RRTMGP memory reduction

Source/PhysicsInterfaces/Radiation/ERF_RRTMGP_Interface.H

@@ -120,10 +120,6 @@ rrtmgp_main (const int ncol, const int nlay,
              real1d_k& t_sfc      , real1d_k& sfc_emis   , real1d_k& lw_src,
              real2d_k& lwp        , real2d_k& iwp        ,
              real2d_k& rel        , real2d_k& rei        , real2d_k& cldfrac,
-             real3d_k& aer_tau_sw , real3d_k& aer_ssa_sw , real3d_k& aer_asm_sw,
-             real3d_k& aer_tau_lw ,
-             real3d_k& cld_tau_sw_bnd, real3d_k& cld_tau_lw_bnd,
-             real3d_k& cld_tau_sw_gpt, real3d_k& cld_tau_lw_gpt,
              real2d_k& sw_flux_up    , real2d_k& sw_flux_dn    , real2d_k& sw_flux_dn_dir,
              real2d_k& lw_flux_up    , real2d_k& lw_flux_dn    ,
              real2d_k& sw_clnclrsky_flux_up, real2d_k& sw_clnclrsky_flux_dn, real2d_k& sw_clnclrsky_flux_dn_dir,

Source/PhysicsInterfaces/Radiation/ERF_RRTMGP_Interface.cpp

@@ -389,10 +389,6 @@ rrtmgp_main (const int ncol, const int nlay,
              real1d_k& t_sfc      , real1d_k& sfc_emis   , real1d_k& lw_src,
              real2d_k& lwp, real2d_k& iwp,
              real2d_k& rel, real2d_k& rei, real2d_k& cldfrac,
-             real3d_k& /*aer_tau_sw*/, real3d_k& /*aer_ssa_sw*/, real3d_k& /*aer_asm_sw*/,
-             real3d_k& /*aer_tau_lw*/,
-             real3d_k& /*cld_tau_sw_bnd*/, real3d_k& /*cld_tau_lw_bnd*/,
-             real3d_k& /*cld_tau_sw_gpt*/, real3d_k& /*cld_tau_lw_gpt*/,
              real2d_k& sw_flux_up, real2d_k& sw_flux_dn, real2d_k& sw_flux_dn_dir,
              real2d_k& lw_flux_up, real2d_k& lw_flux_dn,
              real2d_k& sw_clnclrsky_flux_up, real2d_k& sw_clnclrsky_flux_dn, real2d_k& sw_clnclrsky_flux_dn_dir,
@@ -457,24 +453,8 @@ rrtmgp_main (const int ncol, const int nlay,
     optical_props1_t aerosol_lw;
     aerosol_sw.init(k_dist_sw_k->get_band_lims_wavenumber());
     aerosol_sw.alloc_2str(ncol, nlay);
-    /*
-    Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<3>>({0, 0, 0}, {ncol, nlay, nswbands}),
-                         KOKKOS_LAMBDA (int icol, int ilay, int ibnd)
-    {
-        aerosol_sw.tau(icol,ilay,ibnd) = aer_tau_sw(icol,ilay,ibnd);
-        aerosol_sw.ssa(icol,ilay,ibnd) = aer_ssa_sw(icol,ilay,ibnd);
-        aerosol_sw.g  (icol,ilay,ibnd) = aer_asm_sw(icol,ilay,ibnd);
-    });
-    */
     aerosol_lw.init(k_dist_lw_k->get_band_lims_wavenumber());
     aerosol_lw.alloc_1scl(ncol, nlay);
-    /*
-    Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<3>>({0, 0, 0}, {ncol, nlay, nlwbands}),
-                         KOKKOS_LAMBDA (int icol, int ilay, int ibnd)
-    {
-        aerosol_lw.tau(icol,ilay,ibnd) = aer_tau_lw(icol,ilay,ibnd);
-    });
-    */
 
     // Convert cloud physical properties to optical properties for input to RRTMGP
     optical_props2_t clouds_sw = get_cloud_optics_sw(ncol, nlay,
@@ -483,9 +463,6 @@ rrtmgp_main (const int ncol, const int nlay,
     optical_props1_t clouds_lw = get_cloud_optics_lw(ncol, nlay,
                                                      *cloud_optics_lw_k, *k_dist_lw_k,
                                                      lwp, iwp, rel, rei);
-    //Kokkos::deep_copy(cld_tau_sw_bnd, clouds_sw.tau);
-    //Kokkos::deep_copy(cld_tau_lw_bnd, clouds_lw.tau);
-
     // Do subcolumn sampling to map bands -> gpoints based on cloud fraction and overlap assumption;
     // This implements the Monte Carlo Independent Column Approximation by mapping only a single
     // subcolumn (cloud state) to each gpoint.
@@ -497,21 +474,6 @@ rrtmgp_main (const int ncol, const int nlay,
     auto clouds_lw_gpt = get_subsampled_clouds(ncol, nlay, nlwbands, nlwgpts,
                                                clouds_lw, *k_dist_lw_k, cldfrac, p_lay);
 
-    /*
-    // Copy cloud properties to outputs (is this needed, or can we just use pointers?)
-    // Alternatively, just compute and output a subcolumn cloud mask
-    Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<3>>({0, 0, 0}, {ncol, nlay, nswgpts}),
-                         KOKKOS_LAMBDA (int icol, int ilay, int igpt)
-    {
-        cld_tau_sw_gpt(icol,ilay,igpt) = clouds_sw_gpt.tau(icol,ilay,igpt);
-    });
-    Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<3>>({0, 0, 0}, {ncol, nlay, nlwgpts}),
-                         KOKKOS_LAMBDA (int icol, int ilay, int igpt)
-    {
-        cld_tau_lw_gpt(icol,ilay,igpt) = clouds_lw_gpt.tau(icol,ilay,igpt);
-    });
-    */
-
   // Do shortwave
   rrtmgp_sw(ncol, nlay,
             *k_dist_sw_k, p_lay, t_lay, p_lev, t_lev, gas_concs,
@@ -651,16 +613,28 @@ rrtmgp_sw (const int ncol,
         flux_up    (icol,ilev) = 0;
         flux_dn    (icol,ilev) = 0;
         flux_dn_dir(icol,ilev) = 0;
-        clnclrsky_flux_up    (icol,ilev) = 0;
-        clnclrsky_flux_dn    (icol,ilev) = 0;
-        clnclrsky_flux_dn_dir(icol,ilev) = 0;
         clrsky_flux_up    (icol,ilev) = 0;
         clrsky_flux_dn    (icol,ilev) = 0;
         clrsky_flux_dn_dir(icol,ilev) = 0;
-        clnsky_flux_up    (icol,ilev) = 0;
-        clnsky_flux_dn    (icol,ilev) = 0;
-        clnsky_flux_dn_dir(icol,ilev) = 0;
     });
+    if (extra_clnclrsky_diag) {
+        Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<2>>({0, 0}, {ncol, nlay+1}),
+                             KOKKOS_LAMBDA (int icol, int ilev)
+        {
+            clnclrsky_flux_up    (icol,ilev) = 0;
+            clnclrsky_flux_dn    (icol,ilev) = 0;
+            clnclrsky_flux_dn_dir(icol,ilev) = 0;
+        });
+    }
+    if (extra_clnsky_diag) {
+        Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<2>>({0, 0}, {ncol, nlay+1}),
+                             KOKKOS_LAMBDA (int icol, int ilev)
+        {
+            clnsky_flux_up    (icol,ilev) = 0;
+            clnsky_flux_dn    (icol,ilev) = 0;
+            clnsky_flux_dn_dir(icol,ilev) = 0;
+        });
+    }
     Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<3>>({0, 0, 0}, {ncol, nlay+1, nbnd}),
                          KOKKOS_LAMBDA (int icol, int ilev, int ibnd)
     {
@@ -933,15 +907,27 @@ rrtmgp_lw (const int ncol,
     Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<2>>({0, 0}, {ncol, nlay+1}),
                          KOKKOS_LAMBDA (int icol, int ilev)
     {
-        flux_up(icol, ilev)           = 0;
-        flux_dn(icol, ilev)           = 0;
-        clnclrsky_flux_up(icol, ilev) = 0;
-        clnclrsky_flux_dn(icol, ilev) = 0;
-        clrsky_flux_up(icol, ilev)    = 0;
-        clrsky_flux_dn(icol, ilev)    = 0;
-        clnsky_flux_up(icol, ilev)    = 0;
-        clnsky_flux_dn(icol, ilev)    = 0;
+        flux_up(icol, ilev)        = 0;
+        flux_dn(icol, ilev)        = 0;
+        clrsky_flux_up(icol, ilev) = 0;
+        clrsky_flux_dn(icol, ilev) = 0;
     });
+    if (extra_clnclrsky_diag) {
+        Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<2>>({0, 0}, {ncol, nlay+1}),
+                             KOKKOS_LAMBDA (int icol, int ilev)
+        {
+            clnclrsky_flux_up(icol, ilev) = 0;
+            clnclrsky_flux_dn(icol, ilev) = 0;
+        });
+    }
+    if (extra_clnsky_diag) {
+        Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<2>>({0, 0}, {ncol, nlay+1}),
+                             KOKKOS_LAMBDA (int icol, int ilev)
+        {
+            clnsky_flux_up(icol, ilev) = 0;
+            clnsky_flux_dn(icol, ilev) = 0;
+        });
+    }
     Kokkos::parallel_for(Kokkos::MDRangePolicy<Kokkos::Rank<3>>({0, 0, 0}, {ncol, nlay+1, nbnd}),
                          KOKKOS_LAMBDA (int icol, int ilev, int ibnd)
     {

Source/PhysicsInterfaces/Radiation/ERF_Radiation.H

@@ -54,6 +54,10 @@ public:
     // Destructor
     ~Radiation ()
     {
+        // Release k-distribution data and memory pool
+        if (rrtmgp::initialized) {
+            rrtmgp::rrtmgp_finalize();
+        }
         // Note that Kokkos is now finalized in main.cpp
     }
 
@@ -322,8 +326,11 @@ private:
     // Offsets for MultiFab <-> KOKKOS transfer
     amrex::Vector<int> m_col_offsets;
 
-    // Whether we use aerosol forcing in radiation
-    bool m_do_aerosol_rad = true;
+    // Whether we use aerosol forcing in radiation.
+    // Aerosol plumbing is currently not implemented; this hook is retained so a
+    // future aerosol scheme can wire in without reintroducing the parameter.
+    // Setting it true today triggers an abort in the Radiation constructor.
+    bool m_do_aerosol_rad = false;
 
     // Whether we do extra aerosol forcing calls
     bool m_extra_clnsky_diag    = false;
@@ -363,6 +370,9 @@ private:
     // Rad frequency in number of steps
     int m_rad_freq_in_steps = 1;
 
+    // Number of columns to process per RRTMGP chunk (controls peak memory)
+    int m_ncol_chunk = 5000;
+
     // Whether or not to do subcolumn sampling of cloud state for MCICA
     bool m_do_subcol_sampling = true;
 
@@ -442,21 +452,16 @@ private:
     real2d_k sfc_alb_dir;
     real2d_k sfc_alb_dif;
 
-    // 2d size (ncol, nlwbands)
-    real2d_k emis_sfc;
-
+    // Aerosol optical properties. Dormant scaffolding kept so a future aerosol
+    // coupling (e.g. SPA, prescribed aerosol climatology) can populate these
+    // without re-adding members. Only allocated when m_do_aerosol_rad is true,
+    // which currently aborts in the constructor; when aerosol coupling is
+    // wired in, remove the abort and pass these into rrtmgp_main.
+    //
     // 3d size (ncol, nlay, n[sw,lw]bands)
     real3d_k aero_tau_sw;
     real3d_k aero_ssa_sw;
     real3d_k aero_g_sw;
     real3d_k aero_tau_lw;
-
-    // 3d size (ncol, nlay, n[sw,lw]bnds)
-    real3d_k cld_tau_sw_bnd;
-    real3d_k cld_tau_lw_bnd;
-
-    // 3d size (ncol, nlay, n[sw,lw]gpts)
-    real3d_k cld_tau_sw_gpt;
-    real3d_k cld_tau_lw_gpt;
 };
 #endif