This works by targeting C code directly for the GPU via cross-compilation in clang, looks something like this https://godbolt.org/z/hh44a6vKr. The LLVM C library will provide the headers, C library functions, and the kernel that calls the main function, so we only need to compile the DOOM source code targeting my GPU to get an executable. Then it's just a matter of writing a loader utility to launch this program on the GPU using the appropriate vendor runtime libraries (AMD's HSA or CUDA).

Now, it's not entirely on the GPU because we still need the CPU's operating system to do a few things. The loader utility has a thread continuously scanning some memory for 'syscall' requests from the GPU. Currently the only parts running on the CPU for this demo are standard I/O, reading input keys from SDL, and writing the rendered framebuffer to the SDL window. The only modifications made to the DOOM source code itself were done to allow rendering with multiple threads for better FPS.