Segmentation Fault

tkeith7 · June 13, 2011, 7:01pm

Hey guys, I’m new to cuda programming; below is my code for a game of life program. When I run it I get a Segmentation fault error, the debugger says the error is with the hBlockAll[i][j] = 0; line, but I’m not sure what this means or why I am getting this error.

#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>

const int matrix = 10;
void gameOflifeOnDevice();
global void gameOfLifeOnGPU(int** dBlockAll,int** hBlockAll, int matrix);

int main(){

gameOflifeOnDevice();

getchar();

}

void gameOflifeOnDevice(){

int size = matrix * matrix * sizeof(float);

int **hBlockAll,**dBlockAll;



hBlockAll = (int**)malloc(size);
dBlockAll = (int**)malloc(size);
cudaMalloc(&hBlockAll,size);
cudaMalloc(&dBlockAll,size);


for(int i = 1 ; i < matrix ; i ++){
        for(int j = 1 ; j < matrix ; j++){
            hBlockAll[i][j] = 0;
        }
}
cudaMemcpy(&dBlockAll, hBlockAll , size ,cudaMemcpyHostToDevice);



gameOfLifeOnGPU<<<1,100>>>(dBlockAll,hBlockAll,matrix);

cudaMemcpy(&hBlockAll, dBlockAll , size ,cudaMemcpyDeviceToHost);

for (int i = 1 ; i < matrix ; i++){
    for (int j = 1 ; j < matrix ; j++){
        printf("|%c",(hBlockAll[i][j] == 1)? 'x' : '-');
    }
    printf("|\n");
}

free(hBlockAll);
free(dBlockAll);
cudaFree(hBlockAll);
cudaFree(dBlockAll);

}

global void gameOfLifeOnGPU(int** dBlockAll,int** hBlockAll, int matrix){

int idx = (blockIdx.x * blockDim.x) + threadIdx.x;
int idy = (blockIdx.y * blockDim.y) + threadIdx.y;
int countx;
for(int x = 0 ; x <= idx ; x++)
{
    for(int y = 0 ; y <= idy ; y++)
    {
        if (hBlockAll[x][y] == 1)
         {          
              // check block around currsor
              countx += hBlockAll[x][y+1];
              countx += hBlockAll[x][y-1];
              countx += hBlockAll[x+1][y];
              countx += hBlockAll[x+1][y+1];
              countx += hBlockAll[x+1][y-1];
              countx += hBlockAll[x-1][y];
              countx += hBlockAll[x-1][y-1];
              countx += hBlockAll[x-1][y-1];
              
              if (countx < 2) dBlockAll[x][y] = 0;
             
              if (countx > 3) dBlockAll[x][y] = 0;

              if (countx == 2 || countx == 3) dBlockAll[x][y] = 1;
         }
         else{
              // check block around currsor
              countx += hBlockAll[x][y+1];
              countx += hBlockAll[x][y-1];
              countx += hBlockAll[x+1][y];
              countx += hBlockAll[x+1][y+1];
              countx += hBlockAll[x+1][y-1];
              countx += hBlockAll[x-1][y];
              countx += hBlockAll[x-1][y-1];
              countx += hBlockAll[x-1][y-1];
              if (countx > 1 && countx < 4)
              {
                   dBlockAll[x][y] = 1;    
              }   
              
         } // end if
    }
}

}//gameOfLifeOnGPU

tkeith7 · June 13, 2011, 7:01pm

Hey guys, I’m new to cuda programming; below is my code for a game of life program. When I run it I get a Segmentation fault error, the debugger says the error is with the hBlockAll[i][j] = 0; line, but I’m not sure what this means or why I am getting this error.

#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>

const int matrix = 10;
void gameOflifeOnDevice();
global void gameOfLifeOnGPU(int** dBlockAll,int** hBlockAll, int matrix);

int main(){

gameOflifeOnDevice();

getchar();

}

void gameOflifeOnDevice(){

int size = matrix * matrix * sizeof(float);

int **hBlockAll,**dBlockAll;



hBlockAll = (int**)malloc(size);
dBlockAll = (int**)malloc(size);
cudaMalloc(&hBlockAll,size);
cudaMalloc(&dBlockAll,size);


for(int i = 1 ; i < matrix ; i ++){
        for(int j = 1 ; j < matrix ; j++){
            hBlockAll[i][j] = 0;
        }
}
cudaMemcpy(&dBlockAll, hBlockAll , size ,cudaMemcpyHostToDevice);



gameOfLifeOnGPU<<<1,100>>>(dBlockAll,hBlockAll,matrix);

cudaMemcpy(&hBlockAll, dBlockAll , size ,cudaMemcpyDeviceToHost);

for (int i = 1 ; i < matrix ; i++){
    for (int j = 1 ; j < matrix ; j++){
        printf("|%c",(hBlockAll[i][j] == 1)? 'x' : '-');
    }
    printf("|\n");
}

free(hBlockAll);
free(dBlockAll);
cudaFree(hBlockAll);
cudaFree(dBlockAll);

}

global void gameOfLifeOnGPU(int** dBlockAll,int** hBlockAll, int matrix){

int idx = (blockIdx.x * blockDim.x) + threadIdx.x;
int idy = (blockIdx.y * blockDim.y) + threadIdx.y;
int countx;
for(int x = 0 ; x <= idx ; x++)
{
    for(int y = 0 ; y <= idy ; y++)
    {
        if (hBlockAll[x][y] == 1)
         {          
              // check block around currsor
              countx += hBlockAll[x][y+1];
              countx += hBlockAll[x][y-1];
              countx += hBlockAll[x+1][y];
              countx += hBlockAll[x+1][y+1];
              countx += hBlockAll[x+1][y-1];
              countx += hBlockAll[x-1][y];
              countx += hBlockAll[x-1][y-1];
              countx += hBlockAll[x-1][y-1];
              
              if (countx < 2) dBlockAll[x][y] = 0;
             
              if (countx > 3) dBlockAll[x][y] = 0;

              if (countx == 2 || countx == 3) dBlockAll[x][y] = 1;
         }
         else{
              // check block around currsor
              countx += hBlockAll[x][y+1];
              countx += hBlockAll[x][y-1];
              countx += hBlockAll[x+1][y];
              countx += hBlockAll[x+1][y+1];
              countx += hBlockAll[x+1][y-1];
              countx += hBlockAll[x-1][y];
              countx += hBlockAll[x-1][y-1];
              countx += hBlockAll[x-1][y-1];
              if (countx > 1 && countx < 4)
              {
                   dBlockAll[x][y] = 1;    
              }   
              
         } // end if
    }
}

}//gameOfLifeOnGPU

Skybuck · June 14, 2011, 3:13am

My guess is your indexes are out of bounds:

hBlockAll[x-1][y-1];

when x=0 or y=0 the code above goes out of bounds.

You’ll need to fix that first somehow.

Different solutions thinkable… but I’ll leave it up to you to come up with a solution External Image

Skybuck · June 14, 2011, 3:13am

My guess is your indexes are out of bounds:

hBlockAll[x-1][y-1];

when x=0 or y=0 the code above goes out of bounds.

You’ll need to fix that first somehow.

Different solutions thinkable… but I’ll leave it up to you to come up with a solution External Image

tkeith7 · June 14, 2011, 5:34pm

Yes, I figured that would be an issue and I was going to tackle that later on, but the error occurs before those lines are even reached. I think the problem might be that I’m allocating memory for a two-dimensional array wrong. I am use to programming in java and not having to worry about that.

tkeith7 · June 14, 2011, 5:34pm

Yes, I figured that would be an issue and I was going to tackle that later on, but the error occurs before those lines are even reached. I think the problem might be that I’m allocating memory for a two-dimensional array wrong. I am use to programming in java and not having to worry about that.

Skybuck · June 14, 2011, 9:39pm

The problem is probably with this code:

for(int i = 1 ; i < matrix ; i ++){
for(int j = 1 ; j < matrix ; j++){
hBlockAll[i][j] = 0;
}
}

^

As far as I know C/C++ does not provide “multi dimensional index operator” like you seem to think.

Therefore this code is probably totally wrong… and c interprets it as an array to pointers which point to an array of pointers.

But that’s not what your malloc does… your malloc is a 1d array of pointers.

So to solve it you need to do:

hBlockAll[i * Width + j] = 0;

^ something like that.

So in your case something like:
hBlockAll(i * matrix + j] = 0;

Since matrix appears to be your width and height.

But the i and j should start at zero… so to me it seems you a noobie programmer and noobie c programmer External Image :)

Good luck ! External Image =D

I have seen plenty of weird c code by now :)

So what nvidia can learn from this is: “noobies and beginners and average programmers” want to program cuda too…

But C/C++ is probably way to difficult for them.

So NVIDIA would be wise to add other languages like free basic/basic and/or pascal or perhaps even java or anything that’s easier to program External Image

Skybuck · June 14, 2011, 9:39pm

The problem is probably with this code:

for(int i = 1 ; i < matrix ; i ++){
for(int j = 1 ; j < matrix ; j++){
hBlockAll[i][j] = 0;
}
}

^

As far as I know C/C++ does not provide “multi dimensional index operator” like you seem to think.

Therefore this code is probably totally wrong… and c interprets it as an array to pointers which point to an array of pointers.

But that’s not what your malloc does… your malloc is a 1d array of pointers.

So to solve it you need to do:

hBlockAll[i * Width + j] = 0;

^ something like that.

So in your case something like:
hBlockAll(i * matrix + j] = 0;

Since matrix appears to be your width and height.

But the i and j should start at zero… so to me it seems you a noobie programmer and noobie c programmer External Image :)

Good luck ! External Image =D

I have seen plenty of weird c code by now :)

So what nvidia can learn from this is: “noobies and beginners and average programmers” want to program cuda too…

But C/C++ is probably way to difficult for them.

So NVIDIA would be wise to add other languages like free basic/basic and/or pascal or perhaps even java or anything that’s easier to program External Image

tera · June 14, 2011, 11:16pm

C has multi-dimensional arrays. However, the commonly used “a one-dimensional array and a pointer can be used interchangeably” trick doesn’t apply there, so the double pointer [font=“Courier New”]**a[/font] cannot be used in place of a two-dimensional array.

Declare a 2-dimensional array with

int a;

and a pointer to a 2-dimensional array with

int (*p);

Note that SIZE_X and SIZE_Y must be known at compile time. Variable size arrays were only introduced with C99 and AFAIK are not available in CUDA. If you want to set the array size at runtime, you need to flatten the array to a 1-dimensional array as Skybuck showed.

tera · June 14, 2011, 11:16pm

C has multi-dimensional arrays. However, the commonly used “a one-dimensional array and a pointer can be used interchangeably” trick doesn’t apply there, so the double pointer [font=“Courier New”]**a[/font] cannot be used in place of a two-dimensional array.

Declare a 2-dimensional array with

int a;

and a pointer to a 2-dimensional array with

int (*p);

Note that SIZE_X and SIZE_Y must be known at compile time. Variable size arrays were only introduced with C99 and AFAIK are not available in CUDA. If you want to set the array size at runtime, you need to flatten the array to a 1-dimensional array as Skybuck showed.

Topic		Replies	Views
CUDA C++ Segmentation Fault CUDA Programming and Performance	1	901	September 30, 2017
Segmentation error in cudaMalloc - concurrent kernel execution CUDA Programming and Performance	8	1181	January 30, 2012
compile segmentation fault CUDA Programming and Performance	14	14541	September 2, 2010
CUDA C++ Segmentation Fault CUDA Programming and Performance	7	14637	October 1, 2017
Segmentation fault when calling virtual function on host CUDA Programming and Performance	9	2456	September 10, 2019
Segmentation fault (core dumped) CUDA Programming and Performance	4	13062	May 13, 2017
multi dimension array CUDA Programming and Performance	26	32767	February 12, 2010
segmentation fault at the first cudaMalloc with --device-emulation everything was fine CUDA Programming and Performance	10	4321	January 25, 2010
Segmentation fault (core dumped) CUDA Programming and Performance	5	7563	September 14, 2021
Newbie:Trying Matrix Vector Multiplication CUDA Programming and Performance	3	4212	November 10, 2008

Segmentation Fault

Related topics