Abstract: The design and implementation of a SIMD controller for a polymorphous array processor is presented in the paper. The polymorphous array processor is intended for graphics and multimedia applications. To meet the needs for efficient data level parallel computation, we use the state machine approach to design the row, column, and cluster controllers. A row controller dispatches SIMD instructions to processing elements; a column controller performs data and program loading as well as remote data access. The polymorphous machine employs two computation modes, the SIMD mode and the MIMD mode. The controllers are mainly used for SIMD mode of computation. A dedicated hardware circuit design improves the ability to handle parallel data of the processor.
Introduction: The Multimedia Bottleneck
Modern multimedia applications, ranging from high-definition video processing to complex 3D graphics rendering, are characterized by massive data-level parallelism. These tasks often require a processor that can adapt its execution model based on the nature of the data. While Single Instruction, Multiple Data (SIMD) architectures are highly efficient for uniform operations on large data blocks, Multiple Instruction, Multiple Data (MIMD) modes are necessary for handling independent, non-uniform tasks.
The challenge in designing such a "polymorphous" processor lies in the control logic. Efficiently managing a 2D array of processing elements (PEs) without introducing significant latency or hardware complexity is critical. This paper addresses this bottleneck by proposing a hierarchical control system for the PAAG (Polymorphous Array Architecture for Graphics) processor.
Hierarchical Control Architecture
The proposed architecture organizes a 4x4 cluster of 16 Processing Elements (PEs) using a three-tiered controller hierarchy. This structure allows the processor to seamlessly switch between SIMD and MIMD modes while maintaining efficient instruction dispatch and data movement.
The hierarchy consists of three distinct controller types:
- Cluster Controller (CLC): Acts as the primary interface between the processor cluster and external systems. It handles program and data loading and coordinates high-level SIMD operations.
- Row Controllers (RC): Each of the four rows in the array is managed by a dedicated RC. These controllers are responsible for reconfiguring their respective rows into SIMD mode and dispatching instructions to the PEs.
- Column Controllers (CC): Each column features a CC that manages "plumbing" tasks, such as shared memory access and data transfers between PEs and cluster-wide memory.
This division of labor prevents the control logic from becoming a central bottleneck. By offloading row-specific instruction dispatch to the RCs, the CLC can focus on global coordination, reducing overall system latency.
State Machine Precision
To ensure predictable and efficient transitions between different operational states, each controller is implemented as a rigorous Finite State Machine (FSM). This approach allows for precise control over idling, instruction reception, and data transfer phases. Specialized instructions like CALLC (to trigger a SIMD program) and RETC (to signal completion) are integrated into this logic to facilitate smooth operation.
The Row Controller FSM, for instance, manages the lifecycle of SIMD instruction execution. It transitions from an IDLE state to receiving instruction packets, reading instructions, and eventually dispatching them to the PEs. This structured logic ensures that the "polymorphous" nature of the hardware—its ability to reconfigure rows on the fly—is handled with mathematical precision.
Furthermore, the design utilizes a unified 32-bit data packet format for both instructions and data. This standardization allows the hardware to reuse the same physical interconnects for multiple purposes, significantly saving chip area while maintaining high communication bandwidth.
Verification and Results
The design was implemented using Register Transfer Level (RTL) code and underwent extensive functional verification. The state machines were tested under various simulated workloads to ensure correct transitions and data integrity. Following verification, the design was synthesized using Synopsys Design Compiler and validated on FPGA hardware.
The results demonstrate that the hierarchical controller approach is functionally sound and highly scalable. The modular nature of the CLC, RC, and CC components suggests that the architecture could be expanded to larger PE arrays beyond the initial 4x4 cluster without a linear increase in control complexity.
Conclusion and Future Work
While the paper successfully validates the architectural design and logic of the SIMD controllers, it primarily focuses on functional correctness. Future research is needed to provide quantitative performance benchmarks, such as speedup factors for specific multimedia kernels (e.g., H.264 encoding) and power consumption metrics compared to traditional GPU or fixed-SIMD architectures. Implementing real-world multimedia kernels on the PAAG processor will be the next step in evaluating its true throughput and energy efficiency in production environments.
Summary updated