BTB索引散列算法的研究與設(shè)計(jì)
發(fā)布時(shí)間:2018-02-25 01:23
本文關(guān)鍵詞: 分支目標(biāo)緩沖 散列索引 XOR散列函數(shù) 分支目標(biāo)地址預(yù)測 分支預(yù)測 出處:《計(jì)算機(jī)研究與發(fā)展》2014年09期 論文類型:期刊論文
【摘要】:分支誤預(yù)測是影響高性能處理器性能進(jìn)一步提升的一個(gè)主要因素.現(xiàn)代處理器采用分支目標(biāo)緩存(branch target buffer,BTB)預(yù)測分支指令的目標(biāo)地址,BTB的預(yù)測精度受限于其命中率.由于程序中分支指令的分布并不均勻,傳統(tǒng)的BTB索引方式無法充分利用BTB資源,從而造成不必要的沖突缺失,影響分支目標(biāo)地址的預(yù)測精度,采用散列索引方式優(yōu)化訪問映射關(guān)系是有效解決方法之一.當(dāng)前大量文獻(xiàn)研究了cache的訪問方式,但對BTB的散列索引算法的專門探討則顯不足.為了消除分支指令的分布空洞,離散分支指令和BTB條目的固有映射關(guān)系,設(shè)計(jì)了用于BTB索引的XOR散列算法和優(yōu)化的bit-select索引算法,使用概率方法對BTB單組最大映射數(shù)期望的上界作了估計(jì),并對這兩種散列索引算法的效果進(jìn)行了模擬評估.實(shí)驗(yàn)結(jié)果表明,散列映射方式能夠較好地避免BTB沖突缺失造成的預(yù)測失敗,XOR散列算法的離散效果更好.
[Abstract]:Branch error prediction (BTB) is one of the main factors that affect the performance of high performance processors. In modern processors, the accuracy of BTB prediction for branch instruction is limited by its hit ratio. Because of the uneven distribution of branch instructions in the program, Traditional BTB indexing method can not make full use of BTB resources, which causes unnecessary conflicts and affects the accuracy of branch target address prediction. Using hash index to optimize access mapping is one of the effective methods to solve this problem. A large number of literatures have studied the access mode of cache, but the special discussion of BTB's hash index algorithm is insufficient. Based on the inherent mapping relationship between discrete branch instructions and BTB items, a XOR hash algorithm for BTB index and an optimized bit-select index algorithm are designed. The upper bound of the expected maximum number of BTB maps is estimated by using the probability method. The experimental results show that the Hash mapping method can avoid the prediction failure caused by the BTB collision loss and the XOR hash algorithm has better discretization effect.
【作者單位】: 國家高性能集成電路設(shè)計(jì)中心;
【基金】:“核高基”國家科技重大專項(xiàng)基金項(xiàng)目(2009ZX01028-002-001)
【分類號】:TP332
,
本文編號:1532484
本文鏈接:http://sikaile.net/kejilunwen/jisuanjikexuelunwen/1532484.html
最近更新
教材專著