Repository navigation
Independent worlds stepped at once: worker slots of their own per world - #106
Merged
Merged
Conversation
A world's steps ran as worker 0 on whatever thread called them, with its tasks as workers 1..n, so two worlds stepped at once on different threads shared every module's per-worker scratch and parallel_for's task records (module globals) -- a crash, as the new test shows on main. Box3D supports stepping independent worlds in parallel. Now (#104): - A world claims a run of worker slots of its own when made (a 64-bit mask, MAX_WORKERS slots) and gives it back when destroyed; with none free it starts at slot 0, shared, as before. - step, the collide pass and the solve enter the world (the thread becomes the world's first slot) and give the thread its index back after; tasks run as the slots on from their caller's; task_context subtracts the world's base (a serial caller outside a pass is worker 0). - parallel_for's task records are the scheduler's own, not globals; enqueue and finish give the caller its index back rather than setting 0. - The bullet pass reads another bullet's flags atomically and sets its own with atomic ors (the other part of #104: harmless in practice, but a data race in C). Worlds are still made and destroyed on one thread at a time, as the reference's are; queries run as the calling thread's slot. test_determinism's test_concurrent_worlds: two two-worker pyramids stepped at once, one on a host thread and one on the main thread, each end where the same world stepped alone does, bit for bit (main crashes).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #104.
Until now a world's step ran as worker 0 on whatever thread called it, with its tasks as workers 1..n. Two worlds stepped at once on different threads therefore shared every module's per-worker scratch, and also
parallel_for's task records, which were module globals. The new test crashes on main for that reason. Box3D supports stepping independent worlds in parallel. This change:MAX_WORKERSslots) and releases them when destroyed. If none are free, it starts at slot 0, shared, as before.step, the collide pass and the solve make the calling thread the world's first slot, then give the thread its index back. Tasks run as the slots following their caller's.task_contextsubtracts the world's base; a serial caller outside a pass is treated as worker 0.parallel_for's task records belong to the scheduler, not module globals.enqueueandfinishrestore the caller's index instead of setting it to 0.Unchanged: worlds are still created and destroyed on one thread at a time, as in upstream, and queries run as the calling thread's slot.
test_determinism'stest_concurrent_worldssteps two two-worker pyramids at once, one on a host thread and one on the main thread. Each ends bit-identical to the same world stepped alone. It passed 10 of 10 runs; main crashes every time.All 28 test programs pass locally on aether 0.772.0, with no warnings.