[LiamPalmqvist]

draw_barriers game lets a draw pass's vertex work overlap the previous pass's fragments, ordered by the game's own barriers. On Ocarina of Time it had been doing nothing since 7f50222: every draw pass got the full barrier (draw-pass entry barriers full 31025 / saved 0 per 300 frames).

Cause

Two things in enc_r_begin_pass / enc_r_end_pass (dm/gfx/nvn/nvn_mtl.m):

1. The relaxed barriers were turned off whenever image family heaps exist (&& !g_image_heap_n), to keep the alias flush off end barriers, where it costs a command allocator about 100 KB per encoder. Ocarina of Time's dynamic resolution always has family heaps.
2. Any nonzero nvnCommandBufferBarrier forced the full consumer barrier. In NVN only ORDER_PRIMITIVES and ORDER_INDIRECT_DATA order earlier work before later vertex or indirect fetch. Ocarina of Time's most common barrier, 0x12 (ORDER_FRAGMENTS | INVALIDATE_TEXTURE, pass dump of frame 1400), asks only for fragment-to-fragment order, which the relaxed end barrier (vertex, mesh and fragment before fragment, dispatch and blit) already gives.

Fix (about 7 lines)

• The guard is removed. The relaxed end barrier flushes to the device only, like m4_end. When aliases exist, the alias flush is a start barrier of the next pass before its fragments (barrierAfterQueueStages:All beforeStages:Fragment visibilityOptions:resource_vis()), following the file's rule that alias flushes go on start barriers (no allocator memory), and vertex work still overlaps.
• g_game_barrier_all: only a barrier with ORDER_PRIMITIVES or ORDER_INDIRECT_DATA forces the full consumer barrier. g_game_barrier keeps its meaning for the slice-merge check.
• Games without aliases behave as before apart from the bit narrowing.
• Docs: a docs/reference/frame-rate.md entry, the draw_barriers row in dm/games/README.md, the Ocarina of Time note (profile row and "Found on this game"), the symptom index.

Measured

Ocarina of Time (0400486024DEC000), M1 Max on AC, kokiri-deck gameplay (t 36-52 s), frame hold off, two runs each, every screenshot in Kokiri Forest:

	frame ms	GPU ms
base	23.15 / 23.55	22.02 / 22.25
guard removed only	22.76 / 22.82	21.57 / 21.68
both changes	21.85 / 21.59	20.38 / 20.23

Bench p50: fps 42.5 → 45.8, frame 23.52 → 21.84 ms, GPU 22.88 → 20.91 ms, frames over budget 88 → 23. Entry barriers saved 0 → about 68 of 103 a frame. In a Metal System Trace, render-to-render idle gaps in one command buffer went from 141 to 72 a frame (2.34 → 1.15 ms). The picture is unchanged: the screenshots match as closely as two base runs match each other (PSNR base vs base 19.8 dB, base vs fix 20.0-20.6 dB, the differences in Navi, particles and leaves); no [RED] lines.

Checked

• make -C dm/host GAME=0400486024DEC000 builds; make -C dm/gfx/tests GAME=0400486024DEC000 run passes, including t_image_alias, t_rt_range* and t_copy_alias.
• test_docs.py passes after docs_symptoms.py.
• regress.py check could not run: the only game built on this Mac is Ocarina of Time, which has no smoke route. The other games with draw_barriers game (Breath of the Wild and others) now also skip the full barrier after fragment-only barriers, as NVN allows, and were not run on this change.

Open

About 37 full entry barriers a frame remain on Ocarina of Time (ORDER_PRIMITIVES barriers, compute in between), and about 1.3 ms a frame of gaps at command-buffer boundaries (segment cuts and the CPU commit).

[dpadr]

A user log on main from an M1 Pro (discussion #1789, @brucewrks; our read of it) shows the same thing in the wild: draw-pass entry barriers full 33320 / saved 0 per 300 frames over a 3 h session.

A second device data point: on an M2 Air, the equivalent fix in #1970 saved ~75 % of entry barriers but didn't move GPU span, so the gain looks device-dependent.

[seedds]

Landed as e775054 on main, with you as co-author. Your code is kept: the guard is gone, the alias flush is a start barrier before the fragments, and only ORDER_PRIMITIVES / ORDER_INDIRECT_DATA force the full barrier. Added on top:

• The relaxed end barrier also waits for MTLStageObject. A task shader's writes weren't covered on that side; this was missing on main too.
• One variable instead of two: g_game_barrier keeps the barrier bits, and GAME_BARRIER_FULL masks them. The slice merge still stops at any barrier.
• Docs: only Ocarina of Time's profile has draw_barriers game (Breath of the Wild's says full), so the 'other games' line is gone. OoT's note no longer says the start-barrier flush 'was not kept'. The entry now has the per-device numbers from Draw passes overlap again while image heaps are alive and after fragment-only game barriers #1970: a second M1 Max and an iPhone gained, an M2 Air and an M4 Pro didn't change.
  Checked: host build, make -C dm/gfx/tests run, and the alias and rt_range checks with DM_NVN_BARRIER=game. Thanks for the measurements and the trace.