How to cast xsimd::batch<uint8_t, 16> to xsimd::batch<uint32_t, 4>

[guyuqi]

Hi @JohanMabille @serge-sans-paille ,

Thank you for your time.

In Arrow, I would like to cast xsimd::batch<uint8_t, 16> to xsimd::batch<uint32_t, 4> for operating some shared data buffer:

using simd_batch_8 = xsimd::batch<uint8_t, 16>;
using simd_batch_32 = xsimd::batch<uint32_t, 4>;

simd_batch_32 rtn;
simd_batch_8 lhs, rhs;

rtn= xsimd::zip_lo(
    static_cast<simd_batch_32>(lhs), 
    static_cast<simd_batch_32>(rhs),);

It failed to compile in Arm Neon, but it works correctly in x86 SSE:

error: no type named ‘type’ in ‘struct std::enable_if<false, void>’
  367 |         using is_array_initializer_t = typename is_array_initializer<T, N, Args...>::type;

I noted that it would check the variadic arguments of template:

template <class... Args, class Enable = detail::is_array_initializer_t<uint8_t, 16, Args...>>

I didn't find similar checking in xsimd_sse_int8.hpp, xsimd_sse_int16.hpp, xsimd_sse_int32.hpp...

It was found in Neon code branches: https://github.com/xtensor-stack/xsimd/blob/master/include/xsimd/types/xsimd_neon_int8.hpp#L49

https://github.com/xtensor-stack/xsimd/blob/master/include/xsimd/types/xsimd_neon_int16.hpp#L49

May I ask the checking of template arguments is necessary here? How to correctly cast them?

Thanks.

[guyuqi]

Hi, @serge-sans-paille Maybe I missed something. Could please share the way to cast <uint8_t, 16> to <uint32_t, 4> in xsimd?

Is it correct to cast by static_cast?

xsimd::batch<uint8_t, 16> batch8;
xsimd::batch<uint32_t, 4> batch32;

//casting by static_cast ?
batch32 = static_cast<xsimd::batch<uint32_t, 4>>(batch8);

[serge-sans-paille]

Yeah, that's fine, it's the same register underneath, see https://godbolt.org/z/x8Trdb1cz

[serge-sans-paille]

ok, it's fine on X86 but not on arm, which looks like a bug to me, let me investigate that.

[guyuqi]

Thanks for your feedback.

ok, it's fine on X86 but not on arm, which looks like a bug to me, let me investigate that.

Yes, static_cast doesn't work on Arm64. I'm working on adding these casing operations to xsimd and would like to submit the PR soon.

[serge-sans-paille]

I'm working on adding these casing operations to xsimd and would like to submit the PR soon.

Great! I'm looking forward to review these

[serge-sans-paille]

@guyuqi I believe you fixed that one in #487, correct?