3 Replies Latest reply on May 22, 2015 4:42 AM by dipak

    Kernel Workspaces

    paign10

      I was developing a kernel for a while on my machine (R9 270X, Catalyst 14.12) and everything run smoothly. But then I got to execute it on another machine (HD 6520G), just to get the following error: clEnqueueNDRangeKernel (CL_INVALID_WORK_GROUP_SIZE).

       

      These were the workspaces I was using: global (52, 4096), local (16, 16). So, the problem was that global[0] was not a multiple of local[0].

       

      The question now is, why was I able to run the kernel on my machine? Is this a feature, or a bug?

        • Re: Kernel Workspaces
          nou

          OpenCL 2.0 relaxed limitation where global work group size must be multiple of local workgroup group size. AFAIK R9 270X does support 2.0 but 6520 does not.

          • Re: Kernel Workspaces
            dipak

            Thanks for reporting this. It seems that "not divisible work-groups" are running fine on GCN cards but not on pre GCN i.e. VLIW/TeraScale cards. As I checked on the following cards, my observation was as below:

            HD 6520G --> failed

            HD 6850 ---> failed

            HD 7870  ----> worked

            R9 290X ---> worked

             

            So, though the relaxation is an OpenCL 2.0 feature, however, it seems that the feature is working fine on any GCN cards. I'll check with concerned team and file a report, if needed.

             

            Regards,