Towards Hardware-AwareTractable LearningofProbabilisticModels