cudamallocpitch how to use cudamallocpitch

Accelerated Computing CUDA CUDA Programming and Performance
1

good afternoon all

i have two dimensional structure in the host memory and i need to transfer it to the device memory by allocating memory in the device and then copying from host to device.

Can some one help me with a sample code of using cudamallocpitch .

my structure looks like this

struct rngen
{
int seed[2];
int seed[2];
int init_seed;
int prime;
int prime_position;
int prime_next;
char *gentype;
int parameter;
int *multiplier;
};

now i have declared struct rngen **new and allocated for new.This new i need to transfer to the device memory first by allocating and then copying it into the device memory.

Can some body help me urgently.

2

The simplePitchLinearTexture example in the SDK has a complete example of using cudaMallocPitch and cudaMemcpy2D/cudaMemset2D.

But be warned that you will not be able to copy that structure from host memory directly to the device because it contains host pointers.

3

thank you for your reply. Then is there any way to copy directly the structure from host to device

waiting for your reply

4

As you have written it, no there is not. The problem is the pointers inside the structure. You cannot use host pointers on the device.

5

As avidday mentioned, you cannot do that. You HAVE to do a ‘deep-copy’ on that structure, in order to transfer all the contents from CPU to GPU (and vice-versa).

6

can you please elaborate on deep copy

waiting for your reply

thanking you

7

Here’s an example on what I meant by deep-copy: (untested)
You copy both the elements of the struct, as well as malloc and cudaMemcpy all the pointers inside it.

struct myStruct {
int* a;
int b;
};
myStruct h_data, d_data;
// ----- DEEP-COPY ----- //
h_data = (myStruct) malloc(sizeof(myStruct) * 1);;
h_data->b = 10;
h_data->a = (int) malloc(sizeof(int) * h_data->b);
cudaMalloc((void) &d_data, sizeof(myStruct) * 1);
cudaMemcpy(&h_data, &d_data, sizeof(myStruct) * 1, cudaMemcpyHostToDevice);
cudaMalloc((void**) &(d_data->a), sizeof(int) * h_data->b);
cudaMemcpy(&(d_data->a), &(h_data->a), sizeof(int) * d_data->b, cudaMemcpyHostToDevice);

8

thank you very much for your help.

So this is the only way an array of structures are copied.?

once again thank you

9

AFAIK, yes, this is the only way one can copy the array of structures… :(

10

Thank you very much for your kind help