Introduction
Convolutional Neural Networks (CNNs) have become the cornerstone of modern computer vision, yet their computational demands often exceed the capabilities of general-purpose processors. FPGAs offer a compelling alternative due to their reconfigurability and energy efficiency. To further optimize performance, researchers frequently employ the Winograd minimal filtering algorithm, which reduces the number of multiplications required for convolutions by transforming them into a different domain.
However, traditional Winograd implementations suffer from two significant drawbacks:
- Rigidity: Most accelerators are hard-coded for a specific kernel size (e.g., 3x3). When encountering 1x1 kernels—prevalent in modern lightweight models—these accelerators either waste resources through zero-padding or require separate, area-intensive hardware.
- Data Movement Complexity: The tile-based processing of Winograd creates irregular memory access patterns, often leading to bandwidth bottlenecks.
WinoCNN addresses these challenges through a novel "kernel sharing" architecture and a parameterized systolic array, achieving significant improvements in both throughput and DSP efficiency.
Methodology: The WinoCNN Architecture
The core of WinoCNN is a specialized Processing Element (WinoPE) integrated into a 2D systolic array, supported by a hierarchical memory subsystem.
Winograd Transformation and Domain Processing
The Winograd algorithm transforms the input data ($d$) and the kernel ($g$) into the Winograd domain using transformation matrices $B^T$ and $G$. The element-wise multiplication ($\odot$) in this domain is then transformed back to the spatial domain ($Y$) using matrix $A^T$.
WinoPE and Kernel Sharing
The primary innovation of WinoCNN is the WinoPE, which enables a single piece of hardware to support multiple kernel sizes (e.g., 1x1, 3x3, 5x5) with near-maximal efficiency. The authors discovered that Winograd transformation matrices for different kernel sizes share common sub-structures when using the same Winograd filter size ($\omega$).
By using a "selection bit," the WinoPE can toggle between these configurations. This "kernel sharing" mechanism ensures that the DSP slices are fully utilized regardless of whether the network is processing a 3x3 or a 1x1 convolution layer, avoiding the typical 8/9ths resource waste seen in fixed-size accelerators.
Parameterized Systolic Array and Memory Subsystem
To minimize data fanout and maximize clock frequency, WinoPEs are organized into a 2D systolic array. Data flows rhythmically between PEs, facilitating natural data reuse for overlapping convolution windows.
Feeding this array is a hierarchical memory subsystem utilizing a specialized BRAM buffer matrix. This matrix uses a unique addressing scheme to provide parallel, "planar" data access, significantly reducing the need to re-fetch overlapping data tiles from external DDR memory.
Evaluation and Results
WinoCNN was evaluated on Xilinx ZCU102 and Ultra96 FPGAs using three benchmark models: VGG-16, Inception-V4, and YoloV2.
Performance and Efficiency
The architecture achieved a peak throughput of 3.1 TOPS on the ZCU102. More importantly, it reached a DSP efficiency of 1.33 GOPS/DSP, representing a 29.1% improvement over previous state-of-the-art Winograd solutions.
| Metric | WinoCNN (ZCU102) | Improvement vs. SOTA |
|---|---|---|
| Peak Throughput | 3.1 TOPS | ~20% |
| DSP Efficiency | 1.33 GOPS/DSP | 29.1% |
Flexibility Across Models
Unlike rigid accelerators that see performance drops in models with mixed kernel sizes (like Inception-V4), WinoCNN maintained high efficiency. It also outperformed Xilinx’s Vitis-AI implementation in both throughput and latency for VGG-16 and YoloV2, despite operating at a lower clock frequency, highlighting the architectural superiority of the kernel-sharing systolic array.
Conclusion
WinoCNN demonstrates that mathematical insights into Winograd transformation matrices can be translated into highly efficient hardware. By enabling kernel sharing within a systolic array, the design overcomes the rigidity of traditional Winograd accelerators. While future work is needed to support highly irregular asymmetric kernels (e.g., 1x7), WinoCNN sets a new standard for DSP efficiency and flexibility in FPGA-based CNN acceleration.
Summary updated