mirror of https://github.com/MobileGL-Dev/MobileGL synced 2026-11-07 10:03:23 +09:00

Files

T

swung0x48 fe748773f3 Docs (P15): hand-off for the next agents (LTO GL CPU Magma 2.52 -> 2.36, Espryt 2.99 -> 2.90; levers table; WIP branches; harness rules).

2026-10-09 12:54:27 -04:00

20 KiB

Raw Blame History

Monolith render-thread diff: dev vs p15 (FCL MC 1.21.5 vanilla, H2, cpuhunt)

The fixed function-level baseline for the monolith CPU cuts (SCOREBOARD.md has the fps reference). Every cut starts from a symbolized profile of the p15 side and ends with a before/after of the same function.

Sources:

Profiles: simpleperf cpu-clock --trace-offcpu --call-graph fp -f 4000 --duration 6, taken after the measurement windows of an H2 run (FCL pinned to cpu2-6).
dev: 08124c99 in runs/safter/dev-Direct{Vulkan,GLES}-r1.
p15: p15t = HEAD 3def5cb2 (p15s 0b96a3ef plus the Magma blit view cache) in runs/pt/t-Direct{Vulkan,GLES}-r1.
Scope and units: render thread only (the GL thread, which in monolith also applies), off-CPU samples dropped, in ms/frame. Frames are the run's measured fps × 6 s; profiling overhead inflates both sides alike.
Reducer: layerdiff.py in the scratchpad.

Attribution: each sample goes to its innermost libMobileGL function, then to one layer.

The leaf DSO decides first: driver (Adreno), libc+TLS (libc, libc++, __emutls_get_address, TLS init), or kernel.
Otherwise the function's namespace decides:
- client frontend: MG_Impl, MG_State;
- pipe record: MG_Pipe validate, tracker and emitters, plus the record channel;
- applier: MG_Record sinks and record emitters, plus MG_Pipe applier state;
- backend: MG_Backend.
Samples with no libMobileGL frame are app-side (JIT/JVM).

Caveat on function rows: some functions were renamed between dev and p15, so a "+" row and a "-" row can be the same work. Read such pairs together:

Espryt FlushPendingRangesNow -> FlushPendingRangesFrom (0.105 -> 0.116);
BackendVertexArrayObject::SyncToBackend -> SyncToBackendFromApplier (0.062 -> 0.058);
BackendFramebufferObject::SyncToBackend -> SyncToBackendByHandle;
Magma SetupDraw -> SetupWireDraw.

Where the excess lives (summary)

layer	Magma diff	Espryt diff	what it is
pipe record	+0.36	+0.35	`MGPipeValidateForVerb` (0.34 inclusive on Magma). Its pieces: framebuffer-state build 0.06, vertex-buffer emit 0.04, global constants 0.03, slot lookups (`FindByLifetimeId` + `Acquire`) 0.05, residual `CopyField` 0.02, tracker 0.02
driver	+0.23	+0.06	Magma: per-frame fixed Vulkan work (below). Espryt: about even once renames are paired
libc+TLS	+0.22	+0.14	`__emutls_get_address` 0.055-0.075 plus `pGLContext` TLS init 0.016 (dev: ~0). Magma: libc under the extra command-buffer and pass begins. Espryt: `QueryCurrentSurfaceSize`
client frontend	+0.12	+0.17	Spread thin. Atomics (`__aarch64_ldadd8_*`) +0.02, state lookups 0.005 each
backend	-0.10	+0.11	Magma's wire backend is cheaper than dev's `SetupDraw` (no XXH64 vertex-input hash). Espryt pays per-handle registry lookups
JIT/JVM (app)	+0.04	+0.08	Not ours; profiling noise between runs

Magma per-frame fixed costs (inclusive ms/frame, dev -> p15):

BeginRenderPass 0.045 -> 0.117.
EndCommandRecording 0.054 -> 0.113.
BeginCommandRecording 0.052 -> 0.089.
EndRenderPass 0.023 -> 0.033.
Present 0.303 -> 0.352.
p15 also has SyncTextureResourceByHandle 0.068, of which UploadPendingWireLevels is 0.038 and FlushPendingCommands 0.025.

Cause: every wire texture upload first submits the frame's recorded work (FlushWirePendingCommandsForTextureUpload) and then uploads in a separate batch. MC updates textures every frame, so the frame pays an extra submit, an extra command-buffer end/begin and a render-pass restart. GpuProgressMarkers::SubmitBracketed is NOT new work: it wraps Present's queue submit, which dev pays inside Present (0.121 vs 0.115).

Espryt-only: QueryCurrentSurfaceSize costs 0.065 ms/frame and dev pays nothing for it. It is two eglQuerySurface calls on every dirty render-state sync (draws 0.048, clears 0.022). It substitutes the surface for a scissor box the app never wrote (and for an empty viewport).

Correction after cut 1 (2026-10-09): what the Magma fixed costs really are

Cut 1 (wire texture uploads stop flushing the frame recording) engaged on device: FlushPendingCommands and SubmitPendingCommandBuffer left the profile, and UploadPendingWireLevels went from 0.038 to 0.016. But Begin/EndCommandRecording and BeginRenderPass did not move. H2 c1: 254.1 -> 249.8 fps, inside the noise.

Their cost is in the driver and scales with what is recorded:

driver call	dev	p15
vkResetCommandBuffer	0.044	0.077
vkEndCommandBuffer	0.053	0.111
vkCmdBeginRenderPass	0.037	0.105

Host call counts per frame (diagnostic counting builds; FCL vanilla; 600-frame windows):

Magma, Vulkan calls: dev 331.6 vs p15 379.
- The draw-level stream is identical and almost free of repeats: DrawIndexed 86.1, BindVertexBuffers 87.1, BindDescriptorSets 87.1, BindPipeline 15.1, UpdateDescriptorSets 14, dynamic state 2 each.
- Render passes: 4.1 -> 12.6.
- Pipeline barriers: 3.4 -> 27.5.
- CopyBufferToImage: 0.1 -> 2.2.
- By call site (unwinding counter):
  - Pass begins: 6.5 at draws, 5.2 at clears, 1 at the blit.
  - Pass ends: 5.3 from ResolveWireImageDescriptor (sampling an image written earlier), 4.2 from TransitionWireImage in the in-pass clear, 1.2 from TransitionWireImage in SetupWireDraw, 2 at the blit.
  - Barriers: 11.8 from TransitionWireImage at draw-pass begin, 6.3 at clear-pass begin, 5.6 from ResolveWireImageDescriptor, 2 at the blit, 1 at present.
- Cause: each new wire draw pass runs TransitionWireImage(GENERAL -> GENERAL) per attachment. That ends the open pass and records a full ALL_COMMANDS memory barrier whether or not a hazard exists.
Espryt, host GLES calls: dev 850.7 (479 repeat their previous arguments) vs p15 835.6 (446).
- p15 does not issue more GLES calls than dev.
- In both, most of the repeats are glVertexAttribPointer 302 + glVertexAttribIPointer 75: the same layout re-pointed at a new VBO on every draw.

Levers beyond the local cuts (evaluated 2026-10-09)

Estimates are ms/frame of FCL GL-thread CPU. "Measured" rows come from H2 A/Bs (SCOREBOARD).

lever	result	status
Pipe batch (FB memo, handle hints, buffer reuse, gate cache, Magma hash TLS)	measured: Magma -0.06, Espryt -0.03	landed `485dff14`, `c85c9439`
ThinLTO + -O3 + ICF/gc (gradle default ON)	measured: Magma -0.16 (2.52 -> 2.36), Espryt -0.09	190580f2 (coordinator pushes)
Magma rename spares 4 -> 12 + fence-poll quiet window	profile ceiling ~0.05-0.06 (Create 0.031, Destroy 0.014, vkGetFenceStatus 0.020)	a0015bad; m2ab A/B
API 29 separate artifact (native ELF TLS)	estimate 0.08-0.1 (emutls 0.054 + pGLContext TLS init 0.024 on Magma)	wip/api29-switch, A/B open
AutoFDO/PGO	estimate 0.1-0.25	with the user
Emulated-TLS work below 29 (mirror, owner-thread path, dual lib)	estimate 0.024-0.08	dropped by the user
App-side JIT delta	+0.065 vs dev (0.708 -> 0.773 app-only), cause unknown	needs a GL entry-count A/B
shared_ptr refcount avoidance in bind paths	0.005-0.01	not started
Espryt VertexAttribPointer split	<=0.04	parked

Cuts, biggest first

Revised order:

1b (Magma): explicit external subpass dependencies on the wire draw passes. Barrier only images last written by transfer or storage, once and coalesced before the pass. Estimate: passes 12.6 -> ~5, barriers 27 -> ~5, and the driver's command-buffer and pass-begin cost with them.
Cut 1 stays local (no measurable gain).

Order from 2026-10-09:

1c deferred clears (PARITY-AUDIT M1).
The remaining Magma barriers.
The pipe-record pieces.
The TLS hoist.

Measured and parked: Espryt glVertexAttribPointer -> glBindVertexBuffer + glVertexAttribFormat.

Ceiling: SyncToBackendFromApplier costs 0.060 ms/frame inclusive, driver calls included (0.015 self). Source: the c4 p15w Espryt profile, cpuhunt.
The repeats are MC's per-draw re-point: the same layout with a new buffer at offset 0, about 4 attributes per buffer. The split would turn 1 glBindBuffer + ~4 pointer calls into 1 glBindVertexBuffer.
Expected saving: ~0.02-0.04 ms/frame. That is below H2 resolution (1.2-2.9% spread).
Cost: a second VAO shadow (format + binding). ES glVertexAttribPointer and glVertexAttribDivisor silently rebind attribute -> binding, so the change is correctness-sensitive.

The original list follows.

Estimates are ms/frame on the H2 render thread. Each cut gets a before/after profile of its function and an H2 A/B.

Magma: record wire texture uploads into the frame's command buffer instead of flush + separate batch. Estimate 0.10-0.15. It removes the extra submit, the command-buffer end/begin and the pass restart. Correctness-sensitive: draw / TexSubImage / draw ordering, and staging lifetime per frame. Red-once with the existing between-draws upload scenarios, plus syncval.
Shared: the pipe-record layer, about 0.35 on both backends. Parts:
- (a) framebuffer-state build per verb: memo on the framebuffer's version;
- (b) slot lookups by lifetime id;
- (c) residual CopyField on monolith, where fillOwed is always true;
- (d) vertex-buffer and global-constant emit. Each part is 0.02-0.06. Profile-driven, one at a time.
Shared: TLS. __emutls_get_address plus TLS init is ~0.07-0.09. The structural fix is native TLS (minSdk, the user's call). The in-code mitigation is to read pGLContext / gPipeInputs once per verb and pass the pointer down.
Espryt: memoize the surface size per draw surface, invalidated at swap and make-current. Estimate 0.06.
Magma: the clear path (glClear 0.285 vs dev ~0.13): first re-profile after cut 1, since much of it is the command-buffer and pass begins cut 1 removes.

Magma (DirectVulkan): +0.88 ms/frame on the render thread (schedstat: 2.84 vs 2.00)

layer	dev	p15	diff
client frontend	0.257	0.376	+0.119
pipe record	0.000	0.363	+0.363
applier	0.000	0.026	+0.026
backend	0.516	0.416	-0.100
driver	0.394	0.622	+0.228
libc+TLS	0.213	0.430	+0.218
kernel (from MobileGL)	0.112	0.115	+0.003
mgl other	0.066	0.049	-0.017
JIT/JVM (app)	0.659	0.697	+0.038
kernel (app)	0.049	0.047	-0.001
total	2.265	3.142	+0.877

Top 15 p15 excess rows:

layer	function (attributed)	dev	p15	diff
libc+TLS	`__emutls_get_address`	0.000	0.075	+0.075
driver	`VkRenderPassManager::BeginRenderPass`	0.029	0.081	+0.052
driver	`FrameContext::EndCommandRecording`	0.045	0.090	+0.045
kernel (from MobileGL)	`GpuProgressMarkers::SubmitBracketed`	0.000	0.045	+0.045
driver	`VulkanRenderer::SetupWireDraw`	0.000	0.040	+0.040
pipe record	`MG_Pipe::MGPipeValidateForVerb`	0.000	0.035	+0.035
backend	`VulkanRenderer::SetupWireDraw`	0.000	0.034	+0.034
driver	`VulkanRenderer::DrawElements`	0.101	0.129	+0.028
pipe record	`MG_Pipe::MGPipeSlotAllocator::FindByLifetimeId`	0.000	0.026	+0.026
driver	`GpuProgressMarkers::SubmitBracketed`	0.000	0.024	+0.024
driver	`UniformManager::ResolveWireImageDescriptor`	0.000	0.023	+0.023
pipe record	`MG_Pipe::MGPipeTracker::Update`	0.000	0.022	+0.022
libc+TLS	`VkRenderPassManager::BeginRenderPass`	0.010	0.032	+0.022
libc+TLS	`FrameContext::BeginCommandRecording`	0.018	0.039	+0.021

Per layer (top 6 each):

layer	function (attributed)	dev	p15	diff
client frontend	`__aarch64_ldadd8_acq_rel`	0.013	0.023	+0.010
client frontend	`MG_State::GLState::FramebufferAttachmentObject::IsEmpty`	0.000	0.008	+0.007
client frontend	`MG_State::GLState::BufferState::GetBindingSlot`	0.000	0.006	+0.006
client frontend	`__aarch64_ldadd8_relax`	0.004	0.010	+0.006
client frontend	`MG_State::GLState::ShareGroupState::GetBufferObject`	0.000	0.005	+0.005
client frontend	`MG_State::GLState::FramebufferObject::CheckCompleteness`	0.000	0.004	+0.004
pipe record	`MG_Pipe::MGPipeValidateForVerb`	0.000	0.035	+0.035
pipe record	`MG_Pipe::MGPipeSlotAllocator::FindByLifetimeId`	0.000	0.026	+0.026
pipe record	`MG_Pipe::MGPipeTracker::Update`	0.000	0.022	+0.022
pipe record	`MG_Pipe::MGPipeFillAccess::CopyField`	0.000	0.021	+0.021
pipe record	`MG_Pipe::MGPipeVertexInputEmitter::EmitVertexBuffers`	0.000	0.018	+0.018
pipe record	`MG_Pipe::MGPipeProgramEmitter::EmitGlobalConstants`	0.000	0.011	+0.011
applier	`MG_Record::RecordVerbSink::OnDrawVbo`	0.000	0.005	+0.005
applier	`MG_Record::{anon}::EmitDrawRecord`	0.000	0.004	+0.004
applier	`MG_Record::{anon}::EmitClear`	0.000	0.003	+0.003
applier	`MG_Record::{anon}::EmitIndexedDraw`	0.000	0.003	+0.003
applier	`MG_Record::{anon}::EmitDrawElements`	0.000	0.003	+0.003
applier	`MG_Pipe::MGPipeApplierState::DrawFramebuffer`	0.000	0.003	+0.003
backend	`VulkanRenderer::SetupWireDraw`	0.000	0.034	+0.034
backend	`VulkanRenderer::GetOrCreatePipelineWithInput`	0.000	0.021	+0.021
backend	`UniformManager::PrepareWireTextureResources`	0.000	0.018	+0.018
backend	`_t const&, std::__ndk1::tuple<unsigned long const&>, std::__ndk1::tuple<> >`	0.000	0.016	+0.016
backend	`UniformManager::ResolveWireImageDescriptor`	0.000	0.014	+0.014
backend	`VkTextureManager::ResolveWireTextureStorage`	0.000	0.012	+0.012
driver	`VkRenderPassManager::BeginRenderPass`	0.029	0.081	+0.052
driver	`FrameContext::EndCommandRecording`	0.045	0.090	+0.045
driver	`VulkanRenderer::SetupWireDraw`	0.000	0.040	+0.040
driver	`VulkanRenderer::DrawElements`	0.101	0.129	+0.028
driver	`GpuProgressMarkers::SubmitBracketed`	0.000	0.024	+0.024
driver	`UniformManager::ResolveWireImageDescriptor`	0.000	0.023	+0.023
libc+TLS	`__emutls_get_address`	0.000	0.075	+0.075
libc+TLS	`VkRenderPassManager::BeginRenderPass`	0.010	0.032	+0.022
libc+TLS	`FrameContext::BeginCommandRecording`	0.018	0.039	+0.021
libc+TLS	`thread-local initialization routine for MG_State::pGLContext`	0.000	0.017	+0.017
libc+TLS	`EndActiveRenderPassOn`	0.000	0.014	+0.014
libc+TLS	`FrameContext::EndCommandRecording`	0.009	0.023	+0.014

Espryt (DirectGLES): +0.93 ms/frame on the render thread (schedstat: 3.13 vs 2.34)

layer	dev	p15	diff
client frontend	0.238	0.404	+0.166
pipe record	0.000	0.349	+0.349
applier	0.000	0.041	+0.041
backend	0.245	0.356	+0.111
driver	0.800	0.861	+0.062
libc+TLS	0.209	0.350	+0.141
kernel (from MobileGL)	0.209	0.221	+0.011
mgl other	0.134	0.113	-0.021
JIT/JVM (app)	0.693	0.774	+0.081
kernel (app)	0.055	0.050	-0.005
total	2.584	3.518	+0.935