Add Safe GPU Api #41
Add Safe GPU Api #41
Conversation
|
Pardon the delay, most of these should be addressed now. Still haven't done documentation, but will wait until we've settled on api shape. |
|
This totally got lost in my history - got this fully back up to date, an improved some of the code to make it more resilient. Needs more documentation and the test causes tracy to crash, which I need to get a debug build of tracy to debug (this was an issue before the fix, the test just makes it trivial to reproduce. |
|
I would be happy to land this as soon as the doc-comments are filled in. Tracy crashes: unfortunately those usually are an issue in the upstream project. I’m of an opinion that we’re not supposed to work around those in the bindings. The only place where any meaningful amount of complex logic is warranted is in the |
|
Cool! Will add the doc comments and such and some end-to-end documentation. I want to at least get a culpret for the tracy crash, to make sure we're not somehow misusing the api, as the builtin integrations seem to work okay. Will build a c repro if it happens to be Tracy's fault, thankfully should be pretty simple. |
|
Traced down the bug, it was actually an assert in tracy that was getting hit. I wasn't using the From looking at the tracy's use of query ids, they have asserts that they are never re-used until the timestamps are written. As such, I'm going to add a freelist for validation and re-use prevention. |
|
Failures unrelated. |
|
I think this is good for final reviews |
| #[cfg(feature = "enable")] | ||
| gpu_start_timestamp: i64, | ||
| #[cfg(feature = "enable")] | ||
| span_freelist: Arc<Mutex<Vec<u16>>>, |
There was a problem hiding this comment.
Are we certain that span IDs are per-context and not global in tracy?
There was a problem hiding this comment.
Yes, they make no attempt to globally de-duplicate them in their own bindings.
|
Merged this in 9ed943b and pushed out a new release. |
|
The commit on main unfortunately appears to be a fairly old iteration of this PR, missing most of the doc comments and some of the fixes. |
Shows you shouldn't be merging stuff late Sunday night lol :)
|
Ah, my bad, looks like squash-merging this on a late sunday night was a mistake :) Yanked 0.16.0, merged the rest of the PR, and released a new one. I believe it correctly contains the rest of this PR. |
Depends on #39
Context Wumpf/wgpu-profiler#15
This adds the GPU api to tracy_client. It merely adds safe interfaces. I am putting this up as a draft to get opinions on architecture.
TODO: