[Question] Combining two vector registers with different LMUL

[Tameem97]

Hi,

It is already discussed in issue #28 but I still can't figure out how to combine two vector. More detail about this is that I want to combine two vuint8mf4_t type vector into single vuint8mf2_t so I can further perform multiplication For example, I have two vector

        vuint8mf4_t v1   = __riscv_vle8_v_u8mf4(arr_1, 8);
        vuint8mf4_t v2   = __riscv_vle8_v_u8mf4(arr_2, 8);

Now I want to combine them into v3 such that v3 first half consist of v1 and second half has v2

        vuint8mf2_t v3 = {v2 : v1};

The method I already search were vset which is only available for segment registers, and vlmul, zext. I required this for multiplication else I have to load vector (which is required to multiply with v3) multiple time in smaller chunk.

Furthermore, this is similar to ARM Neon intrinsics vcombine (Link)

Thanks

[dzaima]

vuint8mf2_t ext1 = __riscv_vlmul_ext_v_u8mf4_u8mf2(v1);
vuint8mf2_t ext2 = __riscv_vlmul_ext_v_u8mf4_u8mf2(v2);
size_t vl = __riscv_vsetvlmax_e8mf2();
vuint8mf2_t res = __riscv_vslideup_vx_u8mf2(ext1, ext2, vl/2, vl);

Though, if you were to use LMUL ≥ 1, e.g. with inputs of vuint8m1_t v1, v2 you could

vuint8m2_t ext1 = __riscv_vlmul_ext_v_u8m1_u8m2(v1);
vuint8m2_t res = __riscv_vset_v_u8m1_u8m2(ext1, 1, v2);

But I would guess that there probably is some alternative way to achieve what you want (especially considering that such vector concatenation is quite a weird operation due to varying VLEN across RISC-V V implementations).

[Tameem97]

Thanks @dzaima for answer. Yes, there are some other ways also but they are taking more intrinsics then I anticipated so this is the only thing which I got stuck with for LMUL<1