Skip to content

Link error with R.dll

November 21, 2015

I’ve taken a crack at compiling the R package gputools on Windows, but I hit a bit of a roadblock. I’ll describe the steps I’ve taken below, hopefully someone will be able to suggest a way forward.

Some background: gputools provides an interface between R and CUDA, the nVidia API for executing code on a graphics card (GPU). On Linux or Mac OS, installation is fairly straightforward because the CUDA compiler nvcc depends on gcc or clang, respectively. However, interfacing between R and CUDA on Windows is an unsolved problem, since the respective toolchains are fundamentally incompatible. The only supported toolchain for compiling R on Windows is MinGW with gcc and gfortran. On the other hand, the nVidia CUDA Toolkit requires certain specific versions of Microsoft Visual Studio.

I installed Visual Studio Community 2013 and nVidia Toolkit 7.5.18 for Windows 10. You can fool the gputools configure script by running it inside a Cygwin64 terminal. This enables you to generate a Makefile for gcc + nvcc, but of course this won’t actually work as I’ve explained above. There are some Makefile entries that you will need to edit manually, as follows:

NVCC = nvcc
CUDA_LIB = $(CUDA_HOME)/lib/x64
R_LIB = -L"$(R_HOME)/bin/x64" -lR
CFLAGS = -O2 -Wall
TARGETS = gputools.dll

Before you run make, add cl.exe to your PATH so that nvcc can find it:

set PATH=%PATH%;C:\Program Files (x86)\Microsoft Visual Studio 11.0\VC\bin

You will also need to create a file vcvars64.bat in the directory VC\bin\amd64, containing only a single line:

CALL setenv /x64

Unfortunately, there is some unportable code in qrdecomp.cu and cuseful.cu that won’t compile in cl:

qrdecomp.cu(802): error: expression must have a constant value
qrdecomp.cu(803): error: expression must have a constant value
2 errors detected in the compilation of "qrdecomp.cpp1.ii".
make: *** [qrdecomp.o] Error 2

../VC/INCLUDE\xlocale(1610): error: identifier "_First1" is undefined
../VC/INCLUDE\xlocale(1610): error: identifier "_Last1" is undefined
../VC/INCLUDE\xlocale(1610): error: identifier "_Count" is undefined
../VC/INCLUDE\xlocale(1845): error: identifier "_First1" is undefined
../VC/INCLUDE\xlocale(1845): error: identifier "_Last1" is undefined
../VC/INCLUDE\xlocale(1845): error: identifier "_Count" is undefined
6 errors detected in the compilation of "cuseful.cpp1.ii".
make: *** [cuseful.o] Error 2

For qrdecomp.cu, you need to allocate the arrays using Calloc (and deallocate using Free). For cuseful.cu, the problem is with STL std::string. I chose to work around this problem using good old char* instead. Most of the calls to that private method use .c_str() anyway. If cl had an equivalent compile flag to -std=gnu99 then these changes wouldn’t be necessary.

This is where I hit the final problem, which currently has me stumped:

cuseful.o : error LNK2019: unresolved external symbol Rprintf referenced in function "void __cdecl printMat(int,int,float const *,char const *)" (?printMat@@YAXHHPEBMPEBD@Z)
gputools.dll : fatal error LNK1120: 15 unresolved externals
make: *** [gputools.dll] Error 2

The version of R that I have installed provides R.dll, but not R.lib. I tried using pexports to create a .def file, but I still end up with unresolved externals as above. It is possible that the only way around this is to compile my own copy of R.dll from source.

 

 

Advertisements

From → R

5 Comments
  1. Daniel permalink

    Great post, I’ve just found it today after trying to do exactly what you did and looking for solutions to the qrdecomp/cuseful issues. Did you ever find a way to get gputools to compile on Windows given the lack of an R.lib?

  2. Sadly, no. I planned to revisit this once CUDA 8.0 was released, since it supports Visual Studio 2015. However, even with Microsoft R Open (MRO) and R Tools for Visual Studio (RTVS), it doesn’t sound like there will ever be a compatible toolchain for CUDA in R on Windows:

    http://stackoverflow.com/questions/37629098/compiler-for-source-packages-in-microsoft-r-open

  3. The documentation for Rth mentions Windows, but it really isn’t clear whether they’ve ever gotten it working on that platform:

    http://heather.cs.ucdavis.edu/~matloff/rth.html

    Another option would be OpenCL, since you might be able to compile that with the Intel INDE for Windows:

    https://software.intel.com/en-us/intel-inde

  4. Daniel permalink

    However, I am not a C/C++ developer, so I made some pretty severe guesses on how to fix qrdecomp.cu and cuseful.cu.

    For qrdecomp.cu, I did this:
    // added after line 803
    (804) float * scales;
    (805) double * Beta;
    (806) scales = Calloc(blockSize, float);
    (807) Beta = Calloc(blockSize, double);
    // added after line 912
    Free(scales);
    Free(Beta);

    For cuseful.cu, I did this:
    (line 104) changed return type of function to be char *
    (line 133) deleted .c_str()
    (line 140) deleted .c_str()
    I also deleted all #include lines at top

    Would you be able to let me know if these are correct and what you used in your code?

Trackbacks & Pingbacks

  1. Fixing compile issues on SPARC Solaris | Matt Moores

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

Let's Look at the Figures

David Firth's blog

Nicholas Tierney

Computational Bayesian statistics

One weiRd tip

Computational Bayesian statistics

Series B'log

discussion blog for JRSS Series B papers

Mad (Data) Scientist

Musings, useful code etc. on R and data science

R-bloggers

R news and tutorials contributed by (750) R bloggers

Another Astrostatistics Blog

The random musings of a reformed astronomer ...

Darren Wilkinson's research blog

Statistics, computing, data science, Bayes, stochastic modelling, systems biology and bioinformatics

CHANCE

Computational Bayesian statistics

StatsLife - Significance magazine

Computational Bayesian statistics

(badness 10000)

Computational Bayesian statistics

Igor's Blog

Computational Bayesian statistics

Statisfaction

I can't get no

Xi'an's Og

an attempt at bloggin, nothing more...

Sam Clifford

Postdoctoral Fellow, Bayesian Statistics, Aerosol Science

Bayesian Research & Applications Group

Frontier Research in Bayesian Methodology & Computation

%d bloggers like this: