AMD has just published a new version of its OpenCL SDK called AMD APP SDK (APP for Accelerated Parallel Processing).
You can download the OpenCL SDK here: AMD APP SDK downloads. More information on AMD APP SDK HERE.
Time for a little update of GPU Caps Viewer 😉
AMD APP SDK 2.4 changelog:
- Improved OpenCL runtime performance:
- Zero copy transfer of data between CPU and GPU for APUs (Windows only)
- Improved PCIe transfer speed
- Improved kernel launch times
- Support for AMD A-series devices
- Support for:
- AMD C-30 Accelerated Processor with AMD Radeon™ HD 6250 graphics card
- AMD C-50 Accelerated Processor with AMD Radeon™ HD 6250 graphics card
- AMD E-240 Accelerated Processor with AMD Radeon™ HD 6310 graphics card
- AMD E-350 Accelerated Processor with AMD Radeon™ HD 6310 graphics card
- AMD Radeon™ HD 6670 graphics card
- AMD Radeon™ HD 6570 graphics card
- AMD Radeon™ HD 6450 graphics card
- Support for CPU/x86 images (Enables the support for image formats, as described in the Khronos OpenCL Specification, to be run on the x86 CPU)
- Support for cl_khr_fp64 on the CPU.
- Cached reads (-fno-alias)
- Binary image format improvements
- DXVA buffer sharing
- FFT performance improvements
- Uninstallation of SDK 2.4 or its components is now done through the Catalyst Install Manager (CIM). See the Installation Notes for a description of the procedure.
- For the AMD APP Profiler v2.2
- Support for OpenCL™ application summary
- Support for profiling with AMD Radeon™ HD6900 series GPUs (requires AMD APP SDK v2.4 and AMD Catalyst 11.2 or newer)
- Improved Session List Explorer
- Improved user documentation (now also accessible online in Documentation)
- For the AMD APP KernelAnalyzer v1.8:
- Support for AMD Catalyst 11.1 – 11.3
- Support for setting the OpenCL compiler build options
- Improved analysis for AMD Radeon™ HD 6970 graphics cards
- Support for displaying x86 assembly code
i don’t get it. Isn’t the runtime already shipped with the drivers? How can the SDK improve CL performance.
Yes AMD APP SDK is shipped with the drivers (you have two kind of downloads for AMD Catalyst: with and without OpenCL). But you can also download the OpenCL SDK in a separate way. The current Cat11.3 drivers come with AMD APP SDK 2.4 RC1. Then if you install AMD APP SDK 2.4 over Cat11.3, you will enjoy the new performance of the OpenCL runtime.
Hmm…
“Zero copy transfer of data between CPU and GPU for APUs”
I think this is the same as NVIDIA CUDA 4.0’s
“No-copy pinning of system memory, a faster alternative to cudaMallocHost()”
This is a very good improvement, I always thought the extra copy was not needed.
I smell a OpenCL 1.2 in the works to add some of the other goodies from CUDA 4.0 like multithreading, GPU-GPU copy etc.