Lines Matching refs:v16u8
29 v16u8 src0, src1, dst0; in ScaleARGBRowDown2_MSA()
33 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ScaleARGBRowDown2_MSA()
34 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ScaleARGBRowDown2_MSA()
35 dst0 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0); in ScaleARGBRowDown2_MSA()
47 v16u8 src0, src1, vec0, vec1, dst0; in ScaleARGBRowDown2Linear_MSA()
51 src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0); in ScaleARGBRowDown2Linear_MSA()
52 src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16); in ScaleARGBRowDown2Linear_MSA()
53 vec0 = (v16u8)__msa_pckev_w((v4i32)src1, (v4i32)src0); in ScaleARGBRowDown2Linear_MSA()
54 vec1 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0); in ScaleARGBRowDown2Linear_MSA()
55 dst0 = (v16u8)__msa_aver_u_b((v16u8)vec0, (v16u8)vec1); in ScaleARGBRowDown2Linear_MSA()
69 v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0; in ScaleARGBRowDown2Box_MSA()
74 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in ScaleARGBRowDown2Box_MSA()
75 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in ScaleARGBRowDown2Box_MSA()
76 src2 = (v16u8)__msa_ld_b((v16i8*)t, 0); in ScaleARGBRowDown2Box_MSA()
77 src3 = (v16u8)__msa_ld_b((v16i8*)t, 16); in ScaleARGBRowDown2Box_MSA()
78 vec0 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src0, (v16i8)src0); in ScaleARGBRowDown2Box_MSA()
79 vec1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src1, (v16i8)src1); in ScaleARGBRowDown2Box_MSA()
80 vec2 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src2, (v16i8)src2); in ScaleARGBRowDown2Box_MSA()
81 vec3 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src3, (v16i8)src3); in ScaleARGBRowDown2Box_MSA()
90 dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0); in ScaleARGBRowDown2Box_MSA()
131 v16u8 src0 = {0}, src1 = {0}, src2 = {0}, src3 = {0}; in ScaleARGBRowDownEvenBox_MSA()
132 v16u8 vec0, vec1, vec2, vec3; in ScaleARGBRowDownEvenBox_MSA()
134 v16u8 dst0; in ScaleARGBRowDownEvenBox_MSA()
141 src0 = (v16u8)__msa_insert_d((v2i64)src0, 0, data0); in ScaleARGBRowDownEvenBox_MSA()
142 src0 = (v16u8)__msa_insert_d((v2i64)src0, 1, data1); in ScaleARGBRowDownEvenBox_MSA()
143 src1 = (v16u8)__msa_insert_d((v2i64)src1, 0, data2); in ScaleARGBRowDownEvenBox_MSA()
144 src1 = (v16u8)__msa_insert_d((v2i64)src1, 1, data3); in ScaleARGBRowDownEvenBox_MSA()
149 src2 = (v16u8)__msa_insert_d((v2i64)src2, 0, data0); in ScaleARGBRowDownEvenBox_MSA()
150 src2 = (v16u8)__msa_insert_d((v2i64)src2, 1, data1); in ScaleARGBRowDownEvenBox_MSA()
151 src3 = (v16u8)__msa_insert_d((v2i64)src3, 0, data2); in ScaleARGBRowDownEvenBox_MSA()
152 src3 = (v16u8)__msa_insert_d((v2i64)src3, 1, data3); in ScaleARGBRowDownEvenBox_MSA()
153 vec0 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src0); in ScaleARGBRowDownEvenBox_MSA()
154 vec1 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src1); in ScaleARGBRowDownEvenBox_MSA()
155 vec2 = (v16u8)__msa_ilvl_b((v16i8)src2, (v16i8)src0); in ScaleARGBRowDownEvenBox_MSA()
156 vec3 = (v16u8)__msa_ilvl_b((v16i8)src3, (v16i8)src1); in ScaleARGBRowDownEvenBox_MSA()
169 dst0 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4); in ScaleARGBRowDownEvenBox_MSA()
182 v16u8 src0, src1, src2, src3, dst0, dst1; in ScaleRowDown2_MSA()
186 src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0); in ScaleRowDown2_MSA()
187 src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16); in ScaleRowDown2_MSA()
188 src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32); in ScaleRowDown2_MSA()
189 src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48); in ScaleRowDown2_MSA()
190 dst0 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in ScaleRowDown2_MSA()
191 dst1 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in ScaleRowDown2_MSA()
203 v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0, dst1; in ScaleRowDown2Linear_MSA()
207 src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0); in ScaleRowDown2Linear_MSA()
208 src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16); in ScaleRowDown2Linear_MSA()
209 src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32); in ScaleRowDown2Linear_MSA()
210 src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48); in ScaleRowDown2Linear_MSA()
211 vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ScaleRowDown2Linear_MSA()
212 vec2 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in ScaleRowDown2Linear_MSA()
213 vec1 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0); in ScaleRowDown2Linear_MSA()
214 vec3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2); in ScaleRowDown2Linear_MSA()
230 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0, dst1; in ScaleRowDown2Box_MSA()
234 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in ScaleRowDown2Box_MSA()
235 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in ScaleRowDown2Box_MSA()
236 src2 = (v16u8)__msa_ld_b((v16i8*)s, 32); in ScaleRowDown2Box_MSA()
237 src3 = (v16u8)__msa_ld_b((v16i8*)s, 48); in ScaleRowDown2Box_MSA()
238 src4 = (v16u8)__msa_ld_b((v16i8*)t, 0); in ScaleRowDown2Box_MSA()
239 src5 = (v16u8)__msa_ld_b((v16i8*)t, 16); in ScaleRowDown2Box_MSA()
240 src6 = (v16u8)__msa_ld_b((v16i8*)t, 32); in ScaleRowDown2Box_MSA()
241 src7 = (v16u8)__msa_ld_b((v16i8*)t, 48); in ScaleRowDown2Box_MSA()
254 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ScaleRowDown2Box_MSA()
255 dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2); in ScaleRowDown2Box_MSA()
268 v16u8 src0, src1, src2, src3, vec0, vec1, dst0; in ScaleRowDown4_MSA()
272 src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0); in ScaleRowDown4_MSA()
273 src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16); in ScaleRowDown4_MSA()
274 src2 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 32); in ScaleRowDown4_MSA()
275 src3 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 48); in ScaleRowDown4_MSA()
276 vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0); in ScaleRowDown4_MSA()
277 vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2); in ScaleRowDown4_MSA()
278 dst0 = (v16u8)__msa_pckod_b((v16i8)vec1, (v16i8)vec0); in ScaleRowDown4_MSA()
294 v16u8 src0, src1, src2, src3, src4, src5, src6, src7, dst0; in ScaleRowDown4Box_MSA()
299 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in ScaleRowDown4Box_MSA()
300 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in ScaleRowDown4Box_MSA()
301 src2 = (v16u8)__msa_ld_b((v16i8*)s, 32); in ScaleRowDown4Box_MSA()
302 src3 = (v16u8)__msa_ld_b((v16i8*)s, 48); in ScaleRowDown4Box_MSA()
303 src4 = (v16u8)__msa_ld_b((v16i8*)t0, 0); in ScaleRowDown4Box_MSA()
304 src5 = (v16u8)__msa_ld_b((v16i8*)t0, 16); in ScaleRowDown4Box_MSA()
305 src6 = (v16u8)__msa_ld_b((v16i8*)t0, 32); in ScaleRowDown4Box_MSA()
306 src7 = (v16u8)__msa_ld_b((v16i8*)t0, 48); in ScaleRowDown4Box_MSA()
315 src0 = (v16u8)__msa_ld_b((v16i8*)t1, 0); in ScaleRowDown4Box_MSA()
316 src1 = (v16u8)__msa_ld_b((v16i8*)t1, 16); in ScaleRowDown4Box_MSA()
317 src2 = (v16u8)__msa_ld_b((v16i8*)t1, 32); in ScaleRowDown4Box_MSA()
318 src3 = (v16u8)__msa_ld_b((v16i8*)t1, 48); in ScaleRowDown4Box_MSA()
319 src4 = (v16u8)__msa_ld_b((v16i8*)t2, 0); in ScaleRowDown4Box_MSA()
320 src5 = (v16u8)__msa_ld_b((v16i8*)t2, 16); in ScaleRowDown4Box_MSA()
321 src6 = (v16u8)__msa_ld_b((v16i8*)t2, 32); in ScaleRowDown4Box_MSA()
322 src7 = (v16u8)__msa_ld_b((v16i8*)t2, 48); in ScaleRowDown4Box_MSA()
341 dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0); in ScaleRowDown4Box_MSA()
358 v16u8 src0, src1, vec0; in ScaleRowDown38_MSA()
366 src0 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 0); in ScaleRowDown38_MSA()
367 src1 = (v16u8)__msa_ld_b((v16i8*)src_ptr, 16); in ScaleRowDown38_MSA()
368 vec0 = (v16u8)__msa_vshf_b(mask, (v16i8)src1, (v16i8)src0); in ScaleRowDown38_MSA()
387 v16u8 src0, src1, src2, src3, out; in ScaleRowDown38_2_Box_MSA()
400 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in ScaleRowDown38_2_Box_MSA()
401 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in ScaleRowDown38_2_Box_MSA()
402 src2 = (v16u8)__msa_ld_b((v16i8*)t, 0); in ScaleRowDown38_2_Box_MSA()
403 src3 = (v16u8)__msa_ld_b((v16i8*)t, 16); in ScaleRowDown38_2_Box_MSA()
408 vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0); in ScaleRowDown38_2_Box_MSA()
409 vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1); in ScaleRowDown38_2_Box_MSA()
410 vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2); in ScaleRowDown38_2_Box_MSA()
411 vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3); in ScaleRowDown38_2_Box_MSA()
436 out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0); in ScaleRowDown38_2_Box_MSA()
457 v16u8 src0, src1, src2, src3, src4, src5, out; in ScaleRowDown38_3_Box_MSA()
470 src0 = (v16u8)__msa_ld_b((v16i8*)s, 0); in ScaleRowDown38_3_Box_MSA()
471 src1 = (v16u8)__msa_ld_b((v16i8*)s, 16); in ScaleRowDown38_3_Box_MSA()
472 src2 = (v16u8)__msa_ld_b((v16i8*)t0, 0); in ScaleRowDown38_3_Box_MSA()
473 src3 = (v16u8)__msa_ld_b((v16i8*)t0, 16); in ScaleRowDown38_3_Box_MSA()
474 src4 = (v16u8)__msa_ld_b((v16i8*)t1, 0); in ScaleRowDown38_3_Box_MSA()
475 src5 = (v16u8)__msa_ld_b((v16i8*)t1, 16); in ScaleRowDown38_3_Box_MSA()
484 vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0); in ScaleRowDown38_3_Box_MSA()
485 vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1); in ScaleRowDown38_3_Box_MSA()
486 vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2); in ScaleRowDown38_3_Box_MSA()
487 vec3 = __msa_hadd_u_h((v16u8)vec3, (v16u8)vec3); in ScaleRowDown38_3_Box_MSA()
488 vec0 += __msa_hadd_u_h((v16u8)vec4, (v16u8)vec4); in ScaleRowDown38_3_Box_MSA()
489 vec1 += __msa_hadd_u_h((v16u8)vec5, (v16u8)vec5); in ScaleRowDown38_3_Box_MSA()
490 vec2 += __msa_hadd_u_h((v16u8)vec6, (v16u8)vec6); in ScaleRowDown38_3_Box_MSA()
491 vec3 += __msa_hadd_u_h((v16u8)vec7, (v16u8)vec7); in ScaleRowDown38_3_Box_MSA()
516 out = (v16u8)__msa_vshf_b(dst_mask, (v16i8)vec1, (v16i8)vec0); in ScaleRowDown38_3_Box_MSA()
530 v16u8 src0; in ScaleAddRow_MSA()