Auto-tuning Dense Matrix Multiplication for GPGPU with Cache