The number returned by clDeviceInfo should be per CU.
Originally posted by: zhuzxy the AMD a8-3850 has 32K local mem, does that belongs to a CU? (that means total 160K total since it has 5 CU)
Yes it is per CU and total is 160K per A8-3850.
And another question, how can I get teh wavefronts number from appprofiler or other tool? If not possible, what's the right way?
1. You get clGetKernelWorkGroupInfo(..., CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE, ...)
2. You can calculate from profiler counters GlobalWorkSize and Wavefronts.