high performance computing on graphics processing units: hgpu.org

The most popular multithreaded languages based on the fork-join concurrency model (CilkPlus, OpenMP) are currently being extended to support other forms of parallelism (vectorization, pipelining and single-instruction-multiple-data (SIMD)). In the SIMD case, the objective is to execute the corresponding code on a many-core device, like a GPGPU, for which the CUDA language is a natural choice. Since the programming concepts of CilkPlus and OpenMP are very different from those of CUDA, it is desirable to automatically generate optimized CUDA-like code from CilkPlus or OpenMP. In this thesis, we propose an accelerator model for annotated C/C++ code together with an implementation that allows the automatic generation of CUDA code. One of the key features of this CUDA code generator is that it supports the generation of CUDA kernel code where program parameters (like number of threads per block) and machine parameters (like shared memory size) are treated as unknown symbols. Hence, these parameters need not to be known at codegeneration-time: machine parameters and program parameters can be respectively determined when the generated code is installed on the target machine. In addition, we show how these parametric CUDA programs can be optimized at compiletime in the form of a case discussion, where cases depend on the values of machine parameters (e.g. hardware resource limits) and program parameters (e.g. dimension sizes of threadblocks). This generation of parametric CUDA kernels requires to deal with non-linear polynomial expressions during the dependence analysis and tiling phase of the input code. To achieve these algebraic calculations, we take advantage of techniques from computer algebra, in particular in the RegularChains library of Maple. Various illustrative examples are provided together with performance evaluation. Our preliminary implementation uses LLVM, Maple and PPCG; moreover, it successfully processes a variety of standard test-examples.