Last half year more time was spent on the Vulkan project then originally anticipated. I will outline the approach we followed, what the current state is and known future quirks and bottlenecks.
Approach
Past milestones
-
being able to cross compile shaders between OpenGL, Metal and Vulkan. [Goal reached December 2022]
-
being able to run compute shaders test cases. [Goal reached February 2023]
-
memory and buffer management [Goal reached March 2023]
-
being able to start Blender (no viewport) [Goal reached May 2023]
-
Most of workbench working [Goal reached May 2023]
Current milestones
- Make sure that other renderengines are able to draw something. Might still have many artifacts, but at least you would be able to see something.
- Workbench-next, Eevee-next, Overlay-next, Grease Pencil, Cycles Viewport, Image engine
Future milestones
After everything can be seen, more time will be spent on adding support for ‘all’ platforms and improving the performance.
Current development is mostly done on discrete GPUs. On a weekly basis I switch between NVIDIA, AMD and Intel GPUs, to make sure that we don’t implement something that cannot be done on another brand of GPUs. This is mostly done on Linux. Of course the Blender community is sending in patches when something is not working on Windows. There are many configurations not working and they should be checked how we can get them fixed.
Vulkan can deliver excellent performance in a gaming pipeline, Blender isn’t a game and therefore isn’t (and cannot) be organized in a similar way as games. For now the performance is only around 5-10% of what I expect the final version will be. Mostly because I have been very conservative when sending commands to the GPU. When commands are schedule the backend will wait until that command is send to the GPU, then it waits until the GPU has finished the command, before scheduling of the next command can start. Vulkan can speed up when sending in multiple commands in a single go, and add barriers between commands that write to a resource, where the resource is being read by another command.
Until now we didn’t implement any performance improvements as most documentation on the internet is about gaming pipelines where developers have more control overthe full pipeline. We have to find our own path how to solve this. This path is getting clearer with each commit we make. Current idea is to have a custom command encoding that will send the commands and barriers at certain moments.
Potential Risks
Timeline
The initial plan is to have a usable experimental backend that can be downloaded and used to collect feedback at the end of 2023. Depending on initial feedback it might even be available in an official release. This means that we would not recommend to use it as the default backend yet.
There are still a lot of tasks that needs to be implemented in the upcoming period; however I have been asked to help with Eevee-next and that might shift the planning 1 or 2 months and that adds pressure to the original planning. It is currently unclear what the impact of this priority change has on the Vulkan Project.
Supporting data types
During the course of the project there has been quite a few gotchas. One of them are if a GPU support a certain data type (GPU_RGB16) and usage (Framebuffer attachment), it can be that the platform would return an error that that the combination isn’t supported. In order to continue we sometimes need to reconfigure those buffers so we don’t need to transform the content of the buffer all the time.
Inside Blender we have more control over the actual buffer data and can update the original data or even phase out problematic data types, but this isn’t always possible. Blender Python API allows add-on developers still to create those buffers. The vulkan backend currently detects invalid usages and tries to transform the data to a compatible format when uploading the data to the GPU.
It is unclear if during the project we will need to implement more work-arounds to ensure that ‘most’ used data-type/usage combinations would be working from an add-on developer point of view.
Final note
A huge thanks to the community developers for lending a hand, providing patches and sharing knowledge. The implementation has been changed based on what we have heard back from you!
During the Blender Conference I will also share more information about the project and go more into the details of Vulkan.