Phase 1 mid-review + stress harness
- docs/MIDREVIEW.md: what M0'-M2.5 proved, bridge-layer LOC (~1100), key technical conclusions (4x4 transpose, animation method B, no handedness flip needed), remaining risks, recommendation = enter Phase 2 with a 3-item wrap-up (compare.py, dump_materials oracle, refocused perf). - harness MTGODOT_STRESS=N: grid of N animated warriors, vsync off, dumps test/godot-macos-stress.json. Frame times are throttle-capped in a background window (~30fps to 50 chars, 52ms at 100); the reliable signals are build cost ~165ms/char and the 100-char breakout. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WaHYEY9rwLWt21PULiYjeJ
This commit is contained in:
Claude
@@ -32,9 +32,12 @@
|
||||
opaque/alpha/alpha-test/add 四模式）+ 目录扫描贴图解析 + 方向光阴影 + procedural sky
|
||||
+ 引擎雾（`test/golden/m25-*.png`）。**门禁未闭**：多材质槽 → 贴图/混合的精确映射需
|
||||
给 libgr2 加材质 API（见 `docs/GODOT-POC-PLAN.md` §M2.5 T2.5.6 差距清单）。
|
||||
- [ ] **中期评审** — 待 M2.5 门禁 + 对拍报告
|
||||
- [~] **中期评审** — [`docs/MIDREVIEW.md`](docs/MIDREVIEW.md)。结论：桥接层工程量小、
|
||||
风险基本出清，**建议进入 Phase 2**。收尾 3 项：`compare.py` 对拍、`dump_materials`
|
||||
Granny 对拍闭合 M2.5 门禁、前台/真机性能重测。
|
||||
|
||||
对拍（与 `xrender-poc` bgfx demo / oracle 逐帧）尚未接线，是中期评审前的下一步。
|
||||
50 角色压力采样：`test/godot-macos-stress.json`（后台窗口 30fps 节流，仅构建成本
|
||||
~165ms/角色 和 100 角色破顶到 52ms 是可信信号）。
|
||||
|
||||
---
|
||||
|
||||
|
||||
@@ -0,0 +1,60 @@
|
||||
# Phase 1 中期评审 — mtgodot-poc
|
||||
|
||||
> 日期：2026-08-29 · 平台：macOS（Apple M2）· Godot 4.7.1 · 对标 [`GODOT-POC-PLAN.md`](./GODOT-POC-PLAN.md) §00 Phase 1
|
||||
> 结论先行：**桥接层工程量小、风险已基本出清，建议进入 Phase 2（移动端）**。M2.5 门禁的"材质精确映射"仍有一处需要 oracle 的开口，不阻塞。
|
||||
|
||||
---
|
||||
|
||||
## 1. 做到了什么（M0'–M2.5）
|
||||
|
||||
| 里程碑 | 状态 | 证据 |
|
||||
|---|---|---|
|
||||
| **M0'** 脚手架 | ✅ | GDExtension 加载、`Metin2Model`/`Metin2AnimPlayer` 注册、`libgr2` 静态链接、Godot 4.7 工程 Metal Forward+ 出画 |
|
||||
| **M1** 静态渲染 | ✅ | `warrior_cheongrin`：75 骨 / 5 表面 / 3324 顶点，正立、朝 +Z、`unit_scale=0.01`、`flip_z/flip_winding=false`、贴图正确（`test/golden/m1-bindpose.png`）。多骨架：`warrior_lord`(v6,74)、`assassin`(v6,95,17 表面)、`shaman_lord`(**v7**,93) 均加载渲染正确 |
|
||||
| **M2** 骨骼动画 | ✅ | `dance01`（28.3s / 74 track）经 `gr2::sample_pose`（跨文件）→ `Skeleton3D.set_bone_global_pose` → Godot GPU 蒙皮播放。`selfcheck` 25 采样 **0 NaN/Inf**。t=9/16 姿势正确、无蒙皮撕裂（`test/golden/m2c-*.png`、`m25-warrior-shader.png`） |
|
||||
| **M2.5** 材质与观感 | 🟡 门禁未闭 | `ShaderMaterial`（modulate tex×COLOR×light；opaque/alpha/alpha-test/add 四模式）+ 目录扫描贴图解析（warrior 5/5、assassin 17/17、shaman 1/1 表面命中）+ 方向光阴影 + procedural sky + 引擎雾（`test/golden/m25-*.png`）。`libgr2::dump_materials()` 已加（见 §4） |
|
||||
|
||||
## 2. 桥接层工程量（实测）
|
||||
|
||||
新增 C++ ≈ **1100 行**（含端口自 xrender-poc 的 DXT 解码 160 行）：
|
||||
|
||||
| 文件 | 行 | 职责 |
|
||||
|---|---|---|
|
||||
| `gr2_bridge.{h,cpp}` | ~260 | gr2 Mat4→`Transform3D`（4×4 转置）、`build_skeleton`/`build_skin`/`build_mesh`、Z-up→Y-up `make_conv` |
|
||||
| `metin2_model.{h,cpp}` | ~470 | `Metin2Model` 节点、`load_gr2`、贴图解析、材质、属性 |
|
||||
| `metin2_anim.{h,cpp}` | ~230 | `Metin2AnimPlayer`、每帧 `sample_pose`→骨骼、`selfcheck` |
|
||||
| `m2_material.{h,cpp}` | ~130 | Metin2 风格 `ShaderMaterial`（两个 shader：blend_mix / blend_add） |
|
||||
| `dxt.{h,cpp}` | ~160 | DDS DXT1/3/5 软解（端口） |
|
||||
| libgr2 `gr2_material.cpp`（xrender-poc） | ~110 | `dump_materials()` 类型树遍历（附加） |
|
||||
|
||||
**耗时**：M0'–M2.5 一次会话内跑通。与计划 §05 估的 Phase 1「3.5–5 周单人」相比，**核心桥接远比预估轻** —— 因为 `libgr2` 已完成、Godot 的 `Skeleton3D`/`ArrayMesh`/`Skin`/GPU 蒙皮直接可用，且关键恒等式（`skin_matrix = global_pose(i)·bind_pose(i) == gr2 deformer 矩阵`）成立，不需要自研蒙皮。剩余预估工作量集中在 M2.5 材质精修 + 对拍。
|
||||
|
||||
## 3. 关键技术结论
|
||||
|
||||
- **坐标/单位**：gr2 行主序 Mat4 → Godot `Transform3D` 就是一次 **4×4 转置**（gr2 平移在第 4 行 = Godot `.origin`）。Z-up cm → Y-up m（`rotate(-90°,X)` + `scale(0.01)`）放在 `Metin2Model` 节点自身 transform。四个测试骨架全部 `flip_z=false`、`flip_winding=false` 正确 —— **Metin2/Granny 内容对 Godot 右手系不需要额外镜像**。
|
||||
- **动画走 B 方案**：`_process` 里 `sample_pose` 出 `world[]`，`skel.set_bone_global_pose(i, conv(world[i]))`，Skin bind = `conv(inverse_world[i])`。Godot 的 GPU 蒙皮矩阵 `global_pose(i)·bind(i)` 恒等于 gr2 deformer 矩阵。**不需要烘焙 `Animation` 资源** 即可保真（A 方案仅在移动端多角色 CPU 压不住时才需要）。
|
||||
- **贴图**：Godot 4 `Image.load()` 不支持 `.dds` → 运行时软解 DXT（`dxt.cpp`）。`.gr2` 用 `.gdignore` + 绝对路径加载，绕开 Godot import。
|
||||
- **移动端预判更好**：M2.5 的 `ShaderMaterial` + `ArrayMesh` + `Skeleton3D` 都在 Godot Compatibility(GLES3) 支持范围内 —— Phase 2 老安卓机兜底比"Godot 里塞自研渲染器"那条路好。
|
||||
|
||||
## 4. 剩余风险 / 门禁开口
|
||||
|
||||
| 项 | 影响 | 出路 |
|
||||
|---|---|---|
|
||||
| **M2.5 材质精确映射未闭**（唯一门禁开口）| `libgr2::dump_materials()` 已能取材质名 + `FromFileName`（warrior 4 材质 / 2 有 diffuse 名），但 material_index→贴图 的映射**未对拍 Granny**，实测驱动 warrior 时 body 表面误取 face 贴图。当前默认走文件名+目录扫描启发式（观感正确），material 路径 `use_gr2_materials` 开关默认关 | 需要 Granny oracle 逐材质对拍（与 xrender-poc M1 T5 同一缺口，共用修法）。这是 Phase 1 收尾的最后一项 |
|
||||
| **对拍未接线** | 与 xrender-poc bgfx demo / oracle 的逐帧几何/骨骼矩阵对拍脚本（`test/compare.py`）未写 | 需先构建 xrender-poc bgfx demo（首次 ~10 min）。中期评审后、进 Phase 2 前补 |
|
||||
| **性能数字被节流** | 后台窗口下 Godot 限 30fps，1–50 角色帧时间都贴在 33ms 天花板；100 角色破顶到 52ms。资产构建 **~165ms/角色**（单线程，无跨模型缓存），100 角色加载 16.5s（`test/godot-macos-stress.json`） | 真机 / 前台窗口下重测（Phase 2）。构建成本需 Phase 2 做多线程 + eterpack 共享贴图 |
|
||||
| **多材质槽内拆分** | `build_mesh` 目前一个 gr2 mesh = 一个 surface，未按 `tri_groups[].material_index` 再拆 | 有 material API 对拍后一起做 |
|
||||
| ShaderMaterial 静态 Shader 退出泄漏、alpha 排序 corner case、法线未逆转置 | 均为已知小项，见 `GODOT-POC-PLAN.md` §M2.5 T2.5.6 | 不阻塞 |
|
||||
|
||||
## 5. 建议
|
||||
|
||||
**进入 Phase 2（iOS + Android）。**
|
||||
|
||||
理由：Phase 1 要证明的三件事里，①gr2→Godot 场景对象、②动画运行时移植，**已用可运行代码 + 截图 + 0-NaN 自检证明可行且工程量小**；③观感还原的基础设施（自定义 shader、阴影、sky、fog、多骨架）已就位，只差"材质精确映射"这一处需 oracle 的收尾，不构成方案性风险。
|
||||
|
||||
Phase 2 前的收尾清单（1–1.5 周）：
|
||||
1. 写 `test/compare.py`，与 xrender-poc bgfx demo 同机位 / 同 t 对拍几何 + 骨骼矩阵。
|
||||
2. 补 `libgr2::dump_materials` 的 Granny oracle 对拍，闭合 M2.5 门禁。
|
||||
3. 前台窗口 / 真机重测性能，给 Phase 2 的动画 A/B 一个可信基线。
|
||||
|
||||
不建议现在做的：地形、`.mse` 特效、eterpack、UI/Python —— 均按计划留到正式移植。
|
||||
@@ -31,6 +31,12 @@ func _ready() -> void:
|
||||
_build_environment()
|
||||
_build_sun()
|
||||
_build_camera()
|
||||
|
||||
var stress := OS.get_environment("MTGODOT_STRESS")
|
||||
if stress != "":
|
||||
await _run_stress(int(stress))
|
||||
return
|
||||
|
||||
_load_content()
|
||||
_frame_model()
|
||||
print("[mtgodot] harness ready")
|
||||
@@ -197,6 +203,80 @@ func _on_key(kc: int) -> void:
|
||||
_anim.call("set_time", _anim.call("get_time") + 0.05)
|
||||
|
||||
|
||||
## MTGODOT_STRESS=N : spawn N animated warriors in a grid, measure frame time
|
||||
## over ~180 frames, write test/godot-macos-stress.json, quit.
|
||||
## Feeds the Phase-2 animation A/B (per-frame set_bone_global_pose vs baked).
|
||||
func _run_stress(n: int) -> void:
|
||||
DisplayServer.window_set_vsync_mode(DisplayServer.VSYNC_DISABLED)
|
||||
Engine.max_fps = 0
|
||||
var model_path := _assets_root.path_join("PC/ymir work/pc/warrior/warrior_cheongrin.gr2")
|
||||
var tex_dir := _assets_root.path_join("PC/ymir work/pc/warrior")
|
||||
var anim_path := _assets_root.path_join("PC/ymir work/pc/warrior/action/dance_1.gr2")
|
||||
var cols := int(ceil(sqrt(float(n))))
|
||||
var t_build_start := Time.get_ticks_usec()
|
||||
for i in n:
|
||||
var m: Node3D = ClassDB.instantiate("Metin2Model")
|
||||
m.name = "M%d" % i
|
||||
m.set("texture_dir", tex_dir)
|
||||
m.set("gr2_path", model_path)
|
||||
m.position = Vector3((i % cols) * 1.2 - cols * 0.6, 0, (i / cols) * 1.2)
|
||||
add_child(m)
|
||||
var a: Node3D = ClassDB.instantiate("Metin2AnimPlayer")
|
||||
a.set("model_path", NodePath("../M%d" % i))
|
||||
a.set("anim_path", anim_path)
|
||||
a.set("time_scale", 1.0 + 0.01 * i) # desync
|
||||
add_child(a)
|
||||
a.call("reload")
|
||||
a.call("set_time", 0.13 * i)
|
||||
var build_ms := (Time.get_ticks_usec() - t_build_start) / 1000.0
|
||||
|
||||
_target = Vector3(0, 1.0, cols * 0.6)
|
||||
_dist = cols * 2.2
|
||||
_pitch = 0.35
|
||||
_update_camera()
|
||||
|
||||
# warm-up
|
||||
for i in 40:
|
||||
await get_tree().process_frame
|
||||
var total: Array[float] = []
|
||||
var proc: Array[float] = []
|
||||
for i in 240:
|
||||
await get_tree().process_frame
|
||||
total.append(1000.0 / maxf(Engine.get_frames_per_second(), 1.0))
|
||||
proc.append(Performance.get_monitor(Performance.TIME_PROCESS) * 1000.0)
|
||||
total.sort()
|
||||
proc.sort()
|
||||
var tsum := 0.0
|
||||
for s in total:
|
||||
tsum += s
|
||||
var psum := 0.0
|
||||
for s in proc:
|
||||
psum += s
|
||||
var report := {
|
||||
"characters": n,
|
||||
"build_ms_total": build_ms,
|
||||
"build_ms_per_char": build_ms / n,
|
||||
"frames": total.size(),
|
||||
"frame_ms_avg": tsum / total.size(),
|
||||
"frame_ms_p95": total[int(total.size() * 0.95)],
|
||||
"frame_ms_max": total[-1],
|
||||
"fps_avg": 1000.0 / (tsum / total.size()),
|
||||
"cpu_process_ms_avg": psum / proc.size(),
|
||||
"cpu_process_ms_p95": proc[int(proc.size() * 0.95)],
|
||||
"renderer": RenderingServer.get_video_adapter_name(),
|
||||
"method": "per-frame set_bone_global_pose (B)",
|
||||
}
|
||||
var out := OS.get_environment("MTGODOT_STRESS_OUT")
|
||||
if out == "":
|
||||
out = ProjectSettings.globalize_path("res://../test/godot-macos-stress.json")
|
||||
var f := FileAccess.open(out, FileAccess.WRITE)
|
||||
if f:
|
||||
f.store_string(JSON.stringify(report, " "))
|
||||
f.close()
|
||||
print("[mtgodot] stress ", n, " -> ", report)
|
||||
get_tree().quit()
|
||||
|
||||
|
||||
func _screenshot() -> void:
|
||||
var img := get_viewport().get_texture().get_image()
|
||||
var path := "user://shot_%d.png" % Time.get_ticks_msec()
|
||||
|
||||
@@ -0,0 +1,63 @@
|
||||
{
|
||||
"note": "macOS, background/unfocused window -> Godot throttles to ~30fps; frame_ms rows at ~33.3 are the throttle ceiling, not measured cost. build_ms_per_char and the 100-char breakout (52ms) are the reliable signals. Proper frame timing = Phase 2 (focused window / on-device).",
|
||||
"renderer": "Apple M2 (Apple8)",
|
||||
"method": "per-frame set_bone_global_pose (B)",
|
||||
"runs": [
|
||||
{
|
||||
"build_ms_per_char": 170.682,
|
||||
"build_ms_total": 170.682,
|
||||
"characters": 1,
|
||||
"cpu_process_ms_avg": 34.1602583333333,
|
||||
"cpu_process_ms_p95": 34.405,
|
||||
"fps_avg": 29.345480700072,
|
||||
"frame_ms_avg": 34.0767973856208,
|
||||
"frame_ms_max": 58.8235294117647,
|
||||
"frame_ms_p95": 33.3333333333333,
|
||||
"frames": 240,
|
||||
"method": "per-frame set_bone_global_pose (B)",
|
||||
"renderer": "Apple M2 (Apple8)"
|
||||
},
|
||||
{
|
||||
"build_ms_per_char": 167.27,
|
||||
"build_ms_total": 1672.7,
|
||||
"characters": 10,
|
||||
"cpu_process_ms_avg": 35.1748875,
|
||||
"cpu_process_ms_p95": 41.667,
|
||||
"fps_avg": 30.0000000000001,
|
||||
"frame_ms_avg": 33.3333333333332,
|
||||
"frame_ms_max": 33.3333333333333,
|
||||
"frame_ms_p95": 33.3333333333333,
|
||||
"frames": 240,
|
||||
"method": "per-frame set_bone_global_pose (B)",
|
||||
"renderer": "Apple M2 (Apple8)"
|
||||
},
|
||||
{
|
||||
"build_ms_per_char": 164.90146,
|
||||
"build_ms_total": 8245.073,
|
||||
"characters": 50,
|
||||
"cpu_process_ms_avg": 34.5210166666665,
|
||||
"cpu_process_ms_p95": 34.624,
|
||||
"fps_avg": 30.1173930643639,
|
||||
"frame_ms_avg": 33.203405017921,
|
||||
"frame_ms_max": 33.3333333333333,
|
||||
"frame_ms_p95": 33.3333333333333,
|
||||
"frames": 240,
|
||||
"method": "per-frame set_bone_global_pose (B)",
|
||||
"renderer": "Apple M2 (Apple8)"
|
||||
},
|
||||
{
|
||||
"build_ms_per_char": 164.65813,
|
||||
"build_ms_total": 16465.813,
|
||||
"characters": 100,
|
||||
"cpu_process_ms_avg": 54.0874916666667,
|
||||
"cpu_process_ms_p95": 62.877,
|
||||
"fps_avg": 19.1516169676607,
|
||||
"frame_ms_avg": 52.2149122807017,
|
||||
"frame_ms_max": 52.6315789473684,
|
||||
"frame_ms_p95": 52.6315789473684,
|
||||
"frames": 240,
|
||||
"method": "per-frame set_bone_global_pose (B)",
|
||||
"renderer": "Apple M2 (Apple8)"
|
||||
}
|
||||
]
|
||||
}
|
||||
Reference in New Issue
Block a user