[BUG] fused_concatenate_kernel can cause illegal memory access

[abellina]

When concatenating via the fused_concatenate_kernel we see an issue where its strided loop can overflow, even if the output number of rows is < 2B. Here is a small scala repro case:

import ai.rapids.cudf.{ColumnVector, Cuda, Scalar, Table}
val s = Scalar.fromByte(1.toByte)
val tbls = (0 until 5).map { _ => new Table(ColumnVector.fromScalar(s, 250 * 1000 * 1000)) }
Table.concatenate(tbls:_*)
Cuda.DEFAULT_STREAM.sync

Will produce:

ai.rapids.cudf.CudaException: an illegal memory access was encountered
  at ai.rapids.cudf.Cuda.streamSynchronize(Native Method)
  at ai.rapids.cudf.Cuda$Stream.sync(Cuda.java:111)
  ... 47 elided

After adding printfs, the grid size is very large and the stride update: https://github.com/rapidsai/cudf/blob/branch-22.04/cpp/src/copying/concatenate.cu#L200 can cause the size_type index to become negative.

It seems the simplest fix is to change output_index to be std::size_t (https://github.com/rapidsai/cudf/blob/branch-22.04/cpp/src/copying/concatenate.cu#L200).

Thanks to @nvdbaranec and @jlowe for debugging this with me.

[abellina]

I'll have this up for review today. I added a test that reproduces it and building cuDF with the solution.

[abellina]

I put up the PR #10344, to address the fused_concatenate_kernel specifically.

I've seen the same bug in other kernels. I am looking for input on how that should be handled. Does it seem we want 1 PR that handles the issue as a whole?

For example:

• fused_concatenate_string_offset_kernel, fused_concatenate_string_chars_kernel
• get_json_object_kernel