docs: record the 2026-10-12 per-frame optimization pass

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
shenlei
2026-10-12 16:46:17 +09:00
co-authored by Claude Opus 5.5
parent 121f1a6307
commit b5a15871f4
+45
View File
@@ -999,3 +999,48 @@ groups, independent reduced branch/leaf/frond LOD meshes, exact frond smoothing,
billboard dimensions/component fades, newer SPT extension semantics, gameplay
collision integration and GPU wind rendering. No DLL runtime or rendered visual
parity comparison was performed. Original AVA2 data and server remain unchanged.
### 2026-10-12 Per-frame CPU/GPU optimization pass
Implemented the prioritized per-frame findings, one commit per item (client
1eccdc87..121f1a63, shared sdk a1b3938..53115d9):
- Render3D mesh cache sweeps only on growth; one shared GPU skinning palette per
instance (pushed once per change); cached part->parent bone maps.
- Text tails/labels: no per-frame codepoint vectors, show-state flag instead of
list search, memoized widths, reused painter text instance.
- Off-screen actor shadows culled; draw lists reserve last frame's peak.
- Full mip chains for GPU scene textures (visible change: distant terrain/props
no longer shimmer; sampler max_lod was zero before).
- GPU tri-groups share one resident mesh and draw index sub-ranges.
- Skins resolved once per skin change on the GPU draw cache; RenderSceneGPU
resolves each texture once and sorts opaque draws by texture handle.
- HardwareVertex packed 88 -> 48 bytes (UBYTE4_NORM colour, SHORT4_NORM normal,
UBYTE4 bones, USHORT4_NORM weights); shaders unchanged.
- Texture/mesh uploads staged and copied in the frame command buffer (one
transfer buffer and copy pass per frame).
- Every draw reads its transform from the instance storage buffer
(firstInstance = batch index); the 288-byte uniform is pushed once per frame.
Texture eviction sorts candidates once.
- Bone lookup by a cached name index; attachments cache their bone index.
- Python member-function names interned for PyCallClassMemberFunc (cache dropped
by Py_AtExit across Py_Finalize/Py_Initialize).
- Preview minimap radar redrawn only when its inputs change.
Verified: Debug macOS/Metal build, 50/50 CTests (including new regressions for
resident index sub-ranges and GPU-cache skin changes). The current origin/main
SpeedTree forest test needs SDK APIs not present in any pushed sdk revision, so
verification used a local SDK shim; speedtree_forest_test was not built.
Not done, by decision: removing the offscreen colour target + blit (the 3D layer
must composite premultiplied over SDL_Renderer 2D layers); baking text outlines
(SDL already batches the five quads; colour would enter the cache key); partial
palette pushes (risky on Vulkan/SPIR-V); sharing streams on the CPU-skinning
tri-group path (would churn resident buffers); a per-frame light/fog uniform
(light is per draw; fog is 32 bytes and nothing is pushed per draw any more).
Outstanding: no Vulkan/Android device run or frame-time capture for this pass;
`frames*0.033f` / `frames/60` animation clocks are frame-based rather than real
time; translucent depth sorting uses each draw's first vertex. Original AVA2
data and the server are unchanged.