CAJ | 학술논문

讨论在Fermi结构GPU使用CUDA对GEMM（单精度和双精度）算法进行优化，以及Fermi体系结构的新特性（如缓存）对性能的影响。GPU缓存一方面可以提高处理器在运行时数据访问的局部性，另一方面使得代码性能对与性能相关算法参数的依赖变得不可预测。自动优化技术可以用来解决这一问题。自动优化的SGEMM和DGEMM代码在Tesla C2050 GPU上达到了563GFlops和253GFlops的性能。代码使用CUDA和C语言进行实现，未进行二进制代码级别的优化。
토론재Fermi결구GPU사용CUDA대GEMM（단정도화쌍정도）산법진행우화，이급Fermi체계결구적신특성（여완존）대성능적영향。GPU완존일방면가이제고처리기재운행시수거방문적국부성，령일방면사득대마성능대여성능상관산법삼수적의뢰변득불가예측。자동우화기술가이용래해결저일문제。자동우화적SGEMM화DGEMM대마재Tesla C2050 GPU상체도료563GFlops화253GFlops적성능。대마사용CUDA화C어언진행실현，미진행이진제대마급별적우화。
Automatic optimization of GEMM algorithm on Fermi GPU and the impact of Fermi GPU’s architectural fea-tures on performance are discussed in this paper. The cache on GPU can not only improve the data access locality of proces-sors,but also make the code performance relying on relative algorithm parameters unpredictable. Auto-tuning can be used to solve this problem. Auto-tuned SGEMM and DGEMM codes achieve 563 GFlops and 253 GFlops respectively on C2050 GPU. The codes are implemented by using CUDA and C language,but the optimization on the binary level is not involved.