gumath 0.2.0dev5 → 0.2.0dev8

data/ext/ruby_gumath/gumath/configure.ac

@@ -41,9 +41,12 @@ AC_SUBST(host)
 
 # Language and compiler:
 AC_LANG_C
-saved_cflags=$CFLAGS
+saved_cflags="$CFLAGS"
+saved_cxxflags="$CXXFLAGS"
 AC_PROG_CC
-CFLAGS=$saved_cflags
+AC_PROG_CXX
+CFLAGS="$saved_cflags"
+CXXFLAGS="$saved_cxxflags"
 
 # ar and ranlib:
 AC_CHECK_TOOL(AR, ar, ar)
@@ -61,6 +64,40 @@ AC_CHECK_HEADER([pthread.h],
 AC_PROG_INSTALL
 AC_SUBST(INSTALL)
 
+# Cuda compiler:
+AC_MSG_CHECKING(for nvcc)
+saved_cc="$CC"
+saved_cflags="$CFLAGS"
+saved_cxxflags="$CXXFLAGS"
+CC=nvcc
+AC_COMPILE_IFELSE([AC_LANG_SOURCE([[
+void
+f(void)
+{
+    return;
+}
+]])],
+[have_nvcc=yes],
+[have_nvcc=no],
+[have_nvcc=undefined])
+CC="$saved_cc"
+CFLAGS="$saved_cflags"
+CXXFLAGS="$saved_cxxflags"
+AC_MSG_RESULT($have_nvcc)
+
+CUDA_CXX=
+CONFIGURE_CUDA_CXXFLAGS=
+if test "$have_nvcc" = yes; then
+    if nvcc -o tools/detect_cuda_arch tools/detect_cuda_arch.cc > /dev/null 2>&1; then
+        CUDA_CXX="nvcc"
+        CUDA_ARCH=`./tools/detect_cuda_arch`
+        CONFIGURE_CUDA_CXXFLAGS="-std=c++11 -arch=compute_$CUDA_ARCH -code=sm_$CUDA_ARCH"
+        AC_DEFINE(HAVE_CUDA, 1, [Define to 1 if you have the nvcc cuda compiler.])
+    fi
+fi
+AC_SUBST(CUDA_CXX)
+AC_SUBST(CONFIGURE_CUDA_CXXFLAGS)
+
 # Add an explicit include directory.
 AC_MSG_CHECKING(for --with-includes)
 AC_ARG_WITH(includes,
@@ -140,7 +177,7 @@ esac
 
 # Substitute variables and generate output:
 if test -z "$LD"; then
-    LD="$CC"
+    LD="$CXX"
 fi
 AC_SUBST(LD)
 AC_SUBST(AR)
@@ -154,6 +191,12 @@ else
     CONFIGURE_CFLAGS="$GM_INCLUDE $GM_WARN $GM_CONFIG $GM_OPT $CFLAGS"
 fi
 
+if test -z "$CXXFLAGS"; then
+    CONFIGURE_CXXFLAGS="$XND_INCLUDE -Wall -Wextra -std=c++11 $GM_OPT -g"
+else
+    CONFIGURE_CXXFLAGS="$XND_INCLUDE -Wall -Wextra -std=c++11 $GM_OPT -g $CXXFLAGS"
+fi
+
 if test -z "$LDFLAGS"; then
     CONFIGURE_LDFLAGS="$GM_LINK $CONFIGURE_LDFLAGS"
 else
@@ -161,6 +204,7 @@ else
 fi
 
 AC_SUBST(CONFIGURE_CFLAGS)
+AC_SUBST(CONFIGURE_CXXFLAGS)
 AC_SUBST(CONFIGURE_LDFLAGS)
 
 AC_OUTPUT

data/ext/ruby_gumath/gumath/libgumath/Makefile

@@ -0,0 +1,236 @@
+
+# ==============================================================================
+#                          Unix Makefile for libxnd
+# ==============================================================================
+
+
+LIBSTATIC = libgumath.a
+LIBNAME = libgumath.so
+LIBSONAME = libgumath.so.0
+LIBSHARED = libgumath.so.0.2.0dev3
+
+CC = gcc
+CXX = g++
+LD = g++
+AR = ar
+RANLIB = ranlib
+CUDA_CXX = 
+
+GM_INCLUDES = ../ndtypes/libndtypes
+
+CONFIGURE_CFLAGS =  -Wall -Wextra -std=c11 -pedantic  -O3  -I /home/sameer/.rvm/gems/ruby-2.4.1/gems/xnd-0.2.0dev7/ext/ruby_xnd/include -I /home/sameer/.rvm/gems/ruby-2.4.1/gems/ndtypes-0.2.0dev6/ext/ruby_ndtypes/include 
+GM_CFLAGS = $(strip -I.. -I$(GM_INCLUDES) $(CONFIGURE_CFLAGS) $(CFLAGS))
+GM_CFLAGS_SHARED = $(GM_CFLAGS) -fPIC
+
+CONFIGURE_CXXFLAGS =  -Wall -Wextra -std=c++11 -O3 -g
+GM_CXXFLAGS = $(strip -I$(GM_INCLUDES) $(CONFIGURE_CXXFLAGS) $(CXXFLAGS))
+GM_CXXFLAGS_SHARED = $(GM_CXXFLAGS) -fPIC
+
+CONFIGURE_LDFLAGS =  -shared -Wl,-soname,libgumath.so.0
+GM_LDFLAGS = $(strip $(CONFIGURE_LDFLAGS) $(LDFLAGS))
+
+CONFIGURE_CUDA_CXXFLAGS = 
+GM_CUDA_CXXFLAGS = $(strip $(CONFIGURE_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS))
+
+default: $(LIBSTATIC) $(LIBSHARED)
+
+
+OBJS = apply.o func.o nploops.o tbl.o thread.o xndloops.o cpu_host_unary.o \
+       cpu_device_unary.o cpu_host_binary.o cpu_device_binary.o common.o \
+       examples.o graph.o quaternion.o pdist.o
+
+SHARED_OBJS = .objs/apply.o .objs/func.o .objs/nploops.o .objs/tbl.o .objs/thread.o .objs/xndloops.o \
+              .objs/cpu_host_unary.o .objs/cpu_device_unary.o .objs/cpu_host_binary.o .objs/cpu_device_binary.o \
+              .objs/common.o .objs/examples.o .objs/graph.o .objs/quaternion.o .objs/pdist.o
+
+ifdef CUDA_CXX
+OBJS += cuda_host_unary.o cuda_device_unary.o cuda_host_binary.o cuda_device_binary.o
+SHARED_OBJS += .objs/cuda_host_unary.o .objs/cuda_device_unary.o .objs/cuda_host_binary.o .objs/cuda_device_binary.o
+endif
+
+
+$(LIBSTATIC): Makefile $(OBJS)
+	$(AR) rc $(LIBSTATIC) $(OBJS)
+	$(RANLIB) $(LIBSTATIC)
+
+$(LIBSHARED): Makefile $(SHARED_OBJS)
+	$(LD) $(GM_LDFLAGS) -o $(LIBSHARED) $(SHARED_OBJS)
+	ln -sf $(LIBSHARED) $(LIBNAME)
+	ln -sf $(LIBSHARED) $(LIBSONAME)
+
+
+apply.o:\
+Makefile apply.c gumath.h
+	$(CC) $(GM_CFLAGS) -c apply.c
+
+.objs/apply.o:\
+Makefile apply.c gumath.h
+	$(CC) $(GM_CFLAGS_SHARED) -c apply.c -o .objs/apply.o
+
+func.o:\
+Makefile func.c gumath.h
+	$(CC) $(GM_CFLAGS) -c func.c
+
+.objs/func.o:\
+Makefile func.c gumath.h
+	$(CC) $(GM_CFLAGS_SHARED) -c func.c -o .objs/func.o
+
+nploops.o:\
+Makefile nploops.c gumath.h
+	$(CC) $(GM_CFLAGS) -c nploops.c
+
+.objs/nploops.o:\
+Makefile nploops.c gumath.h
+	$(CC) $(GM_CFLAGS_SHARED) -c nploops.c -o .objs/nploops.o
+
+tbl.o:\
+Makefile tbl.c gumath.h
+	$(CC) $(GM_CFLAGS) -c tbl.c
+
+.objs/tbl.o:\
+Makefile tbl.c gumath.h
+	$(CC) $(GM_CFLAGS_SHARED) -c tbl.c -o .objs/tbl.o
+
+thread.o:\
+Makefile thread.c gumath.h
+	$(CC) $(GM_CFLAGS) -c thread.c
+
+.objs/thread.o:\
+Makefile thread.c gumath.h
+	$(CC) $(GM_CFLAGS_SHARED) -c thread.c -o .objs/thread.o
+
+xndloops.o:\
+Makefile xndloops.c gumath.h
+	$(CC) $(GM_CFLAGS) -c xndloops.c
+
+.objs/xndloops.o:\
+Makefile xndloops.c gumath.h
+	$(CC) $(GM_CFLAGS_SHARED) -c xndloops.c -o .objs/xndloops.o
+
+cpu_device_unary.o:\
+Makefile kernels/cpu_device_unary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS) -Wno-absolute-value -c kernels/cpu_device_unary.cc
+
+.objs/cpu_device_unary.o:\
+Makefile kernels/cpu_device_unary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS_SHARED) -Wno-absolute-value -c kernels/cpu_device_unary.cc -o .objs/cpu_device_unary.o
+
+cpu_host_unary.o:\
+Makefile kernels/cpu_host_unary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -Wno-absolute-value -c kernels/cpu_host_unary.c
+
+.objs/cpu_host_unary.o:\
+Makefile kernels/cpu_host_unary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -Wno-absolute-value -c kernels/cpu_host_unary.c -o .objs/cpu_host_unary.o
+
+cpu_host_binary.o:\
+Makefile kernels/cpu_host_binary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c kernels/cpu_host_binary.c
+
+.objs/cpu_host_binary.o:\
+Makefile kernels/cpu_host_binary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c kernels/cpu_host_binary.c -o .objs/cpu_host_binary.o
+
+cpu_device_binary.o:\
+Makefile kernels/cpu_device_binary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS) -c kernels/cpu_device_binary.cc
+
+.objs/cpu_device_binary.o:\
+Makefile kernels/cpu_device_binary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS_SHARED) -c kernels/cpu_device_binary.cc -o .objs/cpu_device_binary.o
+
+common.o:\
+Makefile kernels/common.c kernels/common.h gumath.h
+common.o:\
+Makefile kernels/common.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c kernels/common.c
+
+.objs/common.o:\
+Makefile kernels/common.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c kernels/common.c -o .objs/common.o
+
+examples.o:\
+Makefile extending/examples.c gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c extending/examples.c -o examples.o
+
+.objs/examples.o:\
+Makefile extending/examples.c gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c extending/examples.c -o .objs/examples.o
+
+graph.o:\
+Makefile extending/graph.c gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c extending/graph.c -o graph.o
+
+.objs/graph.o:\
+Makefile extending/graph.c gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c extending/graph.c -o .objs/graph.o
+
+quaternion.o:\
+Makefile extending/quaternion.c gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c extending/quaternion.c -o quaternion.o
+
+.objs/quaternion.o:\
+Makefile extending/quaternion.c gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c extending/quaternion.c -o .objs/quaternion.o
+
+pdist.o:\
+Makefile extending/pdist.c gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c extending/pdist.c -o pdist.o
+
+.objs/pdist.o:\
+Makefile extending/pdist.c gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c extending/pdist.c -o .objs/pdist.o
+
+
+# Cuda
+cuda_host_unary.o:\
+Makefile kernels/cuda_host_unary.c kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -Wno-absolute-value -c kernels/cuda_host_unary.c
+
+.objs/cuda_host_unary.o:\
+Makefile kernels/cuda_host_unary.c kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -Wno-absolute-value -c kernels/cuda_host_unary.c -o .objs/cuda_host_unary.o
+
+cuda_device_unary.o:\
+Makefile kernels/cuda_device_unary.cu kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_unary.cu
+
+.objs/cuda_device_unary.o:\
+Makefile kernels/cuda_device_unary.cu kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS_SHARED)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_unary.cu -o .objs/cuda_device_unary.o
+
+cuda_host_binary.o:\
+Makefile kernels/cuda_host_binary.c kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -Wno-absolute-value -c kernels/cuda_host_binary.c
+
+.objs/cuda_host_binary.o:\
+Makefile kernels/cuda_host_binary.c kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -Wno-absolute-value -c kernels/cuda_host_binary.c -o .objs/cuda_host_binary.o
+
+cuda_device_binary.o:\
+Makefile kernels/cuda_device_binary.cu kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_binary.cu
+
+.objs/cuda_device_binary.o:\
+Makefile kernels/cuda_device_binary.cu kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS_SHARED)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_binary.cu -o .objs/cuda_device_binary.o
+
+
+# Coverage
+coverage:\
+Makefile clean runtest
+	./tests/runtest
+	for file in *.c; do gcov -l "$$file" > /dev/null 2>&1; done
+
+FORCE:
+
+clean: FORCE
+	rm -f *.o *.so *.gch *.gcda *.gcno *.gcov *.dyn *.dpi *.lock
+	rm -f $(LIBSTATIC) $(LIBSHARED) $(LIBSONAME) $(LIBNAME)
+	cd .objs && rm -f *.o *.so *.gch *.gcda *.gcno *.gcov *.dyn *.dpi *.lock
+
+distclean: clean
+	rm -f Makefile
+
+
+

data/ext/ruby_gumath/gumath/libgumath/Makefile.in

@@ -10,9 +10,11 @@ LIBSONAME = @LIBSONAME@
 LIBSHARED = @LIBSHARED@
 
 CC = @CC@
+CXX = @CXX@
 LD = @LD@
 AR = @AR@
 RANLIB = @RANLIB@
+CUDA_CXX = @CUDA_CXX@
 
 GM_INCLUDES = @CONFIGURE_INCLUDES@
 
@@ -20,19 +22,31 @@ CONFIGURE_CFLAGS = @CONFIGURE_CFLAGS@
 GM_CFLAGS = $(strip -I.. -I$(GM_INCLUDES) $(CONFIGURE_CFLAGS) $(CFLAGS))
 GM_CFLAGS_SHARED = $(GM_CFLAGS) -fPIC
 
+CONFIGURE_CXXFLAGS = @CONFIGURE_CXXFLAGS@
+GM_CXXFLAGS = $(strip -I$(GM_INCLUDES) $(CONFIGURE_CXXFLAGS) $(CXXFLAGS))
+GM_CXXFLAGS_SHARED = $(GM_CXXFLAGS) -fPIC
+
 CONFIGURE_LDFLAGS = @CONFIGURE_LDFLAGS@
 GM_LDFLAGS = $(strip $(CONFIGURE_LDFLAGS) $(LDFLAGS))
 
+CONFIGURE_CUDA_CXXFLAGS = @CONFIGURE_CUDA_CXXFLAGS@
+GM_CUDA_CXXFLAGS = $(strip $(CONFIGURE_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS))
 
 default: $(LIBSTATIC) $(LIBSHARED)
 
 
-OBJS = apply.o func.o nploops.o tbl.o thread.o xndloops.o unary.o binary.o \
-       examples.o bfloat16.o graph.o quaternion.o pdist.o
+OBJS = apply.o func.o nploops.o tbl.o thread.o xndloops.o cpu_host_unary.o \
+       cpu_device_unary.o cpu_host_binary.o cpu_device_binary.o common.o \
+       examples.o graph.o quaternion.o pdist.o
 
 SHARED_OBJS = .objs/apply.o .objs/func.o .objs/nploops.o .objs/tbl.o .objs/thread.o .objs/xndloops.o \
-              .objs/unary.o .objs/binary.o .objs/examples.o .objs/bfloat16.o .objs/graph.o \
-              .objs/quaternion.o .objs/pdist.o
+              .objs/cpu_host_unary.o .objs/cpu_device_unary.o .objs/cpu_host_binary.o .objs/cpu_device_binary.o \
+              .objs/common.o .objs/examples.o .objs/graph.o .objs/quaternion.o .objs/pdist.o
+
+ifdef CUDA_CXX
+OBJS += cuda_host_unary.o cuda_device_unary.o cuda_host_binary.o cuda_device_binary.o
+SHARED_OBJS += .objs/cuda_host_unary.o .objs/cuda_device_unary.o .objs/cuda_host_binary.o .objs/cuda_device_binary.o
+endif
 
 
 $(LIBSTATIC): Makefile $(OBJS)
@@ -93,21 +107,47 @@ Makefile xndloops.c gumath.h
 Makefile xndloops.c gumath.h
 	$(CC) $(GM_CFLAGS_SHARED) -c xndloops.c -o .objs/xndloops.o
 
-unary.o:\
-Makefile kernels/unary.c gumath.h
-	$(CC) -I. $(GM_CFLAGS) -Wno-absolute-value -c kernels/unary.c
+cpu_device_unary.o:\
+Makefile kernels/cpu_device_unary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS) -Wno-absolute-value -c kernels/cpu_device_unary.cc
+
+.objs/cpu_device_unary.o:\
+Makefile kernels/cpu_device_unary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS_SHARED) -Wno-absolute-value -c kernels/cpu_device_unary.cc -o .objs/cpu_device_unary.o
+
+cpu_host_unary.o:\
+Makefile kernels/cpu_host_unary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -Wno-absolute-value -c kernels/cpu_host_unary.c
+
+.objs/cpu_host_unary.o:\
+Makefile kernels/cpu_host_unary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -Wno-absolute-value -c kernels/cpu_host_unary.c -o .objs/cpu_host_unary.o
 
-.objs/unary.o:\
-Makefile kernels/unary.c gumath.h
-	$(CC) -I. $(GM_CFLAGS_SHARED) -Wno-absolute-value -c kernels/unary.c -o .objs/unary.o
+cpu_host_binary.o:\
+Makefile kernels/cpu_host_binary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c kernels/cpu_host_binary.c
 
-binary.o:\
-Makefile kernels/binary.c gumath.h
-	$(CC) -I. $(GM_CFLAGS) -c kernels/binary.c
+.objs/cpu_host_binary.o:\
+Makefile kernels/cpu_host_binary.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c kernels/cpu_host_binary.c -o .objs/cpu_host_binary.o
 
-.objs/binary.o:\
-Makefile kernels/binary.c gumath.h
-	$(CC) -I. $(GM_CFLAGS_SHARED) -c kernels/binary.c -o .objs/binary.o
+cpu_device_binary.o:\
+Makefile kernels/cpu_device_binary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS) -c kernels/cpu_device_binary.cc
+
+.objs/cpu_device_binary.o:\
+Makefile kernels/cpu_device_binary.cc kernels/common.h gumath.h
+	$(CXX) -I. $(GM_CXXFLAGS_SHARED) -c kernels/cpu_device_binary.cc -o .objs/cpu_device_binary.o
+
+common.o:\
+Makefile kernels/common.c kernels/common.h gumath.h
+common.o:\
+Makefile kernels/common.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -c kernels/common.c
+
+.objs/common.o:\
+Makefile kernels/common.c kernels/common.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -c kernels/common.c -o .objs/common.o
 
 examples.o:\
 Makefile extending/examples.c gumath.h
@@ -117,14 +157,6 @@ Makefile extending/examples.c gumath.h
 Makefile extending/examples.c gumath.h
 	$(CC) -I. $(GM_CFLAGS_SHARED) -c extending/examples.c -o .objs/examples.o
 
-bfloat16.o:\
-Makefile extending/bfloat16.c gumath.h
-	$(CC) -I. $(GM_CFLAGS) -c extending/bfloat16.c -o bfloat16.o
-
-.objs/bfloat16.o:\
-Makefile extending/bfloat16.c gumath.h
-	$(CC) -I. $(GM_CFLAGS_SHARED) -c extending/bfloat16.c -o .objs/bfloat16.o
-
 graph.o:\
 Makefile extending/graph.c gumath.h
 	$(CC) -I. $(GM_CFLAGS) -c extending/graph.c -o graph.o
@@ -150,6 +182,40 @@ Makefile extending/pdist.c gumath.h
 	$(CC) -I. $(GM_CFLAGS_SHARED) -c extending/pdist.c -o .objs/pdist.o
 
 
+# Cuda
+cuda_host_unary.o:\
+Makefile kernels/cuda_host_unary.c kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -Wno-absolute-value -c kernels/cuda_host_unary.c
+
+.objs/cuda_host_unary.o:\
+Makefile kernels/cuda_host_unary.c kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -Wno-absolute-value -c kernels/cuda_host_unary.c -o .objs/cuda_host_unary.o
+
+cuda_device_unary.o:\
+Makefile kernels/cuda_device_unary.cu kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_unary.cu
+
+.objs/cuda_device_unary.o:\
+Makefile kernels/cuda_device_unary.cu kernels/common.h kernels/cuda_device_unary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS_SHARED)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_unary.cu -o .objs/cuda_device_unary.o
+
+cuda_host_binary.o:\
+Makefile kernels/cuda_host_binary.c kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS) -Wno-absolute-value -c kernels/cuda_host_binary.c
+
+.objs/cuda_host_binary.o:\
+Makefile kernels/cuda_host_binary.c kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CC) -I. $(GM_CFLAGS_SHARED) -Wno-absolute-value -c kernels/cuda_host_binary.c -o .objs/cuda_host_binary.o
+
+cuda_device_binary.o:\
+Makefile kernels/cuda_device_binary.cu kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_binary.cu
+
+.objs/cuda_device_binary.o:\
+Makefile kernels/cuda_device_binary.cu kernels/common.h kernels/cuda_device_binary.h gumath.h
+	$(CUDA_CXX) --compiler-options "$(GM_CXXFLAGS_SHARED)" $(GM_CUDA_CXXFLAGS) -c kernels/cuda_device_binary.cu -o .objs/cuda_device_binary.o
+
+
 # Coverage
 coverage:\
 Makefile clean runtest