Package org.lwjgl.opencl

Class INTELDeviceSideAVCMotionEstimation

  • public final class INTELDeviceSideAVCMotionEstimation
extends java.lang.Object
    Native bindings to the intel_device_side_avc_motion_estimation extension.

    Video motion estimation (VME) is defined as of set motion estimation operations that are used to determine the motion vectors, intra estimation angles and macroblock partitioning combination that best describe the transformation to the source macroblock, from blocks in one or more previous reference pictures (inter-prediction), or from other blocks in the same source picture (intra-prediction). It does this by searching for spatial and temporal patterns on the current and various forward and backward reference pictures.

    The goal of this extension is to provide programmers with a fine- grained interface to the AVC VME media sampler in Intel graphics processors. It describes the specification of low-level built-in functions, callable from OpenCL kernels, that facilitate the programming of the VME media sampler to evaluate specific AVC motion estimation operations. If only a coarser-level interface at the level of built-in kernels suffices, then the Intel vendor extensions cl_intel_motion_estimation and cl_advanced_motion_estimation may be considered.

    Built-in functions are defined for all the major operations of the VME media sampler. The major operations of the AVC VME media sampler in Intel Graphics Processors can be described as follows:

    1. Integer motion estimation (IME)

    Perform motion estimation on a given source macroblock in a source image over a single or dual reference window in a reference image, at full-pixel resolution, to determine the best integer motion vectors and their associated distortions, and the best macroblock shape partitioning combination.

    2. Motion estimation refinement (REF)

    Perform refinement operations on the results of IME. The two sub- operations are:

    Fractional motion estimation (FME)

    Perform sub-pixel refinement on the results of an IME operation. half-pixel (HPEL) or quarter-pixel (QPEL) refinements are performed to determine the best sub-pixel motion vectors and their associated distortions.

    Bidirectional motion estimation (BME)

    Perform bidirectional refinement on the results of an IME operation using two reference images to check if the bidirectional mode using two references yields lesser distortions. An FME can optionally be performed implicitly as part of a bidirectional refinement.

    3. Skip and Intra check (SIC)

    Performs the following two sub-operations:

    Skip check (SKC)

    Compute the pixel distortion of a user-specified shape and motion vector combination. The VME media sampler fetches necessary pixels, performs fractional and bidirectional filtering (as necessary), and then computes the distortion between the derived reference and source. The skip decision can optionally be enhanced to include a 4x4 forward transform, the results of which are compared against a user specified threshold to emulate the effects of the forward quantization zeroing effect.

    Intra prediction estimation (IPE)

    Perform intra prediction on a given source macroblock to determine the best intra prediction modes and the best shape partitioning combination.

    Requires CL12 and intel_subgroups.