A Strategy for Automatic Performance Tuning of Stencil Computations on GPUs

Table 4

Breakdown by data loading technique (NVIDIA).

Data loading technique	Speedup			Best count
	Dims	Opts	Hybrid	Dims	Opts	Hybrid
Global without vectorization	0.54	0.54	0.54	0	0	0
Global with vectorization	0.45	0.45	0.45	3	1	2
Image	1.00	1.05	1.04	13	40	36
Local	0.76	0.87	0.86	6	21	20