NPU SDK 6.4.4.3 release

Announcements
24

@1228760706 你可以参考app仓库里面，现在有一个inception的文件夹，你可以参考这个，哪个是直接转换case代码的

25

你的意思是转换工具生成的makefile.linux这个没有什么意义？需要修改之后才可以编译？

26

@1228760706 我们也是在这个基础上适配的，不是没有意义，但是你如果不想手动修改，那你就直接用我们inception文件的编译文件，你现在替换完可以编译了么

27

直接替换之后 报这个错误
fatal error: vnn_global.h: No such file or directory
15 | #include “vnn_global.h”
| ^~~~~~~~~~~~~~
compilation terminated.

28

@1228760706

yan@yan:~/data/git/npu/aml_npu_app/detect_library/inception$ ls
BUILD        inceptionv3.vcxproj  makefile.linux      makefile.target_name  vnn_global.h       vnn_inceptionv3.h   vnn_post_process.h  vnn_pre_process.h
build_vx.sh  main.c               makefile.linux.def  nbg_meta.json         vnn_inceptionv3.c  vnn_post_process.c  vnn_pre_process.c

这里面有一个vnn_global.h，一并拷贝过去

29

jpeglib.h这个文件找不着
vnn_pre_process.c:14:10: fatal error: jpeglib.h: No such file or directory
14 | #include “jpeglib.h”
| ^~~~~~~~~~~
compilation terminated

30

@1228760706

sudo apt install libjpeg9-dev
31

大佬 你们放出来的inception可以编译么？我跑你们的示例报这个错误
/usr/bin/ld: cannot find -lnn_detect
collect2: error: ld returned 1 exit status

32

@1228760706

khadas@Khadas:~/aml_npu_app/detect_library/inception$ ./build_vx.sh 
  COMPILE /home/khadas/aml_npu_app/detect_library/inception/vnn_pre_process.c
  COMPILE /home/khadas/aml_npu_app/detect_library/inception/vnn_inceptionv3.c
vnn_inceptionv3.c: In function ‘vnn_CreateInceptionV3’:
vnn_inceptionv3.c:145:29: warning: unused variable ‘data’ [-Wunused-variable]
  145 |     uint8_t *               data;
      |                             ^~~~
At top level:
vnn_inceptionv3.c:94:17: warning: ‘load_data’ defined but not used [-Wunused-function]
   94 | static uint8_t* load_data
      |                 ^~~~~~~~~
  COMPILE /home/khadas/aml_npu_app/detect_library/inception/main.c
  COMPILE /home/khadas/aml_npu_app/detect_library/inception/vnn_post_process.c
gmake: Nothing to be done for 'all'.

我这边是可以的，针对这个nn_detect,我们有说明

你如果单独编译你的case code，可以直接去掉这个库

33

我照猫画虎改了下build_vx.sh,然后编译通过 但是 跑的时候报
E [compute_node:354]Create node[0] NBG fail
E [vnn_CreateSesrX2Rep:288]CHECK STATUS(-1:A generic error code, used when no other describes the error.)
E [vnn_CreateNeuralNetwork:179]CHECK PTR 179
E [main:212]CHECK PTR 212
问题2：转模型的时候 export那一步输出的top5 的值全是0，这个感觉不正常呀
image

34

@1228760706 第二步这里这个top5全是零是正常的。你的转换参数是什么样的，贴出来看看

35

$export_ovxlib \

--model-input ${NAME}.json \

--data-input ${NAME}.data \

--reorder-channel '2 1 0' \

--channel-mean-value '0 0 0 255' \

--export-dtype quantized \

--model-quantize ${NAME}.quantize \

--optimize VIPNANOQI_PID0XB9  \

--viv-sdk ${ACUITY_PATH}vcmdtools \

--pack-nbg-unify
36

@1228760706 VIM3 这里应该是--optimize VIPNANOQI_PID0X88 \

1 Like
37

好了 改了就可以了 3q（输入为啥非要20个字符，，，）

38

Hi,
It seems that with this new version of the SDK, the NPU is trying to load some kernels from binaries instead of files, as he did in previous versions:

  • The kernels that used to be provided in linux_sdk/build/sdk/nngpu_kernels folder should now be in the libNNGPUBinary.so binary
  • The kernels that used to be provided in linux_sdk/build/sdk/nnvxc_kernels folder should now be in the libNNVXCBinary.so binary
  • The kernels that used to be provided in linux_sdk/build/sdk/ovx12_vxcKernels folder should now be in the libOvx12VXCBinary.so binary

Nonetheless, I didn’t find these new kernel binaries anywhere. Would it be possible to add tthem to the aml-npu package? Without them, some model are broken, as they can’t load the kernels they need.

Thanks!

39

Development board using the 20200530 firmware version, and then use Upgrade The System To Latest Version | Khadas Documentation to update system, then download the latest code repository. Is it OK to compile on the board at the end?

40

We will update this week, please wait.

1 Like
41

hi，
2022年有木有最新的版本压缩包发出来啊？比如6.4.6.2XXX
BR