跳转到主要内容

VCK5000运行Vitis-AI 3.0 WeGO实时量化演示

<font color="#FF8000">作者：Shaoyi Chen，AMD开发工程师</font>

<strong>测试环境</strong>
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295990-01.pn…; alt=""></center>

<strong>简介</strong>

VCK5000在vitis ai 3.0对应的shell版本为xilinx-vck5000-gen4x8-qdma-base_2，对应的XRT和XRM版本为xrt_202220.2.14.354和xrm_202220.1.5.212。旧版本的VCK5000需要首先升级到新的shell上，才能保证vitsi ai的demo正常运行。在平台更新中存在一些需要遵守的规则，例如xilinx_vck5000_gen3x16_xdma_base_1不能直接升级到最新版本，需要首先升级到vck5000_gen4x8_xdma_base_2。On the fly quantization是Vitis AI 3.0新添加的特性，可以执行训练后量化 (PTQ) 以从浮点模型获取量化模型。

<strong>Part1 VCK5000升级</strong>

<strong>Step 1 使用以下命令检查当前板卡的状态</strong>
xbmgmt examine --report platform --device <Management BDF>

输出示例：
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295991-02.pn…; alt=""></center>

如果Flash properties: Type is ospi_versal一定需要更新到ospi_xgq。

只要板卡能被Vivado识别，可以通过下载比特流恢复到这个状态，恢复的方法为：
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295992-03.pn…; alt=""></center>
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295993-04.jp…; alt=""></center>

确认勾选了Erase, Program, and Verify。点击Apply, 然后点击OK。等待烧录完毕即可回到初始状态。

<strong>Step 2 卸载已安装的XRT</strong>

vck5000_gen4x8_xdma_base_2指定的XRT版本为<a target="_blank" href="https://www.xilinx.com/bin/public/openDownload?filename=xrt_202210.2.13…;。如果和这个XRT版本不符，先卸载XRT。

Remove XRT : 'sudo apt remove xrt'
Remove XRT APU if present : 'sudo apt remove xrt-apu'

安装所需的XRT
Install XRT 2.13.478 : https://www.xilinx.com/bin/public/openDownload?filename=xrt_202210.2.13…
Install XRT-APU: https://www.xilinx.com/bin/public/openDownload?filename=xrt-apu_202210…

安装Gen3和Gen4XDMA平台
Install the gen3x16 platform : https://www.xilinx.com/bin/public/openDownload?filename=xilinx-vck5000-…
AND install the gen4x8_xdma platform: https://www.xilinx.com/bin/public/openDownload?filename=xilinx-vck5000-…

<strong>Step 3 更新到xilinx_vck5000_gen3x16_xdma_base_1</strong>

将Gen3的镜像烧录到板卡中
sudo xbmgmt program --base --device <Management BDF> --image <path to install image>

完成后等待一分钟后冷启动电脑。

<strong>Step 4 更新到vck5000_gen4x8_xdma_base_2</strong>

根据链接可以将VCK5000的platform从vck5000_gen3x16_base_1更新到vck5000_gen4x8_xdma_base_2。
https://xilinx.github.io/Alveo-Versal-Platforms/alveoversalplatforms/bu…
sudo /opt/xilinx/firmware/vck5000/gen4x8-xdma/base/migration/migrate.sh <Management BDF>

完成后等待一分钟后冷启动电脑。

<strong>Step 5 更新到vck5000_gen4x8_qdma_base_2</strong>

下一步需要更新到vck5000_gen4x8_qdma_base_2，首先安装对应版本的XRT,XRM和XRT-APU，然后烧写platform，注意安装版本要和vitis-ai 3.0匹配。

安装XRT
https://www.xilinx.com/bin/public/openDownload?filename=xrt_202220.2.14…

安装XRM
https://www.xilinx.com/bin/public/openDownload?filename=xrm_202220.1.5…

安装vck5000_gen4X8_qdma平台

wget
https://www.xilinx.com/bin/public/openDownload?filename=xilinx-vck5000-…

tar xfz shell.tgz
sudo apt install ./xilinx-vck5000-gen4x8-qdma-base_2-20221205_all.deb -y

第一步，只能更新shell。
sudo /opt/xilinx/xrt/bin/xbmgmt program --device 01:00.0 --base shell --image xilinx_vck5000_gen4x8_qdma_base_2

升级完成后等待一分钟冷启动，由于没有更新SC会出现
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295994-05.pn…; alt=""></center>

将SC版本从4.4.33更新到4.4.35 需要更新两次。第一次更新SC会以失败告终。
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295995-06.pn…; alt=""></center>

等待一分钟后再次烧录SC
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295996-07.pn…; alt=""></center>

再等待一分钟后冷启动电脑，然后检查板卡状态。
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295997-08.pn…; alt=""></center>

到此更新完毕。

<strong>Part2 运行Vitis-AI 3.0 demo</strong>

获取Vitis-AI仓库。

git clone <https://github.com/Xilinx/Vitis-AI.git&gt;;

以Pytorch为例子，获取docker容器。

docker pull xilinx/vitis-ai-pytorch-cpu:latest

查询系统中存在的容器

docker images
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295998-09.pn…; alt=""></center>

启动容器。
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-295999-10.pn…; alt=""></center>

根据DPU型号设置vck5000脚本，脚本在setup_board/vck5000目录下
source ./setup.sh DPUCVDX8H_4pe_miscdwc
source ./setup.sh DPUCVDX8H_6pe_dwc
source ./setup.sh DPUCVDX8H_6PE_MISC
source ./setup.sh DPUCVDX8H_8pe_normal

该脚本会设置两个环境变量
XCLBIN_PATH = /opt/xilinx/overlaybins/DPUCVDX8H/4PE
XLNX_VART_FIRMWARE = /opt/xilinx/overlaybins/DPUCVDX8H/4PE/dpu_DPUCVDX8H_4PE_350M_xilinx_vck5000_ge

准备数据集

To use ImageNet, first download it from http://www.image-net.org/. For calibration purpose, only the validation set is needed.

数据集结构
# imagenet/train/
# ├── n01440764
# │ ├── n01440764_10026.JPEG
# │ ├── n01440764_10027.JPEG
# │ ├── ......
# imagenet/val/
# ├── n01440764
# │ ├── ILSVRC2012_val_00000293.JPEG
# │ ├── ILSVRC2012_val_00002138.JPEG
# │ ├── ......
# ├── ......

安装必要的包
pip install -r requirements.txt

准备浮点模型
wget <https://download.pytorch.org/models/inception_v3_google-0cc3c7bd.pth&gt;;

准备测试图像
wget -O wego_example_recipes.tar.gz <http://www.xilinx.com/bin/public/openDownload?filename=wego_example_re

tar xf wego_example_recipes.tar.gz -C /tmp

运行脚本
bash quantize_compile_serialize_run.sh ./inception_v3_google-0cc3c7bd.pth ./val

可以得到最终的结果
<center><img src="https://cdn.eetrend.com/files/2023-03/wen_zhang_/100569455-296000-11.pn…; alt=""></center>

<strong>常见错误</strong>

1.运行
bash quantize_compile_serialize_run.sh ./inception_v3_google-0cc3c7bd.pth ./testval/

出现错误
Traceback (most recent call last):
File "run.py", line 181, in <module>
main()
File "run.py", line 166, in main
wego_mod = get_wego_mod(img_transforms)
File "run.py", line 125, in get_wego_mod
quantized_model = wego_torch.quantize(float_model, [[target_batch] + input_shape], calib_dataloader, calibrator)
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/wego_torch/quantizer.py", line 130, in quantize
for b_idx, batch in tqdm(enumerate(dataloader), total=len(dataloader)):
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/tqdm/std.py", line 1195, in __iter__
for obj in iterable:
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 681, in __next__
data = self._next_data()
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 721, in _next_data
data = self._dataset_fetcher.fetch(index) # may raise StopIteration
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 49, in <listcomp>
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/torch/utils/data/dataset.py", line 290, in __getitem__
return self.dataset[self.indices[idx]]
File "/opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/torchvision/datasets/folder.py", line 229, in __getitem__
path, target = self.samples[index]
IndexError: list index out of range

请往数据集添加更多的测试数据。

2.更新VCK5000到xilinx_vck5000_gen4x8_xdma_base_2出错

EXE: /opt/xilinx/xrt/bin/unwrapped/xbmgmt2
[xbmgmt] ERROR: Failded to update base: Invalid argument
<li>检查platform是否存已经安装且名字输入正确。</li>
<li>联系FAE回退到Gen3X16，重新升级。</li>

参考链接：
https://support.xilinx.com/s/article/XRT-Xilinx-Runtime-version-2021-2-…
https://xilinx.github.io/Alveo-Versal-Platforms/alveoversalplatforms/bu…
https://github.com/Xilinx/Vitis-AI/tree/master/examples/wego
https://xilinx.github.io/Vitis-AI/docs/install/install.html
https://www.xilinx.com/member/vck5000.html#vitis