apriori_compiled_kernel_t::get_attribute should be marked inline

[jhol]

kernel::attribute_value_t apriori_compiled_kernel_t::get_attribute(kernel::attribute_t attribute) const

is defined in src/cuda/api/multi_wrapper_impls/apriori_compiled_kernel.hpp. However, unlike other functions in this file, it is not declared inline, causing linker issues with projects that end up including the header more than once.

[eyalroz]

Not sure how this wasn't caught when linking my example program with two translation units... :-(

Anyway, will be part of the next release (0.6.2).

[jhol]

Yeah I'm not sure either. It works on my host PC building with NixOS, CMake and GCC 11.3. It fails cross-building with Yocto, OE4T, GCC 10.2, and the CMake project is packaged using BitBake.

[eyalroz]

I don't even know what Yocto, OE4T and BitBake even are :-P

Anyway, thanks for reporting this. If you have any non-bug feedback (e.g. likes and dislikes about the wrappers), I'm always interested in more verbose user feedback (by email).

[jhol]

OE4T is "OpenEmbedded for Tegra" - which makes it easier to cross-build things for Tegra boards.

[eyalroz]

@jhol Is it possible to cross-build using OE4T on the "GitHub actions" virtual machines? If you think that's something you could help me set up, we could get GitHub perform that build on every commit.

[jhol]

I think it would make more sense just to break the toolchain down and figure out why it behaves differently, because the only bit we care about is what AArch64 gcc is doing. 99.9% of everything else OE4T does is irrelevant, and would be a headache to maintain.