[REVIEW] BUG Calling gunrock cmake using explicit -D options, re-enabling C++ tests

[rlratzel]

• Changes to call gunrock cmake using explicit -D options for each arch, similar to a prior version of the CMakeLists.txt, since the update to expand a variable into the arch options was being expanded as a single quoted value with spaces that wasn't recognized, resulting in gunrock being built for only compute 60 support. This caused C++ test crashes.
• Re-enabled C++ tests.

This was tested by building cugraph with this change and running the Pagerank C++ test 1000 times without failures/crashes. Prior to this, the Pagerank C++ test (and others) would crash after 1-50 test run iterations.

The issue and corresponding fix was verified with a standalone reproducer, which can be seen here: https://github.com/rlratzel/cugraph_build_problem

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[ajschmidt8]

@rlratzel, nice job. I created an integration PR to revert these changes in that repo as well. Please let me know if that PR can be merged with this one.

rapidsai/integration#163

[seunghwak]

Wow, so if I am not mistaken, gunrock (built for an outdated architecture) affected the cugraph pagerank/SSSP behavior outside gunrock? Very surprising and great catch!

[rlratzel]

Wow, so if I am not mistaken, gunrock (built for an outdated architecture) affected the cugraph pagerank/SSSP behavior outside gunrock? Very surprising and great catch!

I admittedly don't fully understand why, but when the test binary loads a .so with gunrock code compiled for the old arch, the test binary will occasionally crash in a Thrust call. I image it may not have to be gunrock code, but gunrock exposed the issue with the incorrect arch option. I created a minimal reproducer here to verify the problem and the fix: https://github.com/rlratzel/cugraph_build_problem

I'd like to better understand why we see this particular crash even though the code being built for the old arch isn't even being called, but I figured we should move forward with this as-is if it allows us to re-enable C++ tests sooner.

[codecov-io]

Codecov Report

Merging #1242 into branch-0.17 will not change coverage.
The diff coverage is n/a.

@@             Coverage Diff              @@
##           branch-0.17    #1242   +/-   ##
============================================
  Coverage        56.59%   56.59%           
============================================
  Files               62       62           
  Lines             2564     2564           
============================================
  Hits              1451     1451           
  Misses            1113     1113           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 60af376...3a5ba3f. Read the comment docs.