embeded/i.mx 8m plus2023. 12. 21. 15:51

i.MX Machine Learning User's Guide에 있는거 테스트 해봄

The behavior is as follows:
• If USE_GPU_INFERENCE=1, the graph is executed on the GPU
• Otherwise, the graph is executed on the NPU (if available)
By default, the NPU is used for OpenVX graph execution.

 

전체 실행시간으로는 neon 이 가장 빨랐고, inference 시간으로는 NPU가 가장 빨랐다.

의외로 GPU 가속이 NEON만도 못하다는 충격적인(!) 결과가..

root@imx8mpevk:/usr/bin/tensorflow-lite-2.12.1/examples# time ./label_image 
INFO: Loaded model ./mobilenet_v1_1.0_224_quant.tflite
INFO: resolved reporter
INFO: invoked
INFO: average time: 41.16 ms
INFO: 0.764706: 653 military uniform
INFO: 0.121569: 907 Windsor tie
INFO: 0.0156863: 458 bow tie
INFO: 0.0117647: 466 bulletproof vest
INFO: 0.00784314: 835 suit

real    0m0.173s
user    0m0.520s
sys     0m0.024s
root@imx8mpevk:/usr/bin/tensorflow-lite-2.12.1/examples# time USE_GPU_INFERENCE=0 ./label_image --external_delegate_path=/usr/lib/libvx_delegate.so
INFO: Loaded model ./mobilenet_v1_1.0_224_quant.tflite
INFO: resolved reporter
Vx delegate: allowed_cache_mode set to 0.
Vx delegate: device num set to 0.
Vx delegate: allowed_builtin_code set to 0.
Vx delegate: error_during_init set to 0.
Vx delegate: error_during_prepare set to 0.
Vx delegate: error_during_invoke set to 0.
EXTERNAL delegate created.
INFO: Applied EXTERNAL delegate.
W [HandleLayoutInfer:291]Op 162: default layout inference pass.
INFO: invoked
INFO: average time: 2.861 ms
INFO: 0.768627: 653 military uniform
INFO: 0.105882: 907 Windsor tie
INFO: 0.0196078: 458 bow tie
INFO: 0.0117647: 466 bulletproof vest
INFO: 0.00784314: 835 suit

real    0m3.116s
user    0m2.916s
sys     0m0.195s
root@imx8mpevk:/usr/bin/tensorflow-lite-2.12.1/examples# time USE_GPU_INFERENCE=1 ./label_image --external_delegate_path=/usr/lib/libvx_delegate.so
INFO: Loaded model ./mobilenet_v1_1.0_224_quant.tflite
INFO: resolved reporter
Vx delegate: allowed_cache_mode set to 0.
Vx delegate: device num set to 0.
Vx delegate: allowed_builtin_code set to 0.
Vx delegate: error_during_init set to 0.
Vx delegate: error_during_prepare set to 0.
Vx delegate: error_during_invoke set to 0.
EXTERNAL delegate created.
INFO: Applied EXTERNAL delegate.
W [query_hardware_caps:89]Unsupported evis version
W [HandleLayoutInfer:291]Op 162: default layout inference pass.
INFO: invoked
INFO: average time: 171.261 ms
INFO: 0.784314: 653 military uniform
INFO: 0.105882: 907 Windsor tie
INFO: 0.0156863: 458 bow tie
INFO: 0.00784314: 466 bulletproof vest
INFO: 0.00392157: 835 suit

real    0m1.992s
user    0m1.377s
sys     0m0.103s

'embeded > i.mx 8m plus' 카테고리의 다른 글

sounddevice on arm i.mx8 evk  (0) 2024.05.14
missed: not vectorized: relevant stmt not supported:  (0) 2023.08.31
nxp i.mx8mp win iot part 2  (0) 2023.05.26
nxp i.mx8mp win iot  (0) 2023.05.26
rpmsg-lite  (0) 2023.03.23
Posted by 구차니