• AMD APP SDK 2.6 available for download - first glimpse of OpenCL 1.2 [Update]

    AMD APP SDK 2.6 component versioningJust a quickie this time. AMDs developer page has been updated as of now with the checksums of AMD APP SDK 2.6. The files linked there as of this writing are still 2.5, but if you alter the link to point to 2.6 instead of the linked SDK, you get the brand new APP SDK 2.6. But please try and visit their site first - they should have updated it any minute now. In a blog-entry from a couple of days ago, AMDs Mark Ireton apart from revealing today's launch date for the 2.6 SDK, already outlined some of the changes, AMD APP 2.6 will bring. Most notably, Open CL 1.2 will be on the support list for 2012 where AMD is commited to fully support the most important open GPU-Compute standard on both CPUs and GPUs. The APP SDK 2.6 is to include previews on some of OpenCL 1.2's core features according to AMD.


    Provided that the system requirements did not change too drastically compared to APP SDK 2.5, the Linux releases will be suitable for openSUSE 11.x, Ubuntu 11.04, Red Hat Enterprise Linux 6.x. If you need to know, what version SDK, Runtime, App Profiler, Kernel Analyzer etc. you'll be getting when you install 2.6-SDK, just take a look at the screenshot above. On a sidenote: The APP SDK 2.6 prompts you per default to decompress to c:\AMD\SUPPORT\streamsdk_2-6-RC3_win764 [my bold]

    According to the included „Getting Started Guide“, the following processors are supported by AMDs APP SDK 2.6:

    List of supported processors for AMD APP SDK 2.6 (from AMDs APP-SDK „Getting-Started Guide“).Additionally, SIMD instruction sets like SSE3 and AVX found on modern CPUs are supported and auto-vectorized as well as compiled to take advantage of XOP and FMA4 - AMD exclusives on the recent AMD FX-CPUs (Bulldozer). Crossfire, AMDs Multi-GPU solution, will need some extra love from developers to be properly utilized. Here's a quote from the APP SDK FAQ (also included in the downloads linked above):

    Q: Does the AMD APP SDK v2.6 with OpenCL 1.1 support work on multiple GPUs (ATI
    CrossFire)?

    A: OpenCL applications can explicitly invoke separate compute kernels on multiple compatible
    GPUs in a single system. The partitioning of the algorithm to multiple parallel compute
    kernels must be done by the developer. It is recommended that ATI CrossFire be turned off
    in most system configurations so that AMD APP applications can access all available GPUs
    in the system.

    ATI CrossFire technology allows multiple AMD GPUs to work together on a single graphicsrendering
    tasks. This method does not apply to AMD APP computational tasks because it is
    not compatible with the compute model used for AMD APP applications.“

    [Update Dec. 23rd, 2012]:
    In the meantime, AMD has released the Radeon HD 7970 as well as Catalyst 11.12 WHQL drivers. There has been an information update as to what functions of OpenCL 1.2 are supported. The following list ist from AMDs leaked press deck for Radeon HD 7970 launch:

    OpenCL 1.2 core function developer preview:
    • Host access flags for memory objects enables more efficient buffer handling
    • Pattern based GPU buffer and image initialization using 32-bit scalar or float4 patterns eliminates need for certain buffer/image transfers
    • Memory objects migration supports transfer of buffer prior to need
    • New generalized image creation API
    • Enhanced image/buffer map operations
    • OpenCL 1.2 CPU device partition including partition of a CPU after addition to a context
    • Generalized 1D and 2D images, image arrays, and image<-> buffer interop

    OpenGL 4.2 production - highlights:
    • Enable shaders with atomic counters
    • Save cycles by capturing GPU-tessellated geometry
    • Efficient shader processing
    Comments 1 Comment
    1. Anonymous Coward's Avatar
      Anonymous Coward -
      Also, what should be included in the FAQs:
      Q: Does the AMD APP SDK v2.6 with Catalyst 11.12 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.5 with Catalyst 11.11 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.5 with Catalyst 11.10 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.5 with Catalyst 11.9 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.5 with Catalyst 11.8 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.5 with Catalyst 11.7 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.4 with Catalyst 11.6 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.4 with Catalyst 11.5 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: No.
      Q: Does the AMD APP SDK v2.4 with Catalyst 11.4 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: Sometimes.
      Q: Does the AMD APP SDK v2.4 with Catalyst 11.3 execute kernels simultaneously on multiple GPUs in a single OpenCL context?
      A: Yes, but everything else is slow.
      Q: Does AMD care that they've launched 8 months of broken drivers?
      A: Seems not.
      Q: Can AMD release a stable driver for an existing API before adding lots of extra bells and whistles?
      A: Seems not.