The solidification microstructure of metal 3D-printed products, which determines their quality, is often formed through the competitive growth of columnar dendritic crystals during rapid solidification in the molten pool created by beam irradiation. Since the crystal orientation of the polycrystalline structure significantly affects the mechanical properties, understanding the mechanisms of competitive growth and controlling crystal orientation are essential. However, the formation process of such rapid solidification microstructure occurs at small spatiotemporal scales and under optically opaque conditions, making in-situ observation challenging. Thus, numerical simulation is also an important approach for understanding the formation process of microstructures. The phase-field (PF) model is widely used as a numerical approach capable of accurately representing dendrite growth. By employing the nanometer-scale interface width, the PF model can quantitatively express far-from-equilibrium solidification phenomena such as solute trapping at the solid-liquid interface [1]. However, its requirement for a nanometer-scale computational grid spacing makes it difficult to simulate competitive growth in a micrometer-scale molten pool. To address this issue, we have developed a high-performance computational approach that enables large-scale simulations involving the competitive growth of multiple dendritic crystals. The polycrystalline rapid solidification model is formulated by extending the PF model of rapid solidification of a dilute binary alloy [1] into a multi-phase field (MPF) framework [2]. Furthermore, we implement high-performance computing techniques such as GPU parallel computing and block-structured adaptive mesh refinement (AMR) [3]. We also introduce the active parameter tracking method [4], which significantly reduces memory usage by storing only the necessary PF variables at each grid point. We evaluate the computational performance of the developed method and demonstrate its capability to simulate the competitive growth of columnar dendrites in micrometer-scale molten pool using nanometer-scale numerical grids.