Building Input Adaptive Parallel Applications: A Case Study of Sparse Grid Interpolation
Alin Florindor Murarasu, Josef Weidendorfer · 2012
The well-known power wall resulting in multi-cores requires special techniques for speeding up applications. In this sense, parallelization plays a crucial role. Besides standard serial optimizations, techniques such as input specialization can also bring a substantial contribution to the speedup. By identifying common patterns in the input data, we propose new algorithms for sparse grid interpolation that accelerate the state-of-the-art non-specialized version. Sparse grid interpolation is an inherently hierarchical method of interpolation employed for example in computational steering applications for decompressing high-dimensional simulation data. In this context, improving the speedup is essential for real-time visualization. Using input specialization, we report a speedup of up to 9x over the non-specialized version. The paper covers the steps we took to reach this speedup by means of input adaptivity. Our algorithms will be integrated in fastsg, a library for fast sparse grid interpolation.