Cross compile OpenCV with CUDA on x86-64 host for nvidia tk1 (arm) target using Yocto Project?

Question

I want to build a custom Linux using the Yocto Project for a system using the nvidia tegra k1 processor, which has an arm cortex A 15 (32bit) and a GPU. Furthermore, OpenCV (version >= 3.0) with CUDA support needs to be installed.

Now when Bitbake tries to compile OpenCV with CUDA it crashes because nvcc throws an error Syntax error: word unexpected (expecting ")") which I learned is usually the case if you try to run an nvcc binary that is for the wrong system (e.g. 32bit-nvcc on 64bit-system). So I guess it just runs the nvcc that is suppost for the target system and therefore crashes.

I don't know whether I need to compile using the usual x86-64bit-nvcc of the host system or (if this exists) some cross-compilation-nvcc.
I need to know how I can involve said compiler into bitbake's process that builds OpenCV.

can you use openCV4tegra? Afaik it is much better optimized than plain cuda opencv — Micka
– Micka, Commented Nov 3, 2016 at 21:47
OpenCV4tegra is not an option since an OpenCV version of 3.0 or higher is required. Afaik OpenCV4tegra releases are all 2.4.X. — PeterErnsthaft
– PeterErnsthaft, Commented Nov 4, 2016 at 10:39
Have a look at the 2nd version in elinux.org/Jetson/Installing_OpenCV on how to compile openCV for tk1. — Micka
– Micka, Commented Nov 4, 2016 at 10:45
There is no actual information on cross-compiling at that page except for what it means or am I missing something? — PeterErnsthaft
– PeterErnsthaft, Commented Nov 4, 2016 at 11:00
so your question is about how cross compiling can be used at all? — Micka
– Micka, Commented Nov 4, 2016 at 11:03

PeterErnsthaft · Accepted Answer · 2016-11-07 15:17:35Z

1

I figured it out:

One needs to use an nvcc binary that supports the host architecture. To specify the target architecture on can pass it cross compilation flags (see http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/#cross-compilation). These are passed via the the cmake flag -DCUDA_NVCC_FLAGS
EXTRA_OECMAKE_append = "-DCUDA_NVCC_EXECUTABLE=${PATH_TO_THE_NVCC_BINARY}"

The whole statement in the bbappend file could look like this

EXTRA_OECMAKE_append = " \
    -DCUDA_NVCC_EXECUTABLE=${NVCC_BINARY} \ 
    -DCUDA_NVCC_FLAGS="--compiler-bindir ${GCC_BINARY}"  \
    "

answered Nov 7, 2016 at 15:17

PeterErnsthaft

3956 silver badges14 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Cross compile OpenCV with CUDA on x86-64 host for nvidia tk1 (arm) target using Yocto Project?

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related