As an open, royalty-free framework for writing programs that execute across heterogeneous platforms, OpenCL gives programmers access to a variety of data parallel processors including CPUs, GPUs, the Cell and DSPs. All OpenCL-compliant implementations support a core specification, thus ensuring robust functional portabiity of any OpenCL program. This thesis presents the CUDAtoOpenCL source-to-source tool that translates code from CUDA to OpenCL, thus ensuring portability of applications on a variety of devices. However, current compiler optimizations are not sufficient to translate performance from a single expression of the program onto a wide variety of different architectures. To achieve true performance portability, an open standard like OpenCL needs to be augmented with automatic high-level optimization and transformation tools, which can generate optimized code and configurations for any target device. This thesis presents details of the working and implementation of the CUDAtoOpenCL translator, based on the Cetus compiler framework. This thesis also describes key insights from our studies optimizing selected benchmarks for two distinct GPU architectures: the NVIDIA GTX280 and the ATI Radeon HD 5870. It can be concluded from the generated results that the type and degree of optimization applied to each benchmark need to be adapted to the target architecture specifications. In particular, the different hardware architectures of the basic compute unit, register file organization, on-chip memory limitations, DRAM coalescing patterns and floating point unit throughput of the two devices interact with each optimization differently.