Custom types in OpenCL kernel

Question

Is it possible to use custom types in OpenCL kernel like gmp types (mpz_t, mpq_t, …) ?

To have something like this (this kernel doesn't build just because of #include <gmp.h>) :

#include <gmp.h>
__kernel square(
   __global mpz_t* input,
   __global mpz_t number,
   __global int* output,
   const unsigned int count)
{
   int i = get_global_id(0);
   if(i < count)
       output[i] = mpz_divisible_p(number,input[i]);
}

Maybe by adding different arguments to the fourth parameter (options) of clBuildProgram ?

Or does OpenCL already have types that can handle large numbers ?

score 22 · Answer 1 · answered Jul 22 '10 at 12:20

Generally you can use any types in an OpenCL program. But since imports do not work, you have to re-define them within the same program. For example:

typedef char my_char[8];

typedef struct tag_my_struct
{
    long int        id;
    my_char         chars[2];
    int             numerics[4]
    float           decimals[4];
} my_struct;

__kernel void foo(__global my_struct * input,
                  __global int * output)
{
    int gid = get_global_id(0);
    output[gid] = input[gid].numerics[3]== 2 ? 1 : 0;
}

However, you obviously need to keep the definitions within and outside OpenCL the same. Also make sure the type has the same size on both device and host (using a sizeof(my_struct) should do the trick). In some cases I had to adjust the definitions, to have matching sizes.

To ensure equally sized types, it's a good idea to use the cl_* types in host code (cl_int, cl_long, cl_float2, etc.). — dietr, Sep 20 '10 at 06:21
@dietr It also helps with code clarity, as in "this variable is meant to be passed to a kernel" — Thomas, Nov 23 '13 at 01:27

score 8 · Answer 2 · answered Aug 23 '13 at 22:07

I used VHristov's answer and dietr's comment to get mine working. This code works for me in OpenCL 1.2

kernel

typedef struct tag_my_struct{
  int a;
  char b;
}my_struct;

__kernel void myKernel(__global my_struct *myStruct)
{
    int gid = get_global_id(0);
    (myStruct+gid)->a = gid;
    (myStruct+gid)->b = gid + 1;
}

host

typedef struct tag_my_struct{
  cl_int a;
  cl_char b;
}my_struct;

void runCode() 
{
    cl_int status = 0;
    my_struct* ms = new my_struct[5];

    cl_mem mem = clCreateBuffer(*context, 0, sizeof(my_struct)*5, NULL, &status);
    clEnqueueWriteBuffer(*queue, mem, CL_TRUE, 0, sizeof(my_struct)*5, &ms, 0, NULL, NULL);

    status = clSetKernelArg(*kernel, 0, sizeof(ms), &mem);

    size_t global[] = {5};
    status = clEnqueueNDRangeKernel(*queue, *kernel, 1, NULL, global, NULL, 0, NULL, NULL);

    status = clEnqueueReadBuffer(*queue, mem, CL_TRUE, 0, sizeof(my_struct)*5, ms, 0, NULL, NULL);

    for(int i = 0; i < 5; i++)
        cout << (ms+i)->a << " " << (ms+i)->b << endl;
}

output

0 ☺

1 ☻

2 ♥

3 ♦

4 ♣

Why use `(myStruct+gid)->a = gid;` instead of `myStruct[gid].a = gid;`. That looks ugly for me (however, will give same results) — DarkZeros, Sep 24 '13 at 10:00

Adam S. · Accepted Answer · 2010-03-14T19:07:33.510

4

You can use custom types but anything used in the kernel needs to be specifically written for OpenCL. Check out this website perhaps for how to implement larger precision numbers: FP128

Edit: NVIDIA's CUDA SDK has a complex number data type, it's not ideal but may give you some ideas on how they go about it, OpenCL should be similar.

edited Mar 14 '10 at 19:07

answered Mar 14 '10 at 18:43

Adam S.

1,251
1
12
16

score 3 · Answer 4 · edited May 23 '17 at 10:30

3

If you want to include header files into kernel file, you can add the -l dir as an argument to clBuildProgram, where dir is the directory with the header files.

More explanation here: include headers to OpenCL .cl file

Source: http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clBuildProgram.html

edited May 23 '17 at 10:30

Community

1
1

answered Sep 24 '13 at 01:58

gallo

546
7
10

Custom types in OpenCL kernel

4 Answers4