Thrust vector of structs containing pointers

Question

I have the following data structure:

struct Data {
    int* ptr;
    int a;
    int b;
}

I need an array of such structures on the GPU passed from the host. The easy way to do that is

thrust::host_vector<Data> h;
... // fill vector 'h'
thrust::device_vector<Data> d = h;

The tricky moment is how to allocate the memory for ptr and how to copy data there. Is there any suggestions?

Pointers can be copied just like other values, but presumably you want to dereference pointers in GPU code. The right way to transfer data depends on what you are doing with the pointers. What does ptr point to? Is the data read or written on the GPU? Is it possible that several instances of Data contain pointers to the same object? — Heatsink
– Heatsink, Commented Apr 23, 2012 at 14:25
@Heatsink: ptr is a pointer to the device memory individual for each Data object. I just now realized that I can write something like cudaMalloc( &h[i].ptr,...) and later assign d=h. Is it true? I can't say why it was confusing me before. ) Do you want to get the accepted answer? Please, formulate your comment as an answer then. — AdelNick
– AdelNick, Commented Apr 24, 2012 at 5:56

Devin Lane · Accepted Answer · 2012-04-25 00:00:54Z

1

Updated based on your comment to @Heatsink, if I understand correctly you're saying that given

struct Data {
    int* ptr;
    int a, b;
};

and

vector<Data> data;

data[0].ptr points to GPU memory containing data[0].a and data[0].b If this is correct, then I would recommend the following organization instead:

struct Data {
    int a, b;
};

thrust::host_vector<Data> h;
thrust::device_vector<Data> d = h;

The GPU memory for h[i] is simply d[i]. I would not recommend storing a pointer per-element to the GPU memory, nor would you want to allocate separate GPU memory for each data object (would be horrifically slow.) Your compute code will still probably be faster if you use separate arrays as well.

Generally organizing your data as structure-of-arrays instead of array-of-structures is preferred for several reasons, including alignment and ease of load coalescing.

edited Apr 25, 2012 at 0:00

answered Apr 23, 2012 at 16:59

Devin Lane

1,0441 gold badge9 silver badges15 bronze badges

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

4 Comments

AdelNick Over a year ago

With respect to my problem, it is known that each block will access only 1 element of an array so I can sacrifice the universality to the code clarity )

Devin Lane Over a year ago

@AdelNick I updated my answer to reflect the clarification you provided to Heatsink above.

AdelNick Over a year ago

no, ptr does not store a and b. It is the separate independent piece of memory, so for each Data object I need to allocate memory for ptr and fill it with the data and also assign the values for a and b.

Devin Lane Over a year ago

@AdelNick some information about how much data is stored in ptr and where it comes from would be useful then. My comment about not wanting to separately allocate device memory for each Data object still stands.

Collectives™ on Stack Overflow

Thrust vector of structs containing pointers

1 Answer 1

4 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Your Answer

Sign up or log in

Post as a guest

Related