Repository navigation
Windows (AMD RX 6600 XT): with VK_LAYER_KHRONOS_validation requested, test processes fail at vkCreateInstance with -9, so the required validation runs are impossible #10663
Description
What happens
With VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation (Vulkan SDK 1.4.363), AGC tests that need a device fail at "vkCreateInstance: Vulkan result -9" (VK_ERROR_INCOMPATIBLE_DRIVER); without the variable they pass, and vulkaninfo creates an instance with the same layer.
Title, test or build step
ctest: agc_driver_image_table, agc_driver_sample_order (and other Vulkan tests); guest_video_out_wqhd_detection is skipped instead
AnyPS5 commit
main b84322d, no pull requests on top
System
Windows 11 Pro 10.0.26300, AMD Radeon RX 6600 XT (Navi 23, 0x73ff), AMD driver 32.0.21045.16010, Vulkan SDK 1.4.363.0 (System32 vulkan-1.dll 1.4.363.0), WinLibs GCC 15.2.0 r7
Steps to reproduce
$env:VK_INSTANCE_LAYERS = "VK_LAYER_KHRONOS_validation"
ctest --test-dir build -R "^agc_driver_image_table$" --output-on-failure # fails in 0.02 s
Remove-Item Env:VK_INSTANCE_LAYERS
ctest --test-dir build -R "^agc_driver_image_table$" # passes
Log
[src/VulkanDevice.cpp:1105 VulkanDevice] VulkanDevice constructor window=0000000000000000
[src/VulkanDevice.cpp:1114 VulkanDevice] Vulkan loader loaded library=00007ff90d620000
vkCreateInstance: Vulkan result -9
1/1 Test #781: agc_driver_image_table ...........***Failed 0.02 sec
What was checked
With the layer requested: agc_driver_sample_order and agc_driver_image_table fail in under 0.1 s, guest_video_out_wqhd_detection is skipped ("no display or Vulkan device"); agc_driver_graphics and agc_shader_memory pass (they also pass with VK_DRIVER_FILES pointing at a missing file, so they don't need a device). Without the variable all of them pass.
vulkaninfo --summary with the same VK_INSTANCE_LAYERS lists the GPU, and the layer creates the file named by VK_KHRONOS_VALIDATION_LOG_FILENAME; the test processes create no such file. VK_LOADER_DEBUG=layer prints loader messages for vulkaninfo but nothing for the test processes, so the loader's reason for -9 is not visible here.
Neither log contains "was not found but was requested by env var VK_INSTANCE_LAYERS", so the check in AGENTS.md / CONTRIBUTING.md does not catch this, and the "no new VUID- errors" comparison compares two logs that cannot contain any.
In #9056 another tester reported validation output on Windows with a Radeon 680M (driver 32.0.21030) and layer 1.4.363, so this may depend on the driver or the machine.
AI-assisted: yes
Activity
MarcZX10 commentedon Oct 24, 2026
A useful next discriminator is the actual loader DLL path used by the test process.
VulkanDeviceopensvulkan-1.dllby bare name throughSDL_LoadObject(core/libs/prx/libSceAgcDriver/Execution/src/VulkanDevice.cpp:1107), whereasvulkaninfomay resolve a different copy. Could you compare the loaded module path and file version/hash foragc_driver_image_tableversusvulkaninfo, and check for a build-directory-localvulkan-1.dllorVK_DRIVER_FILES/VK_ICD_FILENAMESoverride? The absence of loader debug output in the test process makes a different loader or process environment worth ruling out before attributing this to the RX 6600 XT driver.brummilab commentedon Oct 24, 2026
False alarm, my mistake, sorry for the noise. @MarcZX10 your
VK_DRIVER_FILESquestion found it:VK_DRIVER_FILESpointing at a missing file was still set in my shell from an earlier probe, so the loader found no driver and returned -9. With it cleared, on the samemainb84322da6,agc_driver_image_tablepasses withVK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation, the layer writes the file named byVK_KHRONOS_VALIDATION_LOG_FILENAME, and it reportsVUID-VkWriteDescriptorSet-descriptorType-00337x2 (the one #9056 addresses). Closing.AI-assisted: yes