Yesterday I was fixing a bunch of misaligned errors(https://devtalk.nvidia.com/default/topic/537588/nvcc-arch-sm_20-causes-access-violations-in-shared-memory), with them all corrected I am now left with a runtime error which only occurs when not using nsight memory checker.
At runtime the error causes a break of Unknown Error (30) at the CUT_CHECK_ERROR after executing the kernel which I was fixing yesterday.;
CUT_CHECK_ERROR("avoid_pedestrians0 failed"); GPUFLAME_avoid_pedestrians<<<grid, threads, sm_size>>>(d_agents, d_pedestrian_locations, d_pedestrian_location_partition_matrix, d_rand48); >>CUT_CHECK_ERROR("avoid_pedestrians1 failed");
Using the nsight analysis and not the memory debugger it picks up a single error
Call ID|Name |CudaError|Start Time (µs)|Duration (µs)|Context ID|Process ID|Thread ID 2263 |cudaDeviceSynchronize|30 |12,862,697.811 |4,551.503 |31 |6244 |6384
I’m unsure how to trace that error back to a source, from looking through the nsight guide It doesn’t make this clear either.
Furthermore visual studio states the error as;
First-chance exception at 0x753dc41f (KernelBase.dll) in PedestrianFLAMEGPU.exe: Microsoft C++ exception: cudaError_enum at memory location 0x0018f778..