Release 1.0

.gitignore

@@ -6,3 +6,4 @@ src/xc_integrator/local_work_driver/host/obara_saika/test/*.x
 src/xc_integrator/local_work_driver/host/obara_saika/generator/integral*
 src/xc_integrator/local_work_driver/host/obara_saika/generator/obara*
 src/xc_integrator/local_work_driver/host/obara_saika/generator/*.x
+*.swp

cmake/gauxc-dep-versions.cmake

@@ -11,13 +11,13 @@ set( GAUXC_CUTLASS_REPOSITORY https://github.com/NVIDIA/cutlass.git )
 set( GAUXC_CUTLASS_REVISION v2.10.0 )
 
 set( GAUXC_EXCHCXX_REPOSITORY https://github.com/wavefunction91/ExchCXX.git )
-set( GAUXC_EXCHCXX_REVISION   21a4700a826ec0beae1311a1d59677393bcb168f )
+set( GAUXC_EXCHCXX_REVISION   f932a1bd3a0ff9d5b1de4cfe7fe86bfb1f20cb74 )
 
 set( GAUXC_GAU2GRID_REPOSITORY https://github.com/dgasmith/gau2grid.git )
 set( GAUXC_GAU2GRID_REVISION   v2.0.6 )
 
 set( GAUXC_INTEGRATORXX_REPOSITORY https://github.com/wavefunction91/IntegratorXX.git )
-set( GAUXC_INTEGRATORXX_REVISION   ea07dedd37e7bd49ea06394eb811599002b34b49 )
+set( GAUXC_INTEGRATORXX_REVISION   cf2917c64916583cef1081011beab3085b66e352 )
 
 set( GAUXC_HIGHFIVE_REPOSITORY https://github.com/BlueBrain/HighFive.git )
 set( GAUXC_HIGHFIVE_REVISION 805f0e13d09b47c4b01d40682621904aa3b31bb8 )

include/gauxc/enums.hpp

@@ -15,6 +15,7 @@ namespace GauXC {
  *  Generally mapped to equivalent enums in IntegratorXX
  */
 enum class RadialQuad {
+  Becke,             ///< Becke radial quadrature
   MuraKnowles,       ///< Mura-Knowles radial quadrature
   MurrayHandyLaming, ///< Murray-Handy-Laming radial quadrature
   TreutlerAldrichs   ///< Treutler-Aldrichs radial quadrature
@@ -38,6 +39,7 @@ enum class AtomicGridSizeDefault {
  *  molecular integration
  */
 enum class XCWeightAlg {
+  NOTPARTITIONED, ///< Not partitioned
   Becke, ///< The original Becke weighting scheme
   SSF,   ///< The Stratmann-Scuseria-Frisch weighting scheme
   LKO    ///< The Lauqua-Kuessman-Ochsenfeld weighting scheme

include/gauxc/load_balancer.hpp

@@ -15,6 +15,7 @@
 #include <gauxc/xc_task.hpp>
 #include <gauxc/util/timer.hpp>
 #include <gauxc/runtime_environment.hpp>
+#include <gauxc/enums.hpp>
 
 namespace GauXC {
 
@@ -27,6 +28,8 @@ namespace detail {
 struct LoadBalancerState {
   bool modified_weights_are_stored = false; 
     ///< Whether the load balancer currently stores partitioned weights
+  XCWeightAlg weight_alg = XCWeightAlg::NOTPARTITIONED; 
+    ///< Weight partitioning scheme used by this LoadBalancer
 };
 
 
@@ -77,6 +80,9 @@ class LoadBalancer {
   /// Return internal timing tracker
   const util::Timer& get_timings() const;
 
+  /// Return the total number of points for local tasks
+  size_t total_npts() const;
+
   /// Return the maximum number of points for local tasks
   size_t max_npts()       const;
 

include/gauxc/molgrid/defaults.hpp

@@ -14,6 +14,7 @@ namespace GauXC {
   double slater_radius_64(AtomicNumber);
   double slater_radius_30(AtomicNumber);
   double clementi_radius_67(AtomicNumber);
+  double uff_radius_103(AtomicNumber);
   double default_atomic_radius(AtomicNumber);
 
   RadialScale default_mk_radial_scaling_factor( AtomicNumber );

include/gauxc/runtime_environment/decl.hpp

@@ -77,6 +77,8 @@ class DeviceRuntimeEnvironment : public RuntimeEnvironment {
   bool owns_memory() const;
   DeviceBackend* device_backend() const;
 
+  void release_buffer();
+  void set_buffer(void* m, size_t sz);
 };
 #endif
 

include/gauxc/shell.hpp

@@ -225,7 +225,6 @@ class alignas(256) Shell {
 };
 
 
-#if 0
 template <typename T>
 inline std::ostream& operator<<( std::ostream& os, const Shell<T>& sh ) {
     os << "GauXC::Shell:( O={" 
@@ -234,9 +233,6 @@ inline std::ostream& operator<<( std::ostream& os, const Shell<T>& sh ) {
     os << "  ";
     os << " {l=" << sh.l() << ",sph=" << sh.pure() << "}";
     os << std::endl;
-    os << " {cr=" << sh.cutoff_radius() << ",cv=" << sh.cutoff_val() 
-	    <<",mr=" << sh.max_radius() << ",mv=" << sh.max_val() << "}";
-    os << std::endl;
 
     for(auto i=0ul; i<sh.nprim(); ++i) {
       os << "  " << sh.alpha()[i];
@@ -246,6 +242,5 @@ inline std::ostream& operator<<( std::ostream& os, const Shell<T>& sh ) {
 
     return os;
 }
-#endif
 
 }

include/gauxc/xc_integrator.hpp

@@ -36,6 +36,10 @@ class XCIntegrator {
   using exc_vxc_type_gks  = std::tuple< value_type, matrix_type, matrix_type, matrix_type, matrix_type >;
   using exc_grad_type = std::vector< value_type >;
   using exx_type      = matrix_type;
+  using fxc_contraction_type_rks = matrix_type;
+  using fxc_contraction_type_uks = std::tuple< matrix_type, matrix_type >;
+  using dd_psi_type   = std::vector< value_type >;
+  using dd_psi_potential_type   = matrix_type;
 
 private:
 
@@ -66,11 +70,19 @@ class XCIntegrator {
   exc_vxc_type_gks  eval_exc_vxc ( const MatrixType&, const MatrixType&, const MatrixType&, const MatrixType&,
                                    const IntegratorSettingsXC& = IntegratorSettingsXC{});
 
-  exc_grad_type eval_exc_grad( const MatrixType& );
+  exc_grad_type eval_exc_grad( const MatrixType&, const IntegratorSettingsXC& = IntegratorSettingsXC{} );
+  exc_grad_type eval_exc_grad( const MatrixType&, const MatrixType&, const IntegratorSettingsXC& = IntegratorSettingsXC{} );
 
   exx_type      eval_exx     ( const MatrixType&, 
                                const IntegratorSettingsEXX& = IntegratorSettingsEXX{} );
 
+  fxc_contraction_type_rks  eval_fxc_contraction ( const MatrixType&, const MatrixType&,
+                                  const IntegratorSettingsXC& = IntegratorSettingsXC{} );
+  fxc_contraction_type_uks  eval_fxc_contraction ( const MatrixType&, const MatrixType&, const MatrixType&, const MatrixType&,
+                                  const IntegratorSettingsXC& = IntegratorSettingsXC{} );
+
+  dd_psi_type eval_dd_psi( const MatrixType&, unsigned );
+  dd_psi_potential_type eval_dd_psi_potential( const MatrixType&, unsigned );
 
   const util::Timer& get_timings() const;
   const LoadBalancer& load_balancer() const;

include/gauxc/xc_integrator/impl.hpp

@@ -76,9 +76,16 @@ typename XCIntegrator<MatrixType>::exc_vxc_type_gks
 
 template <typename MatrixType>
 typename XCIntegrator<MatrixType>::exc_grad_type
-  XCIntegrator<MatrixType>::eval_exc_grad( const MatrixType& P ) {
+  XCIntegrator<MatrixType>::eval_exc_grad( const MatrixType& P, const IntegratorSettingsXC& ks_settings ) {
   if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
-  return pimpl_->eval_exc_grad(P);
+  return pimpl_->eval_exc_grad(P, ks_settings);
+};
+
+template <typename MatrixType>
+typename XCIntegrator<MatrixType>::exc_grad_type
+  XCIntegrator<MatrixType>::eval_exc_grad( const MatrixType& Ps, const MatrixType& Pz, const IntegratorSettingsXC& ks_settings ) {
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+  return pimpl_->eval_exc_grad(Ps, Pz, ks_settings);
 };
 
 template <typename MatrixType>
@@ -89,6 +96,37 @@ typename XCIntegrator<MatrixType>::exx_type
   return pimpl_->eval_exx(P,settings);
 };
 
+template <typename MatrixType>
+typename XCIntegrator<MatrixType>::fxc_contraction_type_rks
+  XCIntegrator<MatrixType>::eval_fxc_contraction( const MatrixType& P, const MatrixType& tP, 
+                                               const IntegratorSettingsXC& ks_settings ) { 
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+  return pimpl_->eval_fxc_contraction(P, tP, ks_settings);
+};
+
+template <typename MatrixType>
+typename XCIntegrator<MatrixType>::fxc_contraction_type_uks
+  XCIntegrator<MatrixType>::eval_fxc_contraction( const MatrixType& Ps, const MatrixType& Pz, 
+                           const MatrixType& tPs, const MatrixType& tPz, const IntegratorSettingsXC& ks_settings ) { 
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+  return pimpl_->eval_fxc_contraction(Ps, Pz, tPs, tPz, ks_settings);
+};
+
+template <typename MatrixType>
+typename XCIntegrator<MatrixType>::dd_psi_type
+  XCIntegrator<MatrixType>::eval_dd_psi(const MatrixType& P, unsigned max_Ylm) {
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+  return pimpl_->eval_dd_psi(P, max_Ylm);
+}
+
+template <typename MatrixType>
+typename XCIntegrator<MatrixType>::dd_psi_potential_type
+  XCIntegrator<MatrixType>::eval_dd_psi_potential(const MatrixType& X, unsigned max_Ylm) {
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+  return pimpl_->eval_dd_psi_potential(X, max_Ylm);
+}
+
+
 template <typename MatrixType>
 const util::Timer& XCIntegrator<MatrixType>::get_timings() const {
   if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();

include/gauxc/xc_integrator/replicated/impl.hpp

@@ -159,13 +159,27 @@ typename ReplicatedXCIntegrator<MatrixType>::exc_vxc_type_gks
 
 template <typename MatrixType>
 typename ReplicatedXCIntegrator<MatrixType>::exc_grad_type 
-  ReplicatedXCIntegrator<MatrixType>::eval_exc_grad_( const MatrixType& P ) {
+  ReplicatedXCIntegrator<MatrixType>::eval_exc_grad_( const MatrixType& P, const IntegratorSettingsXC& ks_settings ) {
 
   if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
 
   std::vector<value_type> EXC_GRAD( 3*pimpl_->load_balancer().molecule().natoms() );
   pimpl_->eval_exc_grad( P.rows(), P.cols(), P.data(), P.rows(),
-                         EXC_GRAD.data() );
+                         EXC_GRAD.data(), ks_settings );
+
+  return EXC_GRAD;
+
+}
+
+template <typename MatrixType>
+typename ReplicatedXCIntegrator<MatrixType>::exc_grad_type 
+  ReplicatedXCIntegrator<MatrixType>::eval_exc_grad_( const MatrixType& Ps, const MatrixType& Pz, const IntegratorSettingsXC& ks_settings ) {
+
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+
+  std::vector<value_type> EXC_GRAD( 3*pimpl_->load_balancer().molecule().natoms() );
+  pimpl_->eval_exc_grad( Ps.rows(), Ps.cols(), Ps.data(), Ps.rows(), Pz.data(), Pz.rows(),
+                         EXC_GRAD.data(), ks_settings );
 
   return EXC_GRAD;
 
@@ -184,6 +198,67 @@ typename ReplicatedXCIntegrator<MatrixType>::exx_type
 
   return K;
 
+}
+template <typename MatrixType>
+typename ReplicatedXCIntegrator<MatrixType>::fxc_contraction_type_rks
+  ReplicatedXCIntegrator<MatrixType>::eval_fxc_contraction_( const MatrixType& P, 
+    const MatrixType& tP, const IntegratorSettingsXC& ks_settings ) {
+
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+  matrix_type FXC( P.rows(), P.cols() );
+
+  pimpl_->eval_fxc_contraction( P.rows(), P.cols(), P.data(), P.rows(),
+                        tP.data(), tP.rows(),
+                        FXC.data(), FXC.rows(), ks_settings );
+
+  return FXC;
+}
+
+template <typename MatrixType>
+typename ReplicatedXCIntegrator<MatrixType>::fxc_contraction_type_uks
+  ReplicatedXCIntegrator<MatrixType>::eval_fxc_contraction_( const MatrixType& Ps, const MatrixType& Pz, 
+    const MatrixType& tPs, const MatrixType& tPz, const IntegratorSettingsXC& ks_settings ) {
+
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+  matrix_type FXCs( Ps.rows(), Ps.cols() );
+  matrix_type FXCz( Pz.rows(), Pz.cols() );
+
+  pimpl_->eval_fxc_contraction( Ps.rows(), Ps.cols(), Ps.data(), Ps.rows(),
+                        Pz.data(), Pz.rows(),
+                        tPs.data(), tPs.rows(),
+                        tPz.data(), tPz.rows(),
+                        FXCs.data(), FXCs.rows(),
+                        FXCz.data(), FXCz.rows(), ks_settings );
+
+  return std::make_tuple( FXCs, FXCz );
+
+}
+
+template <typename MatrixType>
+typename ReplicatedXCIntegrator<MatrixType>::dd_psi_type
+  ReplicatedXCIntegrator<MatrixType>::eval_dd_psi_( const MatrixType& P, unsigned max_Ylm ) {
+
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+
+  const size_t natoms = pimpl_->load_balancer().molecule().natoms();
+  const size_t Ylm_sz = (max_Ylm + 1) * ( max_Ylm + 1);
+  std::vector<value_type> ddPsi(natoms * Ylm_sz, 0.0);
+  pimpl_->eval_dd_psi(P.rows(), P.cols(), P.data(), P.rows(), max_Ylm, ddPsi.data(), Ylm_sz);
+  return ddPsi;
+}
+
+template <typename MatrixType>
+typename ReplicatedXCIntegrator<MatrixType>::dd_psi_potential_type
+  ReplicatedXCIntegrator<MatrixType>::eval_dd_psi_potential_( const MatrixType& X, unsigned max_Ylm ) {
+
+  if( not pimpl_ ) GAUXC_PIMPL_NOT_INITIALIZED();
+
+  const size_t nbf = pimpl_->load_balancer().basis().nbf();
+  matrix_type Vddx(nbf, nbf);
+  Vddx.setZero(); 
+  pimpl_->eval_dd_psi_potential(X.rows(), X.cols(), X.data(), max_Ylm, Vddx.data());
+  return Vddx;                      
+
 }
 
 }

include/gauxc/xc_integrator/replicated/replicated_xc_integrator_impl.hpp

@@ -74,11 +74,30 @@ class ReplicatedXCIntegratorImpl {
                               value_type* VXCx, int64_t ldvxcx,
                               value_type* EXC, const IntegratorSettingsXC& ks_settings ) = 0;
 
-  virtual void eval_exc_grad_( int64_t m, int64_t n, const value_type* P,
-                               int64_t ldp, value_type* EXC_GRAD ) = 0;
+  virtual void eval_exc_grad_( int64_t m, int64_t n, const value_type* P, int64_t ldp, 
+                               value_type* EXC_GRAD, const IntegratorSettingsXC& ks_settings ) = 0;
+  virtual void eval_exc_grad_( int64_t m, int64_t n, const value_type* P, int64_t ldps, 
+                               const value_type* Pz, int64_t lpdz, value_type* EXC_GRAD, const IntegratorSettingsXC& ks_settings ) = 0;
   virtual void eval_exx_( int64_t m, int64_t n, const value_type* P,
                           int64_t ldp, value_type* K, int64_t ldk,
                           const IntegratorSettingsEXX& settings ) = 0;
+  virtual void eval_fxc_contraction_( int64_t m, int64_t n, 
+                            const value_type* P, int64_t ldp,
+                            const value_type* tP, int64_t ldtp,
+                            value_type* FXC, int64_t ldfxc,
+                            const IntegratorSettingsXC& ks_settings )=0;
+  virtual void eval_fxc_contraction_( int64_t m, int64_t n, 
+                            const value_type* Ps, int64_t ldps,   
+                            const value_type* Pz, int64_t ldpz,
+                            const value_type* tPs, int64_t ldtps,
+                            const value_type* tPz, int64_t ldtpz,
+                            value_type* FXCs, int64_t ldfxcs,
+                            value_type* FXCz, int64_t ldfxcz,
+                            const IntegratorSettingsXC& ks_settings )=0;
+  virtual void eval_dd_psi_( int64_t m, int64_t n, const value_type* P, int64_t ldp, unsigned max_Ylm, 
+                             value_type* ddPsi, int64_t ldPsi ) = 0;
+  virtual void eval_dd_psi_potential_( int64_t m, int64_t n, const value_type* X, unsigned max_Ylm,
+                             value_type* Vddx) = 0;
 
 public:
 
@@ -130,13 +149,36 @@ class ReplicatedXCIntegratorImpl {
                      value_type* EXC, const IntegratorSettingsXC& ks_settings );
 
 
-  void eval_exc_grad( int64_t m, int64_t n, const value_type* P,
-                      int64_t ldp, value_type* EXC_GRAD );
+  void eval_exc_grad( int64_t m, int64_t n, const value_type* P, int64_t ldp, 
+                      value_type* EXC_GRAD, const IntegratorSettingsXC& ks_settings );
+  void eval_exc_grad( int64_t m, int64_t n, const value_type* Ps, int64_t ldps, 
+                      const value_type* Pz, int64_t ldpz, value_type* EXC_GRAD, const IntegratorSettingsXC& ks_settings );
 
   void eval_exx( int64_t m, int64_t n, const value_type* P,
                  int64_t ldp, value_type* K, int64_t ldk,
                  const IntegratorSettingsEXX& settings );
 
+  void eval_fxc_contraction( int64_t m, int64_t n, const value_type* P,
+                      int64_t ldp,
+                      const value_type* tP, int64_t ldtp,
+                      value_type* FXC, int64_t ldfxc,
+                      const IntegratorSettingsXC& ks_settings );
+
+  void eval_fxc_contraction( int64_t m, int64_t n, const value_type* Ps,
+                      int64_t ldps,
+                      const value_type* Pz, int64_t ldpz,
+                      const value_type* tPs, int64_t ldtps,
+                      const value_type* tPz, int64_t ldtpz,
+                      value_type* FXCs, int64_t ldfxcs,
+                      value_type* FXCz, int64_t ldfxcz,
+                      const IntegratorSettingsXC& ks_settings );
+
+  void eval_dd_psi( int64_t m, int64_t n, const value_type* P,
+                     int64_t ldp, unsigned max_Ylm, 
+                     value_type* ddPsi, int64_t ldPsi );
+  void eval_dd_psi_potential( int64_t m, int64_t n, const value_type* X, unsigned max_Ylm, 
+                      value_type* Vddx );
+
   inline const util::Timer& get_timings() const { return timer_; }
 
   inline std::unique_ptr< LocalWorkDriver > release_local_work_driver() {

include/gauxc/xc_integrator/replicated_xc_integrator.hpp

@@ -33,6 +33,10 @@ class ReplicatedXCIntegrator : public XCIntegratorImpl<MatrixType> {
   using exc_vxc_type_gks   = typename XCIntegratorImpl<MatrixType>::exc_vxc_type_gks;
   using exc_grad_type  = typename XCIntegratorImpl<MatrixType>::exc_grad_type;
   using exx_type       = typename XCIntegratorImpl<MatrixType>::exx_type;
+  using fxc_contraction_type_rks   = typename XCIntegratorImpl<MatrixType>::fxc_contraction_type_rks;
+  using fxc_contraction_type_uks   = typename XCIntegratorImpl<MatrixType>::fxc_contraction_type_uks;
+  using dd_psi_type       = typename XCIntegratorImpl<MatrixType>::dd_psi_type;
+  using dd_psi_potential_type       = typename XCIntegratorImpl<MatrixType>::dd_psi_potential_type;
 
 private:
 
@@ -46,8 +50,13 @@ class ReplicatedXCIntegrator : public XCIntegratorImpl<MatrixType> {
   exc_vxc_type_rks  eval_exc_vxc_ ( const MatrixType&, const IntegratorSettingsXC& ) override;
   exc_vxc_type_uks  eval_exc_vxc_ ( const MatrixType&, const MatrixType&, const IntegratorSettingsXC&) override;
   exc_vxc_type_gks  eval_exc_vxc_ ( const MatrixType&, const MatrixType&, const MatrixType&, const MatrixType&, const IntegratorSettingsXC& ) override;
-  exc_grad_type eval_exc_grad_( const MatrixType& ) override;
+  exc_grad_type eval_exc_grad_( const MatrixType&, const IntegratorSettingsXC& ) override;
+  exc_grad_type eval_exc_grad_( const MatrixType&, const MatrixType&, const IntegratorSettingsXC& ) override;
   exx_type      eval_exx_     ( const MatrixType&, const IntegratorSettingsEXX& ) override;
+  fxc_contraction_type_rks  eval_fxc_contraction_ ( const MatrixType&, const MatrixType&, const IntegratorSettingsXC& ) override;
+  fxc_contraction_type_uks  eval_fxc_contraction_ ( const MatrixType&, const MatrixType&, const MatrixType&, const MatrixType&, const IntegratorSettingsXC&) override;
+  dd_psi_type   eval_dd_psi_( const MatrixType& , unsigned ) override;
+  dd_psi_potential_type   eval_dd_psi_potential_( const MatrixType& , unsigned ) override;
   const util::Timer& get_timings_() const override;
   const LoadBalancer& get_load_balancer_() const override;
   LoadBalancer& get_load_balancer_() override;