Implement bindless buffers on metal by mate-h · Pull Request #9081 · gfx-rs/wgpu

mate-h · 2026-02-20T08:11:09Z

Connections
Resolves #6741

Description
Problem: Metal on macOS did not support bindless storage buffers.

In wgpu-hal, encode buffer binding arrays using MTLArgumentEncoder. Metal expects encoded pointers, not raw resource IDs. Creates an argument buffer from an encoder, encode each buffer, and register it in resources_to_use.

In naga MSL writer, emit device T* for buffer binding array elements (which are pointers to device memory) instead of the constant layout used for textures and samplers. For struct members behind binding arrays, emit -> instead of . because the elements are pointers.

In the adapter, expose the BUFFER_BINDING_ARRAY feature and raise max_buffers_per_stage on supporting devices, with special handling for MTLCaptureDevice. This is a separate fix for Xcode frame capture, when some functions fail to invoke, in these cases fallback to feature table based limits. I referenced the metal feature table PDF.

Testing

Tested with Bevy Solari on metal / macOS which makes use of bindless storage buffers and raytracing acceleration structures.

Squash

Checklist

Run cargo fmt.
Run taplo format.
Run cargo clippy --tests. If applicable, add:
- --target wasm32-unknown-unknown
Run cargo xtask test to run tests.
If this contains user-facing changes, add a CHANGELOG.md entry.

inner-daemons · 2026-02-20T19:14:25Z

I'm excited to see this and plan on checking out the code. Thanks a ton for working on it!

Vecvec

A couple of things from a read through.

wgpu-hal/src/metal/adapter.rs

inner-daemons

Ok so lots of thoughts here, mostly questions. Overall looks good. More comments would be great.

Big thing though is snapshot tests. I can't really review this change until I know what it generates and know that what it generates is correct. There might even be existing tests that you can just add METAL as a target to.

Thanks for putting in the work. I look forward to using this myself, and I'd like to get this landed soon!

inner-daemons · 2026-02-26T21:03:30Z

CHANGELOG.md


 - Use autogenerated `objc2` bindings internally, which should resolve a lot of leaks and unsoundness. By @madsmtm in [#5641](https://github.com/gfx-rs/wgpu/pull/5641).
 - Implements ray-tracing acceleration structures for metal backend. By @lichtso in [#8071](https://github.com/gfx-rs/wgpu/pull/8071).
+- Added support for bindless storage buffers (buffer binding arrays) on Metal. By @mate-h in [#9081](https://github.com/gfx-rs/wgpu/pull/9081).


Tiny nit but maybe put this under "new features" rather than "changes"?

inner-daemons · 2026-02-26T21:14:29Z

naga/src/back/msl/writer.rs

+                    crate::TypeInner::Struct { .. } => {
+                        write!(
+                            out,
+                            "device {ARGUMENT_BUFFER_WRAPPER_STRUCT}<device {base_tyname}*>*"


Can you add a comment explaining this special case

inner-daemons · 2026-02-26T21:15:12Z

naga/src/proc/mod.rs

                crate::Expression::FunctionArgument(_) => return None,
-                // There are no other expressions that produce pointer values.
-                _ => unreachable!(),
+                _ => return None,


Why this change?

inner-daemons · 2026-02-26T21:18:38Z

wgpu-hal/src/metal/adapter.rs

            {
-                Some(device.hasUnifiedMemory())
+                // On Xcode frame capture this call causes a crash, fall back to feature table.
+                if Self::is_capture_mtl_device(device) {


Technically this might belong in another PR but its fine here IMO. Anyway, why Apple6 specifically? And tbh, it'd be fine to just report None here if we don't know for sure

inner-daemons · 2026-02-26T21:31:52Z

wgpu-hal/src/metal/adapter.rs

        );
+        features.set(
+            F::BUFFER_BINDING_ARRAY,
+            self.msl_version >= MTLLanguageVersion::Version3_0


I can't find any reason why this shouldn't be 2.0? Though I guess you copied the above so that should also be changed.

Also, I believe this might impact Features::PARTIALLY_BOUND_BINDING_ARRAY?

inner-daemons · 2026-02-26T21:35:09Z

wgpu-hal/src/metal/conv.rs

+            wgt::BufferBindingType::Uniform => MTLResourceUsage::Read,
+            wgt::BufferBindingType::Storage { read_only } => {
+                if *read_only {
+                    MTLResourceUsage::Read
+                } else {
+                    MTLResourceUsage::Read | MTLResourceUsage::Write
+                }
+            }


Suggested change

wgt::BufferBindingType::Uniform => MTLResourceUsage::Read,

wgt::BufferBindingType::Storage { read_only } => {

if *read_only {

MTLResourceUsage::Read

} else {

MTLResourceUsage::Read | MTLResourceUsage::Write

}

}

wgt::BufferBindingType::Uniform | wgt::BufferBindingType::Storage { read_only: true }=> MTLResourceUsage::Read,

wgt::BufferBindingType::Storage { read_only: false }=> MTLResourceUsage::Write,

inner-daemons · 2026-02-26T21:36:08Z

wgpu-hal/src/metal/conv.rs

+        wgt::BufferBindingType::Uniform => MTLBindingAccess::ReadOnly,
+        wgt::BufferBindingType::Storage { read_only } => {
+            if *read_only {
+                MTLBindingAccess::ReadOnly
+            } else {
+                MTLBindingAccess::ReadWrite
+            }
+        }


Suggested change

wgt::BufferBindingType::Uniform => MTLBindingAccess::ReadOnly,

wgt::BufferBindingType::Storage { read_only } => {

if *read_only {

MTLBindingAccess::ReadOnly

} else {

MTLBindingAccess::ReadWrite

}

}

wgt::BufferBindingType::Uniform | wgt::BufferBindingType::Storage { read_only: true } => MTLBindingAccess::ReadOnly,

wgt::BufferBindingType::Storage { read_only: false } => MTLBindingAccess::ReadWrite,

inner-daemons · 2026-02-26T21:38:43Z

naga/src/back/msl/writer.rs

We will definitely need some snapshot tests here. You can read through naga/tests/in/wgsl to get a good idea of how to create these.

inner-daemons · 2026-02-26T21:40:24Z

wgpu-hal/src/metal/device.rs

+                                let buffers = &desc.buffers[entry.resource_index as usize..]
+                                    [..count as usize];


Suggested change

let buffers = &desc.buffers[entry.resource_index as usize..]

[..count as usize];

let buffers = &desc.buffers[entry.resource_index as usize..(entry.resource_index + count) as usize];

inner-daemons · 2026-02-26T21:44:51Z

wgpu-hal/src/metal/device.rs

+                                let argument_desc = conv::pointer_array_argument_descriptor(
+                                    count,
+                                    conv::map_binding_access(&ty),
+                                );
+
+                                let encoder = device
+                                    .newArgumentEncoderWithArguments(&NSArray::from_retained_slice(
+                                        &[argument_desc],
+                                    ))
+                                    .unwrap();
+                                let aligned_length = wgt::math::align_to(
+                                    encoder.encodedLength(),
+                                    encoder.alignment(),
+                                );
+                                argument_buffer = device
+                                    .newBufferWithLength_options(
+                                        aligned_length,
+                                        MTLResourceOptions::HazardTrackingModeUntracked
+                                            | MTLResourceOptions::StorageModeShared,
+                                    )
+                                    .unwrap();


This entire section of code (and everything aroudn it) needs more comments. I don't fully understand why we're creating buffers, not putting them anywhere, etc.

mate-h added 3 commits February 20, 2026 00:01

Implement bindless buffers on metal

a1439aa

Fix CI and changelog

1aafa4a

Update changelog

96ed0d9

inner-daemons self-assigned this Feb 20, 2026

Vecvec reviewed Feb 20, 2026

View reviewed changes

wgpu-hal/src/metal/adapter.rs Show resolved Hide resolved

wgpu-hal/src/metal/adapter.rs Outdated Show resolved Hide resolved

Address feedback

84c28e6

inner-daemons suggested changes Feb 26, 2026

View reviewed changes

inner-daemons removed their assignment Feb 27, 2026

		let buffers = &desc.buffers[entry.resource_index as usize..]
		[..count as usize];

Conversation

mate-h commented Feb 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

inner-daemons commented Feb 20, 2026

Uh oh!

Vecvec left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

inner-daemons left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

mate-h commented Feb 20, 2026 •

edited

Loading

inner-daemons left a comment •

edited

Loading