PageRank radixsort error on directed graphs. about gunrock HOT 3 CLOSED

This issue might not related to cub sort actually. Some invalid reads are reported by memcheck. And the answers are not always correct even for the undirected graph, need to check in details.

from gunrock.

@ydwu Could you fix this? On a different topic, I'm having exactly the same error in my code using CUB.
I just realized that whenever I include thrust/sort.h in my code as well, I get this error. By not including that, everything looks fine and works perfectly.

from gunrock.

Hi @sashkiani, this problem was originally observed running bin/pagerank market directed_graph.mtx, but now I actually cannot reproduce the same issue when I switch back using CUB's sort (for safety I changed it to mgpu::sort in current master branch). Changing pr_enactor.cuh L393 back to:

util::CUBRadixSort<Value, VertexId>(
        false, graph_slice->nodes,
        problem->data_slices[0]->d_rank_curr,
        problem->data_slices[0]->d_node_ids);

Running bin/pagerank does not incur any problem for all datasets we have. I guess it was fixed somehow near commit e7dbca5 but not sure why.

Wired enough though, I did more tests and observe that in gunrock simple interface test if I run BFS and/or BC before calling PageRank, this error will occur, but other two primitives SSSP and CC does not have any impact. However, if I comment out L29 and L30, this error disappear. In my case, I did't include any other library such as thrust/sort.h, and I didn't touch any input of PageRank function call, can you find any connections with your situation?

Also I recall that "CUDA error 9" is possibly because of CUDA launch kernel parameter settings.

from gunrock.