Bug 315092 – Some WebGPU content not rendering in Safari

• Component: WebGPU
• Status: RESOLVED FIXED

[Ashley Gullen]

We make Construct, a browser-based game development tool (www.construct.net). Based on a customer report we have found some content that appears to render incorrectly when using WebGPU in Safari only. The same WebGPU content renders correctly in Chrome on Windows and macOS, Firefox Nightly on Windows and macOS, and Safari when rendering with WebGL, so we suspect a Safari WebGPU bug.

FWIW, some of the content that does not render uses a repeat texture mode for a tiling effect, but I don't know if that's relevant to the issue.

WebGL content: https://downloads.scirra.com/labs/bugs/ioswebgpu/webgl/

WebGPU content: https://downloads.scirra.com/labs/bugs/ioswebgpu/webgpu/

Observed result:

Some content missing from WebGPU rendering in Safari on macOS only. This also reproduces in TP 243.

Expected result:

Safari on macOS to render WebGPU content the same as WebGL or Chrome/Firefox WebGPU result.

[Radar WebKit Bug Importer]

<rdar://problem/177443835>

[Mike Wyrzykowski]

Created attachment 479824 [details]
stencilCompareAlways

Disabling stencil test, i.e., setting .stencilCompareFunction = MTLCompareFunctionAlways works around the issue

[Mike Wyrzykowski]

Oh likely the bug is this.

The spec for setStencilReference https://www.w3.org/TR/webgpu/#dom-gpurenderpassencoder-setstencilreference
setStencilReference(reference)
Sets the [[stencilReference]] value used during stencil tests with the "replace" GPUStencilOperation.

But this site is not using replace, it's using equal. So we should ignore the value, 1, getting passed to GPURenderPassEncoder.setStencilReference

[Mike Wyrzykowski]

This fixes the issue:

 % git diff ./Source/WebGPU/WebGPU/RenderPassEncoder.mm
diff --git a/Source/WebGPU/WebGPU/RenderPassEncoder.mm b/Source/WebGPU/WebGPU/RenderPassEncoder.mm
index 8ed3d2fb6bcf..ac7e59406b48 100644
--- a/Source/WebGPU/WebGPU/RenderPassEncoder.mm
+++ b/Source/WebGPU/WebGPU/RenderPassEncoder.mm
@@ -562,7 +562,7 @@ void RenderPassEncoder::setCachedRenderPassState(id<MTLRenderCommandEncoder> com
     }
     if (m_blendColor)
         [commandEncoder setBlendColorRed:m_blendColor->r green:m_blendColor->g blue:m_blendColor->b alpha:m_blendColor->a];
-    if (m_stencilReferenceValue) {
+    if (m_stencilReferenceValue && m_pipeline && m_pipeline->depthStencilDescriptor().frontFaceStencil.stencilFailureOperation == MTLStencilOperationReplace) {
         [commandEncoder setStencilReferenceValue:*m_stencilReferenceValue];
         m_stencilReferenceValue = std::nullopt;
     }

I just need to see if I should check the back stencil face too and what to do when both front and back due not have replace.

[Mike Wyrzykowski]

Pull request: https://github.com/WebKit/WebKit/pull/65650

[Mike Wyrzykowski]

So the root cause is actually:

    The spec says to treat undefined loadOps with read only storage as loads,
    but we were defaulting to clear which is wrong.

landing PR shortly

[EWS]

Committed 313943@main (442e2c0e0168): <https://commits.webkit.org/313943@main>

Reviewed commits have been landed. Closing PR #65650 and removing active labels.