mirror of
https://github.com/sven-n/MuMain
synced 2026-08-15 20:32:27 -04:00
0702144c84
Two changes bundled together: moving skeletal skinning off the CPU onto the
GPU, and retiring the fixed-function pipeline (FFP) in favor of a real
backend-agnostic render hardware interface (RHI) with a GL implementation.
GPU bone skinning
- Per-vertex skeletal skinning (BMD::Transform()'s CPU loop, previously run
for every vertex of every visible body, every frame) moves to the GPU:
the CPU now only computes and uploads the (much smaller) per-bone matrix
palette; the vertex shader does the per-vertex transform in parallel.
- Measured result: `Transform()` self-time at a standard crowded spot
dropped from ~0.25ms to 0.01-0.02ms -- a 15-20x reduction. This is exactly
the cost that scales worst in an MMO client (town crowds, mass PK/boss
events), so the win is disproportionately larger in the scenarios that
used to hurt most.
- Cloth cape meshes are the one remaining CPU-skinned case (pinned-vertex
physics read/write-back), deliberately deferred rather than forced into
this pass.
Eliminating the fixed-function pipeline
- Render/Core/ImmediateRenderer.{h,cpp} ("IR"): a streaming-VBO bridge that
mirrors the legacy glBegin/glEnd/glVertex/glColor/glTexCoord immediate-mode
API 1:1 (decomposing GL_QUADS to GL_TRIANGLES), so the ~hundreds of legacy
immediate-mode call sites across UI/effects code didn't need a per-site
rewrite -- they draw through IR, which streams into a VBO and issues a real
shader-backed glDrawArrays instead of relying on FFP's glBegin/glEnd.
- Render/Core/GlobalUBO.{h,cpp} / SceneUBO.{h,cpp} / BoneUBO.{h,cpp}: replace
the FFP matrix stack (glMatrixMode/glLoadMatrix/glMultMatrix) and per-bone
transform state with explicit uniform buffers the shaders read instead.
- Render/Core/GLColorIntercept.h: glColor*() under Core Profile is a no-op
and glGetFloatv(GL_CURRENT_COLOR, ...) is illegal, so ~340 RenderColor()/
RenderBitmap()-family call sites that used to read back "current color"
via that query now read a shadow g_CurrentColor mirrored by macro-
intercepted glColor* wrappers instead.
- Render/Core/BindState.{h,cpp}: single-source-of-truth glUseProgram/
glBindTexture/glBindVertexArray wrapper with last-bound-id caching, so
shader/texture/VAO state is tracked explicitly instead of relying on GL's
implicit FFP-era binding state.
- tools/check_gl_wrapper_monopoly.py, wired into CMake as an always-rerun
build target (CheckGLWrapperMonopoly): fails the build if a raw gl*() call
appears outside Render/ that isn't on its allowlist, so no new FFP call
site can sneak back in. The current allowlist is a deliberately documented
set of deferrals (screenshot readback, one-time startup queries, event fog
params), not oversights.
- ~60 supporting files (Models/Effects/UI/GameMaps/Terrain/etc.) updated to
route through the new wrapper family instead of calling gl* directly.
RHI abstraction
- Render/RHI/{RHI,RHI_GL}.{h,cpp}: buffer/texture/uniform-block/draw
abstraction so rendering code targets a stable interface rather than raw
GL calls -- a GL implementation today, sized to add other backends later
without touching call sites.
- Render/Shaders/{BMDMeshShader,TerrainShader,PassthroughShader,
PlanarShadowShader,ItemSpecularShader}.{h,cpp} (+ glsl/*.vert/.frag):
GLSL shader pairs replacing the fixed-function mesh/terrain/UI-sprite/
planar-shadow/item-specular rendering paths -- this is also what the GPU
skinning above actually runs inside of.
|
||
|---|---|---|
| .. | ||
| Properties/PublishProfiles | ||
| .gitignore | ||
| Common.xslt | ||
| ConnectionManager.ChatServer.Custom.cs | ||
| ConnectionManager.ChatServerFunctions.cs | ||
| ConnectionManager.ClientToServer.Custom.cs | ||
| ConnectionManager.ClientToServerFunctions.cs | ||
| ConnectionManager.ConnectServerFunctions.cs | ||
| ConnectionManager.cs | ||
| ConnectionWrapper.cs | ||
| GenerateBindingsHeader.xslt | ||
| GenerateEnums.xslt | ||
| GenerateExtensionsDotNet.xslt | ||
| GenerateFunctions.xslt | ||
| GenerateFunctionsHeader.xslt | ||
| MUnique.Client.Library.csproj | ||
| NativeInterop.cs | ||