2 Replies Latest reply on Aug 25, 2011 1:19 PM by oscarbarenys1

    What I expect from FSA IL over CAL..

      Not mentioned by AMD

      Hi there a some requested features long ago for CAL not fixed still which impact OpenCL world so that's what I expect FSA IL to have fixed over CAL:

      *MacOS support.. support it AMD even Nvidia supports CUDA on MacOS in addition to OpenCL and that makes sense to expose features on Macos way before they are standarized and hope this direct low level hardware access isn't negated in Mac platforms now that almost all GPUs are AMD cards.

      *No requeriment to run X11 to run FSAIL kernels similar to CUDA that requires only kernel module to be loaded.. (even CUDA on MacOS supports that feature)i.e. running on a text console..

      *Also no requeriment to have screen attached though that seems fixed on Windows and for other OS above requeriment should fix that..

      *Concurrent kernel execution that seems supported on Cayman but no by CAL..

      *OUT of order queues etc..

      *As FSAIL should support true HW debugging I hope AMD adds some gdb port similar to CUDA gdb port..

      *No watchdog timer so GPUs supporting FSAIL can run simultaneously with windows manager without locking it in large kernels

        • What I expect from FSA IL over CAL..

          Still more wishes:

          *provide some FSAIL backend to LLVM now that PTX LLVM support seems mature and OpenCL in Clang is evolving fast so people can work on developing high level tools that target FSAIL..

          *Fix finally multigpu support on all OSes..


            • What I expect from FSA IL over CAL..

              I hope also FSA brings this over CAL:

              *New in CUDA 4.0 is copy mem between "contexts" (cudamemcpyPeer) even for 1 device.. just add this to completeness to FSA API as CAL doesn't support it.. also allows/exposes a "fast copy path" between devices in a simple call..

              *Expose/document all features in FSA as OCL exposes.. for example seems OCL GL interop is via some CAL GL interop but that support isn't publicily exposed..

              *I hope also DX interop is exposed from version 9 to 11 similar to CUDA a CAL has only DX9 and DX10 interop..

              *New in CUDA 4.0 is creation of "writable texture" (in CUDA called surface) from OpenGL or DX textures using a flag in mapping the texture resource so this in AMD parlance would be allow mapping a graphics tex to UAV interop.. This allows compute to "graphics API" texture directly..