saumwacht-godot/tools/luneraria/lune_benchmark.gd

extends Node
## Frame and simulation cost in a real match scene at representative maximum occupation:
## six armies at the supply cap and a round-13 wave on all three lanes (synthetic fixture,
## gold granted to the fixture only). Two camera setups are measured one after the other:
## a lane view (Mid, Blau side, grotto in the background) and the whole-map overview,
## where every lane, tower, base and the grotto are drawn at once. VSync off; the
## figures apply to this machine only.
## Run: godot --path . --resolution 1920x1080 res://tools/luneraria/lune_benchmark.tscn -- --out=res://docs/luneraria/logs/g1/perf.json
const Game = preload("res://scripts/luneraria/lune_game.gd")
const PHASE_SECONDS: float = 6.0

var out: String = ""
## Start of the current frame's main-thread work (set by a node that processes first).
var frame_start: int = 0

class FrameClock:
	extends Node
	var owner_bench: Node
	func _process(_delta: float) -> void:
		owner_bench.frame_start = Time.get_ticks_usec()

func _ready() -> void:
	for arg: String in OS.get_cmdline_user_args():
		if arg.begins_with("--out="):
			out = arg.substr(6)
	var clock := FrameClock.new()
	clock.owner_bench = self
	clock.process_priority = -1000
	add_child(clock)
	call_deferred("run")

func mean(values: Array) -> float:
	var total: float = 0.0
	for value: Variant in values:
		total += float(value)
	return total / maxf(1.0, values.size())

func measure(game: Node3D, state: LuneMatch) -> Dictionary:
	for i: int in 30:
		await get_tree().process_frame
	var frames: Array[float] = []
	var cpu: Array[float] = []
	var gpu: Array[float] = []
	var draws: Array[float] = []
	var primitives: Array[float] = []
	var objects: Array[float] = []
	var mobile: Array[int] = []
	var last: int = Time.get_ticks_usec()
	var end: int = Time.get_ticks_msec() + int(PHASE_SECONDS * 1000.0)
	while Time.get_ticks_msec() < end and state.phase == LuneMatch.Phase.COMBAT:
		await RenderingServer.frame_post_draw
		var now: int = Time.get_ticks_usec()
		frames.append((now - last) / 1000.0)
		last = now
		# Main-thread time from the first _process of the frame until the frame was drawn.
		cpu.append((now - frame_start) / 1000.0)
		gpu.append(RenderingServer.viewport_get_measured_render_time_gpu(get_viewport().get_viewport_rid()))
		draws.append(Performance.get_monitor(Performance.RENDER_TOTAL_DRAW_CALLS_IN_FRAME))
		primitives.append(Performance.get_monitor(Performance.RENDER_TOTAL_PRIMITIVES_IN_FRAME))
		objects.append(Performance.get_monitor(Performance.RENDER_TOTAL_OBJECTS_IN_FRAME))
		mobile.append(state.combat.mobile_count())
	frames.sort()
	return {"frames": frames.size(), "frame_ms_mean": mean(frames), "frame_ms_p95": frames[int(frames.size() * .95)] if not frames.is_empty() else -1.0,
		"cpu_frame_ms_mean": mean(cpu), "gpu_ms_mean": mean(gpu), "draw_calls_mean": mean(draws), "primitives_mean": mean(primitives),
		"objects_mean": mean(objects), "mobile_units_mean": mean(mobile), "camera_zoom": game.camera.zoom}

func run() -> void:
	Soundscape.enabled = false
	DisplayServer.window_set_vsync_mode(DisplayServer.VSYNC_DISABLED)
	var t0: int = Time.get_ticks_usec()
	var game: Node3D = Game.new()
	add_child(game)
	game.start({"seed": 5, "human_slot": 0})
	var build_ms: float = (Time.get_ticks_usec() - t0) / 1000.0
	var state: LuneMatch = game.match_state
	state.round_number = 13
	for slot: int in range(6):
		state.players[slot].draft.gold = 100000
		state.players[slot].draft.roster.clear()
		LuneBot.plan(state, slot, "ausgewogen")
	game._on_ready()
	RenderingServer.viewport_set_measure_render_time(get_viewport().get_viewport_rid(), true)
	game.speed_index = 1
	game.camera.look_at_point(Vector3(-10, 0, 10), 34.0, true)
	var lane_view: Dictionary = await measure(game, state)
	game.camera.overview(true)
	var overview: Dictionary = await measure(game, state)
	# Pure simulation cost of one fixed step at this occupation, measured separately.
	var probe := LuneMatch.new()
	probe.setup({"seed": 5, "human_slot": -1})
	probe.round_number = 13
	for slot: int in range(6):
		probe.players[slot].draft.gold = 100000
		LuneBot.plan(probe, slot, "ausgewogen")
	probe.begin_combat()
	var start_units: int = probe.combat.mobile_count()
	var s1: int = Time.get_ticks_usec()
	for i: int in 200:
		probe.step_combat()
	var step_ms: float = (Time.get_ticks_usec() - s1) / 1000.0 / 200.0
	var report := {"window": get_window().size, "renderer": RenderingServer.get_current_rendering_method(), "gpu": RenderingServer.get_video_adapter_name(),
		"cpu": OS.get_processor_name(), "engine": Engine.get_version_info().string, "scene_build_ms": build_ms,
		"mobile_units_at_start": start_units, "buildings": 18, "sim_step_ms": step_ms, "sim_speed": "2x (40 steps/s)",
		"lane_view": lane_view, "overview": overview,
		"video_mem_mb": Performance.get_monitor(Performance.RENDER_VIDEO_MEM_USED) / 1048576.0,
		"texture_mem_mb": Performance.get_monitor(Performance.RENDER_TEXTURE_MEM_USED) / 1048576.0,
		"static_mem_mb": Performance.get_monitor(Performance.MEMORY_STATIC) / 1048576.0}
	var path: String = out if not out.is_empty() else "res://docs/luneraria/logs/perf-%dx%d.json" % [get_window().size.x, get_window().size.y]
	DirAccess.make_dir_recursive_absolute(ProjectSettings.globalize_path(path.get_base_dir()))
	var file := FileAccess.open(path, FileAccess.WRITE)
	file.store_string(JSON.stringify(report, "\t"))
	file.close()
	print("LUNE_BENCH ", JSON.stringify(report))
	get_tree().quit()