Hi
I have tried running NVidia's OpenCL example of scan (prefixsum) on an Intel CPU but without luck.
The program fails with a memory error.
If I remove the barrier synchronization lines in the following part of the code it runs without errors, but the results are offcause wrong:
[cpp]inline uint scan1Inclusive(uint idata, __local uint *l_Data, uint size){
uint pos = 2 * get_local_id(0) - (get_local_id(0) & (size - 1));
l_Data[pos] = 0;
pos += size;
l_Data[pos] = idata;
for(uint offset = 1; offset < size; offset <<= 1){
barrier(CLK_LOCAL_MEM_FENCE); //Fails with Intel openCL
uint t = l_Data[pos] + l_Data[pos - offset];
barrier(CLK_LOCAL_MEM_FENCE); //Fails with Intel openCL
l_Data[pos] = t;
}
return l_Data[pos];
}[/cpp] I have tested other code using barriers that worked fine. The main difference is that this example is using a three layer nesting of inlining. The inserted code is the innermost function.
The code runs fine on a NVidia graphics card and on CPU compiled with the AMD openCL compiler.
-Jens
