Use less device memory when checkpointing particles (#3238)

This chunks the operation per box rather than per rank so as to use less
device memory.

The proposed changes:
- [ ] fix a bug or incorrect behavior in AMReX
- [ ] add new capabilities to AMReX
- [ ] changes answers in the test suite to more than roundoff level
- [ ] are likely to significantly affect the results of downstream AMReX
users
- [ ] include documentation in the code and/or rst files, if appropriate

---------

Co-authored-by: Axel Huebl <axel.huebl@plasma.ninja>
Co-authored-by: Ann Almgren <asalmgren@lbl.gov>

Src/Particle/AMReX_WriteBinaryParticleData.H

@@ -189,16 +189,12 @@ packIOData (Vector<int>& idata, Vector<ParticleReal>& rdata, const PC& pc, int l
     const Long rChunkSize = AMREX_SPACEDIM + num_output_real;
     rdata.resize(np*rChunkSize);
 
-    typename PC::IntVector  idata_d(idata.size());
-    typename PC::RealVector rdata_d(rdata.size());
-
     typename PC::IntVector write_int_comp_d(write_int_comp.size());
     typename PC::IntVector write_real_comp_d(write_real_comp.size());
     Gpu::copyAsync(Gpu::hostToDevice, write_int_comp.begin(), write_int_comp.end(),
                    write_int_comp_d.begin());
     Gpu::copyAsync(Gpu::hostToDevice, write_real_comp.begin(), write_real_comp.end(),
                    write_real_comp_d.begin());
-    Gpu::Device::streamSynchronize();
 
     const auto write_int_comp_d_ptr = write_int_comp_d.data();
     const auto write_real_comp_d_ptr = write_real_comp_d.data();
@@ -211,6 +207,9 @@ packIOData (Vector<int>& idata, Vector<ParticleReal>& rdata, const PC& pc, int l
         typename PC::IntVector offsets(np_tile);
         int num_copies = Scan::ExclusiveSum(np_tile, pflags.begin(), offsets.begin(), Scan::retSum);
 
+        typename PC::IntVector  idata_d(num_copies*iChunkSize);
+        typename PC::RealVector rdata_d(num_copies*rChunkSize);
+
         const auto flag_ptr = pflags.data();
 
         auto idata_d_ptr = idata_d.data();
@@ -224,11 +223,11 @@ packIOData (Vector<int>& idata, Vector<ParticleReal>& rdata, const PC& pc, int l
             const auto p = ptd.getSuperParticle(pindex);
 
             if (flag_ptr[pindex]) {
-                std::size_t iout_index = (pindex+poffset)*iChunkSize;
+                std::size_t iout_index = pindex*iChunkSize;
                 packParticleIDs(&idata_d_ptr[iout_index], p, is_checkpoint);
                 iout_index += 2;
 
-                std::size_t rout_index = (pindex+poffset)*rChunkSize;
+                std::size_t rout_index = pindex*rChunkSize;
                 for (int j = 0; j < AMREX_SPACEDIM; j++) {
                   rdata_d_ptr[rout_index] = p.pos(j);
                   rout_index++;
@@ -267,12 +266,14 @@ packIOData (Vector<int>& idata, Vector<ParticleReal>& rdata, const PC& pc, int l
             }
         });
 
+        Gpu::copyAsync(Gpu::deviceToHost, idata_d.begin(), idata_d.end(),
+                       idata.begin() + typename PC::IntVector::difference_type(poffset));
+        Gpu::copyAsync(Gpu::deviceToHost, rdata_d.begin(), rdata_d.end(),
+                       rdata.begin() + typename PC::RealVector::difference_type(poffset));
+        Gpu::Device::streamSynchronize();
+
         poffset += num_copies;
     }
-
-    Gpu::copyAsync(Gpu::deviceToHost, idata_d.begin(), idata_d.end(), idata.begin());
-    Gpu::copyAsync(Gpu::deviceToHost, rdata_d.begin(), rdata_d.end(), rdata.begin());
-    Gpu::Device::streamSynchronize();
 }
 
 template <class PC>