This paper introduces DG-PIC, a novel framework for domain generalized point cloud understanding that addresses the challenges of performance drops on unseen data due to distribution shifts across different domains. Traditional domain generalization (DG) techniques often focus on a single task and neglect the potential of testing data, while In-Context Learning (ICL) typically relies on high-quality data and is limited to a single dataset. The paper proposes a multi-domain multi-task setting that handles multiple domains and tasks within a unified model, enhancing generalizability across various tasks and domains at testing time. DG-PIC introduces dual-level source prototype estimation, considering both global and local features, and a dual-level test-time feature shifting mechanism that leverages macro-level domain semantic information and micro-level patch positional relationships. The framework does not require model updates during testing and can handle unseen domains and multiple tasks, including point cloud reconstruction, denoising, and registration. The paper also presents a benchmark for this new setting, with comprehensive experiments demonstrating that DG-PIC outperforms state-of-the-art techniques significantly. The contributions include a novel multi-domain multi-task setting, two innovative dual-level modules, and a new benchmark for evaluating performance in this setting.This paper introduces DG-PIC, a novel framework for domain generalized point cloud understanding that addresses the challenges of performance drops on unseen data due to distribution shifts across different domains. Traditional domain generalization (DG) techniques often focus on a single task and neglect the potential of testing data, while In-Context Learning (ICL) typically relies on high-quality data and is limited to a single dataset. The paper proposes a multi-domain multi-task setting that handles multiple domains and tasks within a unified model, enhancing generalizability across various tasks and domains at testing time. DG-PIC introduces dual-level source prototype estimation, considering both global and local features, and a dual-level test-time feature shifting mechanism that leverages macro-level domain semantic information and micro-level patch positional relationships. The framework does not require model updates during testing and can handle unseen domains and multiple tasks, including point cloud reconstruction, denoising, and registration. The paper also presents a benchmark for this new setting, with comprehensive experiments demonstrating that DG-PIC outperforms state-of-the-art techniques significantly. The contributions include a novel multi-domain multi-task setting, two innovative dual-level modules, and a new benchmark for evaluating performance in this setting.