Undefine references of functions from CUDA - reference

Question

I am looking for a way to use cufft.h a CUDA toolkit which perform GPU parallelization of fast fourier transform.
First of all, I downloaded cuda library and cufft through synaptic.
Then I used the sample program from the cufft documentation from NVidia.
my cuda library is located at /usr/local/cuda-9.0 on my laptop.
I added those include :
1 #include <iostream>
2 #include <cstdio>
3 #include "/usr/local/cuda-9.0/include/cuda.h"
4 #include "/usr/local/cuda-9.0/include/cuda_runtime_api.h"
5 #include "/usr/local/cuda-9.0/include/cufft.h"
I compile like this :
g++ -Wall main.cpp -o main
and get undefine references error for each cuda-like functions (cudaMalloc,cudaGetLastError, etc...)
I am pretty young about library implementation and I don't understand what should I do to include properly this cuda-cufft library...
The nvidia documentation talk about filename.cu but I don't know what this is about...
Thank you for your time :)
n.b : I added cuda.h and cuda_runtime_api.h after reading a forum (I forgot which it was). Apparently, only cuda_runtime_api.h is necessary (I tried without cuda.h and get the same errors).

Answer 1

Here is a complete sample code (that doesn't do anything useful) and a sample g++ compile command that will properly compile and link the code:
$ cat t1338.cpp
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <cuda_runtime.h>
#include <cufft.h>
int main() {
size_t work_size;
int fft_sz = 32; // Size of each FFT
int num_ffts = 1; // How many FFTs to do
cufftComplex *in_buf_h, *in_buf_d, *out_buf_d;
// Allocate buffers on host and device
in_buf_h = new cufftComplex[fft_sz*num_ffts];
cudaMalloc(&in_buf_d, fft_sz*num_ffts*sizeof(cufftComplex));
cudaMalloc(&out_buf_d, fft_sz*num_ffts*sizeof(cufftComplex));
cudaMemset(out_buf_d, 0, fft_sz*num_ffts*sizeof(cufftComplex));
// Fill input buffer with zeros and copy to device
memset(in_buf_h, 0, fft_sz*num_ffts*sizeof(cufftComplex));
cudaMemcpy(in_buf_d, in_buf_h, fft_sz*num_ffts*sizeof(cufftComplex), cudaMemcpyHostToDevice);
// Plan num_ffts of size fft_sz
cufftHandle plan;
cufftCreate(&plan);
cufftMakePlan1d(plan, fft_sz, CUFFT_C2C, num_ffts, &work_size);
// Execute the plan. We don't actually care about values.
cufftExecC2C(plan, in_buf_d, out_buf_d, CUFFT_FORWARD);
// Sync the device to flush the output
cudaDeviceSynchronize();
return 0;
}
$ g++ t1338.cpp -I/usr/local/cuda/include -L/usr/local/cuda/lib64 -lcudart -lcufft
$
Your include statements are probably OK as-is, but I have used a format that says "search on the standard path for this file" and then I identify an addition to the standard path with
-I/usr/local/cuda/include
However your compile command is definitely missing the necessary link apparatus. You need to specify where to find the libraries (the path) with -L and then indicate the specific libraries to include, which are both the CUDA runtime library (-lcudart) and also the CUFFT library (-lcufft):
-L/usr/local/cuda/lib64 -lcudart -lcufft
The CUDA toolkit normally gets installed with sample codes which will have sample Makefiles you can inspect, or just compile those projects to see typical compilation command usage.
As I mentioned, this source code is incomplete. It doesn't do anything useful. It is just to demonstrate proper compilation behavior. In particular, I've omitted proper error checking, which I recommend you include in your actual codes.
Depending on whether your install created a symbolic link or not, you may need to change the above paths to:
-I/usr/local/cuda-9.0/include
and
-L/usr/local/cuda-9.0/lib64 -lcudart -lcufft