Accelerating Tsunami simulation with FPGA and GPU through automatic compilation

There have been lots of efforts in accelerating computation with FPGA and GPU. In this talk our results on accelerating Tsunami simulation with FPGA and GPU are reported. The approaches to acceleration are a little bit different between with FPGA and with GPU. In both cases, starting with the commonly used Tsunami simulation program, the program has been modified differently for FPGA and GPU. For GPU we use typical approach using CUDA compiler framework. A series of transformation applied to the original program realizes better use of GPU and finally the simulation is speed up by 8 times over single cores. In the case of FPGA, we manually extract large data flow graphs (DFGs) from the program, and they are compiled into FPGA circuits automatically by a commercially available compiler. The key issue here is how large DFG without any control can be extracted which needs some analysis on the original definition of Tsunami simulation, i.e., its differential equations. With this approach 25 times speed up over single cores has been realized.