Make shared libs for CUDA modules depend on host and device checksums
When building a CudaModule, the shared lib is not re-linked if the cache is hit for both the host and the device module. However, this does not mean that the host module was also linked against that very device module we found in cache, which is what we want. To make sure we don't accidentially reuturn a shared lib with wrong linkage when we hit the cache twice, the name of the shared lib contains the hex checksums for both the host and device code. This allows shared libraries for the same host module linked against different device modules to co-exist in cache without the need to re-compile the object files.
Loading
Please sign in to comment