When trying to debug a program built with CUDA 10.1, I get the error cudaErrorLaunchOutOfResources only if I install the higher driver version. This is across different graphics cards, including a GTX 1050, GTX 1060, RTX 2080 Ti (tested on new driver only). Setting the maximum register count to 1 stops the kernel from running out of resources.
Is there a difference in debugging resource usage between different driver versions, or is this a regression which I should report as a bug?