Fix memory traffic in GTH/SCT

openacc/psb_c_oacc_vect_mod.F90

@@ -425,7 +425,6 @@ contains
     class is (psb_i_vect_oacc)
       if (ii%is_host()) call ii%sync()
       if (y%is_host()) call y%sync()
-      !$acc update device(y%combuf)
       call inner_sctb(n,y%combuf(i:i+n-1),beta,y%v,ii%v(i:i+n-1))
       call y%set_dev()
       acc_done = .true.
@@ -444,6 +443,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       complex(psb_spk_)    :: beta,x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async 
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -488,6 +488,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       complex(psb_spk_) :: beta, x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -561,7 +562,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async
     end subroutine inner_gth
   end subroutine c_oacc_gthzbuf
   
@@ -605,7 +606,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async     
     end subroutine inner_gth
   end subroutine c_oacc_gthzv_x
 

openacc/psb_d_oacc_vect_mod.F90

@@ -425,7 +425,6 @@ contains
     class is (psb_i_vect_oacc)
       if (ii%is_host()) call ii%sync()
       if (y%is_host()) call y%sync()
-      !$acc update device(y%combuf)
       call inner_sctb(n,y%combuf(i:i+n-1),beta,y%v,ii%v(i:i+n-1))
       call y%set_dev()
       acc_done = .true.
@@ -444,6 +443,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       real(psb_dpk_)    :: beta,x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async 
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -488,6 +488,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       real(psb_dpk_) :: beta, x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -561,7 +562,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async
     end subroutine inner_gth
   end subroutine d_oacc_gthzbuf
   
@@ -605,7 +606,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async     
     end subroutine inner_gth
   end subroutine d_oacc_gthzv_x
 

openacc/psb_i_oacc_vect_mod.F90

@@ -79,7 +79,6 @@ contains
     class is (psb_i_vect_oacc)
       if (ii%is_host()) call ii%sync()
       if (y%is_host()) call y%sync()
-      !$acc update device(y%combuf)
       call inner_sctb(n,y%combuf(i:i+n-1),beta,y%v,ii%v(i:i+n-1))
       call y%set_dev()
       acc_done = .true.
@@ -98,6 +97,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       integer(psb_ipk_)    :: beta,x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async 
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -142,6 +142,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       integer(psb_ipk_) :: beta, x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -215,7 +216,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async
     end subroutine inner_gth
   end subroutine i_oacc_gthzbuf
   
@@ -259,7 +260,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async     
     end subroutine inner_gth
   end subroutine i_oacc_gthzv_x
 

openacc/psb_l_oacc_vect_mod.F90

@@ -81,7 +81,6 @@ contains
     class is (psb_i_vect_oacc)
       if (ii%is_host()) call ii%sync()
       if (y%is_host()) call y%sync()
-      !$acc update device(y%combuf)
       call inner_sctb(n,y%combuf(i:i+n-1),beta,y%v,ii%v(i:i+n-1))
       call y%set_dev()
       acc_done = .true.
@@ -100,6 +99,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       integer(psb_lpk_)    :: beta,x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async 
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -144,6 +144,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       integer(psb_lpk_) :: beta, x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -217,7 +218,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async
     end subroutine inner_gth
   end subroutine l_oacc_gthzbuf
   
@@ -261,7 +262,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async     
     end subroutine inner_gth
   end subroutine l_oacc_gthzv_x
 

openacc/psb_s_oacc_vect_mod.F90

@@ -425,7 +425,6 @@ contains
     class is (psb_i_vect_oacc)
       if (ii%is_host()) call ii%sync()
       if (y%is_host()) call y%sync()
-      !$acc update device(y%combuf)
       call inner_sctb(n,y%combuf(i:i+n-1),beta,y%v,ii%v(i:i+n-1))
       call y%set_dev()
       acc_done = .true.
@@ -444,6 +443,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       real(psb_spk_)    :: beta,x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async 
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -488,6 +488,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       real(psb_spk_) :: beta, x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -561,7 +562,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async
     end subroutine inner_gth
   end subroutine s_oacc_gthzbuf
   
@@ -605,7 +606,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async     
     end subroutine inner_gth
   end subroutine s_oacc_gthzv_x
 

openacc/psb_z_oacc_vect_mod.F90

@@ -425,7 +425,6 @@ contains
     class is (psb_i_vect_oacc)
       if (ii%is_host()) call ii%sync()
       if (y%is_host()) call y%sync()
-      !$acc update device(y%combuf)
       call inner_sctb(n,y%combuf(i:i+n-1),beta,y%v,ii%v(i:i+n-1))
       call y%set_dev()
       acc_done = .true.
@@ -444,6 +443,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       complex(psb_dpk_)    :: beta,x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async 
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -488,6 +488,7 @@ contains
       integer(psb_ipk_) :: n, idx(:)
       complex(psb_dpk_) :: beta, x(:), y(:)
       integer(psb_ipk_) :: k
+      !$acc update device(x(1:n)) async
       !$acc parallel loop 
       do k = 1, n
         y(idx(k)) = x(k) + beta *y(idx(k))
@@ -561,7 +562,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async
     end subroutine inner_gth
   end subroutine z_oacc_gthzbuf
   
@@ -605,7 +606,7 @@ contains
         y(k) = x(idx(k))
       end do
       !$acc end parallel loop
-      !$acc update self(y)      
+      !$acc update self(y(1:n)) async     
     end subroutine inner_gth
   end subroutine z_oacc_gthzv_x