提出了一种新的适用于处理器的硅前性能验证平台的基准程序实现方法。方法的主要思想是利用现成的广泛使用的测试程序集合,通过降低工作负载,采用基于基本块的划分、归并方式,将多个基于相同特征点的代码片段作为一个基准检测点,这些抽象的检测点构成了基准程序库。该方法将复杂的处理器内部行为的一致性判断转换为性能的宏观统计分析,充分利用了已有的权威测试基准集,无需重新编写性能验证平台的基准程序,既扩大了验证程序的规模,又节省了大量的劳动,同时可以针对验证样本通过分析系统自动展开验证工作,减少了人工核对的工作量。
A method to construct a middle scale testbench for processors' performance debug during the pre-silicon verification stage was proposed. This approach makes use of the prevalent processor benchmarks and replaces the complete test with light workloads. The instruction streams are clustered into several checkpoints according to basic block partition and categorization, and all of these checkpoints make up a novel benchmark. This approach converts the complex performance diagnostic method, which is based on waveform observation of the elaborate architecture, to statistical analysis of checkpoints. This method does not require a new medium scale testbench, so extra work and cost can be avoided. Moreover this can prompt automated verification process with less manual inspection.