【摘要】 当前GPU的体系结构为高性能计算提供了良好的可编程性。为了得到众核GPU高性能程序设计的一般方法,探索GPU程序性能优化技术,对在GPU上进行高性能程序设计的经验进行了总结。通过基准测试,得到GPU性能指标,对GPU程序设计进行指导。使用CUDA对单精度矩阵乘法和FFT进行性能优化,前一个算法是计算密集型任务,后一个算法是带宽密集型任务。在NVIDIA GeForce GTX280 GPU上,矩阵乘法算法达到393 Gflop/s的峰值速度,比CUBLAS 2.0数学库提高了5%;对于一些维度的FFT计算也取得了较好的性能。更多还原

【Abstract】 The optimization technique of GPU program performance is investigated for obtaining the common method to design many-core GPU high-performance program.The authors′ experiences in improving the performance of two key algorithms: single-precision matrix-matrix multiplication subprogram(SGEMM of BLAS) and single precision FFT using CUDA are discussed in this paper.The former is computation intensive,while the latter is memory bandwidth or communication-intensive.The peak speed of 393 Gflops was achieved on NVIDIA GeForce GTX280 GPU for the former.It is about 5% faster than the CUBLAS 2.0 library.Better FFT performance was obtained for a range of dimensions.Some common principles are discussed for the design and implementation of many-core algorithms.更多还原