def test(tile_size, n):
opt_args = [test_in] + [test_in]*NDIM + [test_opt_out]
compiler = CCompiler()
compiler.tile_size = tile_size
optimized = analyzer.repackage(sten, NDIM, n)
kern = compiler.compile(optimized)
kern(*opt_args)
t = -time.time()
kern(*opt_args)
t += time.time()
print(tile_size, n, t/ITER)
