>>720526828
you shouldn't have to crowd source anything. the dev/publisher QA should have a representative sample of all supported GPUs. similar GPUs themselves share essentially the same architecture that only differs quantitatively in terms of compute units, memory size, etc. which should be controllable as variables (in the compiler). so, you should be able to either borrow the hardware or just ask QA to run the compilers, and you now have the binaries for all supported chips.
>but what about new chips
they should be backwards-compatible for at least several generations, and if they're not, then it's not your problem. worst case, you can still run the compiler, but it only runs for unsupported chips.
the bigger point being that all the advantages of compiling on the local machine could also be applied to the base (CPU) program, but nobody expects you to compile that on your machine, and fuck you if have AMD 3DNow when the program was only optimized for Intel SSE.