Vulkan multiview: primitive fully clipped in view 0 is incorrectly discarded in view 1

Gaming and Visualization Technologies APIs Vulkan
, ,
1

I am observing what appears to be incorrect per-view primitive clipping behavior in Vulkan multiview on NVIDIA.

Hardware
GPU NVIDIA GeForce RTX 5070 Laptop GPU
OS / Driver Model Windows, WDDM
NVIDIA Driver 610.88
CUDA UMD 13.3
VRAM 8 GB (8151 MiB reported)
OS Win11 - 10.0.26200

Setup

  • Vulkan multiview rendering with two views
  • Vertex shader uses BuiltIn ViewIndex
  • Per-view user clip distances are written through BuiltIn ClipDistance
  • The generated SPIR-V declares Capability MultiView and Capability ClipDistance
  • The SPIR-V passes spirv-val
  • The issue is reproducible on NVIDIA under Windows
  • ANGLE is only the frontend used to generate the Vulkan/SPIR-V workload

The failure is asymmetric between view 0 and view 1.

Test 1:

ViewIndex == 0:
ClipDistance[0] = +1.0

ViewIndex == 1:
ClipDistance[0] = -1.0

Result:

  • view 0 rasterizes the primitive
  • view 1 rejects the primitive

This is the expected behavior.

Test 2:

ViewIndex == 0:
ClipDistance[0] = -1.0

ViewIndex == 1:
ClipDistance[0] = +1.0

Expected result:

  • the primitive is clipped/rejected in view 0
  • the same primitive is retained and rasterized in view 1

Actual result:

  • the primitive is absent from both views

The other ClipDistance components are positive in these tests, so only one user clip plane is involved.

ViewIndex itself is behaving correctly. I verified per-view shader execution independently by producing different ordinary vertex outputs for the two ViewIndex values, and those results are correct.

The generated SPIR-V also shows the expected data flow:

BuiltIn ViewIndex
    -> per-view clip-distance computation
    -> BuiltIn ClipDistance

There is no shared clip-distance storage between the two view invocations, and the module validates successfully with spirv-val.

The observed behavior therefore appears to be that when a primitive is completely rejected by a user clip plane for view 0, that primitive is also removed from the multiview rendering operation before view 1 can independently retain it.

In other words, the observed behavior is effectively:

view 0 fully outside user clip plane
view 1 inside user clip plane
-> primitive rejected for both views

whereas I would expect clipping and primitive survival to be evaluated independently for each active view in the multiview render pass.

The opposite case works correctly:

view 0 inside
view 1 fully outside
-> view 0 rasterizes, view 1 rejects

As a workaround, I currently avoid BuiltIn ClipDistance for view 0 and mask the unused region using depth instead, while continuing to use normal user clipping for view 1.

Is this a known NVIDIA Vulkan multiview issue involving BuiltIn ClipDistance / per-view primitive clipping?