WIP: perf(macos): a pass's leading whole clears as its attachments' LOAD_OP_CLEAR, off by default (RX_VK_LOAD_CLEAR=1) #780

Closed

mooki wants to merge 80 commits from mooki/Longshot:mac/load-clear into main

pull from: mooki/Longshot:mac/load-clear

merge into: MagicBeanDev:main

MagicBeanDev:main

Conversation 1 Commits 80 Files changed 181 +32211 -786

mooki commented 2026-11-15 11:55:46 +00:00

Member

• Copy link
• Copy Markdown

Type: perf · Risk: low · Tested on: macOS 26.7.1, Apple M4 Max (MoltenVK 1.4.2 and MoltenVK main 5189f8de + argbuf-slots), arm64 and x86_64 under Rosetta · Needs: macOS, Apple silicon with less memory bandwidth than an M4 Max (a hold with the switch on); Windows and Linux, any GPU (build and start)

Why

Every render pass LS records begins by loading each attachment (LOAD_OP_LOAD). On Apple's tile GPUs a load copies the attachment
from memory into tile memory, and when the pass's first commands clear the attachments whole, that copy is thrown away at once; the
clear is then one more MoltenVK quad, with the state it disturbs. In run #31's field-dawn hold 37.8 passes a frame began with whole
clears: 46.8 attachments, 76 MB of the 650 MB loaded a frame (11.7%), loaded only to be overwritten. This PR can make those clears the
attachments' LOAD_OP_CLEAR, Metal's clear load action, which only fills the tiles, wherever that draws exactly what the clears draw.
On an M4 Max it measured under the run-to-run noise, so it ships off, behind RX_VK_LOAD_CLEAR=1, for GPUs with less memory
bandwidth to try. It is part of the macOS port's performance work (MacPerf §4 #7). See #420.

What changed

With RX_VK_LOAD_CLEAR=1 a pass whose first operations are whole clears begins with them as its load ops: its vkCmdBeginRendering
is held while only those clears and changes of state come, and recorded with LOAD_OP_CLEAR for the cleared attachments before
anything else goes into the command buffer. Everything else is recorded as before; a clear that cannot ride on a load op stays a
vkCmdClearAttachments. Off by default on every driver (Integration 40, from the E-series): unset or 0, every pass is recorded as
before.

• the pass: gpu/vk/vk_draw.cpp Pass gains load_clear::Held lead (the ten VkRenderingAttachmentInfo LS gives
  vkCmdBeginRendering, the render area, held) and lead_cmd; begin_pass(cmd, at, lead) fills lead instead of stack structs
  and, for a clear that may fold, holds the begin; pass_settle records a held begin, called by begin_pass for anything that is not
  a foldable clear on the same targets, by end_pass right before vkCmdEndRendering, and by the replay (draw_pass_settle).
• the replay: gpu/vk/vk_replay.cpp replay() settles a held pass before any operation outside keeps_lead (the clears and the
  operations that record nothing on the GPU, a set of bits), before it records or admits anything. Nothing reaches the command buffer
  between a pass's begin and its settle, so the command buffer is the one recorded before, less the folded clears, with those
  attachments' load ops changed.
• the fold, only where it is exact (gpu/vk/load_clear.h, new; clear_color_target, clear_depth_stencil, lead_why): every
  attachment the render area's size and single-sampled, no BC slice or raw view of BC (Targets::whole_extent in targets():
  MoltenVK's clear load action clears each attachment whole, past a smaller render area); the clear of the whole render area and of
  every channel or the aspects it names (the depth's and the stencil's load ops each their own); a value MoltenVK's two ways write
  alike (its clear load action and its vkCmdClearAttachments quad: no conversion that rounds, per format); the pass's begin not yet
  recorded.
• the switch: gpu/vk/vk_vendor.h vendor fix load_clear, off on every driver (x.load_clear = false), RX_VK_LOAD_CLEAR=1 turns
  it on; the device line vk: a pass's leading whole clears as its attachments' load ops (LOAD_OP_CLEAR): off (RX_VK_LOAD_CLEAR) on
  MoltenVK, on wherever it is on (gpu/vk/vk_main.cpp), and the first pass that folds says so once. The bench's B1 arm:
  RX_VK_BENCH_AB=load_clear (the load ops off for the run, on in the bench's B block: A is the clears in the pass, B the load ops,
  on the same frame).
• the default: off, its own commit 7c6f0e9c, Integration 40's 2b4b7b8e for these files (vk_vendor.h, the docs rows,
  test_load_clear).
• the measurement: #704's L2 line (gpu/vk/perf_lines.h, gpu/vk/vk_perf.cpp) also counts what each leading whole clear became:
  a load op, switched off, other attachment sizes or samples, its value or format, the pass begun already (a frame, and in MB).
• docs: docs/development.md (the RX_VK_LOAD_CLEAR row, the bench arm, L2), gpu/vk/VULKAN.md. Tests: tests/test_load_clear.py
  (new; it pins the default off on every driver and 1 on); tests/test_perf_lines.py reads begin_pass's new signature.

Measurements

Verdict: doesn't help measurably on an M4 Max: in the field-dawn hold all on but LoadClear (E4c) and all on (E2) both ran
33.7 fps and 29.67 ms a frame, GPU 19.2 against 19.1 ms, under the run-to-run noise of 0.17 ms of frame, 0.0 ms of GPU and 0.2 fps
(E7), though every one of the 46.8 leading whole clears a frame folded (76 MB of loads skipped). So it is off by default. It is exact
on the GPU where probed: LC-01 drew every folded clear's attachments texel for texel as the clears did, on both MoltenVK builds.

Baseline: main does not start on macOS. The E-series runs are the integration build I39 7a7726fe (this PR's code with
#782's and #783's, and the other integrated macOS topics; LoadClear on by default there), arm64, one run each: before
is E4c, all on but LoadClear (RX_VK_LOAD_CLEAR=0), after is E2, all three on. This head's LoadClear functions are I39's text; only
the default differs (off here, as Integration 40 made it). LC-01 records, in a standalone MoltenVK process per build, each pass as LS
records it with the switch off (A: LOAD_OP_LOAD, then vkCmdClearAttachments, then a draw over the cleared values) and on (B:
LOAD_OP_CLEAR, the same draw), with the PR's own load_clear.h deciding what folds, and compares every texel of every layer,
level, slice and aspect read back.

Hardware: macOS 26.7.1 (25G241), Apple M4 Max (40-core GPU), MoltenVK main 5189f8de + argbuf-slots (c170a745; LC-01 also MoltenVK
1.4.2), 16-core CPU, 64 GB
Scene: the E-series: the field-dawn save, slot 00, entered by START and A×5, then 600 s held in play, windowless; the instrument set
on in every run (RX_VK_PERF_LINES=1, RX_VK_GPU_TIMELINE=1, the present trace); quiet required before each start, the other lanes
paused. LC-01: 373 cases (every colour and depth-stencil format LS can fold, each foldable value; an array layer, a mip level with
and without depth, a 3D slice, the depth alone, the stencil alone, two and three colours; three cases of other attachment sizes) and
64 sweeps of the values LS refuses
Method: the E-series: the medians of the hold's 120-present periods from +20 s (E4c 162 periods, E2 141; the bench's and the chain's
periods left out of E2); one run each, the noise from E7 (a repeat of E2); runs mac-arm64-7a7726fe-06 (E4c), -02 (E2), -08
(E7); reports reports/2026-11-15-mac-perf-E4c.md, -E2.md, -E7.md, the series reports/2026-11-15-mac-perf-before-after.md.
LC-01: reports/LoadClear-files/lc01/runs-lc01-arm64/result.json (787dc0cf…), released once on arm64 by the parent; review
f00090ed…, build 00ac1f0a…, plan a651ba74…
How to reproduce: the series' ls-run.py prepare/run commands and common switches (reports/2026-11-15-mac-perf-runs-i39.md
§2-§3), with RX_VK_LOAD_CLEAR=0 for the before run; LC-01's release command (reports/2026-11-14-load-clear.md §7)

field-dawn hold, 600 s	Before, RX_VK_LOAD_CLEAR=0: integration @ 7a7726fe (E4c)	After, load_clear on: integration @ 7a7726fe (E2)
fps (median [min-max] of the periods)	33.7 [32.2-34.3]	33.7 [32.7-34.4]
GPU ms a frame	19.2 [18.7-19.8]	19.1 [18.6-19.8]
frame ms	29.67	29.67
L2 leading whole clears a frame	46.8, every one switched off	46.8, every one a load op (76.0 MB of loads skipped)

E4c: 162 periods; fps 33.7 [32.2-34.3]; frame ms 29.6736; GPU ms 19.2 [18.7-19.8]
E2: 141 periods; fps 33.7 [32.7-34.4]; frame ms 29.6736; GPU ms 19.1 [18.6-19.8]
E7: run-to-run noise: 0.17 ms of frame (29.50 against 29.67), 0.0 ms of GPU, 0.2 fps

LC-01, each MoltenVK build	Before, A (LOAD and the clears): load-clear @ e0d3af8f	After, B (LOAD_OP_CLEAR): load-clear @ e0d3af8f
Q1: every foldable format and value	327 cases drawn (44 skipped: formats the GPU does not render to)	equal to A in all 327, every texel
Q3: layer, mip, 3D slice, depth alone, stencil alone, two and three colours	drawn	equal to A
Q4: a depth smaller than the colour (X.smaller_depth)	drawn	differs in 512 texels, as predicted: LS does not fold it (whole_extent)
Q2: the values LS refuses (sweeps)	drawn	differ where predicted (half floats 101-197 of 206, D16 195 of 300, 32-bit integers past 2^24)

LC-01 result.json: all_as_predicted 1; main-argbuf-slots and v1.4.2: q1_exact 1 (327 cases), q3_exact 1, q4_as_predicted 1,
controls_equal 1, mvk_errors 0

Run #31's prediction (about 0.2-0.5 ms of the GPU frame if most clears fold) did not hold on the M4 Max: all folded, 0.1 ms of GPU.
A Mac with less memory bandwidth may gain more [inference, not measured].

Flags and settings

One new switch, and one new arm of #704's bench switch; off by default.

Flag	What it is	Before this PR	After this PR	Effect	Helps / doesn't help / untested	Recommendation
RX_VK_LOAD_CLEAR=0 or 1	a pass's leading whole clears as its attachments' load ops	no switch: LOAD_OP_LOAD and vkCmdClearAttachments	unset or 0: off; 1 on, on any driver	fewer attachment loads and clear quads where a pass begins with whole clears; the same pixels	exact in LC-01 (both MoltenVK builds); under the noise on an M4 Max (E4c → E2: 0.0 ms of frame, 0.1 ms of GPU); untested on Macs with less bandwidth and off MoltenVK (with these values Vulkan's rules make it exact on every driver)	off (the default); 1 to try where memory bandwidth is short
RX_VK_BENCH_AB=load_clear	the bench's B arm for this change, with RX_VK_BENCH	no such arm	the run with the load ops off, the bench's B block with them on	A: the clears in the pass; B: the load ops; the verdict says whether B's picture is A's	no verdict in the E-series (the field frame never repeats under replay)	for measuring only

Testing

Tested: on macOS 26.7.1, Apple M4 Max:

• At this head (7c6f0e9c, on #704 46eabf24, main e090cc48): test_load_clear passes on arm64 and x86_64 (2,000 sequences
  through the bench's arms, 779 clears made load ops in B; the default off, 1 on), and fails on #704 with this head's test, both
  arches (no keeps_lead); test_perf_lines (178 checks), test_vendor_fixes, test_bank_gen, test_parallel_dry_marks,
  test_same_bind, test_bc_raw_targets, test_pass_hazard pass; test_env_on fails as on main e090cc48 (main's own
  RX_VK_PRESENT_AFTER_SIGNAL, #588); vk_draw.cpp, vk_main.cpp, vk_perf.cpp, vk_replay.cpp compiled whole for both arches,
  0 warnings (reports/PubWave-files/PubLate5/PerfTrio/proof/, compile/load-clear-7c6f0e9c/).
• At the topic's own head (macos/load-clear 0d15403f, on the integration branch I38 3f872b0c, whose changed lines are this
  head's line for line but the default): tests/test_load_clear.py (LS's own code cut from the source against a model of MoltenVK's
  passes: 25 fixed sequences command by command, 6,000 random sequences switch off against on, 2,000 through the bench's arms, the
  value rules over 40 formats, whole_extent over 16 target sets): pass; 19 of 19 mutants killed; full suite 586 passed, 1 failed
  (test_menu_i18n, as on I38), 16 skipped (reports/2026-11-14-load-clear.md §8).
• With #782 and #783 on the integration branch (Integration 39, 7a7726fe): test_load_clear with the widening and the
  mesh batches in its fixture, both arches, among 103 of 103 checks; full suite 588 passed, 1 failed (test_menu_i18n), 16 skipped
  (reports/2026-11-14-integration-2.md §Integration 39).
• The E-series runs on I39 (Measurements) and LC-01 on the GPU.
• Off Apple (unifdef -U__APPLE__, #704 against this head): every changed file differs, the change being gated by the switch
  (reports/PubWave-files/PubLate5/PerfTrio/offapple-load-clear.txt).
• The wave's full suite: see the run_all box.

Not tested: a game run at this head (the E-series ran I39, LoadClear on by default there); a Mac with less memory bandwidth than
an M4 Max; the fold on any driver but MoltenVK; not built on Windows or Linux; its test request covers it.

Must be tested:

• macOS, Apple silicon with less memory bandwidth than an M4 Max (an M1 or M2): the field-dawn hold with RX_VK_LOAD_CLEAR=1 and
  RX_VK_PERF_LINES=1 against the default, expecting L2's fates to show folds and the GPU time per frame to fall by more than the
  noise, if it is to be turned on there (by hand: the owner's hardware).
• Windows and Linux, any GPU: build this PR (head 7c6f0e9c, it carries #704 and what it carries) and #704's head the same way,
  start each to the title, expecting no LOAD_OP_CLEAR device line, every other line as #704's, and the same picture (by hand: test
  request T-XXXX).

Risks

What could break: with RX_VK_LOAD_CLEAR=1 only: a pass's vkCmdBeginRendering moves after its leading clears, which become its
load ops. A command recorded between a held begin and its settle would land outside the pass; the replay settles before every
operation that records, and the test checks it on every sequence. Folds are limited to values and extents LC-01 showed exact. Off
(the default), a pass's rendering info is built in the pass (lead) and its begin recorded at once, as before.

Turning it off: it is off by default (unset RX_VK_LOAD_CLEAR, or 0), or revert the PR.

Compatibility: Nothing changes for players outside macOS. No file in the shader stamp changed; no MKWVK_API_VERSION change.

Reviewer notes

• Review 46eabf24..7c6f0e9c (3 commits): cherry-pick -x of macos/load-clear e0d3af8f and 0d15403f onto #704, then
  7c6f0e9c (own commit: off by default, Integration 40's 2b4b7b8e line for line for these files, but for the RX_VK_BENCH_AB row:
  the same change to its load_clear sentence, on the row without the widen8 and mesh_batch arms the stacked PRs add). The
  picks' conflicts were adjacency only, and only this topic's lines were kept: vk_vendor.h (the field, default and switch beside
  #652's compiler_service and #678's index8_size), vk_main.cpp (beside #678's vk: 8-bit indices line), vk_draw.cpp's
  includes (#680's, #701's, #703's, #705's, #707's, #702's), the docs/development.md row (beside #678's; main e090cc48's newer
  RX_VK_NATIVE_FADD row kept). Every changed line of the picks is the topic's, line for line.
• Depends on #704 (perf_lines.h's L2 counters, perf_lines::bench_b, the perf hooks this PR extends) and what it carries (#679,
  #677, #663 …).
• Stacked on this PR: #782 carries Integration 39's fix A (11fe233a: one pass_keeps for this PR's pass_on and Widen8's;
  the widening recorded right before a held begin), #783 its fix B (7a7726fe: this PR's test with the mesh batches).
• Sibling items (whichever lands second): with #678, the vendor fields and the docs rows sit next to each other (keep both; resolved
  in #782's merge ffb5dfba).
• Owner item: with these values LOAD_OP_CLEAR is exact on every driver by Vulkan's rules; turning it on by default anywhere (a Mac
  with less bandwidth that measures faster, other tile GPUs) is a separate PR.
• Conflicts with main: this head is on main e090cc48; main has moved since, and a delta sweep follows.

Issues

See #420

Checklist

Everyone

• One topic.
• The title is a Conventional Commit (type(scope): description) of the status line's type.
• Why gives the reason plainly and links the issues that fit (the tracker searched).
• Measurements has a verdict, or one line saying why there is nothing to measure.
• Testing names what was tested, what wasn't, and what must still be tested, where and how. (Test requests not open yet.)
• Rebased on current main. (On main e090cc48 through #704; main has moved since: a delta sweep follows.)
• python -I -B tests/run_all.py passes (any failure that also fails on main is named). (On wave P = #783's head 94fca509, which carries this PR, #782 and their bases: 585 passed, 31 failed, 18 skipped against main e090cc48's 544/32/17 on this Mac; the 31 failures are main's own and nothing that passes on main fails; test_collect_wide now passes; 40 new tests pass; reports/PubWave-files/PubLate5/wave/full-P-94fca509/.)
• No game files, dumps, build output or generated files.
• third_party/ untouched (a new folder with a PROVENANCE.md instead).
• The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled.

When it applies

• AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.
• Switches documented (docs/development.md; the README for a menu row) and shipped at their recommended value.
• A game-free test added for the fix, failing on main and passing here.
• A check under tests/ that reads changed source updated in this PR, not weakened.
• Docs updated.

**Type:** perf · **Risk:** low · **Tested on:** macOS 26.7.1, Apple M4 Max (MoltenVK 1.4.2 and MoltenVK main 5189f8de + argbuf-slots), arm64 and x86_64 under Rosetta · **Needs:** macOS, Apple silicon with less memory bandwidth than an M4 Max (a hold with the switch on); Windows and Linux, any GPU (build and start) ## Why Every render pass LS records begins by loading each attachment (`LOAD_OP_LOAD`). On Apple's tile GPUs a load copies the attachment from memory into tile memory, and when the pass's first commands clear the attachments whole, that copy is thrown away at once; the clear is then one more MoltenVK quad, with the state it disturbs. In run #31's field-dawn hold 37.8 passes a frame began with whole clears: 46.8 attachments, 76 MB of the 650 MB loaded a frame (11.7%), loaded only to be overwritten. This PR can make those clears the attachments' `LOAD_OP_CLEAR`, Metal's clear load action, which only fills the tiles, wherever that draws exactly what the clears draw. On an M4 Max it measured under the run-to-run noise, so it ships off, behind `RX_VK_LOAD_CLEAR=1`, for GPUs with less memory bandwidth to try. It is part of the macOS port's performance work (MacPerf §4 #7). See #420. ## What changed With `RX_VK_LOAD_CLEAR=1` a pass whose first operations are whole clears begins with them as its load ops: its `vkCmdBeginRendering` is held while only those clears and changes of state come, and recorded with `LOAD_OP_CLEAR` for the cleared attachments before anything else goes into the command buffer. Everything else is recorded as before; a clear that cannot ride on a load op stays a `vkCmdClearAttachments`. Off by default on every driver (Integration 40, from the E-series): unset or `0`, every pass is recorded as before. - the pass: `gpu/vk/vk_draw.cpp` `Pass` gains `load_clear::Held lead` (the ten `VkRenderingAttachmentInfo` LS gives `vkCmdBeginRendering`, the render area, `held`) and `lead_cmd`; `begin_pass(cmd, at, lead)` fills `lead` instead of stack structs and, for a clear that may fold, holds the begin; `pass_settle` records a held begin, called by `begin_pass` for anything that is not a foldable clear on the same targets, by `end_pass` right before `vkCmdEndRendering`, and by the replay (`draw_pass_settle`). - the replay: `gpu/vk/vk_replay.cpp` `replay()` settles a held pass before any operation outside `keeps_lead` (the clears and the operations that record nothing on the GPU, a set of bits), before it records or admits anything. Nothing reaches the command buffer between a pass's begin and its settle, so the command buffer is the one recorded before, less the folded clears, with those attachments' load ops changed. - the fold, only where it is exact (`gpu/vk/load_clear.h`, new; `clear_color_target`, `clear_depth_stencil`, `lead_why`): every attachment the render area's size and single-sampled, no BC slice or raw view of BC (`Targets::whole_extent` in `targets()`: MoltenVK's clear load action clears each attachment whole, past a smaller render area); the clear of the whole render area and of every channel or the aspects it names (the depth's and the stencil's load ops each their own); a value MoltenVK's two ways write alike (its clear load action and its `vkCmdClearAttachments` quad: no conversion that rounds, per format); the pass's begin not yet recorded. - the switch: `gpu/vk/vk_vendor.h` vendor fix `load_clear`, off on every driver (`x.load_clear = false`), `RX_VK_LOAD_CLEAR=1` turns it on; the device line `vk: a pass's leading whole clears as its attachments' load ops (LOAD_OP_CLEAR): off (RX_VK_LOAD_CLEAR)` on MoltenVK, `on` wherever it is on (`gpu/vk/vk_main.cpp`), and the first pass that folds says so once. The bench's B1 arm: `RX_VK_BENCH_AB=load_clear` (the load ops off for the run, on in the bench's B block: A is the clears in the pass, B the load ops, on the same frame). - the default: off, its own commit `7c6f0e9c`, Integration 40's `2b4b7b8e` for these files (`vk_vendor.h`, the docs rows, `test_load_clear`). - the measurement: #704's `L2` line (`gpu/vk/perf_lines.h`, `gpu/vk/vk_perf.cpp`) also counts what each leading whole clear became: a load op, switched off, other attachment sizes or samples, its value or format, the pass begun already (a frame, and in MB). - docs: `docs/development.md` (the `RX_VK_LOAD_CLEAR` row, the bench arm, `L2`), `gpu/vk/VULKAN.md`. Tests: `tests/test_load_clear.py` (new; it pins the default off on every driver and `1` on); `tests/test_perf_lines.py` reads `begin_pass`'s new signature. ## Measurements **Verdict:** doesn't help measurably on an M4 Max: in the field-dawn hold all on but LoadClear (E4c) and all on (E2) both ran **33.7 fps and 29.67 ms a frame, GPU 19.2 against 19.1 ms**, under the run-to-run noise of 0.17 ms of frame, 0.0 ms of GPU and 0.2 fps (E7), though every one of the 46.8 leading whole clears a frame folded (76 MB of loads skipped). So it is off by default. It is exact on the GPU where probed: LC-01 drew every folded clear's attachments texel for texel as the clears did, on both MoltenVK builds. Baseline: main does not start on macOS. The E-series runs are the integration build I39 `7a7726fe` (this PR's code with #782's and #783's, and the other integrated macOS topics; LoadClear on by default there), arm64, one run each: before is E4c, all on but LoadClear (`RX_VK_LOAD_CLEAR=0`), after is E2, all three on. This head's LoadClear functions are I39's text; only the default differs (off here, as Integration 40 made it). LC-01 records, in a standalone MoltenVK process per build, each pass as LS records it with the switch off (A: `LOAD_OP_LOAD`, then `vkCmdClearAttachments`, then a draw over the cleared values) and on (B: `LOAD_OP_CLEAR`, the same draw), with the PR's own `load_clear.h` deciding what folds, and compares every texel of every layer, level, slice and aspect read back. Hardware: macOS 26.7.1 (25G241), Apple M4 Max (40-core GPU), MoltenVK main 5189f8de + argbuf-slots (`c170a745`; LC-01 also MoltenVK 1.4.2), 16-core CPU, 64 GB Scene: the E-series: the field-dawn save, slot 00, entered by START and A×5, then 600 s held in play, windowless; the instrument set on in every run (`RX_VK_PERF_LINES=1`, `RX_VK_GPU_TIMELINE=1`, the present trace); quiet required before each start, the other lanes paused. LC-01: 373 cases (every colour and depth-stencil format LS can fold, each foldable value; an array layer, a mip level with and without depth, a 3D slice, the depth alone, the stencil alone, two and three colours; three cases of other attachment sizes) and 64 sweeps of the values LS refuses Method: the E-series: the medians of the hold's 120-present periods from +20 s (E4c 162 periods, E2 141; the bench's and the chain's periods left out of E2); one run each, the noise from E7 (a repeat of E2); runs `mac-arm64-7a7726fe-06` (E4c), `-02` (E2), `-08` (E7); reports `reports/2026-11-15-mac-perf-E4c.md`, `-E2.md`, `-E7.md`, the series `reports/2026-11-15-mac-perf-before-after.md`. LC-01: `reports/LoadClear-files/lc01/runs-lc01-arm64/result.json` (`787dc0cf…`), released once on arm64 by the parent; review `f00090ed…`, build `00ac1f0a…`, plan `a651ba74…` How to reproduce: the series' `ls-run.py prepare`/`run` commands and common switches (`reports/2026-11-15-mac-perf-runs-i39.md` §2-§3), with `RX_VK_LOAD_CLEAR=0` for the before run; LC-01's release command (`reports/2026-11-14-load-clear.md` §7) | field-dawn hold, 600 s | Before, RX_VK_LOAD_CLEAR=0: integration @ 7a7726fe (E4c) | After, load_clear on: integration @ 7a7726fe (E2) | |---|---|---| | fps (median [min-max] of the periods) | 33.7 [32.2-34.3] | 33.7 [32.7-34.4] | | GPU ms a frame | 19.2 [18.7-19.8] | 19.1 [18.6-19.8] | | frame ms | 29.67 | 29.67 | | `L2` leading whole clears a frame | 46.8, every one `switched off` | 46.8, every one a load op (76.0 MB of loads skipped) | ``` E4c: 162 periods; fps 33.7 [32.2-34.3]; frame ms 29.6736; GPU ms 19.2 [18.7-19.8] E2: 141 periods; fps 33.7 [32.7-34.4]; frame ms 29.6736; GPU ms 19.1 [18.6-19.8] E7: run-to-run noise: 0.17 ms of frame (29.50 against 29.67), 0.0 ms of GPU, 0.2 fps ``` | LC-01, each MoltenVK build | Before, A (LOAD and the clears): load-clear @ e0d3af8f | After, B (LOAD_OP_CLEAR): load-clear @ e0d3af8f | |---|---|---| | Q1: every foldable format and value | 327 cases drawn (44 skipped: formats the GPU does not render to) | equal to A in all 327, every texel | | Q3: layer, mip, 3D slice, depth alone, stencil alone, two and three colours | drawn | equal to A | | Q4: a depth smaller than the colour (`X.smaller_depth`) | drawn | differs in 512 texels, as predicted: LS does not fold it (`whole_extent`) | | Q2: the values LS refuses (sweeps) | drawn | differ where predicted (half floats 101-197 of 206, D16 195 of 300, 32-bit integers past 2^24) | ``` LC-01 result.json: all_as_predicted 1; main-argbuf-slots and v1.4.2: q1_exact 1 (327 cases), q3_exact 1, q4_as_predicted 1, controls_equal 1, mvk_errors 0 ``` Run #31's prediction (about 0.2-0.5 ms of the GPU frame if most clears fold) did not hold on the M4 Max: all folded, 0.1 ms of GPU. A Mac with less memory bandwidth may gain more [inference, not measured]. ### Flags and settings One new switch, and one new arm of #704's bench switch; off by default. | Flag | What it is | Before this PR | After this PR | Effect | Helps / doesn't help / untested | Recommendation | |---|---|---|---|---|---|---| | `RX_VK_LOAD_CLEAR=0` or `1` | a pass's leading whole clears as its attachments' load ops | no switch: `LOAD_OP_LOAD` and `vkCmdClearAttachments` | unset or `0`: off; `1` on, on any driver | fewer attachment loads and clear quads where a pass begins with whole clears; the same pixels | exact in LC-01 (both MoltenVK builds); under the noise on an M4 Max (E4c → E2: 0.0 ms of frame, 0.1 ms of GPU); untested on Macs with less bandwidth and off MoltenVK (with these values Vulkan's rules make it exact on every driver) | off (the default); `1` to try where memory bandwidth is short | | `RX_VK_BENCH_AB=load_clear` | the bench's B arm for this change, with `RX_VK_BENCH` | no such arm | the run with the load ops off, the bench's B block with them on | A: the clears in the pass; B: the load ops; the verdict says whether B's picture is A's | no verdict in the E-series (the field frame never repeats under replay) | for measuring only | ## Testing **Tested:** on macOS 26.7.1, Apple M4 Max: - At this head (`7c6f0e9c`, on #704 `46eabf24`, main e090cc48): `test_load_clear` passes on arm64 and x86_64 (2,000 sequences through the bench's arms, 779 clears made load ops in B; the default off, `1` on), and fails on #704 with this head's test, both arches (no `keeps_lead`); `test_perf_lines` (178 checks), `test_vendor_fixes`, `test_bank_gen`, `test_parallel_dry_marks`, `test_same_bind`, `test_bc_raw_targets`, `test_pass_hazard` pass; `test_env_on` fails as on main e090cc48 (main's own `RX_VK_PRESENT_AFTER_SIGNAL`, #588); `vk_draw.cpp`, `vk_main.cpp`, `vk_perf.cpp`, `vk_replay.cpp` compiled whole for both arches, 0 warnings (`reports/PubWave-files/PubLate5/PerfTrio/proof/`, `compile/load-clear-7c6f0e9c/`). - At the topic's own head (`macos/load-clear` `0d15403f`, on the integration branch I38 `3f872b0c`, whose changed lines are this head's line for line but the default): `tests/test_load_clear.py` (LS's own code cut from the source against a model of MoltenVK's passes: 25 fixed sequences command by command, 6,000 random sequences switch off against on, 2,000 through the bench's arms, the value rules over 40 formats, `whole_extent` over 16 target sets): pass; 19 of 19 mutants killed; full suite 586 passed, 1 failed (`test_menu_i18n`, as on I38), 16 skipped (`reports/2026-11-14-load-clear.md` §8). - With #782 and #783 on the integration branch (Integration 39, `7a7726fe`): `test_load_clear` with the widening and the mesh batches in its fixture, both arches, among 103 of 103 checks; full suite 588 passed, 1 failed (`test_menu_i18n`), 16 skipped (`reports/2026-11-14-integration-2.md` §Integration 39). - The E-series runs on I39 (Measurements) and LC-01 on the GPU. - Off Apple (`unifdef -U__APPLE__`, #704 against this head): every changed file differs, the change being gated by the switch (`reports/PubWave-files/PubLate5/PerfTrio/offapple-load-clear.txt`). - The wave's full suite: see the run_all box. **Not tested:** a game run at this head (the E-series ran I39, LoadClear on by default there); a Mac with less memory bandwidth than an M4 Max; the fold on any driver but MoltenVK; not built on Windows or Linux; its test request covers it. **Must be tested:** - macOS, Apple silicon with less memory bandwidth than an M4 Max (an M1 or M2): the field-dawn hold with `RX_VK_LOAD_CLEAR=1` and `RX_VK_PERF_LINES=1` against the default, expecting `L2`'s fates to show folds and the GPU time per frame to fall by more than the noise, if it is to be turned on there (by hand: the owner's hardware). - Windows and Linux, any GPU: build this PR (head `7c6f0e9c`, it carries #704 and what it carries) and #704's head the same way, start each to the title, expecting no `LOAD_OP_CLEAR` device line, every other line as #704's, and the same picture (by hand: test request T-XXXX). ## Risks **What could break:** with `RX_VK_LOAD_CLEAR=1` only: a pass's `vkCmdBeginRendering` moves after its leading clears, which become its load ops. A command recorded between a held begin and its settle would land outside the pass; the replay settles before every operation that records, and the test checks it on every sequence. Folds are limited to values and extents LC-01 showed exact. Off (the default), a pass's rendering info is built in the pass (`lead`) and its begin recorded at once, as before. **Turning it off:** it is off by default (unset `RX_VK_LOAD_CLEAR`, or `0`), or revert the PR. **Compatibility:** Nothing changes for players outside macOS. No file in the shader stamp changed; no `MKWVK_API_VERSION` change. ## Reviewer notes - Review `46eabf24..7c6f0e9c` (3 commits): `cherry-pick -x` of `macos/load-clear` `e0d3af8f` and `0d15403f` onto #704, then `7c6f0e9c` (own commit: off by default, Integration 40's `2b4b7b8e` line for line for these files, but for the `RX_VK_BENCH_AB` row: the same change to its `load_clear` sentence, on the row without the `widen8` and `mesh_batch` arms the stacked PRs add). The picks' conflicts were adjacency only, and only this topic's lines were kept: `vk_vendor.h` (the field, default and switch beside #652's `compiler_service` and #678's `index8_size`), `vk_main.cpp` (beside #678's `vk: 8-bit indices` line), `vk_draw.cpp`'s includes (#680's, #701's, #703's, #705's, #707's, #702's), the `docs/development.md` row (beside #678's; main e090cc48's newer `RX_VK_NATIVE_FADD` row kept). Every changed line of the picks is the topic's, line for line. - Depends on #704 (`perf_lines.h`'s `L2` counters, `perf_lines::bench_b`, the perf hooks this PR extends) and what it carries (#679, #677, #663 …). - Stacked on this PR: #782 carries Integration 39's fix A (`11fe233a`: one `pass_keeps` for this PR's `pass_on` and Widen8's; the widening recorded right before a held begin), #783 its fix B (`7a7726fe`: this PR's test with the mesh batches). - Sibling items (whichever lands second): with #678, the vendor fields and the docs rows sit next to each other (keep both; resolved in #782's merge `ffb5dfba`). - Owner item: with these values `LOAD_OP_CLEAR` is exact on every driver by Vulkan's rules; turning it on by default anywhere (a Mac with less bandwidth that measures faster, other tile GPUs) is a separate PR. - Conflicts with main: this head is on main e090cc48; main has moved since, and a delta sweep follows. ## Issues See #420 ## Checklist ### Everyone - [x] One topic. - [x] The title is a Conventional Commit (`type(scope): description`) of the status line's type. - [x] Why gives the reason plainly and links the issues that fit (the tracker searched). - [x] Measurements has a verdict, or one line saying why there is nothing to measure. - [x] Testing names what was tested, what wasn't, and what must still be tested, where and how. (Test requests not open yet.) - [ ] Rebased on current `main`. (On main e090cc48 through #704; main has moved since: a delta sweep follows.) - [x] `python -I -B tests/run_all.py` passes (any failure that also fails on `main` is named). (On wave P = #783's head `94fca509`, which carries this PR, #782 and their bases: 585 passed, 31 failed, 18 skipped against main e090cc48's 544/32/17 on this Mac; the 31 failures are main's own and nothing that passes on main fails; `test_collect_wide` now passes; 40 new tests pass; `reports/PubWave-files/PubLate5/wave/full-P-94fca509/`.) - [x] No game files, dumps, build output or generated files. - [x] `third_party/` untouched (a new folder with a `PROVENANCE.md` instead). - [x] The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled. ### When it applies - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. - [x] Switches documented (docs/development.md; the README for a menu row) and shipped at their recommended value. - [x] A game-free test added for the fix, failing on `main` and passing here. - [x] A check under `tests/` that reads changed source updated in this PR, not weakened. - [x] Docs updated.

mooki added 80 commits 2026-11-15 11:55:47 +00:00

fix(runtime): preserve guest pages when discarding on macOS ... c71ae93353

Reserve the fixed guest window with a non-overwriting Mach mapping. Reclaim
complete owned host pages with fresh anonymous mappings and zero only the
partial-page edges, preserving adjacent 4 KiB guest pages on 16 KiB hosts.
Use the actual host granule for non-guest allocation/protection and report
macOS discard failures instead of returning success.

Add game-free source-linked arm64/Rosetta tests covering collisions, repeated
free/recommit, unaligned ranges, failure paths and resident-memory reduction.
Keep Linux and Windows paths unchanged. Build-list integration is separate.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(runtime): retain guest-page rounding in Darwin memory wrappers ... 23ae3872be

Keep the POSIX layer's outward 4 KiB MEM_DECOMMIT and inward 4 KiB MEM_RESET
contracts above the byte-exact Darwin helper. Check arithmetic before rounding,
and keep separate real-memory assertions for helper and wrapper ranges.

Default the Darwin regression test to the native architecture. Both CPU routes
remain explicitly selectable for acceptance without requiring Rosetta or arm64
execution from every ordinary macOS suite run.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(runtime): cut main's int hos_commit in the Darwin guest-memory test ... d4d89421d9

Main's #324 (a4c77174) made hos_commit in host/mkw_glue.c return whether
the range was committed, so its definition now starts "int hos_commit(".
The Darwin guest-memory test cut the function by its old "void" head and
stopped with a ValueError on main's source, on this branch and on main alike,
before compiling anything. Cut it by the new head.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(runtime): follow main's HLE call counts in the Darwin guest-memory fixture ... ae36a96bb0

Main's HOS_HLE_STATS (#153) made HLE() a counting wrapper: a static
HleCount named by its first field, hos_hle_stats and hos_hle_counted. The
fixture's cut of hos_mem.c now gets those two (off: each body runs
directly) and main's -Wno-missing-field-initializers, as main's other
C fixtures pass. The guest-memory checks are unchanged.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(runtime): translate macOS fault contexts on both CPU routes ... 4ddb671f77

Decode Darwin signal contexts and access faults for arm64 and x86_64,
copy supported context edits back on signal return, and prevent recursive
callback dispatch when a fault handler itself faults. Keep the non-Apple
bridge text and the x64-shaped CONTEXT layout unchanged.

Add source-linked, game-free owned-memory fault tests for both targets,
including page repair, PC redirection, prior handlers and an inaccessible
handler stack. Document the native PC/SP contract and remaining limits.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(runtime): keep Darwin fault checks native by default ... c60d8d05d1

Do not make the ordinary Apple Silicon suite depend on Rosetta. Default
to the running Python process's architecture and retain explicit repeated
--arch options for dual-route qualification. Document that command and
count only files when reporting crash-report directory changes.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(runtime): classify the fault bridge's fatal _exit in main's exit inventory ... 1b028c8419

Main's crash records (#75) added an exhaustive inventory of host process
exits to test_crash_record.py, which scans every host/**/*.c on every
platform. darwin_fault.c's nested-fault _exit(128 + sig) is fatal, so it
is listed as one and exempt from the clean-end mark, like hos_thread.c's
ExitProcess(3). Without it the inventory fails on Windows and Linux too.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): add macOS Vulkan loader and device admission ... 624e909b79

Select a local Vulkan loader or MoltenVK with RX_VK_LIBRARY and keep SDL and
renderer startup on that same library. Capture SDL's instance extensions on
the window thread, enable advertised portability extensions, and name missing
admission requirements and VkResults without changing Windows or Linux paths.

Keep the non-Apple host API at 21 for existing prebuilt renderers; append the
two macOS callbacks only in Apple's version 22. Cover pure decisions, legacy
platform expectations, ABI layout, and the SDL2/SDL3 bridge without a GPU.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): use semantic admission contracts ... 45df17e92d

Remove the SDL source-string wiring and source-index ordering assertions while
keeping the sixteen compiled behavioral cases and surface adapter compile.
Replace the comment-sensitive API header hash with fixed baseline field names,
byte offsets, versions and sizes, retaining typed Apple callback checks and
both platforms' actual admission decisions and negative controls.

Production code is unchanged. Prior full-suite and device receipts remain
bound to 818e33b1; only the affected game-free tests were rerun.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): follow main's LoadLibraryA and float-add choice in the admission fixture ... 1b306e47e5

The fixture cuts LoadLibraryA and create_device out of main's source. Since
#34, LoadLibraryA keeps dlerror()'s text in load_error, and since #191
native_fadd_wanted also takes the device ID and driver version, so the cut
code no longer compiled against the fixture's stand-ins on this base. Give
it load_error and a dlerror stand-in, and the four-argument stand-in. No
expectation changed.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): pin the admission ABI at main's version 22 and macOS 23 ... 9aa916d0ce

Main's custom output size (#51) took MKWVK_API_VERSION 22 on every
platform and appended take_text at offset 352. The macOS loader
callbacks now follow it: the rebased header says 23 on Apple and 22
elsewhere. The fixed contract follows: non-Apple version 22, size 360
and 45 named offsets (take_text added); Apple version 23, size 376,
vulkan_library and vulkan_extensions at 360 and 368.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(macos): storage buffer tables get descriptor sets on MoltenVK ... 9778fd0ee8

MoltenVK 1.4.2 binds a push descriptor set's resources one by one, never
through a Metal argument buffer (MVKDescriptorSet.mm pickArgumentBufferMode
returns Off for every push layout, and the pipeline names that set to
SPIRV-Cross as discrete). The SPIR-V back end declares a storage buffer
table (count > 1: Info::global_tables, the task and mesh programs'
c0+0x880 table) as one fixed-size array per typed view, all at one binding
(DefineSsbos; any global memory access uses the U32, U32x2 and U32x4
views), and SPIRV-Cross 6c09849f aliases such arrays only inside an
argument buffer. So every pushed program with a table failed: 256
"Attempting to alias same binding with a descriptor array which is not
implemented through argument buffers" errors and as many failed compute
pipelines in the first Mac start (mac-arm64-a5f2d6f9-01).

vk_vendor.h gets a fix, set_for_tables, on by default only for
VK_DRIVER_ID_MOLTENVK, with RX_VK_SET_FOR_TABLES to force it either way.
ensure_layout's push decision moves into layout_pushes(): with the fix on,
a set that has an array of storage buffers is not pushed but allocated
from the recording's pool, the path programs over the push limit already
take, where MoltenVK's sets are Metal 3 argument buffers. Every other set
is still pushed, the bindings are the same either way, and the SPIR-V,
the translator stamp and the program cache are unchanged. The vendor
fixes' log line at device creation says whether it is on, and a MoltenVK
run with MVK_CONFIG_USE_METAL_ARGUMENT_BUFFERS=0 is warned that these
programs cannot translate.

tests/test_set_for_tables.py compiles vk_draw.cpp's layout code (kPushSize
to the end of ensure_layout) on made-up programs and devices: on MoltenVK
a set with a table (compute, task, fragment) is a set, a set without one
(texture arrays and a constant buffer of count 2 included) is pushed;
ten other drivers push as before; the switch works both ways. It fails
on the base with 7 of 79 cases. tests/test_vendor_fixes.py adds a MoltenVK
device and the switch, and checks that the fix stays out of the
translation's profile and stamp. Windows and Linux: same layouts as
before; the log line gains one field.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 0c2fea2a303a5158d12ebc87bcd28467486486ef)

fix(macos): programs reading storage buffer lengths get descriptor sets on MoltenVK ... ef43dffcc8

MoltenVK 1.4.2 writes the buffer sizes SPIRV-Cross reads for OpArrayLength
in a discretely bound set only in bindDescriptorSets
(MVKCommandEncoderState.mm), never in pushDescriptorSet, and it binds every
push set discretely. So every pushed storage buffer's length read 0
whatever its range, while an ordinary set's (a Metal 3 argument buffer
with its own sizes) was exact (GPU probe,
reports/2026-11-14-null-range-probe.md). The back end measures storage
buffers in its global memory search (DefineGlobalMemoryFunctions:
native_global_bounds' candidates and every single storage buffer), so in a
pushed program no buffer ever held a global address: global stores and
atomics were dropped, the atomics returned 0, and loads the page table
does not serve read 0.

vk_vendor.h gets set_for_lengths, on by default only for
VK_DRIVER_ID_MOLTENVK, with RX_VK_SET_FOR_LENGTHS to force it either way.
The new gpu/vk/spirv_lengths.h (no Vulkan, as fallback_size_policy.h)
walks a module's words for OpArrayLength.
Each Stage gets array_length, noted beside its phis wherever the program
cache handles its SPIR-V: program_cache_save after translation (before the
program is published), a whole record read, and complete_from for a
compute program read without its code. The words are not kept anywhere a
layout could scan them later, and a light compute program's layout is made
before its code is read, so until then (array_length -1) a stage counts as
reading a length where it uses global memory: outside segmented variants,
which the cache never keeps, that is the only place the back end measures
a storage buffer. layout_pushes(p, bindings, total) then refuses a push
for any program with a length in any stage. Every other program, every
other driver, the SPIR-V, the translator stamp and the program cache are
unchanged.

The vendor fixes' log line says whether it is on, and at close one line
says how many programs it gave a set (and how many of those were decided
before their code was read).

tests/test_set_for_lengths.py builds real back end modules
(tests/set_for_lengths_ir.cpp: compute and vertex global memory with a
single storage buffer, 18 OpArrayLength each; global memory with nothing to
search; direct storage in compute, vertex and fragment stages, none), then
compiles vk_draw.cpp's layout code with their descriptors and notes:
138 layouts on MoltenVK and ten other drivers, with the switch unset, 0
and 1, stages not read yet, and the bindings equal pushed or not; plus the
renderer's scan on the same modules and on words that are not a module,
and the count kept for the log. test_vendor_fixes.py covers the new fix
on 17 devices, both ways of the switch, and its absence from the
translation's profile and stamp. The program cache fixtures
(program_cache_one_variant_test.cpp, program_cache_cleanup_test.cpp), which
compile vk_progcache.cpp's functions, include spirv_lengths.h beside
fallback_size_policy.h.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit bf5ca961c91479da13e8923387012ef6dc7fb2c7)

fix(macos): stages with more than 8 read_write textures get descriptor sets on MoltenVK ... bff8bdc326

On MoltenVK 1.4.2 a pushed set is bound outside an argument buffer
(MVKDescriptorSet.mm pickArgumentBufferMode: push layouts are Off), and
there Metal's compiler takes at most 8 read_write textures in one function:
TexelDescProbe TD5 (40 storage images and 9 storage texel buffers in one
compute stage) failed pipeline creation pushed with "number of textures
with read_write access exceeds maximum supported (8)", and worked in an
ordinary set (Metal 3 argument buffers). The device reports
maxPerStageDescriptorStorageImages 256 (MoltenVK sets
maxPerStageStorageTextureCount to its argument buffers' texture count), so
nothing in Vulkan's limits stops it.

Which descriptors count: SPIRV-Cross 6c09849f declares a storage image
(sampled 2), storage texel buffers included as texture_buffer since
MoltenVK sets texture_buffer_native, access::read_write unless the
variable is decorated NonWritable or NonReadable (spirv_msl.cpp
image_type_glsl), and it never adds those decorations itself. The eden back
end decorates none (DefineImages, DefineImageBuffers), although Info knows
is_read and is_written. So every storage image and storage texel buffer
counts, by descriptor; sampled images and uniform texel buffers do not.
The limit is per Metal function, that is per stage.

layout_pushes now refuses a push under the existing set_for_tables fix
when any stage's storage images and storage texel buffers, counted by the
bindings' descriptorCount, exceed 8. It is the same cause and remedy as a
storage buffer table (a shape a pushed set cannot take on MoltenVK, where
the pipeline fails; an ordinary set takes it), and RX_VK_SET_FOR_TABLES
covers both, so the vendor fixes' log line is unchanged. ensure_layout
records where each stage's bindings begin, so the mesh fallback's task
and mesh stages, both compute, are counted apart. The programs moved for
it are counted and logged at close (draw_layouts_report). Every other
driver, the SPIR-V, the translator stamp and the program cache are
unchanged.

tests/test_set_for_tables.py gains 55 layouts: 8 and 9 storage images,
5+3 and 5+4 images and texel buffers, 9 texel buffers, an image array of
8 and of 9, graphics stages of 8+8 (pushed) and 0+9, the fallback's task
and mesh stages of 5+5 (pushed) and 4+9, and 41 read-only textures beside
8 storage images (pushed), on MoltenVK; the 9s on the ten other drivers
(pushed); the switch both ways; and the count kept for the close. It
fails on 7b82e199. test_vendor_fixes.py follows the new call and checks
the count reaches the close's log.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 162aca1b3cce50adde9250f14b209a4e5b4c1331)

Merge mac/fault-contexts (#436) into mac/guest-16k ... fc1857fc3b

This PR needs the macOS fault-context PR (#436); its commits are kept as they are.
The one conflict, host/posix/win32_posix.c's Apple include block, keeps both
darwin_mem.h (the guest-memory PR's) and darwin_fault.h (#436's).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): plan where an oversize guest span is cut into import segments ... 697a993cbd

A connected guest span that the views of one operation need can be larger than
one device allocation (41,747,087,360 bytes on MoltenVK 1.4.2, M4 Max). It is
then imported as disjoint, adjacent segments of at most the cap.
gpu/vk/span_segments.h (rules only, no Vulkan) chooses the cuts: multiples of
the import granularity, the fewest segments, the least cost of the views a cut
splits (gathers by their bytes, then uniform, texel and storage buffers), a
cost for moving a previous cut, and the lowest addresses on ties. It maps any
guest range onto the segments, and splits a texel view at cuts and at
maxTexelBufferElements, reporting 12-byte seams. Invalid input gets an
explicit refusal. Nothing uses it yet: vk_memory.cpp takes it later.

tests/test_span_segments.py checks it against brute force at caps of 48 KiB,
64 KiB, 96 KiB and 1 MiB with 16 KiB granularity, plus the 48 GiB example, the
512 GiB guest window and the top of the u64 range (--arch arm64 / x86_64 on
macOS).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 11e53b66ffbdddd0cfeaa7a043361a337299d855)

fix(macos): serialize guest backing and native-page watch lifetimes ... cf1f48feb2

Keep address-region and page frees owned until discard or in-place zeroing
finishes. Route guest mutations from each image to one executable-owned
registry, and pin complete imported backing through duplicate and overlapping
GPU memory lifetimes without querying the renderer under ownership locks.

Use real Mach query ranges and protections, preserve byte-exact 4 KiB guest
ownership and 64 KiB commit semantics, and handle all native watch subpages
with checked fault recovery and conservative dirty coverage. Replace racy
bulk dirty-byte reads with same-granule atomics.

Add game-free native arm64/Rosetta fixtures for two-image dispatch, actual
faults, aliases, failures, and deterministic free/reallocate/import races.
Carry prior RSS evidence rather than repeat unchanged experiments.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): let the segment planner add segments that pay for themselves ... aae09e20f7

With the fewest segments, a large gathered view can be cut on every draw
although one more segment would keep it whole: a 32 GiB index extent at
[8, 40) GiB in a 48 GiB domain at the 41,747,087,360-byte cap is split by
every two-segment cut list. plan() now weighs any number of segments, each
one beyond the fewest at Weights::segment (2^32 bytes-equivalent, above a
move), and returns the least cost, then the fewest segments, then the lowest
cuts. Every cut of that plan is a mark (the first multiple of g at or above a
demand's end, or a previous cut) or ceil(next cut - cap), so the search runs
over those candidates and is exact over [k_min, k_min + min(marks,
C0 / segment)]; a least plan over 4096 segments is refused. Plan::extra
reports the added segments.

The test brute-forces the same range of segment counts, checks a DP over
every count, and adds the 32 GiB extent (three segments, nothing gathered).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 189d0636cb5e90e0d98de397d0b9676a5f8354c2)

fix(macos): retain pending watch faults across protection changes ... 79d253120a

Keep a cancelled-writable watch state so an already pending diagnostic fault
is acknowledged without rearming or changing permissions. Check the real
mapping before acknowledgement; intentional read-only/no-access stays closed.

Freeze affected watch pages once, preflight the complete native range against
Mach maximum permissions, and issue one protection operation. Preserve prior
watch states when either preflight or the native operation fails.

Exercise the real Win32 vectored dispatcher, signal-handler retention and the
next genuine watch fault. Cover permission failures both before preflight and
after a deterministic barrier, with real Mach maximum-protection changes and
no manufactured VM results, on native arm64 and Rosetta.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(macos): make the guest-memory changes Apple-only; Windows and Linux keep main's code ... 10d05689a1

The two commits before this one also changed Windows and Linux: the
dirty-page scan read its marks with byte atomics, the renderer's zeroed
guest memory was marked written, an empty range consumed no mark,
allocation failures in the scan and in the address-region bookkeeping
were fatal, hos_region_alloc refused a bad size or alignment,
hos_mark_written clamped its length, and VirtualQuery checked its
pointer and set last_error. Nobody has timed the scan there, so each of
these is now inside #if defined(__APPLE__), with main's code in the
#else. unifdef -U__APPLE__ of every host/ file this branch changes
equals main's byte for byte, and so do their preprocessed Linux and
Windows token streams. Issue #432 is the version for every platform.

test_collect_wide's fixture compiles the scan the platform builds: byte
atomics, and an empty range consuming nothing, on macOS; main's word
loads elsewhere.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

docs(macos): describe the guest window's one owner, import pins and native-page watches ... fceb934f11

The guest-memory section still described the groundwork alone: it said
the write watch and VirtualQuery were separate work and that the
region allocator's unlock-before-discard race remained. Describe the
executable's single owner and how each image reaches it, frees that
keep their range until the bytes are gone, import pins, the real
VirtualQuery and VirtualProtect, the native-page watch and the atomic
dirty scan, that all of it is macOS only (issue #432), and the test's
two narrower modes.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(vk): scan the import table as far back as its largest live import ... e37808fba1

Every scan down vk_memory.cpp's import table by start (pool_lookup's
covering import and merge plan, bg_drain, prefetch_chunks,
import_in_background, memory_note_gpu_write and memory_guest_imported)
stopped once an import ended more than 1 GiB below its address, on the
belief that no import is larger than that. pool_lookup's own merge escape
leaves an import above 1 GiB over the smaller ones inside it ("left
overlapping"), and the device allows imports up to max_import_size: a scan
stopped at a small import inside a large one, missed the large one, and
the same bytes were imported again, two memories over the same bytes.

import_reach.h counts the live imports by span class; its bound is above
every live span, within twice the largest. imported_add and imported_take,
now the table's only ways in and out, keep the count. A scan stops only
where no import further down can reach its address, so it also stops
sooner than before when the imports are small. The covering-import checks
in pool_lookup and import_in_background look past the last import at or
below the range for the same reason.

Tests: test_import_reach.py runs the actual scans with a 4 GiB import over
a 16 MiB chunk, left by the merge escape (imports are recorded ranges: no
payload, no Vulkan, no game). Its lookup, query, note, drain, background
and prefetch cases fail on main 5e072fc8 and pass here, and a source check
keeps every change to the table in the two helpers. test_import_windows
runs pool_lookup with the helpers; test_mirror_gpu_only names the new
publication point.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): the readback refresh fixture runs memory_note_gpu_write with the import table's helpers ... 639ab978be

test_parallel_rb_refresh cuts memory_note_gpu_write out of vk_memory.cpp,
whose scan now stops at the live imports' reach: the fixture includes
import_reach.h, cuts live_reach, imported_add and imported_take beside it,
and puts its pool in the table through imported_add. It failed to build
at 11e9ee44 ("use of undeclared identifier 'import_reach'").

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): descriptors bound as pieces in the SPIR-V back end and program variants ... 68a6ea49c7

A guest span larger than one device allocation (the cap on MoltenVK/M4 Max is 41,747,087,360 bytes) is bound as
pieces: one descriptor per allocation segment, at the descriptor's own binding (reports/2026-11-14-full-range-design.md
sections 4.1-4.4). This is the shader half; the draw side (descriptors, layout, gathers) comes later.

- runtime_info.h: SegEntry {Storage|Table|TexBuf|ImgBuf, desc, pieces} and RuntimeInfo::segmented.
- Back end: a segmented storage buffer is an array of pieces at its binding (a table count * pieces, entry k's piece j
  at k * pieces + j); its loads, stores and atomics call DontInline helpers made after main, which find the piece by
  the pieces' OpArrayLength in order, with constant indices. The global memory fallbacks sum the pieces' lengths for a
  segmented descriptor's bound end and branch to its piece (single, native candidate, table written out and in the
  loop). Texel and image buffers are arrays too, their fetches, reads, writes and atomics through helpers that pick
  the piece by OpImageQuerySize, and a texel buffer's size query sums the pieces. ImageQuery was declared for every
  module already. With RuntimeInfo::segmented empty nothing new is made: the SPIR-V is byte-identical.
- vk_shader: SegSig (entries, hash) on Program, seg_sig and seg_variant; program_segmented finds or translates the
  variant of a plain program with plain's texture questions and answers and checks its descriptors are plain's;
  program_pick, mesh_program_locked, program_find_plain, loaded_variants and portable_variant keep the plain and
  segmented variants apart; every stage's RuntimeInfo::segmented comes from Program::seg.
- vk_progcache: a segmented variant is never written to the program cache nor listed among the loaded programs.
- tests/test_seg_shader.py + tests/seg_shader_ir.cpp: synthetic IR programs through the real back end, plain and
  segmented; structure checks, spirv-val (Vulkan 1.3, uniform buffer standard layout), and with --base-tree a byte
  comparison of the plain modules against an older tree. Five source-reading checks follow the new Program field and
  the atomics' new branch.

The translator stamp changes with this (eden/, vk_shader.cpp, sizeof(Program)): every program cache is rebuilt once.
Land it in the same release as the vk_api.h change that already does so.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit b7928876dab846485f667c1c934d34d25ecef035)

fix(vk): MslArrayPtr: no pointer into a storage buffer array is passed to a function ... 407a0cfc93

On MoltenVK a segmented storage program with a compare-exchange loop (an
increment, decrement or float add the device does not have) failed pipeline
creation: NullProbe revision 02's NS2, "no matching function for call",
&ssbo0_4[0u] given to a device _66 * parameter. SPIRV-Cross's MSL (6c09849f,
MoltenVK 1.4.2) passes a storage buffer pointer to a function by value, and
takes the address of the argument's expression: for an element of a buffer
array, which MSL holds as an array of pointers, that is the address of the
pointer (reports/2026-11-14-msl-arrayptr.md).

A descriptor bound as pieces now has its loops written out in the piece,
inside the segmented helper (CasLoopHere, CasStoreHere: CasLoop's loop,
attempt for attempt, its variables made first in the helper), instead of
calling the CasLoop function or write_storage_cas_loop_func with a pointer
into the pieces' array. Every other program is as before byte for byte.

Tests: test_seg_shader.py checks that no module passes a pointer into a
buffer descriptor array to a function. The two checks that cut the back
end's source follow the new loop and parameter.

Ported from macos/msl-arrayptr, where it was committed on NullShader, to the
pieces alone: here every storage access that reaches a segmented helper is
in pieces, so the loops are always written out. Left for the NullShader
change: the bound-whole branch that still calls the function with its
block, the guarded helper's cas_loop, and the tests that need NullShader or
tests/spirv_run.h (the CPU runs of the loops and of a storage buffer table,
--changed-segmented and the *.cpurun comparison).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit bd396010887afab9520181009db1b56a4cc425e3)

feat(vk): texel codecs in the SPIR-V back end and program variants ... b78e421a97

A texel buffer whose format the GPU gives no texel buffer view (MoltenVK 1.4.2 on Apple GPUs: R32G32B32_SFLOAT, the
R8G8B8 formats and the 16-bit packed ones) is read and written through an R8, R16 or R32 UINT view of the same bytes,
with the format's decode and encode in the shader (reports/2026-11-14-texel-formats-design.md section 4).

- Back end: RuntimeInfo::texel_codecs (TexelEntry: a texture or image buffer descriptor and its Shader::TexelCodec),
  honoured with Profile::support_texel_codecs only. Carrier images, per-codec decode and encode functions of integer
  operations alone (normalized components through tables of the exact floats, writes rounded to nearest even from the
  float's bits), whole-texel bounds (reads past the last texel 0, writes dropped, the size the elements over k), an
  integer format's own view read as U32 or S32 texels, and composition with the segmented piece arrays (the helpers
  take each descriptor's own image type). Typed storage accesses and atomics keep the view of their declared format:
  a codec there is refused. With no codec the SPIR-V is byte for byte as before.
- Renderer: gpu/vk/texel_codec.h, the plan shared by the variants and (next) the views; texel_codecs_on (macOS, from
  the format query, RX_VK_TEXEL_CODECS); each texel buffer's codec a variant answer (desc | Program::kTexelAnswer, so
  the program cache and the pipeline state list carry it as they are), checked by still_valid; "txcd" in the
  translator stamp when on; program cache format 2.
- Tests: tests/seg_shader_ir.cpp and tests/spirv_run.h emit every codec plain and in pieces, validate it, run its
  decode and encode on the CPU against reference conversions (every input of the 8- and 16-bit formats, random and
  boundary sets for 32-bit) and run its entry point over texel buffers in memory; the GPU probe's codec module.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 783eb1898b91d505ff22a31f337256c5047d0177)

fix(macos): a window without Vulkan extensions ends as a clean exit, as main's exit inventory asks ... f270209f99

Main's #75 classifies every host process exit (test_crash_record's
inventory, read from the source on every platform): a normal exit marks
the clean end first, a fatal one is listed as fatal. The macOS
ExitProcess(1) after hos_darwin_vulkan_capture() fails is the same kind
of exit as the SDL_CreateWindow failure just above it (a message, then
the end), so it marks the clean end the same way, and the inventory
lists display_sdl.c's second ExitProcess(1). Windows and Linux compile
the same code as before (the exit is inside #if defined(__APPLE__)).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

Merge mac/vk-admission (#438) into mac/renderer-correctness ... 621f3762bf

Carries the vk-admission pull request's commits, which this one needs
(MkwVkHost's Apple callbacks, darwin_vulkan.h), on top of guest-16k.
host/posix/win32_posix.c: its Apple include block keeps darwin_mem.h,
darwin_fault.h and darwin_vulkan.h.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(macos): pin renderer imports through final resource ownership ... 60b3d083be

Append Apple23 full-backing import lifetime callbacks while preserving the
non-Apple21 layout. Keep buffer views move-only, release buffers before the
last memory owner unpins backing, and preserve completion-gated retirement.
Unpublished failures unwind; published roots retain process lifetime rather
than calling Vulkan from unsafe static teardown.

Check and unwind post-device startup allocations. Resolve the shared audio
shim through the executable-owned SDL service on Darwin without unresolved
weak imports or linker relaxation. Retain the existing loader/SDL behavior.

Add game-free owner, ABI, source-topology and strict-link consumers, including
both native arm64 and x86_64 routes. This slice does not claim to fix separate
host-import alias visibility or implement the final alias admission planner.

Assisted-by: GPT-6 Astra
Co-Authored-By: GPT-6 Astra <[email protected]>
Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(macos): cut the non-Apple Pool and import route in three more fixtures ... dcd47c89ee

044635ea gave vk_memory.cpp's Pool an Apple-only DarwinImportView member and
import_pool an Apple route (darwin_import_pool) right after its guards. It
moved test_import_buffer_binding and test_texel_buffer_ownership onto the
Windows/Linux layout (#if defined(__APPLE__) -> #if 0: the Apple owner has
its own darwin_import_lifetime_test), but three more fixtures cut the same
code and stopped compiling on macOS. Each passes at 58c5c5aa (044635ea's
parent) and fails at 044635ea:

- test_direct_buffer_writes and test_parallel_rb_refresh: their cut of
  struct Pool named DarwinImportView. They now cut Pool's Windows/Linux
  layout, the same one they compiled before.
- nx2/test_import_windows: its cut of import_pool's guards, which ends at
  the import, took in the unterminated Apple #if. It now ends at the Apple
  route, which must stay one #if just before the import; the fixture's own
  continuation models the Windows/Linux import as before.

No assertion, case or check is removed or loosened. test_parallel_rb_refresh
and test_import_windows also need the renderer-core extractions (next).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(vk): drain the recorder before startup rollback destroys the allocator ... e072777e3e

044635ea added vma_startup_rollback (vk_vma.cpp, run after a failed
replay_start), which calls vmaDestroyAllocator. vk_vma.cpp includes
vk_work.h last so that RX_VK_RECORDER=defer/thread drains before its
destructions, and every such boundary call in a vk_work.h file must be in
RX_REC_BOUNDARIES with a drain-first macro. This one was not, and
test_recorder_sync caught it ("vk_vma.cpp calls ['vmaDestroyAllocator'],
which vk_work.h does not drain before"): it passes at 58c5c5aa (044635ea's
parent) and fails at 044635ea.

Add it to the list and to vk_work.h, next to vmaDestroyBuffer and
vmaDestroyImage. With the recorder off (the default) boundary() is a
relaxed load and a branch. Rollback runs before any image or buffer exists,
so nothing should be held there today, but the allocator is no longer
destroyed while a recorder could still hold packets. The test is unchanged
(sources: 42 boundaries drained, was 41).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(macos): follow main's #324 and #250 in the Darwin renderer fixture ... 8fa0934f6a

tests/test_darwin_renderer.py cuts vk_memory.cpp's code into
darwin_import_lifetime_test.cpp. Two of main's changes since this
branch's first base broke that cut on the rebase, before any check ran:

- #324 reworded the comment after imports_give_back's drain of the
  evicted pools ("(longshot) Off by default on Windows ..." is now
  "(longshot) Default by OS and vendor ..."). The cut ends at that
  comment, so it now looks for the new words. The cut code is the same
  drain loop as before.
- #250 (sparse mirrors) made retire_pool erase a retired pool from
  sparse_wanted. The fixture's stand-ins gain an empty sparse_wanted,
  beside mirror_wanted, as main declares it (std::vector<Pool*>).

No assertion, case or expected output changes.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): follow the moved Pool in test_mirror_gpu_only's cut ... b73b1501c7

107fda29 moves the new pool into the import table
(imported.emplace(p.lo, std::move(p))) instead of copying it.
tests/test_mirror_gpu_only.py ends its cut of pool_lookup's merge at
"imported.emplace(p.lo, p)", so it stopped with "substring not found"
(full-suite receipt full-91f9a58d: the only check that passes on main
and fails here). It now ends at the moved form. The check is the same:
the merged pool must take in the gpu_only and gpu_written marks of the
pools it absorbs before it goes into the table.

On the macOS integration branch this was hidden: renderer-core's
53b8914d rewrote the same line for its imported_add. When that pull
request rebases onto this one, its line replaces this one.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(macos): the import pin callbacks are macOS host API version 24 ... 5a831dca0a

Main's custom output size (#51) took MKWVK_API_VERSION 22 everywhere, so
the macOS loader callbacks (#438) became version 23 after take_text, and
this branch's guest_import_pin and guest_import_unpin, appended after
them, become version 24. The Apple table is now 392 bytes with the four
callbacks at 360, 368, 376 and 384; the admission contract and the
renderer startup fixture pin exactly that. Windows and Linux stay at 22.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(macos): vk_memory.cpp compiles on macOS: #324's evicted_pools declaration is not Apple's ... f66d0a87b8

Main's #324 (a4c77174) declares `extern std::deque<…> evicted_pools;` above
memory_guest_imported, which reads it on Windows and Linux. This branch
defines evicted_pools on macOS as a process-owned reference
(`auto& evicted_pools = *new std::deque<…>;`), so after the rebase onto
6c9edffe vk_memory.cpp did not compile for macOS ("redefinition of
'evicted_pools' with a different type"). The game-free fixtures cut the
functions out and never compile the whole file, so they did not see it.

The declaration is now for Windows and Linux only: on macOS
memory_guest_imported answers through darwin_guest_imported, which comes after
the definition. Windows and Linux compile the same text as before.

First made on mac/renderer-core, where this branch is merged.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 444e05c119a261b36c7c762abf9d26194b0ee417)

Merge mac/renderer-correctness (#491) into mac/renderer-core a58333bb9f

Merge fix/import-lookback (#431) into mac/renderer-core ... c6d49c67dd

# Conflicts:
#	gpu/vk/vk_memory.cpp
#	tests/test_mirror_gpu_only.py
#	tests/test_parallel_rb_refresh.py

fix(vk): join #431's import table helpers with #491's macOS Pool ... 118a66f2e7

#431 (fix/import-lookback) makes imported_add and imported_take the import
table's only ways in and out. #491 (mac/renderer-correctness) gives the
macOS Pool an import view (DarwinImportView) that owns the imported memory
and cannot be copied, defines the table on macOS as a process-owned map,
and releases imports through destroy_import. Merged, they do not build or
test on macOS:

- imported_add took a const Pool& and copied it. It now takes Pool&& and
  moves it into the table (the macOS stack's form, 53b8914d). Its callers
  already move (the merge kept #491's moves); the fixtures that call it with
  a local pool move it, or a copy where they read the pool afterwards
  (nx2/test_import_windows.py). Windows and Linux insert the same Pool at
  the same key: a move in place of a copy.
- test_import_reach.py cut the Pool and the scans with their Apple guards,
  so on macOS the fixture met DarwinImportView. It now resolves the Apple
  guards to the other side (test_vk_darwin_admission.py's select, as
  test_darwin_renderer.py does; the macOS route is that test's), cuts
  destroy_import where the source has it, and its source check does not
  count macOS's definition of the table as a change to it.
- test_darwin_renderer.py's Darwin fixture cuts live_reach, imported_add and
  imported_take (darwin_import_lifetime_test.cpp includes import_reach.h),
  and its queue-slot order check finds imported_take where the source has it.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(macos): import NOACCESS guest pages and name a missing audio service ... 01aad9ff48

RendererSliceReview findings on 044635ea, forwarded by the parent.

darwin_import_commit refused a whole import when any page in its range was
committed but NOACCESS. Darwin's VirtualQuery reports the demand-paged guest
window committed with each page's own protection, so an explicit NOACCESS or
a diagnostic watch page dropped every binding in that range to zeros, where
the pre-pin path tried the import and let the driver's VkResult decide. The
CPU protection of a committed page is CPU-side state: such pages are now left
exactly as they are and the import proceeds; only an unmapped page or a failed
query refuses. Whether MoltenVK/Metal imports a NOACCESS host range is left to
its own result; a GPU arm for the parent-released probe settles it.

waveOutOpen on macOS returned MMSYSERR_ERROR without a word when the
executable provides no hos_posix_audio_open. The lookup now names the missing
service once, on stderr; nothing else changes.

Tests: darwin_import_lifetime_test.cpp imports a range with a committed
NOACCESS page (no commit, no reprotection); test_darwin_audio_service.py
counts the line: once with the provider missing, never with it present. Both
fail on an archived 73f90266.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 9e3832639a33e22dc8672e80a64354eb729c4c48)

feat(vk): admit an operation's guest buffers before it binds them (alias isolation, macOS) ... 1bcfec58cb

Two host imports of the same bytes read stale data within one submission on MoltenVK. Under alias
isolation (set in memory_init on Apple only) every operation's guest buffers are admitted before it
binds the first, and no import is replaced while recorded work may still reach it:

- Seam: memory_admit / replay_admit on the replay thread (replay_mu, no parallel worker): draw()
  (mesh records, then its buffers once the pipeline is known), dispatch() (the upscale stand-in, then
  its buffers), the replay's signal, buffer copy, fill and uniform update, the textures of every
  copy and clear, and copy_b2t's buffer side once its texture is resolved.
- Resolver: one side-effect-free description of what each binding path binds, from the binders' own
  helpers (uniform_bind_size, storage_binding, texel_buffer_extent, buffer_extent, the *_bytes
  helpers), including the GPU unswizzle's source (texture_unswizzle_source), the temporal
  upscaler's actor motion (temporal_actor_demands) and the diagnostic readers' copies
  (RX_VK_DRAW_STATS, RX_VK_DUMP). tests/test_resolver_coverage.py holds every gpu_buffer_bind /
  gpu_buffer call site of the renderer to its resolver demand and its seam.
- Covering reuse: a range a live import holds needs nothing; on Apple a range inside a live import's
  backing becomes a new view on that memory. Backing extents count for alias conflicts.
- Handoff: end the pass, flush the uniform updates before this place, device-to-host barrier,
  submit, wait, then import (old memories retired and kept to completion, page-table changes
  deferred), host-to-device barrier.
- Cohorts: different operations whose union passes the cap get one memory each, with a handoff.
- Under batch admission the dry pass's view-making marks are skipped (its command buffer is then
  recorded in order); RX_VK_POOL_END_IMPORTED is ignored under alias isolation (it imported to
  answer a resolver's question).

Off alias isolation (Windows, Linux) nothing of this runs and every binding path is unchanged.

Tests: test_alias_admission, test_admission_handoff, test_resolver_coverage (new); the Apple
spare-backing reuse and backing-overlap cases in darwin_import_lifetime_test; extractions of
test_import_chunk_default, test_fallback_depth_hold, test_vulkan_copy_span(_gpu_compile),
test_draw_side_effects and test_vk_occlusion_queries follow the moved code, every assertion kept
(the last two also check the admission seam). test_import_reach.py's main() call is restored: the
53b8914d copy ran nothing.

Published on main 6c9edffe with #491 and #431 below it (mac/renderer-core):
- actor_reads binds main's index bytes (count << itype) for every index
  type, as main's actor_motion does. The macOS stack's copy bound nothing
  for an index type outside 0..2, which would change Windows and Linux too;
  such a draw does not reach actor_motion (bones_used refuses it when the
  draw is noted) unless its model key matches a draw of another index type,
  so that guard is left out of this macOS PR.
- vk_temporal.cpp: #59 (actor motion pairs copies by place) copies a pose
  read at its draw from the stage; the copy from the game's memory, its
  else branch, binds through actor_reads.
- vk_memory.cpp imports_give_back keeps #324's retire grace beside alias_ref.
- test_import_reach.py keeps #431's source check and prefetch case and takes
  the admission excerpts and --arch. Its main() call needed no restoring:
  #431's copy has it.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit ae1d7812c12f3e64941da0c34bb3541eb695f6db)

feat(vk): import a span past one device allocation as segments (SegMem, alias isolation) ... 46873da415

A connected guest span one operation needs that is larger than max_import_size (41,747,087,360 bytes
on MoltenVK 1.4.2, M4 Max) was left unplanned (admit_over_cap): its bindings came back empty. It is
now a segmented domain (reports/2026-11-14-full-range-design.md 2.1, 2.3, 5.3-5.5), in place of
that branch:

- memory_admit plans it as disjoint, adjacent segments where span_segments::plan cuts them (the
  operation's views and classes as demands, the domain's previous cuts kept unless moving one pays,
  priced extra segments), each an ordinary Pool of at most the cap: one import, its own page-table
  entries, mirror, write-backs and retirement. Only segments holding a demand are imported; every
  live import over their bytes is retired, after a handoff when recorded work may reach it.
- A view across a cut is never one memory: pool_lookup refuses it (counted, logged), not owed to an
  admission. gpu_span_bind (vk_internal.h SpanRef) gives its pieces, one per segment, for the
  consumers' failure branches; gpu_span_pieces and gpu_span_layout describe them binding nothing.
- Domains change only at an admission, never while parallel workers run (debug-checked with the
  design's invariants, seg_invariants); a batch past the cap is admitted as segments instead of
  being recorded in order. memory_admit loses its batch flag and AdmissionStats its splits.
- span_unused_piece: the hook the NullDesc design fills (MoltenVK has no nullDescriptor: an unused
  piece slot of a storage table cannot be null).
- import_planned retires a taken-in import only while it is still over the plan's bytes (two
  segment plans of one admission may name the same old import).

Off alias isolation nothing of this runs.

Tests: test_segmented_domains (the 48 GiB A/B case at the real cap, the priced extra segment,
injected small caps, growth with the cuts kept, a moved cut after a handoff, frozen under workers,
a refused domain reported, nothing off alias isolation; fails on the planner commit and on the
archived base); test_resolver_coverage lists gpu_span_bind's bindings and holds future callers to
a failure branch.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 9fb9b3b4a60af49e54031d9c7707cdaf71fa5117)

test(vk): follow the renderer-core merge's moved code in three extractions ... 2724ac9194

The renderer-core merge (f216d287) moved code three fixtures cut out of the
sources. ae1d7812 updated the extractions it listed; these three it missed
(its named set held no nx2/ tests, and the other two already failed on its
base for the reason the previous commit fixes). The code moved; what it does
did not change, so each fixture follows it and keeps every assertion and case:

- nx2/test_cbuf_table_extent: ae1d7812 moved uniform_range's `used` lines,
  unchanged, into uniform_used(), which the resolver shares, and
  uniform_range calls it. The test searched for those lines after
  uniform_range( (ValueError). It now cuts uniform_used's body, from the
  same `uint32_t used = index < Shader::Info::MAX_CBUFS` line, and
  uniform_range's own `used` up to the driver banks, and runs both as
  uniform_range does.
- tests/test_parallel_rb_refresh: ae1d7812 moved build_descriptors'
  bare-entry guess and binding range into storage_binding() (b.addr, b.size,
  b.bare, ...). The fixture now cuts storage_binding's guess and range and
  build_descriptors' binding from them, through memory_note_gpu_write, the
  lines it cut before. 53b8914d made memory_note_gpu_write's scan stop by
  import_reach.h's live_reach, so the fixture includes the header, cuts
  live_reach and imported_add, and adds its pool through imported_add.
- nx2/test_import_windows: pool_lookup and the code above it now read
  53b8914d's import_reach (live_reach, imported_add, imported_take) and
  ae1d7812's admission state and alias isolation globals. The fixture cuts
  them (DemandClass and AdmissionStats from vk_internal.h), with isolation
  off as on Windows and Linux and alias_conflict false (as
  tests/import_reach_test.cpp does). Its retire_pool spy and its two seeded
  imports use imported_take/imported_add, the table's only ways in and out.

The three print what they printed at 58c5c5aa.

Published on main 6c9edffe (mac/renderer-core): #431 below this branch
already cuts live_reach, imported_add and imported_take into
nx2/test_import_windows and test_parallel_rb_refresh (its rb_table.inc), so
those two keep #431's cuts and add only what follows ae1d7812: the alias
isolation globals, DemandClass and AdmissionStats, and alias_conflict's
stub; storage_binding's cut.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 64a84e4dbc10017f44cd16fc0fbd7440135ee498)

fix(vk, macos): admit a parallel batch's uniform updates in order, and keep its own updates out of its handoff ... 6368f9306a

Two findings from the renderer-core review (reports/2026-11-14-renderer-core-review.md), both in vk_replay.cpp's
parallel-batch admission. Neither path is reachable off Apple: admit_fallback and batch_admit are set only where
memory_admission_on() is true.

P2.5: a parallel replay's batch its dry pass could not know is recorded on the replay thread in order (Mode::Worker,
admit_fallback), each operation admitted as the serial replay admits it, except NOP_UPDATE_UB, whose worker branch broke out
before op_admit. Its destination is bound at the recording's end (flush_ub_updates -> gpu_write); one that needed a growth over
reached memory was refused by the guard and written on the CPU out of GPU order. The worker branch now admits it under
admit_fallback, before the update counts as passed, so a handoff there writes only the updates before it.

P3.1: a batch admitted at once whose admission hands off called admission_handoff() with rmode Serial, which wrote every listed
uniform update into the recording that ends, including the ones the batch's own dry pass had listed (from ub0). admission_handoff
takes the end of the updates to write (ub_end, default all: unchanged for every other caller), and the batch passes ub0.

Tests: tests/test_batch_admission.py (new) runs the NOP_UPDATE_UB case, op_admit, replay_admit, admission_handoff and the batch's
handoff statement cut out of the source; on bacb7c59 both of its cases fail for those reasons. tests/test_admission_handoff.py
cuts admission_handoff out by its name only (its signature gained the parameter).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 39542f89f4443e14b70fb1e46bdaa3e0e330f2ea)

fix(vk, macos): keep every byte the page table reached mapped when an admission retires imports ... ffb586ae1c

Findings P1 and P2.2 of the renderer-core review (reports/2026-11-14-renderer-core-review.md). Both are in vk_memory.cpp's
admission planner, which runs only under alias isolation (memory_admit returns at once elsewhere: macOS only).

P1, the page-table holes. A plan retires every live import it overlaps (retire_pool zeroes the import's whole range), but two plans
did not cover what they retired:
- the cohort (plan_cohort) retired the imports it overlaps and does not need without growing over them;
- a segment (seg_plan) retired an import that reaches into a segment no demand needs.
Shaders that follow the guest page table then miss on those bytes. The fix, by the rule that no byte reachable before an admission
is left without memory after it (design 5.5):
- admit_rest: the bytes of each retired import that no plan holds are planned with the first plan that retires it, imported
  exactly, before that plan's import (one that fails leaves everything as it was, as mirrors_trim does), and published after it;
  they keep the import's marks (gpu_only, gpu_written, trims, last_value) and count in the handoff's conflict test;
- seg_plan: a segment that replaces live imports makes every segment over those imports needed, to a fixpoint;
- the SegMem span takes in every live import over it, so a segment retires only imports its domain holds whole;
- the cohort's threshold is the device's import cap (max_import_size) under the admission, not kMergeMax: a merge up to the cap is
  one import (pool_lookup keeps kMergeMax: plan_import's merge_max defaults to it);
- admit_apply: a plan that is not made keeps the imports it would retire, and no later plan of the admission retires them.

P2.2, covered demands across a domain's boundary or cut. A range a live import already holds was left out of the domain: across the
domain's edge, the segment's plan retired its import and the guard then refused its binding; across a cut, one live import bound
it whole while gpu_span_pieces reported its pieces. The span's live-import take-in covers the first. For the second, a covered
range that straddles a cut seeds its domain's span (its segments are planned, the crossing import retired with the handoff the
conflict test asks for).

Tests: tests/test_admission_coverage.py with tests/admission_coverage_test.cpp (new; the admission cut out of the source, recorded
imports and page table, injected caps), cases merge-table, cohort-table, segment-table, domain-edge, covered-straddle. Each fails on
bacb7c59 for the reviewed reason (reports/CoreFix-files/repro-base-bacb7c59) and passes here on arm64 and x86_64.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit f29396c79600467b11fbb935a529a4c2511cf533)

fix(vk, macos): one view per byte of an Apple spare backing, and no endless retry of a failed view ... 0c2ace70ce

Findings P2.1 and P3.3 of the renderer-core review (reports/2026-11-14-renderer-core-review.md). The code is Apple-only
(#if defined(__APPLE__) in vk_memory.cpp's admission planner).

P2.1: covering_backing returns the owner's backing for a demand across the end of the owner's view, and the view plan was
[max(want_lo, b->lo), min(want_hi, b->hi)): (a) at the owner's own key, imported_add kept the owner, the new view was destroyed,
memory_admit returned Ready and the guard refused the binding at every admission; (b) at another key, the view overlapped the
owner (eviction of either zeroed page-table entries the other served; seg_invariants broke inside a domain). A failed
darwin_import_view looped the same way. Now:
- view_plan grows the view over every live view on the same backing it overlaps; those are retired when it is made (the same
  memory: alias_conflict excludes it, no handoff; retired views are kept until their work is done) and their gpu_only and
  gpu_written marks carried. A live import of another memory over it means no view: the range is planned as pool_lookup imports it.
- admit_apply makes the view first, then retires what it takes in, then inserts it, and checks the insertion.
- A view that fails is logged, its backing is remembered (weak_ptr) as one that takes no view, and its range is imported on its own
  (the binding, or the guard's debt and the next admission's handoff), never as a view on it again.
P3.3: admit_st.reused counts only a view that went into the table.

Tests: tests/darwin_import_lifetime_test.cpp (the real Apple import code, test_darwin_renderer.py, both arches) gains spare-cross
(a range across a one-chunk owner's end), spare-reuse-count, spare-overlap (a two-chunk owner) and spare-view-failure (the driver
refuses a second buffer on a memory). Each fails on bacb7c59 for the reviewed reason (reports/CoreFix-files/repro-base-bacb7c59)
and passes here; the fixture takes --case NAME to run one alone (the driver's --process-exit is unchanged).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 0c1d2aea03e17a3a10b0c1aa6ef69461d8d405d9)

fix(vk, macos): bg_drain publishes no background import over this epoch's retired memory or across a domain ... 3aff819a69

Finding P2.4 of the renderer-core review (reports/2026-11-14-renderer-core-review.md). Under alias isolation (macOS) bg_drain
dropped a background import only when it overlapped a live import or one sat at its start; it never asked alias_conflict, so an
import over bytes a memory retired, evicted or queued dead in the current handoff epoch still held (alias_ref) was published:
two memories over the same bytes in one epoch, the stale read alias isolation prevents. It could also publish a chunk across a
segmented domain's cut. Reachable only with the opt-in background imports (RX_VK_PREFETCH_CHUNKS, RX_VK_ASYNC_IMPORT,
RX_VK_BG_IMPORT).

Under alias isolation bg_drain now also drops (counted in st_prefetch_dropped) an import when alias_conflict holds over its
memory's bytes (on Apple its whole backing), or when it is over a segmented domain (seg_overlaps). The new branch runs only when
alias_isolation is set: off Apple the condition is unchanged. alias_conflict's and seg_overlaps' declarations move up beside the
alias isolation globals, and seg_overlaps' definition into SegMem's state block (above pool_lookup), so bg_drain (and every
fixture that cuts it out with them) sees both.

Tests: tests/admission_coverage_test.cpp gains bg-epoch (an import given back after work reached it, a background import of its
bytes), bg-domain (a background chunk across a cut after the segments were given back) and bg-elsewhere (off alias isolation,
published as before); tests/darwin_import_lifetime_test.cpp gains apple-bg-epoch (Apple's alias_conflict, real import code). The
first two and apple-bg-epoch fail on bacb7c59 for the reviewed reason (reports/CoreFix-files/repro-base-bacb7c59); all pass here
on arm64 and x86_64.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 755ac7e8757fdd30d4809ae1142e51232fae2a9e)

fix(vk, macos): admission merges no import past 1 GiB again, so mirror_make can still mirror it ... 7ce8a80966

Finding N1 of the core-fix review (reports/2026-11-14-core-fix-review.md). f29396c7 let the admission merge imports up to the
device's import cap (plan_import(..., max_import_size), and the plans' merge at max_import_size). mirror_make refuses a VRAM mirror
for any pool over 1 GiB, and mirrors are on by default for a discrete, non-GCN GPU (an Intel Mac with an RDNA Radeon, which alias
isolation also covers) and with RX_VK_MIRROR=1: such a merged pool lost its mirror, every binding in it reading host memory.

The cap was never needed for correctness once admit_rest keeps a cohort's leftover bytes mapped, so the admission's threshold is
kMergeMax again (cohort plus leftovers past it). plan_import loses the merge_max parameter that only the admission used; it, its
comment and pool_lookup's calls are as on bacb7c59. No correctness reason to keep the cap was found: a cohort and its leftovers
import the same bytes a merge would, each piece no larger than the import it came from.

Off Apple: unchanged (admit_plan runs only under alias isolation; plan_import is textually bacb7c59's again).

Tests: tests/admission_coverage_test.cpp gains merge-mirror (the seam of two live 768 MiB imports at the real cap: no import the
admission makes is larger than mirror_make's limit, which the driver reads from mirror_make itself, and every byte stays mapped).
It fails on 755ac7e8 for that reason (reports/CoreFix-files/repro2-base-755ac7e8) and passes here on arm64 and x86_64;
merge-table (the same seam: every byte mapped, one handoff) passes on both.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 9ccc7c19864991be11d4437211eecf3e430b2a7d)

fix(vk, macos): a segment keeps a crossing import's bytes as leftovers clipped at the cuts, not as whole segments ... 3e60977da0

Finding N2 of the core-fix review (reports/2026-11-14-core-fix-review.md). f29396c7's closure in seg_plan marked every segment over
an import a needed segment retires as needed, and each such segment was then imported whole (up to the device's cap, 41.7 GB of
guest address space on an M4 Max) only to keep that import's few bytes in it mapped; and when such a forced segment's import failed
after the needed one's, the crossing import was already retired.

The closure is gone: seg_plan plans only the segments the operation's ranges need. The retired imports' bytes in segments no range
needs are leftovers, as a cohort's are: admit_rest now takes segment plans too and cuts every piece that lies in a domain planned in
this admission at that domain's cuts, so each piece lies in one segment, crosses no cut and overlaps nothing (invariant 1 holds),
and costs no more than the import it came from. They are made before the segment's own import and published after it retires the
import (admit_apply, unchanged), so a failure leaves everything as it was.

Off Apple: unreachable (seg_plan and admit_rest run only from the admission, under alias isolation).

Tests: tests/admission_coverage_test.cpp gains segment-rest (a chunk across the cut into a 100 MiB segment nobody needs, retired by
the segment the operation needs: its 4 MiB there stay one import clipped at the cut, nothing else of that segment is imported, every
byte stays mapped, no import crosses a cut). It fails on 755ac7e8 for that reason (reports/CoreFix-files/repro2-base-755ac7e8) and
passes here on arm64 and x86_64; segment-table, domain-edge and covered-straddle (the other segment cases) pass on both.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 596f0ff4063da99505e7d5df789580fb4cde89f7)

fix(vk, macos): each new import over a retired import's bytes inherits its share of that import's VRAM mirror ... 4309c50216

Open item 1 of the core-fix review (reports/2026-11-14-core-fix-review.md), more reachable after N1 (9ccc7c19): with mirrors on
(a discrete, non-GCN GPU such as an Intel Mac's RDNA Radeon, or RX_VK_MIRROR=1), an import an admission retires whose bytes go to
several new imports (a cohort and that import's leftovers, or two segments over an import across their cut) handed what the GPU
wrote into its mirror and nowhere else to the first of them whose mirror was made only: mirror_make ends a retired pool's wait unless
it is trimmed, so the others' mirrors were filled from guest memory without it.

admit_apply now does what mirrors_trim does for a trim's parts:
- it counts, per import the plans retire, the new pools its bytes go to (the plans over it and its leftovers);
- a retired import with more than one, whose mirror was on (inherit_pending, set by retire_pool), is marked trimmed and counted in
  trims_waiting once the plan that retires it is made, so each new pool takes its share when its mirror is made, and trims_settle
  lets it go once every live pool over its range has its own (or after kIdleEpochs, as before);
- a leftover of an import whose mirror was on wants a mirror at once (MIRROR_WANTED, in mirror_wanted), as a trim's parts do.
trims_waiting is declared ahead of the admission region (its definition, with mirrors_trim, is unchanged).

Mirrors off (Apple Silicon's default): no pool is ever MIRROR_ON, so retire_pool sets no inherit_pending, nothing is marked and no
mirror is wanted: unchanged (the counting runs, without effect). Off Apple: unreachable (admit_apply runs only under alias isolation).

Tests: tests/admission_coverage_test.cpp gains mirror-cohort (the seam's cohort retires a mirrored import: the cohort and the leftover
both inherit, the leftover wants a mirror, trims_settle, cut out of the source, lets the old one go) and mirror-segments (two needed
segments over a mirrored import across their cut); mirror_made models mirror_make's inheritance, whose two tests the driver checks
are in mirror_make verbatim. Both fail on 596f0ff4 for the reviewed reason (reports/CoreFix-files/repro3-base-596f0ff4) and pass here
on arm64 and x86_64; mirror-off (mirrors off: nothing marked, nothing wanted) passes on both. alias_admission_test.cpp,
segmented_domains_test.cpp and darwin_import_lifetime_test.cpp get a trims_waiting stub only (no assertion changed).

Published on main 6c9edffe (mac/renderer-core): darwin_import_lifetime_test's
trims_waiting stub sits beside #491's sparse_wanted stub (main's #250).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 6aa7c2af0fb4e9d0dc57d6eeb8018aaac07ed234)

test(vk): the admission fixtures follow main's #250 sparse mirrors ... d64b26a285

Main's #250 (sparse mirrors) makes retire_pool drop a retired pool from
sparse_wanted, the pools a binding found a sparse block with no memory in.
test_alias_admission, test_segmented_domains and test_admission_coverage cut
retire_pool out of vk_memory.cpp with the admission, so their fixtures did
not compile on main 6c9edffe ("use of undeclared identifier
'sparse_wanted'"). Each gets the stub darwin_import_lifetime_test already has
(#491). No assertion changes.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

fix(vk, macos): the actor motion's resolver follows #59: a pose read at its draw is not copied ... d792779940

Main's #59 (actor motion pairs a model's copies by place) takes a skinned
draw's pose as the draw read it (temporal_note_skinned_draw, ub_read), through
a host-visible stage, and copies the bones and origin from the game's memory
only for a draw whose pose was not read, or with no stage. actor_reads and
temporal_actor_demands (alias isolation's resolver for the upscaler's actor
motion) still demanded the bones and origin of every skinned draw: a
superset, so admissions on macOS could import or hand off for bytes nothing
binds.

pose_read_at_draw(s, k, staged) is now the one test, actor_motion's (with its
stage) and the resolver's (with the stage actor_motion takes next): the
resolver demands a draw's bones and origin only where actor_motion copies
them, and its vertices and indices always. With no stage made yet the
resolver demands the bones and origin too, since actor_motion's making of the
stage may fail. test_resolver_coverage holds both to that test.

Off alias isolation (Windows, Linux) the resolver does not run, and
actor_motion's test is the same expression as before.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): test_split_ub_flush follows flush_ub_updates' end parameter ... 61766e3eca

Main's #189 check reads flush_ub_updates by its signature. The admission (6e765c8d) gives it an end, the first update not
written, for a handoff in a dry-passed command buffer, defaulting to SIZE_MAX, so the check's anchor was not found and it
stopped with a ValueError. It now reads the new signature and still requires what it guarded: the writes start at
ub_first_of_recording and, with the default end every caller in maybe_split and finish_recording uses, reach the last
update (end = std::min(end, ub_updates.size())).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

Merge mac/texel-codec (#P21) into mac/seg-contract ... 1ebb97e7f1

This PR needs the texel codec PR (#P21: texel_codec::Plan, which
texel_view_pieces takes) and the renderer-core PR (#P19, the base:
SegMem, gpu_span_bind, SpanLayout). Both are kept as they are; the
merge had no conflicts.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): the segmented consumers' frozen contract (C0) ... 4c16dbc4cc

The interfaces the four segmented-consumer lanes (SegBind, SegXfer,
TexelDesc, NullShader; reports/2026-11-14-seg-consumer-plan.md section 3)
code against, frozen before they fork. Inert: no caller of anything new,
no SPIR-V, layout, push range or stream change, sizeof(Program) unchanged.

- vk_internal.h / vk_memory.cpp, SegMem's consumer interface:
  gpu_span_any() (a relaxed count of the live domains, stored in
  seg_commit; domains are kept for the run), gpu_span (gpu_buffer's
  flavour of gpu_span_bind: import buffers, no written marks, no mirror),
  SpanLayout::bind_stamp (gpu_bind_stamp(): a segment can be imported
  again with its cuts kept, so cut_gen alone never keys piece buffers),
  span_unused_piece's semantics ({null_storage, 0, 1}), and the SegArena
  API (SegArenaRef, seg_arena_take, seg_gather, seg_scatter; stub bodies
  after stream_alloc that give no room and log, SegBind's to fill).
- Review fixes in the same interface (renderer-core review P2-3, P3.2):
  every piece of gpu_span_bind carries its buffer's device address (the
  import's or the mirror's; 0 only with bufferDeviceAddress off), and its
  one-piece path stores the translated guest address, not a virtual one.
- vk_texture.h/.cpp: texel_view_pieces / texel_piece_views, stubs that
  return 1 and false (TexelDesc's to fill).
- vk_shader.h: Program::presence_words, in what was padding.
- Back end: PRESENCE_PUSH_OFFSET (88), Profile::emulate_null_descriptors,
  Bindings::presence / presence_read with the bit numbering rule.
- tests: test_resolver_coverage lists gpu_span_bind / gpu_span sites too
  (scoped to gpu/vk: gpu/render has an rx::gpu_span of its own), checks
  that gpu_span follows its own gpu_buffer, and splits SITES into
  SITES_BASE plus one block per lane. test_segmented_domains executes the
  contract (the count 0, 1, 2 -> 1 on a merge, 0; gpu_span's pieces;
  addresses with the fixture's gpu_buffer_bind returning none as the real
  one does; the bind stamp; a virtual address's guest). test_seg_shader
  accepts a base tree of this back end's generation and compares every
  module.
- docs/segmented-binding.md: the gate (every new behaviour Apple-only,
  other platforms byte-identical) and one heading per lane.

Conflict on this branch (not with main): tests/test_seg_shader.py's
docstring, against the array_pointer_calls paragraph #P20's
msl-arrayptr port added; both kept, this paragraph first.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 95fa06d4e6526601bd2e1bc88c2f8f5720855350)

Merge mac/mvk-rwtex (#P28) into mac/seg-consumers ... 2b9dd377f3

The segmented binding PR needs the MoltenVK descriptor-set fixes (#P17
storage buffer tables, #P26 storage buffer lengths; #P28 read_write
textures, which it lands after): ensure_layout's layout_pushes decides
the push, and seg-bind's counts go through it. Their commits are kept as
they are; the merge had no conflicts.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): texel codecs' views, texel piece views, null-descriptor stand-ins and presence words (TexelDesc, L3) ... 74549ce92b

The renderer's half of plan section 6 (reports/2026-11-14-seg-consumer-plan.md), on B2 08cccafb:

- vk_texture.cpp texel_buffer_view(t, v, storage, plan): the plan's view (the format's own, an alias, a typed
  access's declared format, or an R8/R16/R32 UINT carrier), whole-texel ranges at most maxTexelBufferElements of
  the view's elements, plan.zero as a cached 16-byte zero view; key (format, offset, range) with the backing as
  before, every view through one cache helper (texel_view_locked).
- texel_view_pieces / texel_piece_views (C0.4): cut at a segmented domain's cuts (span_segments::texel_pieces, in
  view elements) and, on Apple, at runs of maxTexelBufferElements; each piece view on its segment's buffer (gpu_span
  after gpu_buffer); a stale count, an unbound piece, a seam or a written seam refused loudly (the element absent).
- gpu/vk/null_dummies.h (new, pure): the stand-ins' tables by view type x class x multisample, storage format, and
  the texel zero view's format from the variant's texel answer. vk_texture.cpp null_dummies_make makes them once
  (1x1 zero images, views, four samplers, texel zero views; vkGetPhysicalDeviceImageFormatProperties first; one
  clear submit), only on Apple without nullDescriptor (make_dummies).
- vk_draw.cpp: texel_views with texel_plan / texel_plan_native, P piece views for a segmented variant's entry, the
  zero view of the declared format for an absent element; sampled and storage-image stand-ins, null samplers, the
  default sampler for a real view whose sampler failed; DescWriter presence bits (C0.8 numbering), the graphics
  words kept per thread, push_presence at offset 88.
- Non-Apple, and Apple with nullDescriptor: writes byte-identical to B2 (checked against the B2 cut).

Tests: test_texel_buffer_ownership (extended), test_texel_views_plan, test_null_dummies, test_null_descriptors,
test_null_layout_push (new), SITES_TEXEL; docs: texel-buffer-formats.md (new), texel-buffer-ownership.md,
segmented-binding.md TexelDesc.

Conflict on main: tests/test_texel_buffer_ownership.py's eviction cut
ends at imports_give_back's "(longshot) Default by OS and vendor"
comment, as main's #324 reworded it ("Off by default" before).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 36fe0a1359b31095fcb8d40d3c2f502a105d6b1b)

test(vk): the texel ownership fixture names the case that failed ... 17ce0ff24c

A failed case printed only a CalledProcessError; it now prints the fixture's own FAILED line and the case.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit cb8dea7c94ef0b0c002c9102d6c78c48bc778b7e)

feat(vk): SegBind's descriptor core, gathers and arena (L1) ... 5f75a29085

On a segmented domain (macOS), vk_draw.cpp now binds storage descriptors across a segment's cut as piece arrays in the
segmented variant program_segmented makes, gathers uniform buffers and fixed-function inputs into a per-recording arena at
the pre-pass point, and fills C0's arena stubs. On Apple, unbound storage descriptors take a zero-length range of a new
64 KiB null_storage buffer (span_unused_piece set), and unbound uniform buffers take 64 KiB of zero_buffer, which loses
its STORAGE usage. Other systems keep every write, layout and push range as before.

- ensure_layout and the descriptor counts take piece counts from Program::seg; push range 88 + 4 * presence_words
- build_descriptors: single, table (k*P + j, prefix rule), bare/wide/shrink per piece; a refused array writes nothing
- uniform_range: gather queued (SegBuild), null_ubo on Apple
- draw()/dispatch(): seg_sig_of + program_segmented after the seam, always waited for; same_desc never reuses a
  gathered or stale piece set; vertex/index/indirect/mesh-args/dispatch-args gathers; index extent demanded (O9)
- pipeline_note and the compute notes skip segmented variants
- tests: test_seg_bind (new), resolver coverage SITES_BIND, and the fixtures and anchors the change moved

The next commits add test_seg_gathers, test_seg_notes, the desc_counts, push_state and vertex_bind
extensions and the docs section. (Reworded from the topic's "wip:" commit; the content is the topic's.)

Conflicts on main 6c9edffe: #252's compute-variant path in dispatch()
(program_cache_complete_ask, on_worker, made_how; compute_optimize at
the bind) meets this commit's rename of the pre-seam program to
`plain`. #252's lines are kept here as main has them, naming `p`; the
notes take SegBind's guard and #252's optimize and bind follow it.
They follow the rename in this PR's last commit, which vk_draw.cpp
needs to compile; tests/test_program_cache_one_variant.py the same.
tests/test_set_for_tables.py (#P28's read_write cases): CASES counts
both sides' cases.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit f40b77f71a47ee130cbe0df476d5e59f00fcef03)

fix(vk): a draw reusing its descriptors drops an earlier build's refusal (SegBind) ... 34df68891a

draw()'s pre-pass point cleared only the gathers an earlier build had queued when the draw reused its descriptors. The build's
lost flag stayed: a build or signature that was refused (seg_lost: no arena room, a piece array that could not be written, a
descriptor in more than 255 pieces) left its own draw out, and then every later draw that reused the previous set while a domain
existed was left out too, with no log line of its own. The whole SegBuild is now reset there (its gathered/pieces/stamp fields are
read only at the bind of a fresh build). Apple-only in effect: off alias isolation no build is ever refused.

tests/test_seg_gathers.py's "stale" case fails without it.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit a38cf86cc766d3254539965265c9bf3d56123b87)

test(vk): SegBind's gathers and notes fixtures, its parts of three checks, and its docs section ... d4865e157e

Plan 4.3 items 2, 3, 6 and 7 (reports/2026-11-14-seg-consumer-plan.md):
- tests/test_seg_gathers.py + tests/seg_gathers_test.cpp: the source's own cuts (RecordingRes and the arena with
  draw_begin_recording's reset and give-back, SegBuild, the size helpers, the bound-state caches, and draw()'s mesh-args
  rewrite, pre-pass gathers, vertex streams and indirect/indexed binds; dispatch()'s grid gather and binding) over a command
  recorder and SegProbe's domain. 76 checks: the order end_pass, barrier, copies, barrier, reader; vertex sized and WHOLE
  (clipped to the domain); direct indices; the indexed-indirect extent; records across a cut; native mesh records; the
  mesh-args rewrite per piece; a queued uniform gather; the 12-byte grid; count words in place and a straddling one left out;
  aligned arena offsets; arena made, grown, reset, given back, and full (a loud skip); nothing kept in kept_vb, vertex_set or
  last_ib; no stale refusal after a reusing draw. --base compares the common path's recorded commands with B2's cut; --json
  writes seg-gathers.json for SegBindProbe. Against B2 (--source) it builds and fails.
- tests/test_seg_notes.py: pipeline_note and dispatch()'s note lines cut out: a segmented variant makes no state_record,
  PipeNote or compute note; a plain one is noted as B2's cut is (--base).
- test_desc_counts.py, test_push_state.py, test_vertex_bind.py: --source, and SegBind's source rules (piece counts with
  seg_layout_count; the push range kPushSize + 4 * presence_words and uses_push; gathered streams and indices never reach
  vertex_bind_reuse, kept_vb, vertex_set or last_ib). Each fails against B2.
- docs/segmented-binding.md: the SegBind (L1) section.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit a0eccd31253378f8beb5963da4d4a6e5f340e670)

test(vk): test_set_for_tables reads its cut from --source ... 6e983aa6b0

The segmented single storage array's case (plan 4.3 item 5) had no negative: the check always cut its own tree. With --source
<B2 archive> it fails as required ("moltenvk.segmented_single want=set got=push").

Conflict with #P28 (read_write textures; not with main): the docstring
keeps #P28's paragraph with this usage line, and main() cuts from
--source's vk_draw.cpp with #P28's read_write count (as in the
macOS integration, Integration 17).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 5824f6828e21cf810dcceaa6be425a9224b15994)

test(vk): test_mesh_fallback's topology anchor follows draw()'s plain ... 7a498e5865

SegBind renamed draw()'s program for the bound state to `plain` (after the seam `p` is it or its segmented variant), so the
source rule "topo = pipeline_topology(p, topo);" no longer matched and the check failed in the full suite (a new failure
against B2's full run). The rule now reads `plain`; what it checks is unchanged.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 2e1eb52102ae7a889af7a0aa770f448833d36ebf)

test(vk): test_set_for_lengths' layout fixture takes SegBind's seg_layout_count and presence_words ... 8811999777

ensure_layout, which the fixture compiles from vk_draw.cpp (kPushSize
to its end), now counts each descriptor through SegBind's
seg_layout_count and grows the push range by Program::presence_words.
test_set_for_lengths (mvk-sizes) was written after SegBind's base, so
its fixture had neither, and it no longer built after the merge of
SegBind into macos/integrated (a34be2da). It gets the same stubs
SegBind gave test_set_for_tables' fixture: presence_words = 0, and a
seg_layout_count that keeps every count (its modules are plain
variants). No assertion changed.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 071c628f14ec46d0c55b1e5ef1a2941040a362aa)

feat(vk): SegXfer, the transfer consumers of a range across a segment's cut (L2) ... 6a383d426a

A guest range across a SegMem segment's cut is never one memory: gpu_buffer_bind
and gpu_buffer return it empty. Each SegXfer consumer's failure branch now binds
it piece by piece (gpu_span_bind, or gpu_span for the unswizzle's gpu_buffer
reader) after its operation's admission, and records the work per piece or
through the recording's SegArena (C0's declarations; SegBind fills the bodies):

- vk_replay.cpp: gpu_write as one vkCmdUpdateBuffer per piece (unaligned stays
  on the CPU); NOP_COPY_B2B as regions between the sides' pieces, one
  vkCmdCopyBuffer per buffer pair; NOP_CLEAR_BUF as one fill per piece.
- vk_xfer.cpp: B2T gathers the span into the arena and copies from it; T2B
  gathers in first (row padding keeps the guest's bytes), copies the image into
  the arena and scatters every byte back; bufferOffset a multiple of the texel
  block; stamp and texture_gpu_wrote over the span.
- vk_texture.cpp: the GPU unswizzle gathers its source into the arena in the
  upload command buffer, ahead of the dispatches, at the storage alignment.
- vk_temporal.cpp: the actor's bones and origin copied per piece, its vertices
  and indices gathered and read through the arena's device address.
- vk_draw.cpp: the frame dump, StatsQuery copies and mesh-indirect convert per
  piece; the mesh fallback's 8-byte indirect input gathered, by address.

No arena room is refused loudly (error log, counter), never silently. Every
path tests gpu_span_any() first: with no domain (all platforms but macOS under
alias isolation) every consumer records exactly what it did before, which the
new fixtures check against B2's recorded commands. The arithmetic is in the
new pure header gpu/vk/span_xfer.h.

Tests: tests/test_seg_xfer.py (+ seg_xfer_test.cpp), copy_b2t in
tests/test_vulkan_copy_span.py (+ copy_b2t_span_test.cpp), the gpu compile
fixture given SegMem's types, and SITES_XFER in test_resolver_coverage.py.
Docs: docs/segmented-binding.md § SegXfer.

Conflict on main 6c9edffe: vk_temporal.cpp actor_motion, #59 (actor
motion pairs copies by place): a pose read at its draw is copied from
the stage (placed), and only otherwise from the game's memory. The
piece-by-piece copy of a pose across a segment's cut goes in that
else branch, where the bones and origin are bound; the kept pose
takes #59's placed and place.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit db3c9bf0c7c9181001522991d2447321362b5724)

fix(vk): dispatch()'s compute-variant path (#252) names SegBind's pre-seam plain ... 1e1a3b3e0d

Semantic rework for main 6c9edffe, after the picks. SegBind renames the
program for the bound state to `plain` before the admission seam (after
it, `p` is that program or its segmented variant). Main's #252 added
the compute-variant path on those lines (program_cache_complete_ask,
on_worker, made_how), kept as main has it in the seg-bind pick, which
left vk_draw.cpp not compiling ("use of undeclared identifier 'p'", 7
lines). Those lines now name `plain`: the CAS and the ask, the job, the
cached pipeline, and the census and give-up after the 50 ms wait.

Behaviour is #252's: the plain program is the one read without its
code and made whole there. A segmented variant is translated, not read
from the cache, and compiled synchronously after the seam through
cached_compute, which takes #252's Intel gate itself (Windows Intel
only; segmented variants exist on macOS only). At the bind, #252's owed
optimized build and compute_optimized stay on `p`: a variant is never
built quick, so it binds its own pipeline.

The checks that read these lines follow:
test_program_cache_one_variant (its first-CAS, cached-pipeline and
complete anchors; its code_words check now looks for both names),
test_intel_compiler_gate (the ask's anchor), and test_seg_notes, whose
cut of the compute notes now ends where #252 queues the owed optimized
build (at the bind on a tree without it; its between() helper went with
the old cut). No check is weakened.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): the seg_gathers fixture stubs #162's rec::OwnSlot ... 03d88f59d7

Fixture follow-up for main 6c9edffe. tests/test_seg_gathers.py cuts
draw_begin_recording into its fixture, and main's #162 (the recorder
thread's own slot resets, RX_VK_RECORDER_OWN) marks the descriptor pool
reset there with a rec::OwnSlot scope, so the fixture did not build
("use of undeclared identifier 'rec'"). It gets the stub the other
fixtures that cut that function have (draw_stream_visibility_test,
recording_admission_test): a scope that does nothing. No check changed.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): SegXfer's fixture and resolver sites follow #59's actor motion ... 5a5f49731b

Semantic rework for main 6c9edffe. Main's #59 (actor motion pairs
copies by place) takes a pose read at its draw through a stage, and
copies the bones and origin from the game's memory only otherwise;
#P19's rework (pose_read_at_draw, shared by actor_motion and
temporal_actor_demands) makes the resolver demand them only there. The
SegXfer pick puts the per-piece pose copy in that else branch.

- tests/seg_xfer_test.cpp: the fixture's State takes #59's members
  (Skin::pose and kNoPose, pose_bytes, poses_time, Pose::placed and
  place, pose_stages, pose_stage_at, poses_read, poses_copied) and the
  <cmath> and <unordered_set> actor_motion now needs; its skins keep
  kNoPose, so its cases still take the copy from the game's memory,
  across a cut as before.
- tests/test_seg_xfer.py cuts pose_read_at_draw into the fixture too
  (when the tree has it).
- tests/test_resolver_coverage.py: SITES_XFER's four actor_motion sites
  expect the resolver's loop as SITES_BASE's do since #P19's rework
  (from pose_read_at_draw's 2 or 0), and the bones' site the same
  decision in both functions.

28 of 28 SegXfer cases and 74 binding sites in 19 functions pass again;
no expectation was dropped.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): SegXfer's gpu_write checks follow #329's ordered inline writes ... c22c9a9a11

Fixture follow-up for main 6c9edffe. Main's #329 (order overlapping
inline buffer updates) puts a full barrier before each vkCmdUpdateBuffer
in gpu_write, and tests that with test_inline_write_order, which cuts
gpu_write alone.

- tests/test_inline_write_order.py: SegXfer's span path in gpu_write
  calls gpu_span_any, gpu_span_bind and gpu_write_pieces; the fixture
  stubs them as they are off macOS and with no domain (gpu_span_any()
  false), so it tests #329's order as before. Its three variants keep
  their results (the barrier and the work count still caught).
- tests/test_seg_xfer.py: replay.gpu_write.common, recorded on the
  topic's base (08cccafb) without #329's barriers, is recorded again:
  the common path at this branch equals its base's (--base), on arm64
  and x86_64. The other eight common digests are unchanged.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): count polygon modes and restart-off strip draws on Apple, log-only ... 648232f043

KosmicKrisp has no fillModeNonSolid and MoltenVK no POINT polygons, and Metal restarts an indexed strip or fan at the
all-ones index whatever Vulkan's restart enable says. Before anything is fixed, count whether the game needs it
(mode_census.h, Apple only; nothing built or drawn changes):

- the graphics pipelines built, where st_pipelines counts them, and the draws recorded, where st_draws counts them, by
  NVN's polygon mode, with the first 8 LINE and POINT pipelines (program, pipeline key, topology) and programs named;
- the indexed strip and fan draws with restart off; a direct one's indices are read whole, a page at a time, through
  memory_read_as_submitted for the index type's all-ones value (0xFF, 0xFFFF, 0xFFFFFFFF). An indirect one's range is
  the GPU's and a range the CPU cannot read is not guessed: both are counted as not scanned.

Any other draw costs restart_off_strip's compares and its polygon mode's count on the thread's own counters. Printed
with the frame statistics every 120 presents ("at present N") and at close ("at close"). The two tests that compile
build_graphics and draw()'s binds from the source get a stand-in for the counts.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 1bd30dceb348909ca03c4cf12709dccbb0900f4a)

fix(macos): draw count draws through a record rewrite where the device has no drawIndirectCount ... 8512186c2f

MoltenVK 1.4.2 reports drawIndirectCount off, and its vkCmdDrawIndirectCount and
vkCmdDrawIndexedIndirectCount are empty functions (vulkan.mm:2390-2414). draw() called
them whenever the game's draw had a count address, so on the Mac every such draw was
dropped without a word (MT-01's u8_indirect_count step drew nothing).

Where the device lacks the feature (and only there; RX_VK_COUNT_REWRITE=1 forces it on
a device with it, to compare), a count draw's records are rewritten at draw()'s
pre-pass point, after the segment gathers and outside the pass, between two full
barriers: shaders/count_draws.comp copies the game's records under the count the GPU
reads into the recording's arena and zeroes the rest up to maxDrawCount (instanceCount
0: nothing drawn), and a plain vkCmdDraw(Indexed)Indirect draws maxDrawCount records
from there. Record i keeps its draw index and its own words, the count is clamped to
maxDrawCount by construction, and records at or past the count are never read. A draw
the rewrite cannot make (records off 4-byte boundaries, no pipeline, no arena room) is
left out loudly, never given to the empty function. Devices with drawIndirectCount
record what they did before.

A vk: count draws line with the frame statistics counts them a frame (indexed, 8-bit,
tessellated, pass breaks, maxDrawCount), only where the rewrite is on; vk_main says
when the device has no drawIndirectCount.

tests/test_count_draws.py runs the shader's own record code (count_draws_record.glsl,
shared with the shader) and count_draws.h's dispatch plan on the CPU over every count
from 0 to past maxDrawCount, strides, offsets, alignments, indexed and not, split
dispatches; checks the SPIR-V header against its source and the renderer's source.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit b97c61c845eb98bdd0fae4a7f2ea117cf2434d8e)

fix(macos): the count draw report in its own block; SegBind's fixture cuts the rewrite ... 7bd60d4cc0

The full suite at b97c61c8 failed two checks the change broke:
- test_mode_census holds the census call within three lines of draw_skip_report: the
  count draws' report now has its own block after the census's.
- test_seg_gathers compiles draw()'s pre-pass and indirect binds from the source: its
  fixture gets count_draws.h, the draw's topology and a count_rewrite stand-in that
  records what the rewrite is given, and a case on a device without drawIndirectCount:
  the rewrite reads SegBind's gathered records and the count in place, after the
  gathers' barrier and before the pass, and the draw reads the arena's rewritten
  records with the plain indirect call, never the driver's count draw.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit f09886c799272b781b2c3c1d4d99eed8da2719be)

fix(macos): buffer updates copied from the recording's stream on MoltenVK ... 5496309259

MoltenVK (1.4.2, its newest release) stages vkCmdUpdateBuffer's bytes in its
command pool's temporary MTLBuffers when it encodes, and every direct draw
whose pipeline reads the draw index zeroes the first word of the temporary
MTLBuffer its draw-index word lies in, not that word (MVKCmdDraw.mm:212 and
:495; fixed upstream after the release in bc0d5bf733). An update staged at
that buffer's start lost its first word before the GPU copied it. CR-01
(reports/2026-11-14-copy-race.md) measured it: all 1,010 failing trials lost
exactly the vkCmdUpdateBuffer word, in both argument-buffer modes, only with
a direct draw encoded after the update in the same pool (none with the pass's
draw left out or recorded first), while copies from the probe's own memory
landed in every trial.

cmd_update_buffer (vk_draw.cpp) records the renderer's buffer updates: on
MoltenVK the bytes go into the recording's stream (16-aligned) and a
vkCmdCopyBuffer from there takes the update's place in the command order;
without room, outside the recording's command buffer or with no recording it
records vkCmdUpdateBuffer as before, logged and counted. Every other driver
records vkCmdUpdateBuffer exactly as before. gpu_write and gpu_write_pieces
(vk_replay.cpp: event signals, uniform updates, the stream's writes) record
through it. The page table's update (vk_memory.cpp memory_table_flush) stays:
it is in the upload command buffer, whose pool records no draw.

Tests: tests/test_update_staging.py cuts cmd_update_buffer out of the source
and runs it on a model of the stream and of MoltenVK's clobber (fails before:
cmd_update_buffer is missing), and checks that gpu/vk has no other
vkCmdUpdateBuffer. tests/seg_xfer_test.cpp's model is not MoltenVK: its
cmd_update_buffer is vkCmdUpdateBuffer, and the common path's commands are
the base's (COMMON_SHA256 unchanged). VULKAN.md says so.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit e0e1ea92ed2f6e9e6d15ce1a088125cd8b6b34d9)

fix(macos): the count-draw rewrite on macOS only; Windows and Linux give count draws to the driver as before ... 61a098772e

b97c61c8 asked the rewrite of drawIndirectCount alone, on every system: a
Windows or Linux device without the feature would have changed from the
driver's vkCmdDraw*IndirectCount (what main calls there) to the rewrite. The
motivation and every measurement (CD-01/CD-02, CR-01, runs #25/#26) are
MoltenVK's, whose two count-draw functions are empty, so the rewrite, its
device line and its report are now macOS's alone (count_rewrite_on returns
false elsewhere; vk_main.cpp's line inside __APPLE__). Taking the rewrite to
every device without the feature is a change for every platform and needs
its own PR. RX_VK_COUNT_REWRITE=1 still forces the rewrite on a Mac whose
device has the feature (KosmicKrisp, MoltenVK 1.4.3).

test_count_draws checks both gates; VULKAN.md and count_draws.h say so.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

docs(macos): RX_VK_COUNT_REWRITE in the switch table ... b961c7e6ec

The topic named its switch in gpu/vk/VULKAN.md only; docs/development.md's
switch table gets its row: macOS only, off by default, the rewrite on a device
with drawIndirectCount too, to compare.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

test(vk): test_inline_write_order (main's, #329) has gpu_write's cmd_update_buffer ... 1ca0ed4792

Main's test_inline_write_order compiles vk_replay.cpp's gpu_write with a
recording spy; since e0e1ea92 gpu_write records its chunks through
vk_draw.cpp's cmd_update_buffer, which the fixture lacked (it stopped
compiling; the integration branch never had this test). Its stand-in is
cmd_update_buffer on any driver but MoltenVK: vkCmdUpdateBuffer through the
same work wrapper, so the barrier and work-count negatives still fail as
they must. test_update_staging runs the MoltenVK side.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

feat(vk): log-only frame cost lines, a GPU timeline and the bench's A/B exactness arm, all off by default ... 2120971489

What the Mac's 33 ms frame is made of could not be read from a run: there was no reason for each render pass end, no attachment
traffic, no count of what MoltenVK splits, no timing of the frame's command buffers against the CPU's clock, and no way to show a
change exact on the same frame. These are MacPerf's section 5 measurements (reports/2026-11-14-mac-perf.md), log-only:

- RX_VK_PERF_LINES=1: every 120 presents L1 (render pass ends by reason, and the passes begun again on the same attachments), L2
  (bytes loaded and stored by the passes, leading whole clears), L3 (MoltenVK only: 8-bit indexed draws, fans, passes with
  occlusion queries), L4 (NVN's discards, recorded into the command buffers only with this switch as NOP_DISCARD, and what reads a
  discarded target before it is rewritten), L6 (mesh fallback draws, their splits and batchable runs); every 600 presents L5
  (readback waits by thread and kind, and the part the present poll slept through).
- RX_VK_GPU_TIMELINE=1|2: the recordings' existing timestamps put on the CPU's clock with VK_KHR_calibrated_timestamps (asked for
  only with the switch): per frame, the GPU busy and idle times, the gaps between parts, submit to start, the last part to the
  game's present, the present to the next frame's first part; per frame type the counts; at 2, a timestamp pair per pass and
  dispatch and each type's top spans. Bounded memory, made only with the switch.
- RX_VK_BENCH_AB=pass_barrier|dedupe_off|chunk:<n> with RX_VK_BENCH: the replays in blocks A, B, A, each replay's GPU time and a
  hash of the window texture after it, and whether B's picture is A's bit for bit.

Off, each hook is one load of a flag set before the first replay; nothing drawn changes on or off. The host's present-poll sleep
clock includes the sleep in progress (read only while the FPS Breakdown or these switches time sleeps). Tests:
tests/test_perf_lines.py (perf_lines.h on the CPU, the source wiring, the hooks' cost off and on); fixtures that cut the touched
code get the stand-ins.

(Picked onto the count-draw rewrite, main's newer invalidate() with RX_VK_READBACK_LATE and RX_VK_ASYNC kept: L5's wait kinds
sit in its pending branch. The topic's added count on the readback line, the waits for a whole submission whose pages were
written in place, prints on every driver by default, so it is its own PR on main, fix/readback-whole-waits, not in this commit.
Only vk_perf.h is included here: the topic's other include lines were its base's.)

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 2c20c030308c637379690a3b4c9532eb888ddd53)

test(vk): test_inline_write_order (main's, #329) includes perf_lines.h for full_barrier's bench arm ... 46eabf24e4

PerfInstr's full_barrier asks perf_lines::bench_b("dedupe_off") (RX_VK_BENCH_AB's B arm; false unless the bench runs it).
main's test_inline_write_order (#329) compiles full_barrier cut out of vk_texture.cpp, and did not compile at this head
("use of undeclared identifier 'perf_lines'"; found by the wave's full suite: the topic's base predates the test). It now
includes perf_lines.h, as the topic did for test_dump_readback's fixture; its three variants (the current code and two negative
controls) run as before, both compilers.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

perf(macos): a pass's leading whole clears as its attachments' LOAD_OP_CLEAR on MoltenVK (MacPerf #7) ... 5ff63b446d

A render pass whose first operations clear whole attachments now begins with those attachments' LOAD_OP_CLEAR instead of
LOAD_OP_LOAD followed by vkCmdClearAttachments. On a tile GPU the load reads the attachment into tile memory and the clear then
overwrites it; Metal's clear load action only fills the tiles. Run #31's hold: 37.8 passes a frame begin with whole clears, 46.8
attachments' loads (76 MB of the 650 MB loaded a frame) are dead.

- begin_pass holds a pass's vkCmdBeginRendering when a whole clear opens it (Pass::lead, load_clear.h Held); each leading whole
  clear becomes its attachment's load op (the depth's and the stencil's each its own); pass_settle records the begin before anything
  else: begin_pass for a draw or another clear, end_pass, and the replay before any operation that is neither a clear nor a change of
  state (keeps_lead). Nothing is recorded between a pass's begin and its settle, so the command buffer is the one before, less the
  folded clears.
- Only where it is exact: every attachment the render area's size and single-sampled, no BC slice or raw view of BC (Targets::
  whole_extent: MoltenVK uses Metal's clear load action once the render area is the last attachment's level-0 extent, and that
  clears every attachment whole), the clear of the whole render area and every channel or aspect, and a value both of MoltenVK's
  ways write alike (load_clear.h: float formats' +0 and own normal values, integers in range and within 2^24, normalized 0 and 1,
  -1 for signed, D32's +0 and normal values, the unorm depths' +0 and 1). Everything else keeps vkCmdClearAttachments.
- MoltenVK's driver only (vk_vendor.h load_clear, RX_VK_LOAD_CLEAR=0/1). RX_VK_BENCH_AB=load_clear: off for the run but in the
  bench's B block. PerfInstr's L2 line counts what each leading whole clear became.
- tests/test_load_clear.py: the values format by format, whole_extent, and fixed and 8000 random sequences replayed with LS's own
  code against a model of MoltenVK's passes (the same memory, Vulkan's rules, the same commands but the folded clears).

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit e0d3af8f616b3e5829b900affd4531432147f0de)

fix(macos): load_clear's kept operations as a set of bits, its device line from the vendor fixes ... 5dbdde41d6

The full suite at e0d3af8f failed four source checks this change broke:
- test_bank_gen, test_parallel_dry_marks and test_same_bind find replay()'s cases by the first "case NOP_..." text in vk_replay.cpp,
  which keeps_lead's case labels now came before. keeps_lead is a set of bits (op_bit) instead, with no case labels.
- test_env_on: the device line read RX_VK_LOAD_CLEAR with getenv (set-means-on). It is now said on MoltenVK, or wherever the fix is
  on, from the vendor fixes' block (the switch itself is read by vendor_fixes_env, 0 off).

tests/test_load_clear.py reads the set of bits; its fixture cuts op_bit with keeps_lead.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>
(cherry picked from commit 0d15403fb9ae2823a44b8c4f0e046b5571485a37)

fix(vk): LoadClear off by default (mirrors macos/integrated 2b4b7b8e, I40) ... 7c6f0e9cee

LoadClear's half of macos/integrated's 2b4b7b8e (Integration 40), on the parent's word from the E-series
(reports/2026-11-15-mac-perf-before-after.md, field-dawn save, 600 s holds on an M4 Max; run-to-run noise 0.17 ms):
all on but LoadClear (E4c) 33.7 fps, GPU 19.2 ms, against all on (E2) 33.7 fps, GPU 19.1 ms. Measured under the noise,
and only what is measured faster is adopted, so it is off by default:

- vk_vendor.h: load_clear default false on every driver; RX_VK_LOAD_CLEAR=1 turns it on.
- docs/development.md: the switch row and the bench row say so.
- test_load_clear: the fix off on MoltenVK by default, =1 on; the source check pins x.load_clear = false.

The changed lines are 2b4b7b8e's for these files, but for the bench row (RX_VK_BENCH_AB): the same change to its
load_clear sentence, on the row as this branch has it, without the widen8 and mesh_batch arms the later topics add
(with them, at mac/mesh-batch, the row is 2b4b7b8e's). Widen8's half is its own commit on mac/widen8.

Assisted-by: Claude Opus 5.5
Co-Authored-By: Claude Opus 5.5 <[email protected]>

mooki referenced this pull request 2026-11-15 11:55:52 +00:00

WIP: fix(macos): restart-off 8-bit indices drawn as NVN does on MoltenVK; LS's widening ahead of a pass opt-in (RX_VK_WIDEN8=1) #782

mooki referenced this pull request 2026-11-15 11:55:57 +00:00

WIP: perf(macos): batch the mesh fallback's draws of a pass on MoltenVK #783

mooki added the

area/renderer

kind

performance

platform/macos

priority

high

labels 2026-11-15 13:59:46 +00:00

mooki force-pushed mac/load-clear from 7c6f0e9cee to 12fe15f62d 2026-11-15 15:05:27 +00:00 Compare

mooki referenced this pull request 2026-11-15 17:40:20 +00:00

WIP: perf(macos): Mac performance on MoltenVK #852

mooki commented 2026-11-15 18:09:22 +00:00

Author

Member

• Copy link
• Copy Markdown

Folded into #852 (perf(macos): Mac performance on MoltenVK)

Folded into #852 (perf(macos): Mac performance on MoltenVK)

mooki closed this pull request 2026-11-15 18:09:23 +00:00

Some checks failed

Hide checks

CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Has been skipped

Details

CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Has been skipped

Details

CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Has been skipped

Details

Third-party parts / The same as upstream (pull_request) Has been skipped

Details

Website / Website on Cloudflare (pull_request) Has been skipped

Details

PR description / Description follows the template (pull_request_target) Successful in 6s

Details

CI / Conventional Commit title (pull_request) Successful in 4s

Details

CI / Workflow lint (actionlint) (pull_request) Successful in 6s

Details

Guide screenshots / Guide images and local UI source review (pull_request) Failing after 6s

Details

Third-party parts / As their notes say (pull_request) Successful in 10s

Details

Pull request closed

Please reopen this pull request to perform a merge.

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels   

area/accessibility

Barriers for people with disabilities

  

area/audio

Sound and music

  

area/input

Controllers, keyboard, mouse

  

area/install

Installing, updating, building from your game

  

area/menu

In-game menu and settings

  

area/renderer

Vulkan renderer, shaders, effects

  

area/saves

Save files

  

area/upscaling

DLSS, FSR, frame generation

  

beanstalk: approved @ 1697603

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 2c5a62e

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 4fd82b7

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 64d6d09

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 6f91fae

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 70a78f9

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 8015c9a

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 8d361dc

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ 9d8356e

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ abf2227

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ d82f202

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ e5f1e40

Set by Beanstalk for show; the farm's database is what counts

  

beanstalk: approved @ ee8ab5f

Set by Beanstalk for show; the farm's database is what counts

  

ci: approved

A maintainer approved CircleCI for this fork PR (sticky; removed after 14 quiet days)

  

good first issue

A good start for a new contributor

  

gpu/amd

AMD GPUs

  

gpu/igpu

Integrated GPUs

  

gpu/intel

Intel GPUs

  

gpu/nvidia

NVIDIA GPUs

  

help wanted

Extra hands or knowledge wanted

  

kind

bug

Something doesn't work as it should

  

kind

chore

Build, CI, dependencies, tidying

  

kind

docs

Documentation

  

kind

feature

A new feature or an improvement

  

kind

performance

Frame rate, frame time, loading, memory

  

kind

question

A question rather than a change

  

needs measurements

Waiting for before/after measurements

  

platform/linux

Affects Linux (a change for every system carries both Windows and Linux)

  

platform/macos

macOS (Apple silicon or Intel Macs)

  

platform/steam-deck

Steam Deck

  

platform/windows

Affects Windows (a change for every system carries both Windows and Linux)

  

preview

Maintainers: deploy a preview of this PR's website (read its changes first). Removed after the run.

  

priority

critical

Crashes, data loss, or blocks most players

  

priority

high

Important: next in line

  

priority

low

Nice to have

  

regression

Worked in an earlier build, broken in a newer one: highest priority, above other bugs

  

status

blocked

Waits on something else first

  

status

confirmed

Reproduced or agreed on

  

status

duplicate

Already reported or proposed

  

status

invalid

Not a problem in this project

  

status

needs-info

Waiting for more information from the reporter

  

status

wontfix

Will not be worked on

  

testing wanted

Wanted: a test on other hardware (posts a test request in #test-requests)

  

ci: approved

No labels

area/accessibility

area/audio

area/input

area/install

area/menu

area/renderer

area/saves

area/upscaling

beanstalk: approved @ 1697603

beanstalk: approved @ 2c5a62e

beanstalk: approved @ 4fd82b7

beanstalk: approved @ 64d6d09

beanstalk: approved @ 6f91fae

beanstalk: approved @ 70a78f9

beanstalk: approved @ 8015c9a

beanstalk: approved @ 8d361dc

beanstalk: approved @ 9d8356e

beanstalk: approved @ abf2227

beanstalk: approved @ d82f202

beanstalk: approved @ e5f1e40

beanstalk: approved @ ee8ab5f

ci: approved

good first issue

gpu/amd

gpu/igpu

gpu/intel

gpu/nvidia

help wanted

kind

bug

kind

chore

kind

docs

kind

feature

kind

performance

kind

question

needs measurements

platform/linux

platform/macos

platform/steam-deck

platform/windows

preview

priority

critical

priority

high

priority

low

regression

status

blocked

status

confirmed

status

duplicate

status

invalid

status

needs-info

status

wontfix

testing wanted

ci: approved

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

MagicBeanDev/Longshot!780

No description provided.