espressif / esp-nn

Optimised Neural Network functions for Espressif chipsets
Apache License 2.0
146 stars 24 forks source link

Is there perfomance gain for plain ESP32? #1

Closed AIWintermuteAI closed 2 years ago

AIWintermuteAI commented 2 years ago

Very exciting development! I see benchmark numbers comparing ANSI C vs Assembly optimized for ESP32-S3. Have you tried comparison benchmarks for ESP-NN ANSI C vs vanilla tflite-micro ops? Is there any performance gain in that case for ESP32 (plain, not S3 version)?

vikramdattu commented 2 years ago

Hi @AIWintermuteAI

Very exciting development! I see benchmark numbers comparing ANSI C vs Assembly optimized for ESP32-S3. Have you tried comparison benchmarks for ESP-NN ANSI C vs vanilla tflite-micro ops? Is there any performance gain in that case for ESP32 (plain, not S3 version)?

Thanks for the interest! The gain is about 2x when ANSI C versions of ESP-NN are compared with tflite-ops with no changes. This flag when commented out, uses vanilla tflite-micro kernels. When enabled, it uses ANSI C versions OR assembly optimized ESP32-S3.

AIWintermuteAI commented 2 years ago

Thanks for lightning speed reply! Excited to use ESP-NN then. I'll close the issue.