Selecting Mobile Gpu Convolutions

Decide between standard and separable (depthwise+pointwise) convolutions for mobile-GPU (OpenCL / TFLite delegate) deployment using three measured heuristics — first-frame time tracks OpenCL kernel count (not FLOPs), warmup gain tracks arithmetic intensity, steady-state time tracks FLOPs ÷ GPU utilization. Use at model-design time to narrow conv choices and on-device to confirm. Read-only advisory — heuristics from one anonymized field benchmark; verify on the target runtime. Reversal conditions (large channels / kernel fusion) apply.

vemodalen-x Updated 0 repo stars
VEMO_SKILLS skill lifecycle flow

File contents

vemodalen-x/VEMO_SKILLS commit 34b3bac5b2

Frequently asked questions

npx get-skill add vemodalen-x/selecting-mobile-gpu-convolutions