How do I build GDAL with OpenCL against a custom CUDA installation?

Question

I am trying to build GDAL against OpenCL to get GPU accelerated raster operations, depending on Cuda. I managed to get ./configure to complete and start compiling with make. However, when compiling gdalinfo, the compilation fails because it can't find the OpenCL symbols.

I'm pretty sure it is because a -lOpenCL flag is missing somewhere, but (1) I don't know where; and (2) I don't know what to change. So my question is how to correctly configure the build commands to use OpenCL?

The build failure error:

echo '#!/bin/sh' > gdal-config
echo 'CONFIG_LIBS="libgdal.la"' >> gdal-config
echo 'CONFIG_DEP_LIBS=" -lcrypto -lpoppler -ljson-c -lfreexl -lqhull -lqhull -L/usr/lib/x86_64-linux-gnu -lgeos_c -lwebp -lodbc -lodbcinst -lkmlbase -lkmldom -lkmlengine -lkmlxsd -lkmlregionator -lexpat -lxerces-c -lpthread -lopenjp2 -ljasper -L/usr/lib -L/usr/lib/x86_64-linux-gnu/hdf5/serial -lnetcdf -lhdf5_hl -lhdf5 -lpthread -lsz -lz -ldl -lm -lcurl -L/usr/lib/x86_64-linux-gnu/hdf5/serial -lhdf5 -lmfhdfalt -ldfalt -logdi -lgif -ljpeg -lgeotiff -ltiff -lpng -L/usr/lib/x86_64-linux-gnu -lpq -lzstd -llzma -lz -lpthread -lm -lrt -ldl  -lspatialite -lsqlite3 -lpcre   -L/usr/lib/x86_64-linux-gnu -lcurl        cuda    -lxml2     "' >> gdal-config
echo 'CONFIG_PREFIX="gdal-2.4.2"' >> gdal-config
echo 'CONFIG_CFLAGS="-Iport -Igcore -Ialg -Iogr -Iogr/ogrsf_frmts -Ifrmts/vrt -Iapps"' >> gdal-config
echo 'CONFIG_DATA="data"' >> gdal-config
echo 'CONFIG_VERSION="'`cat ../VERSION`'"' >> gdal-config
echo 'CONFIG_OGR_ENABLED=yes' >> gdal-config  # Deprecated.  Always true.
echo 'CONFIG_GNM_ENABLED=yes' >> gdal-config
echo 'CONFIG_FORMATS=" gxf gtiff hfa aigrid aaigrid ceos ceos2 iso8211 xpm sdts raw dted mem jdem envisat elas fit vrt usgsdem l1b nitf bmp airsar rs2 ilwis rmf leveller sgi srtmhgt idrisi gsg ingr ers jaxapalsar dimap gff cosar pds adrg coasp tsx terragen blx msgn til r northwood saga xyz hf2 kmlsuperoverlay ctg e00grid zmap ngsgeoid iris map cals safe sentinel2 derived prf sigdem ignfheightasciigrid mrf webp wcs wms plmosaic wmts rda eeda grib bsb openjpeg jpeg2000 netcdf hdf5 hdf4 gif jpeg png pcraster  pcidsk rik ozi pdf rasterlite mbtiles postgisraster arg"' >> gdal-config
cat gdal-config.in >> gdal-config
chmod a+x gdal-config
/bin/bash libtool --mode=compile --silent --tag=CXX g++ -std=c++11 -Iport -Igcore -Ialg -Iogr -Iogr/ogrsf_frmts -Ignm -Iapps -DHAVE_AVX_AT_COMPILE_TIME -DHAVE_SSSE3_AT_COMPILE_TIME -DHAVE_SSE_AT_COMPILE_TIME -g -O2  -Wall -Wextra -Winit-self -Wunused-parameter -Wformat -Werror=format-security -Wno-format-nonliteral -Wlogical-op -Wshadow -Werror=vla -Wdate-time -Wfloat-conversion -Wmissing-declarations -Wnon-virtual-dtor -Woverloaded-virtual -fno-operator-names -Wzero-as-null-pointer-constant -Wsuggest-override  -Ifrmts/vrt -DGNM_ENABLED -Iport -I/usr/include/openjpeg-2.1 -I/usr/include  -DGDAL_COMPILATION -DHAVE_XERCES -I/usr/include -I/usr/include/xercesc -Iport -I/usr/include/openjpeg-2.1 -I/usr/include  -DGDAL_COMPILATION -I/usr/include/json-c -Iogr/ogrsf_frmts/geojson -Iogr/ogrsf_frmts/generic -Ignm -DHAVE_GEOS=1 -I/usr/include -c -o gdalinfo_bin.lo gdalinfo_bin.cpp
/bin/bash libtool --mode=link --silent g++ -std=c++11  gdalinfo_bin.lo  libgdal.la  -o gdalinfo
.libs/libgdal.so: undefined reference to `clCreateProgramWithSource'
.libs/libgdal.so: undefined reference to `clReleaseKernel'
.libs/libgdal.so: undefined reference to `clGetSupportedImageFormats'
.libs/libgdal.so: undefined reference to `clGetKernelWorkGroupInfo'
.libs/libgdal.so: undefined reference to `clCreateImage2D'
.libs/libgdal.so: undefined reference to `clEnqueueMapBuffer'
.libs/libgdal.so: undefined reference to `clEnqueueNDRangeKernel'
.libs/libgdal.so: undefined reference to `clGetPlatformIDs'
.libs/libgdal.so: undefined reference to `clGetDeviceInfo'
.libs/libgdal.so: undefined reference to `clCreateBuffer'
.libs/libgdal.so: undefined reference to `clReleaseCommandQueue'
.libs/libgdal.so: undefined reference to `clReleaseContext'
.libs/libgdal.so: undefined reference to `clReleaseEvent'
.libs/libgdal.so: undefined reference to `clGetDeviceIDs'
.libs/libgdal.so: undefined reference to `clReleaseMemObject'
.libs/libgdal.so: undefined reference to `clCreateCommandQueue'
.libs/libgdal.so: undefined reference to `clBuildProgram'
.libs/libgdal.so: undefined reference to `clFinish'
.libs/libgdal.so: undefined reference to `clGetProgramBuildInfo'
.libs/libgdal.so: undefined reference to `clEnqueueReadBuffer'
.libs/libgdal.so: undefined reference to `clReleaseProgram'
.libs/libgdal.so: undefined reference to `clCreateContext'
.libs/libgdal.so: undefined reference to `clSetKernelArg'
.libs/libgdal.so: undefined reference to `clCreateKernel'
collect2: error: ld returned 1 exit status
GNUmakefile:82: recipe for target 'gdalinfo' failed
make[1]: *** [gdalinfo] Error 1
make[1]: Leaving directory 'apps'
GNUmakefile:112: recipe for target 'apps-target' failed
make: *** [apps-target] Error 2

I know these symbols are correctly installed, because they are listed if I do for example:

$ export LD_LIBRARY_PATH=/usr/local/cuda/lib64   # Inform GDAL of Cuda
$ export PATH=/usr/local/cuda/bin:$PATH          # Inform GDAL of Cuda
$ cat /usr/local/cuda/include/CL/*  | grep -i clCreateProgramWithSource
clCreateProgramWithSource(cl_context        /* context */,
    ... )

I also think GDAL is correctly configured to know about these, because ./configure --with-opencl --with-opencl-include=/usr/local/cuda/include/ --with-opencl-lib=cuda can find OpenCL.

Therefore, I think it fails because the following command doesn't have -lOpenCL added to it, while apparently gdalinfo does want it.

/bin/bash libtool --mode=link --silent g++ -std=c++11  gdalinfo_bin.lo  libgdal.la  -o gdalinfo

So my final question is: How can I modify the build configuration to include -lOpenCL in more places?

Answer 1

You don't need to modify any GDAL makefiles to use the OpenCL - the GDAL build script is pretty good in finding everything you need for that. The only option for the configure script you need is --with-opencl=yes - this yes is essential. The options --with-opencl-include=ARG and --with-opencl-lib=ARG are useful only when your OpenCL installation is not standard - I didn't use them at all and got everything linked correctly.

The CUDA has its own version of OpenCL headers and the loader library, so you can use them even when there are no standard OpenCL (provided by a Linux software updater, for example - the apt) on your box. In case of the CUDA was installed in some non-standard location (for example - in your home directory) you have to tell the GDAL configure script about that:

./configure \
  --with-opencl=yes \
  --with-opencl-include=${HOME}/apps/cuda/targets/x86_64-linux/include \
  --with-opencl-lib="-L${HOME}/apps/cuda/targets/x86_64-linux/lib -lOpenCL" \
  <other options if needed>

Also, you don't need to tell anything to the configure script about your CUDA - the OpenCL loader library should find all the necessary CUDA libraries at run time using its internal mechanics.