Benchmarking GPUs to tune dense linear algebra