LeNet AIE tutorial hangs with EN_TRACE=1 option while running on VCK190 hardware

[Hari-MathWorks]

Hello Everyone,

I'm trying out the 2022.2 version 01-aie_lenet_tutorial. I was able to build the example and successfully deployed on to the hardware. Thanks for the detailed steps in the tutorial page.

However, When I tried to see the throughput measurement details by building the application using the "EN_TRACE=1" option mentioned in the tutorial link, application hangs at the runtime.

Please see the below screenshot for the Tera Term command window output.

Is this behavior expected? Let me know is there anyway to bypass this behavior and see the throughput details.

Thanks, Hari

[SURUTHI1605]

Can you share the vitis version , build name and run commands you had used?

[Hari-MathWorks]

Below are version and commands details: Vitis Version: 2022.2

01-aie_lenet_tutorial build steps used:

make kernels TARGET=hw EN_TRACE=1 make graph TARGET=hw EN_TRACE=1 make xsa TARGET=hw EN_TRACE=1 make application TARGET=hw EN_TRACE=1 make package TARGET=hw EN_TRACE=1 Runtime steps:

Followed the same steps mentioned in the tutorial page. It seems application hanged while running the "dma hls"

Basically, I would like to see the benchmark results for the AI Engine. Let me know if you need any information.

[SURUTHI1605]

Can you share the result of "which vivado" or "which vitis" after sourcing the vitis 2022.2 build?

[Hari-MathWorks]

Output of "which vitis": /home/hnalla/Desktop/XilinxVersalAI/installDir/Vitis/2022.2/bin/vitis

Output of platforminfo --list | grep -m 1 -A 9 vck190_base

[SURUTHI1605]

Please share the folders in base_platforms repository.

[SURUTHI1605]

You're using "xilinx_vck190_base_dfx_202220_1" platform to run this design. But it has been designed for "xilinx_vck190_base_202220_1" platform. Please use the below mentioned platform and try to run the design.

Confirm you have the VCK190 Production Base Platform. "platforminfo --list | grep -m 1 -A 9 vck190_base"

Output of the above command should be as follows:

"baseName": "xilinx_vck190_base_202220_1",

Thanks

[Hari-MathWorks]

Hi @SURUTHI1605 ,

Below are the folders presented in the base_platforms:

It seems, I already have "Xilinx_vck_190_base_202220_1" folder. If I use the vck190_base_2022 then it will show correct base name. Below is the output of platforminfo --list | grep -m 1 -A 9 vck190_base_2022

I also confirmed that in the compilation logs, platforms files are picked from the Xilinx_vck_190_base_202220_1 folder.

[Hari-MathWorks]

Also, I have modified the core003.cc implementation. It has optimized kernel code. I have modified to use the plain c code for matrix multiplication. I'm doing this exercise to see the performance improvement with the AIE kernel code.

Please see the attached "core03.cc" file to the modifications I made. core03.cc.tar.gz

[SURUTHI1605]

Hi @Hari-MathWorks , from your side have you done any changes other than core03.cc file? For the design we need to check first in hardware emulation platform and then in hardware. Since you've done some changes in AIE kernel we need to check and verify the handshaking signal and output of AIE in emulation platform. I can try this from my side.

Thanks

[Hari-MathWorks]

Hi @SURUTHI1605 ,

If I run the 01-aie_lenet_tutorial without any modifications then it worked fine. I don't see any issues. But, If i use modified(Specifically used plain "C" code for matrix multiplications) the core03.cc mentioned in the above comment then, I was seeing the crash. I'm also looking into this.

I have also experimented by removing matrix multiplication aie code in core03.cc then application running with golden output mismatch error at the end. But, this is expected.

[SURUTHI1605]

Hi @Hari-MathWorks , Are you able to build the design?