Jeonghyung Park

Guillaume Aupy, JeongHyung Park, Padma Raghavan

In this paper, we propose a novel reordering scheme to improve the performance of a Laplacian Mesh Smoothing (LMS). While the Laplacian smoothing algorithm is well optimized and studied, we show how a simple reordering of the vertices of the mesh can greatly improve the execution time of the smoothing algorithm. The idea of our reordering is based on (i) the postulate that cache misses are a very time consuming part of the execution of LMS, and (ii) the study of the reuse distance patterns of various executions of the LMS algorithm. Our reordering algorithm is very simple but allows for huge performance improvement. We ran it on a Westmere-EX platform and obtained a speedup of 75 on 32 cores compared to the single core execution without reordering, and a gain in execution of 32% on 32 cores compared to state of the art reordering. Finally, we show that we leave little room for a better ordering by reducing the L2 and L3 cache misses to a bare minimum.

Jaeyoo Park, Jin Young Choi, Jeonghyung Park et al.

We present a novel OCR-free document understanding framework based on pretrained Multimodal Large Language Models (MLLMs). Our approach employs multi-scale visual features to effectively handle various font sizes within document images. To address the increasing costs of considering the multi-scale visual inputs for MLLMs, we propose the Hierarchical Visual Feature Aggregation (HVFA) module, designed to reduce the number of input tokens to LLMs. Leveraging a feature pyramid with cross-attentive pooling, our approach effectively manages the trade-off between information loss and efficiency without being affected by varying document image sizes. Furthermore, we introduce a novel instruction tuning task, which facilitates the model's text-reading capability by learning to predict the relative positions of input text, eventually minimizing the risk of truncated text caused by the limited capacity of LLMs. Comprehensive experiments validate the effectiveness of our approach, demonstrating superior performance in various document understanding tasks.