Fast Dependent Quantization Using Trellis Pruning, Forward Context Adaptation and Vectorization

Dependent quantization (DQ) is one of the key coding tools in the Versatile Video Coding (VVC) standard. DQ employs two scalar quantizers, with the per-coefficient quantizer selection being governed by a parity-driven 4-state state machine. As the design is normatively enforced, usage of DQ requires a rate-distortion optimized quantization (RDOQ) with a per coefficient decision and state update, e.g. a trellisbased quantization, as initially proposed for the VVC reference software (VTM). Due to its inherent dependencies, including VVCs context selection based on previously encoded coefficient values, and fairly broad search range, the trellis quantization is computationally highly complex. Reducing the complexity of this algorithm is crucial for practical VVC encoders. In this paper we propose a fast dependent quantization trellis search improving the initial design by: trellis pruning of improbable branches, forward adaptive context propagation, and finally a vectorized implementation. The proposed approach, implemented in the open and optimized VVenC encoder, reduces quantization runtime by 37%, allowing an overall 15% encoder speedup in medium preset with no impact on compression performance, in all-intra coding conditions. In random-access conditions, an 9% overall encoder speedup is achieved.