add Cray compiler to build system

- ignore temp files generated by Cray Fortran
- workaround Cray not having reduce

.gitignore

@@ -20,6 +20,8 @@ src/fortran/BabelStream.*
 *.gz
 *.a
 *.mod
+*.cub
+*.ptx
 
 KokkosCore_config.*
 

src/fortran/DoConcurrentStream.F90

@@ -127,7 +127,11 @@ function dot() result(s)
             integer(kind=StreamIntKind) :: i
             ! reduction omitted because NVF infers it and other compilers do not support
             s = real(0,kind=REAL64)
+#ifdef CRAY_THREAD_DOCONCURRENT
+            do i=1,N
+#else
             do concurrent (i=1:N) shared(A,B)
+#endif
                s = s + A(i) * B(i)
             end do
         end function dot

src/fortran/Makefile

@@ -4,6 +4,8 @@ else ifeq ($(COMPILER),oneapi)
     include make.inc.oneapi
 else ifeq ($(COMPILER),gcc)
     include make.inc.gcc
+else ifeq ($(COMPILER),cray)
+    include make.inc.cray
 else
     $(info Set COMPILER={nvhpc,oneapi}. Default is gcc.)
     include make.inc.gcc

src/fortran/OpenMPTargetStream.F90

@@ -101,6 +101,7 @@ subroutine copy()
             do i=1,N
                C(i) = A(i)
             end do
+            !$omp barrier
         end subroutine copy
 
         subroutine add()

src/fortran/build.sh

@@ -1,13 +1,6 @@
 #!/bin/bash
 
-for compiler in oneapi ; do
-    for implementation in DoConcurrent Array OpenMP OpenMPWorkshare OpenMPTarget OpenMPTargetLoop OpenMPTaskloop ; do
-        make COMPILER=${compiler} IMPLEMENTATION=${implementation}
-    done
-done
-#exit
-for compiler in nvhpc ; do
-    for implementation in DoConcurrent Array OpenMP OpenMPWorkshare OpenMPTarget OpenMPTargetLoop OpenMPTaskloop OpenACC OpenACCArray CUDA CUDAKernel ; do
-        make COMPILER=${compiler} IMPLEMENTATION=${implementation}
-    done
+compiler=cray
+for implementation in DoConcurrent Array OpenMP OpenMPWorkshare OpenMPTarget OpenMPTargetLoop OpenMPTaskloop OpenACC OpenACCArray  ; do
+    make COMPILER=${compiler} IMPLEMENTATION=${implementation}
 done

src/fortran/make.inc.cray

@@ -0,0 +1,16 @@
+FC	:= ftn
+FCFLAGS	 = -e F -O3
+
+DOCONCURRENT_FLAG   = -h thread_do_concurrent -DCRAY_THREAD_DOCONCURRENT
+ARRAY_FLAG          = -h autothread
+OPENMP_FLAG         = -h omp
+OPENACC_FLAG        = -h acc
+CUDA_FLAG           =
+SEQUENTIAL_FLAG     =
+
+ifeq ($(IMPLEMENTATION),CUDA)
+    $(error IMPLEMENTATION=$(IMPLEMENTATION) is not supported by this compiler.)
+endif
+ifeq ($(IMPLEMENTATION),CUDAKernels)
+    $(error IMPLEMENTATION=$(IMPLEMENTATION) is not supported by this compiler.)
+endif

src/fortran/run.sh

@@ -1,22 +1,18 @@
 #!/bin/bash
 
-export KMP_HW_PLACES=1s,4c,1t
-export KMP_HW_SUBSET=1s,4c,1t
-export KMP_AFFINITY=compact,granularity=fine #,verbose
+M=64
 
+compiler=cray
 # CPU
-for compiler in oneapi ; do
-    for implementation in DoConcurrent Array OpenMP OpenMPWorkshare OpenMPTaskloop ; do
-        if [ -f BabelStream.${compiler}.${implementation} ] ; then
-            ./BabelStream.${compiler}.${implementation} -s $((1024*1024*64))
-        fi
-    done
+for implementation in DoConcurrent Array OpenMP OpenMPWorkshare OpenMPTaskloop ; do
+    if [ -f BabelStream.${compiler}.${implementation} ] ; then
+        ./BabelStream.${compiler}.${implementation} -s $((1024*1024*${M}))
+    fi
 done
+exit
 # GPU
-for compiler in nvhpc ; do
-    for implementation in DoConcurrent OpenMPTarget OpenMPTargetLoop OpenACC OpenACCArray CUDA CUDAKernel ; do
-        if [ -f BabelStream.${compiler}.${implementation} ] ; then
-            ./BabelStream.${compiler}.${implementation} -s $((1024*1024*64))
-        fi
-    done
+for implementation in OpenMPTarget OpenMPTargetLoop OpenACC OpenACCArray ; do
+    if [ -f BabelStream.${compiler}.${implementation} ] ; then
+        ./BabelStream.${compiler}.${implementation} -s $((1024*1024*${M}))
+    fi
 done