Accelerating Big Integer Arithmetic Using Intel IFMA Extensions
Accelerating Big Integer Arithmetic Using Intel IFMA Extensions
复制标题
使用英特尔 IFMA 扩展加速大整数运算
DOI：
--
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
V. Krasnov
中科院分区：
文献类型：
--
作者：
S. Gueron;V. Krasnov
英特尔最近宣布了一套新的处理器说明，称为AVX512IFMA，该说明执行了整数融合了乘以累积的操作。这些说明在512位寄存器上运行，并计算八个独立的52位未签名的整数乘数，以生成八个104位产品，并将其低/高半堆积到64位容器中。使用这些指令，要求输入转换为（冗余形式）radix 252，然后将输出转换为所需的表示形式。本文展示了几种利用AVX512IFMA指令以加快大型乘法的技术。尽管在撰写本文时尚未提供支持AVX512IFMA的处理器，但我们展示了当前可用的公共工具如何用于估计其潜在的性能益处。例如，基于这些工具，我们希望与当前最佳的方法相比，1024位整数乘法的速度为2倍。
Intel has recently announced a new set of processor instructions, dubbed AVX512IFMA, that carry out Integer Fused Multiply Accumulate operations. These instructions operate on 512-bit registers and compute eight independent 52-bit unsigned integer multiplications, to generate eight 104-bit products, and accumulate their low/high halves into 64-bit containers. Using these instructions requires that inputs are converted to (redundant form) radix 252, and outputs are converted to the desired representation. This paper demonstrates several techniques for leveraging the AVX512IFMA instructions in order to speed up big-integer multiplications. Although processors that support AVX512IFMA are not yet available at the time this paper is written, we show how currently available public tools can be used for estimating their potential performance benefits. For example, based on these tools, we expect a 2x speedup for 1024-bit integer multiplication, over the best currently available method.