lavapipe: a *software* swrast vulkan layer FAQ

(project was renamed from vallium to lavapipe)

I had some requirements for writing a vulkan software rasterizer within the Mesa project. I took some time to look at the options and realised that just writing a vulkan layer on top of gallium's llvmpipe would be a good answer for this problem. However in doing so I knew people would ask why this wouldn't work for a hardware driver.

tl;dr DO NOT USE LAVAPIPE OVER A GALLIUM HW DRIVER,

What is lavapipe?

The lavapipe layer is a gallium frontend. It takes the Vulkan API and roughly translates it into the gallium API.

How does it do that?

Vulkan is a lowlevel API, it allows the user to allocate memory, create resources, record command buffers amongst other things. When a hw vulkan driver is recording a command buffer, it is putting hw specific commands into it that will be run directly on the GPU. These command buffers are submitted to queues when the app wants to execute them.

Gallium is a context level API, i.e. like OpenGL/D3D10. The user has to create resources and contexts and the driver internally manages command buffers etc. The driver controls internal flushing and queuing of command buffers.
 
In order to bridge the gap, the lavapipe layer abstracts the gallium context into a separate thread of execution. When recording a vulkan command buffer it creates a CPU side command buffer containing an encoding of the Vulkan API. It passes that recorded CPU command buffer to the thread on queue submission. The thread then creates a gallium context, and replays the whole CPU recorded command buffer into the context, one command at a time.

That sounds horrible, isn't it slow?

Yes.

Why doesn't that matter for *software* drivers?

Software rasterizers are a very different proposition from an overhead point of view than real hardware. CPU rasterization is pretty heavy on the CPU load, so nearly always 90% of your CPU time will be in the rasterizer and fragment shader. Having some minor CPU overheads around command submission and queuing isn't going to matter in the overall profile of the user application. CPU rasterization is already slow, the Vulkan->gallium translation overhead isn't going to be the reason for making it much slower.
For real HW drivers which are meant to record their own command buffers in the GPU domain and submit them direct to the hw, adding in a CPU layer that just copies the command buffer data is a massive overhead and one that can't easily be removed from the lavapipe layer.

The lavapipe execution context is also pretty horrible, it has to connect all the state pieces like shaders etc to the gallium context, and disconnect them all at the end of each command buffer. There is only one command submission queue, one context to be used. A lot of hardware exposes more queues etc that this will never model.

I still don't want to write a vulkan driver, give me more reasons.

Pipeline barriers:

Pipeline barriers in Vulkan are essential to efficient driver hw usage. They are one of the most difficult to understand and hard to get right pieces of writing a vulkan driver. For a software rasterizer they are also mostly unneeded. When I get a barrier I just completely hardflush the gallium context because I know the sw driver behind it. For a real hardware driver this would be a horrible solution. You spend a lot of time trying to make anything optimal here.

Memory allocation:

Vulkan is built around the idea of separate memory allocation and objects binding to those allocations. Gallium is built around object allocation with the memory allocs happening implicitly. I've added some simple memory allocation objects to the gallium API for swrast. These APIs are in no way useful for hw drivers. There is no way to expose memory types or heaps from gallium usefully. The current memory allocation API works for software drivers because I know all they want is an aligned_malloc. There is no decent way to bridge this gap without writing a new gallium API that looks like Vulkan. (in which case just write a vulkan driver already).

Can this make my non-Vulkan capable hw run Vulkan?

No. If the hardware can't do virtual memory properly, or expose features for vulkan this can't be fixed with a software layer that just introduces overhead.


Comments

  1. Is this going to help with thread scaling on many core cpus like the 3990x? (layman asking)

    ReplyDelete
    Replies
    1. Nope, though I've some patches for llvmpipe that help saturate CPU threads more, so maybe makes sense to just increase the max threads in llvmpipe.
      https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4385

      Delete

Post a Comment

Popular posts from this blog

Fedora 38 LLVM vs Team Fortress 2 (TF2)

tinygrad + rusticl + aco: why not?

nvk: the kernel changes needed