PyTorch DNNL_aarch64 build manual for FUJITSU Software Compiler Package (PyTorch v1.6.0)

Requirements

• FUJITSU Software Compiler Package is already installed.
• The login node has access to the external network.

Verified environment

Hardware	Fujitsu FX1000 / FX700
OS	RedHad 8.1 / Centos 8.1
Compiler	FUJITSU Software Compiler Package

Preparation

1. Checkout from Repository.

 $ git clone https://github.com/fujitsu/pytorch.git
 $ cd pytorch
 $ git checkout -b fujitsu_v1.6.0_for_a64fx origin/fujitsu_v1.6.0_for_a64fx

2. Environment Setting

 $ cd scripts/fujitsu

 Modify the following environment variables in "env.src".
 ################################################
 ## Please change the following to suit your environment.
 ## PREFIX    : The directory where this file is located.
 ## TCSDS_PATH: TCS installation path
 ################################################
 export PREFIX=/home/users/ai/ai0003/pytorch/scripts/fujitsu
 export TCSDS_PATH=/opt/FJSVxtclanga/tcsds-1.2.26

  Modify each batch files to suit your environment.
 #!/bin/bash
 #PJM -L "rscunit=rscunit_ft01,rscgrp=ai-default"
 #PJM -L elapse=04:00:00
 #PJM -L "node=1"
 #PJM -j
 #PJM -S

Build PyTorch

Login Node[Estimated time:4h]

Download

Run checkout script

 $ ./checkout.sh

Build binary files for A64FX.

Run build-script on compute node.

 $ pjsub submit_build.sh

Check the environment

Run the sample programs

 $ pjsub submit_train.sh
 $ pjsub submit_train_multi.sh
 $ pjsub submit_val.sh
 $ pjsub submit_val_multi.sh

   Example of output(submit_train_multi.sh.xxx.out)
    ～
    Model: resnet50
    Batch size: 75
    Number of CPUs: 4
    Running warmup...
    Running benchmark...
    Iter #0: 26.1 img/sec per CPU
    Iter #1: 26.1 img/sec per CPU
    Iter #2: 26.1 img/sec per CPU
    Iter #3: 26.1 img/sec per CPU
    Iter #4: 26.1 img/sec per CPU
    Img/sec per CPU: 26.1 +-0.0
    Total img/sec on 4 CPU(s): 104.5 +-0.1
    ～

Build for OpenNMT

 $ cd opennmt_build_pack

Environment Setting

Modify each batch files to suit your environment.

• submit_build.sh
• opennmt-py/submit_prepare.sh
• opennmt-py/submit_opennmt.sh

Download

 $ ./dataset.sh
 $ ./checkout.sh

Build [Estimated time:1.5h]

 $ pjsub submit_build.sh

Preprocessing for dataset [Estimated time:1h]

 $ cd opennmt-py
 $ pjsub submit_prepare.sh

Check the environment

 $ pjsub submit_opennmt.sh

Copyright

Copyright FUJITSU LIMITED 2020