Hi,
to avoid global synchronisations from cudaMalloc/cudaFree, we are using a custom memory allocator (actually, the CUB caching memory allocator).
One downside is that since the memory is not actually freed, “cuda-memcheck --initcheck” cannot flag the reused memory as “uninitialized”.
Is there any way to mark a memory area explicitly as “uninitialized” (short of calling cudaFree) to let cuda-memcheck do its work ?