Use pointers instead of IDs for tracing #8289

andyleiserson · 2025-10-02T01:08:07Z

This resolves the duplicate storage of Id and Arc versions of commands when tracing that was introduced by encode-on-finish. It now stores the Arc version only, and uses the pointers from the Arcs as object IDs when tracing. It also overhauls the trace player to play back traces in this format.

This fixes tracing of compute and render passes when encoded using the standard APIs. Tracing of render bundles is still broken.

Testing
TBD. Existing player tests, manual tests, and it would be nice to have an automated test of recording a trace but I don't think we do right now.

Squash or Rebase? Rebase, after squashing fixups.

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.

andyleiserson · 2025-10-08T16:23:12Z

This does have all my latest local changes, but there's still a little work to do to fix the CI failures, and I want to revisit the way that polling and user_closures are handled.

teoxoy · 2025-10-13T12:51:11Z

wgpu-core/src/device/global.rs

            // this check can't go in the body of `create_bind_group_layout` since the closure might not get called
            if let Err(e) = device.check_is_valid() {
                break 'error e.into();
            }


I think we can remove this check since create_bind_group_layout now contains it as well.

teoxoy · 2025-10-13T12:55:20Z

wgpu-core/src/resource.rs

        Ok(submit_index)
    }

+    pub fn get_mapped_range(


get_mapped_range on the global is not calling this method.

Thank you for catching this, I missed updating it to call the extracted version.

teoxoy · 2025-10-13T13:10:22Z

wgpu-core/Cargo.toml

 hashbrown.workspace = true
 indexmap.workspace = true
 log.workspace = true
+macro_rules_attribute.workspace = true


We can gate this dependency behind the serde feature.

teoxoy · 2025-10-13T13:19:41Z

wgpu-core/src/id.rs

+        // Unfortunately, because `Arc::as_ptr` returns a pointer to the
+        // contained data, and `Arc` stores reference counts before the data,
+        // we are adding an offset to the pointer here.
+        Self(Arc::as_ptr(arc) as usize, PhantomData)


Out of date comment?

I will clarify it. The issue is that Arc::as_ptr is not doing the equivalent of arc.inner as *const ArcInner<T> (where arc: &Arc<T>), it is doing &arc.inner.data as *const T, which is located at an offset of 16 from the base address of the ArcInner.

teoxoy · 2025-10-13T13:29:33Z

wgpu-core/src/command/compute_command.rs

-        #[cfg_attr(not(any(feature = "serde", feature = "replay")), allow(dead_code))]
+        //#[cfg_attr(not(any(feature = "serde", feature = "replay")), allow(dead_code))]
        color: u32,
        len: usize,
    },

    PopDebugGroup,

    InsertDebugMarker {
-        #[cfg_attr(not(any(feature = "serde", feature = "replay")), allow(dead_code))]


It seems like we kept those guards for the render pass, but not here.

teoxoy · 2025-10-13T14:01:50Z

player/src/lib.rs

+    fn resolve_buffer_id(
+        &self,
+        id: wgc::id::PointerId<wgc::id::markers::Buffer>,
+    ) -> Arc<wgc::resource::Buffer> {
+        self.buffers.get(&id).expect("invalid buffer").clone()
+    }
+
+    pub fn get_buffer(
+        &self,
+        id: wgc::id::PointerId<wgc::id::markers::Buffer>,
+    ) -> Arc<wgc::resource::Buffer> {
+        self.resolve_buffer_id(id)
+    }


Do we need both functions?

teoxoy · 2025-10-13T14:12:41Z

player/src/lib.rs

+                let external_texture = self
+                    .external_textures
+                    .remove(&id)
+                    .expect("invalid external texture");
+                external_texture.destroy();


I don't think the free functions should remove the resource from the map since the spec says it's ok to call destroy multiple times.

Yeah, the naming is confusing because a traced "free" is a destroy, and a traced "destroy" is a drop. I had fixed this for buffers and textures when tracing boat attack, but I missed external textures. I'll also add a comment about the naming mismatch.

teoxoy · 2025-10-13T14:31:06Z

I wanted to mention that the main difference I noticed between this approach and the previous is that we now effectively only record commands after successful validation. I find that this somewhat limits the usefulness of recording and replaying traces since they won't record invalid commands.

I've mostly used the record/replay functionality to debug overly restrictive or wrong validation in Firefox which will no longer be possible with this PR.

We should probably talk about this at the maintainers meeting.

teoxoy · 2025-10-15T15:53:03Z

We talked a bit about this at the maintainers meeting. Would it be possible to trace before validation runs (so that invalid commands get included in the trace)?

teoxoy · 2025-10-17T17:25:15Z

I just realized Fallible::Invalid also has type Arc<String> - we could use that address to be able to record before validation.

Use it to deduplicate some `Foo` / `ArcFoo` / `TraceFoo` definitions.

This updates pass operations. The trace player is temporarily broken.

This updates non-pass command encoder operations and device operations.

andyleiserson · 2025-10-24T19:05:05Z

Since this PR is already quite large, I opened #8429 for the error tracing support.

andyleiserson force-pushed the tracing branch from ab0ed04 to 923162d Compare October 3, 2025 00:03

jimblandy assigned teoxoy Oct 8, 2025

teoxoy reviewed Oct 13, 2025

View reviewed changes

andyleiserson added 5 commits October 22, 2025 16:57

Move set_buffer_data and get_mapped_range out of Global

d935b78

Move surface_configure out of Global

179a276

Move start/stop debugger capture out of Global

d6156a8

Add ReferenceType trait for command structures

94cc67e

Use it to deduplicate some `Foo` / `ArcFoo` / `TraceFoo` definitions.

Create submodule of device::trace for recording support

3664e75

andyleiserson force-pushed the tracing branch from 923162d to 0e96cb8 Compare October 23, 2025 00:00

andyleiserson added 14 commits October 23, 2025 13:45

Store pointers instead of IDs when tracing (1/2)

510baed

This updates pass operations. The trace player is temporarily broken.

Store pointers instead of IDs when tracing (2/2)

b660760

This updates non-pass command encoder operations and device operations.

Replay traces with pointers as IDs

f66ab87

Clean up device poll API

13d3740

Fix debugger capture

be2b30d

Restore PointerId(...) serialized representation

ccb155f

Fix Free* operations

54a2a4f

Fix get_mapped_range duplication

2ccb2c8

Condition macro_rules_attribute dev-dependency on serde feature

4ce4630

Remove redundant device validity check and inaccurate comment

10ed0c4

Clarify comment about Arc::as_ptr

b2c7322

Remove stale attributes

8394c48

Fix serde feature gating

568125f

Remove get_buffer alias of resolve_buffer_id

eeb5fe1

andyleiserson force-pushed the tracing branch from e82d77c to eeb5fe1 Compare October 24, 2025 18:56

andyleiserson mentioned this pull request Oct 24, 2025

Support for storing failed commands in traces #8429

Open

6 tasks

Use pointers instead of IDs for tracing #8289

Are you sure you want to change the base?

Use pointers instead of IDs for tracing #8289

Conversation

andyleiserson commented Oct 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

andyleiserson commented Oct 8, 2025

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!

andyleiserson Oct 22, 2025 • 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!

teoxoy commented Oct 13, 2025

Uh oh!

teoxoy commented Oct 15, 2025

Uh oh!

teoxoy commented Oct 17, 2025

Uh oh!

andyleiserson commented Oct 24, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

andyleiserson commented Oct 2, 2025 •

edited

Loading

andyleiserson Oct 22, 2025 •

edited

Loading