Research Article

A Strategy for Automatic Performance Tuning of Stencil Computations on GPUs

Table 9

Heuristic performance on various input sizes.

Data size per dimension
NVIDIAAMD
32641282565123264128256512

Speedup
Hybrid oracle1.031.041.151.181.190.980.931.121.071.36
Hybrid0.971.041.151.181.190.850.911.111.061.39
Random sample1.001.001.001.001.001.001.001.001.001.00

Correct predictions2777733666