#1540

AMD Radeon 8060S (Windows): vk_overlapping_occlusion triggers GPU TDR and d3d12_r9g9b9e5_uav fails on RENDER_TARGET support

Opendanielbercuopened 2h ago

Environment

  • Commit: f85582e2
  • OS: Windows x64 (MSVC 19.51, Windows SDK 10.0.26100, Vulkan SDK 1.4.363.0)
  • GPU: AMD Radeon(TM) 8060S Graphics (vendorID=0x1002, deviceID=0x1586, subgroup 32..64)
  • Driver: AMD proprietary driver 26.10.1 (LLPC) (driverVersion=2.0.416, Vulkan 1.4.359)

Summary

  • scripts/run_gpu_tests.py: 133/133 passed
  • ctest in runtime/build: 2 hardware-specific failures (100% reproducible):

1. vk_overlapping_occlusion triggers GPU TDR / driver reset

The test stalls for ~24s, triggers a Windows GPU driver reset (black screen), and leaves the destination buffer untouched (0xcdcdcdcdcdcdcdcd):

at byte 32: got 14829735431805717965, wanted 80
device: AMD Radeon(TM) 8060S Graphics

2. d3d12_r9g9b9e5_uav fails on D3D12_FORMAT_SUPPORT1_RENDER_TARGET

[rt] d3d12: upload heaps in VRAM (DEVICE_LOCAL | HOST_VISIBLE, Resizable BAR)
[rt] d3d12: bindless set: 65536 2D images, 4096 samplers, 1024 2D-array, 256 3D, 256 1D, 4096 storage indices in 2D
[rt] d3d12: CheckFeatureSupport feature 3
line 399: (fs.Support1 & D3D12_FORMAT_SUPPORT1_RENDER_TARGET) == 0
device: AMD Radeon(TM) 8060S Graphics

Activity

3
AGVibecoder2h ago

Thanks for the report. There are two parts:

  1. d3d12_r9g9b9e5_uav on RENDER_TARGET support: this looks fixed by #1450 (merged), which reports only the RGB9E5 render targets the backend can actually clear. Could you retest on current main and say whether it still fails?
  2. The vk_overlapping_occlusion TDR: reproducing this needs the Radeon 8060S. Please add the commit you ran, the driver version, and whether it hangs on every run or only sometimes. A LEONIDA_VK_VALIDATION=1 run would also help, if you can do one.
hosienzade451h ago

Software Vulkan control runs and readback finding

I built the exact reported commit f85582e2 and 99f99890 on Linux with the repository’s synthetic occlusion shaders. vk_overlapping_occlusion passed on both revisions with Mesa llvmpipe 25.0.7 and 26.1.6, with synchronization validation on. All four runs returned crossing 80/20, nested 84/16 and 261 summed segments.

I found a separate test-side issue related to the sentinel observation: the destination is mapped, but its ResourceBinding did not set host_readable. A vkCmdPipelineBarrier2 trace on the first overlapping-occlusion submission showed zero transfer-to-host-read barriers before the change and one after it. #1594 marks the mapped outputs in both occlusion tests; both pass with validation on the two Mesa versions.

These control runs did not reproduce the Windows Radeon 8060S timeout. The barrier finding explains a gap in how the test makes its results visible to the CPU; it does not establish the cause of the TDR. Synthetic shaders only, AI-assisted testing.

Ryx691h ago

@danielbercu, could you run the new bisect mode from #1622 on the 8060S? It splits vk_overlapping_occlusion into its seven query patterns, one per submission, so we can see which one hangs the GPU.

git fetch https://github.com/Ryx69/Leonida.git runtime/d3d189-occlusion-tdr
git checkout FETCH_HEAD
cmake --build runtime/build --target vk_overlapping_occlusion_test
cd runtime/build
vk_overlapping_occlusion_test overlapping_occlusion.vert.spv overlapping_occlusion.frag.spv --bisect
vk_overlapping_occlusion_test overlapping_occlusion.vert.spv overlapping_occlusion.frag.spv --bisect --gated

Each phase prints running <phase> before it starts, so the last line before the hang names the pattern. If one hangs, running that phase alone (e.g. --bisect chunked) would confirm it. Both runs may reset the driver, so close anything you don't want interrupted first. Please paste the full output of both runs and your driver version, and a run with LEONIDA_VK_VALIDATION=1 too if you have the Vulkan SDK layers. Thanks!