refactor pw

Author: Qianruipku

source/module_basis/module_pw/pw_basis_k.cpp

@@ -22,7 +22,6 @@ PW_Basis_K::~PW_Basis_K()
     delete[] igl2isz_k;
     delete[] igl2ig_k;
     delete[] gk2;
-    delete[] ig2ixyz_k_;
 #if defined(__CUDA) || defined(__ROCM)
     if (this->device == "gpu") {
         if (this->precision == "single") {
@@ -169,6 +168,7 @@ void PW_Basis_K::setupIndGk()
         syncmem_int_h2d_op()(gpu_ctx, cpu_ctx, this->d_igl2isz_k, this->igl2isz_k, this->npwk_max * this->nks);
     }
 #endif
+    this->get_ig2ixyz_k();
     return;
 }
 
@@ -334,8 +334,12 @@ int& PW_Basis_K::getigl2ig(const int ik, const int igl) const
 
 void PW_Basis_K::get_ig2ixyz_k()
 {
-    delete[] this->ig2ixyz_k_;
-    this->ig2ixyz_k_ = new int [this->npwk_max * this->nks];
+    if (this->device != "gpu")
+    {
+        //only GPU need to get ig2ixyz_k
+        return;
+    }
+    int * ig2ixyz_k_cpu = new int [this->npwk_max * this->nks];
     ModuleBase::Memory::record("PW_B_K::ig2ixyz", sizeof(int) * this->npwk_max * this->nks);
     assert(gamma_only == false); //We only finish non-gamma_only fft on GPU temperarily.
     for(int ik = 0; ik < this->nks; ++ik)
@@ -348,15 +352,12 @@ void PW_Basis_K::get_ig2ixyz_k()
             int ixy = this->is2fftixy[is];
             int iy = ixy % this->ny;
             int ix = ixy / this->ny;
-            ig2ixyz_k_[igl + ik * npwk_max] = iz + iy * nz + ix * ny * nz;
+            ig2ixyz_k_cpu[igl + ik * npwk_max] = iz + iy * nz + ix * ny * nz;
         }
     }
-#if defined(__CUDA) || defined(__ROCM)
-    if (this->device == "gpu") {
-        resmem_int_op()(gpu_ctx, ig2ixyz_k, this->npwk_max * this->nks);
-        syncmem_int_h2d_op()(gpu_ctx, cpu_ctx, this->ig2ixyz_k, this->ig2ixyz_k_, this->npwk_max * this->nks);
-    }
-#endif
+    resmem_int_op()(gpu_ctx, ig2ixyz_k, this->npwk_max * this->nks);
+    syncmem_int_h2d_op()(gpu_ctx, cpu_ctx, this->ig2ixyz_k, ig2ixyz_k_cpu, this->npwk_max * this->nks);
+    delete[] ig2ixyz_k_cpu;
 }
 
 std::vector<int> PW_Basis_K::get_ig2ix(const int ik) const

source/module_basis/module_pw/pw_basis_k.h

@@ -71,8 +71,6 @@ class PW_Basis_K : public PW_Basis
         const bool xprime_in = true
     );
 
-    void get_ig2ixyz_k();
-
   public:
     int nks=0;//number of k points in this pool
     ModuleBase::Vector3<double> *kvec_d=nullptr; // Direct coordinates of k points
@@ -88,8 +86,7 @@ class PW_Basis_K : public PW_Basis
 
     int *igl2isz_k=nullptr, * d_igl2isz_k = nullptr; //[npwk_max*nks] map (igl,ik) to (is,iz)
     int *igl2ig_k=nullptr;//[npwk_max*nks] map (igl,ik) to ig
-    int *ig2ixyz_k=nullptr;
-    int *ig2ixyz_k_=nullptr;
+    int *ig2ixyz_k=nullptr; ///< [npw] map ig to ixyz
 
     double *gk2=nullptr; // modulus (G+K)^2 of G vectors [npwk_max*nks]
 
@@ -108,6 +105,8 @@ class PW_Basis_K : public PW_Basis
     double * d_gk2 = nullptr; // modulus (G+K)^2 of G vectors [npwk_max*nks]
     //create igl2isz_k map array for fft
     void setupIndGk();
+    // get ig2ixyz_k
+    void get_ig2ixyz_k();
     //calculate G+K, it is a private function
     ModuleBase::Vector3<double> cal_GplusK_cartesian(const int ik, const int ig) const;
 

source/module_basis/module_pw/test/test4-4.cpp

@@ -213,13 +213,6 @@ TEST_F(PWTEST,test4_4)
             }
         }
 
-        //check getig2ixyz_k
-        pwtest.get_ig2ixyz_k();
-        for(int igl = 0; igl < npwk ; ++igl)
-        { 
-            EXPECT_GE(pwtest.ig2ixyz_k_[igl + ik * pwtest.npwk_max], 0);
-        }
-
     }
     delete []tmp; 
     delete [] rhor;

source/module_esolver/esolver_ks_pw.cpp

@@ -138,12 +138,6 @@ void ESolver_KS_PW<T, Device>::before_all_runners(UnitCell& ucell, const Input_p
 {
     // 1) call before_all_runners() of ESolver_KS
     ESolver_KS<T, Device>::before_all_runners(ucell, inp);
-#if defined(__CUDA) || defined(__ROCM)
-    if (PARAM.inp.device == "gpu")
-    {
-        this->pw_wfc->get_ig2ixyz_k();
-    }
-#endif
 
     // 3) initialize ElecState,
     if (this->pelec == nullptr)

tests/integrate/102_PW_BPCG/result.ref

@@ -1,8 +1,8 @@
-etotref -3937.5288918827272937
-etotperatomref -1968.7644459414
-totalforceref 21308.015678
-totalstressref 8138673.223144
+etotref -4869.74705201
+etotperatomref -2434.87352600
+totalforceref 5.19483000
+totalstressref 37241.44843500
 pointgroupref C_1
 spacegroupref C_1
 nksibzref 8
-totaltimeref 10.10
+totaltimeref 10.37

tests/integrate/103_PW_15_CF_CS_S1_smallg/result.ref

@@ -1,8 +1,8 @@
-etotref -378.4158765482854960
+etotref -378.4158765482866329
 etotperatomref -126.1386255161
-totalforceref 1005.225100
-totalstressref 2123.397615
+totalforceref 1005.225123
+totalstressref 2123.397548
 pointgroupref C_2v
 spacegroupref C_2v
 nksibzref 1
-totaltimeref 0.56
+totaltimeref 1.05

tests/integrate/121_PW_kspacing/INPUT

@@ -17,3 +17,4 @@ smearing_sigma          0.020000
 
 kspacing                0.6
 pseudo_dir	../../PP_ORB
+pw_seed                 1

tests/integrate/121_PW_kspacing/result.ref

@@ -1,6 +1,6 @@
-etotref -31.02022530858069
-etotperatomref -15.5101126543
+etotref -31.02022535682493
+etotperatomref -15.5101126784
 pointgroupref D_2h
 spacegroupref D_2h
 nksibzref 4
-totaltimeref 1.53
+totaltimeref 3.04