Nowadays, Graphics Processing Units (GPUs) and other types of emerging accelerators have an important role in high-performance computing. These devices can be leveraged in a wide range of applications through using appropriate programming environments such as CUDA and OpenCL which lead to reaching high-performance applications. However, on one hand, programming GPUs is a painful and error-prone task and requires a great amount of expertise especially in low-level architectural features as well as their memory management in order to achieve reasonable performance. On the other hand, enabling running high-level programming languages such as Java with massive computational power of today's GPUs can lessen the burden of this complexity. Considering new features in Java 8 such as lambda functions which are used in Java parallel streams, supporting these new features is vital to use these devices in real applications. In this paper, we introduce a just-in-time compiler, named Bamshad, which ports lambda functions used in Java parallel streams to GPU at runtime. For this, a series of compiler techniques are adopted to transparently eliminate unnecessary data communication between CPUs and GPUs. With our approach, a programmer is not involved in the detailed process of tuning a GPU device for reducing the amount of communication. The experimental results show that the proposed JIT compiler yields 13× improvement in comparison to sequential Java execution for all benchmarks. Also, in comparison to parallel Java, our work yields 3.9× improvement.