ROCm/rocPRIM

Please add gfx906 to support Vega20 in cmakelists

Closed this issue · 10 comments

Can you please add gfx906 to support Vega20 in cmakelists @ https://github.com/ROCmSoftwarePlatform/rocPRIM/blob/master/CMakeLists.txt#L70

Did you run it on Vega20?

Yes, we run it on Vega20 also and hence we need Vega20 support. Please add.

Sure. Just wanted confirmation it works ok. I'll add it to develop today.

Check out https://github.com/ROCmSoftwarePlatform/rocPRIM/tree/develop-gfx906. It will be merged into develop when there's a public release of ROCm with support for gfx906.

Superb...Let me use develop-gfx906 for Vega20 and comeback with observations.

Thanks @jszuppe ,
I tested on Vega20 with develop-gfx906 branch, tests are running well.
But few failures observed. Looks like issue is due to compiler, but not sure.

91% tests passed, 8 tests failed out of 90

Total Test time (real) = 246.16 sec

The following tests FAILED:
9 - rocprim.hc.block_radix_sort (Failed)
24 - rocprim.hc.device_segmented_radix_sort (Failed)
43 - rocprim.hip.block_radix_sort (Failed)
51 - rocprim.hip.device_merge_sort (Failed)
53 - rocprim.hip.device_radix_sort (OTHER_FAULT)
58 - rocprim.hip.device_segmented_radix_sort (Failed)
80 - hipcub.DeviceRadixSort (Failed)
85 - hipcub.DeviceSegmentedRadixSort (Failed)
Errors while running CTest

We would need more info, but I doubt we can help without environment and gpu.

Thanks @jszuppe
Looks like issue is NOT specific to Vega20. Its observed in Vega10/MI25 also and we have logged an internal ticket, found to be compiler issue.

I'll rebase develop-gfx906 on current develop later today.

Are all of these issues solved, including test 51?