From cc71b45e76b0910c75a6c5f81b90944239f1aace Mon Sep 17 00:00:00 2001
From: Pietro Bonfa <p.bonfa@nospam.com>
Date: Tue, 27 Feb 2018 13:54:17 +0100
Subject: [PATCH] Added test for mp_circular_shift_left (GPU code only)

---
 UtilXlib/tests/test_mp_bcast_im_gpu.f90       |  2 +-
 .../test_mp_circular_shift_left_gpu.tmpl      | 73 +++++++++++++++++++
 2 files changed, 74 insertions(+), 1 deletion(-)
 create mode 100644 UtilXlib/tests/test_mp_circular_shift_left_gpu.tmpl

diff --git a/UtilXlib/tests/test_mp_bcast_im_gpu.f90 b/UtilXlib/tests/test_mp_bcast_im_gpu.f90
index ba4ee1544..e1b7ad69f 100644
--- a/UtilXlib/tests/test_mp_bcast_im_gpu.f90
+++ b/UtilXlib/tests/test_mp_bcast_im_gpu.f90
@@ -1,7 +1,7 @@
 #if defined(__CUDA)
 PROGRAM test_mp_bcast_im_gpu
 !
-! Simple program to check the functionalities of test_mp_bcast_i1.
+! Simple program to check the functionalities of test_mp_bcast_im.
 !
     USE cudafor
 #if defined(__MPI)
diff --git a/UtilXlib/tests/test_mp_circular_shift_left_gpu.tmpl b/UtilXlib/tests/test_mp_circular_shift_left_gpu.tmpl
new file mode 100644
index 000000000..fd607deeb
--- /dev/null
+++ b/UtilXlib/tests/test_mp_circular_shift_left_gpu.tmpl
@@ -0,0 +1,73 @@
+! Implemented: i1, iv, im, rm, cm
+
+#if defined(__CUDA)
+
+PROGRAM test_mp_circular_shift_left_{vname}_gpu
+!
+! Simple program to check the functionalities of test_mp_circular_shift_left.
+!  N.B : mp_circular_shift_left is doing a allreduce!
+!
+    USE cudafor
+#if defined(__MPI)
+    USE MPI
+#endif
+    USE mp, ONLY : mp_circular_shift_left
+    USE mp_world, ONLY : mp_world_start, mp_world_end, mpime, &
+                          root, nproc, world_comm
+    USE tester
+    IMPLICIT NONE
+    !
+    TYPE(tester_t) :: test
+    INTEGER :: world_group = 0
+    INTEGER, PARAMETER :: datasize = {datasize}
+    !
+    ! Stuff for comparing with CPU implementation
+    integer :: i
+    REAL(8) :: rnd{size}
+    !
+    ! test variable
+    {type}, DEVICE :: {vname}_d{size}
+    {type} :: {vname}_h{size}
+    {type} :: aux_h{size}
+    
+    !
+    CALL test%init()
+    
+#if defined(__MPI)    
+    world_group = MPI_COMM_WORLD
+#endif
+    CALL mp_world_start(world_group)
+    {vname}_h = mpime
+    {vname}_d = {vname}_h
+    !mp_circular_shift_left    (msg_dest,  msg_sour, mpime, dest, sour, ip, gid)
+    CALL mp_circular_shift_left({vname}_d, 0, world_comm)
+    {vname}_h = {vname}_d
+    !
+    CALL test%assert_{compare}({sumf}({vname}_h - {typeconv}(MODULO(mpime+1, nproc))), {typeconv}(0) )
+    !
+    !
+    ! Test against CPU implementation
+    CALL save_random_seed("test_mp_circular_shift_left_{vname}_gpu", mpime)
+    !
+    DO i = 0, nproc-1
+      CALL RANDOM_NUMBER(rnd)
+      {vname}_h = {typeconv} ( 10.0 * rnd )
+      {vname}_d = {vname}_h
+      CALL mp_circular_shift_left({vname}_d, 0, world_comm)
+      CALL mp_circular_shift_left({vname}_h, 0, world_comm)
+      aux_h = {vname}_d
+      CALL test%assert_{compare}({sumf}({vname}_h) , {sumf}(aux_h) )
+    END DO
+    !
+    CALL collect_results(test)
+    !
+    CALL mp_world_end()
+    !
+    IF (mpime .eq. 0) CALL test%print()
+    !
+END PROGRAM test_mp_circular_shift_left_{vname}_gpu
+#else
+PROGRAM test_mp_circular_shift_left_{vname}_gpu
+    CALL no_test()
+END PROGRAM test_mp_circular_shift_left_{vname}_gpu
+#endif