Allocating array pointers segfaults What am I doing wrong?

Accelerated Computing CUDA CUDA Programming and Performance
1

I have the below code.

I tried in “normal” C first, and then I’m trying to transfer it to CUDA.

When I is 1, the program segfaults. I don’t understand why.

Perhaps I could gain more information from cuda-gdb, but I’m not sure how.

[codebox]int main() {

int **ptrArray;

//ptrArray = (int**)malloc(3 * sizeof(int*));

cudaError_t error = cudaMalloc( (void**) ptrArray, 3 * sizeof(int*));

for(int i = 0; i < 3; i++) {

  //ptrArray[i] = (int*)malloc(10 * sizeof(int));

  error = cudaMalloc( (void**) ptrArray[i], 10 * sizeof(int));

}

TehKernel<<<1, 1>>>(ptrArray);

}[/codebox]

Cuda-dgb:

[codebox][Thread debugging using libthread_db enabled]

[New process 30521]

[New Thread -1211263280 (LWP 30521)]

[New Thread -1215595632 (LWP 30524)]

[New Thread -1215861872 (LWP 30525)]

Program received signal SIGSEGV, Segmentation fault.

[Switching to Thread -1211263280 (LWP 30521)]

0xb7f76773 in cudaMalloc () from /opt/cuda/lib/libcudart.so.2[/codebox]

2

Hi!

ptrArray points to device code. This means that the value of ptrArray must be interpreted as an adress in device mem. But

error = cudaMalloc( (void**) ptrArray[i], 10 * sizeof(int));

is executed in host code. So dereferencing in ptrArray[i] is done on the CPU. And the value of the memory in host at adress (ptrArray + i) is something undefined.

3

[quote name=‘navier-stokes’ post=‘531319’ date=‘Apr 17 2009, 01:44 PM’]

Hi!

ptrArray points to device code. This means that the value of ptrArray must be interpreted as an adress in device mem. But

[codebox] int **ptrArray;

cudaError_t error = cudaMalloc( (void**) ptrArray, 3 * sizeof(int*));

printf(“%d\n”, error);

int *ptrA, *ptrB, *ptrC;

error = cudaMalloc( (void**) ptrA, 10 * sizeof(int));

printf(“%d\n”, error);

error = cudaMalloc( (void**) ptrB, 10 * sizeof(int));

printf(“%d\n”, error);

error = cudaMalloc( (void**) ptrC, 10 * sizeof(int));

printf(“%d\n”, error);

int *ptrS[3];

ptrS[0] = ptrA;

ptrS[1] = ptrB;

ptrS[2] = ptrC;

cudaMemcpy(ptrArray, ptrS, 3 * sizeof(int*), cudaMemcpyHostToDevice);[/codebox]

Thinking then, that if I allocate 4 arrays, and the copy the adresses to three of them to the last, I could achieve what I wanted (Which is sending over several arrays without getting a huge parameter list)

However, I still seg fault on the second Malloc.

Something else I can’t quite figure out is, why gdb claims there is no variable ptrA in current context.

4

Nevermind, my mallocs where wrong >_<

[codebox]global static void TehKernel(int **DArr) {

for(int i = 0; i < 3; i++)

  printf("%p\n", DArr[i]);

}

template

T* SimpleCudaMalloc(const size_t size) {

T* devicePtr = NULL;

cudaError_t error = cudaMalloc((void**)&devicePtr, size * sizeof(T));

if (error == cudaErrorMemoryAllocation) {

  //cout << "CudaMalloc error: " << error << " (" << cudaGetErrorString(error) << ")" << ". Exiting" << endl;

  assert(false);

}

return devicePtr;

}

int main() {

const int popSize = 1024;

int *T, *V, *A;

T = SimpleCudaMalloc(popSize);

V = SimpleCudaMalloc(popSize);

A = SimpleCudaMalloc(popSize);

int **DArr, **HArr;

DArr = SimpleCudaMalloc<int*>(3);

HArr = (int**)malloc(3 * sizeof(int*));

HArr[0] = T;

HArr[1] = V;

HArr[2] = A;

for(int i = 0; i < 3; i++)

  printf("%p\n", HArr[i]);

cudaMemcpy(HArr, DArr, 3 * sizeof(int*), cudaMemcpyHostToDevice);

TehKernel<<<1, 1>>>(DArr);

printf(“\n”);

}[/codebox]

The output when printing from the Kernel is just (nil).

5

And I got that figured out too ^^

Again my copying was wrong. Gonna stick to my templates next time ;)

6

Must be:

int **ptrArray;

   cudaError_t error = cudaMalloc( (void**)&ptrArray, 3 * sizeof(int*));

   printf("%d\n", error);
7

Hi Letharion,

Would you mind sharing your working code? I am also trying to make 2D arrays using memcpy2D but my code is not working. It’s very frustrating.

Dinh

EDIT: Got it to work…Why didn’t you use memcpy2D like this???

cudaMemcpy2D(DArr, popSize*sizeof(int), HArr, popSize*sizeof(int), popSize*sizeof(int), 3, cudaMemcpyHostToDevice);