Heterogeneous computing using FPGA accelerators is a promising approach to boost the performance of application programs within given power consumption. This paper focuses on optimizations targeting FPGA-based reconfigurable dataflow computing platform, and shows how they benefit an application. In order to evaluate them, we use the Himeno benchmark, which is a floating point computation kernel known to be bound by memory bandwidth. To understand the performance characteristics of the benchmark, we compare it with the current state-of-the-art implementation on GPUs. From the results, we find that our implementation with specialized dataflow pipelines outperforms the current state-of-the-art GPU implementations by making full use of memory locality.