@@ -47,24 +51,143 @@ class="cmr-12">GPU capabilities of PSBLAS and AMG4PSBLAS.
-
+
-! build a one-level RAS with overlap 2 and ILU(0) on the local blocks.
-call P%init(’AS’,info)
-call P%set(’SUB_OVR’,2,info)
-call P%build(A,desc_A,info)
-... ...
-! solve Ax=b with preconditioned BiCGSTAB
- call psb_krylov(’BICGSTAB’,A,P,b,x,tol,desc_A,info)
+program amg_d_pde3d
+ use psb_base_mod
+ use amg_prec_mod
+ use psb_krylov_mod
+ use psb_util_mod
+ use psb_gpu_mod
+ use data_input
+ use amg_d_pde3d_base_mod
+ use amg_d_pde3d_exp_mod
+ use amg_d_pde3d_gauss_mod
+ use amg_d_genpde_mod
+ implicit none
+ .......
+ ! GPU variables
+ type(psb_d_hlg_sparse_mat) :: agmold
+ type(psb_d_vect_gpu) :: vgmold
+ type(psb_i_vect_gpu) :: igmold
+
-
+
Listing 5: setup of a one-level Schwarz preconditioner.
+class="content">setup of a GPU-enabled test program part one.
+ We then have to initialize the GPU environment, and pass the appropriate MOLD
+variables to the build methods
+
+
+
+
+
+
+
+
+
+
+ call psb_init(ctxt)
+ call psb_info(ctxt,iam,np)
+ !
+ ! BEWARE: if you have NGPUS per node, the default is to
+ ! attach to mod(IAM,NGPUS)
+ !
+ call psb_gpu_init(ictxt)
+ ......
+ t1 = psb_wtime()
+ call prec%smoothers_build(a,desc_a,info, amold=agmold, vmold=vgmold, imold=igmold)
+
+
+
+
Listing 6: setup of a GPU-enabled test program part two.
+
+
+
+
+ Finally, we convert the input matrix, the descriptor and the vectors, then
+preallocate the preconditioner workspace before entering the Krylov method. At the
+end of the code, we close the GPU environment
+
+
+
+
+
+
+
+
+
+
+ call desc_a%cnv(mold=igmold)
+ call a%cscnv(info,mold=agmold)
+ call psb_geasb(x,desc_a,info,mold=vgmold)
+ call psb_geasb(b,desc_a,info,mold=vgmold)
+ !
+ ! iterative method parameters
+ !
+ call psb_barrier(ctxt)
+ call prec%allocate_wrk(info)
+ t1 = psb_wtime()
+ call psb_krylov(s_choice%kmethd,a,prec,b,x,s_choice%eps,&
+ & desc_a,info,itmax=s_choice%itmax,iter=iter,err=err,itrace=s_choice%itrace,&
+ & istop=s_choice%istopc,irst=s_choice%irst)
+ call prec%deallocate_wrk(info)
+ call psb_barrier(ctxt)
+ tslv = psb_wtime() - t1
+ ......
+ call psb_gpu_exit()
+ call psb_exit(ctxt)
+ stop
+
+
+
+
Listing 7: setup of a GPU-enabled test program part three.
+
+
+
+
+ It is very important to employ solvers that are suited to the GPU, i.e. solvers that
+do NOT employ triangular system solve kernels. Solvers that satisfy this constraint
+include:
+
+ - JACOBI
+
+ - INVK
+
+ - INVT
+
+ - AINV
+and their L1 variants.
diff --git a/docs/src/gettingstarted.tex b/docs/src/gettingstarted.tex
index e4322ce0..7e5e0568 100644
--- a/docs/src/gettingstarted.tex
+++ b/docs/src/gettingstarted.tex
@@ -419,41 +419,182 @@ call P%build(A,desc_A,info)
\subsection{GPU example\label{sec:gpu-example}}
-The code reported in Figure~\ref{fig:gpu-ex1} shows how to set up a
+The code discussed here shows how to set up a
program exploiting the combined GPU capabilities of PSBLAS and
-AMG4PSBLAS.
+AMG4PSBLAS. The code example is availabile in the source distribution
+directory \verb|amg4psblas/tests/gpu|.
+
+First of all, we need to include the appropriate modules and
+declare some auxiliary variables:
\begin{listing}[h!]
\ifpdf
\begin{minted}[breaklines=true,bgcolor=bg,fontsize=\small]{fortran}
-! build a one-level RAS with overlap 2 and ILU(0) on the local blocks.
-call P%init('AS',info)
-call P%set('SUB_OVR',2,info)
-call P%build(A,desc_A,info)
-... ...
-! solve Ax=b with preconditioned BiCGSTAB
- call psb_krylov('BICGSTAB',A,P,b,x,tol,desc_A,info)
+program amg_d_pde3d
+ use psb_base_mod
+ use amg_prec_mod
+ use psb_krylov_mod
+ use psb_util_mod
+ use psb_gpu_mod
+ use data_input
+ use amg_d_pde3d_base_mod
+ use amg_d_pde3d_exp_mod
+ use amg_d_pde3d_gauss_mod
+ use amg_d_genpde_mod
+ implicit none
+ .......
+ ! GPU variables
+ type(psb_d_hlg_sparse_mat) :: agmold
+ type(psb_d_vect_gpu) :: vgmold
+ type(psb_i_vect_gpu) :: igmold
\end{minted}
\else
\begin{center}
\begin{minipage}{.90\textwidth}
{\small
\begin{verbatim}
-! build a one-level RAS with overlap 2 and ILU(0) on the local blocks.
-call P%init('AS',info)
-call P%set('SUB_OVR',2,info)
-call P%build(A,desc_A,info)
-... ...
-! solve Ax=b with preconditioned BiCGSTAB
- call psb_krylov('BICGSTAB',A,P,b,x,tol,desc_A,info)
-\end{verbatim}
+program amg_d_pde3d
+ use psb_base_mod
+ use amg_prec_mod
+ use psb_krylov_mod
+ use psb_util_mod
+ use psb_gpu_mod
+ use data_input
+ use amg_d_pde3d_base_mod
+ use amg_d_pde3d_exp_mod
+ use amg_d_pde3d_gauss_mod
+ use amg_d_genpde_mod
+ implicit none
+ .......
+ ! GPU variables
+ type(psb_d_hlg_sparse_mat) :: agmold
+ type(psb_d_vect_gpu) :: vgmold
+ type(psb_i_vect_gpu) :: igmold
+
+ \end{verbatim}
+}
+\end{minipage}
+\end{center}
+\fi
+\caption{setup of a GPU-enabled test program part one.\label{fig:gpu-ex1}}
+\end{listing}
+
+We then have to initialize the GPU environment, and pass the
+appropriate MOLD variables to the build methods
+\begin{listing}[h!]
+\ifpdf
+\begin{minted}[breaklines=true,bgcolor=bg,fontsize=\small]{fortran}
+ call psb_init(ctxt)
+ call psb_info(ctxt,iam,np)
+ !
+ ! BEWARE: if you have NGPUS per node, the default is to
+ ! attach to mod(IAM,NGPUS)
+ !
+ call psb_gpu_init(ictxt)
+ ......
+ t1 = psb_wtime()
+ call prec%smoothers_build(a,desc_a,info, amold=agmold, vmold=vgmold, imold=igmold)
+
+\end{minted}
+\else
+\begin{center}
+\begin{minipage}{.90\textwidth}
+{\small
+\begin{verbatim}
+ call psb_init(ctxt)
+ call psb_info(ctxt,iam,np)
+ !
+ ! BEWARE: if you have NGPUS per node, the default is to
+ ! attach to mod(IAM,NGPUS)
+ !
+ call psb_gpu_init(ictxt)
+ ......
+ t1 = psb_wtime()
+ call prec%smoothers_build(a,desc_a,info, amold=agmold, vmold=vgmold, imold=igmold)
+
+ \end{verbatim}
}
\end{minipage}
\end{center}
\fi
-\caption{setup of a one-level Schwarz preconditioner.\label{fig:gpu-ex1}}
+\caption{setup of a GPU-enabled test program part two.\label{fig:gpu-ex2}}
\end{listing}
+Finally, we convert the input matrix, the descriptor and the vectors,
+then preallocate the preconditioner workspace before entering the
+Krylov method. At the end of the code, we close the GPU environment
+\begin{listing}[h!]
+\ifpdf
+\begin{minted}[breaklines=true,bgcolor=bg,fontsize=\small]{fortran}
+ call desc_a%cnv(mold=igmold)
+ call a%cscnv(info,mold=agmold)
+ call psb_geasb(x,desc_a,info,mold=vgmold)
+ call psb_geasb(b,desc_a,info,mold=vgmold)
+
+ !
+ ! iterative method parameters
+ !
+ call psb_barrier(ctxt)
+ call prec%allocate_wrk(info)
+ t1 = psb_wtime()
+ call psb_krylov(s_choice%kmethd,a,prec,b,x,s_choice%eps,&
+ & desc_a,info,itmax=s_choice%itmax,iter=iter,err=err,itrace=s_choice%itrace,&
+ & istop=s_choice%istopc,irst=s_choice%irst)
+ call prec%deallocate_wrk(info)
+ call psb_barrier(ctxt)
+ tslv = psb_wtime() - t1
+ ......
+ call psb_gpu_exit()
+ call psb_exit(ctxt)
+ stop
+
+
+\end{minted}
+\else
+\begin{center}
+\begin{minipage}{.90\textwidth}
+{\small
+\begin{verbatim}
+ call desc_a%cnv(mold=igmold)
+ call a%cscnv(info,mold=agmold)
+ call psb_geasb(x,desc_a,info,mold=vgmold)
+ call psb_geasb(b,desc_a,info,mold=vgmold)
+
+ !
+ ! iterative method parameters
+ !
+ call psb_barrier(ctxt)
+ call prec%allocate_wrk(info)
+ t1 = psb_wtime()
+ call psb_krylov(s_choice%kmethd,a,prec,b,x,s_choice%eps,&
+ & desc_a,info,itmax=s_choice%itmax,iter=iter,err=err,itrace=s_choice%itrace,&
+ & istop=s_choice%istopc,irst=s_choice%irst)
+ call prec%deallocate_wrk(info)
+ call psb_barrier(ctxt)
+ tslv = psb_wtime() - t1
+
+ ......
+ call psb_gpu_exit()
+ call psb_exit(ctxt)
+ stop
+
+ \end{verbatim}
+}
+\end{minipage}
+\end{center}
+\fi
+\caption{setup of a GPU-enabled test program part three.\label{fig:gpu-ex3}}
+\end{listing}
+It is very important to employ solvers that are suited
+to the GPU, i.e. solvers that do NOT employ triangular
+system solve kernels. Solvers that satisfy this constraint include:
+\begin{itemize}
+\item \verb|JACOBI|
+\item \verb|INVK|
+\item \verb|INVT|
+\item \verb|AINV|
+\end{itemize}
+and their \verb|L1| variants.
%%% Local Variables:
%%% mode: latex