Crosscompilation for arm64

Peter_Peter · November 13, 2017, 5:43pm

I’m looking for a way to compile (install) pytorch for arm64 platform and as of now I’m not interesting of optimized (GPU-like) version of pytorch, it’s good enough to have CPU version able to run on ARM device. Please can you advice how can I either compile or install such a version of pytorch?
In my case I do not have CUDA available on device but instead I have OpenCL 1.2 FP.

SimonW · November 13, 2017, 6:50pm

I think you can just compile with NO_CUDA=1. Could you try it and see if there is any error?

Peter_Peter · November 13, 2017, 7:40pm

I would love too, but is there install dependencies ? Since my platfrom ti’s kind of minimal on packages …

SimonW · November 13, 2017, 8:00pm

Yes, you would need numpy pyyaml mkl setuptools cmake cffi

Peter_Peter · November 14, 2017, 3:38pm

Is it mkl library for arm architecture? Isn’t Intel lib?

SimonW · November 14, 2017, 5:36pm

Right, you can skip that. Sorry I just copied + pasted from Github install page without checking.

NikhilKanda · December 26, 2018, 7:19am

I’m wondering to know if there’s a way to cross compile Pytorch on Allwinner platform (AArch64 processor). Leads are greatly appreciated.

Also @Peter_Peter, I presume that you’ve compiled or installed pytorch on arm64 by now. Could you please post the steps you have followed for the same?

TIA

smth · December 28, 2018, 11:28pm

@NikhilKanda this is the script I used to compile on AGX Xavier which is an ARM64 platform: https://gist.github.com/soumith/45f61db3379db67414814ad9d5f2a427

NikhilKanda · January 2, 2019, 2:07pm

Hi @smth. Thanks for the info, but I hope that is not cross compilation. Correct me if I’m wrong

smth · January 3, 2019, 1:11am

oh, sorry yes this is not cross-compilation.

I dont think all parts of our build system are even ready for cross-compilation

dinesh_mareedu · May 28, 2019, 11:09am

Can you give me any information about cross compilation of pytorch …

solvingPuzzles · January 6, 2020, 2:54am

I haven’t tried this, but the following tutorial looks relevant:

Tumb1eweed · June 22, 2022, 6:34am

met the same problem, need to cross compile libtorch for arm64 in x86-64 platform, is there any resources about this?

Tchamp · June 27, 2022, 3:13pm

To build libtorch you can follow these instructions:

github.com

pytorch/pytorch/blob/master/docs/libtorch.rst

libtorch (C++-only)
===================

The core of pytorch does not depend on Python. A
CMake-based build system compiles the C++ source code into a shared
object, libtorch.so.

Building libtorch using Python
------------------------------

You can use a python script/module located in tools package to build libtorch
::
   cd <pytorch_root>

   # Make a new folder to build in to avoid polluting the source directories
   mkdir build_libtorch && cd build_libtorch

   # You might need to export some required environment variables here.
   Normally setup.py sets good default env variables, but you'll have to do
   that manually.

This file has been truncated. show original

Personally, I only use CMake for cross-compiling and add a -DCMAKE_TOOLCHAIN_FILE=<path to your cmake toolchain> argument

If you have problems with binaries like mkrename mkdisp mkalias or addSuffix, you will have to compile sleef with your x86_64 compiler and then add for libtorch the cmake argument : -DNATIVE_BUILD_DIR=<path to native sleef build>.

Jesse_Stone · August 1, 2022, 7:43pm

full build step cross compile on arm64

mavs · September 6, 2023, 6:50am

Hi, i wonder if the build system is ready for cross-compilation now, and my terget is Jetson TX2, i’d like to cross compiling libtorch, thx.

mavs · September 6, 2023, 6:59am

Hi, i followed the steps but always encounter the cmake warning ’ Could Not find CUDA(missing:CUDA_INCLUDE_DIRS CUDA_CUDART_LIBRARY)'. Then the USE_CUDA was forced to be OFF, and the output of cross compilation was just libtorch_cpu.so and so on, but no libtorch_cuda.so. Could you give more cross compilation information, thx.