CUDA and gcc compatibility issue

Question

I was getting this error

/usr/local/cuda-5.0/bin/../include/host_config.h:82:2: error: #error -- unsupported GNU version! gcc 4.7 and up are not supported! make: * [src/Throughput.o] Error 1

In the host_config.h they assure compatibility up to the 4.6

#if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ > 6)

#error -- unsupported GNU version! gcc 4.7 and up are not supported!

I have both 4.6 and 4.7

elect@elect-desktop:/usr/local/cuda-5.0/bin$ gcc gcc
gcc-4.7 gcc-nm-4.7 gcc-4.6 gcc-ar-4.7
gcc-ranlib-4.7

Looking on internet they suggest to add a link to the gcc-4.6 in the cuda bin directory.

So I did

elect@elect-desktop:/usr/local/cuda-5.0/bin$ sudo ln -s /usr/bin/gcc-4.6 gcc

And I get another error

**** Build of configuration Debug for project Throughput ****

make all 
Building file: ../src/Throughput.cu
Invoking: NVCC Compiler
nvcc -G -g -O0 -gencode arch=compute_20,code=sm_20 -odir "src" -M -o "src/Throughput.d"     "../src/Throughput.cu"
gcc: error trying to exec 'cc1plus': execvp: No such file or directory
make: *** [src/Throughput.o] Error 1

**** Build Finished ****

Googling again didn't bring me on some clear situations (gcc downgrading, etc)

So I am asking here what is now the problem since CUDA is supposed to be compatible with gcc-4.6 ...

My system:

Ubuntu 12.10 64b
cuda_5.0.35_linux_64_ubuntu11.10-1

This is the tutorial code I am trying to compile at the moment

/**
 * Copyright 1993-2012 NVIDIA Corporation.  All rights reserved.
 *
 * Please refer to the NVIDIA end user license agreement (EULA) associated
 * with this source code for terms and conditions that govern your use of
 * this software. Any use, reproduction, disclosure, or distribution of
 * this software and related documentation outside the terms of the EULA
 * is strictly prohibited.
 */
#include <stdio.h>
#include <stdlib.h>

static const int WORK_SIZE = 256;

/**
 * This macro checks return value of the CUDA runtime call and exits
 * the application if the call failed.
 */
#define CUDA_CHECK_RETURN(value) {                                          \
    cudaError_t _m_cudaStat = value;                                        \
    if (_m_cudaStat != cudaSuccess) {                                       \
        fprintf(stderr, "Error %s at line %d in file %s\n",                 \
                cudaGetErrorString(_m_cudaStat), __LINE__, __FILE__);       \
        exit(1);                                                            \
    } }

__device__ unsigned int bitreverse(unsigned int number) {
    number = ((0xf0f0f0f0 & number) >> 4) | ((0x0f0f0f0f & number) << 4);
    number = ((0xcccccccc & number) >> 2) | ((0x33333333 & number) << 2);
    number = ((0xaaaaaaaa & number) >> 1) | ((0x55555555 & number) << 1);
    return number;
}

/**
 * CUDA kernel function that reverses the order of bits in each element of the array.
 */
__global__ void bitreverse(void *data) {
    unsigned int *idata = (unsigned int*) data;
    idata[threadIdx.x] = bitreverse(idata[threadIdx.x]);
}

/**
 * Host function that prepares data array and passes it to the CUDA kernel.
 */
int main(void) {
    void *d = NULL;
    int i;
    unsigned int idata[WORK_SIZE], odata[WORK_SIZE];

    for (i = 0; i < WORK_SIZE; i++)
        idata[i] = (unsigned int) i;

    CUDA_CHECK_RETURN(cudaMalloc((void**) &d, sizeof(int) * WORK_SIZE));

    CUDA_CHECK_RETURN(cudaMemcpy(d, idata, sizeof(int) * WORK_SIZE, cudaMemcpyHostToDevice));

    bitreverse<<<1, WORK_SIZE, WORK_SIZE * sizeof(int)>>>(d);

    CUDA_CHECK_RETURN(cudaThreadSynchronize());
    // Wait for the GPU launched work to complete
    CUDA_CHECK_RETURN(cudaGetLastError());
    CUDA_CHECK_RETURN(cudaMemcpy(odata, d, sizeof(int) * WORK_SIZE, cudaMemcpyDeviceToHost));

    for (i = 0; i < WORK_SIZE; i++)
        printf("Input value: %u, device output: %u\n", idata[i], odata[i]);

    CUDA_CHECK_RETURN(cudaFree((void*) d));
    CUDA_CHECK_RETURN(cudaDeviceReset());

    return 0;
}

Apart from the CUDA, are you dealing with pure C code? Or is it perhaps C++? — Bart
@talonmies Ah, but isn't that the problem then? A problem with (or absence of) g++? — Bart

talonmies talonmies · Accepted Answer · 2013-02-09T11:48:48

The problem stems from the CUDA toolchain not being able to find a valid C++ compiler. nvcc is only a compiler driver, it requires a working C++ compiler to compile any code.

The most correct way to do this [note you are using an unsupported Linux version, so use this advice at your own risk], is to set up a local directory holding links to a supported compiler suite (this mean matching, supported verions of gcc and g++) and pass the --compiler-bindir argument to nvcc when you compile. For example:

$ ls -l $HOME/cuda/bin
total 16
lrwxr-xr-x  1 talonmies  koti  16 Feb  9 12:41 g++ -> /usr/bin/g++-4.2
lrwxr-xr-x  1 talonmies  koti  16 Feb  9 12:41 gcc -> /usr/bin/gcc-4.2

Here I have a set of links to a supported compiler. I then can compile like this:

$ nvcc --compiler-bindir=$HOME/cuda/bin -c -arch=sm_12 -Xptxas="-v" nanobench.cu 
ptxas info    : 0 bytes gmem
ptxas info    : Compiling entry function '_Z5benchIfLi128000EEvPjPT_i' for 'sm_12'
ptxas info    : Used 5 registers, 28 bytes smem, 12 bytes cmem[1]

This is probably the safest and least invasive way to use alternative compilers where the system compiler is not supported.

CUDA and gcc compatibility issue

2 Answers