I am trying to load an OpenCL program using clCreateProgramWithBinary. Here is what is happening:
1.) It works for the CPU device (all other things the same).
2.) It worked for the GPU device until recent driver update.
3.) The error message returned by the GPU is "out of host memory".
4.) Sometimes after the error also the driver crashes.
5.) The GPU program works fine, when building directly from source (otherwise I would have had no binary loading to complain about in the first place.)
6.) It takes 30 seconds to build from source and previously 1-2s to load prebuilt from the file . This adds quite some delay to the loading of the app.
The CPU device binary is about 1 MB and the GPU binary is about 5MB in size.
I am attaching the GPU binary. It should be failing on latest AMD drivers (AMD R9 200). Any ideas are welcome.