高性能计算HPC优化工具.PDFVIP

高性能计算HPC 优化工具 高性能计算HPC/优化工具 优化工具 kepler汇编器 概述 概述 Kepler Assembler是针对nvidia Kepler架构GPU原生汇编器。可以实现从二进制文件到原生汇编语言生成及其 反向过程。能够帮助用户直接控制指令调度以及寄存器使用情况。kepler Assembler让广大用户挖掘Kepler架 构系列GPU极限性能成为可能。用户可以用过以下地址访问 /tools/assembler 使用手册 1. Introduction Kepler Assembler is an almost full-featured assembler created for NVidia’s Kepler Architecture. It can be used to generate binary files (.cubin) from assembly code, or to dump “sass” from a given cubin. Except for some kernels’ descriptive information, including parameter numbers and kernel names, the “sass” format supported by this tool is definitely the same as those dumped from “cuobjdump”. The function of dumping “sass” code from existing cubin is useful, since it is difficult for most to write whole assembly code from scratch. Thus, dumping assembly code from a cubin compiled by nvcc, tuning them, then generating cubin would be a more convenient, efficient and acceptable way to use this tool. Just like asfermi and maxas, we wrote this tool to get extremely high device utility. The tools provided by NVidia didn’t show enough supportive to our all kinds of needs in processes of pursuing ultimate performance. With Kepler assembler, it becomes possible to control kernel’s register usage directly, to adjust executing order of instructions, and even to change the scheduling details of a kernel. From Fermi to Maxwell, NVidia has brought a number of changes to every new architecture generation, for example, new instruction scheduling strategy, new ISA encoding, and so on. Therefore, such kind of assembler could not equip with enough backward compatibility. Maxas cannot be used on Kepler Architecture, and Kepler assembler won’t work at Fermi architecture, vice vasa. 2. Input Assembly File Format If you want to generate a cubin file with your own se