Fix a number of tests for GPU (#4323)

Author: WeiqunZhang

Tests/MultiBlock/IndexType/main.cpp

@@ -6,18 +6,11 @@
 int MyMain();
 
 int main(int argc, char** argv) {
-#ifdef AMREX_USE_MPI
-    MPI_Init(&argc, &argv);
-#else
-    amrex::ignore_unused(argc,argv);
-#endif
+    amrex::Initialize(argc, argv);
     // Let me throw exceptions for triggering my debugger
-    amrex::Initialize(MPI_COMM_WORLD, std::cout, std::cerr, [](const char* msg) { throw std::runtime_error(msg); });
+    amrex::SetErrorHandler([](const char* msg) { throw std::runtime_error(msg); });
     int ret = MyMain();
     amrex::Finalize();
-#ifdef AMREX_USE_MPI
-    MPI_Finalize();
-#endif
     return ret;
 }
 
@@ -51,19 +44,23 @@ bool ParallelCopyWithItselfIsCorrect(amrex::iMultiFab& mf, const amrex::Box& dom
     const int nx = domain.length(0);
     const int ny = domain.length(1);
     int fails = 0;
+    amrex::ReduceOps<amrex::ReduceOpSum> reduce_op;
+    amrex::ReduceData<int> reduce_data(reduce_op);
+    using ReduceTuple = amrex::GpuTuple<int>;
     for (amrex::MFIter mfi(mf); mfi.isValid(); ++mfi) {
         const amrex::Box section = dest_box & mfi.tilebox();
         if (section.isEmpty()) { continue; }
         auto array = mf.const_array(mfi);
-        amrex::LoopOnCpu(section, [&](int i, int j, int k)
+        reduce_op.eval(section, reduce_data,
+                       [=] AMREX_GPU_DEVICE (int i, int j, int k) -> ReduceTuple
         {
             amrex::Dim3 si = dtos(amrex::Dim3{i,j,k});
             int value = si.x + si.y*nx + si.z*nx*ny;
-            fails += (array(i,j,k) != value);
-
-            AMREX_ASSERT(fails==0);  // If DEBUG, crash on first error.
+            auto r = int(array(i,j,k) != value);
+            return { r };
         });
     }
+    fails += amrex::get<0>(reduce_data.value());
     return fails == 0;
 }
 
@@ -113,19 +110,23 @@ bool ParallelCopyFaceToFace(amrex::iMultiFab& dest, const amrex::Box& domain_des
     int fails = 0;
     const int nx = domain_src.length(0);
     const int ny = domain_src.length(1);
+    amrex::ReduceOps<amrex::ReduceOpSum> reduce_op;
+    amrex::ReduceData<int> reduce_data(reduce_op);
+    using ReduceTuple = amrex::GpuTuple<int>;
     for (amrex::MFIter mfi(dest); mfi.isValid(); ++mfi) {
         const amrex::Box section = dest_box & mfi.tilebox();
         if (section.isEmpty()) { continue; }
         auto darray = dest.const_array(mfi);
-        amrex::LoopOnCpu(section, [&](int i, int j, int k)
+        reduce_op.eval(section, reduce_data,
+                       [=] AMREX_GPU_DEVICE (int i, int j, int k) -> ReduceTuple
         {
             amrex::Dim3 si = dtos(amrex::Dim3{i,j,k});
             int value = si.x + si.y*nx + si.z*nx*ny;
-            fails += (darray(i,j,k) != value);
-
-            AMREX_ASSERT(fails==0); // If in debug, crash on first error.
+            auto r = int(darray(i,j,k) != value);
+            return { r };
         });
     }
+    fails += amrex::get<0>(reduce_data.value());
     return fails == 0;
 }
 

Tests/Particles/DenseBins/inputs

@@ -1,5 +1,6 @@
+amrex.the_arena_is_managed = 1
 nitems = 100000017
 nbins = 5000
 
 #nitems = 20
-#nbins = 4
+#nbins = 4

Tests/Particles/DenseBins/main.cpp

@@ -109,9 +109,7 @@ void testDenseBins ()
     amrex::Vector<int> items(nitems);
     initData(nbins, items);
 
-#ifndef AMREX_USE_SYCL
     testSerial(nbins, items);
-#endif
 #ifdef AMREX_USE_OMP
     testOpenMP(nbins, items);
 #endif

Tests/Particles/NamedSoAComponents/main.cpp

@@ -131,6 +131,10 @@ void addParticles ()
 
 int main (int argc, char* argv[])
  {
+    {
+        amrex::ParmParse pp("amrex");
+        pp.add("the_arena_is_managed", 1);
+    }
     amrex::Initialize(argc,argv);
     {
         addParticles();

Tests/Particles/SOAParticle/main.cpp

@@ -180,6 +180,10 @@ void addParticles ()
 
 int main(int argc, char* argv[])
  {
+    {
+        amrex::ParmParse pp("amrex");
+        pp.add("the_arena_is_managed", 1);
+    }
     amrex::Initialize(argc,argv);
     {
         addParticles< ParticleContainerPureSoA<4, 2> > ();