7 Replies Latest reply on Apr 20, 2011 9:47 AM by Ignus

    Fusion and OpenCL

    Ignus

      Hi!

       

      Maybe I didn't search hard enough, but I haven't really found any developer documentation or guide on programming specifically AMD Fusion processors, so I would like to have some more info on that (I did find OpenCL guides, but not specific guides on optimizing for Fusion).

       

      To be exact, one of my questions:

      * How is the memory between CPU and GPU in Fusion handled in OpenCL? Is a copy/"upload" still done from CPU memory to "GPU memory" (when mapping/unmapping buffers for ex.), or is it possible to just pass a pointer somehow to host-allocated memory to OpenCL calls?

       

      Thank you very much in advance!

       

      --

      Greets,

      I.

       

        • Fusion and OpenCL
          himanshu.gautam

          Thanks for the suggestion. I have reported it to concerned people.

            • Fusion and OpenCL
              Ignus

              Hi!

               

              Any news on this issue?

               

              --

              Greets,

              I.

                • Fusion and OpenCL
                  douglas125

                  Additionally, I think it might be useful if a benchmark software was available to assess the various performance aspects of OpenCL such as:

                   

                  - Data read/write to/from device times;
                  - Coalesced memory read/write times;
                  - Non-coalesced (random) memory read/write times;
                  - Atomic operation times;
                  - Full single-precision (and double when applicable) throughput;
                  - Native single-precision (and double) throughput.

                   

                  This would bring to the scene a much better comparison tool than just run Crysis and see which GPU gets a better FPS (which, honestly, is an absolutely meaningless benchmark to non-players).

                    • Fusion and OpenCL
                      ryta1203

                       

                      Originally posted by: douglas125 Additionally, I think it might be useful if a benchmark software was available to assess the various performance aspects of OpenCL such as:

                       

                      - Data read/write to/from device times; - Coalesced memory read/write times; - Non-coalesced (random) memory read/write times; - Atomic operation times; - Full single-precision (and double when applicable) throughput; - Native single-precision (and double) throughput.

                       

                      This would bring to the scene a much better comparison tool than just run Crysis and see which GPU gets a better FPS (which, honestly, is an absolutely meaningless benchmark to non-players).

                      There are several OpenCL Benchmark suites and other micro-benchmark suites (CAL/IL) available.

                • Fusion and OpenCL
                  MicahVillmow
                  Ignus,
                  You want to read the documentation on 'zero-copy' with the APU's(i.e. Fusion devices).