|
|
|
|
@ -29,17 +29,21 @@ class="cmr-12">4.2 </span></span> <a
|
|
|
|
|
id="x16-150004.2"></a><span
|
|
|
|
|
class="cmr-12">GPU example</span></h4>
|
|
|
|
|
<!--l. 422--><p class="noindent" ><span
|
|
|
|
|
class="cmr-12">The code reported in Figure</span><span
|
|
|
|
|
class="cmr-12"> </span><a
|
|
|
|
|
href="#x16-15001r5"><span
|
|
|
|
|
class="cmr-12">5</span><!--tex4ht:ref: fig:gpu-ex1 --></a> <span
|
|
|
|
|
class="cmr-12">shows how to set up a program exploiting the combined</span>
|
|
|
|
|
class="cmr-12">The code discussed here shows how to set up a program exploiting the combined GPU</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">GPU capabilities of PSBLAS and AMG4PSBLAS.</span>
|
|
|
|
|
class="cmr-12">capabilities of PSBLAS and AMG4PSBLAS. The code example is availabile in the</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">source distribution directory </span><span class="obeylines-h"><span class="verb"><span
|
|
|
|
|
class="cmtt-12">amg4psblas/tests/gpu</span></span></span><span
|
|
|
|
|
class="cmr-12">.</span>
|
|
|
|
|
<!--l. 427--><p class="indent" > <span
|
|
|
|
|
class="cmr-12">First of all, we need to include the appropriate modules and declare some auxiliary</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">variables:</span>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<!--l. 425--><p class="indent" > <a
|
|
|
|
|
<!--l. 429--><p class="indent" > <a
|
|
|
|
|
id="x16-15001r5"></a><hr class="float"><div class="float"
|
|
|
|
|
>
|
|
|
|
|
|
|
|
|
|
@ -47,24 +51,143 @@ class="cmr-12">GPU capabilities of PSBLAS and AMG4PSBLAS.</span>
|
|
|
|
|
|
|
|
|
|
<div class="center"
|
|
|
|
|
>
|
|
|
|
|
<!--l. 437--><p class="noindent" >
|
|
|
|
|
<!--l. 451--><p class="noindent" >
|
|
|
|
|
<div class="minipage"><pre class="verbatim" id="verbatim-10">
|
|
|
|
|
! 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
|
|
|
|
|
 
|
|
|
|
|
</pre>
|
|
|
|
|
<!--l. 448--><p class="nopar" ></div></div>
|
|
|
|
|
<!--l. 473--><p class="nopar" ></div></div>
|
|
|
|
|
<br /> <div class="caption"
|
|
|
|
|
><span class="id">Listing 5: </span><span
|
|
|
|
|
class="content">setup of a one-level Schwarz preconditioner.</span></div><!--tex4ht:label?: x16-15001r5 -->
|
|
|
|
|
class="content">setup of a GPU-enabled test program part one.</span></div><!--tex4ht:label?: x16-15001r5 -->
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
</div><hr class="endfloat" />
|
|
|
|
|
<!--l. 481--><p class="indent" > <span
|
|
|
|
|
class="cmr-12">We then have to initialize the GPU environment, and pass the appropriate MOLD</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">variables to the build methods</span>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<!--l. 483--><p class="indent" > <a
|
|
|
|
|
id="x16-15002r6"></a><hr class="float"><div class="float"
|
|
|
|
|
>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<div class="center"
|
|
|
|
|
>
|
|
|
|
|
<!--l. 499--><p class="noindent" >
|
|
|
|
|
<div class="minipage"><pre class="verbatim" id="verbatim-11">
|
|
|
|
|
  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)
|
|
|
|
|
 
|
|
|
|
|
</pre>
|
|
|
|
|
<!--l. 514--><p class="nopar" ></div></div>
|
|
|
|
|
<br /> <div class="caption"
|
|
|
|
|
><span class="id">Listing 6: </span><span
|
|
|
|
|
class="content">setup of a GPU-enabled test program part two.</span></div><!--tex4ht:label?: x16-15002r6 -->
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
</div><hr class="endfloat" />
|
|
|
|
|
<!--l. 521--><p class="indent" > <span
|
|
|
|
|
class="cmr-12">Finally, we convert the input matrix, the descriptor and the vectors, then</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">preallocate the preconditioner workspace before entering the Krylov method. At the</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">end of the code, we close the GPU environment</span>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<!--l. 524--><p class="indent" > <a
|
|
|
|
|
id="x16-15003r7"></a><hr class="float"><div class="float"
|
|
|
|
|
>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<div class="center"
|
|
|
|
|
>
|
|
|
|
|
<!--l. 553--><p class="noindent" >
|
|
|
|
|
<div class="minipage"><pre class="verbatim" id="verbatim-12">
|
|
|
|
|
  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
|
|
|
|
|
 
|
|
|
|
|
</pre>
|
|
|
|
|
<!--l. 580--><p class="nopar" ></div></div>
|
|
|
|
|
<br /> <div class="caption"
|
|
|
|
|
><span class="id">Listing 7: </span><span
|
|
|
|
|
class="content">setup of a GPU-enabled test program part three.</span></div><!--tex4ht:label?: x16-15003r7 -->
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
</div><hr class="endfloat" />
|
|
|
|
|
<!--l. 588--><p class="indent" > <span
|
|
|
|
|
class="cmr-12">It is very important to employ solvers that are suited to the GPU, i.e. solvers that</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">do NOT employ triangular system solve kernels. Solvers that satisfy this constraint</span>
|
|
|
|
|
<span
|
|
|
|
|
class="cmr-12">include:</span>
|
|
|
|
|
<ul class="itemize1">
|
|
|
|
|
<li class="itemize"><span class="obeylines-h"><span class="verb"><span
|
|
|
|
|
class="cmtt-12">JACOBI</span></span></span>
|
|
|
|
|
</li>
|
|
|
|
|
<li class="itemize"><span class="obeylines-h"><span class="verb"><span
|
|
|
|
|
class="cmtt-12">INVK</span></span></span>
|
|
|
|
|
</li>
|
|
|
|
|
<li class="itemize"><span class="obeylines-h"><span class="verb"><span
|
|
|
|
|
class="cmtt-12">INVT</span></span></span>
|
|
|
|
|
</li>
|
|
|
|
|
<li class="itemize"><span class="obeylines-h"><span class="verb"><span
|
|
|
|
|
class="cmtt-12">AINV</span></span></span></li></ul>
|
|
|
|
|
<!--l. 597--><p class="noindent" ><span
|
|
|
|
|
class="cmr-12">and their </span><span class="obeylines-h"><span class="verb"><span
|
|
|
|
|
class="cmtt-12">L1</span></span></span> <span
|
|
|
|
|
class="cmr-12">variants.</span>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Salvatore Filippone
4 years ago