search

Xilinx / Vitis-AI

Vitis AI is Xilinx’s development stack for AI inference on Xilinx hardware platforms, including both edge devices and Alveo cards.
https://www.xilinx.com/ai
Apache License 2.0
1.49k stars 628 forks source link

Wrong environment variables setup for vitis-ai docker container #604

Closed nhphuong91 closed 2 years ago

nhphuong91 commented 2 years ago

Hi, I found a bug with setup/alveo/setup.sh. When following instruction to setup environment (for Alveo U50) after entering docker container, I notice that the last 2 env variables are wrong:

Vitis-AI /workspace/setup/alveo > source setup.sh DPUCAHX8H
.
.
---------------------
XCLBIN_PATH = /opt/xilinx/overlaybins/DPUCAHX8H
XLNX_VART_FIRMWARE = /opt/xilinx/overlaybins/DPUCAHX8H/dpuv3e_6E300_xilinx_u50_gen3x4_xdma_base_2.xclbin
---------------------

The structure of /opt/xilinx/overlaybins/ has changed as below:

Vitis-AI /workspace/setup/alveo > tree /opt/xilinx/overlaybins/
/opt/xilinx/overlaybins/
├── dpuv3e
│   ├── dpuv3e_10E275_xilinx_u50lv_gen3x4_xdma_base_2.xclbin
│   ├── dpuv3e_14E300_xilinx_u280_xdma_201920_3.xclbin
│   ├── dpuv3e_6E300_xilinx_u50_gen3x4_xdma_base_2.xclbin
│   ├── md5sum_v3e.txt
│   └── waa
│       ├── classification
│       │   └── dpuv3e_pre.hw.xilinx_u50_gen3x4_xdma_2_202010_1.xclbin
│       └── detection
│           └── dpuv3e_pre.hw.xilinx_u50_gen3x4_xdma_2_202010_1.xclbin
├── dpuv3me
│   ├── dpuv3me_1E250_xilinx_u50lv_gen3x4_xdma_base_2.xclbin
│   ├── dpuv3me_1E333_xilinx_u50_gen3x4_xdma_base_2.xclbin
│   ├── dpuv3me_2E250_xilinx_u280_xdma_201920_3.xclbin
│   ├── md5sum_v3me.txt
│   └── waa
│       └── detection
│           └── dpuv3me_pre_post.hw.xilinx_u280_xdma_201920_3.xclbin
├── xclbin
│   └── my_system.xclbin
└── xdnnv3
    ├── xdnn_v3_96x16_2pe_8b_9mb_bank03_2.xclbin
    ├── xdnn_v3_96x16_2pe_8b_9mb_bank03_2.xclbin.json
    ├── xdnn_v3_96x16_2pe_8b_9mb_bank03.xclbin
    ├── xdnn_v3_96x16_2pe_8b_9mb_bank03.xclbin.json
    ├── xdnn_v3_96x16_4pe_8b_9mb_bank0123_2.xclbin
    ├── xdnn_v3_96x16_4pe_8b_9mb_bank0123_2.xclbin.json
    ├── xdnn_v3_96x16_4pe_8b_9mb_bank0123.xclbin
    ├── xdnn_v3_96x16_4pe_8b_9mb_bank0123.xclbin.json
    ├── xdnnv3-aws-v3.2.xclbin
    ├── xdnnv3-aws-v3.2.xclbin.json
    └── xplusml
        ├── detection_preproc
        │   ├── with_jpeg_decoder
        │   │   └── xdnn_v3_96x16_2pe_8b_9mb_bank03.xclbin
        │   └── wo_jpeg_decoder
        │       └── xdnn_v3_96x16_2pe_8b_9mb_bank03.xclbin
        └── optical_flow
            └── dpu_lk_dense_non_pyr_optical_flow_U200.xclbin

14 directories, 25 files

Can anyone check and fix it? Thanks!

bryanloz-xilinx commented 2 years ago

The tar ball used to install the xclbins should create the subdirectories named as DPUCAHX8H, and similar.

what tar.gz file did you use?

nhphuong91 commented 2 years ago

@bryanloz-xilinx I simply follow the instruction here (after started the docker): https://github.com/Xilinx/Vitis-AI/tree/v1.4.1/setup/alveo#dpu-ip-selection

cd /workspace/setup/alveo
source setup.sh DPUCAHX8H

That's all. No download needed.

hanxue commented 2 years ago

Hi @nhphuong91 ,

The /opt/xilinx/overlaybins directory inside Vitis AI CPU/GPU Docker is mapped from the host's directory -

Vitis-AI /workspace > tree -L 2 /opt/xilinx/overlaybins/
/opt/xilinx/overlaybins/
├── DPUCAHX8H
│   ├── dpu_DPUCAHX8H_10PE275_xilinx_u50lv_gen3x4_xdma_base_2.xclbin
│   ├── dwc
│   ├── md5sum_DPUCAHX8H.txt
│   └── waa
└── DPUCVDX8H
    ├── 6pedwc
    └── 8pe

6 directories, 2 files

Perhaps you are using an older version of the Alveo setup.sh script?

The v2.0 release script works fine

(vitis-ai-pytorch) Vitis-AI /workspace/setup/alveo > source setup.sh DPUCAHX8H
------------------
VAI_HOME = /vitis_ai_home
------------------
XILINX_XRT        : /opt/xilinx/xrt
PATH              : /opt/xilinx/xrt/bin:/opt/xilinx/xrm/bin:/opt/xilinx/xrt/bin:/opt/vitis_ai/conda/envs/vitis-ai-pytorch/bin:/opt/vitis_ai/conda/bin:/opt/vitis_ai/conda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
LD_LIBRARY_PATH   : /opt/xilinx/xrt/lib:/opt/xilinx/xrm/lib:/opt/xilinx/xrt/lib:/opt/xilinx/xrt/lib:/usr/lib:/usr/lib/x86_64-linux-gnu:/usr/local/lib:/opt/vitis_ai/conda/envs/vitis-ai-tensorflow/lib
PYTHONPATH        : /opt/xilinx/xrt/python:/opt/xilinx/xrt/python:
---------------------
XILINX_XRT = /opt/xilinx/xrt
---------------------
XILINX_XRM      : /opt/xilinx/xrm
PATH            : /opt/xilinx/xrm/bin:/opt/xilinx/xrt/bin:/opt/xilinx/xrm/bin:/opt/xilinx/xrt/bin:/opt/vitis_ai/conda/envs/vitis-ai-pytorch/bin:/opt/vitis_ai/conda/bin:/opt/vitis_ai/conda/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
LD_LIBRARY_PATH : /opt/xilinx/xrm/lib:/opt/xilinx/xrt/lib:/opt/xilinx/xrm/lib:/opt/xilinx/xrt/lib:/opt/xilinx/xrt/lib:/usr/lib:/usr/lib/x86_64-linux-gnu:/usr/local/lib:/opt/vitis_ai/conda/envs/vitis-ai-tensorflow/lib
---------------------
XILINX_XRM = /opt/xilinx/xrm
---------------------
---------------------
LD_LIBRARY_PATH = /opt/xilinx/xrm/lib:/opt/xilinx/xrt/lib:/opt/xilinx/xrm/lib:/opt/xilinx/xrt/lib:/opt/xilinx/xrt/lib:/usr/lib:/usr/lib/x86_64-linux-gnu:/usr/local/lib:/opt/vitis_ai/conda/envs/vitis-ai-tensorflow/lib
---------------------
  [0000:3b:00.1] : xilinx_u50lv_gen3x4_xdma_base_2 user(inst=128) 
u50lv_ card detected
---------------------
XCLBIN_PATH = /opt/xilinx/overlaybins/DPUCAHX8H
XLNX_VART_FIRMWARE = /opt/xilinx/overlaybins/DPUCAHX8H/dpu_DPUCAHX8H_10PE275_xilinx_u50lv_gen3x4_xdma_base_2.xclbin
---------------------

https://github.com/Xilinx/Vitis-AI/tree/v2.0

nhphuong91 commented 2 years ago

@hanxue, I'm using Alveo U50 now, but it seems VAI 2.0 doesn't support it anymore. Cannot find any supported DPU for U50, only U50LV

hanxue commented 2 years ago

Yes, U50LVis supported, U50 not supported in Vitis AI 2.0

nhphuong91 commented 2 years ago

Any plan to support U50 in the future version of VAI or are you planning to abandon it?

hanxue commented 2 years ago

Sorry, I don;t have an answer to U50 support. Will ask colleagues to comment on VAI product roadmap and platform support.

wanghy-xlnx commented 2 years ago

Hi @nhphuong91 ,

If you are running Vitis AI 2.0 on U50 card, then you will need to generate the .xclbin file yourself from .xo file. Therefore, it is recommended that you use Vitis AI 1.4.1 on the U50 card.

Note the following:

  1. You need to checkout the repo https://github.com/Xilinx/Vitis-AI.git to the v1.4.1 branch
  2. You need to use the 1.4.1.978 version of the docker image according to README
  3. Read Quick Start, you need run setup.sh according Alveo Setup
nhphuong91 commented 2 years ago

@wanghy-xlnx I was able to use VAI 1.4.1 on the U50. And the setup script has an issue where I suggest a temporary solution -> all was mentioned in this ticket. I just want to know if you have plan to support U50 in the future version of VAI or are you planning to drop support for U50 in favor of other products such as U50LV?

wanghy-xlnx commented 2 years ago

Hi @nhphuong91 ,

We will continue to support the U50, including U280 and other cards. However, from Vitis AI 2.0, the supported method has changed from pre-compiling to publishing XO files to facilitate more flexible use and integration by users. This approach allows users to customize the resources of the DPU, as well as integrate with other IP.

In Vitis AI 2.0, U50LV, U55C, etc. provide pre-compiled DPUs and models, which is helpful for users to quickly start and see the execution results. However, this usage method is fixed. If you want to customize the DPU resources, or integrate with other IP, you still need to use the XO file to regenerate the DPU's xclbin and the corresponding models.

PS: The XO release URL is https://github.com/Xilinx/Vitis-AI/tree/master/dsa/DPUCAHX8H-XO