Try   HackMD

Multiple components cannot effectively share a GPUDevice because it is stateful

Issue: https://github.com/gpuweb/gpuweb/issues/3250

Proposal written and proposed by Kai, based on in-depth discussion with editors (Brandon/Myles/Kai) and more input from Brandon.

Previous drafts

Premises

  • It should be possible for multiple "components" using the same device on the same thread to have different error scope state.
  • It's okay if each component can't device.destroy() independently.
    • The per-component device should be a lightweight object that can be handled by GC.
    • This avoids difficult-to-understand tracking of which resources would get destroyed upon destroying one component's device.
  • Error scope state should not be shared across threads by default, as it would be a footgun.
    • For implementing native semantics over JS, it is useful for there to be a way for a device's error state to be shared across multiple threads.

Proposal

(Names are all placeholders.)

  • All objects have an [[errorTarget]], which (roughly speaking) has type GPUDevice?.
    • This means createView(), for example, does not have to move.
  • The only pre-V1 change to today's API is the addition of GPURootDevice. (I have long expected we would want this - I just didn't realize it was a breaking change.)
    • There is only one uncapturederror EventTarget object for the device (GPURootDevice) and it's never possible to duplicate or transfer it. This approach is chosen so [[errorTarget]] can be nullable.
  • When an object is sent to another thread, the error target does not come with it; [[errorTarget]] is null. Errors that happen on this object get sent straight to uncapturederror. To get them to go to an error target, you have to "clone" the object and specify an errorTarget to use.
  • GPUDevice is the error target and can never have a null error target.
    • When you send it to another thread you have to explicitly choose whether it's going to share its error state tracker or get a new one.

Pre-V1

Changes:

  • Add a GPURootDevice subclass of GPUDevice. Move EventTarget to GPURootDevice. Return GPURootDevice from requestDevice().
    • This is done now only because the addition of a new object in the prototype chain is observable. This was planned for multi-threading anyway.
    • This is impossible because it requires multiple inheritance (GPURootDevice : GPUDevice, EventTarget).

Editorial changes:

  • GPUObjectBase gets an internal slot [[errorTarget]], pointing to a "device client" (a thing which can send operations to a device).
  • A new object created from any object this inherits this.[[errorTarget]].
  • All operations send their errors to (GPUObjectBase)this, which passes them through [[errorTarget]] and forwards uncaptured errors to uncapturederror.

Notes:

  • There is exactly one error target per device, and not yet any way to override [[errorTarget]]. No significant implementation work is needed at this stage.

Post-V1: Multi-Client

Changes:

  • Most objects gain a method object.withErrorTarget(GPUDevice errorTarget), which gives you a new JS object with the same underlying object and the provided [[errorTarget]].
    • GPUDevice and GPURootDevice don't have these.
    • Object labels are client-shared state.
  • Add GPUDevice.withNewErrorTracker() which returns a new GPUDevice with a new "device client" with an empty error state tracker.

Post-V1: Multi-Threading

Changes:

  • Make [[errorTarget]] nullable.
    • Objects without an [[errorTarget]] send errors directly to uncapturederror.
  • Most objects are thread-shareable. When deserialized, they have a null [[errorTarget]]. To associate them with an error target, the receiver must call withErrorTarget().
    • Object labels are thread-shared state.
  • When GPUDevice or GPURootDevice is sent, you receive a GPUReceivedDevice. To get a GPUDevice from it, you must choose either .withNewErrorTracker() or .withSharedErrorTracker(). (These can be called multiple times.)
    • You can't get a GPURootDevice from it (that object is tied to the original thread).
    • This affects whether the internal [[errorTarget]] points to the same device "client" or a new one.

Note:

  • Native-style semantics can be implemented with withSharedErrorTracker(). This also eliminates a need in the actual native API to use thread-local storage to track which error scope state to target.

Post-V1: Multi-Queue

New independent queue objects behave the same way as other objects.

Last changed by 
 
WebGPU
https://webgpu.dev
4
267

Read more

WebGPU upload paths (MDN article draft)

writeBuffe

Jul 19, 2023
DRAFT SharedValueTable Proposal

[Serializable] interface SharedValueTable { undefined insert([EnforceRange] unsigned long long key, any value); // value must be serializable any get([EnforceRange] unsigned long long key); undefined remove([EnforceRange] unsigned long long key); }; TODO write more Keeping track of some use cases, and random thoughts

Nov 18, 2022
Order of error scope resolution

Proposed by editors (Myles/Brandon/Kai), and written by Brandon, with edits from Kai. An ongoing concern about use of error scopes is that they may require that operations within the scope finish in-order. At this time, it's unclear if that may prevent certain classes of optimization, such as moving pipeline creation to a separate thread. (See #2119 for prior conversation on that point.) After some discussion the editors feel that in-order resolution is not a requirement for error scopes. This document outline the reasons why, and how we feel error scopes should operate. Proposal The proposal is to remove all constraints on the ordering of the resolution of promises from popErrorScope(), relative to one another and to other promises in the API. This allows implementations use a strict ordering, resolve strictly as early as possible, or anything else. The general behavior that we would expect from an error scope is:

Jul 26, 2022
(Drafts) Multiple components cannot effectively share a GPUDevice because it is stateful

These are drafts. The working proposal can be found here. [TOC] Multiple components, for example multiple graph/plot widgets on a page, cannot effectively share a GPUDevice because it is stateful: It has a global error scope state, which means any error scope spanning an async boundary will capture stuff from other components. This is very difficult for an application to work around. It can be destroyed, which means one component which wants to clean up after itself cannot take advantage of GPUDevice.destroy() to do so. An application can somewhat work around this by tracking its own resource allocations, which is also more flexible for more complex ownership models (like refcounting), though that only allows freeing Buffer/Texture/QuerySet. This can be useful for some other things as well:

Jul 21, 2022
Read more from WebGPU
Published on HackMD