test(M00): add checkpoint perceptual comparison #1

Merged
sindoring merged 2 commits from work/sindo-main-codex/m00-checkpoint-diff into master 2026-07-11 17:57:05 +03:00
9 changed files with 95 additions and 34 deletions
Showing only changes of commit d467ffee7f - Show all commits
+2
View File
@@ -4,6 +4,8 @@
Парные checkpoint PNG можно автоматически проверить через `src/tools/compare_render_checkpoints.gd`; reference-снимки оригинального клиента остаются вне Git, а автоматический tolerance не заменяет human fidelity approval.
Paired run 2026-07-11 подтвердил крупный coordinate/placement gap: некоторые server-derived camera positions оказываются под terrain или внутри WMO/rocks OpenWC. До исправления этого расхождения perceptual metrics измеряют также несовпадение композиции, а не только материалы и свет.
Цель renderer-работы в этом проекте: добиться ощущения производительности оригинального клиента WoW 3.3.5a в Godot, без видимых фризов при переходе ADT -> ADT и без постоянного отката видимых участков к низкому качеству.
Этот документ фиксирует текущее состояние рендера, сделанные оптимизации и практические правила дальнейшей работы.
@@ -66,13 +66,13 @@ Provide a deterministic, headless perceptual comparison for paired renderer chec
## Status
- State: ready
- Done: comparator, runner integration, manifest tolerance contract, synthetic regression, module/baseline documentation, repository gates
- Next: integrator review and paired run when approved build 12340 screenshots become available
- Done: comparator, runner integration, manifest tolerance contract, synthetic regression, five local build 12340 references, three viewpoint calibrations, corrected visual capture and first paired report
- Next: integrator review; coordinate/placement gap should be resolved before tolerance calibration
- Blocked by: real paired screenshots only for final human fidelity approval
## Handoff
- Commit: branch HEAD
- Results: synthetic identical image passed, changed image failed; full M00 dry-run and repository gates passed
- Remaining risks: thresholds need calibration against approved paired captures; animation phase and camera alignment still require human review
- Results: ten pairs compared with no missing candidates; mean error 0.0707..0.1746 and changed ratio 0.5504..0.8187; human review identified under-terrain/inside-geometry camera mismatches
- Remaining risks: coordinate/placement mismatch blocks meaningful tolerance calibration; synthetic animation, dusk and full human approval remain incomplete
- Documentation updated: `docs/RENDER_BASELINE.md`, `docs/modules/world-renderer.md`, `RENDER.md`
+9 -1
View File
@@ -33,7 +33,7 @@
-ReferenceCheckpointDirectory 'D:\private-fixtures\wow-3.3.5a-checkpoints'
```
Результат записывается в `user://render_baseline/visual_comparison.json`. Другой путь задаётся через `-VisualComparisonReport`. `-DryRun` несовместим с визуальным сравнением, потому что не создаёт PNG.
Результат записывается в `user://render_baseline/visual_comparison.json`. Другой путь задаётся через `-VisualComparisonReport`. `-DryRun` несовместим с визуальным сравнением, потому что не создаёт PNG. Runner сохраняет обычный performance capture в нормативном `1280×900`, затем отдельно снимает visual candidates в `2560×1440`; разрешение можно переопределить через `-VisualComparisonWidth` и `-VisualComparisonHeight`, не меняя performance baseline.
Отдельный запуск сравнения:
@@ -45,6 +45,8 @@ godot --headless --path . `
--output user://render_baseline/visual_comparison.json
```
Reference-каталог принимает JPG/JPEG/PNG оригинального клиента с именем checkpoint, например `goldshire_dense_m2.jpg`. Диагностические кадры с префиксом `diagnostic_` игнорируются. Один reference автоматически сопоставляется с `<checkpoint>__cold_process.png` и `<checkpoint>__warm_revisit.png` OpenWC.
Сравнение переводит sRGB в linear color space, вычисляет взвешенную ошибку яркости/цвета для каждого пикселя, среднюю ошибку кадра и долю пикселей выше локального порога. Нормативные defaults находятся в `comparison_budgets` manifest: mean error `0.015`, changed-pixel ratio `0.01`, pixel error threshold `0.04`. Любое превышение, несовпадение размеров, отсутствие пары или пустой reference-каталог возвращает ненулевой exit code.
Автоматический pass не является доказательством parity: animation phase, weather, camera alignment и intentional gaps требуют human approval. Для проверки алгоритма без proprietary данных используется:
@@ -112,3 +114,9 @@ Baseline пока не имеет approved парных кадров ориги
- возможная смена D3D12 на Vulkan при ошибке descriptor heap — фактический backend всегда берётся из `report.json`.
Парное сравнение с клиентом 3.3.5a должно использовать те же map position, local time и weather. Каждый gap регистрируется как terrain/material, placement, animation, lighting, liquid или culling; baseline сам по себе gap не закрывает.
Локальная сессия build 12340 от 2026-07-11 откалибровала три непригодные исходные позиции: dense-M2 camera была полностью закрыта кронами, large-WMO camera находилась у дымохода кузни, liquid camera попадала под terrain оригинального клиента. Manifest теперь содержит проверенные replacement camera positions. Пять approved локальных reference JPG хранятся вне Git; synthetic animated probe и dusk checkpoint всё ещё требуют отдельной процедуры.
Первый paired run после калибровки создал десять сравнений (пять reference × cold/warm), без missing pairs. Все пары ожидаемо превысили строгий tolerance: mean perceptual error `0.0707..0.1746`, changed-pixel ratio `0.5504..0.8187`. Human inspection показал, что это пока не чистая material/lighting ошибка: при тех же WoW-derived координатах OpenWC terrain-overview camera находится под terrain/placements, WMO camera — внутри таверны, liquid camera — внутри скал; ADT и dense-M2 композиции также существенно смещены. До исправления coordinate/placement mismatch эти значения являются gap evidence, а не основанием расширять tolerance.
Capture tool строит camera basis явно из target и world-up. Это исключает неоднозначный roll `look_at` при автоматической съёмке. `ViewportTexture.get_image()` сохраняется без дополнительного vertical flip для Godot 4.6.1.
+4 -2
View File
@@ -194,7 +194,7 @@ Exact exported settings and cache versions remain documented in [`../../RENDER.m
- Unit/contract tests: material mapping, unique-ID dedupe, placement probes, baseline manifest, synthetic perceptual checkpoint diff.
- Integration/E2E: Eastern Kingdoms/Kalimdor streaming scenes and seven cold/warm checkpoints.
- Fidelity evidence: current fixed-position captures and automated paired-PNG metrics exist; approved original-client paired screenshots and human approval are still missing.
- Fidelity evidence: пять локальных build 12340 reference JPG откалибровали terrain/ADT/M2/WMO/liquid viewpoints; automated paired-image metrics exist, но synthetic animation/dusk и полный human approval ещё не закрыты.
- Performance budgets: M00 report records cold/warm p95 and max hitch; no final acceptance threshold yet.
- Manual diagnostics: [`../RENDER_BASELINE.md`](../RENDER_BASELINE.md) and [`../../RENDER.md`](../../RENDER.md).
@@ -222,6 +222,7 @@ Exact exported settings and cache versions remain documented in [`../../RENDER.m
- Monolithic streamer mixes planning, jobs, caches and presentation.
- Direct camera path remains until M01/M03.
- Original-client paired fidelity evidence incomplete.
- Первый paired run выявил coordinate/placement mismatch: несколько совпадающих server-derived camera positions оказываются под terrain или внутри WMO/rocks OpenWC.
- D3D12 descriptor and shutdown RID/resource issues remain.
- M2/WMO/material/particle/ribbon/portal parity incomplete.
- Public API is mostly exported configuration rather than stable contracts.
@@ -239,7 +240,8 @@ Exact exported settings and cache versions remain documented in [`../../RENDER.m
| `src/native/src/*_loader.cpp` | Native binary parsing |
| `src/tools/build_*cache.gd`, `src/tools/bake_*cache.gd` | Offline cache generation |
| `tools/run_render_baseline.ps1` | Unified M00 baseline runner |
| `src/tools/compare_render_checkpoints.gd` | Offline paired-PNG perceptual metrics and JSON pass/fail report |
| `src/tools/compare_render_checkpoints.gd` | Offline JPG/PNG paired-image perceptual metrics and JSON pass/fail report |
| `src/tools/capture_render_checkpoints.gd` | Deterministic no-roll checkpoint camera, performance and visual capture |
## Related decisions and references
+13 -3
View File
@@ -25,7 +25,9 @@ func _capture_async() -> void:
var headless := DisplayServer.get_name().to_lower() == "headless"
var dry_run := args.has("--dry-run") or headless
var viewport: Array = manifest.get("viewport", [1280, 900])
get_root().size = Vector2i(maxi(16, int(viewport[0])), maxi(16, int(viewport[1])))
var viewport_width := int(_arg(args, "--viewport-width", str(viewport[0])))
var viewport_height := int(_arg(args, "--viewport-height", str(viewport[1])))
get_root().size = Vector2i(maxi(16, viewport_width), maxi(16, viewport_height))
var abs_output_dir := ProjectSettings.globalize_path(output_dir)
DirAccess.make_dir_recursive_absolute(abs_output_dir)
@@ -107,7 +109,7 @@ func _capture_async() -> void:
probe.scale = Vector3.ONE * float(checkpoint.get("probe_scale", 1.0))
camera.global_position = _vector3(checkpoint.get("camera", [0.0, 0.0, 0.0]))
camera.look_at(_vector3(checkpoint.get("target", [0.0, 0.0, 0.0])), Vector3.UP)
_orient_camera_without_roll(camera, _vector3(checkpoint.get("target", [0.0, 0.0, 0.0])))
if player != null:
player.global_position = _vector3(checkpoint.get("player", checkpoint.get("target", [0.0, 0.0, 0.0])))
_set_sky_time(world, float(checkpoint.get("time_hours", 13.0)))
@@ -135,7 +137,6 @@ func _capture_async() -> void:
await RenderingServer.frame_post_draw
var image := get_root().get_texture().get_image()
image.flip_y()
var file_name := "%s__%s.png" % [checkpoint_name, pass_name]
var abs_path := abs_output_dir.path_join(file_name)
var err := image.save_png(abs_path)
@@ -196,6 +197,15 @@ func _measure_frames(seconds: float, world: Node) -> Dictionary:
}
func _orient_camera_without_roll(camera: Camera3D, target_position: Vector3) -> void:
var forward := (target_position - camera.global_position).normalized()
var right := forward.cross(Vector3.UP).normalized()
if right.is_zero_approx():
right = Vector3.RIGHT
var corrected_up := right.cross(forward).normalized()
camera.global_basis = Basis(right, corrected_up, -forward)
func _result_record(checkpoint: Dictionary, pass_name: String, load_time_ms: float, metrics: Dictionary, sha256: String) -> Dictionary:
return {
"name": checkpoint.get("name", "checkpoint"),
+40 -20
View File
@@ -1,6 +1,6 @@
extends SceneTree
## Compares paired PNG checkpoint directories without storing proprietary reference images.
## Compares original-client JPG/PNG references with OpenWC checkpoint PNGs.
## Usage: godot --headless --path . --script res://src/tools/compare_render_checkpoints.gd --
## --reference <directory> --candidate <directory> [--output <report.json>]
## [--pixel-threshold 0.04] [--mean-threshold 0.015] [--changed-ratio-threshold 0.01]
@@ -73,22 +73,29 @@ func _parse_arguments(raw_arguments: PackedStringArray) -> Dictionary:
func _compare_directories(reference_directory: String, candidate_directory: String, options: Dictionary) -> Dictionary:
var reference_files := _png_file_names(reference_directory)
reference_directory = ProjectSettings.globalize_path(reference_directory)
candidate_directory = ProjectSettings.globalize_path(candidate_directory)
var reference_files := _reference_image_file_names(reference_directory)
var results: Array[Dictionary] = []
var failed_count := 0
var missing_count := 0
for file_name in reference_files:
var reference_path := reference_directory.path_join(file_name)
var candidate_path := candidate_directory.path_join(file_name)
if not FileAccess.file_exists(candidate_path):
results.append({"name": file_name, "status": "missing_candidate"})
var checkpoint_name := file_name.get_basename()
var candidate_file_names := _candidate_file_names(candidate_directory, checkpoint_name)
if candidate_file_names.is_empty():
results.append({"checkpoint": checkpoint_name, "reference": file_name, "status": "missing_candidate"})
missing_count += 1
continue
var comparison := _compare_images(reference_path, candidate_path, options)
comparison.name = file_name
results.append(comparison)
if comparison.status != "passed":
failed_count += 1
for candidate_file_name in candidate_file_names:
var candidate_path := candidate_directory.path_join(candidate_file_name)
var comparison := _compare_images(reference_path, candidate_path, options)
comparison.checkpoint = checkpoint_name
comparison.reference = file_name
comparison.candidate = candidate_file_name
results.append(comparison)
if comparison.status != "passed":
failed_count += 1
var no_reference_images := reference_files.is_empty()
return {
"schema_version": REPORT_SCHEMA_VERSION,
@@ -149,18 +156,32 @@ func _perceptual_color_error(reference_color: Color, candidate_color: Color) ->
return 0.2126 * red_delta + 0.7152 * green_delta + 0.0722 * blue_delta + 0.25 * alpha_delta
func _png_file_names(directory_path: String) -> Array[String]:
func _reference_image_file_names(directory_path: String) -> Array[String]:
var directory := DirAccess.open(directory_path)
if directory == null:
return []
var file_names: Array[String] = []
for file_name in directory.get_files():
if file_name.to_lower().ends_with(".png"):
var lower_name := file_name.to_lower()
if not lower_name.begins_with("diagnostic_") and (lower_name.ends_with(".png") or lower_name.ends_with(".jpg") or lower_name.ends_with(".jpeg")):
file_names.append(file_name)
file_names.sort()
return file_names
func _candidate_file_names(candidate_directory: String, checkpoint_name: String) -> Array[String]:
var candidates: Array[String] = []
for pass_name in ["cold_process", "warm_revisit"]:
var file_name := "%s__%s.png" % [checkpoint_name, pass_name]
if FileAccess.file_exists(candidate_directory.path_join(file_name)):
candidates.append(file_name)
if candidates.is_empty():
var exact_file_name := "%s.png" % checkpoint_name
if FileAccess.file_exists(candidate_directory.path_join(exact_file_name)):
candidates.append(exact_file_name)
return candidates
func _write_report(output_path: String, report: Dictionary) -> bool:
var absolute_path := ProjectSettings.globalize_path(output_path)
var error := DirAccess.make_dir_recursive_absolute(absolute_path.get_base_dir())
@@ -186,14 +207,13 @@ func _run_self_test() -> int:
reference_image.fill(Color(0.25, 0.5, 0.75, 1.0))
var changed_image := reference_image.duplicate()
changed_image.set_pixel(0, 0, Color.WHITE)
for pair in [
[reference_directory, reference_image],
[identical_directory, reference_image],
[changed_directory, changed_image],
]:
var save_error: Error = pair[1].save_png(String(pair[0]).path_join("synthetic.png"))
if save_error != OK:
push_error("RENDER_CHECKPOINT_DIFF SELF_TEST: cannot save fixture")
if reference_image.save_jpg(reference_directory.path_join("synthetic.jpg"), 1.0) != OK:
push_error("RENDER_CHECKPOINT_DIFF SELF_TEST: cannot save JPG reference")
return 1
for pass_name in ["cold_process", "warm_revisit"]:
if reference_image.save_png(identical_directory.path_join("synthetic__%s.png" % pass_name)) != OK:
return 1
if changed_image.save_png(changed_directory.path_join("synthetic__%s.png" % pass_name)) != OK:
return 1
var options := {
"pixel_threshold": DEFAULT_PIXEL_THRESHOLD,
+4 -3
View File
@@ -5,6 +5,7 @@
"map": "Azeroth",
"quality_preset": "High",
"viewport": [1280, 900],
"fidelity_comparison_viewport": [2560, 1440],
"camera_fov": 62.0,
"default_wait_seconds": 8.0,
"default_measure_seconds": 3.0,
@@ -57,7 +58,7 @@
{
"name": "goldshire_dense_m2",
"coverage": ["dense_m2"],
"camera": [16835.0, 135.0, 26435.0],
"camera": [16956.666, 150.0, 26466.666],
"target": [17015.0, 62.0, 26525.0],
"player": [17015.0, 58.0, 26525.0],
"time_hours": 13.0
@@ -65,7 +66,7 @@
{
"name": "goldshire_inn_large_wmo",
"coverage": ["large_wmo"],
"camera": [16942.0, 82.0, 26503.0],
"camera": [17013.666, 72.0, 26451.666],
"target": [17042.27, 66.0, 26530.91],
"player": [17042.27, 58.0, 26530.91],
"time_hours": 13.0
@@ -73,7 +74,7 @@
{
"name": "elwynn_waterfall_liquid",
"coverage": ["liquid"],
"camera": [16445.0, 125.0, 26295.0],
"camera": [16481.666, 190.0, 26366.666],
"target": [16518.84, 68.0, 26427.27],
"player": [16518.84, 55.0, 26427.27],
"time_hours": 13.0
@@ -20,6 +20,8 @@ func _initialize() -> void:
_expect(int(manifest.get("schema_version", 0)) == 1, "manifest schema must be version 1", failures)
_expect(String(manifest.get("profile", "")) == "Blizzlike335", "manifest profile must be Blizzlike335", failures)
_expect(ResourceLoader.exists(String(manifest.get("scene", ""))), "streaming scene must exist", failures)
var fidelity_viewport: Array = manifest.get("fidelity_comparison_viewport", [])
_expect(fidelity_viewport.size() == 2 and int(fidelity_viewport[0]) > 0 and int(fidelity_viewport[1]) > 0, "fidelity comparison viewport must contain positive width and height", failures)
var covered := {}
var names := {}
+17 -1
View File
@@ -7,6 +7,8 @@ param(
[double]$MeasureSeconds = 3.0,
[string]$ReferenceCheckpointDirectory,
[string]$VisualComparisonReport = "user://render_baseline/visual_comparison.json",
[int]$VisualComparisonWidth = 2560,
[int]$VisualComparisonHeight = 1440,
[switch]$DryRun
)
@@ -63,12 +65,26 @@ try {
if ($DryRun) {
throw "Reference checkpoint comparison requires PNG capture; remove -DryRun"
}
$visualCandidateOutput = "$Output/visual_comparison_candidates"
$visualCaptureArgs = @(
"--path", ".",
"--script", "res://src/tools/capture_render_checkpoints.gd",
"--",
"--output", $visualCandidateOutput,
"--cache-state", $CacheState,
"--revision", $revision,
"--wait", $WaitSeconds.ToString([Globalization.CultureInfo]::InvariantCulture),
"--measure", $MeasureSeconds.ToString([Globalization.CultureInfo]::InvariantCulture),
"--viewport-width", $VisualComparisonWidth.ToString(),
"--viewport-height", $VisualComparisonHeight.ToString()
)
Invoke-GodotStep "visual-capture" $visualCaptureArgs
Invoke-GodotStep "visual-comparison" @(
"--headless", "--path", ".",
"--script", "res://src/tools/compare_render_checkpoints.gd",
"--",
"--reference", $ReferenceCheckpointDirectory,
"--candidate", $Output,
"--candidate", $visualCandidateOutput,
"--output", $VisualComparisonReport
)
}