K-PACT: Kernel Planning for Adaptive Context Switching - A Framework for Clustering, Placement, and Prefetching in Spectrum Sensing

Hasan Umut Suluhan, Jiahao Lin, Serhan Gener, Chaitali Chakrabarti, Ümit Y. Ogras, Ali Akoglu

Abstract

In next-generation wireless communication, the ability to rapidly sense and adapt to the radio frequency (RF) environment is critical. However, traditional reconfigurable hardware architectures suffer from significant latency bottlenecks when switching between different signal-processing tasks. This paper introduces K-PACT, a software planning framework designed to optimize the mapping of spectrum-sensing kernels onto reconfigurable hardware. By utilizing temporal-independence clustering and intelligent instruction memory (IMEM) management, K-PACT reduces context-switching latency by over 98x and off-chip memory fetches by 207x, enabling real-time wideband signal sensing.

Introduction

Spectrum sensing—the process of scanning the RF environment to identify occupied bands and signal types (e.g., OFDM, PSK)—is a foundational component of cognitive radio systems. Modern sensing requirements involve executing complex decision trees, which are sequences of mathematical kernels like Fast Fourier Transforms (FFTs) and matrix inversions, across dozens of subchannels simultaneously.

The primary challenge lies in the dynamic nature of the RF environment. Hardware must constantly adapt to new signal hypotheses by switching between different decision trees. On traditional Coarse-Grained Reconfigurable Architectures (CGRAs), this typically requires a "hard switch": the processor must stop, fetch new instruction code from slow off-chip memory, and reconfigure the chip. This overhead creates a massive latency bottleneck that prevents the real-time processing of wideband signals. K-PACT addresses this by treating reconfiguration as a manageable software planning problem rather than a hardware limitation.

Methodology

K-PACT is a multi-stage planning tool designed to work with modified systolic array architectures, specifically those based on the FluxSPU. The framework introduces several key innovations to mitigate switching overhead:

Temporal-Independence Clustering

Rather than focusing solely on spatial packing, K-PACT identifies kernels that are "mutually exclusive"—kernels that never run simultaneously within a specific decision tree. The framework groups these kernels into clusters that share the same Processing Element (PE), maximizing hardware utilization without increasing the physical footprint.

Multi-Bank IMEM Management and the "Soft Switch"

Each PE is equipped with multiple Instruction Memory (IMEM) banks. K-PACT pre-loads different kernels into these banks during the initialization phase. This architectural shift transforms a "hard switch" (fetching from off-chip memory) into a "soft switch" (simply pointing the PE to a different local memory bank). This transition occurs almost instantly, allowing the system to adapt to new signal environments at nanosecond scales.

Congestion-Aware Placement

To minimize data movement latency, the tool physically places "entry" kernels—the starting points of decision trees—closer to the SRAM data buffers. This strategic placement prevents routing "traffic jams" on the chip and ensures that data flows efficiently through the systolic array.

Multi-Objective Optimization

The planner operates in an optimization loop, balancing the trade-offs between local memory size, total physical chip area, and expected switching latency. This allows designers to find the "Goldilocks" zone for hardware configuration based on specific workload requirements.

Experimental Setup

The researchers evaluated K-PACT using a simulated spectrum-sensing workload on the DS3 (Domain-Specific System-on-Chip) open-source simulator. The workload consisted of 48 concurrent subchannels, each running decision trees composed of kernels such as Energy Detection, Kurtosis Detection, CP Detection, and various signal estimators (OFDM, FSK, PSK, FM). The hardware baseline was a systolic array of PEs synthesized using a 12nm process.

Key Insights and Results

The Performance Impact

Compared to a baseline system lacking intelligent planning and pre-loading, K-PACT achieved transformative results:

  • Off-chip binary fetches: Reduced by 207.81×.
  • Average switching time: Lowered by 98.24×.
  • Per-subband execution time: Improved by 132.92×.
  • Scheduling Overhead: Reduced by 73.77× by pre-planning the layout and reducing the search space for the dynamic scheduler.

Optimization Trade-offs

The study found that increasing local IMEM allows for more "soft switches" but increases PE size, which can eventually slow down data movement. For the specific sensing workload tested, an IMEM size of 4.5KB was found to be optimal. Additionally, the researchers discovered that a Fixed Pre-initialized Placement (FPIP) strategy—where pre-loaded kernels are strictly preserved—outperforms strategies that allow the runtime scheduler to evict kernels, as eviction often triggers a chain reaction of expensive hard switches.

Conclusion and Future Work

K-PACT demonstrates that intelligent software planning can overcome fundamental hardware bottlenecks in reconfigurable computing. By pre-loading kernels and clustering them based on temporal independence, the framework enables the rapid adaptation required for next-generation spectrum sensing.

Future work will focus on making the clustering algorithms more "routing-aware" to further reduce interconnect congestion. Additionally, as decision trees grow in complexity, the authors plan to explore machine learning-based mappers to handle the massive optimization space. The ultimate goal is to validate these simulated gains on physical FPGA or ASIC implementations of the FluxSPU architecture.

Summary updated

BibTeX
@inproceedings{suluhan2025k,
  author = {Hasan Umut Suluhan and Jiahao Lin and Serhan Gener and Chaitali Chakrabarti and Ümit Y. Ogras and Ali Akoglu},
  title = {{K-PACT: Kernel Planning for Adaptive Context Switching - A Framework for Clustering, Placement, and Prefetching in Spectrum Sensing}},
  booktitle = {ICCAD},
  year = {2025},
  doi = {10.1109/iccad66269.2025.11240855},
}
Figure