IMX8MP Neural Processing Unit

The Gateworks Venice SBCs that are using the i.MX8M Plus processors have a built in Neural Processing Unit (NPU) for machine learning.

• All GW74xx use the i.MX8M Plus processor
• Any GW71xx, GW72xx and GW73xx using a GW702x SOM module will use the i.MX8M Plus processor

The IMX8MP NPU IP is VeriSilicon (Vivante VIP8000) and offers 2.3 TOPS of acceleration for inference in endpoint devices; object identification, speech recognition of 40K words, or even medical imaging (MobileNet v1 at 500 images per second). Out of the box, this makes Gateworks boards with NPU capabilities powerful for AI applications on the edge.

NXP Yocto BSP

The easiest way to get started with the NPU is to use a image from the NXP BSP. This image contains the necessary libraries and kernel to interface the NPU with TensorFlow without much configuration. You can either ​ follow the guide to build their image or ​ download a pre-built one (recommended).

This guide assumes you have:

• A Gateworks board with a i.MX8M Plus procesor.
• A NXP account, which is necessary to download their image and models.
• A >= 16GB flash drive, SD card, or other removable block storage to install a Rescue Image, NXP Image, and updated device trees (DTBs) onto the board.

The steps are as generalized as possible to not depend on the boards available RAM to load an image, or the low speeds of JTAG uploading, as the .wic from NXP is >8GB. We will use a ramdisk to boot a "rescue image" fully in RAM, then use dd to write from the removable multimedia (flash drive) to the onboard eMMC (/dev/mmcblk2).

NOTE: In the scripts below, we disable PCIe as a temporary fix to prevent the NXP 6.6.3_1.0.0 kernel from hanging on boot. This is caused by a missing patch necessary to work around a PCIe switch quirk when used on the IMX8MP, which can be found specifically ​here from our kernel.

1. Download the Gateworks Venice Rescue Image to removable multimedia.

Find which device your flash drive/SD card is. For example, /dev/sdc

Then, use dd to flash this image onto your device byte-for-byte.

DEVICE=<your device, no trailing />
wget https://dev.gateworks.com/buildroot/venice/minimal/rescue.img.gz
zcat rescue.img.gz | sudo dd of=${DEVICE} bs=1M oflag=sync 

In this guide, we will have the NXP image and our rescue image on the same drive, so we will resize the partition and file system to fit both. If you're using separate devices, this is not necessary.

sudo parted ${DEVICE} "resizepart 1 -0" # resize the partition to fit the size of the drive
sudo resize2fs ${DEVICE}1 # resize ext fs on device partition 1

2. Download the NXP BSP evaluation kit image to removable multimedia.

On your host machine, install the Linux 6.6.3_1.0.0 image for the i.MX 8M Plus EVK ​ here.

In this download you will find imx-image-full-imx8mpevk.wic, which is a Yocto-generated image with all of the ML libraries. Copy this image to our device.

sudo mount ${DEVICE}1 /mnt
sudo cp imx-image-full-imx8mpevk.wic /mnt/

3. Patch & Build patch Venice DTBs from the Kernel source.

Due to small inconsistencies between the NXP and Gateworks devicetrees for bleeding-edge peripherals, a patch is required until mainline compatibility is reached. The below script gets the patches from the attachments at the bottom of this page.

git clone https://github.com/nxp-imx/linux-imx -b lf-6.6.y
cd linux-imx
wget https://trac.gateworks.com/raw-attachment/wiki/venice/npu/0001-arm64-dts-imx8mp-venice-fix-USB_OC-pinmux.patch
wget https://trac.gateworks.com/raw-attachment/wiki/venice/npu/0002-arm64-dts-imx8mm-venice-gw700x-remove-ddrc.patch
wget https://trac.gateworks.com/raw-attachment/wiki/venice/npu/0003-arm64-dts-freescale-add-Gateworks-venice-board-dtbs.patch
wget https://trac.gateworks.com/raw-attachment/wiki/venice/npu/0004-arm64-dts-imx8mp-venice-gw74xx-enable-gpu-nodes.patch
patch -p1 < 0001-arm64-dts-imx8mp-venice-fix-USB_OC-pinmux.patch
patch -p1 < 0002-arm64-dts-imx8mm-venice-gw700x-remove-ddrc.patch
patch -p1 < 0003-arm64-dts-freescale-add-Gateworks-venice-board-dtbs.patch
patch -p1 < 0004-arm64-dts-imx8mp-venice-gw74xx-enable-gpu-nodes.patch
ARCH=arm64 make imx_v8_defconfig
ARCH=arm64 make dtbs

Copy these patched dtbs to a directory on your flash such as /nxp/, as to not overwrite the ones necessary for booting into the rescue image.

sudo mkdir /mnt/nxp
sudo cp arch/arm64/boot/dts/freescale/*venice*.dtb /mnt/nxp/

Now, the contents of the device should include:

• Rescue image
• Rescue image dtbs
• rootfs.cpio.xz and boot.scr for booting Rescue Image
• nxp/ with new, updated dtbs

4. Boot Rescue Image ramdisk on board

Connect the removable multimedia, in our case a USB stick, to the board before powering. Many boards have built-in SD readers, which would change the device commands slightly.

Connect serial console via JTAG and power on the board. Enter the U-Boot console by stopping autoboot.

Sanity check: is the USB device properly detected?

usb start
part list usb 0

This command should have an expected output like below

Partition Map for USB device 0  --   Partition Type: DOS

Part    Start Sector    Num Sectors     UUID            Type
  1     2048            204800          6fd772a2-01     83 Boot

Override the boot_targets variable temporarily to ensure booting into the Rescue Image, then boot into it. If you are not using usb0, run print boot_targets to see a list.

setenv boot_targets usb0
run bootcmd

If all functions normally, you should be met with a login; login with root and you will enter the shell.

Flash NXP .wic and patched DTBs onto eMMC

You are now booted into the ramdisk rescue image. The next steps are to flash the .wic onto the emmc.

For venice boards the emmc that we are imaging is /dev/mmcblk2 and with only one removable storage device your rescue image with be /dev/sda.

Image the emmc as followes:

mkdir /mnt/src
mkdir /mnt/dst
mount /dev/sda1 /mnt/src
dd if=/mnt/src/imx-image-full-imx8mpevk.wic of=/dev/mmcblk2 bs=16M oflag=sync # this will take a couple of minutes
mount /dev/mmcblk2p1 /mnt/dst
cp /mnt/src/*.dtb /mnt/dst/
cp /mnt/src/nxp/*.dtb /mnt/dst/

This flashes the prebuilt .wic image (both partitions, the kernel and fs) to our eMMC, then also brings over the old and new device trees. Next, we will create the boot script. If the below doesn't copy right, the file can be created/edited in a text editor like vi; just remove the EOF line.

cat <<\EOF > boot.scr.txt
setenv bootargs 'root=/dev/mmcblk2p2'
load mmc 2:1 $kernel_addr_r Image

setenv fdt_addr
setenv fdt_list $fdt_file $fdt_file1 $fdt_file2 $fdt_file3 $fdt_file4 $fdt_file5
setenv load_fdt 'echo Loading $fdt...; load ${devtype} ${devnum}:${distro_bootpart} ${fdt_addr_r} ${prefix}${fdt} && setenv fdt_addr ${fdt_addr_r}'
for fdt in ${fdt_list}; do if test -e ${devtype} ${devnum}:${distro_bootpart} ${prefix}${fdt}; then run load_fdt; fi; done
if test -z "$fdt_addr"; then echo "Warning: Using bootloader DTB"; setenv fdt_addr $fdtcontroladdr; fi
#Disables PCI; patch is needed, otherwise kernel hangs: See note at start of wiki page.
fdt addr $fdt_addr_r && fdt resize && fdt set /soc@0/pcie@33800000 status disabled 
booti $kernel_addr_r - $fdt_addr_r

EOF

'Compile' the boot script txt and flash it onto the MMC

mkimage -A arm64 -T script -C none -d boot.scr.txt /mnt/dst/boot.scr
umount /mnt/dst
umount /mnt/src

Boot into the NXP image.

Power cycle the board, and note the kernel which it boots into. The board should automatically boot into the eMMC image we just flashed, meaning the removable multimedia need not be connected.

If there is an error, look at the logs and the boot scripts in U-Boot.

At this point, all features regarding the Kernel and below are properly enabled. If you have an application that uses TensorFlow, it will run on the NPU or GPU using /usr/lib/libvx_delegate.so. Follow the ​ NXP Machine Learning User's Guide for more information.

Image Classification Example

As per the ​ NXP Machine Learning User's Guide, we will test a simple image labeling script on both the CPU and NPU.

$ cd /usr/bin/tensorflow-lite-2.15.0/examples
$ python3 label_image.py # without NPU acceleration
$ python3 label_image.py -e /usr/lib/libvx_delegate.so # with NPU accerlation via the libvx_delegate external TensorFlow delegate 

Result from either label_image script:

0.878431: military uniform
0.027451: Windsor tie
0.011765: mortarboard
0.011765: bulletproof vest
0.007843: sax

Without the NPU: Image Classification time: 170.5 ms With the NPU: Image Classification: 3.2 ms

Without considering the warmup times, this is a >98% speedup! For every CPU frame, the NPU can process 53.

This data is derived from classifications/sec = 1/(image classification time)

GStreamer Example for Detection

Section 8.2 of the Machine Learning Users Guide details this process, such as how to download the necessary models. After following the download steps, the home/root/nxp-nnstreamer-examples/ directory on your board should have a downloads directory with models and media directories. If not, you need to run the update script on your host to compile the models and scp them to the board.

On your host, execute the following command to have GStreamer take in video over UDP. gst-launch-1.0 udpsrc port=5000 ! application/x-rtp,payload=96 ! rtpjpegdepay ! jpegdec ! autovideosink

Host GStreamer pipeline (SVG)

On your board, execute the following to send a stream over UDP to the host port 5000. This script was derived from Section 8.1 of the Machine Learning Users Guide. The GStreamer command takes in a video input and overlays both bounding boxes and labels on it using TensorFlow and NXP filters.

CAMERA= <your camera device, such as /dev/video2>
HOST_IP= <desktop ip addr>
gst-launch-1.0 v4l2src name=cam_src device=${CAMERA} num-buffers=-1 
! video/x-raw,width=640,height=480,framerate=30/1 
! tee name=t t. 
! queue name=thread-nn max-size-buffers=2 leaky=2 ! imxvideoconvert_g2d 
! video/x-raw,width=300,height=300,format=RGBA ! videoconvert 
! video/x-raw,format=RGB ! tensor_converter 
! tensor_filter framework=tensorflow-lite model=/home/root/nxp-nnstreamer-examples/detection/../downloads/models/detection/ssdlite_mobilenet_v2_coco_quant_uint8_float32_no_postprocess.tflite custom=Delegate:External,ExtDelegateLib:libvx_delegate.so 
! tensor_decoder mode=bounding_boxes option1=mobilenet-ssd option2=/home/root/nxp-nnstreamer-examples/detection/../downloads/models/detection/coco_labels_list.txt option3=/home/root/nxp-nnstreamer-examples/detection/../downloads/models/detection/box_priors.txt option4=640:480 option5=300:300 
! videoconvert ! queue 
! mix. t. 
! queue name=thread-img max-size-buffers=2 leaky=2 ! videoconvert 
! mix. imxcompositor_g2d name=mix latency=30000000 min-upstream-latency=30000000 sink_0::zorder=2 sink_1::zorder=1 
! videoconvert ! jpegenc ! rtpjpegpay ! udpsink host=${HOST_IP} port=5000

GW74xx AI-detection GStreamer pipeline (SVG)

If everything works properly, you should instantly see your video input streamed to your desktop host. After a few seconds of warming up, the bounding boxes from the ​ TensorFlow Detection filter will be overlaid on the video. The stream properties can be changed for different resolutions and framerates; see ​ gstreamer/streaming. NOTE: This example is object detection, which differs from the image classification that we got benchmark data from in the previous section.

Upstream / Mainline NPU support

The IMX8MP has a Verisilicon NPU which is from Vivante. Vivante supports this via their vivante driver (/dev/galcore) but that is not an opensource driver so it would make sense to want to support the NPU with it's opensource equivalent (etnaviv driver). Support for this has been added to the Linux 6.10 kernel.

The userspace library used for GPU is mesa so it makes sense to support the NPU there as well and for hardware accelleration this is done in mesa with 'delegate' libraries. While Support for the VeriSilicon NPU made it into Mesa 24.1.0 via teflon it does not yest support the imx8mp which has a sightly newer version of the VeriSilicon NPU.

Therefore to support the NPU you need to build a custom fork of mesa that is being worked on by ​Tomeu Vizoso with the work being sponsored by ​Ideas On board.

On a Gateworks Venice board with an IMX8MP and an Ubuntu root filesystem:

• update kernel to v6.10 with NPU support

  # update kernel
  cd /tmp && wget https://dev.gateworks.com/venice/kernel/linux-venice-6.10.6.tar.xz && tar -C / -xvf linux*.tar.xz --keep-directory-symlink && reboot
  

• create a non-root user (always a good idea but not required for mesa)

  # create user
  USER=gateworks
  useradd -s /bin/bash $USER
  mkdir /home/$USER; chown $USER /home/$USER
  # allow sudo for this user
  echo "$USER ALL=(ALL) NOPASSWD: ALL" >> /etc/sudoers
  # add render group (needed for some examples)
  usermod -aG render $USER
  # login as this user
  su -l $USER
  

• to avoid a lot of python dependency issues with tensorflow-lite install an older version of python (python3.9) using the ​deadsnakes ppa and create a virtualenv

  sudo add-apt-repository ppa:deadsnakes/ppa
  sudo apt update
  sudo apt install -y python3.9 python3.9-venv
  # create tflite-env virtualenv
  python3.9 -m venv tflite-env
  # activate the venv (repeat or every new shell)
  source tflite-env/bin/activate
  

• build mesa (fork with updated teflon support for imx8mp)

  # install build deps
  sudo apt install -y build-essential git cmake
  sudo apt-get -y build-dep mesa
  # get repo
  git clone https://gitlab.freedesktop.org/tomeu/mesa.git -b etnaviv-imx8mp mesa
  cd mesa
  # mesa requires meson >= 1.1.0 (newer than jammy's), pycparser >= 2.20, and mako
  pip3 install meson pycparser mako
  meson setup build -Dgallium-drivers=etnaviv -Dvulkan-drivers= -Dteflon=true
  meson compile -C build # 20 mins or so on imx8mp
  ldd build/src/gallium/targets/teflon/libteflon.so
  

• install tensorflow lite runtime

  # install tensorflow lite runtime
  pip3 install tflite_runtime
  # clone tensorflow for some examples and assets used below
  cd
  git clone https://github.com/tensorflow/tensorflow.git
  

• run teflon image classification test

  # note this test requires python pillow, python numpy<2.0 and write access to /dev/dri/renderD128
  pip3 install "numpy<2.0" pillow
  groups # make sure your in the render group (or have write access to /dev/dri/renderD128)
  # without teflon (175.335ms on imx8mp)
  python3 ~/mesa/src/gallium/frontends/teflon/tests/classification.py \
   -i ~/tensorflow/tensorflow/lite/examples/label_image/testdata/grace_hopper.bmp \
   -m ~/mesa/src/gallium/targets/teflon/tests/mobilenet_v1_1.0_224_quant.tflite \
   -l ~/mesa/src/gallium/frontends/teflon/tests/labels_mobilenet_quant_v1_224.txt
  # with teflon (7.651ms on imx8mp)
  python3 ~/mesa/src/gallium/frontends/teflon/tests/classification.py \
   -i ~/tensorflow/tensorflow/lite/examples/label_image/testdata/grace_hopper.bmp \
   -m ~/mesa/src/gallium/targets/teflon/tests/mobilenet_v1_1.0_224_quant.tflite \
   -l ~/mesa/src/gallium/frontends/teflon/tests/labels_mobilenet_quant_v1_224.txt \
   -e ~/mesa/build/src/gallium/targets/teflon/libteflon.so
  

• run tensorflow label_image example:

  # we need to patch the label_image example to use tflite_runtime
  cd ~/tensorflow
  cat <<EOF |
  diff --git a/tensorflow/lite/examples/python/label_image.py b/tensorflow/lite/examples/python/label_image.py
  index d26454f921f..08c65962bf1 100644
  --- a/tensorflow/lite/examples/python/label_image.py
  +++ b/tensorflow/lite/examples/python/label_image.py
  @@ -19,7 +19,7 @@ import time
   
   import numpy as np
   from PIL import Image
  -import tensorflow as tf
  +import tflite_runtime.interpreter as tflite
   
   
   def load_labels(filename):
  @@ -85,7 +85,7 @@ if __name__ == '__main__':
           tflite.load_delegate(args.ext_delegate, ext_delegate_options)
       ]
   
  -  interpreter = tf.lite.Interpreter(
  +  interpreter = tflite.Interpreter(
         model_path=args.model_file,
         experimental_delegates=ext_delegate,
         num_threads=args.num_threads)
  
  EOF
  patch -p1
  # without acceleration (175.993ms on imx8mp)
  python3 ~/tensorflow/tensorflow/lite/examples/python/label_image.py \
    -m ~/mesa/src/gallium/targets/teflon/tests/mobilenet_v1_1.0_224_quant.tflite \
    -l ~/mesa/src/gallium/frontends/teflon/tests/labels_mobilenet_quant_v1_224.txt \
    -i ~/tensorflow/tensorflow/lite/examples/label_image/testdata/grace_hopper.bmp
  # with acceleration (14.138ms on imx8mp)
  python3 ~/tensorflow/tensorflow/lite/examples/python/label_image.py \
   -m ~/mesa/src/gallium/targets/teflon/tests/mobilenet_v1_1.0_224_quant.tflite \
   -l ~/mesa/src/gallium/frontends/teflon/tests/labels_mobilenet_quant_v1_224.txt \
   -i ~/tensorflow/tensorflow/lite/examples/label_image/testdata/grace_hopper.bmp \
   -e ~/mesa/build/src/gallium/targets/teflon/libteflon.so