Is there a better way to store pairs of ints? - c

Question

I'm sorry if this is very basic but I'm still learning all that things I can do in C and can't figure out how to do this.
I create pairs of ints in a program and then need to store them. The way I have been doing it so far is by creating a struct:
struct list_el {
short *val; //first value
short *val2; //second value
struct list_el * next;
};
typedef struct list_el item;
I can iterate though the list fine in my normal program but I want to send this to Cuda and am not sure how to transfer the whole struct into Cuda(I know I can make a reference to it). I'm wondering if there's another way I can structure this data so maybe its array? The format I need is in is just simple pairs (something like this 10:5, 20:40, etc..). I thought worst case I can use a char string and have the pairs as characters and then parse them once the main array is in Cuda but I'm wondering if there's a better way create this list of list?

Answer 1

Assuming that you can use two separate arrays, and thinking about how to use/read/write them in CUDA, I will arrange the data in two arrays mainly due to coalesced accesses from global memory wihtin a kernel.
int *h_val1, *h_val2; // allocate arrays in the host and initialize them
Let be N the size of the arrays, allocate the arrays in device memory
int *d_val1, *d_val2;
cudaMalloc( (void**) &d_val1, N * sizeof(int) );
cudaMalloc( (void**) &d_val2, N * sizeof(int) );
and copy data from host to device memory
cudaMemcpy(h_val1, d_val1, N * sizeof(int), cudaMemcpyHostoToDevice);
cudaMemcpy(h_val2, d_val2, N * sizeof(int), cudaMemcpyHostoToDevice);
Configure and launch your kernel to run as much threads as element in the array.
// kernel configuration
dim3 dimBlock = dim3 ( BLK_SIZE, 1, 1 );
dim3 dimGrid = dim3 ( (N / BLK_SIZE) + 1 );
yourKernel<<<dimGrid, dimBlock>>>(d_val1, d_val2);
With this in mind, implement your kernel
__global__ void
yourKernel(int* val1, int* val2, N)
{
// map from threadIdx/BlockIdx to index position
int gid = threadIdx.x + blockIdx.x * blockDim.x;
if (gid < N)
{
int r_val1 = val1[ idx ]; // load from global memory to register
int r_val2 = val2[ idx ]; // load from global memory to register
// do what you need to do with pair val1:val2
}
}
Do not forget to check for errors when calling CUDA functions.

Answer 2

Instead of storing something that references two ints, store something that holds a copy of the ints.
struct list_el {
int val; //first value
int val2; //second value
struct list_el * next;
};
typedef struct list_el item;
Sometimes it is preferable to hold a reference, sometime it is preferable to hold a value. Depending on what you are attempting to do, use the right tool for the job.
By the way, your reference holding struct was only holding references to shorts. To really hold references to ints, you need
struct list_el {
int *val; //reference to first value
int *val2; //reference to second value
struct list_el * next;
};
typedef struct list_el item;
Note that if you hold a reference, the rest of your program should not dispose of the reference's memory before you dispose of the struct reference to prevent accessing memory that is no longer associated with the program (which is an error).
There are other techniques, if you don't want to use list like constructs.
int val[2] = { 1, 2 };
will store two ints, but only two ints.
int val[2][9];
will store nine pairs of two ints, and could easily also be represented as
int val[9][2];
And of course, there is the old standby
int val = 3;
int val2 = 4;

Answer 3

How about just using a two-dimensional array?
int pairs[30][2];
pairs[0][0] = 10;
pairs[0][1] = 5;
// etc.
I'd have to test it, but I think I tested it, and you can even do something like
int pairs[][2] = {{10, 5}, {20, 40}, ...};
for initialization.
NOTE: This method works well if you know how many pairs you will have ahead of time and the number doesn't grow/shrink (in large amounts). If you have a widely variable number of pairs, sticking with a list of structs and using Edwin's answer would probably be better in the long run.

Answer 4

Having a two dimensional array is a good solution, but I am going to answer as if you are keeping your struct solution.
There's nothing wrong with your storing the short ints in a struct, but I would not store the values in short *. To me it is not worth dynamically allocating memory as you need a new structure.
You could have an array of structs to store this data. Here is an example of a fixed size array of item.
#include <stdio.h>
struct list_el {
short val; //first value
short val2; //second value
};
typedef struct list_el item;
item listA[20];
int main()
{
listA[0].val = 1;
listA[0].val2 = 2;
printf("\n%i %i\n", listA[0].val, listA[0].val2);
return 0
}
Even if you make the argument that you won't know in advance how many of these
structs you will have, I would only allocate space for the array like this:
#include <stdio.h>
#include <stdlib.h>
struct list_el {
short val; //first value
short val2; //second value
};
typedef struct list_el item;
item * p_list_el, * pCurStruct;
int main()
{
int idx;
/* p_list_el is the pointer to the array. Don't modify.
pCurStruct can be modified to walk the array. */
p_list_el = malloc(sizeof(item) * 20);
for(idx=0, pCurStruct=p_list_el; idx < 20; idx++)
{
pCurStruct[idx].val = idx;
pCurStruct[idx].val2 = idx + 1;
}
for(idx=0, pCurStruct=p_list_el; idx < 20; idx++)
{
printf("\n%i %i\n", pCurStruct[idx].val, pCurStruct[idx].val2);
}
free(p_list_el);
}