Module motion_estimation 

GPU-accelerated motion estimation for AV1 and VP9 video codecs.

This module provides compute-shader-based motion estimation pipelines suitable for AV1 and VP9 intra/inter frame encoding. The GPU kernels exploit massively parallel block matching to evaluate Sum of Absolute Differences (SAD) and Sum of Squared Differences (SSD) across many candidate motion vectors simultaneously.

Architecture

The pipeline is divided into three GPU dispatch stages:

1. Hierarchical downscale – build a Gaussian pyramid (up to 4 levels) so that large motion is found at low resolution first.
2. Block-match sweep – for every block in the current frame, evaluate all candidate motion vectors within the search window using parallel SAD/SSD kernels dispatched with workgroup-local shared memory (reducing global-memory bandwidth by ~8×).
3. Refinement – perform ±1 / ±½ pixel sub-pixel refinement around the best integer candidate found in stage 2.

Status

The GPU shader dispatch plumbing is present but the WGSL shaders for AV1/VP9-specific block partitions (superblock, transform units, etc.) are stubs. The CPU reference path is fully functional and used for testing / CI.

Structs

BlockMvResult

Motion estimation result for a single block.

FrameMvResult

Full-frame motion estimation result.

MotionEstimationConfig

Configuration for a motion-estimation pass.

MotionEstimator

GPU-accelerated motion estimator.

MotionVector

A 2-D integer motion vector (pixel precision).

SubpixelMv

A 2-D sub-pixel motion vector (1/4-pixel precision, values are in units of 1/4 pixel).

Enums

BlockPartition

Block partition mode used during motion search.

MotionMetric

Cost metric for evaluating motion-vector candidates.

TargetCodec

Codec the motion-estimation result will be used for.