This is a Experimental version of OpenCL by AMD Research, we now recommend you to use The official BVLC Caffe OpenCL branch is over at Caffe branch now at https://github.com/BVLC/caffe/tree/opencl
The problem was debug printf in my opencl code. Apparently, there is memory leak in AMD driver when printf appears in the kernel, even if that printf is never called.
Memory keeps growing while training a net. Testing is turned off, so the problem must be in training.
Using AMD Radeon.