Optimized version with C++. Runs much faster.

README.md

@@ -194,6 +194,9 @@ out over a few seconds. Press any key to bring it back.
 | Mirror-ball count   | `3` / `4`       | How many reflecting mirror spheres are on screen (0–1000) |
 | Shapes              | `N`             | Toggle random ⇄ cycling shape spawns     |
 | Fullscreen          | `F` or `F11`    | Toggle fullscreen                        |
+| Perf HUD            | `F1`            | FPS / frame ms / body & draw-batch counts |
+| VSync               | `F2`            | Toggle VSync (uncap the frame rate)      |
+| Glow mode           | `F3`            | Post-process bloom vs. legacy per-vector glow |
 | Pause               | `Space`         |                                          |
 | Quit                | `Esc`           |                                          |
 
@@ -241,11 +244,36 @@ out over a few seconds. Press any key to bring it back.
   (`spin = 4^(seed · variance)`), so changing it re-spreads every solid's
   rotation rate instantly — `0` makes them all tumble alike, `100` gives a
   very wide range.
-- **CRT glow** wraps each sharp vector in a soft phosphor mist: many faint
-  additive width layers fade outward into a halo, and a few faint enlarged
-  ghost copies bloom that glow into a larger volume than the hardware line-width
-  cap (~10px) could reach on its own. The crisp core is drawn last so the vector
-  stays sharp; the glow setting scales the mist's spread and brightness.
+- **CRT glow** wraps each sharp vector in a soft phosphor mist. Two
+  implementations, switchable live with `F3`:
+  - **Bloom (default, when framebuffer objects are available):** the field's
+    sharp cores are rendered once, then a dual-filter pyramid (downsample, then
+    tent-filter upsample accumulating each scale into the next finer one) builds
+    a smooth, ever-widening, fading halo around the vectors. A fold factor
+    attenuates the wider scales so the halo fades out and the blacks stay clean.
+    This is **one draw per body instead of nine**, dramatically faster at high
+    body counts.
+  - **Legacy:** each vector is redrawn as many faint additive width layers plus
+    a few enlarged ghost copies. Kept as a fallback for GPUs without FBOs.
+
+- **Performance / retained mode.** Geometry is uploaded to vertex/index buffer
+  objects once and drawn with `glDrawElements` instead of per-vertex immediate
+  mode, so the CPU isn't re-submitting millions of vertices per frame. Rigid
+  shapes use static buffers; morphing 4/5/6-D polytopes stream positions into a
+  shared buffer with a static index buffer; clocks and animated objects (whose
+  topology is rebuilt each frame) stay on the immediate path. GL 1.5/3.0 entry
+  points are resolved at run time via `glfwGetProcAddress`, so the same code
+  works on the Linux, Windows, and screensaver builds; both retained mode and
+  bloom fall back gracefully if unavailable.
+- **Frustum culling.** The field is a full sphere around the camera, so most
+  bodies are behind or beside the view. Six view-frustum planes are extracted
+  from the modelview·projection matrix each frame and each body's bounding
+  sphere is tested before the (relatively costly) projection and draw; the cull
+  margin includes the glow spread so edge halos don't pop. Off-screen bodies
+  still advance and recycle — only their drawing is skipped. At a full field
+  (~7200 bodies) this typically draws under ~1000 of them.
+- The `F1` HUD shows FPS, frame time, and live body / drawn / draw-batch /
+  vertex counts; `F2` toggles VSync; `F3` switches bloom vs. legacy glow.
 - **Magnifying glasses** are real lenses. After the field is drawn, the back
   buffer is grabbed into a texture (`glCopyTexSubImage2D`); each lens body is
   then projected to its screen position and drawn as a disc that samples that