Files
MobileGL/docs/Disaggregated/notes/p2/p2-results/magma-v3.md
T
swung0x48 aba329cc53 [Docs] (Disaggregated): P6's plan and contract draft; the working notes move from a local directory into the repository
Two things land together: the plan and contract draft P6 needs before it starts, and the
working notes that until now existed on exactly one machine.

1. P6's plan and contract draft (neither started; not one line of code)

     docs/Disaggregated/P6-SPAWN-PLAN.md      the package plan
     docs/Disaggregated/P6-CONTRACT-DRAFT.md  the contract draft (moves to
                                              MG_Remote/CONTRACT-P6.md when c6 lands)

   The three findings that mattered, all written into the documents:

   a. P6 is SMALLER than its roadmap cell suggests. The transport primitives are already in
      the tree and already tested: SocketDoorbell including hangup detection and the two
      lost-wakeup fences (Doorbell.cpp:211-360, four cases in FdPassingTest), SCM_RIGHTS fd
      passing, ShmSegment::Adopt, and the SurfaceOp / WindowKind schema already defined in
      protocol.fbs. inproc uses a real second shm mapping rather than an alias, and
      SessionRings.h:22-28 says that was done so today is not the first time it runs. P6 is
      assembly plus a control plane plus a process, not "write a socket transport".

   b. The evidence for "P6 is only a transport swap" has EXPIRED. It was bought by P5c, whose
      audit found 59 direct memory accesses in a tree that already looked clean. P5e changed
      the tree again. So the first package, a6, is a read-only audit that writes no code, and
      every contract row that depends on it is marked PENDING a6 rather than decided in
      advance.

   c. P6 does not get a game running on the phone. ANativeWindow* is meaningless in the child,
      and protocol.fbs:186-196's own comment says Android transfers the window out of band -
      that band is P12. P6 lands pbuffer / surfaceless / offscreen, which is what HeadlessGL,
      the split-equivalent lane and trace replay need, and which is exactly what P6's own exit
      gate measures. A real window reaching the server is a named refusal. This gets its own
      section because "P6 is done" would otherwise read as "the game runs now".

   The contract adds rule G: a role may not name a process-local handle. It extends P5c's
   rule E (a role may not name the other role's memory) from memory to identity - EGLDisplay /
   EGLSurface / EGLContext / ANativeWindow* / fd number / pid / mapped address either cross as
   a token this contract defines or they do not cross, and a descriptor crosses only through
   SCM_RIGHTS, where the kernel renumbers it.

   What P5e left for P6 gets its own section. The latch semantics are already designed
   (ARCHITECTURE.md:504) and are copied unchanged. What is new is that after run-ahead the
   client no longer waits while the peer process can DIE: the client holds records that were
   published and will never be applied, and Android's freezer suspends the child without
   closing anything. So slow and dead must be distinguishable WITHOUT a timeout threshold.
   That is P6's only new design; everything else is carriage. It shares a predicate with E1's
   outstanding re-specification (ID-122), and the two should be designed together.

   Two gates carried over from P5e's lessons: G2 requires integration-spawn's test-name set to
   match integration-split's name for name, because a differing set means the lane is quietly
   running less; and the arm-proof gate requires every entry to record the child pid and
   transport=spawn in its OWN private log - ID-124 established that this kind of false green
   really happens, and a spawn lane that actually ran monolith would pass every other gate.
   S2's shape must be settled before it is written, and the document says why: E1 is the
   cautionary case.

2. The working notes move into the repository (docs/Disaggregated/notes/, 415 .md files)

   ROADMAP / CURRENT_STAGE_PROGRESS / MEASUREMENTS cite ruling numbers, audit line numbers and
   the provenance of gate figures from ~/w7/notes/, a directory that exists on one machine and
   is under no version control. A record cited hundreds of times that lives only locally is not
   a record, so it moves in whole, with its directory structure intact.

   Only .md moves. Scripts, run logs, census and gate result JSON, patches, APKs and nm dumps
   stay where they are: they are evidence DATA rather than documents - large, regenerated per
   run, and mostly carrying absolute paths and device serials that mean something only on the
   machine that produced them. The mapping and this split are written in
   docs/Disaggregated/notes/README.md.

   Two exceptions are recorded there too: P6's BRIEF has been promoted to a first-class
   document and keeps no copy under notes/, so the two cannot drift apart; and recovered/ holds
   material salvaged from failed workflow runs, which may duplicate the like-named phase
   directories and is kept because a few of those were the only complete version at the time.

   CI's citation lint takes docs/Disaggregated/*.md only (non-recursive), and notes/ is
   deliberately outside it: those notes' file:line citations were written against their own
   code heads, and resolving them against today's HEAD would fail wholesale - which is not an
   error, it is their correct state as a historical record. To check one, use the revision it
   declares (--rev).

3. Inconsistencies fixed in passing

   - README.md's status line still said P5e was in progress while the stage table already said
     closed.
   - CURRENT_STAGE_PROGRESS's next-step item 1 still listed the exit gate as pending, though
     (a) and (b) were run in the previous commit; rewritten to closed plus the one item that
     actually remains (E1).
   - ROADMAP's P6 row still had the pre-phase plan in three of its cells; it now carries the
     package order, the scope boundary and the two new gates.

Checked with check_doc_citations.py: 125 citations across docs/Disaggregated/*.md, 0 problems.
2026-09-18 23:31:39 -04:00

31 KiB
Raw Blame History

P2 package D — p2/magma result (v3, rework round 3)

Tree: ~/w7/p2-magma, branch p2/magma, from the contract commit 9c6a8a25 (tag p2/contract). HEAD 93cf2249. Not pushed. Build dirs: build-linux (pull, the G1 build), build-push, build-nolegacy (-DMOBILEGL_PIPE_PUSH=ON -DMOBILEGL_PIPE_LEGACY_MEMOS=OFF) and — new this round, review v2 minor 8 — build-verify (-DMOBILEGL_PIPE_VERIFY=ON -DMOBILEGL_ITEST_REQUIRE_GPU=ON, which forces MOBILEGL_PIPE_PUSH=ON).

Verification scaffolding, not deliverables: ~/w7/p2-magma-armed (tmp/magma-armed = p2/magma merged with p2/tracker, re-merged this round at 840f8026, clean) is where the m1 arming evidence and the three red probes were taken. Its source was restored after every probe and re-verified green; git status on both trees is clean apart from the untracked build-nolegacy/ directory.

1. Commits

sha subject
d66f400f [Refactor] (Magma): key the pipeline memo on the render-state CSO handle and stop recomputing a hash the client already computed
553065dc [Refactor] (Magma): drive the dynamic tail from the pushed dynamic version and make the chunk table check DynamicTailKey's inventory
391e1230 [Refactor] (Magma): key the vertex-input cache and the VAO draw memo on {slot, gen} instead of a lifetime id and a heap address
fb410351 [Refactor] (State, Magma): take the backend's raw pointers out of the frontend VAO - the hash and state memos become the factory's own per-slot fields
a0e7c93d [Fix] (Magma): make the legacy-memo lever a startup gate for Magma's own bit, bound the {slot, gen} mint, and keep the all-pull arm free of push-only cost
1e89fd90 [Fix] (Magma): assert rather than assume that a handle indexed into a per-slot table is non-null
38a76e5d [Fix] (Magma): size the {slot, gen} mint by the live working set instead of by a capacity, and give it to the renderer that uses it
93cf2249 [Fix] (Magma): log the mint's high-water at a level a shipped build keeps

The last two are this round. Files touched by the branch are still exactly the six C.3 names (git diff --name-only 9c6a8a25..HEAD), all this package's by C.5: MG_Backend/DirectVulkan/Renderer/{MagmaPipeArms.h, VulkanRenderer.h, VulkanRenderer.cpp, VertexInputStateFactory.h, VertexInputStateFactory.cpp} and MG_State/GLState/VertexArrayState/VertexArrayObject.h.


2. MAJOR M1 — what changed, and the evidence

The review's finding was correct and I take all of it: the r2 MagmaPipeIdentityTable was a fixed 2048/8192-entry 2-way LRU, it introduced eviction where the content-hash memo and the resolved-state memo previously had none (they were unbounded mutable fields on VertexArrayObject), and every green gate in v2 §3 was green below both capacities.

2.1 The fix: no capacity at all, reclamation by age

MagmaPipeIdentityTable (MagmaPipeArms.h) is now a grow-on-demand mint:

  • UnorderedMap<lifetimeId, slotIndex> + a dense Vector<Entry> + a free list, with a one-entry front memo in front of the map probe;
  • OnFrameBoundary() retires slots whose object has not been drawn for kRetireAgeBoundaries = 1024 boundaries, swept every kSweepInterval = 256 — the same two numbers VertexInputStateFactory::OnFrameBoundary uses for the entries those slots key — and returns them to the free list. Gen moves on reuse, so reclamation is exactly as ABA-proof as the LRU was;
  • driven from VulkanRenderer's two existing frame-boundary sites, beside the factory's own.

Age-based reclamation is the stand-in for the frontend death notification P2 has no hook for (v2 MAJOR 3's reason for not using MG_Impl/Pipe/SlotAllocator is unchanged and still stated in the header). Footprint therefore tracks the live drawn working set, not objects ever created — which is what makes "no capacity" affordable.

Consumer tables:

  • VertexInputStateFactory::m_vaoMemos — the two memos that had no capacity before this package — now has none again. It follows the mint through MagmaPipeSlotTable<T>, a chunked table (256 entries/chunk) whose entry addresses never move, which is D12.4's grow-on-demand ask without a relocating Vector. 48 B per live VAO.
  • VulkanRenderer::m_vaoDrawMemoTable — the one memo that did have a capacity before this package — deliberately keeps the base ref's 2048 entries and the base ref's older-frameSerial victim rule, and changes only its key ({slot, gen} instead of a Fibonacci-mixed heap address plus a lifetime-id compare). A VaoDrawMemo is ~450 B (ResolvedVertexBindings dominates), so growing it with the live VAO set is megabytes on an LMK platform; losing one costs one vertex-binding re-resolve, exactly what it cost on the base ref.

Ownership: the two tables are a MagmaPipeIdentityTables member of VulkanRenderer, handed to its VertexInputStateFactory at construction, instead of two function-local statics (review v2 minor 4).

2.2 Evidence 1 — handle churn, the review's own workloads

scratchpad/wf5/p2-magma-r3/churn.py is a verbatim transcription of the landed Acquire / ClaimSlot / OnFrameBoundary, driven with exactly the workloads the review used against r2. Output at ~/w7/p2-magma-r3-churn.txt:

live objects 512 1024 2048 2500 3000 4000 8192 10000 16000
r2 (review, sparse ids) 3.1 % 20.0 % 58.1 % — — — — — —
r2 (review, consecutive) 0 % 0 % 0 % 54.2 % 95.2 % 100 % — 54.2 %¹ 100 %¹
r3, consecutive 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 %
r3, sparse 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 0.0 %

¹ the review's buffer-table row (8192 entries). Identical at 1 and 5 acquisitions per use.

Reclamation, same script: 8 waves × 500 VAOs × 400 frames = 4000 objects ever created, 500 live at a time → 2500 slots minted, 1500 live, 1000 on the free list. The table converges on the working set plus the retirement lag rather than on objects ever created.

2.3 Evidence 2 — the desktop corpus already reaches 69 % of r2's VAO capacity

The mint logs its high-water at power-of-two milestones (see §4, minor 1/D-16). On the armed tree's 40-trace retrace, minecraft-1.21.4-rd12-odinlite-in-world.DirectVulkan emits

MagmaPipeIdentityTable(VertexElementsCso): high-water 1024 slots minted, 1023 live
MagmaPipeIdentityTable(Buffer):            high-water 1024 slots minted, 1023 live

A throwaway 64-slot-granularity probe on the scaffolding tree (peak.sh, source restored, ~/w7/p2-magma-r3-peak-*.log) puts the peak at 1408 live VAOs and 1408 live buffers (±64) for that one trace.

That is 69 % of r2's 2048-entry VAO table on the existing desktop corpus, at render distance 12, and the review's own sparse-id measurement puts r2 at 20 % handle churn from 1024 live objects. So v2 §2 MAJOR 4's "nothing on desktop reaches those working sets" was false against this very corpus, and a device case at a higher render distance or with Sodium would have been past the capacity, not near it. No device time was needed to establish this.

2.4 Evidence 3 — the draw memo, the one table that keeps a capacity

drawmemo.py (~/w7/p2-magma-r3-drawmemo.txt) drives a 2048-entry, 2-way, older-serial-victim table with (a) the base ref's mixed heap addresses and (b) r3's dense {slot, gen}, and measures the steady-state miss rate (first frame discarded):

live VAOs 64 256 512 1024 2048 2500 3000 4096 8192
base ref (address hash) 0.0 % 1.2 % 6.4 % 24.0 % 60.0 % 69.5 % 79.1 % 91.0 % 99.6 %
r3 ({slot, gen} masked) 0.0 % 0.0 % 0.0 % 0.0 % 0.0 % 36.2 % 63.5 % 100 % 100 %

r3 is better everywhere up to 3000 live VAOs and equal-and-hopeless above 4096 (2 × capacity, where LRU on a cyclic pattern cannot win for any indexing — I also measured a skewed mask+mixed-alternate variant, which is no better; the numbers are in the same file). At no working set below 4096 is this table worse than what it replaces, and the two memos that used to be unbounded are no longer in this table's fate at all.

2.5 Evidence 4 — three red probes on the armed tree

All on ~/w7/p2-magma-armed, source git checkout-restored and re-verified 432/432 after each.

probe what expected observed
A kSweepInterval = 1, kRetireAgeBoundaries = 0 — every slot retired at every frame boundary, maximum reclamation churn green (correctness must not depend on the reclamation policy) 432/432
B probe A plus the ++Gen on slot reuse removed red, hoped 432/432 — NOT red. Reported honestly below
C every VAO's handle collapsed to a constant {1,1} (buffers untouched), so every VAO shares one entry of both consumer tables and inherits its predecessor's memos red 6 of 432 fail — DirectVulkan[.AsyncCompile].XfbAfterClipDistanceScenario.{SkipComponentsCaptureAfterClipDistanceWorkloadLines8, …Points1, SkipComponentsCaptureSurvivesEveryClipWorkloadStopPoint}, four of them SEGFAULT. Restored: 432/432

Probe A is the one that matters for M1: correctness is independent of when slots are reclaimed, so the only thing the retirement age buys or costs is memo recomputes.

Probe C is m3's ROADMAP.md:7 half: the suite can see a broken vertex-input identity.

Probe B is a negative result and I am not going to dress it up. Removing the generation bump does not turn the 432 red, because an inherited VaoBackendMemos still has to survive HashConfigVersion != vao.GetConfigVersion() and an inherited VaoDrawMemo still has to survive the per-draw bindings revalidation (frame serial, content hash, live buffer pointers, slice epochs). Only a deliberately constructed ABA — same address, byte-identical attribute configuration, different buffer — defeats all of those at once, and that construction is D18's HandleRecycleScenario, which is package E's file and is not on this tree. So: the Gen discipline is load-bearing by construction and is not covered by any test this package can reach. G8 is the gate; the integrator must not read this package's green 432 as covering it.


3. Verification — every command and its actual result

All from ~/w7/p2-magma at 93cf2249 unless stated. Driver scripts under scratchpad/wf5/p2-magma-r3/, logs under ~/w7/p2-magma-r3*.

# command result
V1 cmake --build build-{linux,push,nolegacy,verify} -j 12 rc 0 in all four
V2 python3 scripts/symbol_report.py --before ~/w7/p2-before-libMobileGL.so --after build-linux/libMobileGL.so --threshold 0 27799 -> 27799 defined symbols: 0 added, 0 removed, 4 resized, 0 renamed; .text +160. The four are the contract's (RenderState::RenderState(), SetCapability, IsCapabilityEnabled, _GLOBAL__sub_I_DirectGLES.cpp) (G1)
V3 same, --before ~/w7/p2-contract/build-linux/libMobileGL.so 0 added, 0 removed, **0 resized**, 0 renamed, .text +0, 27060 of 27060 normalised names unchanged. This package's pull build is byte-identical to the contract's (D-13 still holds after the rewrite)
V4 awk '/^ namespace RenderStateImpl \{/,…/' DirectGLES.cpp | sha256sum d8fd1c48…0efe27, equal to ~/w7/p2-before-syncrenderstate.sha (G5)
V5 grep -rc 'pGLContext' MobileGL/MG_Backend | grep -v ':0$' empty (G13)
V6 python3 scripts/check_include_closure.py 4 probes, 0 skipped, 0 problem(s) (G13)
V7 python3 scripts/gen_pipe.py --check / --self-test rc 0 / 7 negative-control trip(s), positive control OK (G13)
V8 stdio grep over MG_Backend/DirectVulkan + MG_State/.../VertexArrayState no stdio (three hits are Vector<…> inputs( matching the puts( pattern) (G13)
V9 ctest -N names, working grep, four build dirs build-linux / build-push / build-nolegacy = 2367, byte-identical to each other; build-verify = 3185, a strict superset (818 verify-only, 0 missing). vs the 2363 baseline: 0 removed, 4 added, all four the contract's MG_Test/Pipe placeholders (G2 name half, G14)
V10 ctest -L unit -j 8 in all four dirs 1489/1489 in each
V11 ctest -L integration-gpu -j 4 -R DirectVulkan, build-push, default mask 432/432
V12 same, MOBILEGL_PIPE_PUSH=0 (the all-pull control) 432/432
V13 same, MOBILEGL_PIPE_LEGACY_MEMOS=0 at the default mask 432/432
V14 same, MOBILEGL_PIPE_PUSH=0x60 MOBILEGL_PIPE_LEGACY_MEMOS=0 (v1 MAJOR 1's repro) 432/432
V15 same, build-nolegacy 432/432
V16 same, build-linux (pull) 432/432
V17 ctest --test-dir build-verify -L integration-verify -j 4 -R DirectVulkan 409/409 (G4's DirectVulkan half — new this round, review v2 minor 8)
V18 retrace_gate.py --tree ~/w7/p2-magma --lib build-push/libMobileGL.so --out ~/w7/retrace-out/p2-magma-r3f -j 4 --only 'DirectVulkan$' 40/40 PASS, min SSIM 0.993475 (minecraft-1.21.11-main-menu), gate 0.99 (G3, DirectVulkan half)
V19 git diff --name-only 9c6a8a25..HEAD the six files C.3 names, nothing else (C.5)
V20 grep -rn 'BackendStateMemo|BackendAuxMemo|m_backendHashMemo' MobileGL/MG_State MobileGL/MG_Backend 15 hits outside ProgramObject, every one inside a MOBILEGL_PIPE_LEGACY_MEMOS arm (D-8)
V21 grep -rn "MagmaPipeVaoIdentity|MagmaPipeBufferIdentity|MagmaPipeHandleOf|kMagmaVaoIdentityEntries|kVaoMemoSlotCount" MobileGL/ none — the process-global accessors and the two capacity constants are gone (minor 4)
V22 MOBILEGL_PIPE_PUSH=0x20 MOBILEGL_PIPE_LEGACY_MEMOS=0, 13 DirectVulkan cases, MOBILEGL_LOG_FILE_PATH one Fatal{PipeLegacyMemosDisabled} naming bit 6 at startup; ctest reports a visible SKIP, not a mid-frame abort. D14 semantics unchanged
V23 MOBILEGL_PIPE_PUSH=0x7e MOBILEGL_PIPE_LEGACY_MEMOS=0, same 13 cases 13/13 green and one startup WARN naming kMGPipeSubsystemRenderState (bit 0 …) — the silent case review v2 minor 2 found. Control at the default mask: 0 such warnings
V24 armed tree (840f8026): ctest -L integration-gpu -j 4 -R DirectVulkan 432/432
V25 armed tree: 13 CrossFrameBuffer cases with MOBILEGL_LOG_FILE_PATH 0 no-CSO fallback warnings (delivered tree, same cases: 1)
V26 armed tree: retrace_gate.py … --only 'DirectVulkan$' 40/40 PASS; 0 of 40 logs carry the fallback warning (delivered tree: 80 files)
V27 red probes A / B / C (§2.5), each restored and re-verified 432/432 · 432/432 · 6 failures, restored 432/432
V28 ~/w7/p2-magma-armed, ~/w7/p2-magma git status --short clean apart from the untracked build-nolegacy/

Nothing flaked this round: the two …IsActuallyArmedWhenTheEnvironmentPinsItOn cases that failed intermittently in v2 passed in all 13 -j 4 integration runs here.


4. The minors

# disposition
1 — MGLOG_W_ONCE on the per-draw path, cost misdocumented Fixed. The latch is now a mutable Bool m_pipelineCsoFallbackWarned on the renderer and the log is MGLOG_W. The comment says what MOBILEGL_LOG_ONCE_INTERNAL actually is (an unconditional std::atomic_flag::test_and_set, i.e. a locked xchg per evaluation) and why this site may not pay it. Per renderer rather than per process is also the right scope.
2 — MOBILEGL_PIPE_LEGACY_MEMOS=0 with bit 0 clear is silent Fixed as a diagnostic, declared as D-15. MagmaPipeValidateSubsystemConfiguration now names the combination at startup with MGLOG_W. It is not made fatal: bit 0 is not Track H (D14 labels only bits 5/6 that) and unlike bits 5/6 there is a correct answer — the state hash. Reproduced green with the warning at V23.
3 — §4's inference over-stated Fixed. The comment now states exactly what the warning's absence proves ("no draw took the fallback while bit 0 was set") and that with bit 0 clear it proves nothing, because the function returns before the latch and no draw is keyed on a handle either.
4 — process-global identity tables Fixed. MagmaPipeIdentityTables is a VulkanRenderer member handed to its factory; the two function-local statics are gone (V21). Both consumer tables were already per-instance, so all four now share one lifetime and one reclamation clock per context.
5 — 553065dc's subject overclaims Not fixed; argued. The instruction for this round is "as new commits on top", and rewriting that message means rebasing the four commits the review cites by sha, invalidating every citation in magma-review-v2.md and magma-v2.md for a message. The finding is correct and recorded here instead: 553065dc contains only comment and static_asserts; the dynamic tail's re-sourcing is the contract's, not this commit's (D-2 says so). Integrator: if you squash at merge, the honest subject is [Refactor] (Magma): make the chunk table check DynamicTailKey's inventory. My two new commits describe only what they change.
6 — D19's DynamicChunksCoverMagmasDynamicTailKey exists nowhere as a ctest entry Cannot be fixed here; made greppable and routed. MG_Test/Pipe/RenderStateSpansTest.cpp is package A's file (C.5) and MG_IntegrationTest/** is package E's, so this package has no file it may put a ctest entry in. The static_assert block in VulkanRenderer.cpp now names DynamicChunksCoverMagmasDynamicTailKey in its comment so grep finds it, and says explicitly that the integrator must make sure the outcome is not "neither". Blocking-for-integrator item, §8.3.
7 — D17's AccessorCalls re-audit not done Done. Result below.
8 — build-verify never configured Fixed. build-verify is configured and built on this tree; unit 1489/1489 and integration-verify -R DirectVulkan 409/409 (V1, V10, V17). The verify-only assertions are now exercised here, not just proven to compile.
9 — both assertions are compiled out of every build P2 runs Fixed for the one that could be, and the reason the other cannot is now on record. The Gen-wrap defence is on the release path: a slot that reaches generation 2^32-1 is permanently retired (and warned) rather than wrapped. The MagmaPipeSlotIndex null check keeps its release-path ternary. And a new fact the reviewer should have: a DEBUG-log-level build of this tree — the only configuration in which MOBILEGL_ASSERT is live — does not compile at all. MG_Util/Types.h:153,159,170,174 uses MOBILEGL_ASSERT where MGLOG_F is not declared; the file is untouched since the base ref 48268068 and belongs to no package. Log: ~/w7/p2-magma-r3-build-dbgassert.log (configure -DMOBILEGL_LOG_ACTIVE_LEVEL=MOBILEGL_LOG_LEVEL_DEBUG, rc 1, first failures in MG_Util/Converters/MGToGL/ErrorCodeConverter.cpp and MG_State/GLState/ErrorState/Error.cpp, neither of which includes anything this package touches). So no assertion anywhere in MobileGL is exercised by anything today. Routed to the integrator, §8.7.
10 — brief-vs-tree items Re-confirmed and re-recorded in §6, unchanged.
11 — MGPipeApplier() is one process-global Cannot be fixed here; routed and documented at the read site. MG_Pipe/PipeApply.cpp is package A's. A comment at ResolveBoundRenderStateCso now names the hazard (a multi-context process reads whatever CSO another context last bound) so neither review can assume the other caught it. Blocking-for-integrator item, §8.2.

D17 re-audit of the AccessorCalls tally constants (minor 7)

D17 names eight sites; all eight have moved and are re-audited at their current lines. The tallies are unchanged and the audit says why:

site (now) tally what it counts verdict after m1/m2/m3
VulkanRenderer.cpp:5220, :5227 1 the single MGB_CTX->GetPipelineStateVersion() read at :5173, which both arms make still correct. The reviewer's suspicion was that the handle arm no longer performs ComputePipelineStateHash's bulk GetRenderStateParameters() fetch — it does not, but ComputePipelineStateHash has never carried an AddCalls of its own (grep -c AddCalls over its body = 0). So m1's saving is invisible to acc/draw: the counter under-reports the win rather than over-reporting it
:5485 15 the payload-builder walk on a pipeline-memo miss untouched by this package
:6156, :6163 1 / 2 GetRenderStateParametersVersion(), and that plus the bulk parameter fetch that builds DynamicTailKey untouched: m2 re-sourced the value, not the number of reads
:6685 6 the fast path's six unconditional accessor reads untouched
:6718 1 the draw-program read, beside the Draws denominator untouched
:6731 3 the full path's three immediate reads untouched

Rule for the report (D17's precondition): acc/draw may be quoted for P2, but it must be quoted with the note that Magma's largest single removal — one bulk GetRenderStateParameters() per draw whose pipeline-state version moved — is not in the tally, so the acc/draw delta is a lower bound on Magma's side. Lead with the gate hit/miss pairs (MagmaPipelineMemo, MagmaDynamicTail, MagmaDrawFastPath) and the CPU-time series.


5. Deviations from the brief, with reasons

Carried from v2, still true and unchanged: D-1 (the enumeration comment did not move to MGPipeRenderStateSpans.cpp, package A's file), D-2 (the dynamic tail's version gate is re-sourced but not finer-grained; the tree is right against D12.3 and the reviewer confirmed it), D-3 (D12.3's coverage check is static_asserts here, not a ctest entry; see minor 6), D-4 (ScissorTestEnabledMask is pipeline state, so D12.3's blanket claim is false as written and the assertion is inverted), D-6 (a VAO's MGPipeKind is VertexElementsCso), D-8 (the three VertexArrayObject memo triples are #if MOBILEGL_PIPE_LEGACY_MEMOS, not textually deleted, because a literal deletion changes the pull build), D-9 (the eviction epoch is de-globalised, not removed), D-10 (negative control C lives here although C.4 gives HandleRecycleScenario to package E), D-11 (MG_Backend/DirectVulkan no longer includes MG_Impl/Pipe/SlotAllocator.h), D-12 (D12.1's version→hash gate and cached-hash fields survive as the no-CSO fallback's cache), D-13 (the pull build's two pipeline-memo sites keep the base ref's text under #else).

Rewritten or new this round:

  • D-5 (rewritten) — D12.4 asks for a grow-on-demand Vector. Two of the three tables now are grow-on-demand (the mint and m_vaoMemos), which is what the brief asked for; the third, m_vaoDrawMemoTable, deliberately stays fixed at the base ref's 2048 entries with the base ref's victim rule, for the memory and no-regression reasons in §2.1/§2.4. m_vaoMemos grows through a chunked table rather than a Vector because a growing Vector relocates entries the draw path holds references into.
  • D-7 (rewritten) — Magma still mints its own {slot, gen} rather than using MG_Impl/Pipe/SlotAllocator, for v2 MAJOR 3's unchanged reason (nothing in P2 can call its Free). What changed is that the mint is now unbounded-but-reclaimed instead of bounded-with-eviction: it has a Free of its own, driven by age instead of by a death notification. Integrator: when object deletion reaches the client allocator (P3a/P3b) the right end state is still MGPipeSlots() with a live Free and MagmaPipeIdentityTable deleted; the age-based sweep is what to delete with it.
  • D-14 (new) — an acquisition is no longer "a mask plus at most two Uint64 compares". It is a one-entry front-memo compare for every acquisition after a draw's first (five or six of the six a draw makes), and a hash-map probe otherwise. That is still less than the address multiply plus two-way probe the pre-handle arm ran, and the front memo lives inside Acquire so no call site can disagree with another about it.
  • D-15 (new) — D14's runtime row ("false: the legacy arm is never entered") is not literally true for bit 0 with MOBILEGL_PIPE_LEGACY_MEMOS=0, because D14's compile row names ComputePipelineStateHash as part of the pre-handle arm and a no-CSO draw has to key on something. Magma warns at startup instead of aborting. See minor 2.
  • D-16 (new) — D20 says MGLOG_D for anything non-critical. The mint's high-water line is MGLOG_I, at power-of-two milestones from 1024, on the allocate-a-new-slot branch only. Reason: MGLOG_D is compiled out of every shipped build and of every build P2 measures, and this line is the live-object measurement review v2's M1 asks for (§2.3 is that line doing its job on the existing corpus). At most a handful of lines per session; never one on a draw, so ROADMAP.md:7 holds.

6. Tree-versus-brief contradictions

Unchanged from v2 and independently re-confirmed by the reviewer; repeated so the integrator has them in one place.

  1. D12.3's premise about bind_render_state — the applier writes both versions and has to (Espryt's SyncRenderState uses that counter and G5 forbids touching it). Tree wins.
  2. D12.3's claim that every DynamicTailKey input is dynamic — ScissorTestEnabledMask is pipeline state. Tree wins.
  3. D12.1 lists three cached-hash fields; there are five.
  4. D12.1 does not mention the second memo probe in TrySetupDrawFastPath.
  5. Section A's G2/G14 grep is wrong. grep -E '^\s+Test #' matches only 4-digit ids (ctest -N right-aligns) and returns 1368 of 2367 names here. The working form is grep -E '^\s+Test\s+#[0-9]+:' | sed -E 's/^ *Test *#[0-9]+: //'. Fix BRIEF-P2.md before wsl_integrate_p2.sh uses it.
  6. G1/D15's admitted resize set is short by one. _GLOBAL__sub_I_DirectGLES.cpp (−9) is a fourth resized symbol and it is the contract's (V3 proves this package adds none). D15 admits three, so G1 is red by its own text after every merge until D15 is amended.
  7. C.3's verification block prints grep -rn 'BackendStateMemo…' # only ProgramObject's. With D-8 in force the honest check is "every hit is inside a MOBILEGL_PIPE_LEGACY_MEMOS arm, plus ProgramObject's" (V20).
  8. MGLOG_F_ONCE does not exist (MG_Util/Debug/Log.h has _ONCE for D/I/W/E only).
  9. G3's command passes --ssim 0.99 and C.3's passes --backend; ~/w7/retrace_gate.py has neither option (--tree --lib --out -j --only only) and exits 2. Its built-in threshold ran instead. Fix the brief or the script before D.3.
  10. MOBILEGL_PIPE_PUSH=0x20 with MOBILEGL_PIPE_LEGACY_MEMOS=0 is fatal to a DirectVulkan run at startup, naming bit 6. That is D14's own semantics, not a defect: a per-subsystem A/B with the legacy lever off must set the bit belonging to the backend under test.
  11. New this round: a DEBUG-log-level build of feat/disaggregated does not compile (MG_Util/Types.h, untouched since the base ref). Every MOBILEGL_ASSERT in the tree is therefore dead code in every buildable configuration. See minor 9 and §8.7.

7. Flakes

None this round. Thirteen -j 4 integration runs (six configurations on the delivered tree, one verify lane, four probe runs on the armed tree, two restores) produced no intermittent failure. v2's two …IsActuallyArmedWhenTheEnvironmentPinsItOn flakes did not recur.


8. Unfinished, and what the integrator must carry

  1. G11 / D.4.2 is still not run. Nothing in this package can run it. What this round adds is that M1 no longer depends on it: the structure has no capacity to be sized, and the two workloads' live-object counts are now readable from a shipped INFO build's log (MagmaPipeIdentityTable(...): high-water N slots minted, M live). Please grep for that line in the device run and put the two numbers in MEASUREMENTS.md.
  2. Blocking, package A: MG_Pipe::MGPipeApplier() is one process-global g_applier (PipeApply.cpp:99,112), not the per-context CSO store D2 specifies. Magma reads it directly and is its only P2 consumer, so in a multi-context process one renderer's pipeline memo keys on whatever CSO another context last bound. Named in a comment at ResolveBoundRenderStateCso; the fix belongs in package A.
  3. Blocking, D19 outcome: DynamicChunksCoverMagmasDynamicTailKey is static_asserts in VulkanRenderer.cpp here and a named ctest case in package A's RenderStateSpansTest.cpp there. Make sure the outcome is not "neither"; both is fine and intended.
  4. G8 / HandleRecycleScenario is the only thing that covers the Gen discipline. Red probe B (§2.5) shows the 432 integration cases do not. Package E owns it.
  5. D-5 / D-7: revisit the mint when object deletion reaches the client allocator (P3a/P3b).
  6. D-2: decide whether to add a dynamic-only version to the wire. D-1: the enumeration comment's move into MGPipeRenderStateSpans.cpp.
  7. MOBILEGL_ASSERT is dead everywhere. A DEBUG build of the tree does not compile (MG_Util/Types.h uses it before MGLOG_F is declared). Either fix that include order or stop pretending the assertions are a defence. Not this package's file; one-line fix.
  8. Brief fixes: §6.5 (the G2/G14 grep), §6.9 (--ssim / --backend), §6.6 (amend D15's admitted resize set with _GLOBAL__sub_I_DirectGLES.cpp).
  9. 553065dc's subject overclaims — see minor 5 for the honest one if you squash.
  10. A check_include_closure.py probe rooted at an MG_Backend header would actually gate ARCHITECTURE.md:9; today G13 says nothing about it either way. The script is not this package's file.
  11. MEASUREMENTS.md input from this package. The pipeline memo compares an 8-byte handle instead of an 8-byte hash and skips ComputePipelineStateHash entirely (≈50 CombinePipelineStateWord rounds plus one bulk GetRenderStateParameters() fetch that no AccessorCalls tally counts — see the D17 re-audit) on every draw whose pipeline-state version moved. LookupVaoDrawMemo drops a multiply and one of two compares and, below 2048 live VAOs, never collides at all (0.0 % vs the base ref's 6.4 % at 512 and 24.0 % at 1024 — §2.4). New steady-state cost: one hash-map probe per draw per kind behind a one-entry memo, and ~100 B per live VAO plus ~50 B per live buffer of mint and memo storage, reclaimed when the object goes idle. None of it is measured on hardware here; that is D.4.
  12. The armed tree (~/w7/p2-magma-armed, tmp/magma-armed @ 840f8026) is scaffolding and can be deleted after the real integration.