Currently, the majority of devices is heterogeneous and comprises at least a multi-core CPU and a GPU. Exploiting these modules requires programmers to a) assign parallel activities to the different hardware resources, and b) code each activity through target-specific approaches x(e.g., multi-core CPUs and GPUs).Current frameworks do not provide high-productivity while guaranteeing performance comparable to low-level ones. Also, they often lack the easy exploration of possible task assignments to the heterogeneous resources, mainly because the assignment needs to be done in advance and the targets have software incompatibilities.This paper introduces task-DAGs (Directed Acyclic Graph) concepts into PHAST, a single-source data-parallel library that supports multi-core CPUs and NVIDIA GPUs. The resulting framework allows postponing the task-to-target assignment to the testing or runtime phase. In fact, each task is coded in a target-agnostic fashion and can be automatically optimized for a specific platform. The consequent productivity effects are shown for a computer-vision benchmark comparing to a SYCL implementation and performing an automatic task assignment for maximum resources utilization.