cammymoop/semi-vibed-game
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

GPU compute POC: compute shader → Texture2DRD → canvas shader #84

Closed
claude wants to merge 12 commits from feature/gpu-compute-poc into master
pull from: feature/gpu-compute-poc
merge into: cammymoop:master
Conversation 1 Commits 12 Files changed 7 +689
claude commented 2026-03-05 19:25:36 -06:00
Collaborator
Copy link

Summary

  • Adds a standalone proof-of-concept scene in tools/gpu_poc/ for the GPU compute ball rendering pipeline
  • ball_sim.glsl: compute shader animates 2000 balls with Lissajous motion, writes pos/radius/alive + color into a 2×N RGBA32F texture via imageStore
  • ball_render.gdshader: canvas_item shader reads ball data via texelFetch(ball_data, ivec2(0/1, INSTANCE_ID), 0), overrides VERTEX and COLOR per instance, clips to circle in fragment
  • gpu_poc.gd: full RenderingDevice setup — compute pipeline, Texture2DRD bridge, params SSBO, MultiMesh with identity transforms, per-frame buffer_update + dispatch + barrier(); gracefully no-ops when _rd == null (headless/no GPU)
  • gpu_poc.tscn: minimal standalone Node2D scene to run the POC

Test plan

  • Open tools/gpu_poc/gpu_poc.tscn in the Godot editor and run — should show 2000 animated colored circles moving in Lissajous patterns
  • Bump BALL_COUNT to 10k, 50k, 100k to benchmark rendering throughput vs current GDScript multimesh approach
## Summary - Adds a standalone proof-of-concept scene in `tools/gpu_poc/` for the GPU compute ball rendering pipeline - `ball_sim.glsl`: compute shader animates 2000 balls with Lissajous motion, writes pos/radius/alive + color into a 2×N RGBA32F texture via `imageStore` - `ball_render.gdshader`: canvas_item shader reads ball data via `texelFetch(ball_data, ivec2(0/1, INSTANCE_ID), 0)`, overrides `VERTEX` and `COLOR` per instance, clips to circle in fragment - `gpu_poc.gd`: full RenderingDevice setup — compute pipeline, `Texture2DRD` bridge, params SSBO, MultiMesh with identity transforms, per-frame `buffer_update` + dispatch + `barrier()`; gracefully no-ops when `_rd == null` (headless/no GPU) - `gpu_poc.tscn`: minimal standalone Node2D scene to run the POC ## Test plan - [ ] Open `tools/gpu_poc/gpu_poc.tscn` in the Godot editor and run — should show 2000 animated colored circles moving in Lissajous patterns - [ ] Bump `BALL_COUNT` to 10k, 50k, 100k to benchmark rendering throughput vs current GDScript multimesh approach
Proof-of-concept for rendering large ball counts via GPU pipeline:
- ball_sim.glsl: compute shader animates 2000 balls with Lissajous
  motion, writes pos/radius/alive and color into a 2×N RGBA32F texture
  via imageStore
- ball_render.gdshader: canvas_item shader reads ball_data texture via
  texelFetch(ball_data, ivec2(0/1, INSTANCE_ID), 0), overrides VERTEX
  position and COLOR in the vertex stage, clips to circle in fragment
- gpu_poc.gd: RenderingDevice setup (compute pipeline, Texture2DRD
  bridge, params SSBO, MultiMesh with identity transforms), per-frame
  buffer_update + compute dispatch + barrier(); gracefully skips setup
  when _rd == null (headless)
- gpu_poc.tscn: minimal standalone scene

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Remove #[compute] pragma from ball_sim.glsl — only valid when loaded
  via Godot's import system (RDShaderFile), not raw FileAccess text reads
- Replace early return in ball_render.gdshader vertex() with if/else —
  canvas_item vertex functions do not support return statements

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- canvas_item shaders write COLOR directly; ALBEDO/ALPHA are spatial-only
- _rd.barrier() is deprecated in Godot 4.6, RenderingDevice adds barriers
  automatically between compute and rendering passes

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the Lissajous animation POC with a real simulation:

Compute shader (ball_sim.glsl):
- Reads/writes a flat ball state SSBO (pos, vel, radius, alive per ball)
- Each frame: checks explosion events, integrates velocity, bounces off
  arena walls, writes updated state back + render texture
- Colors balls by velocity direction so motion is visible

GDScript (gpu_poc.gd):
- Initializes BALL_COUNT (10000) balls at random positions/velocities
- Ball state SSBO, explosion events SSBO (max 16 per frame), params SSBO
- Timer fires a random explosion every 1.5s; events are uploaded and
  cleared each frame so the SSBO slot is reused
- BALL_COUNT const at the top to tune for benchmarking

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace 2×BALL_COUNT texture (exceeded device max ~8192) with a
TEX_ROW_SIZE×H texture where H = 2*ceil(BALL_COUNT/TEX_ROW_SIZE):

  ball idx → col = idx % 4096,  block = idx / 4096
  row block*2+0 : pos.x, pos.y, radius, alive
  row block*2+1 : color rgba

At 10k balls: 4096×6. At 500k balls: 4096×246.
Updated imageStore in ball_sim.glsl, texelFetch in ball_render.gdshader,
and texture_create dimensions in gpu_poc.gd.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Compute shader (ball_sim.glsl):
- Two new SSBOs: RockData (center.x, center.y, radius per rock) and
  RockParams (count)
- Rock bounce: push-out overlap correction + velocity reflection off
  surface normal; only applies when dot(v, normal) < 0 to avoid sticking
- Color by ball type (inferred from radius match to 4 type radii) using
  the game's actual healthy_color values: white / light-blue / purple /
  salmon for standard / small / big / chain

GDScript (gpu_poc.gd):
- BALL_TYPES const with radius, speed range, and spawn weight per type;
  weighted-random assignment at init
- 8 rocks placed in a ring, sizes cycling 1x/1.75x/2.5x base radius 64
- Rock SSBOs initialized once and never updated (static obstacles)
- _draw_rocks(): fan-triangulated MeshInstance2D circles at rock positions
  so they're visible; z_index=-1 so balls render on top

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
CPU arms random balls; GPU tracks fuse and reports explosions back:

Compute shader (ball_sim.glsl):
- Ball state extended to 7 floats (adds fuse_time at index 6)
- New DetonationOut SSBO at binding 6: int count (atomic) + vec4 events[]
- When fuse_time expires: atomicAdd claims a slot, writes blast center +
  radius (br*2+40), marks ball dead
- Armed balls (fuse_time > 0 at frame start) render red

GDScript (gpu_poc.gd):
- BALL_STRIDE 6 → 7; ball init sets fuse_time = 0
- _arm_random_ball(): patches only the 4-byte fuse_time field of a random
  ball via buffer_update at the exact SSBO byte offset — no full re-upload
- _process(): reads DetonationOut via buffer_get_data (GPU stall / readback
  cost measurement), parses count + events, converts to _pending_explosions
  for the next frame's explosion SSBO upload, then resets count to 0 before
  the next dispatch
- Explosion events SSBO is now driven entirely by detonation readback
  (no more CPU-side random position explosions)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add use_custom_data=true to the MultiMesh. Each instance stores its ball
index in INSTANCE_CUSTOM.r so the vertex shader can look up the correct
texture row independently of instance slot order.

CPU maintains two parallel PackedInt32Arrays (_instance_to_ball,
_ball_to_instance). When a ball is armed, it is swapped into the last
available normal slot (_armed_start-1 and decrements _armed_start).
Armed instances occupy [_armed_start, BALL_COUNT), which renders on top
because MultiMesh uses painter's order (later = in front).

_arm_random_ball now picks ball_idx from _instance_to_ball[randi()%_armed_start]
so it always picks from the normal (non-armed) region, then does the swap
with 2 set_instance_custom_data calls and updates both index arrays.

Vertex shader: replaces INSTANCE_ID lookup with int(round(INSTANCE_CUSTOM.r)).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fuse countdown disabled — armed balls accumulate indefinitely.
Each armed ball checks against the CPU-uploaded armed list for overlap;
if any two armed balls touch, both detonate via atomicAdd.

Compute shader (ball_sim.glsl):
- New ArmedList SSBO at binding 7: float[0]=count, float[1..N]=ball indices
- After physics, each armed ball iterates the armed list, skips self and
  dead balls, checks circle-circle overlap; on hit calls detonate() which
  atomicAdds a det slot and writes (px,py,blast_r,ball_idx) as vec4
- detonate() extracted to a helper function for clarity
- Fuse countdown block present but commented out in spirit (replaced by
  intersection check as the sole detonation trigger)

GDScript (gpu_poc.gd):
- MAX_ARMED = 256; _armed_list_buffer created at startup
- _upload_armed_list(): builds [count, idx0..N] from _instance_to_ball
  [_armed_start..BALL_COUNT) and uploads to SSBO each frame (tiny buffer)
- _arm_random_ball(): sets fuse_time=1.0 (armed marker, no countdown)
- det event .w now carries ball_idx; _remove_from_armed_region() swaps
  the detonated ball back into the normal region and increments _armed_start
- Armed list buffer freed in _exit_tree()

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
cammymoop commented 2026-03-06 10:28:46 -06:00
Owner
Copy link

Successful test

Successful test
cammymoop closed this pull request 2026-03-06 10:28:46 -06:00

Pull request closed

Please reopen this pull request to perform a merge.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
cammymoop/semi-vibed-game!84
No description provided.