Failed to configure with PGI and CUDA unless CC=pg++

[kingchc]

What version of hwloc are you using?

From github source, commit 2742199

Which operating system and hardware are you running on?

CentOS
$ uname -a
Linux xxx 3.10.0-229.el7.x86_64 #1 SMP Fri Mar 6 11:36:42 UTC 2015 x86_64 x86_64 x86_64 GNU/Linux

CPU: Intel(R) Xeon(R) CPU E5-2687W v3 @ 3.10GHz (Haswell)

Details of the problem

I'm trying to use PGI compiler to build hwloc with CUDA support. I got an error message during configuration. Followings are command line I used and the output.

$ ./configure CC=pgcc 'CPP=pgcc -E' CXX=pgc++ F77=pgf77 FC=pgf90 --enable-cuda CPPFLAGS=-I/usr/local/cuda-9.0/include CFLAGS="-I/usr/local/cuda-9.0/include"
...<skip>...
checking cuda.h usability... yes
checking cuda.h presence... yes
checking for cuda.h... yes
checking if CUDA_VERSION >= 3020... yes
checking for cuInit in -lcuda... yes
checking cuda_runtime_api.h usability... no
checking cuda_runtime_api.h presence... no
checking for cuda_runtime_api.h... no
configure: WARNING: Specified --enable-cuda switch, but could not
configure: WARNING: find appropriate support
configure: error: Cannot continue

Here it shows it cannot find cuda_runtime_api.h, but I have confirmed that the file is there at /usr/local/cuda-9.0/include. I have tried different version of PGI (17.10, 18.1, 18.3), but none of them work. However, it works fine with GNU compiler.

Any suggestion is appreciated.
Thanks,

[bgoglin]

Hello
Look for cuda_runtime_api.h in config.log to find out the entire compile line that failed.
In theory, your CFLAGS isn't needed (it's only for compiler flags), only CPPFLAGS should contain -I/path/to/foo/bar.
You could also try setting INCLUDE_PATH=/usr/local/cuda-9.0/include and/or C_INCLUDE_PATH=/usr/local/cuda-9.0/include but I don't know if pgcc honors these variables.

[kingchc]

Thanks for the reply. I have tried INCLUDE_PATH and C_INCLUDE_PATH, unfortunately, none of them works. Looking into the config.log, I found that it's using pgcc instead of pgc++ when processing cuda-related stuff. I guess pgcc somehow is not compatible to CUDA.

configure:24319: result: no
configure:24319: checking cuda_runtime_api.h presence
configure:24319: pgcc -E -I/opt/cuda/9.0/include conftest.c
PGC-F-0249-#error --  --- !!! UNKNOWN COMPILER: please provide a CUDA compatible definition for '__align__' !!! --- (/opt/cuda/9.0/include/crt/host_defines.h: 133)
PGC/x86-64 Linux 17.10-0: compilation aborted

I did some search and apply the following patch to make it work.

$ git diff
diff --git a/config/hwloc.m4 b/config/hwloc.m4
index 63bf11d6..dd9c150c 100644
--- a/config/hwloc.m4
+++ b/config/hwloc.m4
@@ -826,6 +826,7 @@ return clGetDeviceIDs(0, 0, 0, NULL, NULL);
     # CUDA support
     hwloc_have_cuda=no
     hwloc_have_cudart=no
+    AC_LANG_PUSH([C++])
     if test "x$enable_io" != xno && test "x$enable_cuda" != "xno"; then
       AC_CHECK_HEADERS([cuda.h], [
         AC_MSG_CHECKING(if CUDA_VERSION >= 3020)
@@ -860,6 +861,7 @@ return clGetDeviceIDs(0, 0, 0, NULL, NULL);
           ])
         ])
       ])
+    AC_LANG_POP([C++])

       AS_IF([test "$enable_cuda" = "yes" -a "$hwloc_have_cudart" = "no"],
             [AC_MSG_WARN([Specified --enable-cuda switch, but could not])

I hope this will not broke anything.
Thanks,

[bgoglin]

If pgcc doesn't support cuda_runtime_api.h, then some parts of hwloc itself won't build with pgcc (at least topology-cuda.c and tests/hwloc/cuda*.c), and you would need to use CC=pg++ for building hwloc.

[YuhaoYHH]

I'm having the same issue. I checked PGI's installation guide and found that any version pf PGI doesn't support Intel Xeon CPU... Have you solved the issue yet?

[bgoglin]

Hello @ChingKing and @YuhaoYHH
Does this problem still occur?

[kingchc]

Hello @ChingKing and @YuhaoYHH
Does this problem still occur?

I applied the above-mentioned patch to my local repo and proceed. I have confirmed that using CC=pg++ also works. Hope this helps.

[nemequ]

FWIW, it's pretty straightforward to add a PGI build to Travis; see https://github.com/nemequ/pgi-travis

[bgoglin]

This bug is very old and building with your configure line works fine now (just tested with PG compilers 22.3-0). Hence I am closing this.