Question about "Vertex shader runs once per vertex (broadcast to all views)" in multiview docs

[zhywzs]

Hi! First, thank you for the excellent multiview documentation and examples — they're some of the clearest I've found on using VK_KHR_multiview in practice.

I have a question about one specific performance claim in the Multiview Rendering docs:

GPU: Vertex shader runs once per vertex (broadcast to all views)

I've been trying to verify this independently, but I can't find other sources (Vulkan spec, GPU vendor docs, GDC talks, etc.) that explicitly state this "broadcast" behavior at the vertex level. Most multiview documentation focuses on the single-pass draw-call reduction and gl_ViewIndex, but rarely discusses whether the vertex shader output is broadcast vs. replayed per view.

I'd really appreciate clarification on the following:

1. Source / authority: Is this behavior guaranteed by the Vulkan spec (could you point me to the relevant section?), or is it an observed optimization on certain hardware that the doc generalizes? I did find spec language about vertex stages executing "once" in multiview, but I wanted to confirm the "broadcast" framing is canonical.

2. Hardware limitations: Are there GPU architectures or driver stacks where this does not hold — i.e., where the vertex shader still runs per-view despite multiview being enabled? For instance, older desktop GPUs with partial multiview support.

3. Enabling conditions: Beyond VK_KHR_multiview / Vulkan 1.1, are there specific feature bits (e.g., multiviewGeometryShader, multiviewTessellationShader) or VkPhysicalDeviceMultiviewProperties limits that gate this optimization? I'm wondering if the "broadcast" behavior is contingent on anything beyond basic multiview support.

4. Mobile GPUs (tile-based renderers): The doc mentions tile-based renderers benefit "greatly," but does the vertex broadcast behavior work the same way on Mali / Adreno / PowerVR as on desktop IMRs? Or do tile-based architectures handle this differently under the hood (e.g., vertex shading per tile rather than per view)?

This isn't a bug report — the doc is great. I'm just trying to build a more precise mental model of multiview performance across different hardware targets. Thanks for any insight!

[seanharmer]

It should probably say "vertex fetching" rather than "vertex processing" as far as I understand it. We likely need to multiply by different view-projection matrices (indexed by gl_ViewIndex) so the vertex shader has to run twice (thinking about the typical VR case). However, the same input vertex can be passed in twice without having to refetch it via the index-buffer prior to the vertex shader invocation.

[zhywzs]

It should probably say "vertex fetching" rather than "vertex processing" as far as I understand it. We likely need to multiply by different view-projection matrices (indexed by gl_ViewIndex) so the vertex shader has to run twice (thinking about the typical VR case). However, the same input vertex can be passed in twice without having to refetch it via the index-buffer prior to the vertex shader invocation.

😘Thank you for your answer!

I completely agree with what you said! Reusing the index-buffer and running the vertex shader multiple times is the way I can understand it to work. As far as I know, GPU instances can also achieve the effect of vertex data reuse (by using the gl_InstanceIndex), and can be passed to multiple fragment shaders, and there are no output format restrictions imposed by the multiview.

🆘I would like to know if this approach, does multiview have exactly the same performance as the GPU instance?