Bad performance on Hi3516dv300

[ztianlin]

Hi~I have tested your pre-trained Bi-Real Net (18-layer) on Huawei Mate30 Pro and I got an inference time of 20ms which was nice. However, when I built dabnn on Hi3516dv300 (armv7, with NEON open) and ran the same model with net_test, I got an inference time of 2000ms, could you please help me with this "bad performace"?

To reproduce this result:
1.Use the corresponding toolchain file to build dabnn. (Here is my CMAKE_CXX_FLAGS)

SET(CMAKE_CXX_FLAGS " -mfloat-abi=softfp -mfpu=neon-vfpv4 -mcpu=cortex-a7 ${CMAKE_CXX_FLAGS}" )

2.Test the generated net_test on Hi3516dv300 and here is what I got:

/mnt/output/tianlin/dabnn_test/bin # ./net_test
Running main() from /home/tianlin/dabnn/third_party/googletest/googletest/src/gtest_main.cc
[==========] Running 4 tests from 1 test suite.
[----------] Global test environment set-up.
[----------] 4 tests from net
[ RUN ] net.bireal18imagenet_comparison
[ OK ] net.bireal18imagenet_comparison (6526 ms)
[ RUN ] net.bireal18imagenet
[INFO] bireal18imagenet's latency is 2219.43 ms
[ OK ] net.bireal18imagenet (2323 ms)
[ RUN ] net.bireal18imagenetstem_comparison
[ OK ] net.bireal18imagenetstem_comparison (6047 ms)
[ RUN ] net.bireal18imagenetstem
[INFO] bireal18imagenetstem's latency is 2045.26 ms
[ OK ] net.bireal18imagenetstem (2232 ms)
[----------] 4 tests from net (17129 ms total)
[----------] Global test environment tear-down
[==========] 4 tests from 1 test suite ran. (17130 ms total)
[ PASSED ] 4 tests.

[daquexian]

thanks for your report! could you please join the qq group for further communication? Tianlin ZHANG <notifications@github.com> 于 2020年6月9日周二 下午9:27写道：

…

Hi~I have tested your pre-trained Bi-Real Net (18-layer) on Huawei Mate30 Pro and I got an inference time of 20ms which was nice. However, when I built dabnn on Hi3516dv300 (armv7, with NEON open) and ran the same model with *net_test*, I got an inference time of 2000ms, could you please help me with this "bad performace"? To reproduce this result: 1.Use the corresponding toolchain file to build dabnn. (Here is my CMAKE_CXX_FLAGS) SET(CMAKE_CXX_FLAGS " -mfloat-abi=softfp -mfpu=neon-vfpv4 -mcpu=cortex-a7 ${CMAKE_CXX_FLAGS}" ) 2.Test the generated *net_test* on Hi3516dv300 and here is what I got: /mnt/output/tianlin/dabnn_test/bin # ./net_test Running main() from /home/tianlin/dabnn/third_party/googletest/googletest/src/gtest_main.cc [==========] Running 4 tests from 1 test suite. [----------] Global test environment set-up. [----------] 4 tests from net [ RUN ] net.bireal18imagenet_comparison [ OK ] net.bireal18imagenet_comparison (6526 ms) [ RUN ] net.bireal18imagenet [INFO] bireal18imagenet's latency is 2219.43 ms [ OK ] net.bireal18imagenet (2323 ms) [ RUN ] net.bireal18imagenetstem_comparison [ OK ] net.bireal18imagenetstem_comparison (6047 ms) [ RUN ] net.bireal18imagenetstem [INFO] bireal18imagenetstem's latency is 2045.26 ms [ OK ] net.bireal18imagenetstem (2232 ms) [----------] 4 tests from net (17129 ms total) [----------] Global test environment tear-down [==========] 4 tests from 1 test suite ran. (17130 ms total) [ PASSED ] 4 tests. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#72>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACYRZH4BPRJCCKCM3QSCUXLRVY2E3ANCNFSM4NZM76GQ> .