cuda_arch_odr Test

Testing the behavior of linking TU's containing the same kernel compiled with different architectures.

Given two TUs, a.cu and b.cu, where they invoke the same kernel but each is compiled with distinct archictures, which kernel is found when attempting to ODR-use it?

The answer appears to be very complicated and depends on a relationship among:

• Static vs Dynamic linking
• Internal vs External kernel linkage
• Internal vs External linkage of function ODR-using the kernel

Naively, you might expect simply marking the kernel to have internal linkage by annotating it with static would resolve the issue, but it does not. It appears to also depend on the linkage of the enclosing function.

Works?	Linker	kernel() annotation	test()	test() anon namespace
	Static		static	N
	Static		inline	N
Y	Static	static	static	N
	Static	static	inline	N
Y	Static	static		Y
	Static			Y
Y	Dynamic		static	N
	Dynamic		inline	N
Y	Dynamic	static	static	N
	Dynamic	static	inline	N
Y	Dynamic	static		Y
Y	Dynamic			Y

Usage

See the build.sh script which builds/runs the experiment under the various scenarios described in the table.