On-line data processing at the ATLAS general purpose particle detector, which is currently under construction at Geneva, generates demands on computing power that are difficult to satisfy with commodity CPU-based computers. One of the most demanding applications is the recognition of particle tracks that originate from B-quark decays. However, this and many others applications can benefit from parallel execution on field programmable gate arrays (FPGA). After the demonstration of accelerated track recognition with big FPGA-based custom computers, the development of FPGA based coprocessors started in the late 1990's. Applications of FPGA coprocessors are usually partitioned between the host and the tightly coupled coprocessor. The objective of the research that I present in this thesis was the development of software that mediates to applications the access to FPGA coprocessors. I used a software process based on iterative prototyping to cope with the expected changing requirements. Also, I used a strict bottom-up design to create classes that model devices on the coprocessors. Using these low-level classes, I developed tools which were used for bootstrapping, debugging, and firmware update of the coprocessors during their development and maintenance. Measurements show that the software overhead introduced by object-oriented programming and software layering is small. The software-support for six different coprocessors was partitioned into corresponding independent packages, which reuse a set of packages that provide common and basic functions. The steady evolution and use of the software during more than four years shows that the software is maintainable, adaptable, and usable.