Vitis-AI

定制开发板（zynq7020-VART2.5）

参考手册

Vitis-AI-手册-设置定制开发板

具体步骤

VIVADO 工程

在 VIVADO 工程中完成项目创建，获得 XSA 文件

Petalinux 工程 Version：2022.1

创建工程

petalinux-create --type project --template zynq --name zynq7020-dpu

进入项目目录，将 XSA 文拷贝至该目录并进行基本配置

petalinux-config --get-hw-description=./

完成之后会生成基本项目文件，下一步进行定制化配置

拷贝 Vitis-AI github项目中的 recipes-vitis-ai 至 zynq7020-dpu 中的 meta-user

cp <vitis-ai>/setup/petalinux/recipes-vitis-ai <petalinux project>/project-spec/meta-user/

编辑 <petalinux project>/project-spec/meta-user/conf/user-rootfsconfig 文件，附加内容

CONFIG_vitis-ai-library
CONFIG_vitis-ai-library-dev
CONFIG_vitis-ai-library-dbg

配置 rootfs

Select user packages --->
Select [*] vitis-ai-library

执行构建

注意：构建前需保证磁盘可用存储大于40GB

petalinux-build

完成构建

$ petalinux-build 
[INFO] Sourcing buildtools
[INFO] Building project
[INFO] Sourcing build environment
[INFO] Generating workspace directory
INFO: bitbake petalinux-image-minimal
NOTE: Started PRServer with DBfile: /home/pynq/Desktop/pynqz-dpu/build/cache/prserv.sqlite3, Address: 127.0.0.1:40323, PID: 81113
Loading cache: 100% |###############################################################################################################################################| Time: 0:00:00
Loaded 5397 entries from dependency cache.
Parsing recipes: 100% |#############################################################################################################################################| Time: 0:00:00
Parsing of 3597 .bb files complete (3595 cached, 2 parsed). 5399 targets, 554 skipped, 0 masked, 0 errors.
NOTE: Resolving any missing task queue dependencies
Initialising tasks: 100% |##########################################################################################################################################| Time: 0:00:02
Checking sstate mirror object availability: 100% |##################################################################################################################| Time: 0:01:20
Sstate summary: Wanted 675 Local 13 Network 425 Missed 237 Current 1638 (64% match, 89% complete)
NOTE: Executing Tasks
Currently  2 running tasks (2313 of 2313/6032 of 6051)  99% |####################################################################################################NOTE: Tasks Summary: Attempted 6051 tasks of which 6026 didn't need to be rerun and all succeeded.
INFO: Failed to copy built images to tftp dir: /tftpboot
[INFO] Successfully built project

FAQ

vart 编译 test 出现错误

/Vitis-AI-Runtime/VART/vart/xrt-device-handle/test/test_ecmd.cpp:200:24: error: unsigned conversion from 'long long int' to 'long unsigned int' changes value from '824633724928' to '4096' [-Werror=overflow]
|   200 |   unsigned long addr = 0xc000001000;
|       |                        ^~~~~~~~~~~~

问题原因：由于该开发板是 ARM32 构架，这里出现数据溢出问题

解决方案：修改数据类型即可解决问题

<petalinux project>/build/tmp/work/cortexa9t2hf-neon-xilinx-linux-gnueabi/vart/2.5-r0/git/src/Vitis-AI-Runtime/VART/vart/xrt-device-handle/test/test_ecmd.cpp

|   200 |   unsigned long long addr = 0xc000001000;

Vitis-AI-Library 中图像处理相关错误

<petalinux project>/build/tmp/work/cortexa9t2hf-neon-xilinx-linux-gnueabi/vitis-ai-library/2.5-r0/git/src/Vitis-AI-Library/math/src/image_util.cpp

|  4403 |   return (uint8x16_t)__builtin_neon_vshru_nv16qi ((int8x16_t) __a, __b);

问题原因：由于 ARM NEON 指令集的限制 引起的

解决方案：针对该开发板处理器架构重新编写代码

<petalinux project>/build/tmp/work/cortexa9t2hf-neon-xilinx-linux-gnueabi/vitis-ai-library/2.5-r0/git/src/Vitis-AI-Library/math/src/image_util.cpp:void transform_mean_scale()

auto calc_float_mean = [](const uint8x16_t& in_u8, int scale,
                            const uint8x16_t& high_mean,
                            const uint8x16_t& low_mean) -> int8x16_t {
    else if (scale < 0) {
      // scale < 0
    /*
    return vreinterpretq_s8_u8( vsubq_u8(vshrq_n_u8(vqsubq_u8(in_u8, high_mean), -scale),
                   vshrq_n_u8(vqsubq_u8(low_mean, in_u8), -scale)));
    */
    return vreinterpretq_s8_u8(
        vsubq_u8(
            vshlq_u8(vqsubq_u8(in_u8, high_mean), vdupq_n_s8(-scale)),
            vshlq_u8(vqsubq_u8(low_mean, in_u8), vdupq_n_s8(-scale))
        )
    );
    }
  };
  auto calc_int_mean = [](const uint8x16_t& in_u8, int scale,
                          const uint8x16_t& mean) -> int8x16_t {
   else if (scale < 0) {
      // scale < 0
      return vreinterpretq_s8_u8(
        vsubq_u8(
            vshlq_u8(vqsubq_u8(in_u8, mean), vdupq_n_s8(-scale)),
            vshlq_u8(vqsubq_u8(mean, in_u8), vdupq_n_s8(-scale))
        )
    );
    }
  };

这里仅展示修改部分的代码，主要目的是将代码匹配对应的处理器架构

Vitis-AI-Library/math/test/ 中也有相似的一处错误，根据 src 中的修改思路修改代码即可

SiuKiri / Vitis-AI-Development-Records

Vitis-AI #1

Vitis-AI

定制开发板（zynq7020-VART2.5）