Missing header files in src/kernels ?

[xmontagut]

Hello,
When trying to install deconwolf :

[ 61%] Building C object CMakeFiles/dw.dir/src/cl_util.c.o
/home/s_montap01/src/deconwolf-0.3.8/src/cl_util.c:4:10: fatal error: kernels/cl_complex_mul.h: No such file or directory
#include "kernels/cl_complex_mul.h"

In the github repo, there are no .h files under "kernels/", and I cannot find cl_complex_mul.h anywhere. Did I miss something

[tangmc0210]

I met the same problems, too. Expecting to get solved.

[tangmc0210]

[elgw]

Hi!

My apologies, I missed to add info on that.

If you run

make kernels

those files will be generated (if you have xxd installed). I will add the already generated files to the repo in the next update. If you use the dev branch the files are there already.

Cheers,
Erik

[xmontagut]

Thank you for your answer.
It works after modifying the makefile : I try the compilation on a HPC cluster. The "pkg-config OpenCL" was returning an exit status "false".
I had to change "OpenCL" for "opencl-VERSION" to make it work (after loading the same VERSION of NVidia CUDA "module", which is a commonly used tool to configure a user environment for a specific software version in a cluster). As I don't know pkg-config, perhaps there's a better workaround ?
Regards,
Xavier

[elgw]

Hi Xavier,

Thank you for the update. It is great to hear that you got around the problem.

The CMakeList.txt uses the FindOpenCL module which might be more versatile than my manual makefile, did you also test building with cmake?

I have very little experience with HPC clusters and I wouldn't know what to change to facilitate the build procedure. It would be really useful if you could share what system you are using (workload manager, os etc) and the modifications required to build deconwolf, or if you could point to some online resource where the information can be found.

Cheers,
Erik

[xmontagut]

I think there's a little misunderstanding.

I had to modify the "makefile" (tol level directory) in order to generate the kernels header files with "make kernels", but otherwise, for the build, I followed your instructions using CMake (mkdir build && cd build && cmake .. && cmake --build .)

There's just a little trick. For "cmake .." to find OpenCL with "find_package(OpenCL)", as CUDA is not installed in a standard directory in our cluster, here are two workarounds :

• add a symbolic link under "/usr/local/cuda" pointing to the actual location (but you need to be a sysadmin)
• OR add "-DOpenCL_LIBRARY=$CUDADIR/lib -DOpenCL_INCLUDE_DIR=$CUDADIR/include" to "cmake .." - with CUDADIR a variable with the path of the actual CUDA location. In our case, this variable is set when I load the "cuda" module on the cluster, but this is a local configuration.

Cheers,
Xavier

[tangmc0210]

Hi!

My apologies, I missed to add info on that.

If you run

make kernels

those files will be generated (if you have xxd installed). I will add the already generated files to the repo in the next update. If you use the dev branch the files are there already.

Cheers, Erik

It works well! Thanks

[elgw]

Thanks to both of you.

This is fixed in version 0.4.0 so I'm closing this issue for now.

I added a link to Xavier's comment in the INSTALL.md since I think that it can be useful in the future.