6 Replies Latest reply on Dec 4, 2009 2:22 AM by rahulgarg

    OpenCL and 3870


      Is OpenCL not supported on the 3870 cards because compute shader mode is not supported?

      Why isn't compute shader mode supported? The SIMD engines on the 3870 are essentially the same as the 4870 and the 5870 correct?

        • OpenCL and 3870

          IMHO radeon 3xxx does not have local memory.



          The introduction of a 16KB local data share (that's separate from the texture cache) for each SIMD unit allows data to be shared between threads or stream processors, while there's also a global data share that all SIMDs can access, meaning data can be shared right across the GPU.

          • OpenCL and 3870
            Compute shader mode is a hardware feature that did not exist in the HD38XX line of cards. It was introduced with the HD4870. There are also quite a few differences between the SIMD units, one being data sharing between threads as added by HD48XX and improved greatly in the HD5XXX line. Currently our OpenCL implementation does not use hardware local memory on the supported HD4XXX cards because it is owners write and not a fully generate model like HD5XXX.
              • OpenCL and 3870

                Is there any plan to suppoirt OpenCL on the 3xxx series chips? They were supported in the 1.4 Stream SDK after all...

                  • OpenCL and 3870

                    OpenCL wont be supported in 3xxx series chips as they don't have the required hardware to execute OpenCL kernels.

                      • OpenCL and 3870

                        Surely you can use LLVM to translate higher level GPU IL code into something the 3xxx series can handle?

                          • OpenCL and 3870

                            Well OpenCL requires that the hardware have some capability to organize threads into groups that execute on a SIMD and that the thread group have some sort of synchronization ability. 38xx series doesnt have that ability that at all. There is no way to say "group these 512 threads together". RV7xx is the first series from AMD to feature the so-called "compute shaders" which allow you to form thread groups. So OpenCL isnt going to come to 38xx series cards.

                            (Note : not associated with AMD.)