From 93c71c43162fb6663cca9c2f0fff0e8f2ff4c47e Mon Sep 17 00:00:00 2001
From: sfilippone <filippone.salvatore@gmail.com>
Date: Tue, 20 Feb 2024 10:25:31 +0100
Subject: [PATCH] Fix %ZERO() on cuda

---
 cuda/psb_c_cuda_vect_mod.F90 |  6 ++--
 cuda/psb_d_cuda_vect_mod.F90 | 60 ++++++------------------------------
 cuda/psb_d_vectordev_mod.F90 | 13 --------
 cuda/psb_i_cuda_vect_mod.F90 |  6 ++--
 cuda/psb_s_cuda_vect_mod.F90 |  6 ++--
 cuda/psb_z_cuda_vect_mod.F90 |  6 ++--
 6 files changed, 21 insertions(+), 76 deletions(-)

diff --git a/cuda/psb_c_cuda_vect_mod.F90 b/cuda/psb_c_cuda_vect_mod.F90
index 56cc80e6..fca1c616 100644
--- a/cuda/psb_c_cuda_vect_mod.F90
+++ b/cuda/psb_c_cuda_vect_mod.F90
@@ -668,9 +668,9 @@ contains
     use psi_serial_mod
     implicit none 
     class(psb_c_vect_cuda), intent(inout) :: x
-    
-    if (allocated(x%v)) x%v=czero
-    call x%set_host()
+
+    call x%set_scal(czero)
+
   end subroutine c_cuda_zero
 
   subroutine c_cuda_asb_m(n, x, info)
diff --git a/cuda/psb_d_cuda_vect_mod.F90 b/cuda/psb_d_cuda_vect_mod.F90
index f2ef2be3..2220b26c 100644
--- a/cuda/psb_d_cuda_vect_mod.F90
+++ b/cuda/psb_d_cuda_vect_mod.F90
@@ -668,9 +668,9 @@ contains
     use psi_serial_mod
     implicit none 
     class(psb_d_vect_cuda), intent(inout) :: x
-    
-    if (allocated(x%v)) x%v=dzero
-    call x%set_host()
+
+    call x%set_scal(dzero)
+
   end subroutine d_cuda_zero
 
   subroutine d_cuda_asb_m(n, x, info)
@@ -922,56 +922,14 @@ contains
     class(psb_d_vect_cuda), intent(inout)  :: z
     real(psb_dpk_), intent (in)       :: alpha, beta, gamma, delta
     integer(psb_ipk_), intent(out)              :: info
-    integer(psb_ipk_) :: nx, ny, nz
-    logical :: gpu_done
 
-    info = psb_success_
+    call z%psb_d_base_vect_type%abgdxyz(m,alpha,beta,gamma,delta,x,y,info)
+!!$
+!!$    if (x%is_dev()) call x%sync()
+!!$
+!!$    call y%axpby(m,alpha,x,beta,info)
+!!$    call z%axpby(m,gamma,y,delta,info)
     
-    if (.true.) then
-      gpu_done = .false.
-      select type(xx => x)
-      class is (psb_d_vect_cuda)
-        select type(yy => y)
-        class is (psb_d_vect_cuda)
-          select type(zz => z)
-          class is (psb_d_vect_cuda)
-            ! Do something different here 
-            if ((beta /= dzero).and.yy%is_host())&
-                 &  call yy%sync()
-            if ((delta /= dzero).and.zz%is_host())&
-                 &  call zz%sync()
-            if (xx%is_host()) call xx%sync()
-            nx = getMultiVecDeviceSize(xx%deviceVect)
-            ny = getMultiVecDeviceSize(yy%deviceVect)
-            nz = getMultiVecDeviceSize(zz%deviceVect)
-            if ((nx<m).or.(ny<m).or.(nz<m)) then
-              info = psb_err_internal_error_
-            else
-              info = abgdxyzMultiVecDevice(m,alpha,beta,gamma,delta,&
-                   & xx%deviceVect,yy%deviceVect,zz%deviceVect)
-            end if
-            call yy%set_dev()
-            call zz%set_dev()
-            gpu_done = .true.
-          end select
-        end select
-      end select
-      
-      if (.not.gpu_done) then 
-        if (x%is_host()) call x%sync()
-        if (y%is_host()) call y%sync()
-        if (z%is_host()) call z%sync()
-        call y%axpby(m,alpha,x,beta,info)
-        call z%axpby(m,gamma,y,delta,info)
-      end if
-    else
-
-      if (x%is_host()) call x%sync()
-      if (y%is_host()) call y%sync()
-      if (z%is_host()) call z%sync()
-      call y%axpby(m,alpha,x,beta,info)
-      call z%axpby(m,gamma,y,delta,info)
-    end if
   end subroutine d_cuda_abgdxyz
 
   
diff --git a/cuda/psb_d_vectordev_mod.F90 b/cuda/psb_d_vectordev_mod.F90
index c378d8c2..802add96 100644
--- a/cuda/psb_d_vectordev_mod.F90
+++ b/cuda/psb_d_vectordev_mod.F90
@@ -316,19 +316,6 @@ module psb_d_vectordev_mod
     end function axpbyMultiVecDeviceDouble
   end interface
 
-
-  interface abgdxyzMultiVecDevice
-    function abgdxyzMultiVecDeviceDouble(n,alpha,beta,gamma,delta,deviceVecX,&
-         & deviceVecY,deviceVecZ) &
-         & result(res) bind(c,name='abgdxyzMultiVecDeviceDouble')
-      use iso_c_binding
-      integer(c_int)      :: res
-      integer(c_int), value :: n
-      real(c_double), value :: alpha, beta,gamma,delta
-      type(c_ptr), value  :: deviceVecX, deviceVecY, deviceVecZ
-    end function abgdxyzMultiVecDeviceDouble
-  end interface abgdxyzMultiVecDevice
-
   interface axyMultiVecDevice
     function axyMultiVecDeviceDouble(n,alpha,deviceVecA,deviceVecB) &
          & result(res) bind(c,name='axyMultiVecDeviceDouble')
diff --git a/cuda/psb_i_cuda_vect_mod.F90 b/cuda/psb_i_cuda_vect_mod.F90
index df8f4113..684aa324 100644
--- a/cuda/psb_i_cuda_vect_mod.F90
+++ b/cuda/psb_i_cuda_vect_mod.F90
@@ -650,9 +650,9 @@ contains
     use psi_serial_mod
     implicit none 
     class(psb_i_vect_cuda), intent(inout) :: x
-    
-    if (allocated(x%v)) x%v=izero
-    call x%set_host()
+
+    call x%set_scal(izero)
+
   end subroutine i_cuda_zero
 
   subroutine i_cuda_asb_m(n, x, info)
diff --git a/cuda/psb_s_cuda_vect_mod.F90 b/cuda/psb_s_cuda_vect_mod.F90
index 9616b3a6..80c60bc3 100644
--- a/cuda/psb_s_cuda_vect_mod.F90
+++ b/cuda/psb_s_cuda_vect_mod.F90
@@ -668,9 +668,9 @@ contains
     use psi_serial_mod
     implicit none 
     class(psb_s_vect_cuda), intent(inout) :: x
-    
-    if (allocated(x%v)) x%v=szero
-    call x%set_host()
+
+    call x%set_scal(szero)
+
   end subroutine s_cuda_zero
 
   subroutine s_cuda_asb_m(n, x, info)
diff --git a/cuda/psb_z_cuda_vect_mod.F90 b/cuda/psb_z_cuda_vect_mod.F90
index 1153fc61..9f801742 100644
--- a/cuda/psb_z_cuda_vect_mod.F90
+++ b/cuda/psb_z_cuda_vect_mod.F90
@@ -668,9 +668,9 @@ contains
     use psi_serial_mod
     implicit none 
     class(psb_z_vect_cuda), intent(inout) :: x
-    
-    if (allocated(x%v)) x%v=zzero
-    call x%set_host()
+
+    call x%set_scal(zzero)
+
   end subroutine z_cuda_zero
 
   subroutine z_cuda_asb_m(n, x, info)