Vitis-AI 1.4 Flow for Kria KV260 Vision AI Starter Kit

I was asked by my colleague ctammann  to provide a design for the Kria KV260 Vision Starter Kit that would be a power hog for some power testing.

Although all of the Kria apps seems to be using a B3136 DPU, I did remember that Xilinx generated benchmarks for the KV260 using a B4096 DPU.

Looking at the benchmarks taken from the Xilinx Vitis-AI github repository:
https://github.com/Xilinx/Vitis-AI/blob/master/models/AI-Model-Zoo/README.md#performance-on-kria-kv260-som

I figured that the most compute intensive model should be a good candidate to stress the KV260 for power testing.

In the following table, I identify a few candidates highlighted in yellow, including the pointpainting model.

In order to try this out, I first had to get the Vitis-AI 1.4 design for KV260 up and running.

Here are some instructions to quickly get started with the Vitis-AI 1.4 design for KV260:

1. Download the following SD card image:
   
   https://www.xilinx.com/member/forms/download/design-license-xef.html?filename=xilinx-kv260-dpu-v2020.2-v1.4.0.img.gz
   
   
2. Program to 32GB micro SD card with Balena Etcher, then boot
   
   
3. With an internet connection, install the following archives of video and image files:
   
   root@xilinx-k26-starterkit-2020_2:~# wget https://www.xilinx.com/bin/public/openDownload?filename=vitis_ai_runtime_r1.4.0_image_video.tar.gz -O vitis_ai_runtime_r1.4.0_image_video.tar.gz
   root@xilinx-k26-starterkit-2020_2:~#tar -xvzf vitis_ai_runtime_r1.4.0_image_video.tar.gz -C Vitis-AI/demo/VART
   
   root@xilinx-k26-starterkit-2020_2:~#wget https://www.xilinx.com/bin/public/openDownload?filename=vitis_ai_library_r1.4.0_images.tar.gz -O vitis_ai_library_r1.4.0_images.tar.gz
   root@xilinx-k26-starterkit-2020_2:~#tar -xvzf vitis_ai_library_r1.4.0_images.tar.gz -C Vitis-AI/demo/Vitis-AI-Library
   
   root@xilinx-k26-starterkit-2020_2:~#wget https://www.xilinx.com/bin/public/openDownload?filename=vitis_ai_library_r1.4.0_video.tar.gz -O vitis_ai_library_r1.4.0_video.tar.gz
   root@xilinx-k26-starterkit-2020_2:~#tar -xvzf vitis_ai_library_r1.4.0_video.tar.gz -C Vitis-AI/demo/Vitis-AI-Library
   
   
4. Query the DPU core with the xdputil utility:
   
   

   root@xilinx-k26-starterkit-2020_2:~# xdputil query

   {

     "DPU IP Spec":{

         "DPU Core Count":1,

         "DPU Target Version":"v1.4.1",

         "IP version":"v3.3.0",

         "generation timestamp":"2021-06-07 19-15-00",

         "git commit id":"df4d0c7",

         "git commit time":2106071910,

         "regmap":"1to1 version"

     },

     "VAI Version":{

         "libvart-runner.so":"Xilinx vart-runner Version: 1.4.0-fa49b842f283242091476cf8e1ae4d242a2a838e  2021-07-14-07:13:01 ",

         "libvitis_ai_library-dpu_task.so":"Xilinx vitis_ai_library dpu_task Version: 1.4.0-01d12d1134678e1400e683d32e88bc77886f2

         247 2021-07-14 07:14:27 [UTC] ",

         "libxir.so":"Xilinx xir Version: xir-ff89b11dcabb00eef6d148fcf660c8e6d02eb184 2021-07-14-07:12:09",

         "target_factory":"target-factory.1.4.0 ce1b39e329cc06cb7545e8aa39174fb8b9969f0b"

     },

     "kernels":[

     {

         "DPU Arch":"DPUCZDX8G_ISA0_B4096_MAX_BG2",

         "DPU Frequency (MHz)":300,

         "IP Type":"DPU",

         "Load Parallel":2,

         "Load augmentation":"enable",

         "Load minus mean":"disable",

         "Save Parallel":2,

         "XRT Frequency (MHz)":300,

         "cu_addr":"0xa0010000",

         "cu_handle":"0xaaaae4a4e220",

         "cu_idx":0,

         "cu_mask":1,

         "cu_name":"DPUCZDX8G:DPUCZDX8G_1",

         "device_id":0,

         "fingerprint":"0x1000020f6014407",

         "name":"DPU Core 0"

     }

   ]

   }

   root@xilinx-k26-starterkit-2020_2:~#

    

    

5. Optimize the DDR's QoS settings:
   
   

   root@xilinx-k26-starterkit-2020_2:~# cd dpu_sw_optimize/zynqmp/

   root@xilinx-k26-starterkit-2020_2:~/dpu_sw_optimize/zynqmp# ./zynqmp_dpu_optimize.sh

   Start QoS config ...[✔]

   root@xilinx-k26-starterkit-2020_2:~/dpu_sw_optimize/zynqmp# cd -

   root@xilinx-k26-starterkit-2020_2:~#
   
   

6. Turn off verbose:
   
   

   root@xilinx-k26-starterkit-2020_2:~# dmesg -D
   
   

7. Run the pointpainting performance example application for 30 seconds:
   
   

   root@xilinx-k26-starterkit-2020_2:~# cd Vitis-AI/demo/Vitis-AI-Library/samples/pointpainting

    

   root@xilinx-k26-starterkit-2020_2:~/Vitis-AI/demo/Vitis-AI-Library/samples/pointpainting# ./test_performance_pointpainting semanticfpn_nuimage_576_320_pt pointpainting_nuscenes_40000_64_0_pt pointpainting_nuscenes_40000_64_1_pt ./test_performance_pointpainting.list -t 1 -s 30

    

   WARNING: Logging before InitGoogleLogging() is written to STDERR

   I1112 12:36:49.335223  2276 benchmark.hpp:184] writing report to <STDOUT>

   I1112 12:36:49.335584  2276 benchmark.hpp:211] waiting for 0/30 seconds, 1 threads running

   I1112 12:36:59.335726  2276 benchmark.hpp:211] waiting for 10/30 seconds, 1 threads running

   I1112 12:37:09.335916  2276 benchmark.hpp:211] waiting for 20/30 seconds, 1 threads running

   I1112 12:37:19.336179  2276 benchmark.hpp:219] waiting for threads terminated

   FPS=1.2936

    

    

   root@xilinx-k26-starterkit-2020_2:~/Vitis-AI/demo/Vitis-AI-Library/samples/pointpainting# ./test_performance_pointpainting semanticfpn_nuimage_576_320_pt pointpainting_nuscenes_40000_64_0_pt pointpainting_nuscenes_40000_64_1_pt ./test_performance_pointpainting.list -t 4 -s 30

    

   WARNING: Logging before InitGoogleLogging() is written to STDERR

   I1112 12:36:49.335223  2276 benchmark.hpp:184] writing report to <STDOUT>

   I1112 12:36:49.335584  2276 benchmark.hpp:211] waiting for 0/30 seconds, 4 threads running

   I1112 12:36:59.335726  2276 benchmark.hpp:211] waiting for 10/30 seconds, 4 threads running

   I1112 12:37:09.335916  2276 benchmark.hpp:211] waiting for 20/30 seconds, 4 threads running

   I1112 12:37:19.336179  2276 benchmark.hpp:219] waiting for threads terminated

   FPS=2.61017

    

    

8. Query the power metrics:
   
   

   root@xilinx-k26-starterkit-2020_2:~# xmutil platformstats -p

    

   Power Utilization

   SOM total power    :    5350 mW

   SOM total current    :    1068 mA

   SOM total voltage      : 5005 mV

   AMS CTRL

   System PLLs voltage measurement, VCC_PSLL :    1201 mV

   PL internal voltage measurement, VCC_PSBATT :    716 mV

   Voltage measurement for six DDR I/O PLLs, VCC_PSDDR_PLL :    1797 mV

   VCC_PSINTFP_DDR voltage measurement : 840 mV

    

   PS Sysmon

   LPD temperature measurement :    30 C

   FPD temperature measurement (REMOTE) :    30 C

   VCC PS FPD voltage measurement (supply 2) :    842 mV

   PS IO Bank 500 voltage measurement (supply 6) :    1789 mV

   VCC PS GTR voltage :    856 mV

   VTT PS GTR voltage :    1804 mV

    

   PL Sysmon

   PL temperature :    29 C

    

   root@xilinx-k26-starterkit-2020_2:~#

In order to show the power metrics while the AI application is running, the AI application can be run as a background task (with '&') as shown below:

root@xilinx-k26-starterkit-2020_2:~/Vitis-AI/demo/Vitis-AI-Library/samples/pointpainting# ./test_performance_pointpainting semanticfpn_nuimage_576_320_pt pointpainting_nuscenes_40000_64_0_pt pointpainting_nuscenes_40000_64_1_pt ./test_performance_pointpainting.list -t 4 -s 300 &

root@xilinx-k26-starterkit-2020_2:~# xmutil platformstats -p

Using this technique, we were able to measure the SOM total power above 10 W !

Cheers !