OpenCL user defined inline functions

问题:

Is it possible to define my own functions in OpenCL code, in order that the kernels could call them? It yes, where can I see some simple example?

回答1:

Function used to create program is ...

cl_program clCreateProgramWithSource  (     
    cl_context context,
    cl_uint count,
    const char **strings,
    const size_t *lengths,
    cl_int *errcode_ret)

You can place functions inside the strings parameter like this,

float AddVector(float a, float b)
{
    return a + b;
}

kernel void VectorAdd(
    global read_only float* a,
    global read_only float* b,
    global write_only float* c )
{
    int index = get_global_id(0);
    //c[index] = a[index] + b[index];
    c[index] = AddVector(a[index], b[index]);
}

Now you have one user defined function "AddVector" and a kernel function "VectorAdd"

回答2:

Based on the code samples here you can just write functions like:

inline int add(int a,int b)
{
   return a+b;
}

(Eg. look at the .cl file in the DXTC or bitonic sort examples.)

I don't know if that's an nvidia only extension but the OpenCL documentation talks about "auxiliary functions" as well as kernels.

回答3:

OpenCL supports auxiliary functions. See page 19 of this link for examples.

回答4:

I googled around a bit, and just kept coming back to this question :-P

In the end, what I did was use macros, since inlining would be implementation-dependent anyway, and macros don't seem to have any major disadvantage in the context of c99 OpenCL programs? eg:

#define getFilterBoardOffset( filter, inputPlane ) \
    ( ( filter * gInputPlanes + inputPlane ) * gFilterSizeSquared )
#define getResultBoardOffset( n, filter ) \
    ( ( n * gNumFilters + filter ) * gOutputBoardSizeSquared )

instead of:

inline float getFilterBoardOffset( float filter, int inputPlane ) { 
    return ( filter * gInputPlanes + inputPlane ) * gFilterSizeSquared; 
}
inline float getResultBoardOffset( float n, int filter ) { 
    return ( n * gNumFilters + filter ) * gOutputBoardSizeSquared; 
}