A Novel Transform Accelerator With Fast Kernel Selection and Efficient Transform Circuit
A Novel Transform Accelerator With Fast Kernel Selection and Efficient Transform Circuit
Abstract:
The introduction of multiple transform types into the Versatile Video Coding (VVC) standard has yielded notable encoding gains but also resulted in substantial computational burdens, posing two critical challenges for hardware implementation: fast kernel selection and efficient transform circuit design. Existing studies typically address these challenges in isolation, lacking a holistic solution for VVC transform coding. In this paper, we present a groundbreaking transform accelerator that unifies transform kernel selection and multiple transform circuit within a single framework. In terms of algorithms, driven by mechanistic analysis, we propose a decision tree-based kernel selection algorithm that ensures both high decision accuracy and computational efficiency. Additionally, we design a transfer matrix-based approximation algorithm for Discrete Sine Transform Type-7 and a matrix decomposition-based improved computation for Discrete Cosine Transform Type-2, significantly reducing their computational complexity. On the hardware front, we implement a high-performance transform accelerator architecture, which integrates highly pipelined kernel selection and transform computation architectures. With multiple reuse and parallelism strategies, experimental results substantiate impressive performance gains as both theoretical analyses and hardware implementation results reveal that our proposed accelerator achieves a transform accuracy reduction of over 94% with a substantial performance degradation, while maintaining a frequency capability up to 867.5 fps. To the best of our knowledge, this is the first holistic hardware solution for VVC transform coding that addresses both challenges of kernel selection and transform circuit design.
Index Terms —
Video coding, transform accelerator, decision-tree, sparsity optimization, pipelined design.
” Thanks for Visit this project Pages – Register This Project and Buy soon with Novelty “
