diff --git a/CHANGELOG.md b/CHANGELOG.md
index f230ddb7ebc7ea5821955a96aab70963015b5d41..54e78e0a18d9bcd495dec076ee7d2681e8e466ef 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,6 +1,12 @@
# FSL base project changelog
+## 2111.6 (Friday 19th November 2021)
+
+* New convention for compiling C++ files which are part of a CUDA library or
+ executable, using `${CUDACXXFLAGS}`.
+
+
## 2111.5 (Friday 19th November 2021)
* Set `-std=c++14` for `nvcc`-compiled code, as CUDA 11.0 is the first version
diff --git a/README.md b/README.md
index 92ebb8ac0af58af9b639bb0747d19e4c1267b9a5..b24e8fc57bcde99830ff71eac5e600d8a50da81d 100644
--- a/README.md
+++ b/README.md
@@ -18,6 +18,57 @@ behaviour of various aspects of the `base` project. All tests are written
in Python, and are executed with `pytest`.
+## Standard conventions for C/C++/CUDA projects.
+
+Commands for compilation of intermediate object files should have the
+form:
+
+ $(CC) $(CFLAGS) <input/output files> # for .c files
+ $(CXX) $(CXXFLAGS) <input/output files> # for .cc files
+ $(CXX) $(CUDACXXFLAGS) <input/output files> # for .cc files which are to
+ # be linked into CUDA
+ # libraries/executables
+ $(NVCC) $(NVCCFLAGS) <input/output files> # for .cu files
+
+And commands for compilation and linking of executables and libraries
+should have the form:
+
+ $(CC) $(CFLAGS) <input/output files> ${LDFLAGS} # for c libs/exes
+ $(CXX) $(CXXFLAGS) <input/output files> ${LDFLAGS} # for c++ exes
+ $(CXX) $(CXXFLAGS) -shared <input/output files> ${LDFLAGS} # for c++ libs
+ $(NVCC) $(NVCCFLAGS) -shared <input/output files> ${NVCCLDFLAGS} # for CUDA libs
+ $(NVCC) $(NVCCFLAGS) <input/output files> ${NVCCLDFLAGS} # for CUDA exes
+
+
+`LDFLAGS` *must* come at the end, to ensure proper linking.
+
+An example C++ `Makefile` might look like this:
+
+```
+include ${FSLCONFDIR}/default.mk
+
+PROJNAME = myproject
+SOFILES = libmylib.so
+XFILES = myexe1 myexe2
+
+# Instead of explicitly defining this rule,
+# We could instead rely on implicit Make
+# rules for generating *.o files from *.cc
+# files.
+%.o: %.cc
+ $(CXX) $(CXXFLAGS) -c -o $@ $<
+
+libmylib.so: myobj1.o myobj2.o
+ $(CXX) $(CXXFLAGS) -shared -o $@ $^ $(LDFLAGS)
+
+myexe1: myexe1.o myobj1.o
+ $(CXX) $(CXXFLAGS) -o $@ $^ $(LDFLAGS)
+
+myexe2: myexe2.o myobj2.o
+ $(CXX) $(CXXFLAGS) -o $@ $^ $(LDFLAGS)
+```
+
+
## Building FSL CUDA projects.
Some FSL projects (e.g. [`fsl/eddy`(https://git.fmrib.ox.ac.uk/fsl/eddy) use
@@ -42,3 +93,21 @@ using static linking:
The `Makefile` for each FSL CUDA project may provide additional options for
controlling compilation.
+
+An example `Makefile` for a C++/CUDA project might look like this:
+
+```
+include ${FSLCONFDIR}/default.mk
+
+PROJNAME = myproject
+XFILES = myexe
+
+%.o: %.cu
+ $(NVCC) $(NVCCFLAGS) -c -o $@ $<
+
+%.o: %.cc
+ $(CXX) $(CUDACXXFLAGS) -c -o $@ $<
+
+myexe:
+ $(NVCC) $(NVCCFLAGS) -o $@ $< $(NVCCLDFLAGS)
+```
diff --git a/config/buildSettings.mk b/config/buildSettings.mk
index a958de299a892946a94b6a63a52cad1d3041c35f..f01ec640b3bad5304c5bc0c34a1b37e880a2ef68 100644
--- a/config/buildSettings.mk
+++ b/config/buildSettings.mk
@@ -65,6 +65,9 @@ ARCHLIBS =
# armadillo structures may have different byte-alignment
# to equivalent g++-compiled structures (see
# include/armadillo_bits/compiler_setup.hpp)
+#
+# - Set -std=c++14. This is the newest C++ standard
+# supported by nvcc for CUDA<11.0.
ARCHNVCCFLAGS = -DARMA_ALLOW_FAKE_GCC -std=c++14
# Linker flags for CUDA projects.
@@ -171,13 +174,14 @@ ifneq (${NVCC}, )
# "-ccbin" rather than "--compiler-bindir" here,
# because the conda-forge nvcc wrapper checks for
# -ccbin, and adds its own if not present.
- NVCCFLAGS ?= -I${CUDA_HOME}/include \
- ${GENCODEFLAGS} \
- -ccbin $(CXX)
- NVCCLDFLAGS ?= -L${CUDA_HOME}/lib \
- -L${CUDA_HOME}/lib64 \
- -L${CUDA_HOME}/lib/stubs \
- -L${CUDA_HOME}/lib64/stubs
+ CUDACXXFLAGS ?= -I${CUDA_HOME}/include
+ NVCCFLAGS ?= ${GENCODEFLAGS} \
+ -ccbin $(CXX) \
+ ${CUDACXXFLAGS}
+ NVCCLDFLAGS ?= -L${CUDA_HOME}/lib \
+ -L${CUDA_HOME}/lib64 \
+ -L${CUDA_HOME}/lib/stubs \
+ -L${CUDA_HOME}/lib64/stubs
# Link CUDA libraries statically, if compilation
# was invoked with "make CUDA_STATIC=1".
@@ -197,10 +201,10 @@ ifneq (${NVCC}, )
# cublas_Lt_static, in addition to culibos
ifeq ($(patsubst 9.%,,${CUDA_VER}),)
CUBLAS_STATIC = -lcublas_static -lculibos
- NVCCFLAGS += --cudart=static
+ NVCCLDFLAGS += --cudart=static
else
CUBLAS_STATIC = -lcublas_static -lcublasLt_static -lculibos
- NVCCFLAGS += --cudart=static --cudadevrt=static
+ NVCCLDFLAGS += --cudart=static --cudadevrt=static
endif
# Other CUDA toolkit components will
@@ -209,9 +213,9 @@ ifneq (${NVCC}, )
$(subst -lcurand,-lcurand_static,${CUDALIBS}))
else
ifeq ($(patsubst 9.%,,${CUDA_VER}),)
- NVCCFLAGS += --cudart=shared
+ NVCCLDFLAGS += --cudart=shared
else
- NVCCFLAGS += --cudart=shared --cudadevrt=static
+ NVCCLDFLAGS += --cudart=shared --cudadevrt=static
endif
_CUDALIBS += ${CUDALIBS}
endif
diff --git a/config/vars.mk b/config/vars.mk
index dd4358974a964c3ad267dcc4f4ead73acc7dc532..fe402f5c776903a567e364ab629338beac4b9d3c 100644
--- a/config/vars.mk
+++ b/config/vars.mk
@@ -22,13 +22,13 @@ PROJNAME =
USRLDFLAGS = # Linker flags
USRINCFLAGS = # Include directories
-USRCFLAGS = # Compiler flags for C projects
-USRCXXFLAGS = # Compiler flags for C++ projects
+USRCFLAGS = # Compiler flags for C files
+USRCXXFLAGS = # Compiler flags for C++ files
USRCPPFLAGS = # Preprocessor flags
LIBS = # Libraries to link against for C and C++ projects -
# these are incorporated into the final LDFLAGS, below.
-USRNVCCFLAGS = # Compiler flags for CUDA projects
-USRNVCCLDFLAGS = # Linker flags for CUDA projects
+USRNVCCFLAGS = # Compiler flags for CUDA files
+USRNVCCLDFLAGS = # Linker flags for CUDA libraries/executables
CUDALIBS = # CUDA libraries to link against (e.g. curand, cublas, etc) -
# these are incorporated into the final NVCCLDFLAGS, below.
# -lcuda and -lcudart are automatically added, so do not
@@ -84,20 +84,25 @@ INCFLAGS = ${USRINCFLAGS} -I. -I${DEVINCDIR} -I${INCDIR}
# All projects must use these flags for compilation/linking.
# Commands for compilation of intermediate object files
# should have the form:
-#
-# $(CC) $(CFLAGS) <input/output files>
-# $(CXX) $(CXXFLAGS) <input/output files>
-# $(NVCC) $(NVCCFLAGS) <input/output files>
+
+# $(CC) $(CFLAGS) <input/output files> # for .c files
+# $(CXX) $(CXXFLAGS) <input/output files> # for .cc files
+# $(CXX) $(CUDACXXFLAGS) <input/output files> # for .cc files which are to
+# # be linked into CUDA
+# # libraries/executables
+# $(NVCC) $(NVCCFLAGS) <input/output files> # for .cu files
#
# And commands for compilation and linking of executables
# and libraries should have the form:
#
-# $(CC) $(CFLAGS) <input/output files> ${LDFLAGS}
-# $(CXX) $(CXXFLAGS) <input/output files> ${LDFLAGS}
-# $(NVCC) $(NVCCFLAGS) <input/output files> ${NVCCLDFLAGS}
+# $(CC) $(CFLAGS) <input/output files> ${LDFLAGS} # for c libs/exes
+# $(CXX) $(CXXFLAGS) <input/output files> ${LDFLAGS} # for c++ exes
+# $(CXX) $(CXXFLAGS) -shared <input/output files> ${LDFLAGS} # for c++ libs
+# $(NVCC) $(NVCCFLAGS) -shared <input/output files> ${NVCCLDFLAGS} # for CUDA libs
+# $(NVCC) $(NVCCFLAGS) <input/output files> ${NVCCLDFLAGS} # for CUDA exes
+#
+# `LDFLAGS` *must* come at the end, to ensure proper linking.
#
-# LDFLAGS *must* come at the end, to ensure proper linking.
-
# The order in which the final FLAGS variables are
# constructed here is important:
#
@@ -113,12 +118,13 @@ INCFLAGS = ${USRINCFLAGS} -I. -I${DEVINCDIR} -I${INCDIR}
# the default options specified in buildSettings.mk,
# which in turn can override options provided by the
# environment.
-CPPFLAGS += ${ARCHCPPFLAGS} ${USRCPPFLAGS}
-CFLAGS += ${CPPFLAGS} ${ARCHCFLAGS} ${USRCFLAGS} ${INCFLAGS}
-CXXFLAGS += ${CPPFLAGS} ${ARCHCXXFLAGS} ${USRCXXFLAGS} ${INCFLAGS}
-LDFLAGS += ${ARCHLDFLAGS} ${USRLDFLAGS} \
- -L. -L${DEVLIBDIR} -L${LIBDIR} \
- ${LIBS} ${ARCHLIBS}
+CPPFLAGS += ${ARCHCPPFLAGS} ${USRCPPFLAGS}
+CFLAGS += ${CPPFLAGS} ${ARCHCFLAGS} ${USRCFLAGS} ${INCFLAGS}
+CXXFLAGS += ${CPPFLAGS} ${ARCHCXXFLAGS} ${USRCXXFLAGS} ${INCFLAGS}
+CUDACXXFLAGS += ${CXXFLAGS} ${USRNVCCFLAGS}
+LDFLAGS += ${ARCHLDFLAGS} ${USRLDFLAGS} \
+ -L. -L${DEVLIBDIR} -L${LIBDIR} \
+ ${LIBS} ${ARCHLIBS}
# Remove any -std=c++ options, as we are already setting
# -std in ARCHNVCCFLAGS (seee buildSettings.mk), and
# passing another one via --compiler-options will confuse