fix: reject genelpa with GPU device (#7738)

Author: Stardust0831

docs/advanced/input_files/input-main.md

@@ -1160,7 +1160,7 @@
   For numerical atomic orbitals basis,
 
   - lapack: Use LAPACK to diagonalize the Hamiltonian, only used for serial version
-  - genelpa: Use GEN-ELPA to diagonalize the Hamiltonian.
+  - genelpa: Use the CPU-only GEN-ELPA interface to diagonalize the Hamiltonian.
   - scalapack_gvx: Use Scalapack to diagonalize the Hamiltonian.
   - cusolver: Use CUSOLVER to diagonalize the Hamiltonian, at least one GPU is needed.
   - cusolvermp: Use CUSOLVER to diagonalize the Hamiltonian, supporting multi-GPU devices. Note that you should set the number of MPI processes equal to the number of GPUs.

docs/parameters.yaml

@@ -540,7 +540,7 @@ parameters:
       For numerical atomic orbitals basis,
 
       * lapack: Use LAPACK to diagonalize the Hamiltonian, only used for serial version
-      * genelpa: Use GEN-ELPA to diagonalize the Hamiltonian.
+      * genelpa: Use the CPU-only GEN-ELPA interface to diagonalize the Hamiltonian.
       * scalapack_gvx: Use Scalapack to diagonalize the Hamiltonian.
       * cusolver: Use CUSOLVER to diagonalize the Hamiltonian, at least one GPU is needed.
       * cusolvermp: Use CUSOLVER to diagonalize the Hamiltonian, supporting multi-GPU devices. Note that you should set the number of MPI processes equal to the number of GPUs.

source/source_io/module_parameter/read_input_item_elec_stru.cpp

@@ -60,7 +60,7 @@ For plane-wave basis,
 For numerical atomic orbitals basis,
 
 * lapack: Use LAPACK to diagonalize the Hamiltonian, only used for serial version
-* genelpa: Use GEN-ELPA to diagonalize the Hamiltonian.
+* genelpa: Use the CPU-only GEN-ELPA interface to diagonalize the Hamiltonian.
 * scalapack_gvx: Use Scalapack to diagonalize the Hamiltonian.
 * cusolver: Use CUSOLVER to diagonalize the Hamiltonian, at least one GPU is needed.
 * cusolvermp: Use CUSOLVER to diagonalize the Hamiltonian, supporting multi-GPU devices. Note that you should set the number of MPI processes equal to the number of GPUs.
@@ -164,6 +164,13 @@ Then the user has to correct the input file and restart the calculation.)";
                 }
                 else if (ks_solver == "genelpa")
                 {
+                    if (para.input.device == "gpu")
+                    {
+                        ModuleBase::WARNING_QUIT(
+                            "ReadInput",
+                            "ks_solver = genelpa does not support GPU acceleration. "
+                            "Please use ks_solver = elpa with device = gpu.");
+                    }
 #ifndef __ELPA
                     ModuleBase::WARNING_QUIT("Input",
                                              "Can not use genelpa if abacus is not compiled with "

source/source_io/test_serial/read_input_item_test.cpp

@@ -698,6 +698,14 @@ TEST_F(InputTest, Item_test)
         param.input.device = "gpu";
         it->second.reset_value(it->second, param);
         EXPECT_EQ(param.input.ks_solver, "cusolver");
+
+        param.input.ks_solver = "genelpa";
+        param.input.basis_type = "lcao";
+        param.input.device = "gpu";
+        testing::internal::CaptureStdout();
+        EXPECT_EXIT(it->second.check_value(it->second, param), ::testing::ExitedWithCode(1), "");
+        output = testing::internal::GetCapturedStdout();
+        EXPECT_THAT(output, testing::HasSubstr("Please use ks_solver = elpa with device = gpu"));
 #ifdef __ELPA
         param.input.towannier90 = true;
         param.input.basis_type = "lcao_in_pw";