core_kernel_manager.h

// -*-c++-*-
#ifndef STORMM_CORE_KERNEL_MANAGER_H
#define STORMM_CORE_KERNEL_MANAGER_H
 
#include <map>
#include <string>
#ifdef STORMM_USE_HPC
#  ifdef STORMM_USE_CUDA
#    include <cuda_runtime.h>
#  endif
#endif
#include "copyright.h"
#include "Constants/behavior.h"
#include "DataTypes/stormm_vector_types.h"
#include "Math/reduction.h"
#include "Numerics/split_fixed_precision.h"
#include "Potential/energy_enumerators.h"
#include "Structure/structure_enumerators.h"
#include "Synthesis/atomgraph_synthesis.h"
#include "Synthesis/synthesis_enumerators.h"
#include "Topology/atomgraph_enumerators.h"
#include "Trajectory/trajectory_enumerators.h"
#include "gpu_details.h"
#include "kernel_format.h"
 
namespace stormm {
namespace card {
 
using constants::PrecisionModel;
using energy::ClashResponse;
using energy::EvaluateForce;
using energy::EvaluateEnergy;
using energy::getEnumerationName;
using energy::NeighborListKind;
using energy::QMapMethod;
using energy::TinyBoxPresence;
using energy::ValenceKernelSize;
using numerics::AccumulationMethod;
using structure::GridDetail;
using structure::RMSDTask;
using structure::VirtualSiteActivity;
using synthesis::AtomGraphSynthesis;
using synthesis::NbwuKind;
using synthesis::VwuGoal;
using stmath::ReductionGoal;
using stmath::ReductionStage;
using synthesis::VwuGoal;
using topology::ImplicitSolventModel;
using topology::UnitCellType;
using trajectory::IntegrationStage;
  
class CoreKlManager : public KernelManager {
public:

  CoreKlManager(const GpuDetails &gpu_in, const AtomGraphSynthesis *poly_ag);
  CoreKlManager(const GpuDetails &gpu_in, const AtomGraphSynthesis &poly_ag);

  int getArchBlockMultiplier() const;
  
  int2 getValenceKernelDims(PrecisionModel prec, EvaluateForce eval_force, EvaluateEnergy eval_nrg,
                            AccumulationMethod acc_meth, VwuGoal purpose,
                            ClashResponse collision_handling) const;

  int2 getPmeValenceKernelDims(PrecisionModel prec, PrecisionModel neighbor_prec,
                               NeighborListKind neighbor_layout, EvaluateEnergy eval_nrg,
                               AccumulationMethod acc_meth,
                               ClashResponse collision_handling) const;
  
  int2 getIntegrationKernelDims(PrecisionModel prec, AccumulationMethod acc_meth,
                                IntegrationStage process) const;

  int2 getNonbondedKernelDims(PrecisionModel prec, NbwuKind kind, EvaluateForce eval_force,
                              EvaluateEnergy eval_nrg, AccumulationMethod acc_meth,
                              ImplicitSolventModel igb, ClashResponse collision_handling) const;

  int2 getBornRadiiKernelDims(PrecisionModel prec, NbwuKind kind, AccumulationMethod acc_meth,
                              ImplicitSolventModel igb) const;

  int2 getBornDerivativeKernelDims(PrecisionModel prec, NbwuKind kind, AccumulationMethod acc_meth,
                                   ImplicitSolventModel igb) const;

  int2 getDensityMappingKernelDims(QMapMethod approach, PrecisionModel prec, size_t cg_tmat,
                                   int order) const;
 
  int2 getDensityMappingKernelDims(QMapMethod approach, PrecisionModel calc_prec,
                                   PrecisionModel acc_prec, bool overflow, size_t cg_tmat,
                                   int order) const;

  int2 getReductionKernelDims(PrecisionModel prec, ReductionGoal purpose,
                              ReductionStage process) const;

  int2 getVirtualSiteKernelDims(PrecisionModel prec, VirtualSiteActivity purpose) const;

  int2 getRMSDKernelDims(PrecisionModel prec, RMSDTask order) const;

  int2 getPMEPairsKernelDims(PrecisionModel coord_prec, PrecisionModel calc_prec,
                             NeighborListKind grid_configuration, TinyBoxPresence has_tiny_box,
                             EvaluateForce eval_frc, EvaluateEnergy eval_nrg,
                             ClashResponse mitigation) const;

  int2 getMigrationKernelDims(PrecisionModel coord_prec, NeighborListKind grid_configuration,
                              int stage_number, int gpos_bits = 0, int chain_count = 0) const;
 
private:

  ValenceKernelSize valence_kernel_width;

  int nonbond_block_multiplier_dp;
  int nonbond_block_multiplier_sp;
  
  int gbradii_block_multiplier_dp;
  int gbradii_block_multiplier_sp;

  int gbderiv_block_multiplier_dp;
  int gbderiv_block_multiplier_sp;
  
  std::vector<int> gen_qmap_block_multiplier_dp;
  std::vector<int> gen_qmap_block_multiplier_sp;
  std::vector<int> sac_qmap_block_multiplier_dp;
  std::vector<int> sac_qmap_block_multiplier_sp;
  
  int reduction_block_multiplier;
  
  ValenceKernelSize virtual_site_kernel_width;

  int rmsd_block_multiplier_dp;
  int rmsd_block_multiplier_sp;

  void catalogValenceKernel(PrecisionModel prec, EvaluateForce eval_force, EvaluateEnergy eval_nrg,
                            AccumulationMethod acc_meth, VwuGoal purpose,
                            ClashResponse collision_handling, ValenceKernelSize kwidth,
                            const std::string &kernel_name = std::string(""));
 
  void catalogValenceKernel(PrecisionModel prec, PrecisionModel neighbor_prec,
                            NeighborListKind neighbor_layout, EvaluateEnergy eval_nrg,
                            AccumulationMethod acc_meth, ClashResponse collision_handling,
                            const std::string &kernel_name = std::string(""));

  void catalogIntegrationKernel(PrecisionModel prec, AccumulationMethod acc_meth,
                                ValenceKernelSize kwidth, IntegrationStage process,
                                const std::string &kernel_name = std::string(""));
  
  void catalogNonbondedKernel(PrecisionModel prec, NbwuKind kind, EvaluateForce eval_force,
                              EvaluateEnergy eval_nrg, AccumulationMethod acc_meth,
                              ImplicitSolventModel igb, ClashResponse collision_handling,
                              const std::string &kernel_name = std::string(""));

  void catalogBornRadiiKernel(PrecisionModel prec, NbwuKind kind, AccumulationMethod acc_meth,
                              ImplicitSolventModel igb,
                              const std::string &kernel_name = std::string(""));

  void catalogBornDerivativeKernel(PrecisionModel prec, NbwuKind kind, AccumulationMethod acc_meth,
                                   ImplicitSolventModel igb,
                                   const std::string &kernel_name = std::string(""));

  void catalogGeneralQMapKernel(PrecisionModel prec, size_t cg_tmat, int order,
                                const std::string &kernel_name);

  void catalogShrAccQMapKernel(PrecisionModel calc_prec, PrecisionModel acc_prec,
                               bool overflow, size_t cg_tmat, int order,
                               const std::string &kernel_name);

  void catalogReductionKernel(PrecisionModel prec, ReductionGoal purpose, ReductionStage process,
                              int subdivision, const std::string &kernel_name = std::string(""));

  void catalogVirtualSiteKernel(PrecisionModel prec, VirtualSiteActivity purpose,
                                ValenceKernelSize kwidth,
                                const std::string &kernel_name = std::string(""));

  void catalogRMSDKernel(PrecisionModel prec, RMSDTask order,
                         const std::string &kernel_name = std::string(""));

  void catalogPMEPairsKernel(PrecisionModel coord_prec, PrecisionModel calc_prec,
                             NeighborListKind grid_configuration, TinyBoxPresence has_tiny_box,
                             EvaluateForce eval_frc, EvaluateEnergy eval_nrg,
                             ClashResponse mitigation,
                             const std::string &kernel_name = std::string(""));

  void catalogMigrationKernel(PrecisionModel coord_prec, NeighborListKind grid_configuration,
                              int stage_number, int gpos_bits,
                              const std::string &kernel_name = std::string(""));
};

int nonbondedBlockMultiplier(const GpuDetails &gpu, UnitCellType unit_cell, PrecisionModel prec,
                             ImplicitSolventModel igb);

int gbRadiiBlockMultiplier(const GpuDetails &gpu, PrecisionModel prec);

int gbDerivativeBlockMultiplier(const GpuDetails &gpu, PrecisionModel prec);

std::vector<int> densityMappingBlockMultiplier(const GpuDetails &gpu, PrecisionModel prec,
                                               QMapMethod approach);
  
int reductionBlockMultiplier();

int virtualSiteBlockMultiplier(PrecisionModel prec);

int rmsdBlockMultiplier(PrecisionModel prec);

int pmePairsBlockMultiplier(const GpuDetails &gpu, PrecisionModel coord_prec,
                            PrecisionModel calc_prec);

int migrationBlockMultiplier();

std::string valenceKernelWidthExtension(PrecisionModel prec, ValenceKernelSize kwidth);
  
std::string valenceKernelKey(PrecisionModel prec, EvaluateForce eval_force,
                             EvaluateEnergy eval_nrg, AccumulationMethod acc_meth,
                             VwuGoal purpose, ClashResponse collision_handling,
                             ValenceKernelSize kwidth);
 
std::string valenceKernelKey(PrecisionModel prec, PrecisionModel neighbor_prec,
                             NeighborListKind neighbor_layout, EvaluateEnergy eval_nrg,
                             AccumulationMethod acc_meth, ClashResponse collision_handling);

std::string integrationKernelKey(PrecisionModel prec, AccumulationMethod acc_meth,
                                 ValenceKernelSize kwidth, IntegrationStage process);
  
std::string nonbondedKernelKey(PrecisionModel prec, NbwuKind kind, EvaluateForce eval_force,
                               EvaluateEnergy eval_nrg, AccumulationMethod acc_meth,
                               ImplicitSolventModel igb, ClashResponse collision_handling);

std::string appendBornKernelKey(PrecisionModel prec, NbwuKind kind, AccumulationMethod acc_meth,
                                ImplicitSolventModel igb);
  
std::string bornRadiiKernelKey(PrecisionModel prec, NbwuKind kind, AccumulationMethod acc_meth,
                               ImplicitSolventModel igb);

std::string bornDerivativeKernelKey(PrecisionModel prec, NbwuKind kind,
                                    AccumulationMethod acc_meth, ImplicitSolventModel igb);

std::string appendQMapKernelKey(PrecisionModel prec, size_t cg_tmat, int order);

std::string generalQMapKernelKey(PrecisionModel prec, size_t cg_tmat, int order);

std::string shrAccQMapKernelKey(PrecisionModel calc_prec, PrecisionModel acc_prec,
                                bool overflow, size_t cg_tmat,  int order);

std::string reductionKernelKey(PrecisionModel prec, ReductionGoal purpose, ReductionStage process);

std::string virtualSiteKernelKey(PrecisionModel prec, VirtualSiteActivity process,
                                 ValenceKernelSize kwidth);

std::string rmsdKernelKey(PrecisionModel prec, RMSDTask order);

std::string pmePairsKernelKey(PrecisionModel coord_prec, PrecisionModel calc_prec,
                              NeighborListKind grid_configuration, TinyBoxPresence has_tiny_box,
                              EvaluateForce eval_frc, EvaluateEnergy eval_nrg,
                              ClashResponse mitigation);

std::string migrationKernelKey(PrecisionModel coord_prec, NeighborListKind grid_configuration,
                               int stage_number, int gpos_bits);
 
} // namespace card
} // namespace stormm
 
#endif