Repository navigation
Conversation
…ven while earlier work runs: Ocarina of Time -0.85 ms a frame on an M1 Ocarina of Time holds one frame in flight: its main thread waits on the fence of the frame it just submitted (the wait ledger: SyncWait on the newest fence 0.72 times a frame, 17.4 ms), so at each frame boundary the GPU has nothing queued until the next frame's first segment is committed. sub_cut's busy rule commits a segment only once no earlier one is in flight, which held that first segment back until the previous frame's last ran out, and then until the next cut point: a counter-free Metal System Trace on an M1 showed 0.4-3 ms of GPU idle there every frame. Now a submission's first segment is cut after DM_NVN_CUT_FIRST_N draws (8) and committed at once (DM_NVN_CUT_FIRST, on by default; =0 off, ab for the DM_AB phases), so it runs while the previous frame's last work finishes. Measured (0400486024DEC000, Kokiri Forest, MacBook M1 8-core GPU, fullscreen, in a tree with other local changes, among them the image-heap alias flush once a submission of the companion PR, which lets command buffers overlap at all): DM_NVN_CUT_FIRST=ab:2,3/4 in 300-frame phases, frame p50 36.33 / 36.41 / 36.37 ms with it against 37.19 / 37.33 / 37.41 without; with the phases swapped (ab:0/2, 450-frame phases) 36.48 / 36.53 / 36.51 against 37.57 / 37.38 / 37.35: -0.85 ms, lower in every phase both ways. Without the overlap the alias flush on every barrier forbids, an earlier test of committing every segment at once gained nothing, so the gain here likely needs that change. Not run: this branch's host in game, any test suite, the iPhone, other games. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
github-actions
Bot
added
nvn
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
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.
What
A submission's first segment is cut after
DM_NVN_CUT_FIRST_Ndraws (8) and committed at once, even while earlier segments are still running (DM_NVN_CUT_FIRST, on by default;=0off,abfor theDM_ABphases).Why
Ocarina of Time holds one frame in flight: its main thread waits on the fence of the frame it just submitted (wait ledger: SyncWait on the newest fence 0.72 times a frame, 17.4 ms). At each frame boundary the GPU has nothing queued until the next frame's first segment is committed.
sub_cut's busy rule commits a segment only once no earlier one is in flight. That held the first segment back until the previous frame's last ran out, and then until the next cut point. A counter-free Metal System Trace on an M1 showed 0.4–3 ms of GPU idle at every frame boundary. Committed early, the first segment runs while the previous frame's last work finishes.Performance
Ocarina of Time (0400486024DEC000), Kokiri Forest, MacBook M1 (8-core GPU), fullscreen, in a tree with other local changes. Those include the image-heap alias flush once a submission (companion PR), which lets command buffers overlap at all.
ab:2,3/4, 300-frame phasesab:0/2, 450-frame phasesAbout −0.85 ms a frame, lower in every phase both ways.
An earlier test of committing every segment at once gained nothing while the alias flush on every barrier serialized the GPU. So this gain probably depends on the companion change.
Switches (in
docs/env_switches.tsv)DM_NVN_CUT_FIRST=0|1|ab: on by default.DM_NVN_CUT_FIRST_N=<draws>: the first segment's size (default 8).Not run
🤖 Generated with Claude Code