Question

I have this code and it crashes every 1-2th launch.
I have tried use malloc/cudaMallocHost/cudeMalloc but it was useless. It think it happens due to manual cufftComplex initialization but prove it a can’t because without data I can’t get fft. Could you help me eliminate this crashes?

#include <stdio.h>
#include <string.h>
#include <iostream>
#include <fstream>
#include <conio.h>
#include <cuda.h>
#include <cufft.h>

using namespace std;

int main(int argc, char **argv)
{
cufftHandle plan;
cufftComplex *data;
cufftComplex *digits;
cufftComplex *h_data;

cudaMallocHost((void**)&digits, sizeof(cufftComplex)*8);
digits[0].x = 12.5f; digits[0].y = 0.0f;
digits[1].x = 66.23f; digits[1].y = 0.0f;
digits[2].x = 35.1f; digits[2].y = 0.0f;
digits[3].x = 16.7f; digits[3].y = 0.0f;
digits[4].x = 14.83f; digits[4].y = 0.0f;
digits[5].x = 55.1f; digits[5].y = 0.0f;
digits[6].x = 11.7f; digits[6].y = 0.0f;
digits[7].x = 18.83f; digits[7].y = 0.0f;

cudaMalloc((void**)&data, sizeof(cufftComplex)*8);

cudaMemcpy(data, digits, sizeof(cufftComplex)*8, cudaMemcpyHostToDevice);

if (cufftPlan1d(&plan, 8, CUFFT_C2C, 1) != CUFFT_SUCCESS) {
    fprintf(stderr, "Cuda: cufftPlan1d CUFFT_C2C failed\n");
    return 1;
}

if (cufftExecC2C(plan, data, data, CUFFT_FORWARD) != CUFFT_SUCCESS) {
    fprintf(stderr, "Cuda: cufftExecC2C CUFFT_FORWARD failed\n");
    return 1;
}

if (cudaMalloc((void**)&h_data, sizeof(cufftComplex)*8) != cudaSuccess) {
    fprintf(stderr, "Cuda: cudaMalloc((void**)&h_data failed\n");
    return 1;
}

cudaMemcpy(h_data, data, sizeof(cufftComplex)*8, cudaMemcpyDeviceToHost);


printf("\nOriginal:\n");
for(int i = 0; i < 8; ++i){
    printf("\nRe:%2.5f   Im:%2.5f", digits[i].x, digits[i].y);
} 

printf("\n\n1D-FFT:\n");
for(int i = 0; i < 8; ++i){
    printf("\nRe:%2.5f   Im:%2.5f", h_data[i].x, h_data[i].y);
} 

cudaFree(digits);
cudaFree(data);
cudaFree(h_data);
cufftDestroy(plan);
}

Answer 1

You basic problem is improper mixing of host and device memory pointers. You have assigned the address of a device memory allocation (using cudaMalloc) to h_data , but are trying to use it as a pointer to an address in host memory. That won’t work and is producing the host segmentation fault you are seeing. Your example should look something like:

#include <cstdlib>
#include <cuda_runtime.h>
#include <cufft.h>

int main(int argc, char **argv)
{
    cufftHandle plan;
    cufftComplex *data, *digits, *h_data;

    digits = (cufftComplex *)malloc(sizeof(cufftComplex)*8);
    digits[0].x = 12.5f; digits[0].y = 0.0f;
    digits[1].x = 66.23f; digits[1].y = 0.0f;
    digits[2].x = 35.1f; digits[2].y = 0.0f;
    digits[3].x = 16.7f; digits[3].y = 0.0f;
    digits[4].x = 14.83f; digits[4].y = 0.0f;
    digits[5].x = 55.1f; digits[5].y = 0.0f;
    digits[6].x = 11.7f; digits[6].y = 0.0f;
    digits[7].x = 18.83f; digits[7].y = 0.0f;

    cudaMalloc((void**)&data, sizeof(cufftComplex)*8);
    cudaMemcpy(data, digits, sizeof(cufftComplex)*8, cudaMemcpyHostToDevice);

    if (cufftPlan1d(&plan, 8, CUFFT_C2C, 1) != CUFFT_SUCCESS) {
        fprintf(stderr, "Cuda: cufftPlan1d CUFFT_C2C failed\n");
        return 1;
    }

    if (cufftExecC2C(plan, data, data, CUFFT_FORWARD) != CUFFT_SUCCESS) {
        fprintf(stderr, "Cuda: cufftExecC2C CUFFT_FORWARD failed\n");
        return 1;
    }

    h_data = (cufftComplex *)malloc(sizeof(cufftComplex)*8);
    cudaMemcpy(h_data, data, sizeof(cufftComplex)*8, cudaMemcpyDeviceToHost);

    printf("\nOriginal:\n");
    for(int i = 0; i < 8; ++i){
        printf("\nRe:%2.5f   Im:%2.5f", digits[i].x, digits[i].y);
    } 

    printf("\n\n1D-FFT:\n");
    for(int i = 0; i < 8; ++i){
        printf("\nRe:%2.5f   Im:%2.5f", h_data[i].x, h_data[i].y);
    } 

    free(digits);
    free(h_data);
    cudaFree(data);
    cufftDestroy(plan);
}

Note that you should use plain malloc or the C++ new operator to allocate host side memory rather than cudaMallocHost, unless you understand very well what the latter API does and why you are using it.