I’ve been running into problems trying to break into a kernel launched via the graphs API using the legacy debugger on an older GTX 970. I’ve been testing with the simpleCudaGraphs example. The program compiles and runs without issue, but when I attempt to step in the the kernel I get the following output:
Stepper Failed: Unable to find function of PC 0xffffffffffffffff
and VS displays “Frame not in module”.
I have no problems debugging kernels launched normally with the 970. I’ve also tested the simpleCudaGraphs examples with the next-gen debugger on another PC with a newer Turing card and have had no issues. The executable is being built with device debug info (-G).
Just wondering if the legacy debugger is meant to support this situation or not. I am using VS 2019 with CUDA toolkit 11.3.1 and NSight Visual Studio Edition 2021.1.1.