FastLBP: GPU Framework Accelerates Program Analysis Using Belief Propagation

Belief propagation, a foundational inference technique in probabilistic graphical models, has long struggled with computational scalability when applied to analyzing complex software programs. Existing attempts to move this approach onto GPUs have faced limitations in both flexibility and integration with logical constraints.

The Challenge#

Traditional CPU-based belief propagation becomes prohibitively slow when handling large-scale program analysis tasks. Prior GPU implementations sacrifice generality by failing to support the diverse update strategies that real-world program analysis demands, and they struggle to incorporate logical constraints efficiently.

Introducing FastLBP#

A research team has developed FastLBP, a GPU-accelerated belief propagation framework specifically designed for program analysis. The system introduces three key innovations:

• A unified representation for specifying flexible update strategies needed in program analysis
• A dependency analysis algorithm that enables parallel execution of computations
• An efficient GPU implementation using thread-per-message assignment and memory-optimized data structures

Performance Gains#

Testing on two benchmark systems demonstrates significant speedups:

• SmartFL: 17.42× faster than CPU-based methods and 6.14× faster than existing GPU approaches
• BINGO: 2.82× faster than CPU-based approaches

Importantly, these improvements come without sacrificing accuracy—the framework maintains the same analytical quality as its predecessors while running substantially faster.

Why It Matters#

FastLBP represents a meaningful advance in bridging the gap between theoretical inference methods and practical program analysis. By supporting update strategies that competing GPU solutions cannot handle, it brings greater generality to the field while delivering the performance gains that modern software analysis requires.