Rendering Inaccuracy occuring on RDNA 3 Cards using Vulkan in Ryujinx (Xenoblade 2)

[Samsterr]

As per the image below this has been an issue which i'm unsure is related to #53 . The textures tend to wobble which seems like some inaccuracy in the way that it is rendered. Textures closer to or in the cloud itself seem to exhibit this effect.
Do let me know if you require any more info such a video and what not. Both Yuzu and Ryujinx exhibit this issue

Config: 9800X3D and 7900XTX, Xenoblade 2, 2.1.0 using Greemdev's 1.3.1 Ryujinx Version. 1440P

[pperucki-AMD]

It’s unclear to me what exactly is incorrect in the attached screenshot. Could you point it out with an arrow so I know where to focus? If feasible, please also share a video that demonstrates the issue more clearly.

[Samsterr]

apologies, i have attached a video of the issue. The textures near the clouds seem to flicker

2025-09-02.19-41-08.mp4

[danaYatsuta]

This exact issue also occurs on RDNA 2 (RX 6900 XT in my case) on Ryujinx 1.3.2 and Eden 0.0.3. Former yuzu dev discussed this issue here, in case this is of any help. The thread links progress report from yuzu on deactivated website, but it can be read here.

Edit: this actually occurs when setting renderer in Ryujinx/Eden to OpenGL too, at least for me on linux.

[JPikachu]

Eden 0.0.3, can confirm this happens on RDNA 2 Linux (Steam Deck)

I will look back on the info and post it here later but out the top of my head one thing I remember is this is partially in some way due to how Nintendo used a form of ray tracing like effects exclusive to Nvidia GPUs to make the clouds.

[JPikachu]

Here is a video of it

Video_2025-08-27_21-31-38.mp4

[pperucki-AMD]

We determined this is a hardware limitation due to missing support for D24 depth formats. A future driver release is expected to include a workaround, but we cannot guarantee it will fully resolve the issue.

[Shinmegumi]

Part of our testing to try and figure out what is happening is how AMD vs NVIDIA handles float values on these textures. In Linux, if lavapipe is used, the clouds render correctly. When using RADV, this issue occurs. It also occurs on Windows drivers. It looks like the driver is defaulting to double or something else to calculate the vector coordinates and it leads to the math being different between AMD and NVIDIA.

I can get with others who have helped troubleshoot the issue to see if we can provide more information on what we found.

[majosamaso]

I would like add my observations I observed while inspecting a frame under RenderDoc

The frame intends renders the sea of clouds which is the part under the horizon in the background:

For that, the frame prepares a mesh with very large coordinates that ends up rendering the bottom part

Then the fragment shader gets the depth buffer to decide which part draw or not

Right now I don't have captures of the fragment that uses the texture but clearly is involved in the dealignment present on the image proofed by adding in units in the x axis in the debugger.

The UV of that texture is computed from the coordinates of in_attr2 generated by vertex shader

After touching a bit these coordinates, I end up with the conclusion that there is a precision problem because small increments on the coords given make non linear steps in the alignment of the drawn portion

Finally, I would like to expose a weird behaviour of the varyings when are declared as flat dvec4