Neither More Nor Less: Optimizing Thread-level Parallelism for GPGPUs