Fails to compile on NVCC CUDA 11.0 #53

ghost · 2021-02-04T13:06:40Z

The following code snippet will not compile in Nvidia's compiler.

  #include <boost/signals2/signal.hpp>
  
  // TODO: This is an example of a library function
  void fnBoostPreprocessorHeader()
  {
  }

The compiler chokes somewhere inside of the include statement without any additional code except the snippet above. The following error is returned:

C1012 unmatched parenthesis: missing ')' BoostPreprocessorHeader in C:\vcpkg\installed\x64-windows-static\include\boost\preprocessor\slot\detail\shared.hpp on line 27

I have spoken to the nvidia compiler group and they responded with the following:

We checked the issue and confirmed that it is a Boost issue .The error is emitted by Microsoft cl.exe compiler. Something in the    Boost headers is checking for one of the CUDA macros (__CUDACC__ / __NVCC__) and enabling code that is not accepted by the Microsoft compiler used for preprocessing.
We suggest you track a ticket to Boost with this reproducer . Thank you.

Nvidia

The text was updated successfully, but these errors were encountered:

fmhess · 2021-02-05T17:08:44Z

Would you try adding

#define BOOST_PP_VARIADICS 0

at the top of your file before the include? And also try setting it to 1? I ask this because Boost.Preprocessor also has a check for CUDA in it, related to enabling variadic support.

ghost · 2021-02-09T19:14:06Z

@fmhess Neither variation of your suggestion worked. Any other work around suggestions?

fmhess · 2021-02-09T21:39:38Z

On Tue, Feb 9, 2021 at 11:14 AM dahubley ***@***.***> wrote: @fmhess Neither variation of your suggestion worked. Any other work around suggestions?

Well, I do development on Linux but I do have a windows machine. Is it possible for me to download a free nvcc/msvc environment so I can reproduce the problem? Boost doesn't seem to have any nvcc setups in their array of test machines. From your error message, it seems BOOST_PP_VALUE is getting set to something strange in boost/preprocessor/slot/detail/shared.hpp . So you could try narrowing this down to a Boost.Preprocessor issue, it must be happening during the expansion of one of the BOOST_PP macros used in a signals2 header.

…

-- Frank

ghost · 2021-02-09T21:56:20Z

@fmhess I below is a link to the NVCC compiler.
NVCC Download

You can install the visual studio plugin to make the set up easier.

Would it be possible to get boost to start compiling against NVCC? Heterogeneous computing (i.e. GPGPU) is growing in popularity in various fields where C++ is the dominate language. Also hiding dependencies can be a very laborious process for zero gain.

Thanks for your help!
Drew

fmhess · 2021-02-09T23:41:55Z

I'll try out nvcc tomorrow. I just maintain Signals2, so you should probably take your more general concerns to the boost-devel mailing list. You could volunteer to run the boost regression tests on your platform (see https://www.boost.org/development/testing.html ) then all the library maintainers would be able to see if their tests pass on nvcc on windows. Actually, something else you could try now would be to run the regression tests for Boost.Preprocessor and see if they pass.

…

On Tue, Feb 9, 2021, 16:56 dahubley ***@***.***> wrote: @fmhess <https://github.com/fmhess> I below is a link to the NVCC compiler. NVCC Download <https://developer.nvidia.com/cuda-downloads> You can install the visual studio plugin to make the set up easier. Would it be possible to get boost to start compiling against NVCC? Heterogeneous computing (i.e. GPGPU) is growing in popularity in various fields where C++ is the dominate language. Also hiding dependencies can be a very laborious process for zero gain. Thanks for your help! Drew — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#53 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AARKTXYGAOMBY64BRBKZIN3S6GVRJANCNFSM4XCZNEIA> .

eldiener · 2021-02-10T18:45:21Z

Boost preprocessor now assumes variadic macro support so BOOST_PP_VARIADICS is always set to 1.

ghost · 2021-02-10T18:50:30Z

@eldiener is this a recent change?

fmhess · 2021-02-11T03:55:38Z

I installed the latest nvcc 11.2, visual studio community 2019, and boost-1.75 and I couldn't reproduce your problem:

PS C:\Users\fhess\source\repos\CudaRuntime1\CudaRuntime1> nvcc -I C:\Users\fhess\Downloads\boost_1_75_0\boost_1_75_0\ -c .\Source.cpp
Source.cpp
PS C:\Users\fhess\source\repos\CudaRuntime1\CudaRuntime1> ls


    Directory: C:\Users\fhess\source\repos\CudaRuntime1\CudaRuntime1


Mode                LastWriteTime         Length Name
----                -------------         ------ ----
d-----        2/11/2021   2:44 AM                x64
-a----        2/11/2021   3:18 AM           4721 CudaRuntime1.vcxproj
-a----        2/11/2021   2:40 AM            168 CudaRuntime1.vcxproj.user
------        2/11/2021   2:40 AM           3819 kernel.cu
-a----        2/11/2021   3:22 AM            131 Source.cpp
-a----        2/11/2021   3:23 AM           2140 Source.obj


PS C:\Users\fhess\source\repos\CudaRuntime1\CudaRuntime1> nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Thu_Jan_28_19:41:49_Pacific_Standard_Time_2021
Cuda compilation tools, release 11.2, V11.2.142
Build cuda_11.2.r11.2/compiler.29558016_0

Can you try nvcc 11.2 and boost-1.75 and see if your problem still exists?

ghost · 2021-02-11T14:51:12Z

Unfortunately we are no closer to finding the issue. I ran the compiler on the file from command line. The result was the same. I have added to cl.exe verison, CUDA verison, and compiler output for you.

PS C:\Users\Drew Hubley\source\repos\BoostPreprocessorHeader\BoostPreprocessorHeader> nvcc -I "C:\Users\Drew Hubley\Downloads\boost_1_75_0\boost_1_75_0" -c func.cu
func.cu
C:/Users/Drew Hubley/Downloads/boost_1_75_0/boost_1_75_0\boost/preprocessor/slot/detail/shared.hpp(27): fatal error C1012: unmatched parenthesis: missing ')'
PS C:\Users\Drew Hubley\source\repos\BoostPreprocessorHeader\BoostPreprocessorHeader> nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2020 NVIDIA Corporation
Built on Mon_Nov_30_19:15:10_Pacific_Standard_Time_2020
Cuda compilation tools, release 11.2, V11.2.67
Build cuda_11.2.r11.2/compiler.29373293_0
PS C:\Users\Drew Hubley\source\repos\BoostPreprocessorHeader\BoostPreprocessorHeader> cl.exe --version
Microsoft (R) C/C++ Optimizing Compiler Version 19.27.29111 for x64
Copyright (C) Microsoft Corporation.  All rights reserved.

cl : Command line warning D9002 : ignoring unknown option '--version'
cl : Command line error D8003 : missing source filename
PS C:\Users\Drew Hubley\source\repos\BoostPreprocessorHeader\BoostPreprocessorHeader> cl.exe
Microsoft (R) C/C++ Optimizing Compiler Version 19.27.29111 for x64
Copyright (C) Microsoft Corporation.  All rights reserved.
PS C:\Users\Drew Hubley\source\repos\BoostPreprocessorHeader\BoostPreprocessorHeader>

fmhess · 2021-02-11T18:30:41Z

Ok, I can reproduce the problem if I change the source code file extension to ".cu". What I see is BOOST_PP_VALUE is set to BOOST_PP_TUPLE_ELEM_O_1(BOOST_PP_TUPLE_ELEM_O_1(2, 0, (3,(0,10,<boost/function/detail/function_iterate.hpp>))), 0, BOOST_PP_TUPLE_ELEM_O_1(2, 1, (3,(0,10,<boost/function/detail/function_iterate.hpp>)))) after preprocessing. When compiling a ".cpp" file, BOOST_PP_VALUE evaluates to 0. So it looks like the BOOST_PP_TUPLE_ELEM_0_1 macro is failing to evaluate, it must be undefined when compiling a cuda file?

ghost · 2021-02-11T19:05:27Z

The CUDA compiler will default to the MSVC compiler when asked to compile a cpp file. I think the opening message with the quote from Nvidia might be a clue to what is happening. Is there anything NVCC specific in the macros?

eldiener · 2021-02-11T19:54:17Z

There is no macro called BOOST_PP_TUPLE_ELEM_0_1 or BOOST_PP_TUPLE_ELEM_O_1 in the preprocessor library.

fmhess · 2021-02-11T19:58:50Z

The problem can be generated by preprocessing the following code with nvcc (in a file with the ".cu" extension. The code is adapted from boost/function.hpp ``` #include <boost/preprocessor/iterate.hpp> # define BOOST_PP_ITERATION_PARAMS_1 (3,(0,10,<boost/function/detail/function_iterate.hpp>)) # include BOOST_PP_ITERATE() # undef BOOST_PP_ITERATION_PARAMS_1 #endif ```

…

-- Frank

eldiener · 2021-02-11T20:03:54Z

Which version of nvcc and what is the nvcc command line ?

fmhess · 2021-02-11T20:11:00Z

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Thu_Jan_28_19:41:49_Pacific_Standard_Time_2021
Cuda compilation tools, release 11.2, V11.2.142
Build cuda_11.2.r11.2/compiler.29558016_0

nvcc -I C:\Users\fhess\Downloads\boost_1_75_0\boost_1_75_0 -E .\File.cu > myout.cpp
File.cu
C:/Users/fhess/Downloads/boost_1_75_0/boost_1_75_0\boost/preprocessor/slot/detail/shared.hpp(40): fatal error C1012: unmatched parenthesis: missing ')'

And here is an unmangled and corrected version of the contents of the File.cu test program

#include <boost/preprocessor/iterate.hpp>

#  define BOOST_PP_ITERATION_PARAMS_1 (3,(0,10,<boost/function/detail/function_iterate.hpp>))
#  include BOOST_PP_ITERATE()
#  undef BOOST_PP_ITERATION_PARAMS_1

eldiener · 2021-02-11T23:35:22Z

The example runs fine with all compilers I have tested. Unfortunately there is no toolset for nvcc in bjam. This suggests some sort of bug in nvcc, most likely in its preprocessor. Maybe nvcc is emulating VC++ and VC++ has numerous workarounds to make its default non-standard preprocessor work with the preprocessor library. I could change the preprocessor configuration file for nvcc so that it acts like VC++. But what you are telling me is that nvcc when compiling a .cpp file works fine but when compiling a .cu it does not work fine, so this suggests something else. Maybe nvcc only emulates VC++ when compiling a .cu file, otherwise it invokes its own compiler.

fmhess · 2021-02-12T00:40:12Z

nvcc only defines __CUDACC__ when compiling .cu files (see https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html ). There are various places in Boost.Config and one place in Boost.Preprocessor where it checks for __CUDACC__. Also this bit from include/boost/config/compiler/nvcc.hpp might be relevant:

//
// When compiing .cu files, there's a bunch of stuff that doesn't work with msvc:
//
#if defined(_MSC_VER)
#  define BOOST_NO_CXX14_DIGIT_SEPARATORS
#  define BOOST_NO_CXX11_UNICODE_LITERALS
#endif

ghost · 2021-02-12T02:23:15Z

NVCC complies CPP files using the VC++ compiler. Thus, is why it works, boost is tested again the underlying compiler in that case. NVCC uses it own toolchain only when compiling .cu files or using the -x option.

…

On Thu., Feb. 11, 2021, 8:40 p.m. Frank Mori Hess, ***@***.***> wrote: nvcc only defines __CUDACC__ when compiling .cu files (see https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html ). There are various places in Boost.Config and one place in Boost.Preprocessor where it checks for __CUDACC__. Also this bit from include/boost/config/compiler/nvcc.hpp might be relevant: // // When compiing .cu files, there's a bunch of stuff that doesn't work with msvc: // #if defined(_MSC_VER) # define BOOST_NO_CXX14_DIGIT_SEPARATORS # define BOOST_NO_CXX11_UNICODE_LITERALS #endif — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#53 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIPDMWQOK6F76HVJW4SPAI3S6R2HVANCNFSM4XCZNEIA> .

ghost · 2021-02-12T02:58:07Z

Also, it seems like really poor practice to add compiler switches to code that is not testing against that toolchain. Thoughts?

…

On Thu., Feb. 11, 2021, 8:40 p.m. Frank Mori Hess, ***@***.***> wrote: nvcc only defines __CUDACC__ when compiling .cu files (see https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html ). There are various places in Boost.Config and one place in Boost.Preprocessor where it checks for __CUDACC__. Also this bit from include/boost/config/compiler/nvcc.hpp might be relevant: // // When compiing .cu files, there's a bunch of stuff that doesn't work with msvc: // #if defined(_MSC_VER) # define BOOST_NO_CXX14_DIGIT_SEPARATORS # define BOOST_NO_CXX11_UNICODE_LITERALS #endif — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#53 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIPDMWQOK6F76HVJW4SPAI3S6R2HVANCNFSM4XCZNEIA> .

fmhess · 2021-02-12T03:35:53Z

If I add a #define __CUDACC__ before the boost include I can induce the problem to happen even when the file has a ".cpp" extension. Alternatively, if I do a #undef __CUDACC__ before the boost include I can make the problem go away even when using a ".cu" extension. So the problem is triggered by the boost headers changing behavior based on a check for the __CUDACC__ define.

fmhess · 2021-02-12T03:41:41Z

If I delete the check for (defined __NVCC__ && defined __CUDACC__) in boost/preprocessor/config/config.hpp the problem goes away.

eldiener · 2021-02-12T04:34:41Z

That test was actually saying that if the VC++ default preprocessor is being used but nvcc is compiling cuda, don't use any of the many workarounds needed to deal with VC++ default's preprocessor and variadic macros. So what you are saying is that when nvcc is compiling in cuda mode, the preprocessor needs to treat it as VC++'s default preprocessor as far as dealing with variadic macro support is concerned. I am good with that change and will make it.

eldiener · 2021-02-12T04:39:44Z

I have made the change on the preprocessor 'develop' branch.

ghost · 2021-02-12T14:41:34Z

Successfully accomplished a build of my entire project after applying the specified fix.

fmhess mentioned this issue Feb 5, 2021

Fails to compile on NVCC CUDA 11.0 boostorg/config#365

Closed

ghost mentioned this issue Feb 10, 2021

shared.hpp does not compile in nvidia CUDA boostorg/preprocessor#37

Closed

ghost closed this as completed Feb 12, 2021

ghost mentioned this issue Feb 16, 2021

CUDA NVCC Build Support boostorg/build#711

Open

rdinnager mentioned this issue Nov 9, 2021

Get device capability mlverse/torch#734

Merged

This issue was closed.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fails to compile on NVCC CUDA 11.0 #53

Fails to compile on NVCC CUDA 11.0 #53

ghost commented Feb 4, 2021 •

edited by ghost

Loading

fmhess commented Feb 5, 2021

ghost commented Feb 9, 2021

fmhess commented Feb 9, 2021 via email

ghost commented Feb 9, 2021

fmhess commented Feb 9, 2021 via email

eldiener commented Feb 10, 2021

ghost commented Feb 10, 2021

fmhess commented Feb 11, 2021

ghost commented Feb 11, 2021 •

edited by ghost

Loading

fmhess commented Feb 11, 2021 via email

ghost commented Feb 11, 2021

eldiener commented Feb 11, 2021

fmhess commented Feb 11, 2021 via email

eldiener commented Feb 11, 2021

fmhess commented Feb 11, 2021

eldiener commented Feb 11, 2021

fmhess commented Feb 12, 2021

ghost commented Feb 12, 2021 via email

ghost commented Feb 12, 2021 via email

fmhess commented Feb 12, 2021

fmhess commented Feb 12, 2021

eldiener commented Feb 12, 2021

eldiener commented Feb 12, 2021

ghost commented Feb 12, 2021

Fails to compile on NVCC CUDA 11.0 #53

Fails to compile on NVCC CUDA 11.0 #53

Comments

ghost commented Feb 4, 2021 • edited by ghost Loading

fmhess commented Feb 5, 2021

ghost commented Feb 9, 2021

fmhess commented Feb 9, 2021 via email

ghost commented Feb 9, 2021

fmhess commented Feb 9, 2021 via email

eldiener commented Feb 10, 2021

ghost commented Feb 10, 2021

fmhess commented Feb 11, 2021

ghost commented Feb 11, 2021 • edited by ghost Loading

fmhess commented Feb 11, 2021 via email

ghost commented Feb 11, 2021

eldiener commented Feb 11, 2021

fmhess commented Feb 11, 2021 via email

eldiener commented Feb 11, 2021

fmhess commented Feb 11, 2021

eldiener commented Feb 11, 2021

fmhess commented Feb 12, 2021

ghost commented Feb 12, 2021 via email

ghost commented Feb 12, 2021 via email

fmhess commented Feb 12, 2021

fmhess commented Feb 12, 2021

eldiener commented Feb 12, 2021

eldiener commented Feb 12, 2021

ghost commented Feb 12, 2021

ghost commented Feb 4, 2021 •

edited by ghost

Loading

ghost commented Feb 11, 2021 •

edited by ghost

Loading