For fixing the cuGraph test failures with PCG

[vinaydes]

RAFT's RNG class provides methods that fill a buffer with random numbers belonging to various probability distribution functions. For exampleuniform, normal etc. In this methods, multiple instances of random number generator are used in parallel. Each cuda thread gets it own instance of generator. The instance is initialized with three values seed, subsequence and offset. The value for seed is common for all threads and thread id is used for subsequence. The offset is kept 0 for all instances. To fill the buffer, the threads work in grid strided fashion demonstrated by loop below:

  for (size_t i = tid; i < buffer_length; i += total_number_of_threads) {
    buffer[i] = rng.next();
  }

Due to grid-striding of loops, consecutive elements in the buffer are ith element in consecutive subsequences. For example, 10th and 11th elements in the buffer would be 0th element in 10th and 11th subsequences. In the unit tests for cugraph, this scheme seems to introduce a slight bias in certain cases. Easy fix to the issue is to break the lock step increment of individual subsequences.

This PR sets different offset value for each subsequence by setting offset = subsequence. Note that this change has no effect on the period of the random number generator.

This should fix the cuGraph issue rapidsai/cugraph#2266 for now.

[teju85]

Very interesting find @vinaydes ! Can you also check that this fix doesn't break cuML tests, please?

Also, adding @akifcorduk and @MatthiasKohl to see if this doesn't break anything in cuOpt and cugraph-ops respectively.

[teju85]

Changes LGTM.

[vinaydes]

Yes, I'll run cuML tests with this change.

[vinaydes]

Created cuml PR here rapidsai/cuml#4760 after making sure it works locally on my test setup.

[vinaydes]

Also created cuGraph PR here rapidsai/cugraph#2316

[cjnolet]

@gpucibot merge

[MatthiasKohl]

Interesting, this should not break anything in cugraph-ops. Thanks @vinaydes for the fix!