Incorrect max_push_constant_size in the docs

[Bromles]

The docs for the Limits struct say to expect max_push_contant_size to be 256 bytes on the DX12 backend.

But I encountered a 128 bytes limit on a relatively modern iGPU in an AMD Ryzen 7 PRO 4750U CPU on Windows 11

I propose changing this line in the docs to be identical to the Vulkan backend - 128-256 bytes

Since it's a really tiny change, I've done so in my fork and can open a PR if you'd like to accept it
https://github.com/Bromles/wgpu/blob/trunk/wgpu-types/src/lib.rs#L679

I will attach the full wgpu-info -vv logs from my machine, Adapter 1 is the DX12 hardware with the limits mentioned above (not to be confused with Adapter 2, which is the software DX12 implementation that also has only a 128-byte push constant size limit)

wgpu-info-win11.log

[inner-daemons]

See this comment in wgpu-hal/src/dx12/adapter.rs:

// The push constants are part of the root signature which
// has a limit of 64 DWORDS (256 bytes), but other resources
// also share the root signature:
//
// - push constants consume a `DWORD` for each `4 bytes` of data
// - If a bind group has buffers it will consume a `DWORD`
//   for the descriptor table
// - If a bind group has samplers it will consume a `DWORD`
//   for the descriptor table
// - Each dynamic uniform buffer will consume `2 DWORDs` for the
//   root descriptor
// - Each dynamic storage buffer will consume `1 DWORD` for a
//   root constant representing the dynamic offset
// - The special constants buffer count as constants
//
// Since we can't know beforehand all root signatures that
// will be created, the max size to be used for push
// constants needs to be set to a reasonable number instead.
//
// Source: https://learn.microsoft.com/en-us/windows/win32/direct3d12/root-signature-limits#memory-limits-and-costs
max_push_constant_size: 128,

So, the value of 128 is intentional.

I'm closing this PR, if you have other questions or concerns feel free to reopen it.

[Bromles]

I would argue that this comment doesn't solve the issue, because:

1. It's in the wgpu-hal, not wgpu. The general developer wouldn't know to go look here for that comment, since only wgpu is considered a user-facing API, as far as I know
2. The same general developer would see the mentioned 256-byte limit in the wgpu docs and assume it to be the case, i.e., rely on it in shaders. When in reality the limit will be lower

So, I'm not saying that the 128-byte limit is a problem, if it's better to set it to this particular value - no issues here.

I'm highlighting that the docs saying to "expect size to be 256 bytes on DX12" are an issue, because it's misleading when on a lower level, under the hood, in the HAL, we are setting it to be 128 bytes ourselves.

[Bromles]

Additionally, it seems that the adjacent docs entries on the same struct are mixed up as well

It says that the higher min_subgroup_size is better, and lower max_subgroup_size is better, which is incorrect

[cwfitzgerald]

The upstream spec is settling on a min spec of 64 bytes of push constant (immediate) space, so we should follow along with that in the documentation.