How libdav1d Accelerates AV1 CDEF Decoding

The Constrained Directional Enhancement Filter (CDEF) is a crucial post-processing step in the AV1 video codec designed to eliminate ringing artifacts around sharp edges. However, CDEF is computationally expensive, making it a major bottleneck during playback. The libdav1d decoder, developed by VideoLAN and partners, achieves industry-leading decoding speeds by heavily optimizing this step. This article explores the specific software engineering and hardware acceleration techniques libdav1d uses to speed up the CDEF process.

Highly Optimized Assembly Implementations (SIMD)

The primary method libdav1d uses to accelerate CDEF is hand-written assembly language tailored for modern processors. Because CDEF operates on small, discrete pixel blocks (typically 8x8 or 4x4), it is highly receptive to Single Instruction, Multiple Data (SIMD) parallel processing.

• x86 Architectures: libdav1d features extensive assembly optimizations using AVX2 and AVX-512 instruction sets. These allow the decoder to process multiple pixels simultaneously, significantly reducing the clock cycles required to calculate directional pixel averages.
• ARM Architectures: For mobile devices and Apple Silicon, libdav1d utilizes highly optimized ARM NEON (both 32-bit and 64-bit) assembly, ensuring fluid AV1 playback on power-constrained devices.

Vectorized Direction Search

Before applying the filter, CDEF must determine the primary edge direction for each block from eight possible angles. libdav1d optimizes this direction-search phase by vectorizing the mathematical calculations. Instead of evaluating pixel variances sequentially, the decoder uses SIMD instructions to calculate the sum of squared differences (SSD) for multiple directions at once, drastically reducing the search time.

Advanced Multi-Threading and Pipelining

In modern processors, raw instruction speed is only half the battle; efficient CPU utilization is equally important. libdav1d employs a highly sophisticated, task-based threading model that schedules CDEF execution dynamically:

• Row-Level Pipelining: CDEF cannot run until the deblocking filter is complete. libdav1d does not wait for the entire frame to finish deblocking; instead, it utilizes a fine-grained, row-based dependency tracker. As soon as a row of blocks passes the deblocking stage, CDEF tasks are immediately queued and executed on available CPU threads.
• Lock-Free Scheduling: The custom thread pool in libdav1d relies on lock-free synchronization mechanisms to minimize overhead and prevent worker threads from idling.

Memory Bandwidth and Cache Optimization

Pixel processing is highly dependent on memory bandwidth. If the CPU has to constantly fetch pixel data from system RAM, performance plummets. libdav1d counters this by keeping working pixel data within the CPU’s L1 and L2 caches. By structuring CDEF to run immediately after preceding loop filters in a cache-friendly layout, the decoder minimizes memory bus traffic, resulting in a substantial speedup and lower power consumption.