Note: I am by far no expert on this – (I only spent the past few days trying to wrap my head around automake, and before that, never opened a configure.ac file before) – just wanted to share my experiences.
Below is an excerpt from configure.ac and Makefile.am files I modified to add CUDA to an existing automake/autoconf setup (the .. .. implies other unmodified/existing lines).
# -*- configure.ac -*-
# Process this file with autoconf to produce a configure script.
..
..
AM_INIT_AUTOMAKE(no-dependencies)
AC_CONFIG_HEADER([config.h])
..
..
# initialize array index
N="0"
..
..
#=================================================
# Setup default CUDA paths
CUDA_INSTALL_PATH="/usr/local/cuda"
CUDA_SDK_INSTALL_PATH="/usr/local/NVIDIA_GPU_Computing_SDK"
# Check: enable CUDA?
AC_MSG_CHECKING([whether to use CUDA library])
AC_ARG_ENABLE(cuda,
[AS_HELP_STRING(--enable-cuda,enable CUDA support)],
cudaval="yes",
cudaval="no")
AC_MSG_RESULT([$cudaval])
# check to see if CUDA been selected
if test "$cudaval" == "no"; then
AC_MSG_WARN([Not using CUDA!])
fi
# Setup custom CUDA paths
AC_ARG_WITH([cuda],
[AS_HELP_STRING(
[--with-cuda=PATH],
[prefix where CUDA is installed @<:@default=auto@:>@])],
[CUDA_INSTALL_PATH=$withval],
[with_cuda=auto])
AC_ARG_WITH([cuda-sdk],
[AS_HELP_STRING(
[--with-cuda-sdk=PATH],
[prefix where CUDA SDK is installed @<:@default=auto@:>@])],
[CUDA_SDK_INSTALL_PATH=$withval],
[with_cuda_sdk=auto])
# Setup nvcc flags
if test "$cudaval" = "yes"; then
NVCCFLAGS="$NVCCFLAGS"
CUDA_CFLAGS="$CUDA_CFLAGS"
CUDA_CFLAGS="$CUDA_CFLAGS -I$CUDA_SDK_INSTALL_PATH/C/common/inc/"
CUDA_CFLAGS="$CUDA_CFLAGS -I$CUDA_INSTALL_PATH/include"
CUDA_LDFLAGS="-L$CUDA_INSTALL_PATH/lib64"
CUDA_LIBS="-lcuda -lcudart -lcufft"
NVCCFLAGS="$NVCCFLAGS $CUDA_CFLAGS $CUDA_LDFLAGS $CUDA_LIBS"
NVCC="nvcc"
AC_SUBST(NVCC)
AC_SUBST(NVCCFLAGS)
AC_DEFINE([_CUDA],[1],[Defined if CUDA should be used])
eval "ARRAY${N}='-D_CUDA'"
N=`expr $N + 1`
fi
#=================================================
..
..
# construct compiler flags
N=`expr $N - 1`
TMP0=
eval TMP0="\$ARRAY0"
for i in $(seq 1 $N)
do
eval TMP0="\$TMP0' '\$ARRAY${i}"
done
# set C compiler flags
AC_SUBST(CDFLAGS, $TMP0)
..
..
..
When --enable-cuda is set, the above configure.ac will generate the variables NVCC and NVCCFLAGS for use in an Makefile.am, for example:
#
# -*- Makefile.am -*-
..
..
.cu.o:
$(NVCC) $(NVCCFLAGS) -c -o $@ $<
..
..
In addition, a _CUDA variable gets defined (#define _CUDA 1), for use in any file getting compiled with the above Makefile.am/configure.ac setup.
Hopefully, this can save at least someone else time (as I’ve wasted enough on this), and give a nice pointer in the right direction to anyone interested in adding CUDA to an existing ./configure setup.
Enjoy!
Hi Chris,
what is that CUDA_CFLAGS ? i came acroos the same modification in a project. but when i tried to make after configuration, it returns the error : NVCC@ : command not found.
Hi,
Actually, CUDA_CFLAGS is really just a temporary variable to accumulate the nvcc parameters, and eventually becomes NVCCFLAGS for convenience.
Just make sure nvcc is in your PATH environment variable (ex /usr/local/cuda/bin).
~Chris
Hi Chris, i found that generated makefile is dont track changing of .cu files, so when i change .cu files it has no effect onto ‘make’ that’s why i must use ‘make clean & make’…
P.S. Chris, can you upload the file, or change the view of that, because if i copy that it copies with html special chars.Thanks in advance!
Hi Andrey,
Thanks for the info, not sure how to fix that just yet…
Unfortunately, the project I used this on is no longer available for me to access the code, but here is a raw paste of what is above:
-*- configure.ac -*-
# Process this file with autoconf to produce a configure script.
..
..
AM_INIT_AUTOMAKE(no-dependencies)
AC_CONFIG_HEADER([config.h])
..
..
# initialize array index
N=”0″
..
..
#=================================================
# Setup default CUDA paths
CUDA_INSTALL_PATH=”/usr/local/cuda”
CUDA_SDK_INSTALL_PATH=”/usr/local/NVIDIA_GPU_Computing_SDK”
# Check: enable CUDA?
AC_MSG_CHECKING([whether to use CUDA library])
AC_ARG_ENABLE(cuda,
[AS_HELP_STRING(–enable-cuda,enable CUDA support)],
cudaval=”yes”,
cudaval=”no”)
AC_MSG_RESULT([$cudaval])
# check to see if CUDA been selected
if test “$cudaval” == “no”; then
AC_MSG_WARN([Not using CUDA!])
fi
# Setup custom CUDA paths
AC_ARG_WITH([cuda],
[AS_HELP_STRING(
[–with-cuda=PATH],
[prefix where CUDA is installed @<:@default=auto@:>@])],
[CUDA_INSTALL_PATH=$withval],
[with_cuda=auto])
AC_ARG_WITH([cuda-sdk],
[AS_HELP_STRING(
[–with-cuda-sdk=PATH],
[prefix where CUDA SDK is installed @<:@default=auto@:>@])],
[CUDA_SDK_INSTALL_PATH=$withval],
[with_cuda_sdk=auto])
# Setup nvcc flags
if test “$cudaval” = “yes”; then
NVCCFLAGS=”$NVCCFLAGS”
CUDA_CFLAGS=”$CUDA_CFLAGS”
CUDA_CFLAGS=”$CUDA_CFLAGS -I$CUDA_SDK_INSTALL_PATH/C/common/inc/”
CUDA_CFLAGS=”$CUDA_CFLAGS -I$CUDA_INSTALL_PATH/include”
CUDA_LDFLAGS=”-L$CUDA_INSTALL_PATH/lib64″
CUDA_LIBS=”-lcuda -lcudart -lcufft”
NVCCFLAGS=”$NVCCFLAGS $CUDA_CFLAGS $CUDA_LDFLAGS $CUDA_LIBS”
NVCC=”nvcc”
AC_SUBST(NVCC)
AC_SUBST(NVCCFLAGS)
AC_DEFINE([_CUDA],[1],[Defined if CUDA should be used])
eval “ARRAY${N}=’-D_CUDA'”
N=`expr $N + 1`
fi
#=================================================
..
..
# construct compiler flags
N=`expr $N – 1`
TMP0=
eval TMP0=”\$ARRAY0″
for i in $(seq 1 $N)
do
eval TMP0=”\$TMP0′ ‘\$ARRAY${i}”
done
# set C compiler flags
AC_SUBST(CDFLAGS, $TMP0)
..
..
..
Thanks for the interest, and good luck on your project!
~Chris
Thank you, Chris !