DepositCurrent: atomicAdd -> lockAdd

The new amrex::BaseFab::lockAdd function is an optimized version of atomicAdd for OpenMP. In my testing on Frontier CPUs, it's up to 10x faster.
Hi-PACE · Jan 16, 2024 · 1156f57 · 1156f57
1 parent c41482d
commit 1156f57
Showing 1 changed file with 6 additions and 6 deletions.
diff --git a/src/particles/deposition/PlasmaDepositCurrent.cpp b/src/particles/deposition/PlasmaDepositCurrent.cpp
@@ -301,20 +301,20 @@ DepositCurrent (PlasmaParticleContainer& plasma, Fields & fields, const MultiLas
                 amrex::Box srcbx = dstbx;
                 srcbx -= amrex::IntVect(a_itilex_bs, a_itiley_bs, srcbx.smallEnd(2));
                 if (jx_cmp != -1) {
-                    isl_fab.atomicAdd(tmp_dens[ithread], srcbx, dstbx, 0, jx_cmp, 1);
-                    isl_fab.atomicAdd(tmp_dens[ithread], srcbx, dstbx, 1, jy_cmp, 1);
+                    isl_fab.lockAdd(tmp_dens[ithread], srcbx, dstbx, 0, jx_cmp, 1);
+                    isl_fab.lockAdd(tmp_dens[ithread], srcbx, dstbx, 1, jy_cmp, 1);
                 }
                 if (jz_cmp != -1) {
-                    isl_fab.atomicAdd(tmp_dens[ithread], srcbx, dstbx, 2, jz_cmp, 1);
+                    isl_fab.lockAdd(tmp_dens[ithread], srcbx, dstbx, 2, jz_cmp, 1);
                 }
                 if (rho_cmp != -1) {
-                    isl_fab.atomicAdd(tmp_dens[ithread], srcbx, dstbx, 3, rho_cmp, 1);
+                    isl_fab.lockAdd(tmp_dens[ithread], srcbx, dstbx, 3, rho_cmp, 1);
                 }
                 if (chi_cmp != -1) {
-                    isl_fab.atomicAdd(tmp_dens[ithread], srcbx, dstbx, 4, chi_cmp, 1);
+                    isl_fab.lockAdd(tmp_dens[ithread], srcbx, dstbx, 4, chi_cmp, 1);
                 }
                 if (rhomjz_cmp != -1) {
-                    isl_fab.atomicAdd(tmp_dens[ithread], srcbx, dstbx, 5, rhomjz_cmp, 1);
+                    isl_fab.lockAdd(tmp_dens[ithread], srcbx, dstbx, 5, rhomjz_cmp, 1);
                 }
             }
 #endif