From 864872ecacff43d40eed244a1044c0eb293db117 Mon Sep 17 00:00:00 2001
From: Salvatore Filippone <salvatore.filippone@uniroma2.it>
Date: Sat, 17 Feb 2024 17:28:32 +0100
Subject: [PATCH] Intermediate implementation of abgdxyz on cuda

---
 cuda/psb_d_cuda_vect_mod.F90 | 39 ++++++++++++++++++++++++++++++------
 cuda/spgpu/vector.h          | 15 +++++++++++++-
 2 files changed, 47 insertions(+), 7 deletions(-)

diff --git a/cuda/psb_d_cuda_vect_mod.F90 b/cuda/psb_d_cuda_vect_mod.F90
index 03e65f91..fe5d3a38 100644
--- a/cuda/psb_d_cuda_vect_mod.F90
+++ b/cuda/psb_d_cuda_vect_mod.F90
@@ -923,13 +923,40 @@ contains
     real(psb_dpk_), intent (in)       :: alpha, beta, gamma, delta
     integer(psb_ipk_), intent(out)              :: info
 
-    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)
+
+    info = psb_success_
     
+    if (.false.) then
+      
+      select type(xx => x)
+      type is (psb_d_vect_cuda)
+        ! Do something different here 
+        if ((beta /= dzero).and.y%is_host())&
+             &  call y%sync()
+        if (xx%is_host()) call xx%sync()
+        nx = getMultiVecDeviceSize(xx%deviceVect)
+        ny = getMultiVecDeviceSize(y%deviceVect)
+        if ((nx<m).or.(ny<m)) then
+          info = psb_err_internal_error_
+        else
+          info = axpbyMultiVecDevice(m,alpha,xx%deviceVect,beta,y%deviceVect)
+        end if
+        call y%set_dev()
+      class default
+        ! Do it on the host side
+        if ((alpha /= dzero).and.(x%is_dev()))&
+             & call x%sync()
+        call y%axpby(m,alpha,x%v,beta,info)
+      end select
+      
+    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/spgpu/vector.h b/cuda/spgpu/vector.h
index 179206af..69ffedf0 100644
--- a/cuda/spgpu/vector.h
+++ b/cuda/spgpu/vector.h
@@ -463,7 +463,20 @@ void spgpuDaxpby(spgpuHandle_t handle,
 	double alpha,
 	__device double* x);
 	
-/** 
+
+void spgpuDabgdxyz(spgpuHandle_t handle,
+		   int n,
+		   double alpha,
+		   double beta,
+		   double gamma,
+		   double delta,
+		   __device double* x,
+		   __device double *y,
+		   __device double *z)
+;
+	
+
+  /** 
 * \fn void spgpuDmaxpby(spgpuHandle_t handle, __device double *z, int n, double beta, __device double *y, double alpha, __device double* x, int count, int pitch)
  * Computes the double precision z = beta * y + alpha * x of x and y multivectors. z could be exactly x or y (without offset) or another vector.
  * \param handle the spgpu handle used to call this routine