With Vulkan turning one year old I decided to run some fresh comparison benchmarks of Mesa 17.1-dev RADV (as well as some RadeonSI OpenGL results for reference) compared to AMD's latest public hybrid driver release, the AMDGPU-PRO 16.60.
RADV wasn't around one year ago this week but was only born after the fact in the absence of AMD open-sourcing their Vulkan driver code that they have talked about doing but to this day have still not achieved. Anyhow, this should be an interesting comparison using AMDGPU-PRO 16.60 on Ubuntu 16.04 LTS compared to Linux 4.10 and Mesa 17.1-dev loaded on the same system. A Radeon RX 470 was used for all of the benchmarking this round.
It's been a couple weeks since last running many RADV benchmarks and in that time fast clears were enabled by default and a variety of other improvements made by Bas Nieuwenhuizen, Dave Airlie, and others.
For those curious, on Mesa 17.1-dev Git the Vulkan device features not yet enabled are tessellation shader, sample rate shading, ETC2 texture compression, ASTC LDR texture compression, pipeline statistics query, shader storage image multi-sample, int64, int16, variable multi-sample rate, and inherited queries. RADV also isn't marked as Vulkan 1.0 conformant yet by the Vulkan CTS.
RADV continues working great for Dota 2 while on this test system when trying to run The Talos Principle I was still running into a failure to set the video mode error, something that's been happening recently on my RADV test systems while previously it had worked fine. RADV also works for vkQuake and other Vulkan samples. RADV is also said to be working fine internally at Valve for their Steam VR efforts.
On the following pages are some fresh RADV (and RadeonSI OpenGL reference) versus AMDGPU-PRO 16.60. More RadeonSI OpenGL vs. AMDGPU-PRO 16.60 benchmarks will be posted in a follow-up article.