TL;DR
This paper presents a novel design flow for implementing complex dynamic dataflow applications, like video decoders, on embedded multi-core systems. By introducing a custom architecture and advanced software synthesis techniques, they achieve significant performance improvements, enabling real-time HD video decoding on their specialized hardware.
What Problem Does This Paper Address?
Modern embedded systems are becoming increasingly complex, especially for multimedia applications like video codecs. Traditional approaches of simply increasing clock speeds have hit a power wall, pushing the industry towards multi-core architectures. However, designing for these systems is challenging due to the need for both high performance and low power consumption, while also managing the complexity of data-intensive applications. Dataflow programming offers a flexible way to model parallelism and communication, but efficiently implementing dynamic dataflow models on hardware requires specialized architectures and compilation techniques that go beyond traditional memory-centric approaches.
Methods
The authors propose a comprehensive approach comprising:
- Custom Architecture Model: A distributed memory architecture with Transport-Triggered Architecture (TTA) processor cores. TTA processors are chosen for their efficiency and ability to expose low-level parallelism, which can be exploited by dataflow models.
- Dynamic Dataflow Programming Model: Utilizing a Dataflow Process Network (DPN) model, which allows for data-dependent behavior and is well-suited for complex applications like video codecs.
- Software Synthesis Techniques:
- Branch-Free and Copy-Free Communications: Optimizing the execution of actors by merging reading, processing, and writing steps, and directly accessing FIFO channels to reduce overhead.
- Aligned Communications: Generating two versions of actions (standard and aligned) to exploit vectorization and instruction-level parallelism when data is linearly accessible.
- Multi-level Dynamic Scheduling: Implementing both internal action scheduling (checking firing rules) and external actor scheduling (mapping actors to processors) to manage concurrency and exploit locality.
- Hybrid Memory Architecture: A memory organization with private processor memories for actor execution and shared memories for communication channels, aiming to reduce congestion and power consumption.
Datasets
The paper evaluates their approach using descriptions of MPEG-4 Visual and High Efficiency Video Coding (HEVC) decoders. These are not datasets in the traditional sense but rather complex application models. They use standard video sequences (e.g., Foreman, Old Town Cross, BasketballPass, KristenAndSarah) for testing.
Key Insights and Ideas
- TTA for Dataflow: Transport-Triggered Architecture processors are a good fit for dynamic dataflow programs, offering a balance between flexibility and efficiency.
- Beyond Traditional Memory: A hybrid memory architecture that separates local and shared memory, specifically for communication channels, is crucial for overcoming bottlenecks in dataflow systems.
- Software Synthesis is Key: Advanced techniques like branch-free, copy-free, and aligned communications are essential for unlocking the performance potential of dynamic dataflow on embedded multi-core systems.
- Application-Specific Optimization: The effectiveness of the approach is demonstrated by its application to real-world, complex codecs like HEVC, highlighting the need for tailored solutions.
- Coarse-Grain Parallelism Limits: The performance gains from adding more processors eventually plateau due to inherent data dependencies in video decoding, suggesting that architectural improvements and efficient actor mapping are as important as raw core count.
Key Results
- Real-time HD Decoding: Achieved real-time decoding (40 FPS) of 720p MPEG-4 Visual video sequences on their custom multi-core platform clocked at 1GHz.
- Significant Speedup: Demonstrated a speedup of over 100% compared to previously proposed dataflow implementations for similar tasks.
- Actor Performance Variation: Identified "compute-intensive" actors (e.g., inverse transform) that benefit significantly from parallel processing (up to 3x speedup), and "control-intensive" actors (e.g., entropy decoding) that show less improvement.
- MPEG-4 vs. HEVC: The MPEG-4 Visual decoder implementation achieved significantly higher frame rates (approximately 8x) than the HEVC decoder, attributed to algorithmic complexity, development status, and prior optimization efforts.
- Processor Scaling: For MPEG-4 Visual, maximum performance was reached with 16 processors, with diminishing returns beyond that point.
Limitations and Future Work
- HEVC Performance: The current HEVC implementation does not achieve real-time decoding for high-definition sequences, suggesting further optimization is needed.
- Assembly Kernels: The authors note that incorporating highly optimized assembly kernels could further boost performance, especially for compute-intensive actors.
- Processor Sharing: In realistic implementations, actors will need to share processors, requiring more sophisticated actor mapping strategies to balance computation and communication.
- Simulation Speed: The simulation speed of complex dataflow programs on embedded multi-core platforms is a significant limitation, hindering rapid iteration and debugging.
- Future Work: The paper suggests exploring more optimized implementations, potentially using assembly-level optimizations, and refining actor mapping strategies for scenarios where processors are shared.
Summary updated