In mesa CI, about 10% of test runs generate a single failure in the subgroups category. A quick look at the history shows the following failures:
Typical output for a failure:
1022 / 1023 values passed
All tests in the recent CI history occure on GT3e systems, even though nearly half of the BDW in CI are GT2.
Looking into this deeper, I really don't think it has anything to do with subgroups. Instead, I suspect that the issue here has something to do with SSBO stores from graphics shader stages. I dug through the bspec and couldn't find anything GT3 specific.