Exploring full kokkos builds w/ clang.

tarunprabhu · Nov 9, 2023 · 74cb34f · 74cb34f
1 parent 4dc3221
commit 74cb34f
Show file tree

Hide file tree

Showing 4 changed files with 19 additions and 6 deletions.
diff --git a/kitsune/experiments/inc/cuda.mk b/kitsune/experiments/inc/cuda.mk
@@ -11,7 +11,7 @@ ifneq ($(CUDA_PATH),)
     -O$(KITSUNE_OPTLEVEL)
 
   CLANG_CUDA=$(KITSUNE_PREFIX)/bin/clang 
-  CLANG_CUDA_FLAGS=--language=cuda --no-cuda-version-check --cuda-gpu-arch=$(CUDA_ARCH) \
+  CLANG_CUDA_FLAGS=--language=cuda -x cuda --no-cuda-version-check --cuda-gpu-arch=$(CUDA_ARCH) \
     -O$(KITSUNE_OPTLEVEL)
 
   BUILD_CUDA_EXPERIMENTS=true

diff --git a/kitsune/experiments/inc/kokkos.mk b/kitsune/experiments/inc/kokkos.mk
@@ -4,7 +4,11 @@ ifeq ($(BUILD_CUDA_EXPERIMENTS),true)
   KOKKOS_CUDA_LIBS=-L$(KOKKOS_CUDA_PREFIX)/lib -L$(CUDA_PATH)/lib64 -lkokkoscore -lcudart -ldl
   KOKKOS_NVCC=$(KOKKOS_CUDA_PREFIX)/bin/nvcc_wrapper
   KOKKOS_NVCC_FLAGS?= $(NVCC_CXX_FLAGS) -I$(KOKKOS_CUDA_PREFIX)/include/
-  KOKKOS_CLANG_CUDA_FLAGS?= $(CLANG_CUDA_FLAGS) -std=c++17 -I$(KOKKOS_CUDA_PREFIX)/include/
+
+  KOKKOS_CLANG_CUDA=$(KITSUNE_PREFIX)/bin/clang
+  KOKKOS_CLANG_CUDA_PREFIX?=$(KITSUNE_PREFIX)/opt/kokkos/cuda-clang
+  KOKKOS_CLANG_CUDA_FLAGS?= $(CLANG_CUDA_FLAGS) -std=c++17 -I$(KOKKOS_CLANG_CUDA_PREFIX)/include/
+  KOKKOS_CLANG_CUDA_LIBS=-L$(KOKKOS_CLANG_CUDA_PREFIX)/lib64 -L$(CUDA_PATH)/lib64 -lkokkoscore -lcudart -ldl -lstdc++
 endif
 
 ifeq ($(BUILD_HIP_EXPERIMENTS),true)

diff --git a/kitsune/experiments/raytracer/makefile b/kitsune/experiments/raytracer/makefile
@@ -53,9 +53,9 @@ raytracer-kokkos.hipcc.${host_arch}: raytracer-kokkos.cpp
 
 # cuda version 
 raytracer-cuda.${host_arch}: raytracer.cu
-	@echo $@
-	@$(TIME_CMD) $(NVCC) $(NVCC_CXX_FLAGS) -o $@ $<
-	@$(FILE_SIZE)
+	echo $@
+	$(TIME_CMD) $(NVCC) $(NVCC_CXX_FLAGS) -o $@ $<
+	$(FILE_SIZE)
 
 # hip version 
 raytracer-hip.${host_arch}: raytracer-hip.cpp

diff --git a/kitsune/experiments/saxpy/makefile b/kitsune/experiments/saxpy/makefile
@@ -4,10 +4,13 @@ targets = saxpy-forall.opencilk.${host_arch}
 
 ifeq ($(BUILD_CUDA_EXPERIMENTS),true)
   targets += saxpy-forall.cuda.${host_arch}
+  targets += saxpy-kokkos.nvcc.${host_arch}
+  targets += saxpy-kokkos.clang.${host_arch}
 endif
 
 ifeq ($(BUILD_HIP_EXPERIMENTS),true)
   targets += saxpy-forall.hip.${host_arch}
+  targets += saxpy-kokkos.hip.${host_arch}
 endif
 
 all: ${targets}
@@ -19,7 +22,7 @@ saxpy-forall.opencilk.${host_arch}: saxpy-forall.cpp
 	@$(FILE_SIZE)
 saxpy-forall.cuda.${host_arch}: saxpy-forall.cpp 
 	@echo $@
-	@$(TIME_CMD)  $(KIT_CXX) $(TAPIR_CUDA_FLAGS) -o $@ $< -Xlinker -rpath=$(KITSUNE_PREFIX)/lib
+	$(TIME_CMD)  $(KIT_CXX) -ftime-trace=. $(TAPIR_CUDA_FLAGS) -o $@ $< -Xlinker -rpath=$(KITSUNE_PREFIX)/lib
 	@$(FILE_SIZE)
 saxpy-forall.hip.${host_arch}: saxpy-forall.cpp 
 	@echo $@
@@ -36,6 +39,12 @@ saxpy-kokkos.nvcc.${host_arch}: saxpy-kokkos.cpp
 	@echo $@
 	@$(TIME_CMD) $(KOKKOS_NVCC) $(KOKKOS_NVCC_FLAGS) -o $@ $< $(KOKKOS_CUDA_LIBS) -Xlinker -rpath=$(KOKKOS_CUDA_PREFIX)/lib64
 	@$(FILE_SIZE)
+
+saxpy-kokkos.clang.${host_arch}: saxpy-kokkos.cpp 
+	@echo $@
+	$(TIME_CMD) $(CLANG_CUDA) $(KOKKOS_CLANG_CUDA_FLAGS) -ftime-trace=. -fno-exceptions -o $@ $< $(KOKKOS_CUDA_LIBS) -Xlinker -rpath=$(KOKKOS_CUDA_PREFIX)/lib64
+	@$(FILE_SIZE)
+
 saxpy-kokkos.hipcc.${host_arch}: saxpy-kokkos.cpp 
 	@echo $@
 	@$(TIME_CMD) $(KOKKOS_HIPCC) $(KOKKOS_HIP_FLAGS) -o $@ $< $(KOKKOS_HIP_LIBS) -Xlinker -rpath=$(KOKKOS_HIP_PREFIX)/lib64