(Updated after also testing with Intel.)
I'm running on a MacBook Pro 2016, with Radeon Pro 455 and the latest OS update.
There seems to be a memory leak problem when running an application with the Radeon. Instrumenting shows many leaks in AMDRadeonX4000GLDriver, mostly in gldCopyTextureDataWithQueue but also in gldCopyBufferDataToTextureWithQueue. The vast majority of them are in a worker thread and I'm guessing they have to do with kernel invocations. Running with Intel shows no leaks.
(There's also lack of support for CL_SNORM_INT16 for anything but RGBA, and I reported this through the Catalyst driver feedback page, but it might be worth mentioning here too.)