Undefined references with -lcuda -lcudart on Windows

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi I’m compiling on Windows 10 and is just starting out.
I have been trying to just get something to compile for hours already and no success thus far.
The newest error is loads of undefined references with the flags -lcuda -lcudart
I am obviously missing some library but I don’t really know what.
Any help is appreciated, I’m dying here.

Code (filename=test1.cu)

__global__ void foo(){}

int main()
{
    foo<<<1, 1>>>();
}

Makefile

CC = g++
ALLFLAGS = -L E:\cuda\development\lib\x64 -lcuda -lcudart
CUDAFLAGS = -arch=sm_20

EXEC = program
CUDASOURCES = $(wildcard *.cu)
CUDAOBJECTS = $(CUDASOURCES:.cu=.o)

all: $(CUDAOBJECTS)
        $(CC) $(ALLFLAGS) $(CUDAOBJECTS) -o $(EXEC)

$(CUDAOBJECTS): %.o: %.cu
        nvcc -c $(CUDAFLAGS) $< -o $@

Terminal output

nvcc -c -arch=sm_20 test1.cu -o test1.o
g++ -L E:\cuda\development\lib\x64 -lcuda -lcudart  test1.o -o program
Warning: corrupt .drectve at end of def file
test1.o:(.text$mn+0xf6): undefined reference to `cudaConfigureCall'
test1.o:(.text$mn+0x158): undefined reference to `__cudaUnregisterFatBinary'
test1.o:(.text$mn+0x20f): undefined reference to `__cudaRegisterFunction'
test1.o:(.text$mn+0x22c): undefined reference to `__cudaRegisterFatBinary'
test1.o:(.text$mn+0x27f): undefined reference to `cudaLaunch'
collect2.exe: error: ld returned 1 exit status
makefile:19: recipe for target 'all' failed
mingw32-make: *** [all] Error 1
2

You shouldn’t use g++ on windows with CUDA

It’s recommended to use cl.exe

g++ is not an officially supported host compiler/toolchain for CUDA on windows. This is covered in the installation guide for CUDA on windows

3

Ok will try it out, thx

EDIT: excellent advice, but you forgot to say I might want to shoot myself in the head before I start using cl to spare me the pain

EDIT2: I am quitting by the sheer idiocy of how cl works. By following MSDN word by word, it successfully screwed me over and wasted a couple of hours of my life. I believe it is a criminal offense to not warn me that cl is more carcinogenic than a nuclear bomb going off in your face.

4

How so? I was playing with Windows stuff yesterday and I was able to get a dummy test file with Thrust working. Granted, I wasn’t trying to link against the device runtime or any such thing.

5

You didn’t say where or how the build failed for you. The following seems to work just fine on my machine:

nvcc -c -arch=sm_20 -o test1.obj test1.cu
cl test1.obj /Feprogram.exe /link /LIBPATH:"c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\lib\x64" cudart.lib cuda.lib