In high-end intelligent manufacturing, free-form surface workpieces with complex topologies are usually converted into triangular meshes for tool path generation. However, mainstream iso-scallop methods in CAM systems cannot simultane- ously achieve non-self-intersection and iso-scallop height for triangular mesh machining. To solve this problem, this paper presents a novel ball-end milling path planning method. Starting from the boundary of the machining region, we generate inward offset curves layer by layer. For the offset curves, we define the distance between two points under the scallop-height metric, and equivalently transform the self-intersection of path points into the condition that the scal- lop distance to the previous curve is less than the constraint threshold. Then, based on GPU parallel acceleration, we quickly remove such problematic path points from the offset curves, thereby achieving fast self-intersection removal while strictly satisfying the iso-scallop constraint. Comparisons with industrial software and state-of-the-art methods verify that only the proposed method can stably guarantee both non-self-intersection and iso-scallop height, showing clear superiority.