Merge branch 'develop' of https://github.com/deepmodeling/abacus-develop into hotfix

Author: Qianruipku

.github/workflows/coverage.yml

@@ -15,10 +15,10 @@ jobs:
         uses: actions/checkout@v4
       - name: Install Requirements for Coverage Testing
         run: |
-          apt update && apt install -y lcov
+          apt update && apt install -y lcov gpg
       - name: Building
         run: |
-          cmake -B build -DENABLE_DEEPKS=ON -DENABLE_LIBXC=ON -DBUILD_TESTING=ON -DENABLE_COVERAGE=ON
+          cmake -B build -DENABLE_COVERAGE=ON -DBUILD_TESTING=ON -DENABLE_DEEPKS=ON -DENABLE_LIBXC=ON -DENABLE_LIBRI=ON -DENABLE_PAW=ON -DENABLE_GOOGLEBENCH=ON -DENABLE_RAPIDJSON=ON
           cmake --build build -j`nproc`
           cmake --install build
       - name: Testing

CMakeLists.txt

@@ -436,22 +436,6 @@ if(ENABLE_FLOAT_FFTW)
 endif()
 
 if(ENABLE_DEEPKS)
-  # Torch uses outdated components to detect CUDA arch, causing failure on
-  # latest CUDA kits. Set CMake variable TORCH_CUDA_ARCH_LIST in the form of
-  # "major.minor" if required.
-  find_package(Torch REQUIRED)
-  if(NOT Torch_VERSION VERSION_LESS "2.1.0")
-    set_if_higher(CMAKE_CXX_STANDARD 17)
-  elseif(NOT Torch_VERSION VERSION_LESS "1.5.0")
-    set_if_higher(CMAKE_CXX_STANDARD 14)
-  endif()
-  include_directories(${TORCH_INCLUDE_DIRS})
-  if(MKL_FOUND)
-    list(PREPEND math_libs ${TORCH_LIBRARIES})
-  else()
-    list(APPEND math_libs ${TORCH_LIBRARIES})
-  endif()
-  add_compile_options(${TORCH_CXX_FLAGS})
   target_link_libraries(${ABACUS_BIN_NAME} deepks)
 
   find_path(libnpy_SOURCE_DIR npy.hpp HINTS ${libnpy_INCLUDE_DIR})
@@ -470,6 +454,25 @@ if(ENABLE_DEEPKS)
   add_compile_definitions(__DEEPKS)
 endif()
 
+# Torch uses outdated components to detect CUDA arch, causing failure on
+# latest CUDA kits. Set CMake variable TORCH_CUDA_ARCH_LIST in the form of
+# "major.minor" if required.
+if(ENABLE_DEEPKS OR DEFINED Torch_DIR)
+  find_package(Torch REQUIRED)
+  if(NOT Torch_VERSION VERSION_LESS "2.1.0")
+    set_if_higher(CMAKE_CXX_STANDARD 17)
+  elseif(NOT Torch_VERSION VERSION_LESS "1.5.0")
+    set_if_higher(CMAKE_CXX_STANDARD 14)
+  endif()
+  include_directories(${TORCH_INCLUDE_DIRS})
+  if(MKL_FOUND)
+    list(PREPEND math_libs ${TORCH_LIBRARIES})
+  else()
+    list(APPEND math_libs ${TORCH_LIBRARIES})
+  endif()
+  add_compile_options(${TORCH_CXX_FLAGS})
+endif()
+
 if (ENABLE_CNPY)
   find_path(cnpy_SOURCE_DIR
     cnpy.h
@@ -590,13 +593,14 @@ if(DEFINED DeePMD_DIR)
     add_compile_definitions(__DPMDC)
   else()
     target_link_libraries(${ABACUS_BIN_NAME} DeePMD::deepmd_cc)
-    if(NOT DEFINED TensorFlow_DIR)
-      set(TensorFlow_DIR ${DeePMD_DIR})
-    endif()
-    find_package(TensorFlow REQUIRED)
-    if(TensorFlow_FOUND)
-      target_link_libraries(${ABACUS_BIN_NAME} TensorFlow::tensorflow_cc)
-    endif()
+  endif()
+endif()
+
+if(DEFINED TensorFlow_DIR)
+  find_package(TensorFlow REQUIRED)
+  include_directories(${TensorFlow_DIR}/include)
+  if(TensorFlow_FOUND)
+    target_link_libraries(${ABACUS_BIN_NAME} TensorFlow::tensorflow_cc)
   endif()
 endif()
 

docs/advanced/input_files/input-main.md

@@ -39,7 +39,6 @@
     - [pw\_diag\_thr](#pw_diag_thr)
     - [pw\_diag\_nmax](#pw_diag_nmax)
     - [pw\_diag\_ndim](#pw_diag_ndim)
-    - [diago\_full\_acc](#diago_full_acc)
     - [erf\_ecut](#erf_ecut)
     - [fft\_mode](#fft_mode)
     - [erf\_height](#erf_height)
@@ -779,12 +778,6 @@ These variables are used to control the plane wave related parameters.
 - **Description**: Only useful when you use `ks_solver = dav` or `ks_solver = dav_subspace`. It indicates dimension of workspace(number of wavefunction packets, at least 2 needed) for the Davidson method. A larger value may yield a smaller number of iterations in the algorithm but uses more memory and more CPU time in subspace diagonalization.
 - **Default**: 4
 
-### diago_full_acc
-
-- **Type**: bool
-- **Description**: Only useful when you use `ks_solver = dav_subspace`. If `TRUE`, all the empty states are diagonalized at the same level of accuracy of the occupied ones. Otherwise the empty states are diagonalized using a larger threshold (10-5) (this should not affect total energy, forces, and other ground-state properties).
-- **Default**: false
-
 ### erf_ecut
 
 - **Type**: Real
@@ -925,7 +918,7 @@ calculations.
   - **cg**: cg method.
   - **bpcg**: bpcg method, which is a block-parallel Conjugate Gradient (CG) method, typically exhibits higher acceleration in a GPU environment.
   - **dav**: the Davidson algorithm.
-  - **dav_subspace**: subspace Davidson algorithm
+  - **dav_subspace**: Davidson algorithm without orthogonalization operation, this method is the most recommended for efficiency. `pw_diag_ndim` can be set to 2 for this method.
 
   For atomic orbitals basis,
 

docs/advanced/install.md

@@ -38,13 +38,22 @@ cmake -B build -DENABLE_DEEPKS=1 -DTorch_DIR=~/libtorch/share/cmake/Torch/ -Dlib
 If the Deep Potential model is employed in Molecule Dynamics calculations, the following prerequisites and steps are needed:
 
 - [DeePMD-kit](https://github.com/deepmodeling/deepmd-kit)
-- [TensorFlow](https://www.tensorflow.org/)
+- [TensorFlow](https://www.tensorflow.org/) (optional)
+- [LibTorch](https://pytorch.org/) (optional)
 
+In the simplest case, the `tensorflow_cc` and `torch` libraries are in the same directory as the `deepmd_c`/`deepmd_cc` libraries, then
 ```bash
-cmake -B build -DDeePMD_DIR=~/deepmd-kit -DTensorFlow_DIR=~/tensorflow
+cmake -B build -DDeePMD_DIR=/dir_to_deepmd-kit
 ```
+DeePMD-kit supports TensorFlow backend but its libraries are placed at another directory, then
 
-> `deepmd_c`/`deepmd_cc` and `tensorflow_cc` libraries would be called according to `DeePMD_DIR` and `TensorFlow_DIR`, which is showed in detail in [this page](https://github.com/deepmodeling/deepmd-kit/blob/master/doc/inference/cxx.md). If `TensorFlow_DIR` is not defined, it will be the same as `DeePMD_DIR`. Note that `tensorflow_cc` is not required if `deepmd_c` is found.
+```bash
+cmake -B build -DDeePMD_DIR=/dir_to_deepmd-kit -DTensorFlow_DIR=/dir_to_tensorflow
+```
+Similarly, DeePMD-kit supports PyTorch backend but its libraries are placed at another directory, then
+```bash
+cmake -B build -DDeePMD_DIR=/dir_to_deepmd-kit -DTorch_DIR=/dir_to_pytorch
+```
 
 ## Build with LibRI and LibComm
 
@@ -93,9 +102,9 @@ cmake -B build -DUSE_CUDA=1 -DCMAKE_CUDA_COMPILER=${path to cuda toolkit}/bin/nv
 
 ## Build math library from source
 
-> Note: This flag is **enabled by default**. It will get better performance than the standard implementation on `gcc` and `clang`. But it **will be disabled** when using `Intel Compiler` since the math functions will get wrong results and the performance is also unexpectly poor.
+> Note: We recommend using the latest available compiler sets, since they offer faster implementations of math functions.
 
-To build math functions from source code, instead of using c++ standard implementation, define `USE_ABACUS_LIBM` flag.
+This flag is disabled by default. To build math functions from source code, define `USE_ABACUS_LIBM` flag. It is expected to get a better performance on legacy versions of `gcc` and `clang`.
 
 Currently supported math functions:
  `sin`, `cos`, `sincos`, `exp`, `cexp`
@@ -282,15 +291,21 @@ directly.
 
 > Note: This part is only required if you want to load a trained DeeP Potential and run molecular dynamics with that. To train the DeeP Potential with DP-GEN, no extra prerequisite is needed and please refer to [this page](http://abacus.deepmodeling.com/en/latest/advanced/interface/dpgen.html) for ABACUS interface with DP-GEN.
 
-To compile ABACUS with DeePMD-kit, you need to define `DeePMD_DIR` and `TensorFlow_DIR` in the file `Makefile.vars` or use
+To compile ABACUS with DeePMD-kit, you need to define `DeePMD_DIR` and `TensorFlow_DIR` (TensorFlow Backend, optional) and/or `LIBTORCH_DIR` (PyTorch Backend, optional) in the file `Makefile.vars`.
 
+Or the `tensorflow_cc` and `torch` libraries are in the same directory as the `deepmd_c`/`deepmd_cc` libraries, then
 ```makefile
-make DeePMD_DIR=~/deepmd-kit TensorFlow_DIR=~/tensorflow
+make DeePMD_DIR=/dir_to_deepmd-kit
 ```
+DeePMD-kit supports TensorFlow backend but its libraries are placed at another directory, then
 
-directly.
-
-> `deepmd_c`/`deepmd_cc` and `tensorflow_cc` libraries would be called according to `DeePMD_DIR` and `TensorFlow_DIR`, which is showed in detail in [this page](https://github.com/deepmodeling/deepmd-kit/blob/master/doc/inference/cxx.md). If `TensorFlow_DIR` is not defined, it will be the same as `DeePMD_DIR`. Note that `tensorflow_cc` is not required if `deepmd_c` is found.
+```makefile
+make DeePMD_DIR=/dir_to_deepmd-kit TensorFlow_DIR=/dir_to_tensorflow
+```
+Similarly, DeePMD-kit supports PyTorch backend but its libraries are placed at another directory, then
+```makefile
+make DeePMD_DIR=/dir_to_deepmd-kit Torch_DIR=/dir_to_pytorch
+```
 
 ### Add LibRI Support
 To use new EXX, you need two libraries: [LibRI](https://github.com/abacusmodeling/LibRI) and [LibComm](https://github.com/abacusmodeling/LibComm) and need to define `LIBRI_DIR` and `LIBCOMM_DIR` in the file `Makefile.vars` or use

examples/lr-tddft/lcao_Si2/INPUT

@@ -37,4 +37,3 @@ out_alllog	1
 
 nvirt 19
 abs_wavelen_range  100 175
-#diago_full_acc 1

python/pyabacus/src/py_diago_dav_subspace.hpp

@@ -106,7 +106,7 @@ class PyDiagoDavSubspace
         double tol,
         int max_iter,
         bool need_subspace,
-        std::vector<bool> is_occupied,
+        std::vector<double> diag_ethr,
         bool scf_type,
         hsolver::diag_comm_info comm_info
     ) {
@@ -141,7 +141,7 @@ class PyDiagoDavSubspace
             comm_info
         );
 
-        return obj->diag(hpsi_func, psi, nbasis, eigenvalue, is_occupied, scf_type);
+        return obj->diag(hpsi_func, psi, nbasis, eigenvalue, diag_ethr.data(), scf_type);
     }
 
 private:

python/pyabacus/src/py_diago_david.hpp

@@ -132,7 +132,6 @@ class PyDiagoDavid
             const std::complex<double> *psi_in, 
             std::complex<double> *spsi_out, 
             const int nrow, 
-            const int npw, 
             const int nbands
         ) {
             syncmem_op()(this->ctx, this->ctx, spsi_out, psi_in, static_cast<size_t>(nbands * nrow));

python/pyabacus/src/py_hsolver.cpp

@@ -59,8 +59,8 @@ void bind_hsolver(py::module& m)
                 The maximum number of iterations.
             need_subspace : bool
                 Whether to use the subspace function.
-            is_occupied : list[bool]
-                A list of boolean values indicating whether the band is occupied,
+            diag_ethr : list[float]
+                A list of float values indicating the thresholds of each band for the diagonalization,
                 meaning that the corresponding eigenvalue is to be calculated.
             scf_type : bool
                 Whether to use the SCF type, which is used to determine the
@@ -76,7 +76,7 @@ void bind_hsolver(py::module& m)
         "tol"_a, 
         "max_iter"_a, 
         "need_subspace"_a, 
-        "is_occupied"_a, 
+        "diag_ethr"_a, 
         "scf_type"_a, 
         "comm_info"_a)
         .def("set_psi", &py_hsolver::PyDiagoDavSubspace::set_psi, R"pbdoc(

python/pyabacus/src/pyabacus/hsolver/_hsolver.py

@@ -25,7 +25,7 @@ def dav_subspace(
     tol: float = 1e-2,
     max_iter: int = 1000,
     need_subspace: bool = False,
-    is_occupied: Union[List[bool], None] = None,
+    diag_ethr: Union[List[float], None] = None,
     scf_type: bool = False
 ) -> Tuple[NDArray[np.float64], NDArray[np.complex128]]:
     """ A function to diagonalize a matrix using the Davidson-Subspace method.
@@ -52,10 +52,8 @@ def dav_subspace(
         The maximum number of iterations, by default 1000.
     need_subspace : bool, optional
         Whether to use subspace function, by default False.
-    is_occupied : List[bool] | None, optional
-        The list of occupied bands, by default None. This indicates how many eigenvalues 
-        need to be calculated, starting from the smallest eigenvalue. Only the energy levels 
-        occupied by electrons (occupied) need to be calculated.
+    diag_ethr : List[float] | None, optional
+        The list of thresholds of bands, by default None.
     scf_type : bool, optional
         Indicates whether the calculation is a self-consistent field (SCF) calculation. 
         If True, the initial precision of eigenvalue calculation can be coarse. 
@@ -72,8 +70,8 @@ def dav_subspace(
     if not callable(mvv_op):
         raise TypeError("mvv_op must be a callable object.")
 
-    if is_occupied is None:
-        is_occupied = [True] * num_eigs
+    if diag_ethr is None:
+        diag_ethr = [tol] * num_eigs
 
     if init_v.ndim != 1 or init_v.dtype != np.complex128:
         init_v = init_v.flatten().astype(np.complex128, order='C')
@@ -93,7 +91,7 @@ def dav_subspace(
         tol,
         max_iter,
         need_subspace,
-        is_occupied,
+        diag_ethr,
         scf_type,
         comm_info
     )
@@ -113,7 +111,6 @@ def davidson(
     tol: float = 1e-2,
     max_iter: int = 1000,
     use_paw: bool = False,
-    # is_occupied: Union[List[bool], None] = None,
     # scf_type: bool = False
 ) -> Tuple[NDArray[np.float64], NDArray[np.complex128]]:
     """ A function to diagonalize a matrix using the Davidson-Subspace method.

source/Makefile

@@ -135,16 +135,17 @@ ifdef LIBRI_DIR
     HONG += -D__EXX -DEXX_H_COMM=2 -DEXX_DM=3 -DTEST_EXX_LCAO=0 -DTEST_EXX_RADIAL=1
 endif
 
+ifdef LIBNPY_DIR
+    CNPY_INCLUDE_DIR = -I${LIBNPY_DIR}/include
+    HONG += -D__DEEPKS
+    INCLUDES += $(CNPY_INCLUDE_DIR)
+endif
+
 ifdef LIBTORCH_DIR
- ifdef LIBNPY_DIR
-      LIBTORCH_INCLUDE_DIR =  -I${LIBTORCH_DIR}/include -I${LIBTORCH_DIR}/include/torch/csrc/api/include
-      LIBTORCH_LIB_DIR= ${LIBTORCH_DIR}/lib
-      LIBTORCH_LIB = -L${LIBTORCH_LIB_DIR} -ltorch -lc10 -Wl,-rpath,${LIBTORCH_LIB_DIR} -Wl,--no-as-needed,"${LIBTORCH_LIB_DIR}/libtorch_cpu.so" -Wl,--as-needed ${LIBTORCH_LIB_DIR}/libc10.so -lpthread -Wl,--no-as-needed,"${LIBTORCH_LIB_DIR}/libtorch.so" -Wl,--as-needed 
-      CNPY_INCLUDE_DIR = -I${LIBNPY_DIR}/include
-      HONG += -D__DEEPKS
-      INCLUDES += $(LIBTORCH_INCLUDE_DIR) $(CNPY_INCLUDE_DIR)
-      LIBS += $(LIBTORCH_LIB)
-  endif
+    LIBTORCH_INCLUDE_DIR =  -I${LIBTORCH_DIR}/include -I${LIBTORCH_DIR}/include/torch/csrc/api/include
+    INCLUDES += $(LIBTORCH_INCLUDE_DIR)
+    LIBTORCH_LIB_DIR= ${LIBTORCH_DIR}/lib
+    LIBS += -L${LIBTORCH_LIB_DIR} -ltorch -lc10 -Wl,-rpath,${LIBTORCH_LIB_DIR} -Wl,--no-as-needed,"${LIBTORCH_LIB_DIR}/libtorch_cpu.so" -Wl,--as-needed ${LIBTORCH_LIB_DIR}/libc10.so -lpthread -Wl,--no-as-needed,"${LIBTORCH_LIB_DIR}/libtorch.so" -Wl,--as-needed 
 endif
 
 ifdef DeePMD_DIR
@@ -162,20 +163,19 @@ ifdef DeePMD_DIR
         LIBS += -L${DeeP_LIB_DIR} -ldeepmd_c -Wl,-rpath=${DeeP_LIB_DIR}
     else
         LIBS += -L${DeeP_LIB_DIR} -ldeepmd_cc -Wl,-rpath=${DeeP_LIB_DIR}
-
-        #==========================
-        # TensorFlow package
-        #==========================
-        ifndef TensorFlow_DIR
-            TensorFlow_DIR = ${DeePMD_DIR}
-        endif
-        TensorFlow_LIB_DIR = ${TensorFlow_DIR}/lib
-        LIBS += -L${TensorFlow_LIB_DIR} -ltensorflow_cc -Wl,-rpath=${TensorFlow_LIB_DIR}
-        TensorFlow_INCLUDE_DIR = ${TensorFlow_DIR}/include
-        INCLUDES += -I${TensorFlow_INCLUDE_DIR}
     endif
 endif
 
+ifdef TensorFlow_DIR
+    #==========================
+    # TensorFlow package
+    #==========================
+    TensorFlow_LIB_DIR = ${TensorFlow_DIR}/lib
+    LIBS += -L${TensorFlow_LIB_DIR} -ltensorflow_cc -Wl,-rpath=${TensorFlow_LIB_DIR}
+    TensorFlow_INCLUDE_DIR = ${TensorFlow_DIR}/include
+    INCLUDES += -I${TensorFlow_INCLUDE_DIR}
+endif
+
 ifdef PEXSI_DIR
     OBJS_ABACUS += ${OBJS_HSOLVER_PEXSI}
     INCLUDES += -I${PEXSI_DIR}/include -I${PARMETIS_DIR}/include -I${DSUPERLU_DIR}/include