AccumulateTopology uses uninitialized global memory?

[icoderaven]

Hi @ramakarl

In the process of trying to debug my application today, I ran cuda-memcheck with cuda-memcheck --tool initcheck and discovered that when using AccumulateTopology() I get a whole host of global reads for uninitialized memory. To replicate, this is easily seen if running it on the gPointFusion example cuda-memcheck --tool initcheck ./gPointFusion Is this a bug? I don't see this when using RebuildTopology(), but I'd much rather not write additional logic to store the active brick center locations and then add a new point cloud to those points.

Tracing back the error message it seems like something from within the RadixSort CuDPP methods.

========= Uninitialized __global__ memory read of size 4
=========     at 0x000001c8 in void thrust::cuda_cub::cub::RadixSortScanBinsKernel<thrust::cuda_cub::cub::DeviceRadixSortPolicy<__int64, thrust::cuda_cub::cub::NullType, int>::Policy700, int>(thrust::cuda_cub::cub::NullType*, int)
=========     by thread (374,0,0) in block (0,0,0)
=========     Address 0x7fc3305a69d8
=========     Saved host backtrace up to driver entry point 
=========     Host Frame:/usr/lib/x86_64-linux-gnu/libcuda.so.1 (cuLaunchKernel + 0x2cd) [0x24f90d]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so [0x458172]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so [0x458367]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so [0x4870f5]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so (_ZN6thrust8cuda_cub3cub23RadixSortScanBinsKernelINS1_21DeviceRadixSortPolicyIxNS1_8NullTypeEiE9Policy700EiEEvPT0_i + 0xdf) [0x3c3d0f]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so (_ZN6thrust8cuda_cub3cub15DeviceRadixSort8SortKeysIxEE9cudaErrorPvRmRNS1_12DoubleBufferIT_EEiiiP11CUstream_stb + 0x168e) [0x425d1e]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so (_ZN6thrust8cuda_cub4sortINS0_3tagENS_10device_ptrIxEENS_4lessIxEEEEvRNS0_16execution_policyIT_EET0_SB_T1_ + 0xe6) [0x42b086]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so (_Z7runSortIxEvPT_PjmPK18CUDPPRadixSortPlan + 0x57) [0x432197]
=========     Host Frame:/home/icoderaven/packages/gvdb2/build/shared_cudpp/lib/libcudpp_cu9.2.so (cudppRadixSort + 0x3e) [0xe003e]
=========     Host Frame:/home/icoderaven/packages/gvdb2/install/lib/libgvdb.so (_ZN4nvdb10VolumeGVDB15RadixSortByByteEii + 0xb9) [0x80fd9]
=========     Host Frame:/home/icoderaven/packages/gvdb2/install/lib/libgvdb.so (_ZN4nvdb10VolumeGVDB17ActivateBricksGPUEifNS_9Vector3DFEiNS_9Vector3DIE + 0x357) [0x82f57]
=========     Host Frame:/home/icoderaven/packages/gvdb2/install/lib/libgvdb.so (_ZN4nvdb10VolumeGVDB15RebuildTopologyEifNS_9Vector3DFE + 0x180) [0x8c1b0]
=========     Host Frame:/home/icoderaven/packages/gvdb2/install/lib/libgvdb.so (_ZN4nvdb10VolumeGVDB18AccumulateTopologyEifNS_9Vector3DFEi + 0xa0) [0x8c300]

[ramakarl]

It may be an issue in CuDPP.

[icoderaven]

Even after the move away from cudpp, this error still shows up while using thrust (specifically when casting the device pointer to host in gvdb_cutils.cu) but seems to be not much of an in issue. I'm not sure if there are performance/safety implications, however.

[NBickford-NV]

Ah, that's interesting to hear! I'll probably see if I can take time to run through it and see where Thrust is using uninitialized memory soon. (You probably noticed that I closed most of the issues that were about CUDPP earlier today - I left this one open because it seemed like it could still be an issue, so that's good to hear!)

[icoderaven]

I remember looking at it a couple of weeks back, and there were two instances - First, in createLinearMem initcheck complained about not initialising memory - I think that was benign, not sure how much of an impact memsetting it to 0 would have (didn't profile) Second was the location I mentioned, and I suspect it's got to do with unified memory used in thrust, but I don't have much experience there, so I left it there and didn't update this thread. Your recent activity did cause me to post, as you correctly guessed!

[icoderaven]

BTW, I think I buried this in one of my previous issues that I closed two weeks back, but in AccumulateTopology, when adding a new set of points that would increase the spanning cover and requiring a new root node to be created, the library segfaults. I wrote a quick fix that works for now for me that involves reparenting the root node, but might be worth a second look (and I think should definitely merged into mainline) https://github.com/icoderaven/gvdb-voxels/commit/b3256a921deed84995ab74f1bd726616a1d804e7