[Refactor] Remove unnecessary manual MPI_Reduce in hsolver dav code (#7314)

reduce_pool template already has single precision (std::complex<float>) instantiation with MPI_IN_PLACE, so manually creating swap buffers and calling MPI_Reduce for complex types is unnecessary.

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

Author: Cstandardlib

source/source_hsolver/diago_dav_subspace.cpp

@@ -19,6 +19,7 @@
 
 #ifdef __MPI
 #include <mpi.h>
+#include "source_base/parallel_comm.h"
 #endif
 
 using namespace hsolver;
@@ -583,62 +584,9 @@ void Diago_DavSubspace<T, Device>::cal_elem(const int& dim,
         // Only on dsp hardware need an extra space to reduce data
         mtfunc::dsp_dav_subspace_reduce(hcc, scc, nbase, this->nbase_x, this->notconv, this->diag_comm.comm);
 #else
-        auto* swap = new T[notconv * this->nbase_x];
-
-        syncmem_complex_op()(swap, hcc + nbase * this->nbase_x, notconv * this->nbase_x);
-
-        if (std::is_same<T, double>::value)
-        {
-            Parallel_Reduce::reduce_pool(hcc + nbase * this->nbase_x, notconv * this->nbase_x);
-            Parallel_Reduce::reduce_pool(scc + nbase * this->nbase_x, notconv * this->nbase_x);
-        }
-        else
-        {
-            if (base_device::get_current_precision(swap) == "single")
-            {
-                MPI_Reduce(swap,
-                           hcc + nbase * this->nbase_x,
-                           notconv * this->nbase_x,
-                           MPI_COMPLEX,
-                           MPI_SUM,
-                           0,
-                           this->diag_comm.comm);
-            }
-            else
-            {
-                MPI_Reduce(swap,
-                           hcc + nbase * this->nbase_x,
-                           notconv * this->nbase_x,
-                           MPI_DOUBLE_COMPLEX,
-                           MPI_SUM,
-                           0,
-                           this->diag_comm.comm);
-            }
-
-            syncmem_complex_op()(swap, scc + nbase * this->nbase_x, notconv * this->nbase_x);
-
-            if (base_device::get_current_precision(swap) == "single")
-            {
-                MPI_Reduce(swap,
-                           scc + nbase * this->nbase_x,
-                           notconv * this->nbase_x,
-                           MPI_COMPLEX,
-                           MPI_SUM,
-                           0,
-                           this->diag_comm.comm);
-            }
-            else
-            {
-                MPI_Reduce(swap,
-                           scc + nbase * this->nbase_x,
-                           notconv * this->nbase_x,
-                           MPI_DOUBLE_COMPLEX,
-                           MPI_SUM,
-                           0,
-                           this->diag_comm.comm);
-            }
-        }
-        delete[] swap;
+        assert(this->diag_comm.comm == POOL_WORLD);
+        Parallel_Reduce::reduce_pool(hcc + nbase * this->nbase_x, notconv * this->nbase_x);
+        Parallel_Reduce::reduce_pool(scc + nbase * this->nbase_x, notconv * this->nbase_x);
 #endif
     }
 #endif

source/source_hsolver/diago_david.cpp

@@ -6,6 +6,7 @@
 
 #include "source_hsolver/kernels/hegvd_op.h"
 #include "source_base/kernels/math_kernel_op.h"
+#include "source_base/parallel_comm.h"
 
 
 using namespace hsolver;
@@ -613,25 +614,8 @@ void DiagoDavid<T, Device>::cal_elem(const int& dim,
     {
         ModuleBase::matrixTranspose_op<T, Device>()(nbase_x, nbase_x, hcc, hcc);
 
-        auto* swap = new T[notconv * nbase_x];
-        syncmem_complex_op()(swap, hcc + nbase * nbase_x, notconv * nbase_x);
-        if (std::is_same<T, double>::value)
-        {
-            Parallel_Reduce::reduce_pool(hcc + nbase * nbase_x, notconv * nbase_x);
-        }
-        else
-        {
-            if (base_device::get_current_precision(swap) == "single") {
-                MPI_Reduce(swap, hcc + nbase * nbase_x, notconv * nbase_x, MPI_COMPLEX, MPI_SUM, 0, diag_comm.comm);
-            }
-            else {
-                MPI_Reduce(swap, hcc + nbase * nbase_x, notconv * nbase_x, MPI_DOUBLE_COMPLEX, MPI_SUM, 0, diag_comm.comm);
-            }
-
-        }
-        delete[] swap;
-
-        // Parallel_Reduce::reduce_complex_double_pool( hcc + nbase * nbase_x, notconv * nbase_x );
+        assert(diag_comm.comm == POOL_WORLD);
+        Parallel_Reduce::reduce_pool(hcc + nbase * nbase_x, notconv * nbase_x);
 
         ModuleBase::matrixTranspose_op<T, Device>()(nbase_x, nbase_x, hcc, hcc);
     }

source/source_hsolver/test/diago_david_float_test.cpp

@@ -1,5 +1,6 @@
 #include"source_hsolver/diago_david.h"
 #include"source_hsolver/diago_iter_assist.h"
+#include "source_base/parallel_comm.h"
 #include"source_pw/module_pwdft/hamilt_pw.h"
 #include"diago_mock.h"
 #include "source_psi/psi.h"
@@ -85,7 +86,7 @@ class DiagoDavPrepare
 		phm = new hamilt::HamiltPW<std::complex<float>>(nullptr, nullptr, nullptr, nullptr, nullptr, nullptr);
 
 #ifdef __MPI 
-        const hsolver::diag_comm_info comm_info = {MPI_COMM_WORLD, mypnum, nprocs};
+        const hsolver::diag_comm_info comm_info = {POOL_WORLD, mypnum, nprocs};
 #else
        	const hsolver::diag_comm_info comm_info = {mypnum, nprocs};
 #endif

source/source_hsolver/test/diago_david_real_test.cpp

@@ -1,5 +1,6 @@
 #include"source_hsolver/diago_david.h"
 #include"source_hsolver/diago_iter_assist.h"
+#include "source_base/parallel_comm.h"
 #include"source_pw/module_pwdft/hamilt_pw.h"
 #include"diago_mock.h"
 #include "source_psi/psi.h"
@@ -84,7 +85,7 @@ class DiagoDavPrepare
         phm = new hamilt::HamiltPW<double>(nullptr, nullptr, nullptr, nullptr, nullptr, nullptr);
 
 #ifdef __MPI 
-        const hsolver::diag_comm_info comm_info = {MPI_COMM_WORLD, mypnum, nprocs};
+        const hsolver::diag_comm_info comm_info = {POOL_WORLD, mypnum, nprocs};
 #else
         const hsolver::diag_comm_info comm_info = {mypnum, nprocs};
 #endif

source/source_hsolver/test/diago_david_test.cpp

@@ -1,5 +1,6 @@
 #include"source_hsolver/diago_david.h"
 #include"source_hsolver/diago_iter_assist.h"
+#include "source_base/parallel_comm.h"
 #include"source_pw/module_pwdft/hamilt_pw.h"
 #include"diago_mock.h"
 #include "source_psi/psi.h"
@@ -89,7 +90,7 @@ class DiagoDavPrepare
 		phm = new hamilt::HamiltPW<std::complex<double>>(nullptr, nullptr, nullptr, nullptr, nullptr, nullptr);
 
 #ifdef __MPI 
-        const hsolver::diag_comm_info comm_info = {MPI_COMM_WORLD, mypnum, nprocs};
+        const hsolver::diag_comm_info comm_info = {POOL_WORLD, mypnum, nprocs};
 #else
         const hsolver::diag_comm_info comm_info = {mypnum, nprocs};
 #endif