Feature: make force and stress of sDFT support GPU (#5487)

* refactor force in sdft

* refactor stress in sDFT

* make stress_ekin GPU

* finish sdft GPU

* fix compile

* add annotations

* fix bug of stress and force

* modify

Author: Qianruipku

source/Makefile.Objects

@@ -617,6 +617,7 @@ OBJS_PARALLEL=parallel_common.o\
     parallel_grid.o\
     parallel_kpoints.o\
     parallel_reduce.o\
+    parallel_device.o
 
 OBJS_SRCPW=H_Ewald_pw.o\
     dnrm2.o\
@@ -640,6 +641,7 @@ OBJS_SRCPW=H_Ewald_pw.o\
     forces_cc.o\
     forces_scc.o\
     fs_nonlocal_tools.o\
+    fs_kin_tools.o\
     force_op.o\
     stress_op.o\
     wf_op.o\

source/module_base/CMakeLists.txt

@@ -50,6 +50,7 @@ add_library(
     parallel_global.cpp
     parallel_comm.cpp
     parallel_reduce.cpp
+    parallel_device.cpp
     spherical_bessel_transformer.cpp
     cubic_spline.cpp
     module_mixing/mixing_data.cpp

source/module_base/module_device/device.cpp

@@ -191,27 +191,30 @@ else { return "cpu";
 }
 }
 
-int get_device_kpar(const int &kpar) {
+int get_device_kpar(const int& kpar, const int& bndpar)
+{
 #if __MPI && (__CUDA || __ROCM)
-  int temp_nproc;
-  MPI_Comm_size(MPI_COMM_WORLD, &temp_nproc);
-  if (temp_nproc != kpar) {
-    ModuleBase::WARNING("Input_conv",
-                        "None kpar set in INPUT file, auto set kpar value.");
-  }
-  // GlobalV::KPAR = temp_nproc;
-  // band the CPU processor to the devices
-  int node_rank = base_device::information::get_node_rank();
+    int temp_nproc = 0;
+    int new_kpar = kpar;
+    MPI_Comm_size(MPI_COMM_WORLD, &temp_nproc);
+    if (temp_nproc != kpar * bndpar)
+    {
+        new_kpar = temp_nproc / bndpar;
+        ModuleBase::WARNING("Input_conv", "kpar is not compatible with the number of processors, auto set kpar value.");
+    }
+    
+    // get the CPU rank of current node
+    int node_rank = base_device::information::get_node_rank();
 
-  int device_num = -1;
+    int device_num = -1;
 #if defined(__CUDA)
-  cudaGetDeviceCount(&device_num);
-  cudaSetDevice(node_rank % device_num);
+  cudaGetDeviceCount(&device_num); // get the number of GPU devices of current node
+  cudaSetDevice(node_rank % device_num); // band the CPU processor to the devices
 #elif defined(__ROCM)
   hipGetDeviceCount(&device_num);
   hipSetDevice(node_rank % device_num);
 #endif
-  return temp_nproc;
+  return new_kpar;
 #endif
   return kpar;
 }

source/module_base/module_device/device.h

@@ -40,7 +40,7 @@ std::string get_device_info(std::string device_flag);
  * @brief Get the device kpar object
  * for module_io GlobalV::KPAR
  */
-int get_device_kpar(const int& kpar);
+int get_device_kpar(const int& kpar, const int& bndpar);
 
 /**
  * @brief Get the device flag object
@@ -50,6 +50,12 @@ std::string get_device_flag(const std::string& device,
                             const std::string& basis_type);
 
 #if __MPI
+/**
+ * @brief Get the rank of current node
+ *        Note that GPU can only be binded with CPU in the same node
+ * 
+ * @return int 
+ */
 int get_node_rank();
 int get_node_rank_with_mpi_shared(const MPI_Comm mpi_comm = MPI_COMM_WORLD);
 int stringCmp(const void* a, const void* b);

source/module_base/parallel_common.cpp

@@ -11,15 +11,16 @@ void Parallel_Common::bcast_string(std::string& object) // Peize Lin fix bug 201
 {
     int size = object.size();
     MPI_Bcast(&size, 1, MPI_INT, 0, MPI_COMM_WORLD);
-    char* swap = new char[size + 1];
+    
     int my_rank;
     MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
-    if (0 == my_rank)
-        strcpy(swap, object.c_str());
-    MPI_Bcast(swap, size + 1, MPI_CHAR, 0, MPI_COMM_WORLD);
+    
     if (0 != my_rank)
-        object = static_cast<std::string>(swap);
-    delete[] swap;
+    {
+        object.resize(size);
+    }
+
+    MPI_Bcast(&object[0], size, MPI_CHAR, 0, MPI_COMM_WORLD);
     return;
 }
 

source/module_base/parallel_device.cpp

@@ -0,0 +1,38 @@
+#include "parallel_device.h"
+#ifdef __MPI
+namespace Parallel_Common
+{
+void bcast_data(std::complex<double>* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Bcast(object, n * 2, MPI_DOUBLE, 0, comm);
+}
+void bcast_data(std::complex<float>* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Bcast(object, n * 2, MPI_FLOAT, 0, comm);
+}
+void bcast_data(double* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Bcast(object, n, MPI_DOUBLE, 0, comm);
+}
+void bcast_data(float* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Bcast(object, n, MPI_FLOAT, 0, comm);
+}
+void reduce_data(std::complex<double>* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Allreduce(MPI_IN_PLACE, object, n * 2, MPI_DOUBLE, MPI_SUM, comm);
+}
+void reduce_data(std::complex<float>* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Allreduce(MPI_IN_PLACE, object, n * 2, MPI_FLOAT, MPI_SUM, comm);
+}
+void reduce_data(double* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Allreduce(MPI_IN_PLACE, object, n, MPI_DOUBLE, MPI_SUM, comm);
+}
+void reduce_data(float* object, const int& n, const MPI_Comm& comm)
+{
+    MPI_Allreduce(MPI_IN_PLACE, object, n, MPI_FLOAT, MPI_SUM, comm);
+}
+}
+#endif

source/module_base/parallel_device.h

@@ -1,39 +1,34 @@
+#ifndef __PARALLEL_DEVICE_H__
+#define __PARALLEL_DEVICE_H__
 #ifdef __MPI
 #include "mpi.h"
 #include "module_base/module_device/device.h"
+#include "module_base/module_device/memory_op.h"
 #include <complex>
-#include <string>
-#include <vector>
 namespace Parallel_Common
 {
-void bcast_complex(std::complex<double>* object, const int& n, const MPI_Comm& comm)
-{
-    MPI_Bcast(object, n * 2, MPI_DOUBLE, 0, comm);
-}
-void bcast_complex(std::complex<float>* object, const int& n, const MPI_Comm& comm)
-{
-    MPI_Bcast(object, n * 2, MPI_FLOAT, 0, comm);
-}
-void bcast_real(double* object, const int& n, const MPI_Comm& comm)
-{
-    MPI_Bcast(object, n, MPI_DOUBLE, 0, comm);
-}
-void bcast_real(float* object, const int& n, const MPI_Comm& comm)
-{
-    MPI_Bcast(object, n, MPI_FLOAT, 0, comm);
-}
+void bcast_data(std::complex<double>* object, const int& n, const MPI_Comm& comm);
+void bcast_data(std::complex<float>* object, const int& n, const MPI_Comm& comm);
+void bcast_data(double* object, const int& n, const MPI_Comm& comm);
+void bcast_data(float* object, const int& n, const MPI_Comm& comm);
+void reduce_data(std::complex<double>* object, const int& n, const MPI_Comm& comm);
+void reduce_data(std::complex<float>* object, const int& n, const MPI_Comm& comm);
+void reduce_data(double* object, const int& n, const MPI_Comm& comm);
+void reduce_data(float* object, const int& n, const MPI_Comm& comm);
 
-template <typename T, typename Device>
 /**
- * @brief bcast complex in Device
+ * @brief bcast data in Device
  * 
+ * @tparam T: float, double, std::complex<float>, std::complex<double>
+ * @tparam Device 
  * @param ctx Device ctx
  * @param object complex arrays in Device
  * @param n the size of complex arrays
  * @param comm MPI_Comm
  * @param tmp_space tmp space in CPU
  */
-void bcast_complex(const Device* ctx, T* object, const int& n, const MPI_Comm& comm, T* tmp_space = nullptr)
+template <typename T, typename Device>
+void bcast_dev(const Device* ctx, T* object, const int& n, const MPI_Comm& comm, T* tmp_space = nullptr)
 {
     const base_device::DEVICE_CPU* cpu_ctx = {};
     T* object_cpu = nullptr;
@@ -56,7 +51,7 @@ void bcast_complex(const Device* ctx, T* object, const int& n, const MPI_Comm& c
         object_cpu = object;
     }
 
-    bcast_complex(object_cpu, n, comm);
+    bcast_data(object_cpu, n, comm);
 
     if (base_device::get_device_type<Device>(ctx) == base_device::GpuDevice)
     {
@@ -70,7 +65,7 @@ void bcast_complex(const Device* ctx, T* object, const int& n, const MPI_Comm& c
 }
 
 template <typename T, typename Device>
-void bcast_real(const Device* ctx, T* object, const int& n, const MPI_Comm& comm, T* tmp_space = nullptr)
+void reduce_dev(const Device* ctx, T* object, const int& n, const MPI_Comm& comm, T* tmp_space = nullptr)
 {
     const base_device::DEVICE_CPU* cpu_ctx = {};
     T* object_cpu = nullptr;
@@ -93,7 +88,7 @@ void bcast_real(const Device* ctx, T* object, const int& n, const MPI_Comm& comm
         object_cpu = object;
     }
 
-    bcast_real(object_cpu, n, comm);
+    reduce_data(object_cpu, n, comm);
 
     if (base_device::get_device_type<Device>(ctx) == base_device::GpuDevice)
     {
@@ -105,7 +100,9 @@ void bcast_real(const Device* ctx, T* object, const int& n, const MPI_Comm& comm
     }
     return;
 }
+
 }
 
 
+#endif
 #endif

source/module_elecstate/elecstate_pw_sdft.cpp

@@ -14,14 +14,13 @@ void ElecStatePW_SDFT<T, Device>::psiToRho(const psi::Psi<T, Device>& psi)
     ModuleBase::TITLE(this->classname, "psiToRho");
     ModuleBase::timer::tick(this->classname, "psiToRho");
     const int nspin = PARAM.inp.nspin;
+    for (int is = 0; is < nspin; is++)
+    {
+        setmem_var_op()(this->ctx, this->rho[is], 0, this->charge->nrxx);
+    }
 
     if (GlobalV::MY_STOGROUP == 0)
     {
-        for (int is = 0; is < nspin; is++)
-        {
-            setmem_var_op()(this->ctx, this->rho[is], 0, this->charge->nrxx);
-        }
-
         for (int ik = 0; ik < psi.get_nk(); ++ik)
         {
             psi.fix_k(ik);

source/module_elecstate/module_charge/charge_extra.cpp

@@ -47,9 +47,20 @@ void Charge_Extra::Init_CE(const int& nspin, const int& natom, const int& nrxx,
 
     if (pot_order > 0)
     {
-        delta_rho1.resize(this->nspin, std::vector<double>(nrxx, 0.0));
-        delta_rho2.resize(this->nspin, std::vector<double>(nrxx, 0.0));
-        delta_rho3.resize(this->nspin, std::vector<double>(nrxx, 0.0));
+        // delta_rho1.resize(this->nspin, std::vector<double>(nrxx, 0.0));
+        // delta_rho2.resize(this->nspin, std::vector<double>(nrxx, 0.0));
+        // delta_rho3.resize(this->nspin, std::vector<double>(nrxx, 0.0));
+        // qianrui replace the above code with the following code.
+        // The above code cannot passed valgrind tests, which has an invalid read of size 32.
+        delta_rho1.resize(this->nspin);
+        delta_rho2.resize(this->nspin);
+        delta_rho3.resize(this->nspin);
+        for (int is = 0; is < this->nspin; is++)
+        {
+            delta_rho1[is].resize(nrxx, 0.0);
+            delta_rho2[is].resize(nrxx, 0.0);
+            delta_rho3[is].resize(nrxx, 0.0);
+        }
     }
 
     if(pot_order == 3)

source/module_esolver/esolver_ks_pw.cpp

@@ -623,7 +623,6 @@ void ESolver_KS_PW<T, Device>::cal_stress(ModuleBase::matrix& stress)
                   &this->sf,
                   &this->kv,
                   this->pw_wfc,
-                  this->psi,
                   this->__kspw_psi);
 
     // external stress