Reproducible and Accurate Matrix Multiplication for GPU Accelerators