CNA-BUG-099: 24 classic SDL_GPU CTests still fail in CNA's latest recorded run; CNA's records name a cause for only three of them

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

CNA's SDL_GPU records list 26 classic ^SdlGpu CTests still failing just before TARGET; two are analysed in their own entries. Of the other 24, covering formats, render targets, render state and recovery, the records give a cause for two (a test that clears depth on a DepthFormat::None device, and a double free in the constructor failure path), a guess for a third, and nothing for the other 21.

Identifier

CNA-BUG-099

Category

Bug

Subsystem

Graphics & renderers

Status

Open

Verified against

CNA 009d40f5 (009d40f5dd085c4e674d3479675fac84b12b3e0a)

Severity

Medium (a triage suggestion, not a project priority)

Evidence basis

Recorded by CNA: CNA's own recorded run, not repeated here

Tests touching this area

Yes: see Current tests

Affected contract

SDL_GPU renderer behaviour pinned by the classic ^SdlGpu CTest set (renderer examples, shared parity fixtures, EasyGL oracle cases and SdlGpuIndexedDrawRangeTest)

Expected behaviour

Each registered ^SdlGpu CTest states part of the renderer's contract; the parity fixtures and EasyGL oracle cases are the same checks other renderers are held to. SDL_GPU is declared Supported in GraphicsBackendMaturity.hpp, so a routine run of its own suite is expected to pass.

Actual behaviour at TARGET

• PSG-0009 (plan_pre_sdlgpu_closeout.md) recorded 202 passing and 27 failing of 229 classic SDL_GPU CTests. The SDL_GPU modern-graphics and street workstreams re-ran the set after their changes and report "the same 27 by name" (SMG-0042: 204 / 27 of 231 with two new tests) until STREETS-0007 repaired SdlGpu_Smoke's stale checks; the last record before TARGET (STREETS-0008, plan_street_sdlgpu.md) is 26 failures, "none new". Two of the 26 are analysed elsewhere (SdlGpu_OcclusionQuery_Limitation under CNA-BUG-012, SdlGpu_2D_HeadlessDriver under CNA-BUG-203). The other 24, grouped here by name only:
• Formats and targets: SdlGpu_Texture2D_FormatMatrix, SdlGpu_Texture2D_FormatMatrixFallback, SdlGpu_TextureCube_FormatMatrix, SdlGpu_TextureCube_FormatMatrixFallback, SdlGpu_Parity_compressed_cube, SdlGpu_Parity_compressed_cube_fallback, SdlGpu_RenderTargetCube_Formats, SdlGpu_RenderTargetCube_MsaaMip, SdlGpu_RenderTarget_SetData, SdlGpu_TextureFilterMipContract, SdlGpu_EasyGLOracle_surface_format_throws.
• State and draws: SdlGpu_RenderState, SdlGpu_DepthBias, SdlGpu_Deferred_Scissor, SdlGpu_DrawLineTopology, SdlGpu_EasyGLOracle_depthstencilstate_stencil_twosided, SdlGpu_EasyGLOracle_draw_range_validation, SdlGpu_Parity_dual_texture_terms, SdlGpuIndexedDrawRangeTest.MutatingSourceBuffersAfterQueuingDoesNotChangeQueuedDraws, SdlGpuIndexedDrawRangeTest.RejectsIndexedRangesOutsideTheBoundBuffers.
• Device, presentation and lifetime: SdlGpu_BackbufferFormat, SdlGpu_ConstructorExceptionSafety, SdlGpu_SwapchainRecovery, SdlGpu_DepthlessCube_SpriteBatchCompatibility.
• CNA's records name a cause for only three of the 24. SdlGpu_BackbufferFormat: plan_gpu_test_isolation.md (GTI-0010) records that the test clears depth on a DepthFormat::None device, which GraphicsDevice::Clear refuses under XNA's rule (a stale test expectation). SdlGpu_ConstructorExceptionSafety: plan_sdlgpu_modern_graphics.md (SMG-0041 and its classic-debt backlog) records an abort with "free(): double free detected in tcache 2", identical on the unmodified renderer sources, and calls it a real bug rather than test debt; that cause is also published on its own as CNA-BUG-253. SdlGpu_Parity_dual_texture_terms: PSG-0009 only suggests it may be the XNA opaque-black null-texture rule. The other 21 have no recorded cause.

Source locations

• plans/plan_pre_sdlgpu_closeout.md — PSG-0009: the 27 failing classic SDL_GPU CTests by name; PSG-0010: the configuration they belong to
• plans/plan_street_sdlgpu.md — STREETS-0007 and STREETS-0008: the same set re-run, 26 failures remain
• plans/plan_sdlgpu_modern_graphics.md — SMG-0042 and earlier re-runs: the 27 failures identical by name
• plans/plan_sdlgpu.md — SDLGPU-11: SdlGpu_SwapchainRecovery recorded passing when it was added
• modules/renderers/sdl-gpu/examples/CMakeLists.txt — registrations of the SdlGpu_* example, parity and EasyGL-oracle CTests
• modules/graphics/tests/Microsoft/Xna/Framework/Graphics/SdlGpuIndexedDrawRangeTests.cpp — the two failing SdlGpuIndexedDrawRangeTest cases
• modules/core/include/CNA/GraphicsBackendMaturity.hpp — SDL_GPU declared Supported

Evidence

CNA's own record executed the set, on one machine: tools/platform/run_gpu_tests_private.sh (private Weston and rootful Xwayland), an AMD Radeon 780M under RADV Mesa 25.0.7, SDL_gpu on its Vulkan driver, a Release build with CNA_SDL_GPU_SHADERCROSS=OFF and libshaderc present (PSG-0010). The record calls them "Real renderer results, none of them a dead test", and STREETS-0008 re-ran all 27 on a stashed baseline to show they predate its change. No per-test output or cause is kept in the tree, except PSG-0009's remark that SdlGpu_Parity_dual_texture_terms may be the XNA opaque-black null-texture rule that WebGPU's twin failed until WMG-0025. SdlGpu_SwapchainRecovery passed 5/5 when SDLGPU-11 added it, so at least that one regressed later. Nothing was re-run for this entry: whether each failure is a renderer defect or a stale expectation is not established, which is why this entry is one cluster rather than 24 diagnoses. No CI workflow runs any SDL_GPU test.

A record from twelve days before that baseline points the other way: the closeout row of plans/plan_sdlgpu.md (2026-09-11, the sweep after SDLGPU-133) reports all 194 registered native SDL_GPU integration CTests passing sequentially with SDL_VIDEODRIVER=offscreen and no display, while the classic baseline above (PSG-0009, 2026-09-23) was taken under a private Weston and Xwayland. The two records count different sets (194 registrations against 229 tests matching ^SdlGpu) and neither says whether the difference is a regression after 11 September or the display route, so this entry reports the failures as recorded and does not attribute them.

Independent re-verification: CNA's own record executed the set, on one machine: run_gpu_tests_private.sh (private Weston and rootful Xwayland), an AMD Radeon 780M under RADV Mesa 25.0.7, SDL_gpu on its Vulkan driver, a Release build with CNA_SDL_GPU_SHADERCROSS=OFF and libshaderc present (PSG-0010). The record calls them "Real renderer results, none of them a dead test", and STREETS-0008 re-ran all 27 on a stashed baseline to show they predate its change. Nothing was re-run for this entry: the evidence is recorded by CNA, not reproduced here, and whether each failure is a renderer defect, a stale expectation or specific to that machine is not established. Earlier records were greener: remediation/REMEDIATION_INDEX.md (REMED-GFX-180, 2026-07-31) reports a fully green 102-test SDL_GPU shard, and SdlGpu_SwapchainRecovery passed 5/5 when SDLGPU-11 added it, so some of these are later regressions or environment-specific. No CI workflow runs any SDL_GPU test.

Focused reproduction

# The command CNA's own record uses; not executed for this entry.
tools/platform/run_gpu_tests_private.sh cmake-build-sdlgpu -R '^SdlGpu' -j4

Current tests

The failing CTests themselves, registered in modules/renderers/sdl-gpu/examples/CMakeLists.txt and SdlGpuIndexedDrawRangeTests.cpp. They run only where someone runs the classic set on a GPU host; no workflow names SDL_GPU.

Regression test

The tests exist; what is missing is triage. Record each failure's cause in the SDL_GPU plan, fix the renderer or the stale expectation so the CTest turns green, and put the classic ^SdlGpu set behind a recorded gate (a GPU-capable CI lane or a release checklist) so the count cannot silently stay at 26.

Blast radius

SDL_GPU only. The areas these tests pin (texture and cube-map formats and their fallbacks, compressed cubes, cube render targets with MSAA and mips, RenderTarget2D SetData, filtering and mips, render state, depth bias, deferred scissor, line topology, two-sided stencil, draw-range validation, dual-texture terms, indexed draw ranges, back-buffer formats, constructor exception safety and swapchain recovery) may diverge from the tested contract. Which of them affect a real game is unknown until each is triaged. Other renderers are unaffected.

Workaround

No workaround is known.