diff --git a/Kokkos.make b/Kokkos.make index e46236e..7dd6af8 100644 --- a/Kokkos.make +++ b/Kokkos.make @@ -5,7 +5,7 @@ ifndef DEVICE define device_help Set DEVICE to change flags (defaulting to OpenMP). Available devices are: - OpenMP, Serial, Pthreads, Cuda + OpenMP, Serial, Pthreads, Cuda, HIP endef $(info $(device_help)) @@ -36,7 +36,9 @@ ifndef COMPILER define compiler_help Set COMPILER to change flags (defaulting to GNU). Available compilers are: - GNU INTEL CRAY PGI ARMCLANG + GNU INTEL CRAY PGI ARMCLANG HIPCC + + Note: you may have to do `export CXX=\path\to\hipcc` in case Kokkos detects the wrong compiler endef $(info $(compiler_help)) @@ -44,6 +46,7 @@ COMPILER=GNU endif COMPILER_ARMCLANG = armclang++ +COMPILER_HIPCC = hipcc COMPILER_GNU = g++ COMPILER_INTEL = icpc -qopt-streaming-stores=always COMPILER_CRAY = CC @@ -56,14 +59,17 @@ Set TARGET to change to offload device. Defaulting to CPU. Available targets are: CPU (default) GPU + endef $(info $(target_help)) TARGET=CPU endif ifeq ($(TARGET), GPU) +ifneq ($(COMPILER), HIPCC) CXX = $(NVCC_WRAPPER) endif +endif OBJ = main.o KokkosStream.o CXXFLAGS = -O3 diff --git a/OpenMP.make b/OpenMP.make index 753b14f..a3ec90f 100644 --- a/OpenMP.make +++ b/OpenMP.make @@ -16,7 +16,7 @@ ifndef TARGET define target_help Set TARGET to change device (defaulting to CPU). Available targets are: - CPU NVIDIA AMD + CPU NVIDIA AMD INTEL_GPU endef $(info $(target_help)) @@ -36,7 +36,7 @@ CXX = $(COMPILER_$(COMPILER)) FLAGS_GNU = -O3 -std=c++11 -march=native FLAGS_GNU_PPC = -O3 -std=c++11 -mcpu=native -FLAGS_INTEL = -O3 -std=c++11 -xHOST -qopt-streaming-stores=always +FLAGS_INTEL = -O3 -std=c++11 FLAGS_CRAY = -O3 -std=c++11 FLAGS_CLANG = -O3 -std=c++11 FLAGS_XL = -O5 -qarch=auto -qtune=auto -std=c++11 @@ -61,6 +61,9 @@ OMP_CLANG_NVIDIA = -DOMP_TARGET_GPU -fopenmp=libomp -fopenmp-targets=nvptx64-nvi OMP_GNU_NVIDIA = -DOMP_TARGET_GPU -fopenmp -foffload=nvptx-none OMP_GNU_AMD = -DOMP_TARGET_GPU -fopenmp -foffload=amdgcn-amdhsa +OMP_INTEL_CPU = -xHOST -qopt-streaming-stores=always +OMP_INTEL_INTEL_GPU = -DOMP_TARGET_GPU -qnextgen -fiopenmp -fopenmp-targets=spir64 + ifndef OMP_$(COMPILER)_$(TARGET) $(error Targeting $(TARGET) with $(COMPILER) not supported) endif diff --git a/SYCL.make b/SYCL.make index 3572681..4326da5 100644 --- a/SYCL.make +++ b/SYCL.make @@ -1,17 +1,97 @@ +ifndef COMPILER +define compiler_help +Set COMPILER to change flags (defaulting to GNU). +Available compilers are: + HIPSYCL, DPCPP, COMPUTECPP + -COMPUTECPP_FLAGS = $(shell $(COMPUTECPP_PACKAGE_ROOT_DIR)/bin/computecpp_info --dump-device-compiler-flags) + For HIPSYCL and COMPUTECPP, SYCL_SDK_DIR must be specified, the directory should contain [/lib, /bin, ...] + For DPCPP, the compiler must be on path +endef +$(info $(compiler_help)) +COMPILER=HIPSYCL +endif -sycl-stream: main.o SYCLStream.o SYCLStream.sycl - $(CXX) -O3 -std=c++11 -DSYCL main.o SYCLStream.o $(EXTRA_FLAGS) -L$(COMPUTECPP_PACKAGE_ROOT_DIR)/lib -lComputeCpp -lOpenCL -Wl,--rpath=$(COMPUTECPP_PACKAGE_ROOT_DIR)/lib/ -o $@ +ifndef TARGET +define target_help +Set TARGET to change device (defaulting to CPU). +Available targets are: + CPU AMD NVIDIA + +endef +$(info $(target_help)) +TARGET=CPU +endif + + +ifndef ARCH +define arch_help +Set ARCH to change device (defaulting to ""). +(GPU *only*) Available targets for HIPSYCL are: + For CUDA, the architecture has the form sm_XX, e.g. sm_60 for Pascal. + For ROCm, the architecture has the form gfxYYY, e.g. gfx900 for Vega 10, gfx906 for Vega 20. + +endef + +ifneq ($(COMPILER), DPCPP) +$(info $(arch_help)) +ARCH= + +endif + +endif + +SYCL_COMPUTECPP_SYCLFLAGS = $(shell $(SYCL_SDK_DIR)/bin/computecpp_info --dump-device-compiler-flags) +SYCL_COMPUTECPP_SYCLFLAGS_AMD = $(SYCL_COMPUTECPP_SYCLFLAGS) +SYCL_COMPUTECPP_SYCLFLAGS_CPU = $(SYCL_COMPUTECPP_SYCLFLAGS) +SYCL_COMPUTECPP_SYCLFLAGS_NVIDIA = $(SYCL_COMPUTECPP_SYCLFLAGS) -sycl-target ptx64 +SYCL_COMPUTECPP_SYCLCXX = $(SYCL_SDK_DIR)/bin/compute++ +SYCL_COMPUTECPP_FLAGS = -O3 --std=c++17 +SYCL_COMPUTECPP_LINK_FLAGS = -L$(SYCL_SDK_DIR)/lib -lComputeCpp -lOpenCL -Wl,--rpath=$(SYCL_SDK_DIR)/lib/ +SYCL_COMPUTECPP_INCLUDE = -I$(SYCL_SDK_DIR)/include +SYCL_COMPUTECPP_CXX = g++ +SYCL_COMPUTECPP_DEPS = SYCLStream.sycl + +SYCL_HIPSYCL_SYCLFLAGS_CPU = -O3 --std=c++17 --hipsycl-platform=cpu +SYCL_HIPSYCL_SYCLFLAGS_AMD = -O3 --std=c++17 --hipsycl-platform=rocm --hipsycl-gpu-arch=$(ARCH) +SYCL_HIPSYCL_SYCLFLAGS_NVIDIA = -O3 --std=c++17 --hipsycl-platform=cuda --hipsycl-gpu-arch=$(ARCH) +SYCL_HIPSYCL_SYCLCXX = $(SYCL_SDK_DIR)/bin/syclcc +SYCL_HIPSYCL_FLAGS = $(SYCL_HIPSYCL_SYCLFLAGS_$(TARGET)) +SYCL_HIPSYCL_LINK_FLAGS = -L$(SYCL_SDK_DIR)/lib -Wl,-rpath,$(SYCL_SDK_DIR)/lib +SYCL_HIPSYCL_INCLUDE = +SYCL_HIPSYCL_CXX = $(SYCL_HIPSYCL_SYCLCXX) +SYCL_HIPSYCL_DEPS = + +SYCL_DPCPP_SYCLFLAGS_CPU = -O3 --std=c++17 +SYCL_DPCPP_SYCLFLAGS_NVIDIA = -O3 --std=c++17 -fsycl -fsycl-targets=nvptx64-nvidia-cuda-sycldevice -fsycl-unnamed-lambda +SYCL_DPCPP_SYCLCXX = dpcpp +SYCL_DPCPP_FLAGS = $(SYCL_DPCPP_SYCLFLAGS_CPU) +SYCL_DPCPP_LINK_FLAGS = +SYCL_DPCPP_INCLUDE = +SYCL_DPCPP_CXX = dpcpp +SYCL_DPCPP_DEPS = + + +SYCL_SYCLFLAGS = $(SYCL_$(COMPILER)_SYCLFLAGS_$(TARGET)) +SYCL_SYCLCXX = $(SYCL_$(COMPILER)_SYCLCXX) + +SYCL_FLAGS = $(SYCL_$(COMPILER)_FLAGS) +SYCL_LINK_FLAGS = $(SYCL_$(COMPILER)_LINK_FLAGS) +SYCL_INCLUDE = $(SYCL_$(COMPILER)_INCLUDE) +SYCL_CXX = $(SYCL_$(COMPILER)_CXX) +SYCL_DEPS = $(SYCL_$(COMPILER)_DEPS) + +sycl-stream: main.o SYCLStream.o $(SYCL_DEPS) + $(SYCL_CXX) $(SYCL_FLAGS) -DSYCL main.o SYCLStream.o $(EXTRA_FLAGS) $(SYCL_LINK_FLAGS) -o $@ main.o: main.cpp - $(CXX) -O3 -std=c++11 -DSYCL main.cpp -c -I$(COMPUTECPP_PACKAGE_ROOT_DIR)/include $(EXTRA_FLAGS) -o $@ + $(SYCL_CXX) $(SYCL_FLAGS) -DSYCL main.cpp -c $(SYCL_INCLUDE) $(EXTRA_FLAGS) -o $@ -SYCLStream.o: SYCLStream.cpp SYCLStream.sycl - $(CXX) -O3 -std=c++11 -DSYCL SYCLStream.cpp -c -I$(COMPUTECPP_PACKAGE_ROOT_DIR)/include -include SYCLStream.sycl $(EXTRA_FLAGS) -o $@ +SYCLStream.o: SYCLStream.cpp $(SYCL_DEPS) + $(SYCL_CXX) $(SYCL_FLAGS) -DSYCL SYCLStream.cpp -c $(SYCL_INCLUDE) $(EXTRA_FLAGS) -o $@ SYCLStream.sycl: SYCLStream.cpp - $(COMPUTECPP_PACKAGE_ROOT_DIR)/bin/compute++ -DSYCL SYCLStream.cpp $(COMPUTECPP_FLAGS) -c -I$(COMPUTECPP_PACKAGE_ROOT_DIR)/include -o $@ + $(SYCL_SYCLCXX) -DSYCL SYCLStream.cpp $(SYCL_SYCLFLAGS) -c $(SYCL_INCLUDE) -o $@ .PHONY: clean clean: