I am the developer of an OpenCL 1.2 application that has been used for several years by thousands of users on a wide variety of platforms (Windows, Mac, Linux) and OpenCL implementations (AMD, NVidia, Intel, etc.). Our application no longer behaves correctly on AMD Adrenalin drivers 23.7.1 and later (AMD-APP 3570.0) when run on Radeon RX 6000 Series GPUs (and perhaps other AMD GPUs). The application runs fine on the same hardware with the AMD Linux driver, and on all other OpenCL platforms.
After extensive testing, our development team has narrowed the issue down to a single large OpenCL device kernel that performs an image processing task. The 23.7.1 OpenCL compiler appears to be doing very aggressive optimization, and we suspect that these optimizations may be causing the issue. The kernel runs fine when optimizations are disabled, or when the work group size is set to 1. Otherwise the kernel exhibits bizarre behavior, such as skipping loop iterations.
Is there any documentation available regarding what changes were made to the OpenCL compiler between Adrenalin drivers versions 23.5.2 and 23.7.1? I'm not prepared to post the kernel here at this time, but could provide it via another channel if helpful. We would appreciate any other advice on how to resolve this problem with our application when run on the recent AMD Windows drivers.
