【FAQ】Course-Lab_Finn

[Yuoto]

👋 LAB-FINN Q&A

Feel free to ask questions about LAB-FINN here!

• Ask questions you’re wondering about.
• Share ideas.
• Engage with other community members.
• Welcome others and are open-minded. Remember that this is a community we
  build together 💪.

For more discussions about FINN, please refer to the FINN Official Disscussions. Here we mainly focus on LAB-FINN in HLS course.

Highly Recommended:

Please use the Netron tool and open the .onnx files generated in each stage. This would give you greater details about the notebook python code.

Some Frequently Asked Questions:

Installation

• Please use Vivado 2022.1 and Vitis_hls 2022.1, this is currently compatible with the latest FINN v0.8 version.

• Ubuntu 20.04 issues
  For Ubuntu 20.04 users, some versions of Vivado (e.g. 2020.1) do not officially support Ubuntu 20.04. However, we tested the FINN end2end notebook and didn't run into errors currently.
  ※ Note that for Ubuntu 20.04, the Vivado GUI installer will run into the error
  Exception in thread "SPLASH_LOAD_MESSAGE" java.lang.IllegalStateException: no splash screen available.
  To cope with this, we can install in batch mode, see here for more batch mode installation details.

• FINN setup guide & Y2k22 bug (Y2k22 bug is only for Vivado version < 2022.1)
  Please refer to LAB-FINN setup guide w/wo VM

Jupyter Notebook

• In the below code (This issue is fixed in the FINN v0.8 version)

model = model.transform(DeployToPYNQ(ip, port, username, password, target_dir))

There might be errors due to the lack of pynq drivers.
"TypeError: stat: path should be string, bytes, os.PathLike or integer, not NoneType"
To cope with this, just add the code below

from finn.transformation.fpgadataflow.make_pynq_driver import MakePYNQDriver
model = ModelWrapper(build_dir + "/tfc_w1_a1_post_synthesis.onnx")
model = model.transform(MakePYNQDriver(platform="zynq-iodma"))

see here for details.

Pynq

• Issue with ssh-copy-id: (This issue is removed since we employ the PYNQ online renting service)
  "mktemp: failed to create file via template '/tmp/home_dir/.ssh/ssh-copy-id_id.XXXXXXXXXX': No such file or directory"
  To resolve this, use the following command line in the FINN docker environment

$ xxx@finn_dev_xxx:/workspace/finn/ssh_keys$ mkdir /tmp/home_dir/.ssh
$ xxx@finn_dev_xxx:/workspace/finn/ssh_keys$ ssh-copy-id -i id_rsa.pub [email protected] -p XXXX

see here for details.

• ssh: Could not resolve hostname (This issue is removed since we employ the PYNQ online renting service)
  Please beware of the IP environment variables, since sometimes the value stored inside is not as you expected.

• FileNotFoundError: No such file or directory: xxx/output.npy
  Please go to pynq board, find the generated driver.py file and execute it with root privilege.

$ su
$ python3 driver.py

• Error "No module named 'pynq'"
  please install pynq directly with

$ python3 -m pip install pynq

see here for details.

• No device found
  If you're using pynq image version >= 2.7, you might encounter this problem during running

$ sudo python3 driver.py

To deal with this, change the user to root (su) and not xilinx

$ su
$ python3 driver.py

• TypeError: 'NoneType' object is not iterable, during jupyter notebook throughput_test_remote (This issue is removed since we employ the PYNQ online renting service)
  

This is due to the same problem mentioned above If you're using pynq image version >= 2.7.
See here for discussion details and explanation.
To deal with this, there are 2 steps,

1. switch to root in pynq manually，then type in the following command in pynq:
   (switch to pynq deployment directory) (e.g. cd /home/xilinx/finn_dev_huayang/pynq_deployment_apoe927o)

>>> su
>>> python3 driver.py --exec_mode=throughput_test --batchsize=1000 --bitfile=resizer.bit --inputfile=input.npy --outputfile=output.npy --platform=zynq-iodma

2. Go back to the jupyter notebook，open a clean code cell below pynq deployment part，enter the below code segment。

import os, sys
import subprocess
from finn.core.modelwrapper import ModelWrapper

build_dir = "/workspace/finn"
model = ModelWrapper(build_dir + "/tfc_w1_a1_pynq_deploy.onnx")
timeout=None
pynq_ip = model.get_metadata_prop("pynq_ip")
pynq_port = int(model.get_metadata_prop("pynq_port"))
pynq_username = model.get_metadata_prop("pynq_username")
pynq_password = model.get_metadata_prop("pynq_password")
pynq_target_dir = model.get_metadata_prop("pynq_target_dir")
deployment_dir = model.get_metadata_prop("pynq_deploy_dir")
deployment_folder = os.path.basename(os.path.normpath(deployment_dir))
platform = model.get_metadata_prop("platform")
assert platform in ["alveo", "zynq-iodma"]
bitfile = model.get_metadata_prop("bitfile")
bitfile = os.path.basename(bitfile)
if pynq_password == "":
        if "zynq" in platform:
                 raise Exception("PYNQ board remote exec needs password for sudo")
        else:
                local_prefix = ""  # assume we are using an ssh key
                warnings.warn("Empty password, make sure you've set up an ssh key")
else:
        local_prefix = "sshpass -p %s " % pynq_password

# remove any pre-existing metrics file
try:
        os.remove("{}/nw_metrics.txt".format(deployment_dir))
except FileNotFoundError:
        pass

cmd = local_prefix + "scp -P{} {}@{}:{}/{}/nw_metrics.txt {}".format(
        pynq_port,
        pynq_username,
        pynq_ip,
        pynq_target_dir,
        deployment_folder,
        deployment_dir,
    )
bash_command = ["/bin/bash", "-c", cmd]
process_compile = subprocess.Popen(bash_command, stdout=subprocess.PIPE)
process_compile.communicate(timeout=timeout)

try:
        with open("{}/nw_metrics.txt".format(deployment_dir), "r") as file:
                res = eval(file.read())
except FileNotFoundError:
        print('FileNotFoundError: no such file  nw_metrics.txt')
        sys.exit()
print("Network metrics:")
for key in res:
        print(str(key) + ": " + str(res[key]))
II = 64
# frequency in MHz
f_MHz = 100
# expected throughput in MFPS
expected_throughput = f_MHz / II
# measured throughput (FPS) from throughput test, converted to MFPS
measured_throughput = res["throughput[images/s]"] * 0.000001
# peformance
print("We reach approximately " + str(round((measured_throughput / expected_throughput)*100)) + "% of the ideal performance.")

Error with Brevitas/ONNX

• TypeError: export() got an unexpected keyword argument 'enable_onnx_checker'
  Since the newer version of PyTorch removed "enable_onnx_checker", we need to comment line77 in the following file manually:
  /path/to/usr/home/.local/lib/python3.8/site-packages/brevitas/export/onnx/finn/manager.py line77

Unless a newer version of Brevitas (currently is 0.7.1, I've checked that their dev branch fixed this issue) is provided， or we'll have to use PyTorch of version <=1.9.
See here for details.

[yunchenlo]

助教 您好，

我嘗試在 docker 環境內添加 ssh public key，發現會有以下錯誤，請問助教有遇過這樣的問題嗎？

/usr/bin/ssh-copy-id: ERROR: No user exists for uid 1046

我覺得有兩個地方怪怪的：

1. docker 的 user name 是 I have no name!
2. 我在 server 上添加 public key 沒問題，進到 docker 環境內添加就掛掉

錯誤碼如下：

[Yuoto]

我發現原因了，因為我是用HLS01幫大家加入docker group，但我們是用NIS系統帳號資料都放在HLS01上面，我以為docker資料會共享但其實沒辦法。 如果登進別台server HLS02, 03, 04 都會有這個問題。
我剛剛已經把四台都確認過了，請再試一次看看!

[yunchenlo]

收到了，謝謝

[yunchenlo]

助教不好意思，好像還是沒有解掉。謝謝！

錯誤碼如下：

I have no name!@finn_dev_nthu04:/workspace/finn/ssh_keys$ ssh-copy-id -i id_rsa.pub [email protected] -p 10800
/usr/bin/ssh-copy-id: INFO: Source of key(s) to be installed: "id_rsa.pub"
/usr/bin/ssh-copy-id: INFO: attempting to log in with the new key(s), to filter out any that are already installed

/usr/bin/ssh-copy-id: ERROR: No user exists for uid 1046

[Yuoto]

hmmm 看起來這個問題一時不好解，如果可以的話先你們先用 HLS01 這台server。這台很多人在用，所以可能等我確認其他server沒問題後再回覆!

[yunchenlo]

ok, 助教辛苦了！

我先去 HLS01

允辰 敬上

[ChihyuLin0211]

助教你好

我在VM ubuntu中的docker 環境內添加 ssh public key時有遇到以下錯誤，"connection closed by remote host"，請問要如何處理，謝謝

[Yuoto]

同學好，你試試看不要加sudo 因為已經是在root環境了

[yunchenlo]

Dear 助教，

我在清大板子測試 part III throughput evaluation 時，
執行res = throughput_test_remote(model, 10000) 時，會出現 res 是 none：

driver, pynq, numpy, bitstring 都有裝好，也可以進到板子上用 su 身份執行 validate.py

請問助教有其他方法可以評估 throughput 嗎？謝謝
允辰 敬上

[Yuoto]

在FINN環境裡面應該適用他提供的API會比較準確。 我可能需要更多資訊，先前你試這份檔案的時候會有這個問題嗎? 還是只有在清大的板子才有? 或是你前面有改什麼folding參數才變成這樣?

[yunchenlo]

1. 我現在都是在清大板子，所以可能跟 SD card image 版本有關 (台大板子我無法安裝 pynq 、無法部署，所以我現在沒有在試了)
2. 沒有調 folding 參數也會這樣
3. 最剛開始在台大板子試的時候沒有這個問題 這個時候，謝謝！

允辰 敬上

[Yuoto]

同學好，
感謝回報，看起來pynq端下次也要統一版本，目前這個一定要用root才可以使用driver的問題影響滿大的，我會請課程助教延後這份bonus作業繳交的時間。

關於這個問題，一樣是因為他 throughput_test_remote 裡面跟pynq溝通時用的指令是舊的指令
可以在以下FINN github repository中發現他用的是sudo 而不是 root，且python版本是3.6 不是3.8
https://github.com/Xilinx/finn-base/blob/dev/src/finn/core/throughput_test.py

解決辦法:
我目前找不到可以直接在python呼叫時切換成root的方法
因此可能執行到這個部分的時候要分成兩階段

1. 手動在 pynq 裡面切換root使用者，然後手動輸入以下指令 (原本是python code幫忙寫的)
   (切換到pynq deployment資料夾) (e.g. cd /home/xilinx/finn_dev_huayang/pynq_deployment_apoe927o)
   su
   (輸入root密碼)
   python3 driver.py --exec_mode=throughput_test --batchsize=1000 --bitfile=resizer.bit --inputfile=input.npy --outputfile=output.npy --platform=zynq-iodma

以上是利用pynq driver code執行 thoughput test，並且指定 input, output 資料的位置， 以及合成完的hardware 資訊。
因為裡面牽涉到DRAM access時間，因此只能靠driver code來做這件事。 上述指令執行後，pynq端會吐出一份類似 log 檔的文字檔，名稱是 "nw_metrics.txt"，下一步就是把這份統整了pynq執行完的資料再回報給原本的jupyter notebook。

2. 回到原本的 jupyter notebook，開一個空格，輸入以下code。 基本上下面的code就是原本finn提供的 throughput_test_remote 剩下的程式碼。只是我們把中間從python送指令到pynq的那個段落切開來而已。

import os, sys
import subprocess
from finn.core.modelwrapper import ModelWrapper

build_dir = "/workspace/finn"
model = ModelWrapper(build_dir + "/tfc_w1_a1_pynq_deploy.onnx")
timeout=None
pynq_ip = model.get_metadata_prop("pynq_ip")
pynq_port = int(model.get_metadata_prop("pynq_port"))
pynq_username = model.get_metadata_prop("pynq_username")
pynq_password = model.get_metadata_prop("pynq_password")
pynq_target_dir = model.get_metadata_prop("pynq_target_dir")
deployment_dir = model.get_metadata_prop("pynq_deploy_dir")
deployment_folder = os.path.basename(os.path.normpath(deployment_dir))
platform = model.get_metadata_prop("platform")
assert platform in ["alveo", "zynq-iodma"]
bitfile = model.get_metadata_prop("bitfile")
bitfile = os.path.basename(bitfile)
if pynq_password == "":
        if "zynq" in platform:
                 raise Exception("PYNQ board remote exec needs password for sudo")
        else:
                local_prefix = ""  # assume we are using an ssh key
                warnings.warn("Empty password, make sure you've set up an ssh key")
else:
        local_prefix = "sshpass -p %s " % pynq_password

# remove any pre-existing metrics file
try:
        os.remove("{}/nw_metrics.txt".format(deployment_dir))
except FileNotFoundError:
        pass

cmd = local_prefix + "scp -P{} {}@{}:{}/{}/nw_metrics.txt {}".format(
        pynq_port,
        pynq_username,
        pynq_ip,
        pynq_target_dir,
        deployment_folder,
        deployment_dir,
    )
bash_command = ["/bin/bash", "-c", cmd]
process_compile = subprocess.Popen(bash_command, stdout=subprocess.PIPE)
process_compile.communicate(timeout=timeout)

try:
        with open("{}/nw_metrics.txt".format(deployment_dir), "r") as file:
                res = eval(file.read())
except FileNotFoundError:
        print('FileNotFoundError: no such file  nw_metrics.txt')
        sys.exit()
print("Network metrics:")
for key in res:
        print(str(key) + ": " + str(res[key]))
II = 64
# frequency in MHz
f_MHz = 100
# expected throughput in MFPS
expected_throughput = f_MHz / II
# measured throughput (FPS) from throughput test, converted to MFPS
measured_throughput = res["throughput[images/s]"] * 0.000001
# peformance
print("We reach approximately " + str(round((measured_throughput / expected_throughput)*100)) + "% of the ideal performance.")