);
239 s3[31] = highbd_idct_sub_dual(s1[0], s2[31]);
375 int32x4x2_t in[8], s1[32], s2[32], s3[32], out[32];
381 s1[16] = multiply_shift_and_narrow_s32_dual(in[1], cospi_31_64);
382 s1[31] = multiply_shift_and_narrow_s32_dual(in[1], cospi_1_64);
385 s1[19] = multiply_shift_and_narrow_s32_dual(in[7], -cospi_25_64);
386 s1[28] = multiply_shift_and_narrow_s32_dual(in[7], cospi_7_64);
388 s1[20] = multiply_shift_and_narrow_s32_dual(in[5], cospi_27_64);
389 s1[27] = multiply_shift_and_narrow_s32_dual(in[5], cospi_5_64);
391 s1[23] = multiply_shift_and_narrow_s32_dual(in[3], -cospi_29_64);
392 s1[24] = multiply_shift_and_narrow_s32_dual(in[3], cospi_3_64);
402 s1[4] = multiply_shift_and_narrow_s32_dual(in[4], cospi_28_64);
403 s1[7] = multiply_shift_and_narrow_s32_dual(in[4], cospi_4_64);
405 s1[17] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], -cospi_4_64,
406 s1[31], cospi_28_64);
407 s1[30] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], cospi_28_64,
408 s1[31], cospi_4_64);
411 s1[18] = multiply_accumulate_shift_and_narrow_s32_dual(s1[19], -cospi_28_64,
412 s1[28], -cospi_4_64);
413 s1[29] = multiply_accumulate_shift_and_narrow_s32_dual(s1[19], -cospi_4_64,
414 s1[28], cospi_28_64);
416 s1[21] = multiply_accumulate_shift_and_narrow_s32_dual(s1[20], -cospi_20_64,
417 s1[27], cospi_12_64);
418 s1[26] = multiply_accumulate_shift_and_narrow_s32_dual(s1[20], cospi_12_64,
419 s1[27], cospi_20_64);
421 s1[22] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_12_64,
422 s1[24], -cospi_20_64);
423 s1[25] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_20_64,
424 s1[24], cospi_12_64);
427 s1[0] = multiply_shift_and_narrow_s32_dual(in[0], cospi_16_64);
439 s2[16] = highbd_idct_add_dual(s1[16], s1[19]);
441 s2[17] = highbd_idct_add_dual(s1[17], s1[18]);
442 s2[18] = highbd_idct_sub_dual(s1[17], s1[18]);
444 s2[19] = highbd_idct_sub_dual(s1[16], s1[19]);
446 s2[20] = highbd_idct_sub_dual(s1[23], s1[20]);
447 s2[21] = highbd_idct_sub_dual(s1[22], s1[21]);
449 s2[22] = highbd_idct_add_dual(s1[21], s1[22]);
450 s2[23] = highbd_idct_add_dual(s1[20], s1[23]);
452 s2[24] = highbd_idct_add_dual(s1[24], s1[27]);
453 s2[25] = highbd_idct_add_dual(s1[25], s1[26]);
454 s2[26] = highbd_idct_sub_dual(s1[25], s1[26]);
455 s2[27] = highbd_idct_sub_dual(s1[24], s1[27]);
457 s2[28] = highbd_idct_sub_dual(s1[31], s1[28]);
458 s2[29] = highbd_idct_sub_dual(s1[30], s1[29]);
459 s2[30] = highbd_idct_add_dual(s1[29], s1[30]);
460 s2[31] = highbd_idct_add_dual(s1[28], s1[31]);
463 s1[5] = sub_multiply_shift_and_narrow_s32_dual(s1[7], s1[4], cospi_16_64);
464 s1[6] = add_multiply_shift_and_narrow_s32_dual(s1[4], s1[7], cospi_16_64);
466 s1[8] = highbd_idct_add_dual(s2[8], s2[11]);
467 s1[9] = highbd_idct_add_dual(s2[9], s2[10]);
468 s1[10] = highbd_idct_sub_dual(s2[9], s2[10]);
469 s1[11] = highbd_idct_sub_dual(s2[8], s2[11]);
470 s1[12] = highbd_idct_sub_dual(s2[15], s2[12]);
471 s1[13] = highbd_idct_sub_dual(s2[14], s2[13]);
472 s1[14] = highbd_idct_add_dual(s2[13], s2[14]);
473 s1[15] = highbd_idct_add_dual(s2[12], s2[15]);
475 s1[18] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_8_64,
477 s1[29] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], cospi_24_64,
480 s1[19] = multiply_accumulate_shift_and_narrow_s32_dual(s2[19], -cospi_8_64,
482 s1[28] = multiply_accumulate_shift_and_narrow_s32_dual(s2[19], cospi_24_64,
485 s1[20] = multiply_accumulate_shift_and_narrow_s32_dual(s2[20], -cospi_24_64,
487 s1[27] = multiply_accumulate_shift_and_narrow_s32_dual(s2[20], -cospi_8_64,
490 s1[21] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_24_64,
492 s1[26] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_8_64,
496 s2[0] = highbd_idct_add_dual(s1[0], s1[7]);
497 s2[1] = highbd_idct_add_dual(s1[0], s1[6]);
498 s2[2] = highbd_idct_add_dual(s1[0], s1[5]);
499 s2[3] = highbd_idct_add_dual(s1[0], s1[4]);
500 s2[4] = highbd_idct_sub_dual(s1[0], s1[4]);
501 s2[5] = highbd_idct_sub_dual(s1[0], s1[5]);
502 s2[6] = highbd_idct_sub_dual(s1[0], s1[6]);
503 s2[7] = highbd_idct_sub_dual(s1[0], s1[7]);
505 s2[10] = sub_multiply_shift_and_narrow_s32_dual(s1[13], s1[10], cospi_16_64);
506 s2[13] = add_multiply_shift_and_narrow_s32_dual(s1[10], s1[13], cospi_16_64);
508 s2[11] = sub_multiply_shift_and_narrow_s32_dual(s1[12], s1[11], cospi_16_64);
509 s2[12] = add_multiply_shift_and_narrow_s32_dual(s1[11], s1[12], cospi_16_64);
511 s1[16] = highbd_idct_add_dual(s2[16], s2[23]);
512 s1[17] = highbd_idct_add_dual(s2[17], s2[22]);
513 s2[18] = highbd_idct_add_dual(s1[18], s1[21]);
514 s2[19] = highbd_idct_add_dual(s1[19], s1[20]);
515 s2[20] = highbd_idct_sub_dual(s1[19], s1[20]);
516 s2[21] = highbd_idct_sub_dual(s1[18], s1[21]);
517 s1[22] = highbd_idct_sub_dual(s2[17], s2[22]);
518 s1[23] = highbd_idct_sub_dual(s2[16], s2[23]);
522 s3[26] = highbd_idct_sub_dual(s1[29], s1[26]);
523 s3[27] = highbd_idct_sub_dual(s1[28], s1[27]);
524 s2[28] = highbd_idct_add_dual(s1[27], s1[28]);
525 s2[29] = highbd_idct_add_dual(s1[26], s1[29]);
530 s1[0] = highbd_idct_add_dual(s2[0], s1[15]);
531 s1[1] = highbd_idct_add_dual(s2[1], s1[14]);
532 s1[2] = highbd_idct_add_dual(s2[2], s2[13]);
533 s1[3] = highbd_idct_add_dual(s2[3], s2[12]);
534 s1[4] = highbd_idct_add_dual(s2[4], s2[11]);
535 s1[5] = highbd_idct_add_dual(s2[5], s2[10]);
536 s1[6] = highbd_idct_add_dual(s2[6], s1[9]);
537 s1[7] = highbd_idct_add_dual(s2[7], s1[8]);
538 s1[8] = highbd_idct_sub_dual(s2[7], s1[8]);
539 s1[9] = highbd_idct_sub_dual(s2[6], s1[9]);
540 s1[10] = highbd_idct_sub_dual(s2[5], s2[10]);
541 s1[11] = highbd_idct_sub_dual(s2[4], s2[11]);
542 s1[12] = highbd_idct_sub_dual(s2[3], s2[12]);
543 s1[13] = highbd_idct_sub_dual(s2[2], s2[13]);
544 s1[14] = highbd_idct_sub_dual(s2[1], s1[14]);
545 s1[15] = highbd_idct_sub_dual(s2[0], s1[15]);
547 s1[20] = sub_multiply_shift_and_narrow_s32_dual(s3[27], s2[20], cospi_16_64);
548 s1[27] = add_multiply_shift_and_narrow_s32_dual(s2[20], s3[27], cospi_16_64);
550 s1[21] = sub_multiply_shift_and_narrow_s32_dual(s3[26], s2[21], cospi_16_64);
551 s1[26] = add_multiply_shift_and_narrow_s32_dual(s2[21], s3[26], cospi_16_64);
553 s2[22] = sub_multiply_shift_and_narrow_s32_dual(s3[25], s1[22], cospi_16_64);
554 s1[25] = add_multiply_shift_and_narrow_s32_dual(s1[22], s3[25], cospi_16_64);
556 s2[23] = sub_multiply_shift_and_narrow_s32_dual(s3[24], s1[23], cospi_16_64);
557 s1[24] = add_multiply_shift_and_narrow_s32_dual(s1[23], s3[24], cospi_16_64);
560 out[0] = highbd_idct_add_dual(s1[0], s2[31]);
561 out[1] = highbd_idct_add_dual(s1[1], s2[30]);
562 out[2] = highbd_idct_add_dual(s1[2], s2[29]);
563 out[3] = highbd_idct_add_dual(s1[3], s2[28]);
564 out[4] = highbd_idct_add_dual(s1[4], s1[27]);
565 out[5] = highbd_idct_add_dual(s1[5], s1[26]);
566 out[6] = highbd_idct_add_dual(s1[6], s1[25]);
567 out[7] = highbd_idct_add_dual(s1[7], s1[24]);
568 out[8] = highbd_idct_add_dual(s1[8], s2[23]);
569 out[9] = highbd_idct_add_dual(s1[9], s2[22]);
570 out[10] = highbd_idct_add_dual(s1[10], s1[21]);
571 out[11] = highbd_idct_add_dual(s1[11], s1[20]);
572 out[12] = highbd_idct_add_dual(s1[12], s2[19]);
573 out[13] = highbd_idct_add_dual(s1[13], s2[18]);
574 out[14] = highbd_idct_add_dual(s1[14], s1[17]);
575 out[15] = highbd_idct_add_dual(s1[15], s1[16]);
576 out[16] = highbd_idct_sub_dual(s1[15], s1[16]);
577 out[17] = highbd_idct_sub_dual(s1[14], s1[17]);
578 out[18] = highbd_idct_sub_dual(s1[13], s2[18]);
579 out[19] = highbd_idct_sub_dual(s1[12], s2[19]);
580 out[20] = highbd_idct_sub_dual(s1[11], s1[20]);
581 out[21] = highbd_idct_sub_dual(s1[10], s1[21]);
582 out[22] = highbd_idct_sub_dual(s1[9], s2[22]);
583 out[23] = highbd_idct_sub_dual(s1[8], s2[23]);
584 out[24] = highbd_idct_sub_dual(s1[7], s1[24]);
585 out[25] = highbd_idct_sub_dual(s1[6], s1[25]);
586 out[26] = highbd_idct_sub_dual(s1[5], s1[26]);
587 out[27] = highbd_idct_sub_dual(s1[4], s1[27]);
588 out[28] = highbd_idct_sub_dual(s1[3], s2[28]);
589 out[29] = highbd_idct_sub_dual(s1[2], s2[29]);
590 out[30] = highbd_idct_sub_dual(s1[1], s2[30]);
591 out[31] = highbd_idct_sub_dual(s1[0], s2[31]);