search

majianjia / nnom

A higher-level Neural Network library for microcontrollers.
Apache License 2.0
815 stars 235 forks source link

Is 0.4.3 NNOM c backend slower than CMSIS-5.9.0? #182

Open bfs18 opened 1 year ago

bfs18 commented 1 year ago

I test the mnist-simple model in rt-thread environment and run model predict for 1000 times. using CMSIS 5.9.0 backend, it consumes 1423 ms using NNOM 0.4.3 C backend, it consumes 1392 ms The document says that replacing c backend with CMSIS would bring 5 times speed-up, however, the performance is similar according to my test. What's the reason? NNOM updated or I didn't set CMSIS macros correctly?

majianjia commented 1 year ago

Probably you havent enable CMSIS's NN acceleration, i.e. you are running CMSIS on their C backend instead of SIMD assemblys. You may check CMSIS instruction or the note under HWC format

bfs18 commented 1 year ago

UPDATE: When -Ofast is added, using CMSIS 5.9.0 backend, reduces to 369 ms.

Hi Jia, thanks for your advice. I add 

ARM_MATH_DSP,ARM_MATH_CM7,__FPU_PRESENT=1

using CMSIS 5.9.0 backend, reduces to 1055 ms. I used the rt-thread qemu-vexpress-a9 bsp. Is this normal?

Probably you havent enable CMSIS's NN acceleration, i.e. you are running CMSIS on their C backend instead of SIMD assemblys. You may check CMSIS instruction or the note under HWC format