[Doc] Improve Jetson tutorial install doc

[PeterH0323]

Motivation

Improve mmdeploy build on Jetson doc: docs/en/tutorials/how_to_install_mmdeploy_on_jetsons.md

Modification

I follow the doc to build mmdeploy on Jetson nano (4G version, Jetpack 4.6.1), and I add some solution or details when I can't get it from the doc.

Still need to do

• Cmake version of JetPack
• Cmake update guide detail
• Conda install path to ~/archiconda and show the installation detail

Checklist

• Pre-commit or other linting tools are used to fix the potential lint issues.
• The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
• If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
• The documentation has been modified accordingly, like docstring or example tutorials.

[PeterH0323]

@AllentDan Pls take a look. Thx! 😄

[AllentDan]

LGTM

[AllentDan]

as I've forgotten the exact paths of tensorrt and cudnn, could you make sure they are right for TENSORRT_DIR and CUDNN_DIR on TX2 or Xavier?

[PeterH0323]

I only have Jetson Nano 😢
So I modefied it to:

Install MMDeploy

Using git to clone MMDeploy source code.

git clone -b master https://github.com/open-mmlab/mmdeploy.git MMDeploy
cd MMDeploy
git submodule update --init --recursive

We need the path of TensorRT and cuDNN path for MMDeploy installation.

• for Jetson Nano:

export TENSORRT_DIR=/usr/src/tensorrt
export CUDNN_DIR=/etc/alternatives

Build MMDeploy from source:

mkdir -p build && cd build
cmake .. \
    -DCMAKE_CXX_COMPILER=g++-7 \
    -DMMDEPLOY_BUILD_SDK=ON \
    -DMMDEPLOY_BUILD_SDK_PYTHON_API=ON \
    -DMMDEPLOY_TARGET_DEVICES="cuda;cpu" \
    -DMMDEPLOY_TARGET_BACKENDS="trt" \
    -DMMDEPLOY_CODEBASES=all \
    -Dpplcv_DIR=${PPLCV_DIR}/cuda-build/install/lib/cmake/ppl \
    -DTENSORRT_DIR=${TENSORRT_DIR} \
    -DCUDNN_DIR=${CUDNN_DIR}

make -j$(nproc) && make install

It is okay?

[tehkillerbee]

@PeterH0323 Are you sure your TENSORRT_DIR is correct? On my Jetson, TensorRT is detected at:

-- Found TensorRT headers at /usr/include/aarch64-linux-gnu
-- Found TensorRT libs at /usr/lib/aarch64-linux-gnu/libnvinfer.so;/usr/lib/aarch64-linux-gnu/libnvinfer_plugin.so

I set TENSORRT_DIR to /usr/include/aarch64-linux-gnu

[PeterH0323]

Hi, @tehkillerbee.
I am using Jetson Nano. May I ask what is the type of your Jetson?

[tehkillerbee]

@PeterH0323 I have checked on both AGX Xavier and NX with JetPack 4.6. /usr/src/tensorrt does exist - but it does not contain headers or libs as far as I can see. When you build MMDeploy, does it list where TensorRT headers, libs has been detected?

[PeterH0323]

@tehkillerbee Done, the script show it can find trt when I set it to /usr/src/tensorrt, but thx for you reminding, I change both -DTENSORRT_DIR and -DCUDNN_DIR to /usr/include/aarch64-linux-gnu 😄

[PeterH0323]

@lzhangzz Pls take a look. Thx! 😄

[lzhangzz]

@PeterH0323 @AllentDan have you guys tested the SDK demo on Jetson Nano?

We are seeing reports from the WeChat group that ppl.cv is reporting "missing suitable binary for exectuion on deivce" error. This maybe caused by the missing gencode for sm_53 in cuda.cmake from both ppl.cv and mmdeploy.

[lvhan028]

The version of python we got from the previous step. is not a sentence.
May use "Ensure the python version in the command is the same as the above"

[PeterH0323]

Done.

Meanwhile, I added JetPack version at the beginning of the doc

Note: The JetPack we use is 4.6, and the default python version of it is 3.6

[lvhan028]

May use "Since MMDeploy is using mmcv 1.4.0, you can install it from source as below:"

[PeterH0323]

Done

[AllentDan]

below -> above?

[PeterH0323]

Fixed

[lzhangzz]

I believe NVIDIA have never announced these models as "main Jetson boards".

maybe "which we have verifed on the following models" is better

[PeterH0323]

Done