Is it possible to remove LLVM and JIT part of the code and have a vanilla C++/C implementation?
That could be the starting point to write a user-space character device driver (FUSD). Just to emulate enough of the cards to make portable CUDA applications, and applications you could at least test without GPU. No need for it to be fast, as long as it runs it should be fine in this use case.
Is it possible to remove LLVM and JIT part of the code and have a vanilla C++/C implementation?
That could be the starting point to write a user-space character device driver (
FUSD). Just to emulate enough of the cards to make portable CUDA applications, and applications you could at least test without GPU. No need for it to be fast, as long as it runs it should be fine in this use case.