diff options
| -rw-r--r-- | build/linux/acinclude.m4 | 6 | ||||
| -rw-r--r-- | build/linux/configure.ac | 8 |
2 files changed, 11 insertions, 3 deletions
diff --git a/build/linux/acinclude.m4 b/build/linux/acinclude.m4 index 313e849..5022e5b 100644 --- a/build/linux/acinclude.m4 +++ b/build/linux/acinclude.m4 @@ -74,7 +74,7 @@ rm -f conftest.cu conftest.o conftest.nvcc.out ]) dnl ASTRA_FIND_NVCC_ARCHS(archs-to-try,cppflags-to-extend,output-list) -dnl Architectures should be of the form 10, 20, 30, 35, +dnl Architectures should be of the form 10,20,30,35, dnl and should be in order. The last accepted one will be used for PTX output. dnl All accepted ones will be used for cubin output. AC_DEFUN([ASTRA_FIND_NVCC_ARCHS],[ @@ -88,7 +88,10 @@ _ACEOF NVCC_lastarch="none" NVCC_extra="" NVCC_list="" +astra_save_IFS=$IFS +IFS=, for arch in $1; do + IFS=$astra_save_IFS NVCC_opt="-gencode=arch=compute_$arch,code=sm_$arch" $NVCC -c -o conftest.o conftest.cu $$2 $NVCC_opt >conftest.nvcc.out 2>&1 && { NVCC_lastarch=$arch @@ -96,6 +99,7 @@ for arch in $1; do NVCC_list="${NVCC_list:+$NVCC_list, }$arch" } done +IFS=$astra_save_IFS if test $NVCC_lastarch != none; then NVCC_extra="$NVCC_extra -gencode=arch=compute_${NVCC_lastarch},code=compute_${NVCC_lastarch}" $3="$NVCC_list" diff --git a/build/linux/configure.ac b/build/linux/configure.ac index 4c9685b..11a8ff2 100644 --- a/build/linux/configure.ac +++ b/build/linux/configure.ac @@ -121,11 +121,15 @@ if test x"$HAVECUDA" = xyes; then fi AC_MSG_RESULT($HAVECUDA) +AC_ARG_WITH(cuda_compute, [[ --with-cuda-compute=archs comma separated list of CUDA compute models (optional)]],,) if test x"$HAVECUDA" = xyes; then AC_MSG_CHECKING([for nvcc archs]) dnl 10 11 12 13 20 30 32 35 37 50 - ASTRA_FIND_NVCC_ARCHS([10 12 20 30 35 50],NVCCFLAGS,NVCCARCHS) - AC_MSG_RESULT($NVCCARCHS) + if test x"$with_cuda_compute" = x; then + with_cuda_compute="10,12,20,30,35,50" + fi + ASTRA_FIND_NVCC_ARCHS([$with_cuda_compute],NVCCFLAGS,NVCCARCHS) + AC_MSG_RESULT([$NVCCARCHS]) fi |