Application of Many-core Accelerators for Problems in Astronomy and Physics
Naohito Nakasato · 2011
Recently, many-core accelerators are developing so fast that the computing devices attract researchers who are always demanding faster computers.Since many-core accelerators such as graphic processing unit (GPU) are nothing but parallel computers, we need to modify an existing application program with specific optimizations (mostly parallelization) for a given accelerator.In this paper, we describe our problem-specific compiler system for many-core accelerators, specifically, GPU and GRAPE-DR.GRAPE-DR is another many-core accelerators device that is specially targeted scientific applications.In our compiler, we focus a compute intensive problem expressed as two-nested loop.Our compiler asks a user to write computations in the inner-most loop.All details related to parallelization and optimization techniques for a given accelerator are hidden from the user point of view.Our compiler successfully generates the fastest code ever for astronomical N-body simulations with the performance of 2.6 TFLOPS (single precision) on a recent GPU.Another successful application on both GPU and GRAPE-DR is the evaluation of a multi-dimensional integral in quadruple precision.The program generated by our compiler runs at a speed of 15 QD-GFLOPS on GPU and 4 QD-GFLOPS on GRAPE-DR.The performance obtained so far is more than 50-200 times faster than a conventional CPU.