Table of contents
Collapse the table of content
Expand the table of content

Supporting Multiple Processors

Last Updated: 2/14/2017

User-mode display drivers on multiple-processor computers can let the Microsoft Direct3D runtime handle multiple-processor optimizations, or the drivers can perform their own multiple-processor optimizations.

Runtime-Handled Multiple-Processor Optimizations

The multiple-processor optimizations that are handled by Direct3D runtime are enabled only on drivers that support the LockAsync, UnlockAsync, and Rename functions. These functions enable the multiple-processor optimizations to work well with applications that frequently lock dynamic resources. The LockAsync and UnlockAsync functions--along with the GetQueryData function--must be reentrant on drivers that expose a DDI version of 0x0000000B or greater. The driver returns the DDI-version value in the DriverVersion member of the D3D10DDIARG_OPENADAPTER structure in a call to the driver's OpenAdapter function. When the runtime calls a driver function in a reentrant manner, one thread can execute inside that function while another thread that references the same display device executes inside of another driver function.

The Direct3D runtime uses multiple-processor optimizations in some situations to offload work to a separate processor and improve computer performance. When multiple-processor optimizations are enabled, an additional software layer is added between the Direct3D runtime and the user-mode display driver. This software layer intercepts all calls that the Direct3D runtime would otherwise make to the user-mode display driver's functions.

Instead of calling the user-mode display driver directly, the software layer queues commands into batches that a worker thread asynchronously processes. However, the software layer cannot batch all calls that are made to the user-mode display driver's functions. In particular, the software layer cannot batch calls to functions that return information (for example, CreateResource). When the software layer must call one of these types of driver functions, it flushes all queued commands through the worker thread, and then the software layer calls the driver function on the main application thread.

Driver-Handled Multiple-Processor Optimizations

If a driver will perform its own multiple-processor optimizations, it must not implement LockAsync, UnlockAsync, and Rename functions. In this situation, the driver must call the pfnSetAsyncCallbacksCb function to notify the runtime whether the runtime will start or stop receiving calls to the runtime's callback functions from a worker thread.

If the driver performs its own multiple-processor optimizations, it should follow the same policy that the Direct3D runtime uses when it determines to enable multiple-processor optimizations. This policy enables fair sharing of system resources across all processes. In particular, the driver should disable multiple-processor optimizations in the following situations:

  • The application runs in windowed mode.

  • The computer contains only one processor (or processor core); the driver should disable optimizations on single-processor computers with hyper-threading.

  • The application requested that no multiple-processor optimizations be enabled, or the application uses software-vertex processing; this information is passed to the driver's CreateDevice function.

If vendors want to enable multiple-processor optimizations in one of these situations, they should first contact Microsoft.

Send comments about this topic to Microsoft

© 2017 Microsoft