[mtl] Refined passes #2288
[mtl] Refined passes #2288
Conversation
|
hmm, I detect a performance regression here, investigating the cause now... |
|
Good news - the regression is solely caused by the extra commit here, so I dropped it. It's an interesting piece that's not obvious to me: so building one gigantic chain of commands and recording them in one go is slower than calling |
| @@ -759,12 +759,29 @@ impl<'a> PreRender<'a> { | |||
| PreRender::Void => (), | |||
| } | |||
| } | |||
|
|
|||
| fn issue_many<'b, I>(&mut self, commands: I) | |||
There was a problem hiding this comment.
Not really important for now, but it seems like we should be able to collapse these functions by being generic over the command type
src/backend/metal/src/command.rs
Outdated
| { | ||
| { | ||
| let (mut pre, switch) = self.switch_compute(); | ||
| assert!(switch); |
There was a problem hiding this comment.
Why is this assertion guaranteed? It seems valid to have an encoder left open but call {fill|copy}_buffer, although I'm probably missing something here
There was a problem hiding this comment.
good point! my assumption was wrong here
|
bors=grovesNL |
|
bors r=grovesNL |
2288: [mtl] Refined passes r=grovesNL a=kvark Prevents rebinding of all the state and resources in case of consecutive compute dispatches. ~~Also, tries to make a list of commands for graphics pipeline setup and pass it as one big thing (instead of calling `pre.issue` for each individual one). This is piggy-backing on the `issue_many` function of the base commit here.~~ Edit: removed this commit for causing a performance regression. Changes the sink API to be more clear about how the passes are started/ended, it's now the same across compute/render and distinct from blits. PR checklist: - [ ] `make` succeeds (on *nix) - [x] `make reftests` succeeds - [x] tested examples with the following backends: metal - [ ] `rustfmt` run on changed code Co-authored-by: Dzmitry Malyshau <[email protected]>
2292: Metal command recording option r=grovesNL a=kvark Based on #2288 Implements first bits of #2236 Allows the client to control the recording option without re-building gfx-rs. Also hides the remote sink behind a feature. We can't CI-test it until SSheldon/rust-dispatch#10 is accepted and published. PR checklist: - [ ] `make` succeeds (on *nix) - [ ] `make reftests` succeeds - [ ] tested examples with the following backends: - [ ] `rustfmt` run on changed code Co-authored-by: Dzmitry Malyshau <[email protected]>
2292: Metal command recording option r=grovesNL a=kvark Based on #2288 Implements first bits of #2236 Allows the client to control the recording option without re-building gfx-rs. Also hides the remote sink behind a feature. We can't CI-test it until SSheldon/rust-dispatch#10 is accepted and published. PR checklist: - [ ] `make` succeeds (on *nix) - [ ] `make reftests` succeeds - [ ] tested examples with the following backends: - [ ] `rustfmt` run on changed code Co-authored-by: Dzmitry Malyshau <[email protected]>
Prevents rebinding of all the state and resources in case of consecutive compute dispatches.
Also, tries to make a list of commands for graphics pipeline setup and pass it as one big thing (instead of callingEdit: removed this commit for causing a performance regression.pre.issuefor each individual one). This is piggy-backing on theissue_manyfunction of the base commit here.Changes the sink API to be more clear about how the passes are started/ended, it's now the same across compute/render and distinct from blits.
PR checklist:
makesucceeds (on *nix)make reftestssucceedsrustfmtrun on changed code