pulling in HIP support, and extended OpenMP offload support

makefiles/c_app.mk

@@ -9,7 +9,7 @@ APP_ENTRY_MPI_BASENAME := $(basename $(APP_ENTRY_MPI))
 APP_ENTRY_OP := $(APP_ENTRY_BASENAME)_op.cpp
 APP_ENTRY_MPI_OP := $(APP_ENTRY_MPI_BASENAME)_op.cpp
 
-ALL_VARIANTS := seq genseq vec openmp openmp4 cuda cuda_hyb
+ALL_VARIANTS := seq genseq vec openmp openmp4 cuda cuda_hyb hip
 ALL_VARIANTS += $(foreach variant,$(ALL_VARIANTS),mpi_$(variant))
 ALL_VARIANTS := $(foreach variant,$(ALL_VARIANTS),$(APP_NAME)_$(variant))
 
@@ -28,6 +28,10 @@ ifeq ($(HAVE_C),true)
   ifeq ($(HAVE_CUDA),true)
     BASE_BUILDABLE_VARIANTS += cuda cuda_hyb
   endif
+
+  ifeq ($(HAVE_HIP),true)
+    BASE_BUILDABLE_VARIANTS += hip
+  endif
 endif
 
 BUILDABLE_VARIANTS :=
@@ -69,7 +73,7 @@ all: $(BUILDABLE_VARIANTS)
 
 clean:
 	-$(RM) $(ALL_VARIANTS)
-	-$(RM) -r seq vec openmp openmp4 cuda openacc
+	-$(RM) -r seq vec openmp openmp4 cuda hip openacc
 	-$(RM) *_op.cpp
 	-$(RM) .generated .generated
 	-$(RM) *.d
@@ -109,6 +113,8 @@ CUDA_HYB_SRC := $(APP_ENTRY_OP) \
 MPI_CUDA_HYB_SRC := $(APP_ENTRY_MPI_OP) \
 	cuda/$(APP_NAME)_mpi_hybkernels_cpu.o cuda/$(APP_NAME)_mpi_hybkernels_gpu.o
 
+HIP_SRC := $(APP_ENTRY_OP) hip/$(APP_NAME)_kernels.o
+MPI_HIP_SRC := $(APP_ENTRY_MPI_OP) hip/$(APP_NAME)_mpi_kernels.o
 
 # $(1) = variant name
 # $(2) = additional flags
@@ -134,10 +140,14 @@ $(eval $(call RULE_template, openmp,   $(OMP_CPPFLAGS),
 $(eval $(call RULE_template, openmp4,  $(OMP_OFFLOAD_CPPFLAGS) -DOP2_WITH_OMP4, OPENMP4,    ))
 $(eval $(call RULE_template, cuda,,                                             CUDA,    MPI_CUDA))
 $(eval $(call RULE_template, cuda_hyb, $(OMP_CPPFLAGS),                         CUDA,    MPI_CUDA))
+$(eval $(call RULE_template, hip,,                                              HIP,     MPI_HIP))
 
 $(APP_NAME)_cuda: cuda/$(APP_NAME)_kernels.o
 $(APP_NAME)_mpi_cuda: cuda/$(APP_NAME)_mpi_kernels.o
 
+$(APP_NAME)_hip: hip/$(APP_NAME)_kernels.o
+$(APP_NAME)_mpi_hip: hip/$(APP_NAME)_mpi_kernels.o
+
 $(APP_NAME)_cuda_hyb: cuda/$(APP_NAME)_hybkernels_gpu.o cuda/$(APP_NAME)_hybkernels_cpu.o
 $(APP_NAME)_mpi_cuda_hyb: cuda/$(APP_NAME)_mpi_hybkernels_gpu.o cuda/$(APP_NAME)_mpi_hybkernels_cpu.o
 
@@ -147,6 +157,12 @@ cuda/$(APP_NAME)_kernels.o: .generated
 cuda/$(APP_NAME)_mpi_kernels.o: .generated
 	$(NVCC) $(NVCCFLAGS) $(OP2_INC) -c cuda/$(APP_ENTRY_MPI_BASENAME)_kernels.cu -o $@
 
+hip/$(APP_NAME)_kernels.o: .generated
+	$(HIPCXX) $(HIPFLAGS) $(OP2_INC) -c hip/$(APP_ENTRY_BASENAME)_kernels.cpp -o $@
+
+hip/$(APP_NAME)_mpi_kernels.o: .generated
+	$(HIPCXX) $(HIPFLAGS) $(OP2_INC) -c hip/$(APP_ENTRY_MPI_BASENAME)_kernels.cpp -o $@
+
 cuda/$(APP_NAME)_hybkernels_gpu.o: .generated
 	$(NVCC) $(NVCCFLAGS) -DOP_HYBRID_GPU -DGPUPASS $(OP2_INC) \
 		-c cuda/$(APP_ENTRY_BASENAME)_hybkernels.cu -o $@

makefiles/common.mk

@@ -78,6 +78,13 @@ ifeq ($(MAKECMDGOALS),config)
     $(call info_bold,> C/C++ CUDA compiler $(TEXT_NOTFOUND); skipping search for CUDA libraries)
   endif
 
+  ifeq ($(CONFIG_HAVE_C_HIP),true)
+    $(call info_bold,> C/C++ HIP compiler $(TEXT_FOUND) ($(CONFIG_HIP)); looking for the HIP libraries)
+    include $(DEPS_DIR)/hip.mk
+  else
+    $(call info_bold,> C/C++ HIP compiler $(TEXT_NOTFOUND); skipping search for HIP libraries)
+  endif
+
   $(info )
 
   ifeq ($(CONFIG_HAVE_MPI_C),true)
@@ -139,6 +146,7 @@ ifneq ($(MAKECMDGOALS),clean)
   $(info .   C: $(if $(HAVE_C),$(CC),not found))
   $(info .   C++: $(if $(HAVE_C),$(CXX),not found))
   $(info .   CUDA: $(if $(HAVE_C_CUDA),$(NVCC),not found))
+  $(info .   HIP:  $(if $(HAVE_C_HIP),$(HIPCXX),not found))
   $(info .   Fortran: $(if $(HAVE_F),$(FC),not found))
   $(info )
   $(info MPI compilers:)
@@ -161,14 +169,15 @@ ifneq ($(MAKECMDGOALS),clean)
   $(info .   C: $(CFLAGS))
   $(info .   C++: $(CXXFLAGS))
   $(info .   CUDA: $(NVCCFLAGS))
+  $(info .   HIP:  $(HIPFLAGS))
   $(info .   Fortran: $(FFLAGS))
   $(info )
 endif
 
-OP2_LIBS_SINGLE_NODE := seq cuda openmp openmp4
+OP2_LIBS_SINGLE_NODE := seq cuda hip openmp openmp4
 OP2_FOR_LIBS_SINGLE_NODE := $(foreach lib,$(OP2_LIBS_SINGLE_NODE),f_$(lib))
 
-OP2_LIBS_MPI := mpi mpi_cuda
+OP2_LIBS_MPI := mpi mpi_cuda mpi_hip
 OP2_FOR_LIBS_MPI := $(foreach lib,$(OP2_LIBS_MPI),f_$(lib))
 
 OP2_LIBS := hdf5 $(OP2_LIBS_SINGLE_NODE) $(OP2_LIBS_MPI)
@@ -199,3 +208,6 @@ $(foreach lib,$(OP2_LIBS_MPI),$(eval $(call OP2_LIB_template,$(lib),\
 
 OP2_LIB_CUDA += $(CUDA_LIB)
 OP2_LIB_MPI_CUDA += $(CUDA_LIB)
+
+OP2_LIB_HIP += $(HIP_LIB)
+OP2_LIB_MPI_HIP += $(HIP_LIB)

makefiles/compilers.mk

@@ -2,6 +2,7 @@ ifdef OP2_COMPILER
   OP2_C_COMPILER ?= $(OP2_COMPILER)
   OP2_F_COMPILER ?= $(OP2_COMPILER)
   OP2_C_CUDA_COMPILER ?= nvhpc
+  OP2_C_HIP_COMPILER ?= hip
 endif
 
 # Process CUDA_GEN and NV_ARCH until CUDA_GEN is a whitespace separated list of
@@ -28,6 +29,10 @@ ifdef OP2_C_CUDA_COMPILER
   include $(MAKEFILES_DIR)/compilers/c_cuda/$(OP2_C_CUDA_COMPILER).mk
 endif
 
+ifdef OP2_C_HIP_COMPILER
+  include $(MAKEFILES_DIR)/compilers/c_hip/$(OP2_C_HIP_COMPILER).mk
+endif
+
 ifdef OP2_F_COMPILER
   include $(MAKEFILES_DIR)/compilers/fortran/$(OP2_F_COMPILER).mk
 endif
@@ -48,6 +53,13 @@ ifneq ($(shell which $(CONFIG_NVCC) 2> /dev/null),)
   CONFIG_HAVE_C_CUDA := true
 endif
 
+ifneq ($(shell which $(CONFIG_HIP) 2> /dev/null),)
+  CONFIG_HIP != which $(CONFIG_HIP)
+  CONFIG_HIPCXX = $(CONFIG_HIP)
+  CONFIG_HIPCXX = $(CONFIG_HIP)
+  CONFIG_HAVE_C_HIP := true
+endif
+
 ifneq ($(shell which $(CONFIG_FC) 2> /dev/null),)
   CONFIG_FC != which $(CONFIG_FC)
   CONFIG_HAVE_F := true

makefiles/compilers/c/clang.mk

@@ -4,3 +4,7 @@ CONFIG_CC := clang
 CONFIG_CXX := clang++
 
 CONFIG_CXXLINK ?= -lc++
+
+CONFIG_CPP_HAS_OMP_OFFLOAD ?= true
+OMP_OFFLOAD_CXXFLAGS = -target x86_64-pc-linux-gnu -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a
+CONFIG_OMP_OFFLOAD_CXXFLAGS = -target x86_64-pc-linux-gnu -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a

makefiles/compilers/c/cray.mk

@@ -23,4 +23,6 @@ CONFIG_OMP_CPPFLAGS ?= -fopenmp
 CONFIG_CPP_HAS_OMP ?= true
 
 # CONFIG_OMP_OFFLOAD_CPPFLAGS ?=
-CONFIG_CPP_HAS_OMP_OFFLOAD ?= false
+CONFIG_CPP_HAS_OMP_OFFLOAD ?= true
+OMP_OFFLOAD_CXXFLAGS = -fopenmp
+CONFIG_OMP_OFFLOAD_CXXFLAGS = -fopenmp

makefiles/compilers/c/nvhpc.mk

@@ -20,5 +20,5 @@ CONFIG_CPP_HAS_OMP ?= true
 GPU_FFLAG := -gpu=fastmath,ptxinfo,lineinfo
 $(foreach arch,$(CUDA_GEN),$(eval GPU_FFLAG := $(GPU_FFLAG),cc$(arch)))
 
-CONFIG_OMP_OFFLOAD_CPPFLAGS ?= -mp=gpu $(GPU_FFLAG)
+CONFIG_OMP_OFFLOAD_CXXFLAGS ?= -mp=gpu $(GPU_FFLAG)
 CONFIG_CPP_HAS_OMP_OFFLOAD ?= true

makefiles/compilers/c_hip/hip.mk

@@ -0,0 +1,9 @@
+CONFIG_HIP ?= hipcc
+
+ifndef DEBUG
+  HIP_OPT := -Ofast
+else
+  HIP_OPT := -g -O0
+endif
+
+CONFIG_HIPFLAGS ?= -x hip --offload-arch=$(HIP_ARCH) $(HIP_OPT)

makefiles/dependencies/hip.mk

@@ -0,0 +1,33 @@
+ifdef HIP_INSTALL_PATH
+  HIP_INC_PATH := -I$(HIP_INSTALL_PATH)/include
+  HIP_LIB_PATH := -L$(HIP_INSTALL_PATH)/lib64 -L$(HIP_INSTALL_PATH)/lib
+endif
+
+HIP_TEST = $(CONFIG_HIP) $(HIP_INC_PATH) \
+                    $(DEPS_DIR)/tests/hip.cpp $(HIP_LIB_PATH) $(HIP_LINK) \
+                    -o $(DEPS_DIR)/tests/hip
+
+$(file > $(DEP_BUILD_LOG),$(HIP_TEST))
+$(shell $(HIP_TEST) >> $(DEP_BUILD_LOG) 2>&1)
+
+ifneq ($(.SHELLSTATUS),0)
+  HIP_LINK ?= #-lculibos -lpthread -lrt -ldl
+
+  $(file >> $(DEP_BUILD_LOG),$(HIP_TEST))
+  $(shell $(HIP_TEST) >> $(DEP_BUILD_LOG) 2>&1)
+endif
+
+ifeq ($(.SHELLSTATUS),0)
+  $(shell rm -f $(DEPS_DIR)/tests/hip)
+
+  $(call info_bold,  > HIP libraries $(TEXT_FOUND) (link flags: $(or $(HIP_LINK), none)))
+
+  CONFIG_HAVE_HIP := true
+
+  CONFIG_HIP_INC := $(strip $(HIP_INC_PATH) $(HIP_DEF))
+  CONFIG_HIP_LIB := $(strip $(HIP_LIB_PATH) $(HIP_LINK))
+else
+  $(call info_bold,  > HIP libraries $(TEXT_NOTFOUND):)
+  $(info $(file < $(DEP_BUILD_LOG)))
+  $(info )
+endif

makefiles/dependencies/tests/hip.cpp

@@ -0,0 +1,7 @@
+#include <hip/hip_runtime.h>
+
+int main() {
+    int deviceCount = 0;
+    auto err = hipGetDeviceCount(&deviceCount);
+    return err;
+}

op2/Makefile

@@ -32,6 +32,14 @@ ifeq ($(HAVE_C),true)
     endif
   endif
 
+  ifeq ($(HAVE_HIP),true)
+    OP2_BUILDABLE_LIBS += hip
+
+    ifeq ($(HIP_IS_BUILDABLE),true)
+      OP2_BUILDABLE_LIBS += mpi_hip
+    endif
+  endif
+
   ifeq ($(MPI_IS_BUILDABLE),true)
     OP2_BUILDABLE_LIBS += mpi
   endif
@@ -110,6 +118,10 @@ OP2_CUDA := $(OP2_BASE) $(addprefix $(OBJ)/,\
 	cuda/op_cuda_decl.o \
 	cuda/op_cuda_rt_support.o)
 
+OP2_HIP := $(OP2_BASE) $(addprefix $(OBJ)/,\
+	hip/op_hip_decl.o \
+	hip/op_hip_rt_support.o)
+
 OP2_FOR_CUDA := $(OP2_CUDA) $(OP2_FOR_BASE_CUDA) $(addprefix $(OBJ)/fortran/,\
 	op2_for_rt_wrappers_cuda.o \
 	cudaConfigurationParams.o)
@@ -153,6 +165,18 @@ OP2_MPI_CUDA := $(OP2_BASE) $(addprefix $(OBJ)/,\
 	externlib/op_util.o \
 	externlib/op_renumber.o)
 
+OP2_MPI_HIP := $(OP2_BASE) $(addprefix $(OBJ)/,\
+        hip/op_hip_rt_support+mpi.o \
+        mpi/op_mpi_core.o \
+        mpi/op_mpi_part_core.o \
+        mpi/op_mpi_hip_decl.o \
+        mpi/op_mpi_hip_rt_support.o \
+        mpi/op_mpi_hip_kernels.o \
+        mpi/op_mpi_hdf5.o \
+        mpi/op_mpi_util.o \
+        externlib/op_util.o \
+        externlib/op_renumber.o)
+
 OP2_FOR_MPI_CUDA := $(OP2_MPI_CUDA) $(OP2_FOR_BASE_MPI_CUDA) $(addprefix $(OBJ)/fortran/,\
 	op2_for_rt_wrappers_cuda.o \
 	cudaConfigurationParams.o)
@@ -186,6 +210,12 @@ $(OBJ)/cuda/%+mpi.o: src/cuda/%.cpp | $(OBJ)
 $(OBJ)/cuda/%.o: src/cuda/%.cpp | $(OBJ)
 	$(CXX) $(CXXFLAGS) $(INC) -DSET_CUDA_CACHE_CONFIG -c $< -o $@
 
+$(OBJ)/hip/%+mpi.o: src/hip/%.cpp | $(OBJ)
+	$(MPICXX) $(CXXFLAGS) $(HIPFLAGS) $(INC) -DOPMPI -c $< -o $@
+
+$(OBJ)/hip/%.o: src/hip/%.cpp | $(OBJ)
+	$(CXX) $(CXXFLAGS) $(HIPFLAGS) $(INC) -c $< -o $@
+
 $(OBJ)/openmp4/%.o: src/openmp4/%.cpp | $(OBJ)
 	$(CXX) $(CXXFLAGS) $(OMP_OFFLOAD_CXXFLAGS) $(INC) -c $< -o $@
 
@@ -195,6 +225,15 @@ $(OBJ)/mpi/%.o: src/mpi/%.cpp | $(OBJ)
 $(OBJ)/mpi/%.o: src/mpi/%.cu | $(OBJ)
 	$(NVCC) $(NVCCFLAGS) $(INC) -c $< -o $@
 
+$(OBJ)/mpi/op_mpi_hip_decl.o: src/mpi/op_mpi_hip_decl.cpp | $(OBJ)
+	$(MPICXX) $(CXXFLAGS) $(HIPFLAGS) $(INC) -c $< -o $@
+
+$(OBJ)/mpi/op_mpi_hip_rt_support.o: src/mpi/op_mpi_hip_rt_support.cpp | $(OBJ)
+	$(MPICXX) $(CXXFLAGS) $(HIPFLAGS) $(INC) -c $< -o $@
+
+$(OBJ)/mpi/op_mpi_hip_kernels.o: src/mpi/op_mpi_hip_kernels.cpp | $(OBJ)
+	$(MPICXX) $(CXXFLAGS) $(HIPFLAGS) $(INC) -c $< -o $@
+
 $(OBJ)/fortran/%+mpi.o: src/fortran/%.c | $(OBJ)
 	$(MPICC) $(CFLAGS) $(INC) -DOPMPI -c $< -o $@
 

op2/include/op_cuda_rt_support.h

@@ -84,9 +84,9 @@ void cutilDeviceInit_mpi(int argc, char **argv, int mpi_rank);
  * routines to move arrays to/from GPU device
  */
 
-void op_mvHostToDevice(void **map, int size);
+void op_mvHostToDevice(void **map, size_t size);
 
-void op_cpHostToDevice(void **data_d, void **data_h, int size);
+void op_cpHostToDevice(void **data_d, void **data_h, size_t size);
 
 void op_cuda_get_data(op_dat dat);