Introducing new OFFLOAD_TARGET to controll cuda or hip backend in DBCSR (see https://github.com/cp2k/cp2k/pull/1582)

arch/CRAY-XC50_gnu.psmp

@@ -19,6 +19,7 @@ GNU_PATH    = /store/psi/psi01/mkrack/daint/gpu/gnu/8.3.0
 MPI_PATH    = $(GNU_PATH)/mpich
 
 GPUVER      = P100
+OFFLOAD_TARGET = cuda
 
 CC          = cc
 CXX         = CC

exts/Makefile.inc

@@ -7,12 +7,24 @@ EXTSDEPS_LIB  = $(LIBEXTSDIR)/dbcsr/libdbcsr$(ARCHIVE_EXT)
 EXTSDEPS_MOD = $(OBJEXTSDIR)/dbcsr/dbcsr_api.mod $(OBJEXTSDIR)/dbcsr/dbcsr_tensor_api.mod
 $(EXTSDEPS_MOD) : ; # override builtin .mod rule to prevent circular dependency
 
+# Set Acceleration flags
+DBCSR_USE_ACCEL := ""
+DBCSR_ACC       := ""
+DBCSR_ACCFLAGS  := ""
 ifneq (,$(findstring __DBCSR_ACC,$(OFFLOAD_FLAGS)))
-ACC := $(OFFLOAD_CC)
-ACCFLAGS := $(OFFLOAD_FLAGS)
+# By default, we use CUDA if OFFLOAD_TARGET is not specified
+ifeq (,$(OFFLOAD_TARGET))
+DBCSR_USE_ACCEL=cuda
 else
-ACC := ""
-ACCFLAGS := ""
+DBCSR_USE_ACCEL = $(OFFLOAD_TARGET)
+endif
+ifeq (cuda,$(DBCSR_USE_ACCEL))
+DBCSR_ACC      := $(OFFLOAD_CC)
+DBCSR_ACCFLAGS := $(OFFLOAD_FLAGS)
+else ifeq (hip,$(DBCSR_USE_ACCEL))
+DBCSR_ACC := $(CXX)
+DBCSR_ACCFLAGS := $(CXXFLAGS)
+endif
 endif
 
 extversions: dbcsrversion
@@ -25,14 +37,16 @@ dbcsr:
 	   ARCHFILE=$(ARCHDIR)/$(ARCH).$(ONEVERSION) \
 	   LIBDIR=$(LIBEXTSDIR)/$@ \
 	   OBJDIR=$(OBJEXTSDIR)/$@ \
-	   ACC="$(ACC)" \
-	   ACCFLAGS="$(ACCFLAGS)"
+	   USE_ACCEL="$(DBCSR_USE_ACCEL)" \
+	   ACC="$(DBCSR_ACC)" \
+	   ACCFLAGS="$(DBCSR_ACCFLAGS)"
 
 dbcsrversion:
 	@$(MAKE) -C $(EXTSHOME)/dbcsr -f $(EXTSHOME)/build_dbcsr/Makefile \
 	   ARCHFILE=$(ARCHDIR)/$(ARCH).$(ONEVERSION) \
-	   ACC="$(ACC)" \
-	   ACCFLAGS="$(ACCFLAGS)" \
+	   USE_ACCEL="$(DBCSR_USE_ACCEL)" \
+	   ACC="$(DBCSR_ACC)" \
+	   ACCFLAGS="$(DBCSR_ACCFLAGS)" \
 	   version
 
 dbcsrclean:
@@ -41,6 +55,7 @@ dbcsrclean:
 	   ARCHFILE=$(ARCHDIR)/$(ARCH).$(ONEVERSION) \
 	   LIBDIR=$(LIBEXTSDIR)/$@ \
 	   OBJDIR=$(OBJEXTSDIR)/$@ \
-	   ACC="$(ACC)" \
-	   ACCFLAGS="$(ACCFLAGS)" \
+	   USE_ACCEL="$(DBCSR_USE_ACCEL)" \
+	   ACC="$(DBCSR_ACC)" \
+	   ACCFLAGS="$(DBCSR_ACCFLAGS)" \
 	   clean

tools/toolchain/scripts/generate_arch_files.sh

@@ -239,6 +239,7 @@ gen_arch_file() {
 GPUVER        = \${GPUVER}
 OFFLOAD_CC    = \${NVCC}
 OFFLOAD_FLAGS = \${NVFLAGS}
+OFFLOAD_TARGET = cuda
 EOF
   fi
 
@@ -248,6 +249,7 @@ EOF
 GPUVER        = \${GPUVER}
 OFFLOAD_CC    = \${ROCM_PATH}/hip/bin/hipcc
 OFFLOAD_FLAGS = \${HIP_FLAGS} \${HIP_INCLUDES}
+OFFLOAD_TARGET = hip
 EOF
   fi