[dpadr]

Summary

• The queue scheduler replays each piece of an NVN queue as its own Metal command buffer, and alternates ready pieces of different queues round robin.
• A game with an async compute queue (Ocarina of Time's queue 5 holds all of its frame's compute; queue 1 holds every render pass) is replayed render, compute, render, compute. That makes 22 compute↔render switches a frame, at 35–46 µs median each on an M2, with a long compute → render tail.
• With DM_NVN_QDEFER=1, a ready piece of the other class is held while another queue offers a ready wait or a piece of the class in progress. Offline, one OoT frame's queue order replayed with its waits goes from 22 switches to 5, the fewest those waits allow.
• The held piece runs at once:
  • for a CPU fence wait (nvn_sync_wait);
  • before a present or another frame's work;
  • when nothing else qualifies;
  • after DM_NVN_QDEFER_US (default 3000).
• Ordering across queues stays the game's event and sync waits, and one queue's order never changes, so NVN's guarantees hold.
• Log: a [nvn] queue defer line gives the holds a frame and why held pieces ran.

With #2250 and e775054 (in this branch's base; not in the measured build)

• Compute dispatches can be ordered by the game's own barriers (profile key dispatch_barriers game): Kirby Air Riders' particle systems go from 6.6-9.8 ms to 0.4-1.2 ms a frame #2250 (dispatches ordered by the game's barriers): its state lives in the copy encoder. The range sets and g_blit_ops reset in blit_open, and each piece is its own command buffer, whose first copy encoder opens with a full queue barrier.
  • E1 moves whole pieces only, never work within one, so the order Compute dispatches can be ordered by the game's own barriers (profile key dispatch_barriers game): Kirby Air Riders' particle systems go from 6.6-9.8 ms to 0.4-1.2 ms a frame #2250 decides is unchanged. A game barrier flag left at the end of a piece is harmless: the next encoder's first dispatch takes no barrier anyway.
  • Compute dispatches can be ordered by the game's own barriers (profile key dispatch_barriers game): Kirby Air Riders' particle systems go from 6.6-9.8 ms to 0.4-1.2 ms a frame #2250 shortens the work inside queue 5's compute pieces (dispatches overlap). E1 cuts the switches between pieces. They save different things.
• e775054 (relaxed draw-pass entry with draw_barriers game):
  • The one piece of barrier state that outlived a piece was g_last_end_relaxed, global across NVN queues. A queue's pass right after another queue's pass could skip the full entry barrier, although its work after an event wait must see the other queue's writes in its vertex and indirect stages too.
  • Round robin already did this; E1 makes such hand-overs more common. The second commit clears the flag in mtl_queue_activate, with a t_queue_origin case.
  • For OoT (the only draw_barriers game profile) it costs nothing: queue 5 is compute only.
  • In OoT, E1 runs queue 1's render pieces back to back, where e775054's relaxed entry applies. So the two may add up.
• Not measured with either: the A/B below ran on a cut7 build (a406a81f), which has neither Compute dispatches can be ordered by the game's own barriers (profile key dispatch_barriers game): Kirby Air Riders' particle systems go from 6.6-9.8 ms to 0.4-1.2 ms a frame #2250 nor e775054. Their interplay with E1 is covered by the reasoning above and by the new t_queue_origin case (the second commit), not by a measurement.

Measurements

Apple M2 MacBook Air (16 GB), macOS 27.0. Ocarina of Time 0400486024DEC000 v131072 on our cut7 build (a406a81f: an older main plus OoT-specific host changes, shader chain bf6035002b98), handheld 1440x810, 60 fps cap. Route4 to frame 12000, in-run A/B DM_NVN_QDEFER=ab with DM_AB=240: 21 phases on and 21 off, each on phase against the mean of its two off neighbours, bootstrap 95 %.

Measure	Change with QDEFER on	95 % range
GPU idle within the frame	−0.72 ms	[−4.28, −0.49]
GPU span (first to last work)	−2.2 ms	[−4.28, −0.03]
Frame p90	−3.3 ms	[−10.2, −0.12]
Frame p50	−1.77 ms	[−4.16, +0.28], the interval crosses 0
GPU busy	+0.15 ms	noise

• The run had no hang (hangs.py CLEAN) and no new [RED] kinds.
• It ran on the cut7 build, partly on battery.
• We'll rebase onto current main and re-measure once main's post-load slowdown (Ocarina of Time on an M1 Pro: slow motion after ~80 min on main 25c2546, finalized-texture table puts jump from <1 to thousands a frame #2213) is resolved. Until then current main isn't a stable base to A/B on.

Tests

• t_wait_event: a new queue-defer case (each policy reason, the cap's edge, op classes, a recorded copy-only piece classed compute).
• check_t_wait_event and check_t_queue_lifetime: also run with DM_NVN_QDEFER=1.
• test_env_switches: OK.
• t_queue_origin: the queue-switch barrier case, also run with DM_NVN_BARRIER=game (sensitivity).
• Quick-lane run on this branch (qdmtests2, 61967a5): all 0 failed.
  • t_wait_event: 116 checks in each of its 3 runs (ENC_THREAD 0, 1, and 1 with QDEFER=1).
  • t_queue_lifetime: 38 checks in each of its 3 runs.
  • t_queue_origin: 399, 399, and 400 with DM_NVN_BARRIER=game, where the sensitivity check runs.
  • test_env_switches: OK.

Regression (rule: scripts/regress.py check --games)

• Not run. The only local NVN game packages the harness knows, Animal Well and Mina, are stale: they were split before shader packages carried their keys, and main's f7a5452 (Ocarina of Time (0400486024DEC000) black screen with sound on c78d7375 / chain 89f44f9b6bb0: the present pass pairs VS a5b1bd74e47458d7 (writes loc0 only) with FS 29e3c9dd92b8a571 (reads loc1/loc2) #1841) refuses them, so the check stops before any comparison. UFO 50 is GL, which this change doesn't reach. Ocarina of Time's coverage is the A/B above.
• Reach: only when the switch is on (off by default); nothing changes with it off. With the switch off, the only code change on the replay path is classifying ops at record time.

🤖 Generated with Claude Code