4 Replies Latest reply on Jun 4, 2013 5:31 AM by himanshu.gautam

    cl_khr_fp16 support?

    yurtesen

      Is cl_khr_fp16 not supported on any AMD hardware? I am getting the following error?

       

      error: can't enable all OpenCL extensions or unrecognized OpenCL extension

        #pragma OPENCL EXTENSION cl_khr_fp16 : enable

                                                       ^

      Thanks,

      Evren

        • Re: cl_khr_fp16 support?
          himanshu.gautam

          "clinfo" should print all supported extensions on your device. Please check out...

          Also, the spec says:

          "If the cl_khr_fp16 extension is not supported, CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF must return 0."

           

           

           

          • Re: cl_khr_fp16 support?
            yurtesen

            Let me rephrase, is this not supported on any AMD hardware at all? Can you give an example to AMD hardware which supports it? I am checking this on some tahiti cards and cpu etc. and nothing seem to be supporting it?

              • Re: cl_khr_fp16 support?
                himanshu.gautam

                Hi yurtsen,

                No AMD GPU currently supports this extension to the best of my knowledge. We have plans on supporting it with future GPU generations. As always, it is a lot of work to enable this extension, so it has its own priority as of now. If anyone direly needs this extension, they can explain their motivations here, which may lead in increasing the priority of this feature.

                  • Re: cl_khr_fp16 support?
                    himanshu.gautam

                    Meanwhile,

                    It is quite possible to implement a custom floating point format yourself. You just have to convert it to IEEE754 floating point format before commencing arithmetic/transcedental operations on them...

                    At the end, you can wrap it back to your format and write back results.

                    Just an idea...

                     

                    And, Converting to IEEE754 would just involve Logical and arithemtic operations that would go down well on GPU....

                     

                    -

                    Bruhaspati