Closed adamtassier closed 2 months ago
Regarding the crash in the kernel, could you try lowering the amount of allocated GPU memory?
Reduce the 0.96 at the end to something like 0.7.
size_t cptr_buffer_persistent_bytes = static_cast<size_t>(static_cast<double>(availableMem) * 0.96);
I think @JolifantoBambla had the same issue in linux when reserving too much memory.
I had already tried lowering the amount of memory allocated and am still getting this in cuda-gdb:
Thread 22 "SimLOD" received signal CUDA_EXCEPTION_14, Warp Illegal Address.
0x00007fffd3e573c0 in kernel_construct<<<(16,1,1),(256,1,1)>>> ()
cuda block (0, 0, 0) thread (0, 0, 0)
Maybe its due to my T1200 card (the weakest mobile quadro card available). Only support compute capability 75 as well which I changed in the compileProgram opts. to no avail.
Ill try too see if I can get it to run on another PC with a better card.
Hi, while trying to compile on linux (fedora 38) I encountered these compilation problems (maybe in part because GCC13 is a bit stricter):
cuEventQuery
's return type isCUresult
but it was being compared withcudaError::cudaSuccess
instead ofCUresult::CUDA_SUCCESS
(both enum values were 0 anyway)format
calls were ambiguous betweenstd::format
andfmt::format
. Thefmt
library was always compiled even when not used. The use of the__cpp_lib_format
wasn't entirely correct: it itself is defined in<format>
and<version>
. For some reason the macro was defined in one of the files but not the other, leading to both being included.Unfortunately even though everything compiles now it still crashes in the voxel tree construction kernel for me.