Lines Matching refs:val
27 v128 val[2]; member
31 SIMD_INLINE uint32_t v256_low_u32(v256 a) { return v128_low_u32(a.val[0]); } in v256_low_u32()
33 SIMD_INLINE v64 v256_low_v64(v256 a) { return v128_low_v64(a.val[0]); } in v256_low_v64()
37 SIMD_INLINE v128 v256_low_v128(v256 a) { return a.val[0]; } in v256_low_v128()
39 SIMD_INLINE v128 v256_high_v128(v256 a) { return a.val[1]; } in v256_high_v128()
43 t.val[1] = hi; in v256_from_v128()
44 t.val[0] = lo; in v256_from_v128()
67 v128_store_unaligned(p, a.val[0]); in v256_store_unaligned()
68 v128_store_unaligned((uint8_t *)p + 16, a.val[1]); in v256_store_unaligned()
72 v128_store_aligned(p, a.val[0]); in v256_store_aligned()
73 v128_store_aligned((uint8_t *)p + 16, a.val[1]); in v256_store_aligned()
101 return v128_dotp_su8(a.val[1], b.val[1]) + v128_dotp_su8(a.val[0], b.val[0]); in v256_dotp_su8()
105 return v128_dotp_s16(a.val[1], b.val[1]) + v128_dotp_s16(a.val[0], b.val[0]); in v256_dotp_s16()
109 return v128_dotp_s32(a.val[1], b.val[1]) + v128_dotp_s32(a.val[0], b.val[0]); in v256_dotp_s32()
113 return v128_hadd_u8(a.val[1]) + v128_hadd_u8(a.val[0]); in v256_hadd_u8()
117 sad128_internal val[2]; member
122 t.val[1] = v128_sad_u8_init(); in v256_sad_u8_init()
123 t.val[0] = v128_sad_u8_init(); in v256_sad_u8_init()
132 t.val[1] = v128_sad_u8(s.val[1], a.val[1], b.val[1]); in v256_sad_u8()
133 t.val[0] = v128_sad_u8(s.val[0], a.val[0], b.val[0]); in v256_sad_u8()
138 return v128_sad_u8_sum(s.val[1]) + v128_sad_u8_sum(s.val[0]); in v256_sad_u8_sum()
142 ssd128_internal val[2]; member
147 t.val[1] = v128_ssd_u8_init(); in v256_ssd_u8_init()
148 t.val[0] = v128_ssd_u8_init(); in v256_ssd_u8_init()
156 t.val[1] = v128_ssd_u8(s.val[1], a.val[1], b.val[1]); in v256_ssd_u8()
157 t.val[0] = v128_ssd_u8(s.val[0], a.val[0], b.val[0]); in v256_ssd_u8()
162 return v128_ssd_u8_sum(s.val[1]) + v128_ssd_u8_sum(s.val[0]); in v256_ssd_u8_sum()
166 return v256_from_v128(v128_or(a.val[1], b.val[1]), in v256_or()
167 v128_or(a.val[0], b.val[0])); in v256_or()
171 return v256_from_v128(v128_xor(a.val[1], b.val[1]), in v256_xor()
172 v128_xor(a.val[0], b.val[0])); in v256_xor()
176 return v256_from_v128(v128_and(a.val[1], b.val[1]), in v256_and()
177 v128_and(a.val[0], b.val[0])); in v256_and()
181 return v256_from_v128(v128_andn(a.val[1], b.val[1]), in v256_andn()
182 v128_andn(a.val[0], b.val[0])); in v256_andn()
186 return v256_from_v128(v128_add_8(a.val[1], b.val[1]), in v256_add_8()
187 v128_add_8(a.val[0], b.val[0])); in v256_add_8()
191 return v256_from_v128(v128_add_16(a.val[1], b.val[1]), in v256_add_16()
192 v128_add_16(a.val[0], b.val[0])); in v256_add_16()
196 return v256_from_v128(v128_sadd_s8(a.val[1], b.val[1]), in v256_sadd_s8()
197 v128_sadd_s8(a.val[0], b.val[0])); in v256_sadd_s8()
201 return v256_from_v128(v128_sadd_u8(a.val[1], b.val[1]), in v256_sadd_u8()
202 v128_sadd_u8(a.val[0], b.val[0])); in v256_sadd_u8()
206 return v256_from_v128(v128_sadd_s16(a.val[1], b.val[1]), in v256_sadd_s16()
207 v128_sadd_s16(a.val[0], b.val[0])); in v256_sadd_s16()
211 return v256_from_v128(v128_add_32(a.val[1], b.val[1]), in v256_add_32()
212 v128_add_32(a.val[0], b.val[0])); in v256_add_32()
216 return v256_from_v128(v128_add_64(a.val[1], b.val[1]), in v256_add_64()
217 v128_add_64(a.val[0], b.val[0])); in v256_add_64()
221 return v256_from_v128(v128_padd_u8(a.val[1]), v128_padd_u8(a.val[0])); in v256_padd_u8()
225 return v256_from_v128(v128_padd_s16(a.val[1]), v128_padd_s16(a.val[0])); in v256_padd_s16()
229 return v256_from_v128(v128_sub_8(a.val[1], b.val[1]), in v256_sub_8()
230 v128_sub_8(a.val[0], b.val[0])); in v256_sub_8()
234 return v256_from_v128(v128_ssub_u8(a.val[1], b.val[1]), in v256_ssub_u8()
235 v128_ssub_u8(a.val[0], b.val[0])); in v256_ssub_u8()
239 return v256_from_v128(v128_ssub_s8(a.val[1], b.val[1]), in v256_ssub_s8()
240 v128_ssub_s8(a.val[0], b.val[0])); in v256_ssub_s8()
244 return v256_from_v128(v128_sub_16(a.val[1], b.val[1]), in v256_sub_16()
245 v128_sub_16(a.val[0], b.val[0])); in v256_sub_16()
249 return v256_from_v128(v128_ssub_s16(a.val[1], b.val[1]), in v256_ssub_s16()
250 v128_ssub_s16(a.val[0], b.val[0])); in v256_ssub_s16()
254 return v256_from_v128(v128_ssub_u16(a.val[1], b.val[1]), in v256_ssub_u16()
255 v128_ssub_u16(a.val[0], b.val[0])); in v256_ssub_u16()
259 return v256_from_v128(v128_sub_32(a.val[1], b.val[1]), in v256_sub_32()
260 v128_sub_32(a.val[0], b.val[0])); in v256_sub_32()
264 return v256_from_v128(v128_sub_64(a.val[1], b.val[1]), in v256_sub_64()
265 v128_sub_64(a.val[0], b.val[0])); in v256_sub_64()
269 return v256_from_v128(v128_abs_s16(a.val[1]), v128_abs_s16(a.val[0])); in v256_abs_s16()
273 return v256_from_v128(v128_abs_s8(a.val[1]), v128_abs_s8(a.val[0])); in v256_abs_s8()
284 return v256_from_v128(v128_mullo_s16(a.val[1], b.val[1]), in v256_mullo_s16()
285 v128_mullo_s16(a.val[0], b.val[0])); in v256_mullo_s16()
289 return v256_from_v128(v128_mulhi_s16(a.val[1], b.val[1]), in v256_mulhi_s16()
290 v128_mulhi_s16(a.val[0], b.val[0])); in v256_mulhi_s16()
294 return v256_from_v128(v128_mullo_s32(a.val[1], b.val[1]), in v256_mullo_s32()
295 v128_mullo_s32(a.val[0], b.val[0])); in v256_mullo_s32()
299 return v256_from_v128(v128_madd_s16(a.val[1], b.val[1]), in v256_madd_s16()
300 v128_madd_s16(a.val[0], b.val[0])); in v256_madd_s16()
304 return v256_from_v128(v128_madd_us8(a.val[1], b.val[1]), in v256_madd_us8()
305 v128_madd_us8(a.val[0], b.val[0])); in v256_madd_us8()
309 return v256_from_v128(v128_avg_u8(a.val[1], b.val[1]), in v256_avg_u8()
310 v128_avg_u8(a.val[0], b.val[0])); in v256_avg_u8()
314 return v256_from_v128(v128_rdavg_u8(a.val[1], b.val[1]), in v256_rdavg_u8()
315 v128_rdavg_u8(a.val[0], b.val[0])); in v256_rdavg_u8()
319 return v256_from_v128(v128_rdavg_u16(a.val[1], b.val[1]), in v256_rdavg_u16()
320 v128_rdavg_u16(a.val[0], b.val[0])); in v256_rdavg_u16()
324 return v256_from_v128(v128_avg_u16(a.val[1], b.val[1]), in v256_avg_u16()
325 v128_avg_u16(a.val[0], b.val[0])); in v256_avg_u16()
329 return v256_from_v128(v128_min_u8(a.val[1], b.val[1]), in v256_min_u8()
330 v128_min_u8(a.val[0], b.val[0])); in v256_min_u8()
334 return v256_from_v128(v128_max_u8(a.val[1], b.val[1]), in v256_max_u8()
335 v128_max_u8(a.val[0], b.val[0])); in v256_max_u8()
339 return v256_from_v128(v128_min_s8(a.val[1], b.val[1]), in v256_min_s8()
340 v128_min_s8(a.val[0], b.val[0])); in v256_min_s8()
349 return v256_from_v128(v128_blend_8(a.val[1], b.val[1], c.val[1]), in v256_blend_8()
350 v128_blend_8(a.val[0], b.val[0], c.val[0])); in v256_blend_8()
354 return v256_from_v128(v128_max_s8(a.val[1], b.val[1]), in v256_max_s8()
355 v128_max_s8(a.val[0], b.val[0])); in v256_max_s8()
359 return v256_from_v128(v128_min_s16(a.val[1], b.val[1]), in v256_min_s16()
360 v128_min_s16(a.val[0], b.val[0])); in v256_min_s16()
364 return v256_from_v128(v128_max_s16(a.val[1], b.val[1]), in v256_max_s16()
365 v128_max_s16(a.val[0], b.val[0])); in v256_max_s16()
369 return v256_from_v128(v128_min_s32(a.val[1], b.val[1]), in v256_min_s32()
370 v128_min_s32(a.val[0], b.val[0])); in v256_min_s32()
374 return v256_from_v128(v128_max_s32(a.val[1], b.val[1]), in v256_max_s32()
375 v128_max_s32(a.val[0], b.val[0])); in v256_max_s32()
379 return v256_from_v128(v128_ziphi_8(a.val[0], b.val[0]), in v256_ziplo_8()
380 v128_ziplo_8(a.val[0], b.val[0])); in v256_ziplo_8()
384 return v256_from_v128(v128_ziphi_8(a.val[1], b.val[1]), in v256_ziphi_8()
385 v128_ziplo_8(a.val[1], b.val[1])); in v256_ziphi_8()
389 return v256_from_v128(v128_ziphi_16(a.val[0], b.val[0]), in v256_ziplo_16()
390 v128_ziplo_16(a.val[0], b.val[0])); in v256_ziplo_16()
394 return v256_from_v128(v128_ziphi_16(a.val[1], b.val[1]), in v256_ziphi_16()
395 v128_ziplo_16(a.val[1], b.val[1])); in v256_ziphi_16()
399 return v256_from_v128(v128_ziphi_32(a.val[0], b.val[0]), in v256_ziplo_32()
400 v128_ziplo_32(a.val[0], b.val[0])); in v256_ziplo_32()
404 return v256_from_v128(v128_ziphi_32(a.val[1], b.val[1]), in v256_ziphi_32()
405 v128_ziplo_32(a.val[1], b.val[1])); in v256_ziphi_32()
409 return v256_from_v128(v128_ziphi_64(a.val[0], b.val[0]), in v256_ziplo_64()
410 v128_ziplo_64(a.val[0], b.val[0])); in v256_ziplo_64()
414 return v256_from_v128(v128_ziphi_64(a.val[1], b.val[1]), in v256_ziphi_64()
415 v128_ziplo_64(a.val[1], b.val[1])); in v256_ziphi_64()
419 return v256_from_v128(a.val[0], b.val[0]); in v256_ziplo_128()
423 return v256_from_v128(a.val[1], b.val[1]); in v256_ziphi_128()
439 return v256_from_v128(v128_unziplo_8(a.val[1], a.val[0]), in v256_unziplo_8()
440 v128_unziplo_8(b.val[1], b.val[0])); in v256_unziplo_8()
444 return v256_from_v128(v128_unziphi_8(a.val[1], a.val[0]), in v256_unziphi_8()
445 v128_unziphi_8(b.val[1], b.val[0])); in v256_unziphi_8()
449 return v256_from_v128(v128_unziplo_16(a.val[1], a.val[0]), in v256_unziplo_16()
450 v128_unziplo_16(b.val[1], b.val[0])); in v256_unziplo_16()
454 return v256_from_v128(v128_unziphi_16(a.val[1], a.val[0]), in v256_unziphi_16()
455 v128_unziphi_16(b.val[1], b.val[0])); in v256_unziphi_16()
459 return v256_from_v128(v128_unziplo_32(a.val[1], a.val[0]), in v256_unziplo_32()
460 v128_unziplo_32(b.val[1], b.val[0])); in v256_unziplo_32()
464 return v256_from_v128(v128_unziphi_32(a.val[1], a.val[0]), in v256_unziphi_32()
465 v128_unziphi_32(b.val[1], b.val[0])); in v256_unziphi_32()
471 _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(a.val[0]), in v256_unziplo_64()
472 _mm_castsi128_pd(a.val[1]), 0)), in v256_unziplo_64()
473 _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(b.val[0]), in v256_unziplo_64()
474 _mm_castsi128_pd(b.val[1]), 0))); in v256_unziplo_64()
476 return v256_from_v64(v128_low_v64(a.val[1]), v128_low_v64(a.val[0]), in v256_unziplo_64()
477 v128_low_v64(b.val[1]), v128_low_v64(b.val[0])); in v256_unziplo_64()
484 _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(a.val[0]), in v256_unziphi_64()
485 _mm_castsi128_pd(a.val[1]), 3)), in v256_unziphi_64()
486 _mm_castpd_si128(_mm_shuffle_pd(_mm_castsi128_pd(b.val[0]), in v256_unziphi_64()
487 _mm_castsi128_pd(b.val[1]), 3))); in v256_unziphi_64()
489 return v256_from_v64(v128_high_v64(a.val[1]), v128_high_v64(a.val[0]), in v256_unziphi_64()
490 v128_high_v64(b.val[1]), v128_high_v64(b.val[0])); in v256_unziphi_64()
499 return v256_from_v128(v128_unpackhi_u8_s16(a.val[0]), in v256_unpacklo_u8_s16()
500 v128_unpacklo_u8_s16(a.val[0])); in v256_unpacklo_u8_s16()
504 return v256_from_v128(v128_unpackhi_u8_s16(a.val[1]), in v256_unpackhi_u8_s16()
505 v128_unpacklo_u8_s16(a.val[1])); in v256_unpackhi_u8_s16()
513 return v256_from_v128(v128_unpackhi_s8_s16(a.val[0]), in v256_unpacklo_s8_s16()
514 v128_unpacklo_s8_s16(a.val[0])); in v256_unpacklo_s8_s16()
518 return v256_from_v128(v128_unpackhi_s8_s16(a.val[1]), in v256_unpackhi_s8_s16()
519 v128_unpacklo_s8_s16(a.val[1])); in v256_unpackhi_s8_s16()
523 return v256_from_v128(v128_pack_s32_s16(a.val[1], a.val[0]), in v256_pack_s32_s16()
524 v128_pack_s32_s16(b.val[1], b.val[0])); in v256_pack_s32_s16()
528 return v256_from_v128(v128_pack_s32_u16(a.val[1], a.val[0]), in v256_pack_s32_u16()
529 v128_pack_s32_u16(b.val[1], b.val[0])); in v256_pack_s32_u16()
533 return v256_from_v128(v128_pack_s16_u8(a.val[1], a.val[0]), in v256_pack_s16_u8()
534 v128_pack_s16_u8(b.val[1], b.val[0])); in v256_pack_s16_u8()
538 return v256_from_v128(v128_pack_s16_s8(a.val[1], a.val[0]), in v256_pack_s16_s8()
539 v128_pack_s16_s8(b.val[1], b.val[0])); in v256_pack_s16_s8()
551 return v256_from_v128(v128_unpackhi_u16_s32(a.val[0]), in v256_unpacklo_u16_s32()
552 v128_unpacklo_u16_s32(a.val[0])); in v256_unpacklo_u16_s32()
556 return v256_from_v128(v128_unpackhi_s16_s32(a.val[0]), in v256_unpacklo_s16_s32()
557 v128_unpacklo_s16_s32(a.val[0])); in v256_unpacklo_s16_s32()
561 return v256_from_v128(v128_unpackhi_u16_s32(a.val[1]), in v256_unpackhi_u16_s32()
562 v128_unpacklo_u16_s32(a.val[1])); in v256_unpackhi_u16_s32()
566 return v256_from_v128(v128_unpackhi_s16_s32(a.val[1]), in v256_unpackhi_s16_s32()
567 v128_unpacklo_s16_s32(a.val[1])); in v256_unpackhi_s16_s32()
571 return v256_from_v128(v128_cmpgt_s8(a.val[1], b.val[1]), in v256_cmpgt_s8()
572 v128_cmpgt_s8(a.val[0], b.val[0])); in v256_cmpgt_s8()
576 return v256_from_v128(v128_cmplt_s8(a.val[1], b.val[1]), in v256_cmplt_s8()
577 v128_cmplt_s8(a.val[0], b.val[0])); in v256_cmplt_s8()
581 return v256_from_v128(v128_cmpeq_8(a.val[1], b.val[1]), in v256_cmpeq_8()
582 v128_cmpeq_8(a.val[0], b.val[0])); in v256_cmpeq_8()
586 return v256_from_v128(v128_cmpgt_s16(a.val[1], b.val[1]), in v256_cmpgt_s16()
587 v128_cmpgt_s16(a.val[0], b.val[0])); in v256_cmpgt_s16()
591 return v256_from_v128(v128_cmplt_s16(a.val[1], b.val[1]), in v256_cmplt_s16()
592 v128_cmplt_s16(a.val[0], b.val[0])); in v256_cmplt_s16()
596 return v256_from_v128(v128_cmpeq_16(a.val[1], b.val[1]), in v256_cmpeq_16()
597 v128_cmpeq_16(a.val[0], b.val[0])); in v256_cmpeq_16()
601 return v256_from_v128(v128_cmpgt_s32(a.val[1], b.val[1]), in v256_cmpgt_s32()
602 v128_cmpgt_s32(a.val[0], b.val[0])); in v256_cmpgt_s32()
606 return v256_from_v128(v128_cmplt_s32(a.val[1], b.val[1]), in v256_cmplt_s32()
607 v128_cmplt_s32(a.val[0], b.val[0])); in v256_cmplt_s32()
611 return v256_from_v128(v128_cmpeq_32(a.val[1], b.val[1]), in v256_cmpeq_32()
612 v128_cmpeq_32(a.val[0], b.val[0])); in v256_cmpeq_32()
618 uint8x16x2_t p = { { vreinterpretq_u8_s64(x.val[0]), in v256_shuffle_8()
619 vreinterpretq_u8_s64(x.val[1]) } }; in v256_shuffle_8()
621 vreinterpretq_s64_u8(vqtbl2q_u8(p, vreinterpretq_u8_s64(pattern.val[1]))), in v256_shuffle_8()
623 vqtbl2q_u8(p, vreinterpretq_u8_s64(pattern.val[0])))); in v256_shuffle_8()
625 uint8x8x4_t p = { { vget_low_u8(vreinterpretq_u8_s64(x.val[0])), in v256_shuffle_8()
626 vget_high_u8(vreinterpretq_u8_s64(x.val[0])), in v256_shuffle_8()
627 vget_low_u8(vreinterpretq_u8_s64(x.val[1])), in v256_shuffle_8()
628 vget_high_u8(vreinterpretq_u8_s64(x.val[1])) } }; in v256_shuffle_8()
631 vtbl4_u8(p, vreinterpret_u8_s64(vget_high_s64(pattern.val[1])))), in v256_shuffle_8()
633 vtbl4_u8(p, vreinterpret_u8_s64(vget_low_s64(pattern.val[1])))), in v256_shuffle_8()
635 vtbl4_u8(p, vreinterpret_u8_s64(vget_high_s64(pattern.val[0])))), in v256_shuffle_8()
637 vtbl4_u8(p, vreinterpret_u8_s64(vget_low_s64(pattern.val[0]))))); in v256_shuffle_8()
641 v128 maskhi = v128_cmplt_s8(pattern.val[1], c16); in v256_shuffle_8()
642 v128 masklo = v128_cmplt_s8(pattern.val[0], c16); in v256_shuffle_8()
644 v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[1], c16)), in v256_shuffle_8()
645 v128_shuffle_8(x.val[0], pattern.val[1]), maskhi), in v256_shuffle_8()
646 v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[0], c16)), in v256_shuffle_8()
647 v128_shuffle_8(x.val[0], pattern.val[0]), masklo)); in v256_shuffle_8()
655 vreinterpretq_u8_s64(y.val[0]), in v256_wideshuffle_8()
656 vreinterpretq_u8_s64(y.val[1]), in v256_wideshuffle_8()
657 vreinterpretq_u8_s64(x.val[0]), in v256_wideshuffle_8()
658 vreinterpretq_u8_s64(x.val[1]), in v256_wideshuffle_8()
661 vreinterpretq_s64_u8(vqtbl4q_u8(p, vreinterpretq_u8_s64(pattern.val[1]))), in v256_wideshuffle_8()
663 vqtbl4q_u8(p, vreinterpretq_u8_s64(pattern.val[0])))); in v256_wideshuffle_8()
667 uint8x8x4_t p = { { vget_low_u8(vreinterpretq_u8_s64(x.val[0])), in v256_wideshuffle_8()
668 vget_high_u8(vreinterpretq_u8_s64(x.val[0])), in v256_wideshuffle_8()
669 vget_low_u8(vreinterpretq_u8_s64(x.val[1])), in v256_wideshuffle_8()
670 vget_high_u8(vreinterpretq_u8_s64(x.val[1])) } }; in v256_wideshuffle_8()
671 uint8x8x4_t q = { { vget_low_u8(vreinterpretq_u8_s64(y.val[0])), in v256_wideshuffle_8()
672 vget_high_u8(vreinterpretq_u8_s64(y.val[0])), in v256_wideshuffle_8()
673 vget_low_u8(vreinterpretq_u8_s64(y.val[1])), in v256_wideshuffle_8()
674 vget_high_u8(vreinterpretq_u8_s64(y.val[1])) } }; in v256_wideshuffle_8()
677 p, vreinterpret_u8_s64(vget_high_s64(p32.val[1])))), in v256_wideshuffle_8()
679 p, vreinterpret_u8_s64(vget_low_s64(p32.val[1])))), in v256_wideshuffle_8()
681 p, vreinterpret_u8_s64(vget_high_s64(p32.val[0])))), in v256_wideshuffle_8()
683 p, vreinterpret_u8_s64(vget_low_s64(p32.val[0]))))); in v256_wideshuffle_8()
686 q, vreinterpret_u8_s64(vget_high_s64(pattern.val[1])))), in v256_wideshuffle_8()
688 q, vreinterpret_u8_s64(vget_low_s64(pattern.val[1])))), in v256_wideshuffle_8()
690 q, vreinterpret_u8_s64(vget_high_s64(pattern.val[0])))), in v256_wideshuffle_8()
692 q, vreinterpret_u8_s64(vget_low_s64(pattern.val[0]))))); in v256_wideshuffle_8()
699 v128 maskhi16 = v128_cmpgt_s8(c16, pattern.val[1]); in v256_wideshuffle_8()
700 v128 masklo16 = v128_cmpgt_s8(c16, pattern.val[0]); in v256_wideshuffle_8()
701 v128 maskhi48 = v128_cmpgt_s8(c48, pattern.val[1]); in v256_wideshuffle_8()
702 v128 masklo48 = v128_cmpgt_s8(c48, pattern.val[0]); in v256_wideshuffle_8()
704 v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[1], c48)), in v256_wideshuffle_8()
705 v128_shuffle_8(x.val[0], v128_sub_8(pattern.val[1], c32)), in v256_wideshuffle_8()
707 v128_blend_8(v128_shuffle_8(x.val[1], v128_sub_8(pattern.val[0], c48)), in v256_wideshuffle_8()
708 v128_shuffle_8(x.val[0], v128_sub_8(pattern.val[0], c32)), in v256_wideshuffle_8()
711 v128_blend_8(v128_shuffle_8(y.val[1], v128_sub_8(pattern.val[1], c16)), in v256_wideshuffle_8()
712 v128_shuffle_8(y.val[0], pattern.val[1]), maskhi16), in v256_wideshuffle_8()
713 v128_blend_8(v128_shuffle_8(y.val[1], v128_sub_8(pattern.val[0], c16)), in v256_wideshuffle_8()
714 v128_shuffle_8(y.val[0], pattern.val[0]), masklo16)); in v256_wideshuffle_8()
726 return v256_from_v128(v128_shl_8(a.val[1], c), v128_shl_8(a.val[0], c)); in v256_shl_8()
730 return v256_from_v128(v128_shr_u8(a.val[1], c), v128_shr_u8(a.val[0], c)); in v256_shr_u8()
734 return v256_from_v128(v128_shr_s8(a.val[1], c), v128_shr_s8(a.val[0], c)); in v256_shr_s8()
738 return v256_from_v128(v128_shl_16(a.val[1], c), v128_shl_16(a.val[0], c)); in v256_shl_16()
742 return v256_from_v128(v128_shr_u16(a.val[1], c), v128_shr_u16(a.val[0], c)); in v256_shr_u16()
746 return v256_from_v128(v128_shr_s16(a.val[1], c), v128_shr_s16(a.val[0], c)); in v256_shr_s16()
750 return v256_from_v128(v128_shl_32(a.val[1], c), v128_shl_32(a.val[0], c)); in v256_shl_32()
754 return v256_from_v128(v128_shr_u32(a.val[1], c), v128_shr_u32(a.val[0], c)); in v256_shr_u32()
758 return v256_from_v128(v128_shr_s32(a.val[1], c), v128_shr_s32(a.val[0], c)); in v256_shr_s32()
762 return v256_from_v128(v128_shl_64(a.val[1], c), v128_shl_64(a.val[0], c)); in v256_shl_64()
766 return v256_from_v128(v128_shr_u64(a.val[1], c), v128_shr_u64(a.val[0], c)); in v256_shr_u64()
770 return v256_from_v128(v128_shr_s64(a.val[1], c), v128_shr_s64(a.val[0], c)); in v256_shr_s64()
776 ((n) < 16 ? v256_from_v128(v128_or(v128_shl_n_byte(a.val[1], n), \
777 v128_shr_n_byte(a.val[0], 16 - (n))), \
778 v128_shl_n_byte(a.val[0], (n))) \
780 (n) > 16 ? v128_shl_n_byte(a.val[0], (n)-16) : a.val[0], \
787 ? v256_from_v128(v128_shr_n_byte(a.val[1], n), \
788 v128_or(v128_shr_n_byte(a.val[0], n), \
789 v128_shl_n_byte(a.val[1], 16 - (n)))) \
792 (n) > 16 ? v128_shr_n_byte(a.val[1], (n)-16) : a.val[1])))
798 v256_from_v128(v128_shl_n_8(a.val[1], n), v128_shl_n_8(a.val[0], n))
800 v256_from_v128(v128_shl_n_16(a.val[1], n), v128_shl_n_16(a.val[0], n))
802 v256_from_v128(v128_shl_n_32(a.val[1], n), v128_shl_n_32(a.val[0], n))
804 v256_from_v128(v128_shl_n_64(a.val[1], n), v128_shl_n_64(a.val[0], n))
806 v256_from_v128(v128_shr_n_u8(a.val[1], n), v128_shr_n_u8(a.val[0], n))
808 v256_from_v128(v128_shr_n_u16(a.val[1], n), v128_shr_n_u16(a.val[0], n))
810 v256_from_v128(v128_shr_n_u32(a.val[1], n), v128_shr_n_u32(a.val[0], n))
812 v256_from_v128(v128_shr_n_u64(a.val[1], n), v128_shr_n_u64(a.val[0], n))
814 v256_from_v128(v128_shr_n_s8(a.val[1], n), v128_shr_n_s8(a.val[0], n))
816 v256_from_v128(v128_shr_n_s16(a.val[1], n), v128_shr_n_s16(a.val[0], n))
818 v256_from_v128(v128_shr_n_s32(a.val[1], n), v128_shr_n_s32(a.val[0], n))
820 v256_from_v128(v128_shr_n_s64(a.val[1], n), v128_shr_n_s64(a.val[0], n))
826 sad128_internal_u16 val[2]; member
831 t.val[1] = v128_sad_u16_init(); in v256_sad_u16_init()
832 t.val[0] = v128_sad_u16_init(); in v256_sad_u16_init()
842 t.val[1] = v128_sad_u16(s.val[1], a.val[1], b.val[1]); in v256_sad_u16()
843 t.val[0] = v128_sad_u16(s.val[0], a.val[0], b.val[0]); in v256_sad_u16()
848 return v128_sad_u16_sum(s.val[1]) + v128_sad_u16_sum(s.val[0]); in v256_sad_u16_sum()
852 ssd128_internal_s16 val[2]; member
857 t.val[1] = v128_ssd_s16_init(); in v256_ssd_s16_init()
858 t.val[0] = v128_ssd_s16_init(); in v256_ssd_s16_init()
867 t.val[1] = v128_ssd_s16(s.val[1], a.val[1], b.val[1]); in v256_ssd_s16()
868 t.val[0] = v128_ssd_s16(s.val[0], a.val[0], b.val[0]); in v256_ssd_s16()
873 return v128_ssd_s16_sum(s.val[1]) + v128_ssd_s16_sum(s.val[0]); in v256_ssd_s16_sum()