Bug in avx512f swizzle

[lu1and10]

Hi XSIMD development team,

Thanks for maintaining such a great package.

I believe I’ve found a bug in the AVX-512F swizzle implementation: detail::is_cross_lane appears to return the wrong result on both master and release 14.

For example, with the swizzle mask:

3 2 1 0 7 6 5 4

the logic at

xsimd/include/xsimd/arch/xsimd_avx512f.hpp

Line 2376 in a03a5d9

XSIMD_IF_CONSTEXPR(!detail::is_cross_lane(mask))

evaluates detail::is_cross_lane(...) as true and takes the _mm512_permute_pd(self, imm) path. However, _mm512_permute_pd only permutes within 128-bit lanes, so it is not valid for a mask like 3 2 1 0 7 6 5 4, which requires a cross-128-bit-lane permutation.

This code path is triggered when the built-in shuffle is not available, falling back through the implementation here:

xsimd/include/xsimd/arch/common/xsimd_common_memory.hpp

Line 660 in a03a5d9

#else

.

Thanks.

[DiamonDinoia]

PR #1242 fixes this

[lu1and10]

PR #1242 fixes this

Yes, I think the PR tile is misleading, it's not just fixing the gcc-10 auto simd, it's also fix the avx512f swizzle bug in the actual code for all compilers don't support built in shuffle.

[lu1and10]

You may want to separate these two different issues.

[DiamonDinoia]

We have two problems for the zip in swizzle:

1. The bug in detail::is_cross_lane
2. A bug in gcc-10

Now, merging any of those separately causes CI to fail. We could merge them anyway and have master failing for a few mins. Or we could merge the fixes then merge the changes to CI.

@serge-sans-paille what do you suggest?

[lu1and10]

We have two problems for the zip in swizzle:

1. The bug in detail::is_cross_lane
2. A bug in gcc-10

Now, merging any of those separately causes CI to fail. We could merge them anyway and have master failing for a few mins. Or we could merge the fixes then merge the changes to CI.

@serge-sans-paille what do you suggest?

FYI I guess gcc-11 only has the The bug in detail::is_cross_lane issue

[serge-sans-paille]

let's not be petty: one PR, two commits fixing each aspect, ideally with a commit that showcases the issue and we're good

[lu1and10]

@DiamonDinoia @serge-sans-paille thanks