Frame and Fields¶
In oneVPL terminology, a frame (also referred to as frame surface) contains either a progressive frame or a complementary field pair. If the frame is a complementary field pair, the odd lines of the surface buffer store the top fields and the even lines of the surface buffer store the bottom fields.
Frame Surface Management¶
During encoding, decoding, or video processing, cases arise that require reserving input or output frames for future use. For example, when decoding, a frame that is ready for output must remain as a reference frame until the current sequence pattern ends. The usual method to manage this is to cache the frames internally. This method requires a copy operation, which can significantly reduce performance.
oneVPL has two approaches to avoid the need for copy operations. The legacy approach uses a frame-locking mechanism that works as follows:
The application allocates a pool of frame surfaces large enough to include oneVPL function I/O frame surfaces and internal cache needs. Each frame surface maintains a
Lockedcounter, which is part of the
Lockedcounter is initially set to zero.
The application calls a oneVPL function with frame surfaces from the pool whose
Lockedcounter is set as appropriate for the desired operation. For decoding or video processing operations, where oneVPL uses the surfaces to write, the
Lockedcounter should be equal to zero. If the oneVPL function needs to reserve any frame surface, the oneVPL function increases the
Lockedcounter of the frame surface. A non-zero
Lockedcounter indicates that the calling application must treat the frame surface as “in use.” When the frame surface is in use, the application can read but cannot alter, move, delete, or free the frame surface.
In subsequent oneVPL executions, if the frame surface is no longer in use, oneVPL decreases the
Lockedcounter. When the
Lockedcounter reaches zero, the application is free to do as it wishes with the frame surface.
In general, the application should not increase or decrease the
since oneVPL manages this field. If, for some reason, the application needs to
Locked counter, the operation must be atomic to avoid a race condition.
oneVPL API version 2.0 introduces the
which provides a set of callback functions for the
structure to work with frame surfaces. This interface defines
mfxFrameSurface1 as a
reference counted object which can be allocated by oneVPL or the application. The
application must follow the general rules of operation with reference counted
objects. For example, when surfaces are allocated by oneVPL during
MFXVideoDECODE_DecodeFrameAsync() or with the help of
MFXMemory_GetSurfaceForEncode(), the application must call the corresponding
function for the surfaces that are no longer in use.
Note that the
Locked counter defines read/write access policies
and the reference counter is responsible for managing a frame’s
The second approach to avoid the need for copy operations is based on the
mfxFrameSurfaceInterface and works as follows:
oneVPL or the application allocates a frame surface and the application stores a value of reference counter obtained through
The application calls a oneVPL function with the frame surface. If oneVPL needs to reserve the frame surface it increments the reference counter through the
mfxFrameSurfaceInterface::AddRefcall. When the frame surface is no longer in use by the oneVPL it decrements reference counter through the
mfxFrameSurfaceInterface::Releasecall which returns the reference counter to the original value.
The application checks the reference counter of the frame surface and when it is equal to the original value after allocation, it can reuse the reference counter for subsequent operations.