Implemented a CGRA designed to accelerate matrix-matrix multiplication via systolic arrays in the Structural Simulation Toolkit (SST). Achieved 5–24× speedup over RTL simulation on a single core, and up to 42× speedup by exploiting SST's built-in multi-core parallelism.
Searches program memory traces with hill climbing, simulated annealing, and evolutionary algorithms to identify recurring access patterns that can guide memory prefetching.
Code
Explores accelerating explainable random forest inference by moving simple node comparisons into memory, reducing bus traffic and exposing parallelism in XAI workloads.
A taint-analysis-based compiler framework that verifies data obliviousness of programs, ensuring memory access patterns are independent of sensitive inputs. Applies loop unrolling, loop fusion, and FPLICM optimizations to generate data-oblivious code.
A curated benchmark spanning image, text, and tabular datasets, with metrics for evaluating the fidelity and performance of SHAP, LIME, and interpretable machine-learning models.
Code
Evaluates graph-reordering techniques that improve locality and execution speed for power-law graphs on baseline systems and the OMEGA graph accelerator.
Code
A sensor-equipped glove, gesture-classification pipeline, and Kotlin mobile application for translating Ethiopian Sign Language gestures into speech in real time.
Code