passing cuda device pointer to host function

Question

I have a program I'm working on. I'm new with CUDA and C,so it really has been a bumpy ride for me. I'm trying to copy a struct into the device. And then I'm trying to get the struct back to host by copying it to the device. Below is the code :

typedef struct {
    int row;
    int col;
    float *arr;
    int numElements;
} Matrix;

Matrix *RMatrix = //definition here

Matrix *d_RMatrix;

    copyMatrix(d_RMatrix, RMatrix, hostToDevice);

    Matrix *check = createMatrix(0, 0, NULL, 0);

    copyMatrix(check, d_RMatrix, deviceToHost);

and the definition of copyMatrix :

void copyMatrix (Matrix *copyTo, Matrix *copyFrom, Copy_type type) 
{

    if(type == hostToDevice) {

        // create temporary host matrix and array
        Matrix *copyFrom_h = createMatrix(copyFrom->row, copyFrom->col, NULL, copyFrom->numElements);

        // allocate device memory, pointing to array in host. Copy array to device memory
        cudaMalloc((void**) &copyFrom_h->arr, sizeof(float) * copyFrom_h->numElements);
        cudaMemcpy(copyFrom_h->arr, copyFrom->arr, sizeof(float) * copyFrom_h->numElements, cudaMemcpyHostToDevice);

        // copy the temporary memory to device
        cudaMalloc((void**) &copyTo, sizeof(Matrix));
        cudaMemcpy(copyTo, copyFrom_h, sizeof(Matrix), cudaMemcpyHostToDevice);

        copyFrom_h = NULL;
        free(copyFrom_h);

    }

    else if(type == deviceToHost) {

        cudaMemcpy(copyTo, copyFrom, sizeof(Matrix), cudaMemcpyDeviceToHost);

        // allocate space for array in the copy to matrix
        copyTo->arr = makeArray(copyTo->col, copyTo->row);
        cudaMemcpy(copyTo->arr, copyFrom->arr, sizeof(float) * copyTo->numElements, cudaMemcpyDeviceToHost);

    }
}

The error says invalid memory access at 0x3 (value of d_RMatrix) for the 1st call to cudaMemcpy and results in segfault on the 2nd.

Is there anything I'm missing here? Thanks for your help :)

Agent_L · Accepted Answer · 2013-03-21 19:58:28Z

1

In C a pointer is an entity pointing to an object (in this case). Creating a pointer DOES NOT create the object nor allocate space for it.

You have created a pointer Matrix *d_RMatrix; but it doesn't point to any valid object. You got lucky it crashed, because by accident it could manage to actually copy the data into some random place in the memory.

Matrix TheMatrix();
Matrix *PointerToTheMatrix = &TheMatrix;

Or

Matrix *PointerToTheMatrix = createMatrix(...);//remember you will have to delete it eventually!

Function parameters are one way. If you assign something to copyTo inside function, the change will not be visible outside the function.

/edit: I have an idea:

Matrix* CreateMatrixInDevice(Matrix* copyFrom)
{
    Matrix* copyTo = NULL;
    cudaMalloc((void**) &copyTo, sizeof(Matrix));//create outer struct
    cudaMemcpy(copyTo, copyFrom, sizeof(Matrix), cudaMemcpyHostToDevice);//copy data from outer struct
    //the arr element in the device is now INVALID (pointing to host)

    cudaMalloc((void**) &copyTo->arr, sizeof(float) * copyFrom->numElements);//create inner array
    cudaMemcpy(copyTo->arr, copyFrom->arr, sizeof(float) * copyFrom->numElements, cudaMemcpyHostToDevice);//copy matrix data

    return copyTo;
}

edited Mar 21, 2013 at 19:58

answered Mar 21, 2013 at 19:21

Agent_L

5,53738 silver badges37 bronze badges

Sign up to request clarification or add additional context in comments.

5 Comments

Anugerah Erlaut Over a year ago

I allocated space for one using the cudaMalloc function inside the copyMatrix in the device global memory. Does that count as making an object? I'm aware there's this restriction of not using the values of a variable defined in device memory, but I'm not sure that this is a case of that...

Agent_L Over a year ago

@AnugerahErlaut I think you also have a problem with function parameters, I've updated the answer. I think you should try to make a working copy inside one function.

Anugerah Erlaut Over a year ago

Yups, it seems to work if I put the code all together outside the function. Definitely variable scope problem. Will have to rethink the approach (don't want to put all the codes in main). Thanks for your answer :)

Agent_L Over a year ago

@AnugerahErlaut check it now, maybe you'll like this

Anugerah Erlaut Over a year ago

In the end, I decided to flatten everything first and then see if I can group things up into functions. Thanks for the follow up tho :)

Collectives™ on Stack Overflow

passing cuda device pointer to host function

1 Answer 1

5 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

5 Comments

Your Answer

Sign up or log in

Post as a guest

Related