526互联

RK1808上基于PaddleLite调用NPU

发布时间 2023-06-20 17:02:49作者: SheepHuan

1 复现PaddleLite v2.11 demo

为什么复现2.11的版本呢？因为2.12的复现不出来.

2.1 准备

  1. 设备: 九鼎创展 x1808开发板. 默认ssh账户root,密码123456
    • 刷机要点: 按住K1键，然后板子上电，进入LOADER模式，使用瑞芯微提供的AndroidTools升级固件即可。
  2. 软件:
    • uname -a : Linux root 4.4.194 #7 SMP PREEMPT Fri Dec 11 12:06:00 CST 2020 aarch64 GNU/Linux
    • dmesg | grep Galcore | grep version : [ 21.137075] Galcore version 6.4.0.227915
  3. 可执行文件:

2.2 实现

  1. 解压tar.gz
     tar -zxvf PaddleLite-generic-demo_v2_11_0.tar.gz
 tar -zxvf mobilenet_v1_int8_224_per_layer.tar.gz
  2. 将 mobilenet_v1_int8_224_per_layer/放入PaddleLite-generic-demo/image_classification_demo/assets/models
  3. 由于磁盘空间的问题，我在开发板上挂载了一个tf card.
    # 查看磁盘分区
fdisk -l
# 对指定分区进行格式化
mkfs .ext4 /dev/mmcblk1p1
# 创建挂载目录
mkdir -p /mnt/tfcard
# 挂载磁盘
mount /dev/mmcblk1p1 /mnt/tfcard
  4. 修改PaddleLite-generic-demo/image_classification_demo/shell中的run_with_ssh.sh中的工作目录
    ...
# line 25
WORK_SPACE="/mnt/tfcard/test"
...
  5. 执行脚本命令
    • NPU
      # ./run_with_ssh.sh <model dir> <OS> <arch> <device> <ip> <port> <username> <password>
# 调用NPU执行
./run_with_ssh.sh mobilenet_v1_int8_224_per_layer linux arm64 rockchip_npu 172.16.6.98 22 root 123456
      结果
      ...
iter 4 cost: 6.769000 ms
warmup: 1 repeat: 5, average: 6.729200 ms, max: 6.775000 ms, min: 6.679000 ms
results: 3
Top0  Egyptian cat - 0.514779
Top1  tabby, tabby cat - 0.421183
Top2  tiger cat - 0.052648
Preprocess time: 2.457000 ms
Prediction time: 6.729200 ms
Postprocess time: 0.450000 ms
...
    • CPU
      ./run_with_ssh.sh mobilenet_v1_int8_224_per_layer linux arm64 cpu 172.16.6.98 22 root 123456
      结果
      iter 0 cost: 266.263000 ms
iter 1 cost: 266.100006 ms
iter 2 cost: 265.746002 ms
iter 3 cost: 265.558014 ms
iter 4 cost: 265.686005 ms
warmup: 1 repeat: 5, average: 265.870605 ms, max: 266.263000 ms, min: 265.558014 ms
results: 3
Top0  Egyptian cat - 0.512545
Top1  tabby, tabby cat - 0.402567
Top2  tiger cat - 0.067904
Preprocess time: 2.487000 ms
Prediction time: 265.870605 ms
Postprocess time: 0.430000 ms