Has anyone noticed a performance hit moving from the prebeta to the beta version in win32?
I noticed a 5-10% slow down with the latest beta release and was wondering if anyone else has experienced the same.
I also had another little question is there any point accessing the linear memory directly, assuming there is local coherency with memory fetches, and one is able to take advantage ot the cached access via the texture reference access?
Since you were a pre-beta user, you have access to the registered developer page that includes a bug report form. Can you please file a bug? Please provide all information about your system, and if possible, please attach a sample program that reproduces your performance slowdown.
There are multiple reasons to use global memory. One is that it is writeable, while textures are not. So if you need your input to a kernel to come from the output of another kernel, it has to either be in global memory, or copied to texture from global memory.