dm-ray-march-lod-optimization-patterns

---

name: DM Ray March LOD & Optimization Patterns description: Reference for renderer optimizations: screen-space adaptive cell size, dirty-flag incremental baking, camera sensor, and fragment shader stepping. Use when working on dm_scene_ray_march_renderer.py performance.

DM Ray March LOD & Optimization Patterns

---

Architecture Overview

_rebuild() is triggered by:
  - register_field / update_field / unregister_field  (data change)
  - set_field_visible / gc_fields / on_prefs_changed  (visibility/prefs change)
  - _on_zoom_settled()                                (camera zoom change — NEW)

Each _rebuild():
  1. For each dirty field: call bake_sdf_to_volume(field, _compute_cell_size(field, view))
  2. Cache result in _baked_cache[label]
  3. Stack cached volumes into combined atlas (np.zeros)
  4. If atlas dims changed → glTexImage3D (full upload)
     Else → glTexImage3D (still full, same dims, safe fallback)
  5. Update per-field uniforms + bbox proxy coords

---

Screen-Space Adaptive Cell Size

Goal: When a field is small on screen (zoomed out), use coarser cells (faster bake). When large on screen (zoomed in), use finer cells (sharper surface).

Formula (with sqrt power-law falloff):

px_per_world    = vp_height_px / cam_world_height
field_screen_px = field_world_size * px_per_world
linear_cell     = field_world_size / max(field_screen_px / n_texels, 1.0)
linear_ratio    = max(linear_cell / base_cell, 1.0)   # >= 1 (never finer than base)

LOD_EXPONENT = 0.5   # sqrt — change to 1.0 for linear, 0.0 to disable LOD entirely
MAX_RATIO    = 16.0  # hard cap: worst-case = base_cell * sqrt(16) = 4× base
ratio        = min(linear_ratio, MAX_RATIO) ** LOD_EXPONENT
cell_size    = base_cell * ratio

Key insights:

• RayMarchCellSize is a quality floor — adaptive never goes finer than it.
• Linear falloff (exponent=1.0) is too aggressive: 4× zoom-out → 4× cell size → visibly flat shading. Sqrt (0.5) means 4× zoom-out → only 2× cell size.
• MAX_RATIO=16 caps at base_cell * sqrt(16) = 4× regardless of zoom level.

Zoom-out behaviour (base_cell = 2mm):

Zoom-out	Linear (bad)	Sqrt (good)
2×	4mm	2.8mm
4×	8mm	4mm
16×+	20mm	8mm (cap)

Orthographic camera: cam_world_height = cam.height.getValue() Perspective camera: cam_world_height = 2 * dist * tan(cam.heightAngle / 2)

Where dist = cam.position.getValue().length() (distance from origin, or better: distance from camera to field center — but origin approximation is fine).

Getting viewport height:

vp_h = 800.0  # fallback
try:
    viewer = view.getViewer()
    for method in ("getGlxSize", "getSize"):
        if hasattr(viewer, method):
            sz = getattr(viewer, method)()
            vp_h = float(sz[1] if isinstance(sz, (list, tuple)) else sz.height())
            break
except Exception:
    pass

This pattern is used throughout the codebase (input_manager.py, dm_renderer.py).

---

Camera Sensor Pattern

Goal: Trigger _rebuild() when the camera changes. With frustum-clipped baking (RO-014), pan and rotate also change the visible region, so use SoNodeSensor on the camera node (fires on any field change) rather than SoFieldSensor on cam.height only.

Coin3D SoNodeSensor (preferred — covers zoom + pan + rotate):

from pivy.coin import SoNodeSensor

cam = view.getCameraNode()
sensor = SoNodeSensor(callback_fn, userdata)
sensor.attach(cam)   # fires on any camera field change
# To detach: sensor.detach()

SoFieldSensor (zoom-only — only if frustum clipping is NOT implemented):

from pivy.coin import SoFieldSensor, SoOrthographicCamera

cam = view.getCameraNode()
is_ortho = cam.isOfType(SoOrthographicCamera.getClassTypeId())
field = cam.height if is_ortho else cam.position

sensor = SoFieldSensor(callback_fn, userdata)
sensor.attach(field)

Callback signature:

def _on_camera_changed(self, userdata, sensor):
    # Called synchronously during Coin3D traversal — do NOT call _rebuild() here
    # Instead debounce with QTimer
    ...

Debounce pattern (100ms):

from PySide import QtCore

def _on_camera_changed(self, userdata, sensor):
    if self._zoom_timer is not None:
        self._zoom_timer.stop()
    self._zoom_timer = QtCore.QTimer()
    self._zoom_timer.setSingleShot(True)
    self._zoom_timer.timeout.connect(self._on_zoom_settled)
    self._zoom_timer.start(100)

def _on_zoom_settled(self):
    self._zoom_timer = None
    if self._fields:
        self._dirty_fields = set(self._fields.keys())  # all fields need new cell size
        self._rebuild()

Cleanup (in _detach()):

if self._cam_sensor is not None:
    self._cam_sensor.detach()
    self._cam_sensor = None
if self._zoom_timer is not None:
    self._zoom_timer.stop()
    self._zoom_timer = None

---

Dirty Flag Pattern

Goal: Only rebake the field(s) that actually changed.

# State added to __init__:
self._baked_cache  = {}    # label → baked dict from sdf_baker
self._dirty_fields = set() # labels that need rebaking

# Mark only the changed field dirty:
def update_field(self, label, field):
    self._fields[label] = (field, visible)
    self._dirty_fields.add(label)   # ← single field
    self._rebuild()

# Mark all dirty (zoom change, prefs change):
self._dirty_fields = set(self._fields.keys())

# In _rebuild(): only bake if dirty or not cached
for label, f in visible:
    if label in self._dirty_fields or label not in self._baked_cache:
        self._baked_cache[label] = bake_sdf_to_volume(f, cell_size)
self._dirty_fields.clear()

---

Frustum-Clipped Bake Region

Core insight: Adaptive cell size (LOD) still bakes the entire field. Zooming into a 20mm corner of a 500mm sphere still bakes 500mm³ — a ~15,000× waste. Clip the bake region to the camera frustum AABB so only the visible subregion is evaluated.

Order of operations in _rebuild():

1. Compute frustum = _get_view_frustum_aabb(view)
2. For each field: clip = intersect(field.bounding_box(), frustum)
3. If clip is empty → field off-screen, skip baking entirely
4. Compute cell size using the clip region size (not full field size)
5. Call bake_sdf_to_volume(field, cell_size, bbox_override=clip)

World-space frustum AABB from camera parameters:

import numpy as np, math

rot = cam.orientation.getValue()
right   = rot.multVec(coin.SbVec3f(1,  0,  0))
up      = rot.multVec(coin.SbVec3f(0,  1,  0))
forward = rot.multVec(coin.SbVec3f(0,  0, -1))   # -Z is forward in OpenGL

r = np.array([right[0],   right[1],   right[2]])
u = np.array([up[0],      up[1],      up[2]])
f = np.array([forward[0], forward[1], forward[2]])
p = np.array([cam.position.getValue()[0], ...])

near, far = cam.nearDistance.getValue(), cam.farDistance.getValue()

if is_ortho:
    hh = cam.height.getValue() / 2.0
    hw = hh * aspect
    corners = np.array([
        p + f*d + r*sx*hw + u*sy*hh
        for sx in (-1,1) for sy in (-1,1) for d in (near, far)
    ])
else:
    for d in (near, far):
        hh = math.tan(cam.heightAngle.getValue()/2) * d
        hw = hh * aspect
        corners += [p + f*d + r*sx*hw + u*sy*hh for sx in (-1,1) for sy in (-1,1)]
    corners = np.array(corners)

aabb_min = FreeCAD.Vector(*corners.min(axis=0))
aabb_max = FreeCAD.Vector(*corners.max(axis=0))

AABB intersection:

clip_min = FreeCAD.Vector(max(f_min.x, fr_min.x), max(f_min.y, fr_min.y), max(f_min.z, fr_min.z))
clip_max = FreeCAD.Vector(min(f_max.x, fr_max.x), min(f_max.y, fr_max.y), min(f_max.z, fr_max.z))
visible = clip_min.x < clip_max.x and clip_min.y < clip_max.y and clip_min.z < clip_max.z

sdf_baker.py change — add bbox_override param:

def bake_sdf_to_volume(field, cell_size, bbox_override=None):
    if bbox_override is not None:
        mn_raw, mx_raw = bbox_override
    else:
        mn_raw, mx_raw = field.bounding_box()
    # rest unchanged

Shader u_bbox_min/max uniforms — always set to the baked region (already per-field), so the shader correctly clips ray-marching to the visible subregion. No shader changes needed.

Impact table (500mm sphere, 2mm base_cell, viewing 20mm region):

Without frustum clip	With frustum clip
Bake region: 500mm³	Bake region: 20mm³
Grid: 250³ = 15.6M cells	Grid: 10³ = 1000 cells
Bake cost: ~seconds	Bake cost: ~milliseconds

---

IsSubtractive Label Bug

Bug: _rebuild() does doc.getObject("DocName.ObjName") — this always returns None because getObject() expects just the object name, not the full label.

Fix:

def _get_is_subtractive(self, label):
    """Parse 'DocName.ObjName' label and return IsSubtractive."""
    try:
        parts = label.split(".", 1)
        if len(parts) != 2:
            return False
        doc = FreeCAD.getDocument(parts[0])
        obj = doc.getObject(parts[1]) if doc else None
        return bool(getattr(obj, "IsSubtractive", False))
    except Exception:
        return False

---

Fragment Shader Stepping Fix

Problem: min_step = cell * 0.05 forces tiny steps near the surface and for very small SDF values. This wastes iterations.

Fix:

// Old:
float min_step = cell * 0.05;
t += max(min_d, min_step);

// New (no min_step global):
float hit_thresh = cell * 0.1;  // already computed per field
t += max(min_d * 0.9, hit_thresh * 0.5);

0.9 * min_d understeeps slightly to avoid overshooting thin features. hit_thresh * 0.5 is the surface-relative floor — only kicks in when min_d is extremely small (right at the surface), and is proportional to cell size.

Iteration count: Increase from 256 to 512. At 2mm cells with a 200mm box (diagonal ~346mm), 256 * minimum step can be insufficient.

---

glTexSubImage3D Signature (for gl_texture3d.py)

_gl.glTexSubImage3D.argtypes = [
    ctypes.c_uint,   # target  (GL_TEXTURE_3D)
    ctypes.c_int,    # level   (0)
    ctypes.c_int,    # xoffset (0)
    ctypes.c_int,    # yoffset (0)
    ctypes.c_int,    # zoffset (field's Z start in atlas)
    ctypes.c_int,    # width   (atlas width)
    ctypes.c_int,    # height  (atlas height)
    ctypes.c_int,    # depth   (field's Z depth)
    ctypes.c_uint,   # format  (GL_RGBA)
    ctypes.c_uint,   # type    (GL_UNSIGNED_BYTE)
    ctypes.c_void_p, # pixels
]
_gl.glTexSubImage3D.restype = None

Usage rule: Only call glTexSubImage3D if the texture already exists (tex_id != 0) AND the total atlas dimensions have not changed. If dimensions changed, call the full glTexImage3D first, then the sub-update is redundant.

---

Files to Read Before Editing

1. core/dm_scene_ray_march_renderer.py:486-578 — _rebuild() full method
2. core/dm_scene_ray_march_renderer.py:396-440 — public API (register/update/unregister)
3. core/dm_scene_ray_march_renderer.py:52-73 — _attach() / _detach()
4. core/gl_texture3d.py:80-157 — GLTexture3D class (full)
5. core/dm_object.py:115-120 — get_ray_march_cell_size() pattern to follow for new setting
6. commands/cmd_settings.py:33-65 — settings dialog init pattern

---

Invariants — Do NOT Break

• _rebuild() must always update ALL MAX_FIELDS uniform slots (set u_nx[fi] = 0 for unused slots beyond n_fields) — the shader reads up to MAX_FIELDS.
• _baked_cache must be purged in unregister_field() — stale cache causes ghost objects.
• _cam_sensor.detach() must be called in _detach() — leaking a SoFieldSensor after the view is closed causes a crash.
• Camera sensor callback must NOT call _rebuild() directly — use QTimer debounce. Direct calls in Coin3D traversal callbacks can cause re-entrancy crashes.