deepmodeling · mohanchen · Jan 29, 2026 · Jan 28, 2026 · Jan 28, 2026 · Jan 28, 2026
diff --git a/source/source_base/parallel_reduce.cpp b/source/source_base/parallel_reduce.cpp
@@ -233,85 +233,65 @@ void Parallel_Reduce::gather_int_all(int& v, int* all)
     return;
 }
 
-void Parallel_Reduce::gather_min_int_all(const int& nproc, int& v)
+template <>
+void Parallel_Reduce::reduce_min<int>(int& v)
 {
 #ifdef __MPI
-    std::vector<int> all(nproc, 0);
-    MPI_Allgather(&v, 1, MPI_INT, all.data(), 1, MPI_INT, MPI_COMM_WORLD);
-    for (int i = 0; i < nproc; i++)
-    {
-        if (v > all[i])
-        {
-            v = all[i];
-        }
-    }
+    MPI_Allreduce(MPI_IN_PLACE, &v, 1, MPI_INT, MPI_MIN, MPI_COMM_WORLD);
 #endif
 }
 
-void Parallel_Reduce::gather_max_double_all(const int& nproc, double& v)
+template <>
+void Parallel_Reduce::reduce_min<float>(float& v)
 {
 #ifdef __MPI
-    std::vector<double> value(nproc, 0.0);
-    MPI_Allgather(&v, 1, MPI_DOUBLE, value.data(), 1, MPI_DOUBLE, MPI_COMM_WORLD);
-    for (int i = 0; i < nproc; i++)
-    {
-        if (v < value[i])
-        {
-            v = value[i];
-        }
-    }
+    MPI_Allreduce(MPI_IN_PLACE, &v, 1, MPI_FLOAT, MPI_MIN, MPI_COMM_WORLD);
 #endif
 }
 
-void Parallel_Reduce::gather_max_double_pool(const int& nproc_in_pool, double& v)
+template <>
+void Parallel_Reduce::reduce_min<double>(double& v)
 {
 #ifdef __MPI
-    if (nproc_in_pool == 1) 
-    {
-        return;
-    }
-    std::vector<double> value(nproc_in_pool, 0.0);
-    MPI_Allgather(&v, 1, MPI_DOUBLE, value.data(), 1, MPI_DOUBLE, POOL_WORLD);
-    for (int i = 0; i < nproc_in_pool; i++)
-    {
-        if (v < value[i])
-        {
-            v = value[i];
-        }
-    }
+    MPI_Allreduce(MPI_IN_PLACE, &v, 1, MPI_DOUBLE, MPI_MIN, MPI_COMM_WORLD);
+#endif
+}
+
+template <>
+void Parallel_Reduce::reduce_max<float>(float& v)
+{
+#ifdef __MPI
+    MPI_Allreduce(MPI_IN_PLACE, &v, 1, MPI_FLOAT, MPI_MAX, MPI_COMM_WORLD);
+#endif
+}
+
+template <>
+void Parallel_Reduce::reduce_max<double>(double& v)
+{
+#ifdef __MPI
+    MPI_Allreduce(MPI_IN_PLACE, &v, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD);
 #endif
 }
 
-void Parallel_Reduce::gather_min_double_pool(const int& nproc_in_pool, double& v)
+template <>
+void Parallel_Reduce::reduce_max_pool<double>(const int& nproc_in_pool, double& v)
 {
 #ifdef __MPI
     if (nproc_in_pool == 1) 
     {
         return;
     }
-    std::vector<double> value(nproc_in_pool, 0.0);
-    MPI_Allgather(&v, 1, MPI_DOUBLE, value.data(), 1, MPI_DOUBLE, POOL_WORLD);
-    for (int i = 0; i < nproc_in_pool; i++)
-    {
-        if (v > value[i])
-        {
-            v = value[i];
-        }
-    }
+    MPI_Allreduce(MPI_IN_PLACE, &v, 1, MPI_DOUBLE, MPI_MAX, POOL_WORLD);
 #endif
 }
-
-void Parallel_Reduce::gather_min_double_all(const int& nproc, double& v)
+template <>
+void Parallel_Reduce::reduce_min_pool<double>(const int& nproc_in_pool, double& v)
 {
 #ifdef __MPI
-    std::vector<double> value(nproc, 0.0);
-    MPI_Allgather(&v, 1, MPI_DOUBLE, value.data(), 1, MPI_DOUBLE, MPI_COMM_WORLD);
-    for (int i = 0; i < nproc; i++)
+    if (nproc_in_pool == 1) 
     {
-        if (v > value[i])
-        {
-            v = value[i];
-        }
+        return;
     }
+    MPI_Allreduce(MPI_IN_PLACE, &v, 1, MPI_DOUBLE, MPI_MIN, POOL_WORLD);
 #endif
 }
diff --git a/source/source_base/parallel_reduce.h b/source/source_base/parallel_reduce.h
@@ -21,6 +21,14 @@ template <typename T>
 void reduce_pool(T& object);
 template <typename T>
 void reduce_pool(T* object, const int n);
+template <typename T>
+void reduce_min(T& v);
+template <typename T>
+void reduce_max(T& v);
+template <typename T>
+void reduce_min_pool(const int& nproc_in_pool, T& v);
+template <typename T>
+void reduce_max_pool(const int& nproc_in_pool, T& v);
 
 void reduce_int_diag(int& object); // mohan add 2012-01-12
 
@@ -34,13 +42,6 @@ void reduce_double_diag(double* object, const int n);
 void reduce_double_allpool(const int& npool, const int& nproc_in_pool, double& object);
 void reduce_double_allpool(const int& npool, const int& nproc_in_pool, double* object, const int n);
 
-void gather_min_int_all(const int& nproc, int& v);
-void gather_max_double_all(const int& nproc, double& v);
-void gather_min_double_all(const int& nproc, double& v);
-void gather_max_double_pool(const int& nproc_in_pool, double& v);
-void gather_min_double_pool(const int& nproc_in_pool, double& v);
-
-// mohan add 2011-04-21
 void gather_int_all(int& v, int* all);
 
 bool check_if_equal(double& v); // mohan add 2009-11-11

diff --git a/source/source_base/test_parallel/parallel_reduce_test.cpp b/source/source_base/test_parallel/parallel_reduce_test.cpp
@@ -30,9 +30,9 @@
  *   3. ReduceComplexAll:
  *       Tests two variations of reduce_complex_all()
  *   4. GatherIntAll:
- *       Tests gather_int_all() and gather_min_int_all()
+ *       Tests gather_int_all() and reduce_min()
  *   5. GatherDoubleAll:
- *       Tests gather_min_double_all() and gather_max_double_all()
+ *       Tests reduce_min_double() and reduce_max_double()
  *   6. ReduceIntDiag:
  *       Tests reduce_int_diag()
  *   7. ReduceDoubleDiag:
@@ -47,7 +47,7 @@
  *   11.ReduceComplexPool:
  *       Tests two variations of reduce_pool()
  *   12.GatherDoublePool:
- *       Tests gather_min_double_pool() and gather_max_double_pool()
+ *       Tests reduce_min_pool() and reduce_max_pool()
  *
  *
  */
@@ -233,7 +233,7 @@ TEST_F(ParaReduce, GatherIntAll)
     EXPECT_EQ(local_number, array[my_rank]);
     // get minimum integer among all processes
     int min_number = local_number;
-    Parallel_Reduce::gather_min_int_all(nproc, min_number);
+    Parallel_Reduce::reduce_min(min_number);
     for (int i = 0; i < nproc; i++)
     {
         EXPECT_LE(min_number, array[i]);
@@ -256,10 +256,10 @@ TEST_F(ParaReduce, GatherDoubleAll)
     EXPECT_EQ(local_number, array[my_rank]);
     // get minimum integer among all processes
     double min_number = local_number;
-    Parallel_Reduce::gather_min_double_all(nproc, min_number);
+    Parallel_Reduce::reduce_min(min_number);
     // get maximum integer among all processes
     double max_number = local_number;
-    Parallel_Reduce::gather_max_double_all(nproc, max_number);
+    Parallel_Reduce::reduce_max(max_number);
     for (int i = 0; i < nproc; i++)
     {
         EXPECT_LE(min_number, array[i]);
@@ -587,10 +587,10 @@ TEST_F(ParaReduce, GatherDoublePool)
         EXPECT_EQ(local_number, array[mpiContext.rank_in_pool]);
         // get minimum integer among all processes
         double min_number = local_number;
-        Parallel_Reduce::gather_min_double_pool(mpiContext.nproc_in_pool, min_number);
+        Parallel_Reduce::reduce_min_pool(mpiContext.nproc_in_pool, min_number);
         // get maximum integer among all processes
         double max_number = local_number;
-        Parallel_Reduce::gather_max_double_pool(mpiContext.nproc_in_pool, max_number);
+        Parallel_Reduce::reduce_max_pool(mpiContext.nproc_in_pool, max_number);
         for (int i = 0; i < mpiContext.nproc_in_pool; i++)
         {
             EXPECT_LE(min_number, array[i]);

diff --git a/source/source_basis/module_pw/pw_basis_big.h b/source/source_basis/module_pw/pw_basis_big.h
@@ -2,6 +2,7 @@
 #define PW_BASIS_BIG_H
 #include "source_base/constants.h"
 #include "source_base/global_function.h"
+
 #ifdef __MPI
 #include "mpi.h"
 #endif

diff --git a/source/source_cell/k_vector_utils.cpp b/source/source_cell/k_vector_utils.cpp
@@ -245,7 +245,7 @@ void kvec_mpi_k(K_Vectors& kv)
     ModuleBase::GlobalFunc::OUT(GlobalV::ofs_running, "Number of k-points in this process", kv.nks);
     int nks_minimum = kv.nks;
 
-    Parallel_Reduce::gather_min_int_all(GlobalV::NPROC, nks_minimum);
+    Parallel_Reduce::reduce_min(nks_minimum);
 
     if (nks_minimum == 0)
     {

diff --git a/source/source_estate/elecstate_energy.cpp b/source/source_estate/elecstate_energy.cpp
@@ -47,11 +47,10 @@ void ElecState::cal_bandgap()
     { 
         vbm =this->eferm.ef;
     }
-#ifdef __MPI
-    Parallel_Reduce::gather_max_double_all(GlobalV::NPROC, vbm);
-    Parallel_Reduce::gather_min_double_all(GlobalV::NPROC, cbm);
-#endif
-
+    #ifdef __MPI
+    Parallel_Reduce::reduce_max(vbm);
+    Parallel_Reduce::reduce_min(cbm);
+    #endif
     this->bandgap = cbm - vbm;
 }
 
@@ -119,14 +118,12 @@ void ElecState::cal_bandgap_updw()
     { 
         vbm_dw =this->eferm.ef_dw;
     }
-
-#ifdef __MPI
-    Parallel_Reduce::gather_max_double_all(GlobalV::NPROC, vbm_up);
-    Parallel_Reduce::gather_min_double_all(GlobalV::NPROC, cbm_up);
-    Parallel_Reduce::gather_max_double_all(GlobalV::NPROC, vbm_dw);
-    Parallel_Reduce::gather_min_double_all(GlobalV::NPROC, cbm_dw);
-#endif
-
+    #ifdef __MPI
+    Parallel_Reduce::reduce_max(vbm_up);
+    Parallel_Reduce::reduce_min(cbm_up);
+    Parallel_Reduce::reduce_max(vbm_dw);
+    Parallel_Reduce::reduce_min(cbm_dw);
+    #endif
     this->bandgap_up = cbm_up - vbm_up;
     this->bandgap_dw = cbm_dw - vbm_dw;
 }

diff --git a/source/source_estate/occupy.cpp b/source/source_estate/occupy.cpp
@@ -179,10 +179,9 @@ void Occupy::iweights(
             }
         }
     }
-#ifdef __MPI
-    Parallel_Reduce::gather_max_double_all(GlobalV::NPROC, ef);
-#endif
-
+    #ifdef __MPI
+    Parallel_Reduce::reduce_max(ef);
+    #endif
     return;
 }
 
@@ -306,13 +305,11 @@ void Occupy::efermig(const ModuleBase::matrix& ekb,
 
     eup += 2 * smearing_sigma;
     elw -= 2 * smearing_sigma;
-
-#ifdef __MPI
     // find min and max across pools
-    Parallel_Reduce::gather_max_double_all(GlobalV::NPROC, eup);
-    Parallel_Reduce::gather_min_double_all(GlobalV::NPROC, elw);
-
-#endif
+    #ifdef __MPI
+    Parallel_Reduce::reduce_max(eup);
+    Parallel_Reduce::reduce_min(elw);
+    #endif
     //=================
     // Bisection method
     //=================

diff --git a/source/source_io/cal_dos.cpp b/source/source_io/cal_dos.cpp
@@ -56,8 +56,8 @@ void ModuleIO::prepare_dos(std::ofstream& ofs_running,
     }
 
 #ifdef __MPI
-    Parallel_Reduce::gather_max_double_all(GlobalV::NPROC, emax);
-    Parallel_Reduce::gather_min_double_all(GlobalV::NPROC, emin);
+    Parallel_Reduce::reduce_max(emax);
+    Parallel_Reduce::reduce_min(emin);
 #endif
 
     emax *= ModuleBase::Ry_to_eV;

diff --git a/source/source_lcao/module_operator_lcao/op_exx_lcao.hpp b/source/source_lcao/module_operator_lcao/op_exx_lcao.hpp
@@ -3,6 +3,7 @@
 #ifdef __EXX
 
 #include "op_exx_lcao.h"
+#include "source_base/parallel_reduce.h"
 #include "source_io/module_parameter/parameter.h"
 #include "source_lcao/module_ri/RI_2D_Comm.h"
 #include "source_hamilt/module_xc/xc_functional.h"
@@ -244,10 +245,9 @@ OperatorEXX<OperatorLCAO<TK, TR>>::OperatorEXX(HS_Matrix_K<TK>* hsk_in,
                     if (!ifs) { all_exist = 0; break; }
                 }
 // Add MPI communication to synchronize all_exist across processes
-#ifdef __MPI
-                // don't read in any files if one of the processes doesn't have it
-                MPI_Allreduce(MPI_IN_PLACE, &all_exist, 1, MPI_INT, MPI_MIN, MPI_COMM_WORLD);
-#endif                
+                #ifdef __MPI
+                Parallel_Reduce::reduce_min(all_exist);
+                #endif
                 if (all_exist)
                 {
                     // Read HexxR in CSR format
@@ -264,9 +264,9 @@ OperatorEXX<OperatorLCAO<TK, TR>>::OperatorEXX(HS_Matrix_K<TK>* hsk_in,
                     const std::string restart_HR_path_cereal = GlobalC::restart.folder + "HexxR_" + std::to_string(PARAM.globalv.myrank);
                     std::ifstream ifs(restart_HR_path_cereal, std::ios::binary);
                     int all_exist_cereal = ifs ? 1 : 0;
-#ifdef __MPI                    
-                    MPI_Allreduce(MPI_IN_PLACE, &all_exist_cereal, 1, MPI_INT, MPI_MIN, MPI_COMM_WORLD);
-#endif                     
+                    #ifdef __MPI
+                    Parallel_Reduce::reduce_min(all_exist_cereal);
+                    #endif
                     if (!all_exist_cereal)
                     {
                         //no HexxR file in CSR or binary format

diff --git a/source/source_pw/module_pwdft/setup_pwwfc.cpp b/source/source_pw/module_pwdft/setup_pwwfc.cpp
@@ -1,5 +1,6 @@
 #include "source_pw/module_pwdft/setup_pwwfc.h" // pw_wfc
 #include "source_base/parallel_comm.h" // POOL_WORLD
+#include "source_base/parallel_reduce.h" // Parallel_Reduce
 #include "source_io/print_info.h" // print information
 
 void pw::teardown_pwwfc(ModulePW::PW_Basis_K* &pw_wfc)
@@ -52,14 +53,12 @@ void pw::setup_pwwfc(const Input_para& inp,
 			pw_rho.nz);
 
     pw_wfc->initparameters(false, inp.ecutwfc, kv.get_nks(), kv.kvec_d.data());
-
 #ifdef __MPI
     if (inp.pw_seed > 0)
     {
-        MPI_Allreduce(MPI_IN_PLACE, &pw_wfc->ggecut, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD);
+        Parallel_Reduce::reduce_max( pw_wfc->ggecut);
     }
-    // qianrui add 2021-8-13 to make different kpar parameters can get the same
-    // results
+    // qianrui add 2021-8-13 to make different kpar parameters can get the same result
 #endif
 
     pw_wfc->fft_bundle.initfftmode(inp.fft_mode);

diff --git a/source/source_pw/module_stodft/sto_iter.cpp b/source/source_pw/module_stodft/sto_iter.cpp
@@ -203,8 +203,8 @@ void Stochastic_Iter<T, Device>::checkemm(const int& ik,
     if (ik == nks - 1)
     {
 #ifdef __MPI
-        MPI_Allreduce(MPI_IN_PLACE, p_hamilt_sto->emax, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD);
-        MPI_Allreduce(MPI_IN_PLACE, p_hamilt_sto->emin, 1, MPI_DOUBLE, MPI_MIN, MPI_COMM_WORLD);
+        Parallel_Reduce::reduce_max(*p_hamilt_sto->emax);
+        Parallel_Reduce::reduce_min(*p_hamilt_sto->emin);
         MPI_Allreduce(MPI_IN_PLACE, &change, 1, MPI_CHAR, MPI_LOR, MPI_COMM_WORLD);
 #endif
         if (change)

diff --git a/source/source_pw/module_stodft/sto_tool.cpp b/source/source_pw/module_stodft/sto_tool.cpp
@@ -2,6 +2,7 @@
 
 #include "source_base/math_chebyshev.h"
 #include "source_base/parallel_device.h"
+#include "source_base/parallel_reduce.h"
 #include "source_base/timer.h"
 #include "source_io/module_parameter/parameter.h"
 #ifdef __MPI
@@ -103,8 +104,8 @@ void check_che_op<FPTYPE, Device>::operator()(const int& nche_in,
         if (ik == nk - 1)
         {
 #ifdef __MPI
-            MPI_Allreduce(MPI_IN_PLACE, p_hamilt_sto->emax, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD);
-            MPI_Allreduce(MPI_IN_PLACE, p_hamilt_sto->emin, 1, MPI_DOUBLE, MPI_MIN, MPI_COMM_WORLD);
+            Parallel_Reduce::reduce_max(*p_hamilt_sto->emax);
+            Parallel_Reduce::reduce_min(*p_hamilt_sto->emin);
 #endif
             GlobalV::ofs_running << "New Emax " << *p_hamilt_sto->emax << " Ry; new Emin " << *p_hamilt_sto->emin
                                  << " Ry" << std::endl;