4 Replies Latest reply on Aug 12, 2010 5:16 PM by Raistmer

    beta support of 4xxx GPUs in SDK 2.2

    Raistmer
      what exactly in beta?

      Device Type: CL_DEVICE_TYPE_GPU
      Device ID: 4098
      Max compute units: 10
      Max work group size: 256
      Max clock frequency: 750Mhz
      Address bits: 32
      Max memory allocation: 134217728
      Image support: No
      Max size of kernel argument: 1024
      Alignment (bits) of base address: 32768
      Minimum alignment (bytes) for any datatype: 128
      Cache type: None
      Cache line size: 0
      Cache size: 0
      Global memory size: 536870912
      Constant buffer size: 65536
      Max number of constant args: 8
      Local memory type: Global
      Local memory size: 16384
      Profiling timer resolution: 1
      Device endianess: Little
      Available: Yes
      Compiler available: Yes
      Execution capabilities:
      Execute OpenCL kernels: Yes
      Execute native function: No
      Queue properties:
      Out-of-Order: No
      Profiling : Yes
      Platform ID: 02B4D40C
      Name: ATI RV770
      Vendor: Advanced Micro Devices, Inc.
      Driver version: CAL 1.4.739
      Profile: FULL_PROFILE
      Version: OpenCL 1.0 ATI-Stream-v2.2 (302)
      Extensions: cl_khr_icd cl_amd_fp64 cl_khr_gl_sharing cl_amd_device_attribute_query cl_khr_d3d10_sharing

      When bold No become "Yes" ? Will image support be added for 4xxx cards and at what SDK release? What about native functions ? What about binary kernels, will they work for HD4xxx ?
        • beta support of 4xxx GPUs in SDK 2.2
          MicahVillmow
          Raistmer,
          Because of hardware limitations, there is no plan on adding images to the 4XXX series of cards.
            • beta support of 4xxx GPUs in SDK 2.2
              Raistmer
              Originally posted by: MicahVillmow

              Raistmer,

              Because of hardware limitations, there is no plan on adding images to the 4XXX series of cards.


              Should this be read as "Although HD4xxx has shared memory and texture cache we never will expose these hardware resources via OpenCL" ?

              What hardware limitation prevents image support? In what part OpenCL image differs from let say Brook's ordinal (not gather and not scatter) stream so much that it can't be supported by RV770 hardware?

              Such approach just throw out all fast GPU memory besides registers themselves ...
            • beta support of 4xxx GPUs in SDK 2.2
              MicahVillmow
              Raistmer,
              If image support only was for the streaming model, then it could be supported on HD4XXX. However, there are many issues with the OpenCL spec that 4XXX does not comply with. For example, on 8XX hardware there is support for scatter based typed surfaces(write_only images), but this does not exist on 7XX hardware, so write_only images can not be supported in hardware, and that is required for 'image' support.
              • beta support of 4xxx GPUs in SDK 2.2
                Raistmer
                Maybe it's possible to open access to texture cache on HD4xxx via some AMD-specific OpenCL extensions? Let they be more limited then OpenCL image, but at least for some algorithms fast cache memory could be used. It's a shame to have fast memory onboard and have no ability to use it, especially for memory-bound apps like I working on ...