GPU SDK 3.0-beta1 on GCC4.4 Problem with stdarg.h appears with gcc4.4

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi there,

we are running teslas on a ubuntu 9.10 installation with gcc4.4.
This works just fine with the 2.3 version of the SDK when -O2 is
not used for compilation. Our hope was, that
the 4.4 support would get into the next release of the SDK.

However the code is still only compilable without -O2, and it doesn’t
work for some examples at all.

dct8x8 fails due the newly introduced shrUtils.h stuff.
cudafe++ inserts a __builtin_stdarg_start, which is no longer available in
GCC 4.4 (GCC 4.4 Release Series — Changes, New Features, and Fixes - GNU Project):
cudafe++ --m64 --gnu_version=40401 --diag_error=host_device_limited_call --diag_error=ms_asm_decl_not_allowed --parse_templates --gen_c_file_name “/tmp/tmpxft_0000237c_00000000-1_dct8x8.cudafe1.cpp” --stub_file_name “/tmp/tmpxft_0000237c_00000000-1_dct8x8.cudafe1.stub.c” --stub_header_file_name “/tmp/tmpxft_0000237c_00000000-1_dct8x8.cudafe1.stub.h” “/tmp/tmpxft_0000237c_00000000-10_dct8x8.cpp4.ii”

→

/tmp/tmpxft_0000237c_00000000-1_dct8x8.cudafe1.cpp:22293:__builtin_stdarg_start(__args,__f
mt);

This is easily fixed by replacing builtin_stdarg_start by builtin_va_start, however doing this manually would be a little
tedious. Would it be possible to fix this, or workaround this?

The only thing, I can think of right now would be a wrapper around nvcc, which uses nvcc --dryrun to create a script and insert a
sed after each cudafe++ call before executing the script. But this is a little bit clumsy and not very nice.

Thanks!

2

gcc 4.4 isn’t supported yet, so you’re best off rolling back to gcc 4.3 and waiting for 4.4 support (which, as far as I know, is not in 3.0).

3

Well, fine. I know its not supported.

Just tinkering around to see whats possible.

Anyway builtin_stdarg_start seems to be deprecated since gcc4.0, so it might be a good

idea to change it anyway?

If it doesn’t work with 4.4, thats fine. I’m not complaining ;) Just looking for ideas.

Thanks for your quick reply.

4

OK,
so I can compile the SDK 3.0 examples without -O2 and the attached script

nvcc4.4.gz (218 Bytes)

(which is very crude) wrapping nvcc.

5

haraldkl,

Thanks for putting together this script. I have switched to CUDA Toolkit and SDK 3.0 now that’s it’s no longer in beta. I am running Ubuntu 9.10 with gcc4.4, and don’t really want to install an older version of gcc just for the purposes to compiling the SDK examples. Here’s what I have done by way of installation:

  1. I gunzip’ped the file, and copied the nvcc4.4 script to the same directory as nvcc (in my case, /usr/local/cuda/bin), and set executable permissions on the file.

  2. I edited nvcc4.4 to point to the location of nvcc on my system (/usr/local/cuda/bin/nvcc).

  3. I edited NVIDIA_GPU_Computing_SDK/C/common/common.mk for the following line: “NVCC := $(CUDA_INSTALL_PATH)/bin/nvcc4.4”

But I get the following error message from “make verbose=1”

/usr/local/cuda/bin/nvcc4.4 -gencode=arch=compute_10,code="sm_10,compute_10" -gencode=arch=compute_20,code="sm_20,compute_20" --compiler-options -fno-strict-aliasing --compiler-options -fno-inline -po maxrregcount=16 -I. -I/usr/local/cuda/include -I…/…/common/inc -I…/…/…/shared//inc -DUNIX -O0 -o obj/x86_64/release/BlackScholes.cu.o -c BlackScholes.cu

sed: can’t read /tmp/gedit.root.2710946334: No such device or address

sed: couldn’t edit /tmp/keyring-PpTVRw: not a regular file

sed: can’t read /tmp/gedit.root.2710946334: No such device or address

sed: couldn’t edit /tmp/keyring-PpTVRw: not a regular file

sed: can’t read /tmp/gedit.root.2710946334: No such device or address

sed: couldn’t edit /tmp/keyring-PpTVRw: not a regular file

sed: can’t read /tmp/gedit.root.2710946334: No such device or address

sed: couldn’t edit /tmp/keyring-PpTVRw: not a regular file

sed: can’t read /tmp/gedit.root.2710946334: No such device or address

sed: couldn’t edit /tmp/keyring-PpTVRw: not a regular file

/usr/include/c++/4.4/x86_64-linux-gnu/bits/c++locale.h: In function ‘int std::__convert_from_v(__locale_struct* const&, char*, int, const char*, …)’:

/usr/include/c++/4.4/x86_64-linux-gnu/bits/c++locale.h:86: error: ‘__builtin_stdarg_start’ was not declared in this scope

make[1]: *** [obj/x86_64/release/BlackScholes.cu.o] Error 1

make[1]: Leaving directory `/home/cmejia/NVIDIA_GPU_Computing_SDK/C/src/BlackScholes’

make: *** [src/BlackScholes/Makefile.ph_build] Error 2

Do you or anyone else know what is going wrong? I appreciate the fact that you took the time to put together this clever wrapper, but unfortunately I don’t know enough about sed or nvcc to figure out what is going on.

Thanks in advance,

–Chris

6

Hi,

the script should create a file named .cudacompile.tmp.sh by running the original nvcc with the --dry-run option and subsequently perform

some sed scripts upon this. In your case this doesn’t seem to happen. Instead it is trying to act on a file /tmp/gedit.root.2710946334 and

/tmp/keyring-PpTVRw, which is very strange. Are you sure, the script is not corrupted?

I didn’t test the newly released files, so it might also be due to changed behaviour of nvcc --dry-run, though I doubt it.

7

Hi haraldkl,

I just had the problem too – seems the second-to-last line of your script tries to operate on all files in /tmp:

[codebox]sed -i “/^ cudafe++/a sed -i ‘s/__builtin_stdarg_start/__builtin_va_start/g’ /tmp/” .cudacompile.tmp.sh[/codebox]

I haven’t yet figured out what filename pattern to replace your “/tmp/*” with yet, but a quick workaround (if you don’t need any of the stuff in /tmp) is to just delete everything from /tmp between runs.

8

Ah, right, sorry.

You need to modify the cudafe++ generated files, which will be in /tmp.

As I said the workaround is very crude and I didn’t bother finding some more elegant way to identify the generated files…

Anyway the files in question should be modified, even if there is other stuff in /tmp.

You could have look into the .cudacompile.tmp.sh script and the cudafe++ generated files in /tmp and try to figure out, why there

is still a stdarg instead of a va in the files handed over to gcc.

9

I worked around the problem in a simple way. Just add this option to the nvcc command:

-Xcompiler -D__builtin_stdarg_start=__builtin_va_start

In this way the bad symbol __builtin_stdarg_start will be replaced with the correct symbol __builtin_va_start just by the preprocessor.

10

Hey, that’s great! This is obviously a much better solution ;)

It seems I’m not working enough with preprocessors.

Anyway, I put this in nvcc.profile:

[codebox]INCLUDES += “-D__builtin_stdarg_start=__builtin_va_start”[/codebox]

as the documentation says, this flag is passed with the -Xcompiler flag by nvcc.

It looks like the Cuda SDK 3.0 works fine with that.

Is there any better place to put this?

Like a CFLAGS in the same file?

11

I put it in the common.mk:

[codebox]

NVCCFLAGS := --compiler-options -fno-inline -Xcompiler -D__builtin_stdarg_start=__builtin_va_start

[/codebox]

This does the job.

Thanks to the previous posters for this hint!