move '-x cu' into arch files

cjknight · Feb 26, 2025 · 22ae67c · 22ae67c
1 parent 0ea1f49
commit 22ae67c
Show file tree

Hide file tree

Showing 7 changed files with 31 additions and 15 deletions.
diff --git a/gpu/mini-apps/affinity/Makefile b/gpu/mini-apps/affinity/Makefile
@@ -66,7 +66,8 @@ $(C_OBJ): %.o: %.cpp
 	$(CXX) $(CXXFLAGS) -c $<
 
 $(CUDA_OBJ): %.o: pm/%.cpp
-	$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
+	$(CUDA_CXX) $(CUDA_CXXFLAGS) -c $< -o $@
+	#$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
 
 $(PM_OBJ): %.o: pm/%.cpp
 	$(CXX) $(CXXFLAGS) -c $<

diff --git a/gpu/mini-apps/math/Makefile b/gpu/mini-apps/math/Makefile
@@ -77,7 +77,8 @@ $(C_OBJ): %.o: %.cpp
 	$(CXX) $(CXXFLAGS) -c $<
 
 $(CUDA_OBJ): %.o: pm/%.cpp
-	$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
+	$(CUDA_CXX) $(CUDA_CXXFLAGS) -c $< -o $@
+	#$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
 
 $(PM_OBJ): %.o: pm/%.cpp
 	$(CXX) $(CXXFLAGS) -c $<

diff --git a/gpu/mini-apps/transpose/Makefile b/gpu/mini-apps/transpose/Makefile
@@ -34,9 +34,14 @@ CUDA_OBJ2 = $(TMP02:.cpp=.o)
 CUDA_OBJ = $(CUDA_OBJ1) $(CUDA_OBJ2)
 
 # -- subset of src files in src directory
-C_SRC = $(wildcard *.cpp)
+C_SRC1 = $(wildcard *.cpp)
+C_SRC2 = $(filter-out $(CUDA_SRC2), $(wildcard offload/*.cpp))
+C_SRC = $(C_SRC1) $(C_SRC2)
 C_INC = $(wildcard *.h)
-C_OBJ = $(C_SRC:.cpp=.o)
+C_OBJ1 = $(C_SRC1:.cpp=.o)
+TMP03 = $(C_SRC2:offload/%=%)
+C_OBJ2 = $(TMP03:.cpp=.o)
+C_OBJ = $(C_OBJ1) $(C_OBJ2)
 
 # -- subset of src files in src/pm directory
 PM_SRC = $(filter-out $(CUDA_SRC), $(wildcard pm/*.cpp))
@@ -48,9 +53,9 @@ $(info CUDA_SRC is [${CUDA_SRC}])
 $(info CUDA_INC is [${CUDA_INC}])
 $(info CUDA_OBJ is [${CUDA_OBJ}])
 
-#$(info C_SRC is [${C_SRC}])
-#$(info C_INC is [${C_INC}])
-#$(info C_OBJ is [${C_OBJ}])
+$(info C_SRC is [${C_SRC}])
+$(info C_INC is [${C_INC}])
+$(info C_OBJ is [${C_OBJ}])
 
 #$(info PM_SRC is [${PM_SRC}])
 #$(info PM_INC is [${PM_INC}])
@@ -67,14 +72,19 @@ update:
 
 ####################################################################
 
-$(C_OBJ): %.o: %.cpp
+$(C_OBJ1): %.o: %.cpp
+	$(CXX) $(CXXFLAGS) -c $<
+
+$(C_OBJ2): %.o: offload/%.cpp
 	$(CXX) $(CXXFLAGS) -c $<
 
 $(CUDA_OBJ1): %.o: pm/%.cpp
-	$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
+	$(CUDA_CXX) $(CUDA_CXXFLAGS) -c $< -o $@
+	#$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
 
 $(CUDA_OBJ2): %.o: offload/%.cpp
-	$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
+	$(CUDA_CXX) $(CUDA_CXXFLAGS) -c $< -o $@
+	#$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
 
 $(PM_OBJ): %.o: pm/%.cpp
 	$(CXX) $(CXXFLAGS) -c $<
@@ -92,7 +102,7 @@ realclean:
 	rm -f $(EXE)
 	rm -rf $(EXE).dSYM
 	rm -f *.optrpt
-	rm -r pm
+	rm -rf pm
 
 #
 # -- Simple dependencies
@@ -102,3 +112,6 @@ pm_sycl.o : pm/pm_sycl.cpp pm/pm_sycl.h pm/pm.h
 pm_hip.o : pm/pm_hip.cpp pm/pm_hip.h pm/pm.h
 pm_openmp.o : pm/pm_openmp.cpp pm/pm_openmp.h pm/pm.h
 pm_host.o : pm/pm_host.cpp pm/pm_host.h pm/pm.h
+
+offload_cuda.o : offload/offload_cuda.cpp
+offload_sycy.o : offload/offload_sycl.cpp
diff --git a/gpu/src/Makefile b/gpu/src/Makefile
@@ -85,7 +85,8 @@ $(ML_OBJ): %.o: mathlib/%.cpp
 	$(CXX) $(CXXFLAGS) -c $<
 
 $(CUDA_OBJ): %.o: pm/%.cpp
-	$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
+	$(CUDA_CXX) $(CUDA_CXXFLAGS) -c $< -o $@
+	#$(CUDA_CXX) -x cu $(CUDA_CXXFLAGS) -c $< -o $@
 
 #
 # -- Remove *.o and *~ from the directory

diff --git a/gpu/src/arch/midway b/gpu/src/arch/midway
@@ -15,7 +15,7 @@ CXXFLAGS += -I$(CUDA_HOME)/include
 CXXFLAGS += -D_USE_NVTX
 
 CUDA_CXX = nvcc
-CUDA_CXXFLAGS = $(PYTHON_INC)
+CUDA_CXXFLAGS = -x cu $(PYTHON_INC)
 CUDA_CXXFLAGS += -arch=sm_37
 CUDA_CXXFLAGS += -Xcompiler -fopenmp
 #CUDA_CXXFLAGS += -shared -Xcompiler -fPIC 

diff --git a/gpu/src/arch/polaris-gnu-nvcc b/gpu/src/arch/polaris-gnu-nvcc
@@ -20,7 +20,7 @@ CXXFLAGS += -D_SIMPLE_TIMER
 CXXFLAGS += -D_USE_NVTX
 
 CUDA_CXX = nvcc
-CUDA_CXXFLAGS = $(PYTHON_INC)
+CUDA_CXXFLAGS = -x cu $(PYTHON_INC)
 #CUDA_CXXFLAGS += --dryrun
 CUDA_CXXFLAGS += -ccbin=CC
 CUDA_CXXFLAGS += -Xcompiler -fopenmp

diff --git a/gpu/src/arch/sophia-gnu-nvcc b/gpu/src/arch/sophia-gnu-nvcc
@@ -20,7 +20,7 @@ CXXFLAGS += -D_SIMPLE_TIMER
 CXXFLAGS += -D_USE_NVTX
 
 CUDA_CXX = nvcc
-CUDA_CXXFLAGS = $(PYTHON_INC)
+CUDA_CXXFLAGS = -x cu $(PYTHON_INC)
 CUDA_CXXFLAGS += -ccbin=mpicxx -arch=sm_80
 CUDA_CXXFLAGS += -Xcompiler -fopenmp
 CUDA_CXXFLAGS += -D_USE_GPU -D_GPU_CUDA -D_GPU_CUBLAS