diff --git a/pyop2/transforms/auto_tiling.py b/pyop2/transforms/auto_tiling.py index a2d1d28fd..fdbafde85 100644 --- a/pyop2/transforms/auto_tiling.py +++ b/pyop2/transforms/auto_tiling.py @@ -1675,9 +1675,9 @@ def estimated_exec_time(self, tiling_config): """ T_e_r = tiling_config.operator_tile_descriptions[0][0] (quad_tile_len,) = tiling_config.quad_rowtile_lengths - nwi = tiling_config.nthreads_per_cell nwarps = self.get_effective_warps_per_sm(tiling_config) - nblocks = self.get_effective_blocks_per_sm(tiling_config) + if nwarps == 0: + return (np.inf, 0) nsync = self.get_nsync(tiling_config) effective_global_bw = 21 if nwarps > 8 else 20 * (nwarps / 8) effective_shared_bw = ( @@ -1766,8 +1766,6 @@ def estimated_exec_time(self, tiling_config): ) return (total_time, nsync) - return 4.0 / (nwarps) + nsync / nblocks + nwi / 8 - def __call__(self) -> Tuple[ParametricTiling, ...]: from itertools import product