一句话：提出 Graphene，一种能表达 GPU 张量指令所需复杂数据-线程映射的张量 IR，用于优化张量计算。
现代 GPU 硬件加速多维张量的计算与数据移动，但用 CUDA C++ 等语言编写优化张量计算对专家也极具挑战，因为这些语言围绕一维平坦缓冲区，缺乏对多维数据和线程的合理抽象。现有张量 IR 表达力不足，无法表示 GPU 张量指令所需的复杂数据到线程映射。本文提出 Graphene，一种针对 GPU 上优化张量计算的 IR，旨在填补这一表达力空白。摘要未提供具体方法与评估结果。
Modern GPUs accelerate computations and data movements of multi-dimensional tensors in hardware. However, expressing optimized tensor computations in software is extremely challenging even for experts. Languages like CUDA C++ are centered around flat buffers in one-dimensional memory and lack reasonable abstractions for multi-dimensional data and threads. Existing tensor IRs are not expressive enough to represent the complex data-to-thread mappings required by the GPU tensor instructions.
@inproceedings{paperbot10074,
title = {Graphene: An IR for Optimized Tensor Computations on GPUs},
author = {Bastian Hagedorn and Bin Fan and Hanfeng Chen and Cris Cecka and Michael Garland and Vinod Grover},
booktitle = {ASPLOS},
year = {2023},
doi = {10.1145/3582016.3582018}
}