high performance computing on graphics processing units: hgpu.org

As of late massively parallel devices have become mainstream and are widely used in research and industry. But even despite recent advances of the API, programming these devices has proven to be a difficult and error-prone task. We have designed Conflux, an embedded domain-specific language that integrates massively parallel semantics into a high-level programming language. Conflux utilizes operations, type system, and data structures of a host language and provides an extensible framework for mapping those to concrete device-specific implementations. We have prototyped an embedding for C# language and a mapping that translates Conflux code for execution on CUDA GPUs and multicore CPUs. Our approach improves programmability of massively parallel algorithms and makes high-performance more accessible for developers.