在Ubuntu 12.04上构建支持CUDA的openMPI找不到cuda.h [英] Building CUDA-aware openMPI on Ubuntu 12.04 cannot find cuda.h

问题描述

我在Ubuntu 12.04上构建了openMPI 1.8.5，并安装了CUDA 6.5并使用默认样本进行测试。我打算在具有以下配置的单个节点上运行：

I am building openMPI 1.8.5 on Ubuntu 12.04 with CUDA 6.5 installed and tested with default samples. I intend to run it on a single node with following configuration:

Dell Precision T7400

双Xeon X5450

Nvidia GT730 / Tesla C1060

Dell Precision T7400
Dual Xeon X5450
Nvidia GT730/Tesla C1060

发出的配置命令为

The configure command issued was

    $ ./configure --prefix=/usr --with-cuda=/usr/local/cuda

在生成的config.log中，很明显配置脚本找不到/ usr / loca / cuda / include中的cuda.h和cuda_runtime_api.h，它们确实存在。

In the generated config.log, it is clear the the configure script was not able to find cuda.h and cuda_runtime_api.h in /usr/loca/cuda/include, which do exist.

对于cuda.h：

For cuda.h:

    configure:73774: checking cuda.h usability
    configure:73774: gcc -std=gnu99 -c -O3 -DNDEBUG    conftest.c >&5
    conftest.c:645:18: fatal error: cuda.h: No such file or directory
    compilation terminated.
    configure:73774: $? = 1
    configure: failed program was:
    | /* confdefs.h */

对于cuda_runtime_api.h：

For cuda_runtime_api.h:

    configure:73857: checking cuda_runtime_api.h presence
    configure:73857: gcc -E   conftest.c
    conftest.c:612:30: fatal error: cuda_runtime_api.h: No such file or directory
    compilation terminated.
    configure:73857: $? = 1
    configure: failed program was:
    | /* confdefs.h */

我试图将路径更改为特定于版本的目录， usr / loca / cuda-6.5 / cuda，但引发了相同的错误。

I tried to change the path to version-specific directory, i.e. /usr/loca/cuda-6.5/cuda but same error was thrown.

我试图继续安装，ompi_info给了

I tried to proceed to install, and the ompi_info gave

    mca:mpi:base:param:mpi_built_with_cuda_support:value:false

有没有人有类似的经验，可以帮助我吗？非常感谢！

Does anybody had similar experience that can help me out? Many thanks!

推荐答案

好的，我想我解决了这个问题。 conftest.c似乎在/ usr / include中寻找cuda.h，而不是假定的/ usr / local / cuda / include。一旦我创建了一个软链接cuda.h和cuda_runtime_api.h，问题就解决了。

OK, I think I fixed the problem. The conftest.c seems to be looking for cuda.h in /usr/include, instead of the supposed /usr/local/cuda/include. The problem is solved once I created a soft link of cuda.h and cuda_runtime_api.h.

这篇关于在Ubuntu 12.04上构建支持CUDA的openMPI找不到cuda.h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！