88       vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
90       vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
92       vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
94       vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
127           vmlal_u8(vmovl_u8(vget_low_u8(s1)), vget_low_u8(s0), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
129           vmlal_u8(vmovl_u8(vget_high_u8(s1)), vget_high_u8(s0), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
131           vmlal_u8(vmovl_u8(vget_low_u8(s0)), vget_low_u8(s1), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
133           vmlal_u8(vmovl_u8(vget_high_u8(s0)), vget_high_u8(s1), three_u8);  in jsimd_h2v1_fancy_upsample_neon()
247     uint16x8_t s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)),  in jsimd_h2v2_fancy_upsample_neon()
249     uint16x8_t s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)),  in jsimd_h2v2_fancy_upsample_neon()
251     uint16x8_t s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)),  in jsimd_h2v2_fancy_upsample_neon()
253     uint16x8_t s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)),  in jsimd_h2v2_fancy_upsample_neon()
259     uint16x8_t s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)),  in jsimd_h2v2_fancy_upsample_neon()
261     uint16x8_t s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)),  in jsimd_h2v2_fancy_upsample_neon()
263     uint16x8_t s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)),  in jsimd_h2v2_fancy_upsample_neon()
265     uint16x8_t s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)),  in jsimd_h2v2_fancy_upsample_neon()
315       s0colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s0A)), vget_low_u8(s0B),  in jsimd_h2v2_fancy_upsample_neon()
317       s0colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s0A)), vget_high_u8(s0B),  in jsimd_h2v2_fancy_upsample_neon()
319       s0colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s0C)), vget_low_u8(s0B),  in jsimd_h2v2_fancy_upsample_neon()
321       s0colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s0C)), vget_high_u8(s0B),  in jsimd_h2v2_fancy_upsample_neon()
327       s1colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(s1A)), vget_low_u8(s1B),  in jsimd_h2v2_fancy_upsample_neon()
329       s1colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(s1A)), vget_high_u8(s1B),  in jsimd_h2v2_fancy_upsample_neon()
331       s1colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(s1C)), vget_low_u8(s1B),  in jsimd_h2v2_fancy_upsample_neon()
333       s1colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(s1C)), vget_high_u8(s1B),  in jsimd_h2v2_fancy_upsample_neon()
446       uint16x8_t colsum0_l = vmlal_u8(vmovl_u8(vget_low_u8(sA)),  in jsimd_h1v2_fancy_upsample_neon()
448       uint16x8_t colsum0_h = vmlal_u8(vmovl_u8(vget_high_u8(sA)),  in jsimd_h1v2_fancy_upsample_neon()
450       uint16x8_t colsum1_l = vmlal_u8(vmovl_u8(vget_low_u8(sC)),  in jsimd_h1v2_fancy_upsample_neon()
452       uint16x8_t colsum1_h = vmlal_u8(vmovl_u8(vget_high_u8(sC)),  in jsimd_h1v2_fancy_upsample_neon()