ludic/packages/ludic.render3d/drawstats.ludic
Orkuncakilkaya 5d0f83b81c render3d: 671 frame-reachable allocation sites to 161 - declared, split, or made nothing
@alloc_ok with its reason on what is made once per resource and kept (textures, programs, samplers,
views, layouts, memory blocks, the pipeline cache in gvk_pipe_build), on resize and swapchain
remakes, on screenshots and dumps, on a world being set up (streams, layers, water, post, bakes), on
loads (gltf_load, skin_load, tex_load*, png_decode, fonts) and on R3D_PROF / drawstats.
gltf_cached's hit path is its own and makes nothing; the miss (gltf_cached_load) is declared.
m4_look_at (now over m4_look_at_xyz) and m4_inverse work on scalars, and cam_update makes no scratch.
Left: lazy first-use starts, error and debug prints, and scratch made and freed each call (churn,
plan 25.3). Compiled (steady, and ludic deps over the game).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-09-28 15:58:18 +03:00

191 lines
9.6 KiB
Text

# ============================================================================
# drawstats.ludic — every draw a frame makes, by pass, program and kind
# (R3D_DRAWSTATS=<frames>).
#
# Measurement only. Every draw on both backends goes through the five doors in
# gpu.ludic (gpu_draw_mesh, _instanced, _bound_elements, _range, _mesh_indirect);
# each door tallies itself here under the profiler pass that is open (prof_begin)
# and the program that is bound. Nothing about the frame changes.
#
# R3D_DRAWSTATS=<n> count n frames (default 60 when not a number)
# R3D_DRAWSTATS_FROM=<f> start at render frame f (default 90: streaming settled)
# R3D_DRAWSTATS_CSV=<path> the per-row table (default build/drawstats.csv)
#
# The report prints when the window closes, so any run long enough to reach
# f + n frames produces it - a headless shot, a bench, a windowed walk.
#
# An indirect draw is one API call carrying n command records; its "instances"
# column is the record count, and whether the GPU's culling emptied a record is
# not known here (see R3D_VK_PROF for the Vulkan side's own count).
# ============================================================================
const DS_MESH: int = 0
const DS_INSTANCED: int = 1
const DS_PATCH: int = 2
const DS_RANGE: int = 3
const DS_INDIRECT: int = 4
const DS_KINDS: int = 5
# passes, in the order the frame first opens them
# rows: (pass, program, kind)
@alloc_ok("profiling and statistics, only under R3D_PROF / R3D_DRAWSTATS")
function ds_enabled(render3d_st: mut Render3dState) -> bool {
if render3d_st.ds_on < 0 {
render3d_st.ds_on = 0
if r3d_env_has(render3d_st, "R3D_DRAWSTATS") {
render3d_st.ds_on = 1
render3d_st.ds_frames = Text.to_int(r3d_env(render3d_st, "R3D_DRAWSTATS"))
if render3d_st.ds_frames < 2 { render3d_st.ds_frames = 60 }
if r3d_env_has(render3d_st, "R3D_DRAWSTATS_FROM") { render3d_st.ds_from = Text.to_int(r3d_env(render3d_st, "R3D_DRAWSTATS_FROM")) }
render3d_st.ds_pass_names = new []pointer; render3d_st.ds_pass_prog = new []long; render3d_st.ds_pass_tex = new []long
render3d_st.ds_row_pass = new []int; render3d_st.ds_row_prog = new []int; render3d_st.ds_row_kind = new []int
render3d_st.ds_row_draws = new []long; render3d_st.ds_row_inst = new []long; render3d_st.ds_row_idx = new []long
ds_set_pass(render3d_st, render3d_st.ds_outside)
}
}
return render3d_st.ds_on == 1
}
# the profiler's pass boundaries (prof_begin / prof_end), whether or not R3D_PROF is on
@alloc_ok("profiling and statistics, only under R3D_PROF / R3D_DRAWSTATS")
function ds_set_pass(render3d_st: mut Render3dState, name: pointer) -> void {
if render3d_st.ds_on != 1 { return }
render3d_st.ds_last = -1
var i = 0
while i < len(render3d_st.ds_pass_names) {
if render3d_st.ds_pass_names[i] == name { render3d_st.ds_pass_i = i; return }
i += 1
}
push(render3d_st.ds_pass_names, name); push(render3d_st.ds_pass_prog, 0); push(render3d_st.ds_pass_tex, 0)
render3d_st.ds_pass_i = len(render3d_st.ds_pass_names) - 1
}
# once per render frame, before any of its work (prof_gen_frame)
function ds_frame(render3d_st: mut Render3dState) -> void {
if not ds_enabled(render3d_st) { return }
render3d_st.ds_frame_no += 1
ds_set_pass(render3d_st, render3d_st.ds_outside)
if render3d_st.ds_frame_no == render3d_st.ds_from {
render3d_st.ds_counting = true
print(`r3d: drawstats counting frames {render3d_st.ds_from}..{render3d_st.ds_from + render3d_st.ds_frames - 1}`)
}
if render3d_st.ds_frame_no == render3d_st.ds_from + render3d_st.ds_frames {
render3d_st.ds_counting = false
ds_report(render3d_st)
}
}
@alloc_ok("profiling and statistics, only under R3D_PROF / R3D_DRAWSTATS")
function ds_draw(render3d_st: mut Render3dState, kind: int, instances: int, indices: int) -> void {
if render3d_st.ds_on != 1 or not render3d_st.ds_counting { return }
let p = render3d_st.gpu_prog_cur
var r = render3d_st.ds_last
if r < 0 or render3d_st.ds_row_prog[r] != p or render3d_st.ds_row_kind[r] != kind or render3d_st.ds_row_pass[r] != render3d_st.ds_pass_i {
r = -1
var i = 0
while r < 0 and i < len(render3d_st.ds_row_pass) {
if render3d_st.ds_row_pass[i] == render3d_st.ds_pass_i and render3d_st.ds_row_prog[i] == p and render3d_st.ds_row_kind[i] == kind { r = i }
i += 1
}
if r < 0 {
push(render3d_st.ds_row_pass, render3d_st.ds_pass_i); push(render3d_st.ds_row_prog, p); push(render3d_st.ds_row_kind, kind)
push(render3d_st.ds_row_draws, 0); push(render3d_st.ds_row_inst, 0); push(render3d_st.ds_row_idx, 0)
r = len(render3d_st.ds_row_pass) - 1
}
render3d_st.ds_last = r
}
render3d_st.ds_row_draws[r] = render3d_st.ds_row_draws[r] + 1
render3d_st.ds_row_inst[r] = render3d_st.ds_row_inst[r] + instances
render3d_st.ds_row_idx[r] = render3d_st.ds_row_idx[r] + indices
}
function ds_program_change(render3d_st: mut Render3dState, from: int, to: int) -> void {
if render3d_st.ds_on != 1 or not render3d_st.ds_counting or from == to { return }
render3d_st.ds_pass_prog[render3d_st.ds_pass_i] = render3d_st.ds_pass_prog[render3d_st.ds_pass_i] + 1
}
function ds_tex_bind(render3d_st: mut Render3dState) -> void {
if render3d_st.ds_on != 1 or not render3d_st.ds_counting { return }
render3d_st.ds_pass_tex[render3d_st.ds_pass_i] = render3d_st.ds_pass_tex[render3d_st.ds_pass_i] + 1
}
function ds_kind_name(k: int) -> pointer {
if k == DS_MESH { return "mesh" }
if k == DS_INSTANCED { return "instanced" }
if k == DS_PATCH { return "terrain patch" }
if k == DS_RANGE { return "overlay range" }
return "indirect"
}
# a per-frame mean with one decimal
function ds_per(render3d_st: Render3dState, v: long) -> string {
let x = (v * 10) / render3d_st.ds_frames
return `{string(x / 10)}.{string(x % 10)}`
}
function ds_backend(render3d_st: Render3dState) -> pointer {
if render3d_st.gpu_kind == GPU_VK { return "vulkan" }
return "opengl"
}
@alloc_ok("profiling and statistics, only under R3D_PROF / R3D_DRAWSTATS")
function ds_report(render3d_st: mut Render3dState) -> void {
let n_rows = len(render3d_st.ds_row_pass)
let n_pass = len(render3d_st.ds_pass_names)
var total: long = 0
let by_kind = new []long
for k in 0 .. DS_KINDS { push(by_kind, 0) }
let pass_draws = new []long
let pass_inst = new []long
for s in 0 .. n_pass { push(pass_draws, 0); push(pass_inst, 0) }
for r in 0 .. n_rows {
total = total + render3d_st.ds_row_draws[r]
by_kind[render3d_st.ds_row_kind[r]] = by_kind[render3d_st.ds_row_kind[r]] + render3d_st.ds_row_draws[r]
pass_draws[render3d_st.ds_row_pass[r]] = pass_draws[render3d_st.ds_row_pass[r]] + render3d_st.ds_row_draws[r]
pass_inst[render3d_st.ds_row_pass[r]] = pass_inst[render3d_st.ds_row_pass[r]] + render3d_st.ds_row_inst[r]
}
print("")
print(`draw calls per frame ({ds_backend(render3d_st)}, {gl_width()}x{gl_height()}, mean of {render3d_st.ds_frames} frames from {render3d_st.ds_from}): {ds_per(render3d_st, total)}`)
var kinds = ""
for k in 0 .. DS_KINDS { kinds = kinds + ` {ds_kind_name(k)} {ds_per(render3d_st, by_kind[k])}` }
print(` by kind:{kinds}`)
print("")
print(" pass draws instances program changes texture binds")
for s in 0 .. n_pass {
print(` {Text.pad_right(render3d_st.ds_pass_names[s], 24)} {Text.pad_left(ds_per(render3d_st, pass_draws[s]), 9)} {Text.pad_left(ds_per(render3d_st, pass_inst[s]), 11)} {Text.pad_left(ds_per(render3d_st, render3d_st.ds_pass_prog[s]), 17)} {Text.pad_left(ds_per(render3d_st, render3d_st.ds_pass_tex[s]), 15)}`)
}
# rows by draws, most first
let order = new []int
for r in 0 .. n_rows { push(order, r) }
var a = 1
while a < n_rows {
let v = order[a]
var b = a - 1
while b >= 0 and render3d_st.ds_row_draws[order[b]] < render3d_st.ds_row_draws[v] { order[b + 1] = order[b]; b -= 1 }
order[b + 1] = v
a += 1
}
print("")
print(" the 40 largest rows (pass / kind / program):")
print(" draws instances indices pass kind program")
var shown = 0
while shown < 40 and shown < n_rows {
let r = order[shown]
print(` {Text.pad_left(ds_per(render3d_st, render3d_st.ds_row_draws[r]), 9)} {Text.pad_left(ds_per(render3d_st, render3d_st.ds_row_inst[r]), 11)} {Text.pad_left(ds_per(render3d_st, render3d_st.ds_row_idx[r]), 12)} {Text.pad_right(render3d_st.ds_pass_names[render3d_st.ds_row_pass[r]], 24)} {Text.pad_right(ds_kind_name(render3d_st.ds_row_kind[r]), 14)} {gpu_program_key(render3d_st, render3d_st.ds_row_prog[r])}`)
shown += 1
}
var path = "build/drawstats.csv"
if r3d_env_has(render3d_st, "R3D_DRAWSTATS_CSV") { path = r3d_env(render3d_st, "R3D_DRAWSTATS_CSV") }
let f = file_open(path, "wb")
if f == null { print(`r3d: drawstats could not write {path}`); return }
let head = "backend,width,height,frames,pass,kind,program,draws_per_frame,instances_per_frame,indices_per_frame,pass_program_changes_per_frame,pass_texture_binds_per_frame\n"
file_write(f, head, len(head))
for i in 0 .. n_rows {
let r = order[i]
let s = render3d_st.ds_row_pass[r]
let line = `{ds_backend(render3d_st)},{gl_width()},{gl_height()},{render3d_st.ds_frames},"{render3d_st.ds_pass_names[s]}",{ds_kind_name(render3d_st.ds_row_kind[r])},"{gpu_program_key(render3d_st, render3d_st.ds_row_prog[r])}",{ds_per(render3d_st, render3d_st.ds_row_draws[r])},{ds_per(render3d_st, render3d_st.ds_row_inst[r])},{ds_per(render3d_st, render3d_st.ds_row_idx[r])},{ds_per(render3d_st, render3d_st.ds_pass_prog[s])},{ds_per(render3d_st, render3d_st.ds_pass_tex[s])}\n`
file_write(f, line, len(line))
}
file_close(f)
print(` wrote {path} ({n_rows} rows)`)
}