Skip to content

performance.md

Latest commit

 

History

History
106 lines (77 loc) · 3.06 KB

performance.md

File metadata and controls

106 lines (77 loc) · 3.06 KB
Raw

Performance

← Back to README

Key Metrics

Metric Value
View distance 64 chunks (1,024 blocks)
Max loaded chunks 100,000
Max rendered chunks/frame 4,500
Chunk load throttle 32/frame
Mesh generation throttle 24/frame
GPU terrain tasks in-flight 64
Vertex format size 16 bytes (packed)
Vertex mega-buffer 128 MB
Index mega-buffer 64 MB
G-Buffer bandwidth saved ~58 MB/frame (1080p)
GPU-simulated particles 8,192
Frames in-flight 3 (triple buffering)
Block types 27 (with PBR properties)
Biomes 6 (smooth blending)
Dynamic point lights 16 (animated PBR)
Bloom passes 4 mip levels (Kawase)

FPS Tracking

FPS is measured every 0.5 seconds with exponential smoothing (90/10 blend):

smoothFPS = smoothFPS * 0.9 + fps * 0.1

Stats are written to /tmp/voxel_stats.txt every 30 frames:

FPS: 58.2 | Chunks: 2847/4500 | Tris: 1200K | Verts: 890K | GPUwait: 2.1ms | Pos: 123.5,45.0,678.9

GPU Wait Time

The renderer tracks how long the CPU blocks waiting for the GPU:

frameSemaphore.wait()  // Block if 3 frames already in-flight
let waitMs = (CACurrentMediaTime() - waitStart) * 1000.0

Low GPU wait (< 2ms) = GPU-bound. High GPU wait (> 5ms) = CPU has headroom.

Bandwidth Analysis

Without tile memory optimization (traditional IMR):

  • G-Buffer write: 3 textures × 1080p × format size = ~29 MB
  • G-Buffer read: same = ~29 MB
  • Total: ~58 MB/frame just for G-Buffer

With tile memory (this engine):

  • G-Buffer: 0 MB DRAM bandwidth (stays in SRAM)
  • Only HDR output written to DRAM: ~8 MB/frame

Savings: ~50 MB/frame at 1080p, scales with resolution.

Chunk Loading Strategy

Chunks load in a ring pattern from camera center outward:

Ring 0 (closest):  immediate load, LOD 1 (full res)
Ring 1:            queued, LOD 1
Ring 2-10:         LOD 1
Ring 10-24:        LOD 2 (half res)
Ring 24-48:        LOD 4 (quarter res)
Ring 48-64:        LOD 8 (eighth res)

Load queue is rebuilt only when camera enters a new chunk — not every frame.

Frustum Culling Optimization

Culling is not recomputed every frame. It skips if:

  • Camera moved less than 4 blocks
  • Camera rotated less than ~3.6° (forwardDot < 0.998)

Lazy re-cull every 30 frames or when new meshes appear.

Draw Call Reduction

Approach Draw calls/frame
Naive (one per chunk) 4,500
Mega-buffer (one bind) 4,500 (but faster)
ICB (indirect commands) 1

The Indirect Command Buffer encodes all 4,500 drawIndexedPrimitives commands and executes them with a single executeCommandsInBuffer. Amortized cost: ~0.07ms (re-encoded only when chunk list changes).

Greedy Meshing Impact

For a typical chunk with mixed terrain:

Method Vertices Triangles
Naive (6 faces × 4 verts) ~24,000 ~12,000
Greedy meshing ~3,000–6,000 ~1,500–3,000
Reduction ~75–87% ~75–87%

Packed vertex format saves an additional 67% memory (16 bytes vs 48 bytes naive).