Add Mt_Gemm for the nonlocal_pw

source/module_hamilt_pw/hamilt_pwdft/VNL_in_pw.cpp

@@ -271,7 +271,13 @@ void pseudopot_cell_vnl::init(const UnitCell& ucell,
             resmem_sh_op()(s_tab, this->tab.getSize());
             resmem_ch_op()(c_vkb, nkb * npwx);
         }
+        #ifdef __DSP
+        base_device::memory::resize_memory_op_mt<std::complex<double>, base_device::DEVICE_CPU>()
+        (this->z_vkb, this->vkb.size, "Nonlocal<PW>::ps");
+        memcpy(this->z_vkb,this->vkb.c,this->vkb.size*16);
+        #else
         this->z_vkb = this->vkb.c;
+        #endif
         this->d_tab = this->tab.ptr;
         // There's no need to delete double precision pointers while in a CPU environment.
     }

source/module_hamilt_pw/hamilt_pwdft/operator_pw/nonlocal_pw.cpp

@@ -185,7 +185,12 @@ void Nonlocal<OperatorPW<T, Device>>::add_nonlocal_pp(T *hpsi_in, const T *becp,
         int npm = m;
         //<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
         // denghui replace 2022-10-20
-        gemm_op()(
+        #ifdef __DSP
+            ModuleBase::gemm_op_mt<T, Device>()
+        #else
+            gemm_op()
+        #endif
+            (
             transa,
             transb,
             this->npw,
@@ -259,7 +264,12 @@ void Nonlocal<OperatorPW<T, Device>>::act(
             int npm = nbands;
             //<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
             // denghui replace 2022-10-20
-            gemm_op()(
+            #ifdef __DSP
+            ModuleBase::gemm_op_mt<T, Device>()
+            #else
+            gemm_op()
+            #endif
+            (
                 transa,
                 transb,
                 nkb,

source/module_hamilt_pw/hamilt_pwdft/operator_pw/nonlocal_pw.h

@@ -89,8 +89,13 @@ class Nonlocal<OperatorPW<T, Device>> : public OperatorPW<T, Device>
     using gemm_op = ModuleBase::gemm_op<T, Device>;
     using nonlocal_op = nonlocal_pw_op<Real, Device>;
     using setmem_complex_op = base_device::memory::set_memory_op<T, Device>;
+    #ifdef __DSP
+    using resmem_complex_op = base_device::memory::resize_memory_op_mt<T, Device>;
+    using delmem_complex_op = base_device::memory::delete_memory_op_mt<T, Device>;
+    #else
     using resmem_complex_op = base_device::memory::resize_memory_op<T, Device>;
     using delmem_complex_op = base_device::memory::delete_memory_op<T, Device>;
+    #endif
     using syncmem_complex_h2d_op = base_device::memory::synchronize_memory_op<T, Device, base_device::DEVICE_CPU>;
 
     T one{1, 0};

source/module_io/read_input_item_system.cpp

@@ -812,7 +812,12 @@ void ReadInput::item_system()
                 const std::string warningstr = nofound_str(avail_list, "precision");
                 ModuleBase::WARNING_QUIT("ReadInput", warningstr);
             }
-
+            if (para.inp.precision == "single" && para.inp.basis_type == "lcao")
+            {
+                ModuleBase::WARNING_QUIT(
+                    "ReadInput",
+                    "Single precision is not supported for NAO basis,\nPlease use double precision for NAO basis.\n");
+            }
             // cpu single precision is not supported while float_fftw lib is not available
             if (para.inp.device == "cpu" && para.inp.precision == "single")
             {