I get this error with when calling the C++ cudnnLRNCrossChannelForward api.
Illegal access to address (@global)0x1020e962860 detected.
Program received signal CUDA_EXCEPTION_1, Lane Illegal Address.
[Switching focus to CUDA kernel 0, grid 20, block (0,0,0), thread (0,0,0), device 0, sm 0, warp 1, lane 0]
0x00000000162ac6d8 in void cudnn::detail::lrnForward_evenC<5, float, float>(cudnn::detail::LrnForwardParams<float, float>)
This error is thrown from cuda-gdb debugger when cuda memcheck is turned ON.
Could you please give me some advice on how to debug this ?
I have a hunch that there may be an out of memory issue happening here, since the problem does not occur with smaller networks.
In general how do we debug errors thrown from cudnn api, as we don’t have access to source code or the debug symbols.