libgomp/testsuite/libgomp.oacc-fortran/reduction-6.f90

   1 ! { dg-do run }
   2 ! { dg-additional-options "-cpp" }
   3
   4 ! { dg-additional-options "-Wopenacc-parallelism" } for testing/documenting
   5 ! aspects of that functionality.
   6
   7 program reduction
   8   implicit none
   9
  10   integer, parameter    :: n = 100, n2 = 1000, chunksize = 10
  11   integer               :: i, gs1, gs2, ws1, ws2, vs1, vs2, cs1, cs2, hs1, hs2
  12   integer               :: j, red, vred
  13
  14   gs1 = 0
  15   gs2 = 0
  16   ws1 = 0
  17   ws2 = 0
  18   vs1 = 0
  19   vs2 = 0
  20   cs1 = 0
  21   cs2 = 0
  22   hs1 = 0
  23   hs2 = 0
  24
  25   !$acc parallel num_gangs (1000)
  26   !$acc loop reduction(+:gs1, gs2) gang
  27   do i = 1, n
  28      gs1 = gs1 + 1
  29      gs2 = gs2 + 2
  30   end do
  31   !$acc end parallel
  32
  33   !$acc parallel num_workers (4) vector_length (32)
  34   ! { dg-warning "region is vector partitioned but does not contain vector partitioned code" "" { target *-*-* } .-1 }
  35   !$acc loop reduction(+:ws1, ws2) worker
  36   do i = 1, n
  37      ws1 = ws1 + 1
  38      ws2 = ws2 + 2
  39   end do
  40   !$acc end parallel
  41
  42   !$acc parallel vector_length (32)
  43   !$acc loop reduction(+:vs1, vs2) vector
  44   do i = 1, n
  45      vs1 = vs1 + 1
  46      vs2 = vs2 + 2
  47   end do
  48   !$acc end parallel
  49
  50   !$acc parallel num_gangs(8) num_workers(4) vector_length(32)
  51   !$acc loop reduction(+:cs1, cs2) gang worker vector
  52   do i = 1, n
  53      cs1 = cs1 + 1
  54      cs2 = cs2 + 2
  55   end do
  56   !$acc end parallel
  57
  58   ! Verify the results on the host
  59   do i = 1, n
  60      hs1 = hs1 + 1
  61      hs2 = hs2 + 2
  62   end do
  63
  64   if (gs1 .ne. hs1) STOP 1
  65   if (gs2 .ne. hs2) STOP 2
  66
  67   if (ws1 .ne. hs1) STOP 3
  68   if (ws2 .ne. hs2) STOP 4
  69
  70   if (vs1 .ne. hs1) STOP 5
  71   if (vs2 .ne. hs2) STOP 6
  72
  73   if (cs1 .ne. hs1) STOP 7
  74   if (cs2 .ne. hs2) STOP 8
  75
  76   ! Nested reductions.
  77
  78   red = 0
  79   vred = 0
  80
  81   !$acc parallel num_gangs(10) vector_length(32)
  82   !$acc loop reduction(+:red) gang
  83   do i = 1, n/chunksize
  84      !$acc loop reduction(+:red) vector
  85      do j = 1, chunksize
  86         red = red + chunksize
  87      end do
  88   end do
  89   !$acc end parallel
  90
  91   do i = 1, n/chunksize
  92      do j = 1, chunksize
  93         vred = vred + chunksize
  94      end do
  95   end do
  96
  97   if (red .ne. vred) STOP 9
  98 end program reduction