make sdft+bpcg support GPU

Qianruipku · Qianruipku · commit 30bc4ade45ec · 2025-01-27T21:04:00.000+08:00
diff --git a/source/module_base/para_gemm.cpp b/source/module_base/para_gemm.cpp
@@ -241,15 +241,17 @@ void PGemmCN<T, Device>::multiply_col(const T alpha, const T* A, const T* B, con
     {
         T* Cglobal_cpu = nullptr;
         T* Clocal_cpu = C_tmp.data();
-        ;
+        std::vector<T> cpu_tmp;
+        
         if (std::is_same<Device, base_device::DEVICE_GPU>::value)
         {
             delmem_dev_op()(Atmp_device);
 
             syncmem_d2h_op()(Clocal_cpu, C_local, size_C_local);
             delmem_dev_op()(C_local);
 
-            resmem_dev_op()(Cglobal_cpu, size_C_global);
+            cpu_tmp.resize(size_C_global);
+            Cglobal_cpu = cpu_tmp.data();
         }
         else
         {
@@ -269,7 +271,6 @@ void PGemmCN<T, Device>::multiply_col(const T alpha, const T* A, const T* B, con
         if (std::is_same<Device, base_device::DEVICE_GPU>::value)
         {
             syncmem_h2d_op()(C, Cglobal_cpu, size_C_global);
-            delmem_dev_op()(Cglobal_cpu);
         }
     }
     else
diff --git a/tests/integrate/102_PW_BPCG_GPU/INPUT b/tests/integrate/102_PW_BPCG_GPU/INPUT
@@ -15,6 +15,7 @@ smearing_sigma      0.02
 ecutwfc             40
 scf_thr             1e-7
 scf_nmax            100
+bndpar              2
 
 #Parameters (LCAO)
 basis_type          pw
diff --git a/tests/integrate/102_PW_BPCG_GPU/result.ref b/tests/integrate/102_PW_BPCG_GPU/result.ref
@@ -1,8 +1,8 @@
-etotref -4869.7470518350019120
-etotperatomref -2434.8735259175
-totalforceref 5.207670
-totalstressref 37241.465646
+etotref -4869.7470520365577613
+etotperatomref -2434.8735260183
+totalforceref 5.202524
+totalstressref 37241.827525
 pointgroupref C_1
 spacegroupref C_1
 nksibzref 8
-totaltimeref 4.25
+totaltimeref   4.25
diff --git a/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/INPUT b/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/INPUT
@@ -0,0 +1,39 @@
+INPUT_PARAMETERS
+#Parameters (1.General)
+suffix			autotest
+calculation     scf
+esolver_type    sdft
+method_sto      2
+device          gpu
+ks_solver       bpcg
+
+symmetry		0
+pseudo_dir	../../PP_ORB
+
+kpar            1
+bndpar          2
+
+nbands			9
+nbands_sto		all
+
+nche_sto		120
+seed_sto        20000
+cal_force       1
+cal_stress      1
+
+#Parameters (2.Iteration)
+ecutwfc			30
+scf_thr			1e-6
+scf_nmax		20
+
+
+#Parameters (3.Basis)
+basis_type		pw
+
+#Parameters (4.Smearing)
+smearing_method		fd
+smearing_sigma		0.6
+
+#Parameters (5.Mixing)
+mixing_type		plain
+mixing_beta		0.7
diff --git a/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/KPT b/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/KPT
@@ -0,0 +1,4 @@
+K_POINTS
+0
+Gamma
+1 1 1 0 0 0
diff --git a/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/README b/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/README
@@ -0,0 +1,12 @@
+This test for:
+*SDFT
+*Si
+*kpoints 1*1*1
+*10 KS + complete orbitals
+*ks_solver bpcg
+*mixing_type plain
+*mixing_beta 0.7
+*sto_method  2
+*seed_sto > 0
+*bndpar  2
+*kpar    1
diff --git a/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/STRU b/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/STRU
@@ -0,0 +1,19 @@
+ATOMIC_SPECIES
+Si 28 Si.pz-vbc.UPF
+
+LATTICE_CONSTANT
+5  // add lattice constant
+
+LATTICE_VECTORS
+0.5 0 0.5
+0.5 0.5 0
+0 0.5 0.5
+
+ATOMIC_POSITIONS
+Direct 
+
+Si	// Element type	
+0.0	// magnetism
+2	
+0.10 0.00 0.20 1 1 1
+0.5 0.5 0.5 1 1 1
diff --git a/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/jd b/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/jd
@@ -0,0 +1 @@
+test parallel method bndpar with BPCG and GPU
diff --git a/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/result.ref b/tests/integrate/187_PW_SDFT_MALL_BPCG_GPU/result.ref
@@ -0,0 +1,5 @@
+etotref -96.9361135992121490
+etotperatomref -48.4680567996
+totalforceref 248.977364
+totalstressref 230452.899030
+totaltimeref 7.19
diff --git a/tests/integrate/CASES_CPU.txt b/tests/integrate/CASES_CPU.txt
@@ -24,6 +24,7 @@
 101_PW_GTH_CF_CS_Si
 102_PW_DA_davidson
 102_PW_BPCG
+102_PW_BPCG_BP
 102_PW_CG
 102_PW_DS_davsubspace
 102_PW_DS_davsubspace_sca
@@ -123,6 +124,7 @@
 184_PW_BNDKPAR_SDFT_MALL
 184_PW_BNDPAR_SDFT_10S
 184_PW_BNDPAR_SDFT_5D10S
+184_PW_BPCG_SDFT_5D10S
 184_PW_KPAR_SDFT_ALL
 185_PW_SDFT_10D10S_METHD2
 185_PW_SDFT_10S_METHD2
diff --git a/tests/integrate/CASES_GPU.txt b/tests/integrate/CASES_GPU.txt
@@ -3,6 +3,7 @@
 102_PW_BPCG_GPU
 187_PW_SDFT_ALL_GPU
 187_PW_SDFT_MALL_GPU
+187_PW_SDFT_MALL_BPCG_GPU
 187_PW_MD_SDFT_ALL_GPU
 930_NO_BI2SE2CU2O2_GPU
 930_NO_BI2SE2CU2O2_k_GPU

Original file line number	Diff line number	Diff line change
`@@ -241,15 +241,17 @@ void PGemmCN<T, Device>::multiply_col(const T alpha, const T* A, const T* B, con`
`241`	`241`	`{`
`242`	`242`	`T* Cglobal_cpu = nullptr;`
`243`	`243`	`T* Clocal_cpu = C_tmp.data();`
`244`		`- ;`
	`244`	`+ std::vector<T> cpu_tmp;`
	`245`	`+`
`245`	`246`	`if (std::is_same<Device, base_device::DEVICE_GPU>::value)`
`246`	`247`	`{`
`247`	`248`	`delmem_dev_op()(Atmp_device);`
`248`	`249`
`249`	`250`	`syncmem_d2h_op()(Clocal_cpu, C_local, size_C_local);`
`250`	`251`	`delmem_dev_op()(C_local);`
`251`	`252`
`252`		`- resmem_dev_op()(Cglobal_cpu, size_C_global);`
	`253`	`+ cpu_tmp.resize(size_C_global);`
	`254`	`+ Cglobal_cpu = cpu_tmp.data();`
`253`	`255`	`}`
`254`	`256`	`else`
`255`	`257`	`{`
`@@ -269,7 +271,6 @@ void PGemmCN<T, Device>::multiply_col(const T alpha, const T* A, const T* B, con`
`269`	`271`	`if (std::is_same<Device, base_device::DEVICE_GPU>::value)`
`270`	`272`	`{`
`271`	`273`	`syncmem_h2d_op()(C, Cglobal_cpu, size_C_global);`
`272`		`- delmem_dev_op()(Cglobal_cpu);`
`273`	`274`	`}`
`274`	`275`	`}`
`275`	`276`	`else`
-Original file line number
+Diff line change
@@ @@ -0,0 +1,4 @@ @@
 +K_POINTS
 +0
 +Gamma
 +1 1 1 0 0 0
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	`+test parallel method bndpar with BPCG and GPU`