Created attachment 121742 [details]
dmesg after running the uint_add16() kernel

(In reply to Hristo Venev from comment #4)
> Here are two kernels that fail:
> __kernel void uint_div(__global const uint *a,  __global const uint *b,
> __global uint *c){
>     c[0]=a[0]/b[0];
> }
> __kernel void uint_add16(__global const uint *a,  __global const uint *b,
> __global uint *c){
>     for(uint i=0;i<16;i++) c[i]=a[i]+b[i];
> }

I actually can confirm that the 2nd kernel does cause a GPU stall in the radeon
driver (r600/Barts). It causes no problem with amdgpu (radeonsi/Tonga).

I am using kernel 4.5.0-rc3, current llvm 3.8 branch and current mesa git.

As I am not overly good with OpenCL, is this kernel somehow problematic?