gunrock: need eyeballs on OS X Gunrock failure on prerelease ctest

@yzhwang or others, just want a pair of eyeballs on this to tell me if you’ve seen it before. If not, I’ll dig in. This is OS X with CUDA 8 prerelease. It built fine.

(Also I note NOT COMPILED WITH COMPATIBLE PTX VERSION FOR DEVICE 0, which has never occurred before. Problem with cmake? @neoblizz ?)

test 1
      Start  1: TEST_BFS

1: Test command: /Users/jowens/Documents/working/gunrock-build/bin/bfs "market" "/Users/jowens/Documents/working/gunrock/simple_example/bips98_606.mtx" "--undirected" "--src=0"
1: Test timeout computed to be: 9.99988e+06
1: Loading Matrix-market coordinate-formatted graph ...
1:   Reading directly from stored binary CSR arrays ...
1: Done reading (0s).
1:
1: Degree Histogram (7135 vertices, 30380 edges):
1:     Degree   0: 541 (7.58%)
1:     Degree 2^0: 626 (8.77%)
1:     Degree 2^1: 2181 (30.57%)
1:     Degree 2^2: 3085 (43.24%)
1:     Degree 2^3: 621 (8.70%)
1:     Degree 2^4: 80 (1.12%)
1:     Degree 2^5: 0 (0.00%)
1:     Degree 2^6: 1 (0.01%)
1:
1: [/Users/jowens/Documents/working/gunrock/gunrock/app/bfs/bfs_enactor.cuh, 1865 @ gpu 0] BFSEnactor cudaBindTexture labels_tex_ref failed (CUDA error 18: invalid texture reference)
1: [/Users/jowens/Documents/working/gunrock/tests/bfs/test_bfs.cu, 388 @ gpu 0] BFS Enactor Init failed (CUDA error 18: invalid texture reference)
1: Source vertex: 0
1: Using 1 GPU: [ 0 ].
1: NOT COMPILED WITH COMPATIBLE PTX VERSION FOR DEVICE 0
1:
1:  [BFS] finished.
1:  avg. elapsed: 0.0000 ms
1:  iterations: 0
1:  load time: 165.9155 ms
1:  preprocess time: 0.0000 ms
1:  postprocess time: 0.0000 ms
1:  total time: 169.1387 ms
1:
 1/11 Test  #1: TEST_BFS .........................***Failed  Required regular expression not found.Regex=[CORRECT
]  0.18 sec

About this issue

  • Original URL
  • State: closed
  • Created 8 years ago
  • Comments: 21 (21 by maintainers)

Most upvoted comments

Thanks @neoblizz . You are right that the problem was due to multi-pass SSSP. But actually you do not need to disable it since the multi-pass works fine. The reason it failed is because we only used simple regex which assume that the source node is 0 (and when you set the source array to NULL, it will be set to 0 by default) and when I was testing multi-pass feature, I accidentally put 1 there, so it used node 1 as the source node and gave the incorrect answer. The simple fix would be to set the source array to {0}. I will do it on dev_debug and cherry pick it to pre-release. But thanks for quickly spotting where the problem is!