Follow us on Twitter
Follow us on FacebookWindows 8.1 and VS2013 bring GPU computing updates to Direct3D and C++ AMP
@ 2013/07/03Windows 8.1 is bringing a new incremental update to the driver model to WDDM 1.3, which will enable incremental new GPU computing functionality. One of the important pieces is the ability to "map default buffer" (which I will call as MDB), which should be particularly interesting for compute shaders running on APUs/SoCs which combine CPU and GPU on a single chip.
We can explain the feature as follows. In a typical discrete card, GPU has it's own onboard graphics memory. The application allocates memory on the GPU buffer, and the shaders read/write data from this memory. The buffers allocated in GPU memory are called "default buffers" in Direct3D parlance. Let us assume the GPU shader has written some output that you want to read on the CPU. Currently this is done in multiple stages. First, the application allocates a "staging buffer", which is allocated by the Direct3D driver in a special area of system memory such that the GPU can transfer data between the GPU default buffers and staging buffers over the PCI Express bus efficiently. GPU copies the data from GPU buffer to the staging buffer. The CPU then issues a "map" command that allows the CPU to read/write from the staging buffer. This multi-stage process is inefficient for APUs/SoCs where the GPU shares the physical memory with the CPU. In Direct3D 11.2, the staging buffer and the extra copy operation will no longer be required on supported hardware and the CPU will be able to access the GPU buffers directly. Thus, MDB will be a big win for many GPU computing scenarios due to the reduced copy overhead on APUs/SoCs.
We can explain the feature as follows. In a typical discrete card, GPU has it's own onboard graphics memory. The application allocates memory on the GPU buffer, and the shaders read/write data from this memory. The buffers allocated in GPU memory are called "default buffers" in Direct3D parlance. Let us assume the GPU shader has written some output that you want to read on the CPU. Currently this is done in multiple stages. First, the application allocates a "staging buffer", which is allocated by the Direct3D driver in a special area of system memory such that the GPU can transfer data between the GPU default buffers and staging buffers over the PCI Express bus efficiently. GPU copies the data from GPU buffer to the staging buffer. The CPU then issues a "map" command that allows the CPU to read/write from the staging buffer. This multi-stage process is inefficient for APUs/SoCs where the GPU shares the physical memory with the CPU. In Direct3D 11.2, the staging buffer and the extra copy operation will no longer be required on supported hardware and the CPU will be able to access the GPU buffers directly. Thus, MDB will be a big win for many GPU computing scenarios due to the reduced copy overhead on APUs/SoCs.
