Inconsistency with Barriers having workgroup sizes above 256 work items.

Syncleus / aparapi

The New Official Aparapi: a framework for executing native Java and Scala code on the GPU.

http://aparapi.com

Apache License 2.0

465 stars 59 forks source link

Inconsistency with Barriers having workgroup sizes above 256 work items. #156

Open CoreRasurae opened 4 years ago

CoreRasurae commented 4 years ago

For some reason, on some systems, the Barriers are not working properly. Inconsistent results are returned for workgroup sizes of 1024, or 512 on some NVIDIA systems, and 8192 workgroup sizes on some CPU systems. The failing unit test is BarrierSupportTest.

CoreRasurae commented 4 years ago

NVIDIA NVS 4200M and NVIDIA GTX 760 pass the BarrierSupportTest on an Ubuntu 18.04 LTS, with nvidia-390, but a NVIDIA RTX 2070 is failing on a Ubuntu 16.04 LTS with Cuda 10-2, also GTX 1060 is failing on Arch linux.

CoreRasurae commented 4 years ago

In fact I've tested NVIDIA GT 1030 on the same system and the issue above is originated by the driver version. If using NVIDIA GT 1030 with nvidia-390 drivers the tests pass. If using the same NVIDIA GT 1030 on the same system, but with nvidia-440 drivers, the above test fails.