xtensor triggers xsimd error on arm64: <bool, xsimd::neon64> usage of batch type with unsupported type about xsimd HOT 8 CLOSED

A bit more additional info

• All errors in the build log boil down to the instantiation of boolean operations, which seem to be unsupported on neon64 (no bool register declared in https://github.com/xtensor-stack/xsimd/blob/10.0.0/include/xsimd/types/xsimd_neon64_register.hpp).

• Minimal example to reproduce the error on a Raspberry Pi 3 B (the commented line will trigger the error):

#include <xtensor/xtensor.hpp>

int main()
{
    {
        xt::xtensor<double, 1> a{0., 1.};
        xt::xtensor<double, 1> b = a && 0.;
    }

    {
        xt::xtensor<bool, 1> a{false, true};
        //xt::xtensor<bool, 1> b = a && false;
    }

    return 0;
}

from xsimd.

It's a bug in neon64 then :p

from xsimd.

I confirm the following xsimd-only snippet works on x86-64 but not on aarch64

#include <xsimd/xsimd.hpp>

auto pain() {
  xsimd::batch<bool> x,y;
  return x && y;
}

I'll investigate, thanks for bringing that up!

from xsimd.

In that case it's worth noting that xsimd is passing its own tests on arm64,
for v11 in particular, as 11.1.0-1exp4,
https://ci.debian.net/data/autopkgtest/unstable/arm64/x/xsimd/38216389/log.gz
(from https://ci.debian.net/packages/x/xsimd/unstable/arm64/ )

Does that mean xsimd::batch<bool> is not being tested in the xsimd unit tests?

from xsimd.

@JohanMabille We currently don't have support for batch<bool>. We could alias to batch<uint8_t> or try to be smart, à la std::vector<bool>. The actual implementation would be generic for point-to-point operations, but rather complex for shuffles and the like. Likewise memory operations would be slightly more complex, so I advocate for an alias to batch<uint8_t>. What do you think?

from xsimd.

After some extra digging: we actually don't really support xsimd::batch<bool, A> because we try to outsmart the user and decide it should be an xsimd::batch_bool<T, A>. It's all linked to simd_return_type and if I recall correctly, that's something that got baked in for xtensor. @JohanMabille, any thoughts?

from xsimd.

armhf is also affected, apparently the same problem from neon, https://ci.debian.net/data/autopkgtest/unstable/armhf/x/xtensor/38875115/log.gz

from xsimd.

It's all linked to simd_return_type and if I recall correctly, that's something that got baked in for xtensor

Not only for xtensor, it is required to promote types when playing with mixed arithmetic (std::complex<double> and double for instance) and to allow operations on "equivalent" types (like std::complex and xtl::complex for instance).

batch<bool> should definitely not support the same operations as batch<other_arithmetic_type>. I know the following is valid in C++:

bool b1 = true, b2 = false;
bool res = b1 + b2;

but I find it ugly and we should definitely avoid this in xsimd.

Also batch_bool must be constructible from boolscalars, not integers (even if C++ allows implicit conversions, we should not encourgae them in the API), so that's another reason for not aliasing it to batch<uint8_t>.

We could imagine aliasing it to batch_bool<uint8_t>, but then it would feel inconsistent with the return type of batch<double> == batch<double> for instance.

This issue is definitely a bug in the definition of simd_return_type, but we also need to understand why batch<bool> is allowed on x86/x64 architectures. and I don't remember why we allowed register for bool on x86/x64 architecture.

from xsimd.