Bug 266882 – Web Inspector: No errors reported in console when unimplemented WGSL radians() function is used in compute shader

• Status: NEW
• Component: Web Inspector

[Umar]

Created attachment 469200 [details]
Screenshot showing Safari TP running with the unpatched version of my code

I was testing out my project (https://github.com/umar-ahmed/webgpu-pathtracer/) on Safari Technical Preview 185 with the WebGPU feature flag enabled. However, when I loaded it up, I was met with a black screen and no errors in the console.

Using `git bisect`, I was able to narrow down the issue to the use of the radians() function in my WGSL compute shader, which doesn't seem to be supported on this build of Safari (but is supported in Chrome Version 121.0.6167.16 (Official Build) beta (arm64), which I was using to develop). This was the patch that I applied to resolve the issue: https://github.com/umar-ahmed/webgpu-pathtracer/commit/8fdfe37cbb3c38e856feba19275eb023a70b2ff7.

I'm filing this bug as a Web Inspector bug instead of a WebGPU bug because I'm not sure if radians() is supposed to be implemented yet in Safari. If it is, then this bug can be recategorized. But my immediate concern was that I got no compiler/runtime warning when the shader was executed.

[Radar WebKit Bug Importer]

<rdar://problem/120138235>

[Mike Wyrzykowski]

My reading of the specification per https://gpuweb.github.io/gpuweb/#dom-gpudevice-createshadermodule is that getCompilationInfo should return any error messages, see https://gpuweb.github.io/gpuweb/#dom-gpushadermodule-getcompilationinfo as opposed to reporting warnings / errors in alternative means


Does shaderModule.getCompilationInfo() return anything useful when you use radians here? My understanding is shader compilation errors should be reported via that mechanism

[Mike Wyrzykowski]

MSL compilation error: Error Domain=MTLLibraryErrorDomain Code=3 "program_source:181:26: error: use of undeclared identifier 'radians'
    float local46 = tan((radians(parameter18.field2) / 2.));
                         ^
program_source:578:55: error: use of undeclared identifier 'radians'
        float local88 = (global7.field6.field4 / tan((radians(global7.field6.field2) * 0.5)));
                                                      ^
" UserInfo={NSLocalizedDescription=program_source:181:26: error: use of undeclared identifier 'radians'
    float local46 = tan((radians(parameter18.field2) / 2.));
                         ^
program_source:578:55: error: use of undeclared identifier 'radians'
        float local88 = (global7.field6.field4 / tan((radians(global7.field6.field2) * 0.5)));

[Mike Wyrzykowski]

shaderModule.getCompilationInfo won't return anything useful here because we think radians is a MSL function but it is not

[Mike Wyrzykowski]

Created attachment 469220 [details]
WGSL source which generates the error

Attaching source which generates the MSL error

[Umar]

(In reply to Mike Wyrzykowski from comment #2)
> My reading of the specification per
> https://gpuweb.github.io/gpuweb/#dom-gpudevice-createshadermodule is that
> getCompilationInfo should return any error messages, see
> https://gpuweb.github.io/gpuweb/#dom-gpushadermodule-getcompilationinfo as
> opposed to reporting warnings / errors in alternative means
> 
> 
> Does shaderModule.getCompilationInfo() return anything useful when you use
> radians here? My understanding is shader compilation errors should be
> reported via that mechanism

I'm not seeing any messages when I look at the GPUCompilationInfo object resolved from that method.

If I use some other undefined function like `someRandomFunction`, it will output GPUCompilationMessage with `message = "unresolved call target 'someRandomFunction'"` as expected, but not for `radians()`.

[Mike Wyrzykowski]

My apologies, I commented above but shaderModule.getCompilationInfo won't return anything useful here because we assumed radians is a MSL function but it is not.

This will be corrected in the near future.

[Tadeu Zagallo]

I'm fixing this as part of https://bugs.webkit.org/show_bug.cgi?id=266931. The reason I created a new bug is that it should fix all the numeric built-in functions, not just radians.