Convolutional Tiling

Choose an image to get started
Input Image
Kernel
Sobel X
Kernel View
→
Feature Map
-
Tiled Convolution with Halo
Input Image
→
Tile + Halo (extracted)
halo halo halo halo
-
Kernel
3x3 view
→
Convolved (pre-crop)
Only green region kept
→
Final Output
-
Core Tile (kept)
Halo (discarded)
Parallel Tiled Convolution
Input Image
→
All Tiles (parallel)
-
→
Final Output
-
Core Tile (kept)
Halo (discarded)
Parallel Processing
Multi-Channel Convolution (Conv4D)
Input (H×W×C)
R
G
B
Tensor Shape
1×3×H×W
⊛
3×3×3 Kernel
R channel
G channel
B channel
Kernel Shape
1×3×3×3
→
Feature Map
-
Output Shape
1×1×H×W
Red Channel
Green Channel
Blue Channel
Channel-Parallel + Tiled Conv4D
Input Chunks (1×1×H×W tiles)
R
G
B
-
⊛
Per-Chunk Kernels
Channel slice applied
to each spatial tile
→
Partial Sums → Reduce
Reduction
Σ channel partials
→
Final Output
-
Parallel Chunks
Halo Region
Core Output