Examples: macOS / MoltenVK issues

[selimsandal]

1. BindlessSceneViewer

• Crashes with error below when GPU draw call generation is unchecked, even when gpu is not really out of memory

[mvk-error] VK_ERROR_OUT_OF_DEVICE_MEMORY: Lost VkDevice after MTLCommandBuffer "vkQueueSubmit MTLCommandBuffer on Queue 0-0" execution failed (code 3): Caused GPU Address Fault Error (0000000b:kIOGPUCommandBufferCallbackErrorPageFault)
ERROR (SwapChainVK.hpp:356) - VK::Apple M2 Max - AcquireNextTexture(): vkAcquireNextImageKHR returned -4

---

2. Readback

• Half of the tooltip is not showing, was fine before (MoltenVK related?)

---

3. Wrapper

MoltenVK did not release a new version yet so Vulkan SDK still does not have an implementation for vkGetDeviceBufferMemoryRequirements. Also if we can determine a fixed location for libvulkan.dylib on macOS that would be better instead of manually pointing to VulkanSDK path, maybe we can use VulkanSDK env vars?

[selimsandal]

I'm looking at BindlessSceneViewer to try to fix it. But I don't really know how to debug Readback. @dzhdanNV any suggestions?

[dzhdanNV]

Thanks. Readback debugging ideas:

• try increasing queued frames in Streamer:

    streamerDesc.queuedFrameNum = GetQueuedFrameNum() => 5

• try enabling VSYNC:

swapChainDesc.verticalSyncInterval = m_VsyncInterval => 1

[dzhdanNV]

Wrapper: there are 2 issues:

• explain more the problem with vkGetDeviceBufferMemoryRequirements. It seems to be not used
• I got the idea with ENV var. I will try to implement tomorrow

[dzhdanNV]

Also if we can determine a fixed location for libvulkan.dylib on macOS that would be better instead of manually pointing to VulkanSDK path, maybe we can use VulkanSDK env vars?

VULKAN_SDK env var is the only one and it's useless. Vulkan dynamic library must be searchable in system folders automatically, no need to specify the path. It must work on all systems. It's important to use proper name with 1. I committed some polishing changes to address this.

[dzhdanNV]

Crashes with error below when GPU draw call generation is unchecked, even when gpu is not really out of memory

The key phrase is when GPU draw call generation is unchecked. It means:

• when GPU drawcall gen is enabled, the drawcall count is 0 (why?), so nothing is rendered, so "all good", but the screen is empty
• when GPU drawcall gen is disabled, drawing uses non-indirect drawcalls, what renders something, a shader tries to use bindless resources and crashes (why?)

I can't answer these questions. Maybe something is wrong with MVK. Maybe there is a VK misuse on my side (but VK validation is clean).

[selimsandal]

Also if we can determine a fixed location for libvulkan.dylib on macOS that would be better instead of manually pointing to VulkanSDK path, maybe we can use VulkanSDK env vars?

VULKAN_SDK env var is the only one and it's useless. Vulkan dynamic library must be searchable in system folders automatically, no need to specify the path. It must work on all systems. It's important to use proper name with 1. I committed some polishing changes to address this.

When Vulkan SDK is not installed globally I think searching inside system is not possible, by default it does not do a global install.

[selimsandal]

Crashes with error below when GPU draw call generation is unchecked, even when gpu is not really out of memory

The key phrase is when GPU draw call generation is unchecked. It means:

• when GPU drawcall gen is enabled, the drawcall count is 0 (why?), so nothing is rendered, so "all good", but the screen is empty
• when GPU drawcall gen is disabled, drawing uses non-indirect drawcalls, what renders something, a shader tries to use bindless resources and crashes (why?)

I can't answer these questions. Maybe something is wrong with MVK. Maybe there is a VK misuse on my side (but VK validation is clean).

I tried debugging this, scene gets loaded, buffers get loaded but drawcall is always 0. I'll try with very simple shaders and find where the problem starts occurring.

[dzhdanNV]

When Vulkan SDK is not installed globally I think searching inside system is not possible, by default it does not do a global install.

That's true. But this is not what we should fix.

[dzhdanNV]

@vertver has debugged the issue. Thank you very much! I committed a few fixes based on his feedback:

• Metal doesn't support draw indirect with a count buffer (or at least not every Apple device) - added a check for it
• Metal validation detected a texture in an unbound texture array with format RGBA8_UINT which could be implicitly converted to float - actually this texture can't not be accessed in a shader, added a workaround for this (it's ugly but works for now)

But fixes are not enough: there is a bug on MoltenVK side. Currently it can be avoided by a setting a special ENV var.

Please, try the latest.

[vertver]

Addition to dzhdan reply:

I disabled the Argument Buffer because it doesn't work properly on <M3 hardware. However, these changes mean that the resource limits are significantly stricter: 160 bound resources and 128 bound textures. Also, there is still issue with compute shader (validation doesn't give any specific details about GPU crash)