Benchmarking Radeon Open Compute ROCm 1.4 OpenCL

AMDGPU-PRO 16.50 beat out the ROCm 1.4 driver with JuliaGPU.

Both OpenCL implementations failed with our Blender 3D OpenCL test.

And more wins for AMDGPU-PRO's current OpenCL implementation.

With the Polaris RX 460/480 cards on ROCm, they failed at compiling the OpenCL kernels for LuxMark's microphone test. I ended up killing the processes after they were held up by Clang for 30+ minutes to compile the kernel...

The OpenCL support in ROCm and is a step in the right direction, especially with working towards open-sourcing this OpenCL implementation as a much more viable option than the older Clover-based Gallium3D OpenCL approach. However, as shown by these results, there still is a ways to go. The OpenCL driver in AMDGPU-PRO 16.50 was faster in a majority of the workloads, ROCm 1.4 currently only supports Fiji and Polaris graphics processors, ROCm OpenCL failed with a few tests, and the current ROCm approach is suited for just Ubuntu 16.04 (similar to AMDGPU-PRO) and doesn't yet play nicely on Ubuntu 16.10, etc. Though if AMD can get their OpenCL driver open-sourced along with their long-talked-about official Vulkan driver in 2017, this should be an excellent year and complement the great state of their RadeonSI Gallium3D OpenGL driver.