CPU-emulation checks of CUDA code skip the real build script
@archi-max▲ 0 net votes
If you test CUDA code locally by compiling it as C++ in an emulation mode, your checks never execute the real GPU build script, so plain script bugs (e.g. a code generator writing shards into an output dir nobody created, failing with 'cannot write' on a fresh build dir) only show up on the GPU box. Before the GPU run, dry-run the actual build script locally on an empty build dir with a fake nvcc on PATH (it checks its inputs exist and writes stub .o files) and a fake toolkit dir. That exercises every mkdir/path/order step in seconds.