feat: add GPU-accelerated LBVH with radix sort and optimize frontend

Replace CPU-based BVH with GPU-accelerated Linear BVH (LBVH) implementation
using radix sort for morton code ordering. This improves collision detection
performance significantly for large meshes.

Additional optimizations:
- Add numba-accelerated mesh generation (grid/rect faces, cylinder)
- Optimize sphere violation checks using squared distance
- Add early exit to triangle-triangle/edge distance calculations
- Parallelize face-to-vertex weight computation
- Standardize platform detection using platform.system()
- Remove obsolete bvh-alloc-factor parameter

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <[email protected]>

Author: ryichando

build-win-native/build.bat

@@ -139,7 +139,7 @@ set EIGEN_DIR=%DEPS%\eigen-3.4.0
 
 REM Source files
 set CPP_SRCS=%CPP_DIR%\simplelog\SimpleLog.cpp %CPP_DIR%\stub.cpp
-set CU_SRCS=%CPP_DIR%\buffer\buffer.cu %CPP_DIR%\main\main.cu %CPP_DIR%\utility\utility.cu %CPP_DIR%\utility\dispatcher.cu %CPP_DIR%\csrmat\csrmat.cu %CPP_DIR%\contact\contact.cu %CPP_DIR%\energy\energy.cu %CPP_DIR%\eigenanalysis\eigenanalysis.cu %CPP_DIR%\barrier\barrier.cu %CPP_DIR%\strainlimiting\strainlimiting.cu %CPP_DIR%\solver\solver.cu %CPP_DIR%\kernels\reduce.cu %CPP_DIR%\kernels\exclusive_scan.cu %CPP_DIR%\kernels\vec_ops.cu
+set CU_SRCS=%CPP_DIR%\buffer\buffer.cu %CPP_DIR%\main\main.cu %CPP_DIR%\utility\utility.cu %CPP_DIR%\utility\dispatcher.cu %CPP_DIR%\csrmat\csrmat.cu %CPP_DIR%\contact\contact.cu %CPP_DIR%\energy\energy.cu %CPP_DIR%\eigenanalysis\eigenanalysis.cu %CPP_DIR%\barrier\barrier.cu %CPP_DIR%\strainlimiting\strainlimiting.cu %CPP_DIR%\solver\solver.cu %CPP_DIR%\kernels\reduce.cu %CPP_DIR%\kernels\exclusive_scan.cu %CPP_DIR%\kernels\vec_ops.cu %CPP_DIR%\kernels\radix_sort.cu %CPP_DIR%\lbvh\lbvh.cu
 
 REM Compiler flags
 set NVCC_FLAGS=-std=c++17 --expt-relaxed-constexpr --extended-lambda -O3 -rdc=true -shared -Wno-deprecated-gpu-targets

build-win-native/clear-cache.bat

@@ -78,6 +78,21 @@ if exist "!JUPYTER_DIR!" (
     echo   [SKIP] Jupyter directory not found
 )
 
+REM Clear export directory in examples (legacy location cleanup)
+set EXPORT_DIR=!PROJ_ROOT!\examples\export
+if exist "!EXPORT_DIR!" (
+    echo Removing export directory in examples...
+    rmdir /s /q "!EXPORT_DIR!"
+    if exist "!EXPORT_DIR!" (
+        echo   [FAIL] Could not remove !EXPORT_DIR!
+        set HAS_ERROR=1
+    ) else (
+        echo   [OK] Removed !EXPORT_DIR!
+    )
+) else (
+    echo   [SKIP] Export directory in examples not found
+)
+
 echo.
 if %HAS_ERROR%==1 (
     echo === [FAIL] Some caches could not be cleared ===

frontend/_app_.py

@@ -5,6 +5,7 @@
 
 import os
 import pickle
+import platform
 import shutil
 
 from typing import Optional
@@ -187,7 +188,7 @@ def get_data_dirpath():
                 with open(branch_file) as f:
                     git_branch = f.read().strip()
                     if git_branch:
-                        if os.name == 'nt':  # Windows
+                        if platform.system() == "Windows":  # Windows
                             return os.path.join(
                                 base_dir, "local", "share", "ppf-cts", f"git-{git_branch}"
                             )
@@ -212,7 +213,7 @@ def get_data_dirpath():
         except (subprocess.CalledProcessError, FileNotFoundError):
             git_branch = "unknown"
 
-        if os.name == 'nt':  # Windows
+        if platform.system() == "Windows":  # Windows
             return os.path.join(
                 base_dir, "local", "share", "ppf-cts", f"git-{git_branch}"
             )
@@ -243,7 +244,7 @@ def __init__(self, name: str, renew: bool, cache_dir: str = ""):
         if cache_dir:
             self._cache_dir = cache_dir
         else:
-            if os.name == 'nt':  # Windows - use project-relative cache
+            if platform.system() == "Windows":  # Windows - use project-relative cache
                 frontend_dir = os.path.dirname(os.path.realpath(__file__))
                 base_dir = os.path.dirname(frontend_dir)
                 self._cache_dir = os.path.join(base_dir, "cache", "ppf-cts")

frontend/_invisible_collider_.py

@@ -42,51 +42,6 @@ def _check_wall_violation_single(
     return signed_dist < 0.0
 
 
-@njit(cache=True)
-def _check_sphere_violation_single(
-    vertex: np.ndarray,
-    sphere_center: np.ndarray,
-    sphere_radius: float,
-    is_inverted: bool,
-    is_hemisphere: bool,
-) -> bool:
-    """Check if a single vertex violates sphere constraint.
-
-    For normal sphere: vertex must be OUTSIDE (distance >= radius)
-    For inverted sphere: vertex must be INSIDE (distance <= radius)
-    For hemisphere (bowl): above center.y, treat as cylinder (only check horizontal distance)
-
-    Args:
-        vertex: Vertex position (3,)
-        sphere_center: Center of sphere (3,)
-        sphere_radius: Radius of sphere
-        is_inverted: If True, collision is with inside of sphere
-        is_hemisphere: If True, top half is open (bowl shape)
-
-    Returns:
-        True if vertex violates the sphere constraint
-    """
-    # For hemisphere, if vertex is above center.y, project center to vertex's y-level
-    # This creates a cylinder-like region above the hemisphere
-    if is_hemisphere and vertex[1] > sphere_center[1]:
-        center = np.array([sphere_center[0], vertex[1], sphere_center[2]])
-    else:
-        center = sphere_center
-
-    diff = vertex - center
-    dist_sq = diff[0] * diff[0] + diff[1] * diff[1] + diff[2] * diff[2]
-    dist = np.sqrt(dist_sq)
-
-    if is_inverted:
-        # For inverted sphere, vertex must be inside (distance <= radius)
-        # Violation if distance > radius (outside the sphere)
-        return dist > sphere_radius
-    else:
-        # For normal sphere, vertex must be outside (distance >= radius)
-        # Violation if distance < radius (inside the sphere)
-        return dist < sphere_radius
-
-
 @njit(parallel=True, cache=True)
 def _check_wall_violations_parallel(
     vertices: np.ndarray,
@@ -108,19 +63,59 @@ def _check_wall_violations_parallel(
         Number of violations found
     """
     n_verts = len(vertices)
+
+    # Check violations in parallel
     for i in prange(n_verts):
         if not is_pinned[i] and _check_wall_violation_single(
             vertices[i], wall_pos, wall_normal
         ):
             violations[i] = True
 
+    # Parallel reduction for counting
     count = 0
-    for i in range(n_verts):
+    for i in prange(n_verts):
         if violations[i]:
             count += 1
     return count
 
 
+@njit(cache=True)
+def _check_sphere_violation(
+    vertex: np.ndarray,
+    sphere_center: np.ndarray,
+    sphere_radius_sq: float,
+    is_inverted: bool,
+    is_hemisphere: bool,
+) -> bool:
+    """Check if a single vertex violates sphere constraint using squared distance.
+
+    This version avoids sqrt for the check, only computing it when needed.
+    """
+    # For hemisphere, if vertex is above center.y, project center to vertex's y-level
+    if is_hemisphere and vertex[1] > sphere_center[1]:
+        cx = sphere_center[0]
+        cy = vertex[1]
+        cz = sphere_center[2]
+    else:
+        cx = sphere_center[0]
+        cy = sphere_center[1]
+        cz = sphere_center[2]
+
+    dx = vertex[0] - cx
+    dy = vertex[1] - cy
+    dz = vertex[2] - cz
+    dist_sq = dx*dx + dy*dy + dz*dz
+
+    if is_inverted:
+        # For inverted sphere, vertex must be inside (distance <= radius)
+        # Violation if distance > radius (outside the sphere)
+        return dist_sq > sphere_radius_sq
+    else:
+        # For normal sphere, vertex must be outside (distance >= radius)
+        # Violation if distance < radius (inside the sphere)
+        return dist_sq < sphere_radius_sq
+
+
 @njit(parallel=True, cache=True)
 def _check_sphere_violations_parallel(
     vertices: np.ndarray,
@@ -146,14 +141,18 @@ def _check_sphere_violations_parallel(
         Number of violations found
     """
     n_verts = len(vertices)
+    sphere_radius_sq = sphere_radius * sphere_radius
+
+    # Check violations in parallel using squared distance
     for i in prange(n_verts):
-        if not is_pinned[i] and _check_sphere_violation_single(
-            vertices[i], sphere_center, sphere_radius, is_inverted, is_hemisphere
+        if not is_pinned[i] and _check_sphere_violation(
+            vertices[i], sphere_center, sphere_radius_sq, is_inverted, is_hemisphere
         ):
             violations[i] = True
 
+    # Parallel reduction for counting
     count = 0
-    for i in range(n_verts):
+    for i in prange(n_verts):
         if violations[i]:
             count += 1
     return count