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.
23 KiB
P2 package D — p2/magma result (v2, rework round 2)
Tree: ~/w7/p2-magma, branch p2/magma, from the contract commit 9c6a8a25 (tag
p2/contract). Not pushed. Build dirs build-linux (pull, the G1 build), build-push,
build-nolegacy (-DMOBILEGL_PIPE_PUSH=ON -DMOBILEGL_PIPE_LEGACY_MEMOS=OFF).
One extra tree exists for verification only: ~/w7/p2-magma-armed, branch
tmp/magma-armed = p2/magma merged with p2/tracker (clean merge, no conflicts — which
is itself evidence for C.5's ownership split). It is not a deliverable; it exists because
m1 cannot be exercised on a tree with no tracker to bind a render-state CSO (review MAJOR 5),
and it is where the arming evidence in §4 was taken. No file in any other package's tree was
edited.
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 |
The first four are v1's, unchanged. The last two are this round. Files touched are the same
six as v1 — MG_Backend/DirectVulkan/Renderer/{MagmaPipeArms.h, VulkanRenderer.h, VulkanRenderer.cpp, VertexInputStateFactory.h, VertexInputStateFactory.cpp} and
MG_State/GLState/VertexArrayState/VertexArrayObject.h — all this package's by C.5.
2. How each major was fixed
MAJOR 1 — MOBILEGL_PIPE_LEGACY_MEMOS=0 aborted every draw through a non-Track-H bit
MagmaPipeRequireLegacyArm (per-draw MGLOG_F + std::abort()) is deleted. In its place
MagmaPipeValidateSubsystemConfiguration() runs once, from VulkanRenderer::Initialize(),
which is what D14 actually specifies ("a Track-H subsystem whose bit is clear is a startup
Fatal{PipeLegacyMemosDisabled}"). Three boundaries, each of which the per-draw shape got
wrong:
- it checks only bit 6 (
kMGPipeSubsystemMagmaVertexInput), and only when Magma is the backend being brought up, so Espryt's bit 5 cannot kill a DirectVulkan run and vice versa; - bit 0 is not consulted at all. It is not Track H (D14 labels only bits 5/6 that) and it is not a memo re-key — it decides where the pipeline memo's state key comes from. A clear bit 0 now silently selects the state hash;
- the "no CSO bound" case is a fallback, not an abort, in every build (see below).
Every re-keyed Track-H site now asks one helper, MagmaPipeTrackHArmIsHandles(bit), which is
MagmaPipeSubsystemOn(bit) where the legacy arm is compiled and a compile-time true where it
is not (the startup gate has already made a clear bit fatal there). That also removes minor 3:
VertexInputStateFactory::ComputeHash was the one site deciding for itself.
Reproduction of the review's exact commands, on the final commit:
| command | v1 | v2 |
|---|---|---|
MOBILEGL_PIPE_PUSH=0x60 MOBILEGL_PIPE_LEGACY_MEMOS=0 ctest -L integration-gpu -j 4 -R DirectVulkan |
9/9 "Subprocess aborted" on the CrossFrameBuffer subset | 432/432 |
MOBILEGL_PIPE_LEGACY_MEMOS=0 at the default mask |
aborted | 432/432 |
MOBILEGL_PIPE_PUSH=0x20 MOBILEGL_PIPE_LEGACY_MEMOS=0 (Espryt's bit alone) |
aborted mid-draw | startup Fatal naming bit 6, the harness's pre-flight child catches it and the case SKIPS visibly with "no usable GPU/display/ICD for backend DirectVulkan" — the D14 semantics, loud, before any frame |
The no-CSO fallback. delete_render_state clears BoundRenderStateCso
(MG_Pipe/PipeApply.cpp), so the null handle is reachable on any tree, integrated or not.
The fallback is the pre-handle ComputePipelineStateHash where one is compiled, and the
client's own MGPipeComputePipelineSubsetHash over the same 396 pipeline bytes where it is
not (ComputePipelineSubsetStateHashFallback). That is what makes
-DMOBILEGL_PIPE_LEGACY_MEMOS=OFF a runnable configuration: v1's build-nolegacy was
180/432 with Fatal{PipeLegacyMemosDisabled} on every draw, v2's is 432/432.
The fallback logs MGLOG_W_ONCE rather than MGLOG_D_ONCE. W is compiled in at every shipped
log level and _ONCE costs one static bool test, so a run whose pipeline memo never keys on a
CSO handle now says so in its log — which is the thing that made MAJOR 5 invisible, and which
§4 turns into the arming evidence.
MAJOR 2 — the all-pull control arm was contaminated
snap.vaoHandle = ResolveVaoHandle(vao) at both stamping sites is now inside
if (MagmaPipeTrackHArmIsHandles(kMGPipeSubsystemMagmaVertexInput)), so
MOBILEGL_PIPE_PUSH=0 mints nothing and compares nothing for a field that arm never reads.
D14's "reproduces P1's behaviour exactly" holds again, and D.4.3's T2 is back to meaning "P1's
residual fill alone".
MAJOR 3 — unbounded, never-reclaimed growth of the client slot allocator
Magma no longer uses MG_Impl/Pipe/SlotAllocator at all. grep -rn "MGPipeSlots()" MobileGL --include=*.cpp --include=*.h | grep -v MG_Impl/Pipe/SlotAllocator now returns one comment
line and no code. The reason the allocator was wrong here is not a defect in the allocator:
nothing in P2 can call its Free. There is no frontend death notification Magma can hook —
BufferBackendOps::OnDestroy is handed a BackendBufferResource, not the BufferObject, and
only fires for a buffer that ever had one; VertexArrayObject has no hook at all, and adding
one is D13's explicit-destroy work, which covers Espryt's six kinds and not
VertexElementsCso.
In its place, MagmaPipeIdentityTable (in MagmaPipeArms.h): fixed capacity, 2-way
set-associative, indexed by lifetime id, LRU victim inside the set, Gen incremented whenever
a slot changes owner. Two instances — 2048 entries for VertexElementsCso (32 KB), 8192 for
Buffer (128 KB). Bounded, nothing to reclaim, and exactly as ABA-proof as the allocator's
{slot, gen}: a recycled slot always carries a moved generation, so a stale handle cannot
match. It also takes MG_Backend's only MG_Impl include back out (minor 4) and turns the
per-attribute hash-map probe into an array index (minor 2).
MAJOR 4 — the density premise was false and the mitigations had been deleted
Both per-slot tables are now a bijection with the mint rather than a masked direct map:
kVaoDrawMemoSlotCount == kMagmaVaoIdentityEntries and
VertexInputStateFactory::kVaoMemoSlotCount == kMagmaVaoIdentityEntries are static_asserts,
entry i is slot i + kMGPipeFirstAllocatableSlot, and the index is
MagmaPipeSlotIndex(handle) with no mask. Two live VAOs cannot land on one entry at all, so
the birthday-collision scenario the review constructed cannot happen. The mitigations are not
gone, they moved one level down and are now written once: the second candidate is the
identity table's second way, and the frame-serial victim choice is its LRU. The comments that
argued the false density claim are rewritten.
The residual cost this design does carry, stated plainly: with more live objects than the
table holds, an LRU eviction changes a live object's handle. For a VAO that costs a memo
recompute (both tables are caches; correctness is the exact {slot, gen} compare). For a
buffer it changes the vertex-input content hash, which costs a rebuilt
BackendVertexInputState — but ComputeHash is memoised per VAO on its config version, so
that is paid per VAO reconfiguration, not per draw, and the buffer table is sized 4× the VAO
one for it. It is not measured here; nothing on desktop reaches those working sets.
MAJOR 5 — m1 was executed by no test
Fixed by verification, not by code: ~/w7/p2-magma-armed = p2/magma + p2/tracker (D.1's
own order), where the tracker emits create_render_state/bind_render_state and the applier
publishes BoundRenderStateCso. See §4 for the four runs; the short version is that on that
tree 432/432 integration cases and all 40 DirectVulkan retraces run with zero fallback
warnings, i.e. every draw's pipeline memo was keyed on the CSO handle, and a red probe that
collapses the handle to a constant turns 4 of 59 render-state-sensitive scenarios red.
3. Verification — every command and its actual result
All from ~/w7/p2-magma at 1e89fd90 unless stated.
| # | command | result |
|---|---|---|
| V1 | cmake --build build-linux -j 12 |
rc 0 |
| V2 | cmake --build build-push -j 12 |
rc 0 |
| V3 | cmake --build build-nolegacy -j 12 |
rc 0 |
| V4 | 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; the four are the contract's (RenderState::RenderState() +148, SetCapability +77, IsCapabilityEnabled +29, _GLOBAL__sub_I_DirectGLES.cpp −9). This package adds none. (G1) |
| V5 | grep -rc 'pGLContext' MobileGL/MG_Backend | grep -v ':0$' |
empty (G13) |
| V6 | python3 scripts/check_include_closure.py |
4 probes, 0 skipped, 0 problem(s) (G13) |
| V7 | python3 scripts/gen_pipe.py --check / --self-test |
rc 0 / rc 0 (G13) |
| V8 | stdio grep over MG_Backend/DirectVulkan + MG_State/.../VertexArrayState |
none (G13) |
| V9 | awk '/namespace RenderStateImpl {/,/} \/\/ namespace RenderStateImpl/' DirectGLES.cpp | sha256sum |
d8fd1c48…0efe27, equal to ~/w7/p2-before-syncrenderstate.sha (G5) |
| V10 | ctest -N names, corrected grep (§6.5), build-{linux,push,nolegacy} |
2367 / 2367 / 2367, all three identical; vs the baseline: 0 removed, 4 added and all four the contract's MG_Test/Pipe placeholders (G2 name half, G14) |
| V11 | ctest -L unit in build-linux / build-push / build-nolegacy |
1489/1489 in each |
| V12 | ctest -L integration-gpu -j 4 -R DirectVulkan, build-push, default mask |
432/432 |
| V13 | same, MOBILEGL_PIPE_PUSH=0 |
432/432 (one flake on the first run, see §7; clean on rerun) |
| V14 | same, MOBILEGL_PIPE_PUSH=0x60 MOBILEGL_PIPE_LEGACY_MEMOS=0 — the review's MAJOR 1 repro |
432/432 (was 9/9 aborted) |
| V15 | same, MOBILEGL_PIPE_LEGACY_MEMOS=0 at the default mask |
432/432 |
| V16 | same, build-nolegacy |
432/432 (was 180/432, every failure a per-draw Fatal) |
| V17 | same, build-linux (pull) |
432/432 (one flake on the first run; clean on rerun) |
| V18 | retrace_gate.py --tree ~/w7/p2-magma --lib build-push/libMobileGL.so --out ~/w7/retrace-out/p2-magma-r2f -j 4 --only 'DirectVulkan$' |
40/40 PASS, min SSIM 0.993475 (minecraft-1.21.11-main-menu), gate 0.99 (G3, DirectVulkan half) |
| V19 | grep -rn "MGPipeSlots()" MobileGL … | grep -v MG_Impl/Pipe/SlotAllocator |
one comment line, no code (MAJOR 3) |
| V20 | grep -rn "include <MG_Impl" MobileGL/MG_Backend |
only the pre-existing DirectGLES.cpp:22; this package's is gone (minor 4) |
| V21 | 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) |
| V22 | MOBILEGL_PIPE_PUSH=0x20 MOBILEGL_PIPE_LEGACY_MEMOS=0, one DirectVulkan case, MOBILEGL_LOG_FILE_PATH |
one Fatal{PipeLegacyMemosDisabled} naming bit 6 at startup; ctest reports a visible SKIP, not a mid-frame abort |
retrace_gate.py has no --ssim option on this tree (the brief's G3 line passes one); the
script's built-in threshold is what ran. Recorded as a brief-vs-tree difference, not a
deviation.
4. The m1 arming evidence (~/w7/p2-magma-armed, tmp/magma-armed)
p2/magma merged with p2/tracker — the merge is clean, which is the first datum: C.5's
file split holds between D and B.
| # | run | result |
|---|---|---|
| A1 | ctest -L integration-gpu -j 4 -R DirectVulkan, default mask |
432/432 |
| A2 | 13 CrossFrameBufferScenario cases with MOBILEGL_LOG_FILE_PATH |
13/13; occurrences of the fallback warning: 0 |
| A3 | the same 13 on ~/w7/p2-magma (no tracker) — the control |
13/13; fallback warning: 1 |
| A4 | retrace_gate.py … --only 'DirectVulkan$' on the armed tree |
40/40 PASS, min SSIM 0.993475; fallback warning across all 40 logs: 0 (on ~/w7/p2-magma the same grep hits 80 files) |
| A5 | RED PROBE: ResolveBoundRenderStateCso patched to return a constant {1,1}, rebuilt, -R 'DirectVulkan.*(Blend|Stencil|Depth|CrossFrameBuffer|Clear)' |
4 of 59 fail — LayeredCubeMapArrayDepthStencilAttachmentGatesEveryLayerFace, both RenderbufferBlendFormatScenario cases, and one more. Patch reverted, rebuilt: 59/59. |
A2+A3+A4 say the CSO handle arm ran for every draw of 432 integration cases and 40 retraces,
and that the same binary shape makes the absence of the arm observable. A5 is the
ROADMAP.md:7 half: the gate can go red for m1's reason — if the CSO key stops
discriminating render states, the memo hands back the wrong pipeline and the pixels move.
What this still does not cover: build-verify (no verify build on either tree; C.3 asks
for none), the DirectGLES half of G3, and G6–G12 / G11 / all of D.4.
5. Deviations from the brief, with reasons
Carried from v1, still true:
- D-1 —
ComputePipelineStateHash's 20-line enumeration comment did not move toMGPipeRenderStateSpans.cpp(D12.2). That file is package A's. - D-2 — the dynamic tail's version gate is re-sourced but not finer-grained. D12.3 says
GetRenderStateParametersVersion()stops moving on a pipeline-only change; it does not, becauseMGPipeApplyBindRenderStatewrites both versions and has to (Espryt'sSyncRenderStateuses that counter as its all-state detector and G5 forbids touching it). A dynamic-only version on the wire is a contract change. The reviewer independently confirmed the tree is right here. - D-3 — D12.3's coverage check is
static_asserts inVulkanRenderer.cpp, not theRenderStateSpansTest.cppentry D19 names; that file is package A's. Integrator: make sure the outcome is not "neither" (review minor 5). - D-4 —
ScissorTestEnabledMaskis pipeline state, so D12.3's "every fieldDynamicTailKeyreads is dynamic" is false as written; pinned with the assertion inverted. Independently confirmed by the reviewer. - D-6 — a VAO's
MGPipeKindisVertexElementsCso; there is noVertexArraykind. - D-8 — the three
VertexArrayObjectmemo triples are#if MOBILEGL_PIPE_LEGACY_MEMOS, not textually deleted (D12.5 says "deleted outright"): a literal deletion changes the pull build, which G1 admits no resize of.build-nolegacyis where the deletion is real. - D-9 — the eviction epoch is de-globalised, not removed.
- D-10 — negative control C (
MOBILEGL_PIPE_HANDLE_ABA_CONTROL) is implemented here although C.4 givesHandleRecycleScenarioto package E: D18 defines the knob in terms of two guards that live in this package's files.
New or changed this round:
- D-5 (rewritten) — the two per-slot tables stay fixed where D12.4 asks for a
grow-on-demand
Vector, and they are now a bijection with a fixed-capacity mint rather than a masked map. Grow-on-demand only makes sense against an allocator that frees, and nothing in P2 frees aVertexElementsCsoslot. Integrator: revisit when object deletion reaches the client allocator (P3a/P3b) — at that point the mint should becomeMGPipeSlots()again and both tables should be sized offHighWater(). - D-7 (rewritten) — Magma mints its own
{slot, gen}from a backend-local, bounded, self-recycling identity table rather than fromMG_Impl/Pipe/SlotAllocator. It was already a P2 stand-in (the tracker emits no object-class state, so no client handle reaches the backend); what changed is that the stand-in is now bounded and does not leak. The earlier claim that "both acquisition sites sit behind a one-entry memo" was wrong — there are three sites and one had no memo (review minor 2). There is now no memo at any of them and none is needed: an acquisition is a mask plus at most twoUint64compares, which is less than the address multiply and two-way probe the pre-handle arm ran. - D-11 (retired) —
MG_Backend/DirectVulkanno longer includesMG_Impl/Pipe/SlotAllocator.h. The conflict the reviewer named (ARCHITECTURE.md:9— the server process linksMG_Backend- the MGPipe object table and not
MG_State,MG_Impl, glslang) is therefore gone. The remaining push-only include from this package isMG_Pipe/PipeApply.h, which is the server half of the calls and is on the right side of that line. Notecheck_include_closure.pystill has no probe rooted at anMG_Backendheader, so G13 says nothing about this either way — worth a probe, and recorded as a suggestion, not done here (the script is not this package's file).
- the MGPipe object table and not
- D-12 (new) — D12.1 says the version→hash gate and the cached-hash fields "are deleted".
They are not: they are the no-CSO fallback's cache, and that fallback is structurally
required because
delete_render_stateclears the binding. On a tree whose tracker binds a CSO the five words are written once and never read again; they retire with the pull path at P13. This is a direct consequence of fixing MAJOR 1's second guard. - D-13 (new) — the pull build's two pipeline-memo sites keep the base ref's text
statement for statement under
#else. Routing the pull build through the new inline helper instead resizedGetOrCreatePipeline(+104) andTrySetupDrawFastPath(−112) — a real G1 red that this structure removes. Cost: ten duplicated lines, twice.
6. Tree-versus-brief contradictions
- D12.3's premise about
bind_render_state— see D-2. The brief's model of the applier does not matchPipeApply.cppas landed. (Reviewer's minor 7 agrees; the tree wins.) - D12.3's claim about
DynamicTailKey's inputs — see D-4. (Reviewer's minor 6 agrees.) - D12.1 lists three cached-hash fields; there are five (
m_pipelineStateHash{,Valid, Version,ColorCount,SampleCount}). - D12.1 does not mention the second memo probe in
TrySetupDrawFastPath. It exists and is re-keyed identically; leaving it on the hash would have been a silent arm mismatch. - G2/G14's shell command in section A is wrong.
grep -E '^\s+Test #'matches only 4-digit ids (ctest -Nright-aligns), returning 1368 of 2367 names here. The working form isgrep -E '^\s+Test\s+#[0-9]+:' | sed -E 's/^ *Test *#[0-9]+: //'. FixBRIEF-P2.mdbeforewsl_integrate_p2.shuses it. (Reviewer's minor 8 agrees.) - G1's admitted resize set is short by one.
_GLOBAL__sub_I_DirectGLES.cpp −9is a fourth resized symbol and it is the contract's — identical against~/w7/p2-contract/build-linux/libMobileGL.so. D15 admits three, so G1 is red by its own text after every merge until D15 is amended. Integrator's call, not this package's defect. (Reviewer's minor 1.) - 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 aMOBILEGL_PIPE_LEGACY_MEMOSarm, plusProgramObject's" (V21). MGLOG_F_ONCEdoes not exist (MG_Util/Debug/Log.hhas_ONCEfor D/I/W/E only).- G3's command passes
--ssim 0.99;~/w7/retrace_gate.pyhas no such option and exits 2. Its own threshold ran instead. Fix the brief or the script before D.3. MOBILEGL_PIPE_PUSH=0x20(Espryt's bit alone) withMOBILEGL_PIPE_LEGACY_MEMOS=0is still fatal to a DirectVulkan run — but now at startup, naming bit 6, because with bit 6 clear and the legacy arm forbidden Magma's vertex-input site genuinely has no arm. That is D14's own semantics, not a defect; it means a per-subsystem A/B with the legacy lever off must set the bit belonging to the backend under test.
7. Flakes
DirectVulkan.PrimGenReroute…TheRerouteIsActuallyArmedWhenTheEnvironmentPinsItOn and
…PointSizeDemotion…TheDemotionIsActuallyArmedWhenTheEnvironmentPinsItOn failed once each in
four of this round's -j 4 runs (once in build-nolegacy, once in build-push MOBILEGL_PIPE_PUSH=0, once in build-linux, twice in the armed tree's first run) and passed
every time in isolation (3× each) and on every full rerun. build-linux is byte-identical to
the baseline apart from the contract's four resizes, so they cannot be this package's.
Load-sensitive lavapipe flake. Integrator: worth a serial label or a retry-on-failure in
integration-gpu generally. (The reviewer could not reproduce it in v1; this round did, four
times, which supports "load-sensitive" rather than "does not exist".)
8. Unfinished, and follow-ups for the integrator
build-verifywas never run for this package (C.3 asks for none, and neither tree has one). D.3 part 2 remains the authority for G4.- The armed tree is verification scaffolding.
tmp/magma-armed/~/w7/p2-magma-armedcan be deleted after the real integration; keep it only if the D.3 evidence is wanted before B and D are both merged. - D-5 / D-7: revisit the mint when object deletion reaches the client allocator. The
right end state is
MGPipeSlots()with a liveFree, both tables sized offHighWater(), andMagmaPipeIdentityTabledeleted. Nothing else in Magma changes. - D-12: once the tracker is landed and the fallback is provably unreachable in a shipped
configuration, the five cached-hash fields can go — but not before, and the
MGLOG_W_ONCEis the thing that tells you. - D-2: decide whether to add a dynamic-only version to the wire.
- D-1: the enumeration comment's move into
MGPipeRenderStateSpans.cpp. - D-3 / D19:
DynamicChunksCoverMagmasDynamicTailKey—static_asserts here, no ctest entry; make sure package A's file does not also drop it. - Fix
BRIEF-P2.md's G2/G14 grep (§6.5) and G3's--ssimargument (§6.9). - Amend D15's admitted resize set with
_GLOBAL__sub_I_DirectGLES.cpp(§6.6). - Consider a
check_include_closure.pyprobe rooted at anMG_Backendheader, soARCHITECTURE.md:9is actually gated (§5, D-11). MEASUREMENTS.mdinput from this package: the pipeline memo compares an 8-byte handle instead of an 8-byte hash and skipsComputePipelineStateHashentirely (~50CombinePipelineStateWordrounds plus one bulk parameter fetch) on every draw whose pipeline-state version moved;LookupVaoDrawMemodrops a multiply, a two-way probe and one of two compares;VertexInputStateFactorydrops onemutablewrite per VAO reconfiguration. New fixed cost: 160 KB of identity tables in a push build. None of it is measured here — that is D.4.- Not measured, and it cannot be measured on desktop: the identity tables' eviction behaviour above 2048 live VAOs / 8192 live buffers (§2, MAJOR 4). If the device A/B shows an unexplained per-draw cost in a chunk-cycling world, that is the first thing to size up.