Helper attachments made outside the part slots were attached to
PART_MAIN only and dropped whenever the parts were rebound (LOD level
change, shape change, slot update). They are now recorded on the actor,
attached to the drawn instance and re-attached after the slots. A slot
taking the same bone drops the helper (last writer wins, as before);
SetRace/SetModel clear them. LODs missing a helper's bone stay on
PART_MAIN.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Back-to-front order for blended, non-depth-writing, water and additive
draws now uses the bounds centre of the vertices the draw indexes. The
centre of shared (immutable) geometry is cached per index range and
validated against the owning pointer; per-frame geometry is recomputed.
Depth-writing opaque draws keep the first-vertex tie-break.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Port CGrannyLODController level selection: _lod_01.._03 are queued
{lod_03, lod_02, lod_01, main}, loaded lazily, and chosen from the
camera-target/camera distance bands (actors) or the building factor.
Parts, skins and the current motion follow the drawn instance.
Shipped LODs that drop an attaching bone in use (lion2_lod_02/03 lose
"saddle", shaman_miyeom_lod_* lose equip_left/right) stay on PART_MAIN,
and bone queries fall back to PART_MAIN for bones a LOD lacks.
Weapon item LODs are not ported.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Water, network-actor, flying-effect and weapon-trace clocks used
frames*0.033 or frames/60, and the sky box, effect.Update, fly.Update and
render-target previews advanced a fixed 1/30 or 1/60 s per call, so their
speed followed the frame rate. They now use the measured frame delta
(game_clock = sum of dt in main; GetGameFrameDelta for the Python-driven
paths; elapsed SDL ticks for render targets). --frames runs keep their
fixed 0.033 s step. The SpeedTree wind update is left unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Port CGrannyLODController::UpdateLODLevel and
CGrannyModelInstance::Update(dwAniFPS): actors whose model has a
_lod_01.gr2 sibling and stand more than 500 units from the camera
target get 30-120 animation FPS by camera distance (LODDISTANCE_ACTOR
5000, LODDISTANCE_BUILDING 25000 for models of 500+ units), and their
pose time only advances every 120/fps frames, so the SDK's cached pose
is reused in between. Restarted or switched motions re-sample at once.
The _lod_ meshes themselves are still not swapped in.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HardwareScene now tracks the bound vertex/index buffer, pipeline and
sampler within the pass and only rebinds on change (each sampler rebind
allocates a new descriptor set on SDL's Vulkan backend). The opaque sort
orders depth-writing opaque and alpha-tested draws by cull mode, texture
and geometry so consecutive batches share state; blended and
non-depth-writing draws keep back-to-front order.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TerrainHeightmap::AppendCulledDrawRefs references the cached, shared-
geometry patch commands for the visible quadtree patches (writing the
background light into the cache) instead of copying them into a new
vector each frame. BuildCulledDraws keeps its copying behaviour on top
of it for existing callers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The SDL_GPU pipeline renders into its own view-sized colour target and
already composites it over the whole render target, so routing it through
gpu_scene_texture cost one extra same-size copy per frame. The scaled
intermediate is now used only by the SDL geometry fallback; the hardware
composite upscales with linear filtering.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AVA2 builds with ENABLE_MAP_OBJECT_OPTIMIZATION: CArea::
__UpdateLoadedObjectInstances unloads objects whose 2D distance from the
main actor is at least max(OBJECT_DIST, 5000). CPythonBackground::Update
now takes the main actor position (as CPythonApplication passes it) and
sets that loading center on MapObjectManager, which skips farther
objects before their frustum test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MapObjectManager::AppendDrawRefs references each visible object's cached
Render3DDraw (writing the per-frame light into the cache) rather than
copying the whole command (two texture strings, shared_ptrs, matrices)
every frame. main.cpp flattens owned draws and refs in submission order
(FlattenDraws) and passes a pointer list to RenderSceneGPU/RenderScene.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Per-frame UI callbacks (OnUpdate/OnRender/...) looked up methods with
PyObject_GetAttrString, building a new PyString each call. Names are
now interned once and cached; Py_AtExit drops the cache so a
Py_Finalize/Py_Initialize cycle never reuses stale objects.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The per-batch transforms are already uploaded to the storage buffer, so
draws use firstInstance = batch index and the 288-byte uniform header is
pushed once per frame; fragment params are pushed only on change.
Texture eviction sorts inactive images once instead of rescanning the
map per evicted image.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Texture() and Resident() stage their bytes on the CPU; Render copies
everything in one transfer buffer and copy pass (plus mip generation)
at the start of its own command buffer instead of acquiring and
submitting a command buffer per resource.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Colour as UBYTE4_NORM, normals as SHORT4_NORM (unit direction, the
shader renormalises), bones as UBYTE4 and weights as USHORT4_NORM.
Shader inputs keep their float/uint types, so the shaders are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
RenderSceneGPU resolves each draw's texture while classifying it and
sorts opaque draws by the texture handle instead of comparing paths;
invalid draws are dropped before sorting. Bumps sdk (skin cache).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Render3DDraw/HardwareBatch carry first_index/index_count; resident
meshes are uploaded once per model and groups draw with firstIndex.
Bumps sdk to 9ec7fa1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Textures were created with one level while the sampler requested linear
mip filtering (and max_lod defaulted to 0). Allocate the full chain,
generate it after upload, and let the sampler use it, so minified
terrain and props stop shimmering and read less bandwidth. Budget
accounting includes the extra third.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Network actor ground shadows use the actor's own IsVisible frustum
test (its bounds cover the shadow) instead of building 17-vertex
meshes with terrain queries for every actor on the map.
- Shadow ring uses a precomputed sin/cos table and reserved streams.
- Frame and network draw vectors reserve last frame's peak count and
append network draws in one insert.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Font: walk UTF-8 codepoints in place instead of building a vector;
reuse the texture-cache key buffer on the hit path.
- Text tails: show/ShowAll test isShowing (kept equal to list
membership) instead of a linear find per tail, which made the
per-frame ShowAllTextTail O(N^2); memoize label widths measured by
both Arrange and Render.
- Tail painter reuses one CGraphicTextInstance and font name.
- CGraphicTextInstance::Render no longer copies the text unless secret.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Render3DDraw/HardwareBatch hold the SDK's shared palette instead of
copying 16 floats per bone per draw, and HardwareScene re-pushes the
uniform only when the palette pointer changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With more than 1024 live meshes the fixed threshold rescanned the whole
cache on every ToRender3DDraw call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
sdk de8fa40: B-spline loop window, cubic curves, degree-0 first-knot>t.
ActorInstance passes the motion's loop flag to SetAnimationLoop so looping
motions sample across the seam instead of clamping to the end knots.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Correct selection orientation, game-entry lifecycle, mounted NPCs, labels and skill panels. Add the native GPU scene path, resident geometry, bounded texture and Python caches, and collision broad phase.
Pin the independently maintained Metin2 SDK through a submodule and document reproducible setup. Include rendering regressions and macOS packaging improvements.
Validation: client Debug build and 12 targeted client tests passed; SDK build and all 14 SDK tests passed.
- gr2: BitKnit2 decompression, 64-bit LE files, compressed curves
- net: remaining GamePhase handlers, generated GC packet table
- python: event/chat/chr/minimap/guild/quest/messenger/shop-search and
other native modules ported against the AVA2 Windows client
- render: D3D8 fixed-function lighting in Gr2ModelRenderer, terrain
without baked shade; cape/NPC equipment attachment
- tests + audits updated; docs/PORT-PLAN.md and WINDOWS-PARITY.md
Batch 4 (render pipeline: background.*, app.RenderGame/UpdateGame,
camera bindings, main.cpp fake-content removal) is still in progress.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>