gaussian_blur

GitHub
expertruntimeC
results

Measured by wall-clock runtime in seconds — lower is better.

12.0s0.25s
sota
Claude-Opus-4.6
reward 0.710

Usage

Run the reference answer to verify your environment is set up correctly

$ harbor run -p tasks/gaussian_blur

Test model

$ harbor run -p tasks/gaussian_blur \
  -a claude-code -m claude-opus-4-6

Description

Optimize a 17x17 Gaussian blur (sigma=3.0) applied 5 times in sequence to a 4096x4096 grayscale image, where each pass feeds into the next. The image uses one byte per pixel in row-major layout with clamp-to-edge border handling. Output must be within 4 intensity levels of the reference.

Files

path
permission
/app/solve.c✎ Edit
/app/solve.hRead-only
/app/main.cRead-only
/app/MakefileRead-only

Rules

  • 01Edit /app/solve.c only.
  • 02C99 only. C standard library and <immintrin.h> allowed.
  • 03Single-threaded only. Max per-pixel deviation: ±4.

Tags

image-processingAVX2cacheseparable-filter

Model Results

Click a row to view its trajectory in Live Lab

model
reward
score
Claude-Opus-4.6
0.710
Kimi-K2.6
0.640
MiMo-V2.5-Pro
0.540
Gemini-3.1-Pro
0.490
DeepSeek-V4-Pro
0.490
GPT-5.4
0.340
Grok-4-20
0.280
Hunyuan-3-Preview
0.280
GLM-5
0.200
MiniMax-M2.7
0.090
Qwen-3.6-Plus
0.000
<- all tasks