`maxInterStageShaderVariables` should be decremented for _each_ fragment input built-in, not just once

[ErichDonGubler]

In this following inter-stage interface validation routine in the WebGPU spec. (link, emphasis mine):

validating inter-stage interfaces(device, descriptor)

…

Device timeline steps:

…

7. If descriptor.fragment is provided:

   1. Let maxFragmentShaderInputVariables be device.limits.maxInterStageShaderVariables.

   2. For each of the Inter-Stage Builtins that are an input of descriptor.fragment:

      1. Decrement maxFragmentShaderInputVariables by 1.

I interpret this to mean: for each built-in, decrement once per built-in found. However, the webgpu:api,validation,capability_checks,limits,maxInterStageShaderVariables:createRenderPipeline,at_over:* test path in CTS only decrements the maximum once if any of them are set:

cts/src/webgpu/api/validation/capability_checks/limits/maxInterStageShaderVariables.spec.ts

Lines 19 to 21 in d5a131f

	const maxFragmentInputVariables =
	device.limits.maxInterStageShaderVariables -
	(frontFacing || sampleIndex || sampleMaskIn ? 1 : 0);

Which interpretation is right here? EDIT: Confirmed, it should be a deduction per built-in.

[jimblandy]

This seems to be relevant background: gpuweb/gpuweb#2857

[jimblandy]

This comment seems to suggest that the spec means to count each builtin, so the CTS is wrong:

The reference of Vulkan SPEC can be found here:

The number of input and output locations available for a shader input or output interface are limited, and dependent on the shader stage as described in Shader Input and Output Locations.

All variables in both the built-in interface block and the user-defined variable interface count against these limits.

Each effective Location must have a value less than the number of locations available for the given interface, as specified in the "Locations Available" column in Shader Input and Output Locations.

[jimblandy]

Hmm, now I'm not so sure. In #3945, it was Jiawei who changed the CTS to behave that way - but it is also Jiawei who wrote the comment I linked to above. So I'm really lost.

cc: @Jiawei-Shao

[jimblandy]

Ah hah! I think the spec is right, and the CTS wrong.

In gpuweb/gpuweb#5403 (2025-10-21) the spec language was changed from:

    1. If any of the `front_facing`, `sample_index`, or `sample_mask` [=builtins=] are an input of ...

to:

   1. For each of the [=Inter-Stage Builtins=] that are an input of

The old text corresponds to what Jiawei made the CTS do in #3945: if any of the given builtins are used, decrement the max by one, regardless of how many are used. gpuweb/gpuweb#5403 is clearly changing the meaning.

I think that change in meaning was deliberate: gpuweb/gpuweb#5403 fixes gpuweb/gpuweb#5337, in which @teoxoy commented:

At least for Vulkan all builtins count towards the maxFragmentInputComponents limit.

The number of input and output locations available for a shader input or output interface depend on the shader stage as described in Shader Input and Output Locations. All variables in both the built-in interface block and the user-defined variable interface count against these limits.

from section 16.1.4. Location and Component Assignment

[ErichDonGubler]

Aha, @jimblandy did the homework I was going to do today. 😀 Thanks for the assist! Filed #4539 to fix this.

[dneto0]

I agree with the assessment. Thanks for catching this! I think we've seen this in Chrome runs but nobody has dug in on these. (chrome bugs linked in PR)

[kainino0x]

Does this cover all the changes in gpuweb/gpuweb#5403 ? If so the needs-cts-issue label should be removed there. (Also sorry I have not been doing anything with that label, I should come up with a better system for this that is less easy to ignore.)

[ErichDonGubler]

Does this cover all the changes in gpuweb/gpuweb#5403 ? If so the needs-cts-issue label should be removed there. (Also sorry I have not been doing anything with that label, I should come up with a better system for this that is less easy to ignore.)

The PR fixing this has none of them, ATM. I don't think it'll be hard to add; I'll take a look.

[ErichDonGubler]

@kainino0x: I filed #4540.