-
Notifications
You must be signed in to change notification settings - Fork 952
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Reintroduce buffer snatching Part 1 #4867
Conversation
I don't want this marked as draft, it should get reviewed and hopefully landed. The "WIP" in the title was referring to the fact that it's a part of a larger ongoing refactoring, I'll remove it since it is more confusing than helpful. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good! Some organizational notes.
Ah, sorry! |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Nice!
…225e1744fd7. r=webgpu-reviewers,supply-chain-reviewers,teoxoy Changelog: * #4865 fix present mode for wgl By xiaopengli89 in gfx-rs/wgpu#4865 * #4873 Bump ctor from 0.2.5 to 0.2.6 By dependabot[bot] in gfx-rs/wgpu#4873 * #4874 Bump syn from 2.0.40 to 2.0.41 By dependabot[bot] in gfx-rs/wgpu#4874 * #4869 [naga wgsl-out] Include the `f` suffix on `f32` literals. By jimblandy in gfx-rs/wgpu#4869 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4870 [naga wgsl] Let unary operators accept and produce abstract types. By jimblandy in gfx-rs/wgpu#4870 * #4867 Reintroduce buffer snatching Part 1 By nical in gfx-rs/wgpu#4867 * #4882 Bump zerocopy from 0.7.26 to 0.7.31 By dependabot[bot] in gfx-rs/wgpu#4882 * #4878 Buffer snatching part 2 - Refactor create_buffer By nical in gfx-rs/wgpu#4878 * #4815 Add `wgpu` crate features for backends By daxpedda in gfx-rs/wgpu#4815 * #4887 Allow clippy::pattern_type_mismatch By nical in gfx-rs/wgpu#4887 * #4886 Document wgpu & wgpu-core features By Wumpf in gfx-rs/wgpu#4886 * #4826 validation: More detailed on incompatible BGL By scoopr in gfx-rs/wgpu#4826 * #4888 Web: add support for more `RawWindowHandle` variants By daxpedda in gfx-rs/wgpu#4888 * #4890 Bump thiserror from 1.0.50 to 1.0.51 By dependabot[bot] in gfx-rs/wgpu#4890 * #4880 Simplify `ResourceMaps` By nical in gfx-rs/wgpu#4880 * #4891 Make the naga version in trunk as high as the latest published one By nical in gfx-rs/wgpu#4891 * #4893 Avoid allocating memory every time we might log a label By nical in gfx-rs/wgpu#4893 * #4894 Remove some locks in BindGroup By nical in gfx-rs/wgpu#4894 * #4862 Ensure that DeviceLostCallbackC is always called exactly once By bradwerth in gfx-rs/wgpu#4862 * #4900 Support Device fence sharing with dx12 on Windows By sotaroikeda in gfx-rs/wgpu#4900 * #4903 Bump tokio from 1.35.0 to 1.35.1 By dependabot[bot] in gfx-rs/wgpu#4903 * #4895 Check that raw buffers and raw bind groups are valid By nical in gfx-rs/wgpu#4895 * #4901 fix: docs By miaobuao in gfx-rs/wgpu#4901 * #4892 Simplify some code around buffer unmapping By nical in gfx-rs/wgpu#4892 * #4896 Buffer snatching By nical in gfx-rs/wgpu#4896 * #4851 Eagerly release GPU resources when we lose the device. By bradwerth in gfx-rs/wgpu#4851 * #4906 Use nightly for docs By cwfitzgerald in gfx-rs/wgpu#4906 * #4908 Bump syn from 2.0.41 to 2.0.42 By dependabot[bot] in gfx-rs/wgpu#4908 * #4909 Bump profiling from 1.0.12 to 1.0.13 By dependabot[bot] in gfx-rs/wgpu#4909 * #4910 Bump anyhow from 1.0.75 to 1.0.76 By dependabot[bot] in gfx-rs/wgpu#4910 * #4913 Remove id32 Feature By cwfitzgerald in gfx-rs/wgpu#4913 * #4914 Add BGL Deduplication Index Test By cwfitzgerald in gfx-rs/wgpu#4914 * #4921 Fix typo "layout pipeline layout" -> "pipeline layout" By HactarCE in gfx-rs/wgpu#4921 * #4922 Bump winit from 0.29.4 to 0.29.5 By dependabot[bot] in gfx-rs/wgpu#4922 * #4924 Inline `document-features` usage, remove dep. By ErichDonGubler in gfx-rs/wgpu#4924 * #4871 Speed up Naga's `cargo xtask validate wgsl` from 12s to 0.8s By jimblandy in gfx-rs/wgpu#4871 * #4871 Speed up Naga's `cargo xtask validate wgsl` from 12s to 0.8s By jimblandy in gfx-rs/wgpu#4871 * #4871 Speed up Naga's `cargo xtask validate wgsl` from 12s to 0.8s By jimblandy in gfx-rs/wgpu#4871 * #4930 Bump winit from 0.29.5 to 0.29.6 By dependabot[bot] in gfx-rs/wgpu#4930 * #4929 Bump web-time from 0.2.3 to 0.2.4 By dependabot[bot] in gfx-rs/wgpu#4929 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4940 Align `wgpu_types::CompositeAlphaMode` serde serializations to spec By littledivy in gfx-rs/wgpu#4940 * #4936 Bump anyhow from 1.0.76 to 1.0.77 By dependabot[bot] in gfx-rs/wgpu#4936 * #4933 Bump thiserror from 1.0.51 to 1.0.52 By dependabot[bot] in gfx-rs/wgpu#4933 * #4932 Bump syn from 2.0.42 to 2.0.43 By dependabot[bot] in gfx-rs/wgpu#4932 Differential Revision: https://phabricator.services.mozilla.com/D197519
…225e1744fd7. r=webgpu-reviewers,supply-chain-reviewers,teoxoy Changelog: * #4865 fix present mode for wgl By xiaopengli89 in gfx-rs/wgpu#4865 * #4873 Bump ctor from 0.2.5 to 0.2.6 By dependabot[bot] in gfx-rs/wgpu#4873 * #4874 Bump syn from 2.0.40 to 2.0.41 By dependabot[bot] in gfx-rs/wgpu#4874 * #4869 [naga wgsl-out] Include the `f` suffix on `f32` literals. By jimblandy in gfx-rs/wgpu#4869 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4850 [naga wgsl-in] Support abstract operands to binary operators. By jimblandy in gfx-rs/wgpu#4850 * #4870 [naga wgsl] Let unary operators accept and produce abstract types. By jimblandy in gfx-rs/wgpu#4870 * #4867 Reintroduce buffer snatching Part 1 By nical in gfx-rs/wgpu#4867 * #4882 Bump zerocopy from 0.7.26 to 0.7.31 By dependabot[bot] in gfx-rs/wgpu#4882 * #4878 Buffer snatching part 2 - Refactor create_buffer By nical in gfx-rs/wgpu#4878 * #4815 Add `wgpu` crate features for backends By daxpedda in gfx-rs/wgpu#4815 * #4887 Allow clippy::pattern_type_mismatch By nical in gfx-rs/wgpu#4887 * #4886 Document wgpu & wgpu-core features By Wumpf in gfx-rs/wgpu#4886 * #4826 validation: More detailed on incompatible BGL By scoopr in gfx-rs/wgpu#4826 * #4888 Web: add support for more `RawWindowHandle` variants By daxpedda in gfx-rs/wgpu#4888 * #4890 Bump thiserror from 1.0.50 to 1.0.51 By dependabot[bot] in gfx-rs/wgpu#4890 * #4880 Simplify `ResourceMaps` By nical in gfx-rs/wgpu#4880 * #4891 Make the naga version in trunk as high as the latest published one By nical in gfx-rs/wgpu#4891 * #4893 Avoid allocating memory every time we might log a label By nical in gfx-rs/wgpu#4893 * #4894 Remove some locks in BindGroup By nical in gfx-rs/wgpu#4894 * #4862 Ensure that DeviceLostCallbackC is always called exactly once By bradwerth in gfx-rs/wgpu#4862 * #4900 Support Device fence sharing with dx12 on Windows By sotaroikeda in gfx-rs/wgpu#4900 * #4903 Bump tokio from 1.35.0 to 1.35.1 By dependabot[bot] in gfx-rs/wgpu#4903 * #4895 Check that raw buffers and raw bind groups are valid By nical in gfx-rs/wgpu#4895 * #4901 fix: docs By miaobuao in gfx-rs/wgpu#4901 * #4892 Simplify some code around buffer unmapping By nical in gfx-rs/wgpu#4892 * #4896 Buffer snatching By nical in gfx-rs/wgpu#4896 * #4851 Eagerly release GPU resources when we lose the device. By bradwerth in gfx-rs/wgpu#4851 * #4906 Use nightly for docs By cwfitzgerald in gfx-rs/wgpu#4906 * #4908 Bump syn from 2.0.41 to 2.0.42 By dependabot[bot] in gfx-rs/wgpu#4908 * #4909 Bump profiling from 1.0.12 to 1.0.13 By dependabot[bot] in gfx-rs/wgpu#4909 * #4910 Bump anyhow from 1.0.75 to 1.0.76 By dependabot[bot] in gfx-rs/wgpu#4910 * #4913 Remove id32 Feature By cwfitzgerald in gfx-rs/wgpu#4913 * #4914 Add BGL Deduplication Index Test By cwfitzgerald in gfx-rs/wgpu#4914 * #4921 Fix typo "layout pipeline layout" -> "pipeline layout" By HactarCE in gfx-rs/wgpu#4921 * #4922 Bump winit from 0.29.4 to 0.29.5 By dependabot[bot] in gfx-rs/wgpu#4922 * #4924 Inline `document-features` usage, remove dep. By ErichDonGubler in gfx-rs/wgpu#4924 * #4871 Speed up Naga's `cargo xtask validate wgsl` from 12s to 0.8s By jimblandy in gfx-rs/wgpu#4871 * #4871 Speed up Naga's `cargo xtask validate wgsl` from 12s to 0.8s By jimblandy in gfx-rs/wgpu#4871 * #4871 Speed up Naga's `cargo xtask validate wgsl` from 12s to 0.8s By jimblandy in gfx-rs/wgpu#4871 * #4930 Bump winit from 0.29.5 to 0.29.6 By dependabot[bot] in gfx-rs/wgpu#4930 * #4929 Bump web-time from 0.2.3 to 0.2.4 By dependabot[bot] in gfx-rs/wgpu#4929 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4902 [naga xtask] Run validation jobs in parallel, using jobserver. By jimblandy in gfx-rs/wgpu#4902 * #4940 Align `wgpu_types::CompositeAlphaMode` serde serializations to spec By littledivy in gfx-rs/wgpu#4940 * #4936 Bump anyhow from 1.0.76 to 1.0.77 By dependabot[bot] in gfx-rs/wgpu#4936 * #4933 Bump thiserror from 1.0.51 to 1.0.52 By dependabot[bot] in gfx-rs/wgpu#4933 * #4932 Bump syn from 2.0.42 to 2.0.43 By dependabot[bot] in gfx-rs/wgpu#4932 Differential Revision: https://phabricator.services.mozilla.com/D197519
Connections
This is the first step towards addressing #4787.
Description
We need
destroy()
methods to be able to deallocate gpu resources while the associated handle is still alive. This is difficult to do safely because most of wegpu's API can be used from any thread.One of the steps towards our goal is to re-introduce "buffer snatching". The idea is that when, say,buffer.destroy()
is called, we snatch the raw buffer out ofwgpu_core
's buffer struct, and enqueue it to be deallocated after the latest submission the buffer has been used in is complete.Snatching the raw handle out requires some form of locking. @cwfitzgerald's proposed solution is to protect all raw buffer and texture reads behind the read guard of a device-global
RwLock
, and only ever thake a write lock during snatching. The hope is that the overhead of taking a single read guard in most places will be small enough and thatdestroy
will be called concurrently with other wgpu work rarely enough that it won't cause very bad contention.This PR implements the very first step of Connor's idea:
Device
holdssnatchable_lock
.Snatchable<T>
type is introduced. Its content cannot be accessed without an appropriate lock guard.Buffer.raw
is now a ``Snatchable<A::Buffer>` and all associated reads need the guard.The snatching is not actually performed in this PR (yet). It will come in a followup.
Checklist
cargo fmt
.cargo clippy
.cargo xtask test
to run tests.