Wrong detection results for yolov7 on RISC-V with RVV

[vassilijnadarajah]

error log | 日志或报错信息 | ログ

When peforming object detection with yolov7 and NCNN on RISC-V, I get different results depending on if i use NCNN with RVV support or without RVV support. I use the same source code but the NCNN library is compiled differently. The results are the same, when perfoming inference on x86, armv8 or RISC-V without RVV. Just when compiling with RVV, I get wrong results.

Classification with NCNN on RISC-V with RVV

Classification with NCNN on RISC-V without RVV

context | 编译/运行环境 | バックグラウンド

how to reproduce | 复现步骤 | 再現方法

I uploaded an example project with instructions on how to compile with and without RVV support. For the compilation of the NCNN library i followed the steps in the wiki.

more | 其他 | その他

Please let me know, if I need to supply more information. I would really appreciate any help with this. Thank you in advance! 谢谢!

[wzyforgit]

Please run ncnn unit test with RVV to see what goes wrong.

Run ncnn unit test:
cmake -DNCNN_BUILD_TESTS=ON (other parameters) ..
make
make test

[vassilijnadarajah]

Hi @wzyforgit! Thank you for taking the time to reply. I just ran all the unit tests and apart from one test, all of them have passed.

When running test_reduction, I get the error message param is too old, please regenerate! multiple times. I attached a screenshot.

Also for test_squeezenet I get overwrite built-in layer type Convolution and overwrite built-in layer type 6 printed to the console. The test still passes though.

I'd be thankful for any other tips you may be able to give me.

[wzyforgit]

Hi @wzyforgit! Thank you for taking the time to reply. I just ran all the unit tests and apart from one test, all of them have passed.

When running test_reduction, I get the error message param is too old, please regenerate! multiple times. I attached a screenshot.

Also for test_squeezenet I get overwrite built-in layer type Convolution and overwrite built-in layer type 6 printed to the console. The test still passes though.

I'd be thankful for any other tips you may be able to give me.

There is another way here:
0. Keep use RVV param in cmake

1. Remove all RVV files in src/layer/riscv
2. Recover one RVV file in src/layer/riscv
3. Build ncnn and run your yolov7 application
   4.1 If the detection results are wrong, the file you recovered last time have bug, remove it
   4.2 If the detection results are all right, the file you recovered last time is right, keep it
4. Repeat steps 2 to 4 until you find all the problematic files
5. Remove all problematic files and rebuild ncnn into your application

I'm sorry that I'm not good at RVV and C906 core.
If your project is urgent, you can try this method to get maximum speed up and wait for the owner to fix the problems.

[nihui]

param is too old, please regenerate! warning fixed in #5397
I am now trying to reproduce the rvv issue

[nihui]

The bug has been confirmed and can be reproduced

[nihui]

fixed in #5398
Thanks for your report

[vassilijnadarajah]

Thank you so much @nihui! It works now :)