CNA-BUG-264: Vulkan passes XNA's normalised RasterizerState.DepthBias to vkCmdSetDepthBias unscaled, so a realistic bias has no effect

CNA snapshot 009d40f5  ·  Known Issues › Current bugs  ·  source links pinned to 009d40f5

VulkanRenderer hands DepthBias to vkCmdSetDepthBias raw, but that constant factor counts minimum depth steps; EasyGL, OpenGL4, SDL_GPU and Direct3D 11/12 scale XNA's normalised value by the depth format, so -0.0001 offsets nothing on Vulkan.

Identifier

CNA-BUG-264

Category

Bug

Subsystem

Graphics & renderers

Status

Open

Verified against

CNA 009d40f5 (009d40f5dd085c4e674d3479675fac84b12b3e0a)

Severity

Medium (a triage suggestion, not a project priority)

Evidence basis

Source-verified: read at TARGET, not executed

Tests touching this area

Yes: see Current tests

Affected contract

Microsoft::Xna::Framework::Graphics::RasterizerState::DepthBias and SlopeScaleDepthBias on the VULKAN renderer (VulkanRenderer::ApplyRasterizerState)

Expected behaviour

XNA 4.0 is Direct3D 9, where DepthBias is a normalised depth value added to the depth. CNA's renderer contract says so indirectly: IGraphicsRenderer.hpp documents in IRenderTargetRenderer::DepthBufferBitsEXT that an API whose bias is a multiple of the smallest resolvable depth step, naming Vulkan's depthBiasConstantFactor, needs the depth width to convert. FNA3D multiplies the XNA value by 65,535 (16-bit depth) or 16,777,215 (24-bit), and CNA's EasyGL, OpenGL4, SDL_GPU and Direct3D 11/12 renderers convert it as well, so a sample that lifts a flat shadow with DepthBias = -0.0001f lifts it on each of them.

Actual behaviour at TARGET

VulkanRenderer::ApplyRasterizerState stores DepthBias and SlopeScaleDepthBias as given, and the draw replay calls vkCmdSetDepthBias(cb, draw.depthBias, 0.0f, draw.slopeScaleDepthBias) with no depth-format scale; no Vulkan render-target class overrides DepthBufferBitsEXT. A constant factor of -0.0001 is about 1e-4 of one depth step, that is, no bias. vulkan_depth_bias_test.cpp uses -1e6 and -2e3 on purpose (“bias magnitudes are deliberately large”) and so cannot see the defect, and CNA's records close the failing Vulkan_DepthBias case as a test defect (VULKAN-091) without examining the product behaviour. The source comment that says the mapping matches FNA omits the scale. OpenGL4Renderer::ApplyCurrentDepthBias carries the rule Vulkan lacks: passing the XNA value through raw “is indistinguishable from no bias at all”.

Source locations

• modules/renderers/vulkan/src/VulkanRenderer.cpp — ApplyRasterizerState stores the bias raw; the draw replay calls vkCmdSetDepthBias with it
• modules/renderers/vulkan/examples/vulkan_depth_bias_test.cpp — deliberately huge magnitudes, so it cannot see the missing scale
• modules/graphics/include/CNA/Internal/Renderers/Common/IGraphicsRenderer.hpp — IRenderTargetRenderer::DepthBufferBitsEXT documents the conversion that Vulkan's depthBiasConstantFactor needs
• modules/renderers/opengl4/src/OpenGL4Renderer.cpp — ApplyCurrentDepthBias: the FNA3D 16/24-bit scale table
• modules/renderers/sdl-gpu/src/SdlGpuRenderer.cpp — XnaToSdlDepthBiasScale
• docs/easygl_bugs.md — CNA's own record (2026-09-08): Vulkan and PortableGL still carry the unscaled premise

Evidence

Read at 009d40f5; nothing was built or executed. Reported as a RESTORE-ISSUE by the C05 conservation reviewer (an earlier absorption record had closed the extreme-bias Vulkan failure as a test defect) and re-verified by the orchestrator against the Vulkan and OpenGL4 sources, CNA's docs/easygl_bugs.md and the Direct3D 11/12 mapping (D3DCommon, which scales it at this snapshot although CNA's 2026-09-08 note still lists DirectX 11 as unscaled). The same note lists PortableGL, whose glPolygonOffset(slopeScaleDepthBias, depthBias) call is raw in the source too; PortableGL's offset units were not analysed here.

Focused reproduction

No focused reproduction is known. Nothing has been invented here; the evidence above is what exists.

Current tests

Vulkan_DepthBias uses deliberately huge magnitudes and still fails on CNA's own record; no test uses a realistic normalised bias on Vulkan, whereas EasyGL has EasyGLDepthBias.IsScaledByTheDepthBuffersOwnResolution.

Regression test

Draw a quad slightly in front of a coplanar one with DepthBias = -0.0001f on VULKAN and expect it to win the depth test, as on EasyGL; or unit-test the conversion the way EasyGL's EasyGLDepthBias tests do.

Blast radius

Content on VULKAN that relies on DepthBias for decals, planar shadows or z-fight avoidance; any value inside XNA's normalised range has effectively no effect there, while the same content works on the renderers that scale.

Workaround

Prefer offsetting the geometry, or, on VULKAN only, a bias several orders of magnitude larger (the test uses -1e6); both are renderer-specific hacks.