This document describes an attempt to develop a compiler-based approach for computations with symmetric tensors. Given a computation and the symmetries of its input tensors, we derive formulas for random access under a storage scheme that eliminates redundancies; construct intermediate representations to describe the loop structure; and translate this information, using the taco tensor algebra compiler, into code. While we achieve a framework for reasoning about a fairly general class of symmetric computations, the resulting code is not performant when the symmetries are misaligned.
翻译:本文件描述为开发一个基于编译器的计算对称分导体的计算方法而尝试的尝试。 鉴于其输入数导体的计算和对称性, 我们根据一个消除冗余的存储方案为随机访问得出公式; 构建中间表达面来描述循环结构; 并将这一信息, 使用 taco 高压代数编译器, 转换成代码 。 虽然我们实现了关于相当一般的对称计算等级的推理框架, 但当对称对称发生偏差时, 产生的代码并不起作用 。