Auto-tuning for large-scale image processing by dynamic analysis method on multicore platforms

This paper describes a general-purpose method of improving execution performance of the in-memory data, particularly in the case of large-scale image processing on different multicore platforms. To process large-scale arrays, the method of tiling is widely used to achieve high performance. However, frequently accessing the memory system by multithreads is bound to cause system bottleneck. Our optimisation strategies are automatic thread scheduling and data/task partitioning. Those methods that attempt to take advantage of spatial and temporal locality can reduce memory traffic remarkably. According to the hardware configurations, a scheduler automatically partitions the images into tiled blocks of pre-determined size. Then it fuses all the operations for the same blocks to reduce the rate of cache miss. The parallel task execution is more effective than other traditional parallel libraries, such as openMP. Moreover, the optimisation on space-filling curves that optimises the locality of neighbouring tiled blocks can also contribute to the fast memory access.