Compiling CUDA in matlab.

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi.

I have some problems compiling CUDA in matlab. I’m using a small scipt, which looks like this:

function compile(varargin)

!"%VS90COMNTOOLS%vsvars32.bat" & nvcc -c cudaKernel.cu

n=getenv('CUDA_LIB_PATH'); if n(1)=='"', n=n(2:end); end, if n(end)=='"', n=n(1:end-1);end

mex(['-L' n],'-lcudart','-lcufft','cudaKernel.obj',varargin{:});

When I run this I get the error:

>> compile cudaKernel.cpp

Setting environment for using Microsoft Visual Studio 2008 x86 tools. 

cudaKernel.cu 

cudaKernel.cu(180): warning: variable "dreal" was declared but never referenced 

cudaKernel.cu(180): warning: variable "dimagin" was declared but never referenced 

tmpxft_00000804_00000000-3_cudaKernel.cudafe1.gpu 

tmpxft_00000804_00000000-8_cudaKernel.cudafe2.gpu 

cudaKernel.cu(180): warning: variable "dreal" was declared but never referenced 

cudaKernel.cu(180): warning: variable "dimagin" was declared but never referenced 

tmpxft_00000804_00000000-3_cudaKernel.cudafe1.cpp 

tmpxft_00000804_00000000-13_cudaKernel.ii 

Microsoft (R) Incremental Linker Version 9.00.21022.08 

Copyright (C) Microsoft Corporation.  All rights reserved. 

cudaKernel.obj C:\DOCUME~1\MER\LOCALS~1\Temp\mex_dGc1FB\cudaKernel.obj  

MSVCRT.lib(ti_inst.obj) : error LNK2005: "private: __cdecl type_info::type_info(class type_info const &)" (??0type_info@@AEAA@AEBV0@@Z) already defined in LIBCMT.lib(typinfo.obj) 

MSVCRT.lib(ti_inst.obj) : error LNK2005: "private: class type_info & __cdecl type_info::operator=(class type_info const &)" (??4type_info@@AEAAAEAV0@AEBV0@@Z) already defined in LIBCMT.lib(typinfo.obj) 

   Creating library C:\DOCUME~1\MER\LOCALS~1\Temp\mex_dGc1FB\templib.x and object C:\DOCUME~1\MER\LOCALS~1\Temp\mex_dGc1FB\templib.exp 

LINK : warning LNK4098: defaultlib 'MSVCRT' conflicts with use of other libs; use /NODEFAULTLIB:library 

cudaKernel.mexw64 : fatal error LNK1169: one or more multiply defined symbols found 

C:\PROGRA~1\MATLAB\R2008A\BIN\MEX.PL: Error: Link of 'cudaKernel.mexw64' failed. 

??? Error using ==> mex at 207

Unable to complete successfully.

Error in ==> compile at 4

mex(['-L' n],'-lcudart','-lcufft','cudaKernel.obj',varargin{:});

In my cudaKernel.cpp I have a declaration of the function in cudaKernel.cu I want to call, with the extern “C” flag. I have no idea why this won’t work. Maybe it has something with the fact that I can’t compile in cmd. Using nvcc yealds:

c:\cuda\include\host_config.h(115) : fatal error C1083: Cannot open include file: ‘crtdefs.h’: No such file or directory

I’m in the dark here, and would appreciate any help.

Btw. The nvmex function won’t work either. When using nvmex I get:

>> nvmex -f nvmexopt_9.bat cudaKernel.cu

abdelali target arch: win64 

C:\PROGRA~1\MATLAB\R2008A\BIN\NVMEX.PL: Error: Could not find specified options file 

	'nvmexopt_9.bat'. 

??? Error using ==> nvmex at 206

Unable to complete successfully.

>> nvmex -f nvmexopts_9.bat cudaKernel.cu

abdelali target arch: win64 

cudaKernel.cu 

Z:/tekst/MATLAB/ISAR_Tests/cudaKernel.cu(180): warning: variable "dreal" was declared but never referenced 

Z:/tekst/MATLAB/ISAR_Tests/cudaKernel.cu(180): warning: variable "dimagin" was declared but never referenced 

tmpxft_000009fc_00000000-3_cudaKernel.cudafe1.gpu 

tmpxft_000009fc_00000000-8_cudaKernel.cudafe2.gpu 

Z:/tekst/MATLAB/ISAR_Tests/cudaKernel.cu(180): warning: variable "dreal" was declared but never referenced 

Z:/tekst/MATLAB/ISAR_Tests/cudaKernel.cu(180): warning: variable "dimagin" was declared but never referenced 

tmpxft_000009fc_00000000-3_cudaKernel.cudafe1.cpp 

Microsoft (R) Incremental Linker Version 9.00.21022.08 

Copyright (C) Microsoft Corporation.  All rights reserved. 

cudaKernel.obj  

LINK : fatal error LNK1181: cannot open input file 'kernel32.lib' 

C:\PROGRA~1\MATLAB\R2008A\BIN\NVMEX.PL: Error: Link of 'cudaKernel.mexw64' failed. 

??? Error using ==> nvmex at 206

Unable to complete successfully.

Actually I don’t care which of them you know how to fix, ass long as one of them will work, I’ll be happy.

2

I’m getting closer now. When I’m trying to compile now (with nvmex), I get the errors:

tmpxft_00000874_00000000-3_cudaKernel.cudafe1.cpp 

Microsoft (R) Incremental Linker Version 9.00.21022.08 

Copyright (C) Microsoft Corporation.  All rights reserved. 

cudaKernel.obj  

   Creating library C:\DOCUME~1\MER\LOCALS~1\Temp\mex_hVxClX\templib.x and object C:\DOCUME~1\MER\LOCALS~1\Temp\mex_hVxClX\templib.exp 

cudaKernel.obj : error LNK2019: unresolved external symbol __cudaUnregisterFatBinary referenced in function __cudaUnregisterBinaryUtil 

cudaKernel.obj : error LNK2019: unresolved external symbol cudaLaunch referenced in function "enum cudaError __cdecl cudaLaunch<char>(char *)" (??$cudaLaunch@D@@YA?AW4cudaError@@PEAD@Z) 

cudaKernel.obj : error LNK2019: unresolved external symbol cudaSetupArgument referenced in function __device_stub__Z10zeroMatrixP6float2ii 

cudaKernel.obj : error LNK2019: unresolved external symbol __cudaRegisterFunction referenced in function __sti____cudaRegisterAll_45_tmpxft_00000874_00000000_6_cudaK

ernel_cpp1_ii_c9c3f677 

cudaKernel.obj : error LNK2019: unresolved external symbol __cudaRegisterFatBinary referenced in function __sti____cudaRegisterAll_45_tmpxft_00000874_00000000_6_cudaK

ernel_cpp1_ii_c9c3f677 

cudaKernel.obj : error LNK2019: unresolved external symbol cudaMemcpy referenced in function "int __cdecl cudaLoop(float *,float *,int,int)" (?cudaLoop@@YAHPEAM0HH@Z) 

cudaKernel.obj : error LNK2019: unresolved external symbol cudaConfigureCall referenced in function "int __cdecl cudaLoop(float *,float *,int,int)" (?cudaLoop@@YAHPEAM0HH@Z) 

cudaKernel.obj : error LNK2019: unresolved external symbol cudaMalloc referenced in function "int __cdecl cudaLoop(float *,float *,int,int)" (?cudaLoop@@YAHPEAM0HH@Z) 

cudaKernel.mexw64 : fatal error LNK1120: 8 unresolved externals 

C:\PROGRA~1\MATLAB\R2008A\BIN\NVMEX.PL: Error: Link of 'cudaKernel.mexw64' failed. 

??? Error using ==> nvmex at 206

Unable to complete successfully.

I’m compiling with -lcudart and -lcufft, but also get the warnings:

Warning: cudart specified with -l option not found on -L path

Warning: cufft specified with -l option not found on -L path

Any help will be appreciated.

3

I think I’m getting really close now. I now get the error:

>> compile

abdelali target arch: win64 

cudaKernel.cu 

tmpxft_00000fdc_00000000-3_cudaKernel.cudafe1.gpu 

tmpxft_00000fdc_00000000-8_cudaKernel.cudafe2.gpu 

tmpxft_00000fdc_00000000-3_cudaKernel.cudafe1.cpp 

'mt' is not recognized as an internal or external command, 

operable program or batch file.

And when I try to run it, I get:

??? Invalid MEX-file 'z:\tekst\MATLAB\ISAR_Tests\cudaKernel.mexw64': The specified module could not be found.
4

How did you get over the following error:

c:\cuda\include\host_config.h(115) : fatal error C1083: Cannot open include file: ‘crtdefs.h’: No such file or directory

I am facing this while trying to compile on 64-bit Win7.

I will appreciate your reply/help in this regard.

Thanks
Shadab

5

I get the same error as Shadab, on Win 7 64-bit.

I’m using the command:

“C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v3.2\bin\nvcc.exe” -D_MATLAB_ -I"C:\Program Files\MATLAB\R2010b\extern\include" -I"C:\ProgramData\NVIDIA Corporation\NVIDIA GPU Computing SDK 3.2\C\common\inc" -arch sm_11 --ptxas-options=-v -O3 -Iprivate/depth_sensor/ --cuda “private/depth_sensor/pmls.cu” --output-file “C:\Users\OWOODF~1\AppData\Local\Temp\tpc44f628a_a571_4747_80c2_df38d2b8c177.cpp”

pmls.cu

The error is:

c:\program files\nvidia gpu computing toolkit\cuda\v3.2\include\host_config.h(96) : fatal error C1083: Cannot open include file: ‘crtdefs.h’: No such file or directory

Any help would be much appreciated.

6

I don’t remember how I solved my problems. But I have found another solution. Now I use CMAKE to compile, and it does everything for me. Just use CL/gcc, link to the matlab libraries and compile it as a shared library. Then rename the shared library from .so/.dll to what matlab needs.

7

Thanks for the suggestion. I might try it, but for now I solved my problem by using the 32-bit CUDA toolkit instead of the 64-bit one. I’m using MSVC 8.0 (2005).