C标头出现重复符号问题 [英] Duplicate symbol issue with C headers

问题描述

这是我第一次处理CUDA项目，该项目比简单的写入单个源文件和编译例程稍微复杂一些.不出所料，我在使用C标头时遇到了一些问题，即重复的符号.

This is my first time tackling a CUDA project that's slightly more complex than the simple write-single-source-file-and-compile routine. As expected, I'm facing some issues with C headers, namely duplicated symbols.

根据链接器，在多个.cu文件中包含以下头文件会引起冲突:

According to the linker, conflicts arise over the inclusion of the following header file in multiple .cu files:

env_vars.h

#ifndef ENV_VARS_H_
#define ENV_VARS_H_

/*** GLOBAL VARIABLES ***/
unsigned int h_n_osc;
__device__ unsigned int d_n_osc;

/*** CONSTANTS ***/
const double OMEGA_0 = 6.447421494058077e+09;

/* other constants defined in the middle */

#endif

multigpu.cu

#include "env_vars.h"
/* assigns h_n_osc */

adm_matrix.cu

#include "env_vars.h"
/* uses h_n_osc */

在Nsight Eclipse Edition中构建项目会导致链接器抱怨h_n_osc变量被定义了两次:

Building the project in Nsight Eclipse Edition results in the linker complaining about the h_n_osc variable being defined twice:

duplicate symbol _h_n_osc in:
    ./adm_matrix.o
    ./multigpu.o
ld: 1 duplicate symbol for architecture x86_64

通过Internet搜索，我已经意识到将h_n_osc变量的声明移动到multigpu.cu并将其重新声明为adm_matrix.cu中的extern变量(以及以后可能需要的任何地方)解决了问题，实际上做到了.

Searching through the Internet, I've realized that moving the declaration of the h_n_osc variable to multigpu.cu and re-declaring it as an extern variable in adm_matrix.cu (and wherever I might need it later) solves the problem, which in fact it does.

问题已解决，但我想对此进行更深入的研究:

Problem solved, but I'd like to take a deeper look into this:

1. 为什么链接程序也不会抱怨d_n_osc变量?为什么常量(例如OMEGA_0)同样不是问题?
2. 这是否意味着无法将全局变量放在头文件中?
3. 让我最困惑的是，互联网上的许多消息来源都指出，仅当头文件包含变量的定义而其简单的 declaration 应该不构成问题.我很难相信这一点的原因是，即使我的标头只包含一个声明，我也要面对这个问题！我想念什么吗?

1. Why doesn't the linker complain about the d_n_osc variable as well? And why are the constants (such as OMEGA_0) equally not a problem?
2. Does this mean that it is not possible to place global variables in header files?
3. What puzzles me most is that a number of sources over the Internet state that duplicated symbol errors should happen only when the header file contains a definition of the variable, while its simple declaration shouldn't constitute a problem. The reason I have a hard time believing this is that I'm facing the issue even though my header only contains a declaration! Am I missing something?

伙计们，感谢您的耐心等待！

Thanks in advance for your patience, folks!

推荐答案

头文件通常应仅包含 声明性代码. h_n_osc应该在此处声明，而不是定义.

Header files should normally contain only declarative code. h_n_osc should be declared here, not defined.

extern unsigned int h_n_osc;

在至少一个模块中，或者在一个新模块中，您将需要一个定义；例如:

In at least one of your modules, or a new one of its own you will need a definition; for example:

env_vars.cu

#include "env_vars.h"
unsigned int h_n_osc;

然后链接该链接.或者，您当然可以将 definition 放置在现有模块multigpu.cu或adm_matrix.cu中.

Then link that. Alternatively you could of course place the definition in one of the existing modules multigpu.cu or adm_matrix.cu.

我不确定此问题似乎可以解决该问题.

I am not sure of the semantics of the The CUDA __device__ extension, while it may link, it is not necessarily correct; you may end up with each module referencing a separate copy of the device variable; it may be necessary to qualify that with extern as well. This question appears to deal with that issue.

这篇关于C标头出现重复符号问题的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！