@canonizer Have you looked into at the performance impact of this? If not, I can use this as a case study for my new benchmarking code, but that will probably mean this won't make the 1.11.0 release.

This implementation looks good to me, and the performance difference is negligible (<1% difference on GV100/GA100 from @canonizer's benchmarks).

Though now I'm curious whether this is actually expected behavior for users, or if it's better to just let them pre-condition their data when they want these to be treated equivalently. I'll do some polls and confirm whether or not this is something folks expect.

Looks like the stl already does stable sorts this way, so that's a nice, definitive answer :)

@canonizer Can you update the DeviceRadixSort documentation to specify this behavior, and add a regression test to test/test_device_radix_sort.cu? I'll take a look at the other backends in Thrust to see if we can provide this guarantee there, too.

I will update the documentation and add the test.

@canonizer Any updates on this? I'd need to land this by Friday to make the 1.11.0 release, otherwise we can slip to 1.12.0.