Files
MobileGL/docs/Disaggregated
swung0x48 96fa569250 [MG_Backend, MG_Pipe, MG_Remote, MG_IntegrationTest, docs] (P7): the XFB writeback event is sliced, so a large capture stops killing the server
P3b/P4b espryt D1 slice 2, landing under the P7 program (notes/p7/PLAN-PH-P34B-P7.md
§1.2 row "XFB scatter", §3 wave 2-D).

Both XFB writeback producers posted ONE unbounded OnBufferWriteback covering the whole
bound range - ScatterCapturedRecords from its staged vector, ReadbackCapturedRanges from a
live glMapBufferRange mapping. ServerOnBufferWriteback turns each into one SEG_EVENT record
of sizeof(EventBufferWritebackHead) + bytes, and RingProducer::Reserve refuses any record
above Capacity()/2 outright - not "when full" but always, because above half a capacity a
record is placeable at some head offsets and not at others and waiting for room would be a
hang. EventRingBytes defaults to 256 KiB, so any capture target wider than ~128 KiB was
Fatal{EventRingOverflow} and the SERVER PROCESS DIED, on a buffer size, with nothing else
wrong. The producers' own Fatal text already said what to do: raise MOBILEGL_IPC_EVENT_KB
or slice the event at its producer, "as the writeback path already does" - which was true
of the client's request path and of nothing on this side.

MG_Pipe now carries the one number, published the way gMGPipeSegmentResolver is published:
gMGPipeEventRingCapacityBytes, installed by ServerSession::Accept beside the four reverse
producers and released at Close only if it is still ours, and
MGPipeBufferWritebackSliceBytes() which turns it into a per-record width - a quarter of the
ring with a 4 KiB floor, the same number and the same reasoning as the client-linkage
BufferWritebackSliceBytes one ring over. The backend can reach neither MG_Remote::Client
(wrong linkage) nor ServerSession (wrong layer), so the hook is the seam; 0 means "do not
slice", which is monolith, where the callback is a direct call with no ring under it.

Both producers loop against ONE helper, XfbWritebackSliceStep, because a second spelling of
the width is how one of them ends up still killing the server. The readback site slices
inside its mapping and still unmaps once; the scatter site slices the staged vector and
keeps the single glBufferSubData whole, since the ES store has no ring and cutting it would
change how many backend calls one application call makes. In-order delivery makes the last
slice's landing imply every earlier one, so the client needs nothing new.

Red-once (R-16): XfbAfterClipDistanceScenario.{AScatteredCapture,APlainCapture}
IntoAMegabyteRangeReachesTheApplication - a 1 MiB bound range, about eight times what one
record can ever be, with the same six capture records; glBindBufferBase over the whole
store is what makes the mirrored range a megabyte. Two cases for the two producers. With
the slicing reverted all six split entries go red on Fatal{EventRingOverflow}
("kEventBufferWriteback needs 1048600 bytes and SEG_EVENT caps ONE record at 131072") -
aborting the test process under inproc, losing the peer under spawn and tcp - while both
monolith controls stay green.

XfbAfterClipDistanceScenario and XfbRepeatedCaptureScenario now join the three split arms
in full; slice 1 could register only OrphanedCaptureTarget* because the rest could not
survive the unsliced event. XfbRepeatedCaptureScenario's gl_NextBuffer compacted-target
shape is the first thing outside the census to exercise the archive-served stride and
varyings (ProgramArchive->Link's xfb fields) on the server, and it passed on all three arms
with no further change.

Gates: integration-split 202/202, integration-spawn 118/118, integration-tcp 121/121,
ctest -L unit 2408/2408, the monolith and split XFB families 120/120, spawn_lane_parity /
fatal_census (79 sites, 0 unmarked) / link_ratchet (186, unchanged) green, G1 .text
0xa52203 with both nm lists byte-identical to ~/w7/p7-before.
2026-09-22 16:41:54 -04:00
..

MGPipe:MobileGL 前后端拆分

状态:P0–P5e 已收官(P5e 2026-09-19 收官,附一条具名未决:E1 对照需按 ID-114 重新定义),当前在 P5f(一切状态上 wire,计划 P5F-WIRE-COMPLETENESS.md;P6 阻塞于它),P6 计划 P6-SPAWN-PLAN.md、契约草稿 P6-CONTRACT-DRAFT.md,均尚未开工)。P5e:run-ahead 已武装(十二包落地,strict 车道硬绿 179/179、integration-gpu 1357/1357、pull/push/split 三个构建 flavour 全绿;设备上渲染距离 32 下 inproc 已与 monolith 齐平,相对它取代的 lockstep p50 +69%。报告 P5E-RUNAHEAD.md)(P5c 2026-09-17 b88e8487;P5d 三轮 2026-09-18,代码头 1f8de61b——inproc 性能专项,VD12 Minecraft 上 split 从 7-13 fps 拉到 103-106 fps、client 线程 CPU/帧从 15 ms 到 9.2 ms,报告 P5D-INPROC-PERFORMANCE.md)。目标负载(四条 Minecraft A/B trace)已在 Redmi Adreno 830 上以 inproc(独立 apply 线程)双后端渲染,barrier tax 首次实测;P5c 把审计出的 59 处不经 wire 的直接内存访问全部归零(契约 MobileGL/MG_Remote/CONTRACT-P5C.md),inproc 成为只经 wire 交换的诚实两角色。P6 只是传输替换——但这句话的证据是 P5c 时代的,P6 的第一个包是重新证明它的只读审计。当前头、逐门数字与开放项见 CURRENT_STAGE_PROGRESS.md。

性能纪律(2026-09-08 起):逐线程 CPU 与 tracker 绝对 ns 对着 pull 臂基线记录,不作阻塞门;专门的优化阶段排在路线图推完之后。

是什么

MGPipe 是 MobileGL 前端(MG_State + MG_Impl)与后端(MG_Backend:Espryt = DirectGLES、Magma = DirectVulkan)之间的一份显式接口:gallium 形状、句柄寻址、只推不拉。它取代后端每 draw 直接读 MG_State::pGLContext 的做法,让后端拥有自己的状态机,并在此之上把前后端拆到两个线程(inproc,已达成)与两个进程(spawn,P6)。

接口本身是可独立交付的产物:即使 IPC 永不上线,inproc 就是 monolith 的渲染线程。

架构(一段)

应用 GL 调用
  → MG_Impl(GL 语义、错误、shadow)
  → MG_Impl/Pipe/Tracker:在每条 verb 之前 validate,把变化推成 MGPipe 调用
  → MGPipeScreen / MGPipeContext(两张函数指针表,79 条调用,单一真相源 PipeCalls.def)
      monolith:直调 backend 函数          split:发射器写 SEG_CMD ring → server applier(apply 线程)
  → server 对象表(按 {slot, gen} 句柄索引的数组)+ PipeInputs(后端被推送的状态块)
  → MG_Backend(Espryt / Magma),两个后端的 ring / pool / memo / lowering pass 原样不动
  ← MGPipeCallbacks(10 个具名反向回调 + 1 个正向终止符)

三种形态共用同一份 backend 实现:monolith(默认,进程内直调)、inproc(同进程两个线程,CI 形态与渲染线程交付物;P5 起 lockstep verb barrier,P5b 起 71 个后端槽里 54 个发射、15 个具名拒绝)、spawn(P6:fork+execve 出 server 进程,SPSC 共享内存 ring + FlatBuffers 控制面)。

构建与运行

四个 flavour:build-linux(pull,默认)、build-push(-DMOBILEGL_PIPE_PUSH=ON)、build-verify(+MOBILEGL_PIPE_VERIFY 影子比对,MOBILEGL_ITEST_REQUIRE_GPU=1)、build-split(+MOBILEGL_BUILD_DISAGGREGATED +MOBILEGL_BUILD_DISAGGREGATED_INPROC)。split 运行时必须显式设置传输:

MOBILEGL_TRANSPORT=inproc MOBILEGL_ITEST_REQUIRE_GPU=1 ctest --test-dir build-split -L integration-split --output-on-failure
MOBILEGL_TRANSPORT=inproc ctest --test-dir build-split -L integration-gpu --output-on-failure   # 普查车道,只记录

integration-split(DirectGLES.Split.* / DirectVulkan.Split.*)是硬门;integration-gpu 在 inproc 下是带已知 abort / wrong-answer 的普查车道。每条 split entry 带独立 MOBILEGL_LOG_FILE_PATH,阴性控制(scripts/ci/split_negative_controls.sh)只读被选 entry 的私有日志。常用旋钮(全表见 ARCHITECTURE.md 附 A):

变量 默认 用途
MOBILEGL_IPC_STAGE_MB 32 SEG_STAGE 容量(一条记录一个整 blob);内容分块预算 = 本值的 1/4(MGPipeStageChunkBytes(),默认 8 MiB),超预算的 buffer 范围 / 纹理级切成多条记录,故历史 profile 的显式 256 不再是这两条路径的必需
MOBILEGL_IPC_VERB_BARRIER 1 每个 verb 等 appliedSeq == emitSeq;0 只作 E1 阴性控制
MOBILEGL_IPC_AUDIT 0 apply 返回后以 0xDD 填退休 staging,查跨返回持针
MOBILEGL_IPC_STRICT_ERRORS 0 BARRIER-PULLED residual input 提升为具名 Fatal
MOBILEGL_IPC_ADOPT_TIER 2 split 使用 emulated persistent-map 路径;auto/0/1/2
MOBILEGL_IPC_PERSISTENT_BLOCK_KB 64 persistent-map 块推送粒度;0 是 E3(a) 阴性控制

Android 三份 APK flavour:pull、push、split(Gradle 属性 mobilegl.pipePush、mobilegl.buildDisaggregated、mobilegl.buildDisaggregatedInproc;~/w7/notes/tools/wsl_build_p5_apks.sh)。split APK 需运行环境 MOBILEGL_TRANSPORT=inproc,不设置时是 split build 的 monolith control arm(splitctl)。设备 A/B 只用 Redmi 2f7cbe2e,定频协议见 devices/pin-verification-2026-09-07.md。

文件地图

文件 内容
CURRENT_STAGE_PROGRESS.md 当前头实测、P5b 落地内容、开放项、下一步、证据位置、仍在生效的裁定;随每次落地更新
ARCHITECTURE.md 已定稿的设计:句柄与世代、调用目录与生成器、记录约定、tracker、纹理路径、shader 制品、反向通道、后端改造、传输、persistent map 分档、进程 / EGL / 平台、构建与门、P5 / P5b 落地形状、开关表
ROADMAP.md 纪律、两条跑道、P0…P13 阶段表、里程碑、债务表、开放问题
P5F-WIRE-COMPLETENESS.md P5f 计划:gPipeInputs 是什么、拆进程后它怎样;阻碍跨进程的完整清单(BARRIER_PULLED 15 字段、Magma、EGL 控制面、静态量、豁免 scope、反向通道);双块演练机制与出口门
P6-SPAWN-PLAN.md P6 的包计划:树上已有什么(传输原语基本已就绪且有测试)、真正要造什么、P6 与 P12 的边界、七个包与依赖序、出口门与阴性对照
P6-CONTRACT-DRAFT.md P6 契约草稿(规则 G、令牌、控制面帧、死亡与 run-ahead、阶段修正);c6 落地时移为 MobileGL/MG_Remote/CONTRACT-P6.md
MEASUREMENTS.md 逐阶段实测:P0 spike 与基线、P1 verify、P2 门与 DriverBench、P3a / P4a 门与设备 A/B、P5 全门 / E1–E6 / Redmi 四臂、P5b 主机门 / 普查 / 审查 / Redmi 出口
devices/pin-verification-2026-09-07.md 设备定频档案与核验(Redmi 当前口径;小米 / Oppo 历史核验)

代码地图:

路径 作用
MobileGL/MG_Pipe/ PipeCalls.def(目录)、PipeFields.def、Coverage.def、FieldOwnership.def、FillPoints.def、DirtySurface.def;MGPipeTypes.h(payload POD)、MGPipeValueTypes.h、MGPipeHandles.h、MGPipeHostSpan.h、MGPipeCallbacks.h、MGPipeRenderStateSpans、PipeApply、PipeRoute;generated/*.inc(G1–G8 产物,提交进树)
MobileGL/MG_Impl/Pipe/ Tracker.h、PipeFill、SlotAllocator、CsoCache、CompositeResolver、ResourceTracker、SetHashSuppressor、各族 *Emit.h
MobileGL/MG_Backend/MGPipe/ PipeInputs.{h,cpp}(后端被推送的状态块,G5 / G8 产物的 include 点)
MobileGL/MG_Remote/ 仅 MOBILEGL_BUILD_DISAGGREGATED=ON:Protocol/(protocol.fbs)、Transport/(Ring、Doorbell、ShmSegment、FdPassing、Framing、InProcessTransport、ReplySlot、EventRing)、Wire/PipeWireCodec(记录编解码、WireVerbSink)、Client/(BackendObject_Remote、ClientSession、EmitTables、CapsMirror、PersistentMapTracker)、Server/(PipeApplier + ServerVerbSink、ServerLoop、ServerSession、StagedShadow);CONTRACT-P5.md、CONTRACT-P5B.md
scripts/ gen_pipe.py(G1–G7)、gen_pipe_field_ownership.py(G8)、gen_pipe_dirty_surface.py、gen_protocol.py、symbol_report.py(G1)、check_include_closure.py、check_doc_citations.py(本目录 file:line lint)、p3a_/p4a_untouched_regions.sh(G5)、三个阴性对照脚本、ci/(split 阴性控制、普查 JUnit、smoke)
MobileGL/MG_Util/Metrics/PipeStats.{h,cpp} 边界计数器,MOBILEGL_PIPE_STATS=1 开启
MobileGL/Config.h、MobileGL/ConfigLoader.cpp MOBILEGL_PIPE_*、MOBILEGL_TRANSPORT、MOBILEGL_IPC_*
MobileGL/MG_Test/Pipe/、MG_Test/Wire/、MG_IntegrationTest/Harness/Pipe*Peek 目录 / tracker / emitter 单元、wire 层与 remote client/server 单元、白盒断言
tools/spikes/、android-plugin/app/src/trace/cpp/spawn_spike.cpp spike A(Android exec 第二个原生可执行文件)、spike B(跨进程外部内存分档)

术语

  • client / server:前端进程 / 后端进程;monolith 下是同一进程的两个角色,inproc 下是两个线程。
  • verb:会让 server 做事的命令(draw、dispatch、clear、blit、readback、XFB 跨度、query、纹理操作)。推送只发生在 verb 之前的 validate 时刻。
  • class A / B / C:后端槽的三类——A 从 caps mirror 本地回答,B 发射记录,C 具名拒绝(Fatal{UnmigratedVerb}),永不回落到 monolith applier。
  • CSO:常量状态对象(render state、vertex elements、sampler、sampler view、shader),client 侧内容寻址,server 侧按句柄缓存。
  • Track V / Track H:值类读点的迁移(整块 POD 过线)/ 对象类读点的迁移(SharedPtr<前端对象> → 句柄)。
  • MGGen:server 私有的"我重铸了驱动对象"纪元,永不过线;与句柄里的 client 世代严格分开。
  • G1 / G2 / G5 / G14:pull 符号恒等 / pull==push 测试名 / 保护区字节一致 / 测试名只增不删。barrier tax:split − push 的逐线程 CPU 增量(inproc 传输 + verb barrier 的代价)。

历史

本目录此前是一份 328 KB 的实施计划(PLAN.md)加 135 KB 的设计竞赛与评审记录(REVIEW.md);设计定稿后只保留设计与架构本身,评审记录与早期草案留在 git 历史(8b31de2f、1794ac94、8349babe、87ee17c6)。P5b 收官后(ea34bccb 及之前)各文档的完整历史叙述、-O0 设备表与逐轮审查记录同样以 git 历史为准;集成者的逐条裁定长文(ID-1..75)在 ef35ea0c 之前版本的 CURRENT_STAGE_PROGRESS.md。