CUDA 10.0 cross compilation ppc64le

Accelerated Computing CUDA CUDA Setup and Installation
1

hi folks,

i need to cross compile cuda 10.0 base code to ppc64le architecture(host amd64). Do you know where i can find cuda-cross-ppc64le library? Anyone tried ppc64le cross-compilation with cmake?

2

Hi Jan,

Have this same question myself now. Did you every get an answer?

In general: is there a canonical way to cross-compile CUDA 10+ for PowerPC architectures from an x86 host?

3

Hi mafanasiev,

i overcome my problems with creation of docker image for compilation. i got access to PowerPC machine so i compiled there using prepared container.

I think it could be done without PowerPC machine with QEMU virtualization for docker on your pc, so you would not need to find cross compilation libs. You could run compilation environment container on your host pc with virtualization of PowerPC arch.

Look into :

4

Hi Jan,

Thanks for the reply! In the end we went for a different approach which seems to work. Specifically, starting from the quay.io/pypa/manylinux2014_x86_64 image:

  1. Download the CUDA runfile installer for the x86 host (e.g. this one for CUDA 10.0).
  2. Download the IBM Advance Toolchain for Linux.
  3. Download and unpack the cudart-dev libraries for Debian:
curl -L https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/ppc64el/cuda-cudart-dev-10-0_10.0.130-1_ppc64el.deb -o cuda-cudart-dev-10-0_10.0.130-1_ppc64el.deb
ar x cuda-cudart-dev-10-0_10.0.130-1_ppc64el.deb && tar xf data.tar.xz
  1. Run nvcc as follows:
nvcc -ccbin=/opt/at7.1/bin/powerpc64le-linux-gnu-c++ -I/usr/local/cuda/include

Does not seem canonical, but it works. Hope this helps you or others.

Cheers,
Mike.