A Strategy for Automatic Performance Tuning of Stencil Computations on GPUs

Table 1

Optimization abbreviations and allowable values. N denotes the input/output array size in each dimension. X, Y, and Z correspond to the , , and dimensions of the work-item grid, respectively.

Optimization	Abbreviation	Range of values
		Minimum	Maximum
Work-group size	WX, WY, WZ	1	NX, NY, NZ
Block merge factor	BX, BY, BZ	1	NX, NY, NZ
Cyclic merge factor	CX, CY, CZ	1	NX, NY, NZ
Vector width	VX	1	16
Local memory	N/A	0 (no)	1 (yes)
Image memory	N/A	0 (no)	1 ( yes)