Gromacs  2021.3
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
Enumerations | Functions
nbnxm_gpu.h File Reference
#include "gromacs/gpu_utils/gpu_macros.h"
#include "gromacs/math/vectypes.h"
#include "gromacs/mdtypes/locality.h"
#include "gromacs/nbnxm/atomdata.h"
#include "gromacs/utility/basedefinitions.h"
#include "gromacs/utility/real.h"
+ Include dependency graph for nbnxm_gpu.h:
+ This graph shows which files directly or indirectly include this file:

Description

Declare interface for GPU execution for NBNXN module.

Author
Szilard Pall pall..nosp@m.szil.nosp@m.ard@g.nosp@m.mail.nosp@m..com
Mark Abraham mark..nosp@m.j.ab.nosp@m.raham.nosp@m.@gma.nosp@m.il.co.nosp@m.m

Enumerations

enum  eelType : int {
  eelTypeCUT, eelTypeRF, eelTypeEWALD_TAB, eelTypeEWALD_TAB_TWIN,
  eelTypeEWALD_ANA, eelTypeEWALD_ANA_TWIN, eelTypeNR
}
 Nbnxm electrostatic GPU kernel flavors. More...
 
enum  evdwType : int {
  evdwTypeCUT, evdwTypeCUTCOMBGEOM, evdwTypeCUTCOMBLB, evdwTypeFSWITCH,
  evdwTypePSWITCH, evdwTypeEWALDGEOM, evdwTypeEWALDLB, evdwTypeNR
}
 Nbnxm VdW GPU kernel flavors. More...
 

Functions

static bool Nbnxm::useLjCombRule (const int vdwType)
 Returns true if LJ combination rules are used in the non-bonded kernels. More...
 
void Nbnxm::gpu_copy_xq_to_gpu (NbnxmGpu *nb, const struct nbnxn_atomdata_t *nbdata, gmx::AtomLocality aloc)
 Launch asynchronously the xq buffer host to device copy. More...
 
void Nbnxm::gpu_launch_kernel (NbnxmGpu *nb, const gmx::StepWorkload &stepWork, gmx::InteractionLocality iloc)
 Launch asynchronously the nonbonded force calculations. More...
 
void Nbnxm::gpu_launch_kernel_pruneonly (NbnxmGpu *nb, gmx::InteractionLocality iloc, int numParts)
 Launch asynchronously the nonbonded prune-only kernel. More...
 
void Nbnxm::gpu_launch_cpyback (NbnxmGpu *nb, nbnxn_atomdata_t *nbatom, const gmx::StepWorkload &stepWork, gmx::AtomLocality aloc)
 Launch asynchronously the download of short-range forces from the GPU (and energies/shift forces if required).
 
bool Nbnxm::gpu_try_finish_task (NbnxmGpu *nb, const gmx::StepWorkload &stepWork, gmx::AtomLocality aloc, real *e_lj, real *e_el, gmx::ArrayRef< gmx::RVec > shiftForces, GpuTaskCompletion completionKind, gmx_wallcycle *wcycle)
 Attempts to complete nonbonded GPU task. More...
 
float Nbnxm::gpu_wait_finish_task (NbnxmGpu *nb, const gmx::StepWorkload &stepWork, gmx::AtomLocality aloc, real *e_lj, real *e_el, gmx::ArrayRef< gmx::RVec > shiftForces, gmx_wallcycle *wcycle)
 Completes the nonbonded GPU task blocking until GPU tasks and data transfers to finish. More...
 
void Nbnxm::nbnxn_gpu_init_x_to_nbat_x (const Nbnxm::GridSet &gridSet, NbnxmGpu *gpu_nbv)
 Initialization for X buffer operations on GPU. Called on the NS step and performs (re-)allocations and memory copies. !
 
void Nbnxm::nbnxn_gpu_x_to_nbat_x (const Nbnxm::Grid &grid, bool setFillerCoords, NbnxmGpu *gpu_nbv, DeviceBuffer< gmx::RVec > d_x, GpuEventSynchronizer *xReadyOnDevice, gmx::AtomLocality locality, int gridId, int numColumnsMax)
 X buffer operations on GPU: performs conversion from rvec to nb format. More...
 
void Nbnxm::nbnxnInsertNonlocalGpuDependency (const NbnxmGpu *nb, gmx::InteractionLocality interactionLocality)
 Sync the nonlocal stream with dependent tasks in the local queue. More...
 
void Nbnxm::setupGpuShortRangeWork (NbnxmGpu *nb, const gmx::GpuBonded *gpuBonded, gmx::InteractionLocality iLocality)
 Set up internal flags that indicate what type of short-range work there is. More...
 
bool Nbnxm::haveGpuShortRangeWork (const NbnxmGpu *nb, gmx::AtomLocality aLocality)
 Returns true if there is GPU short-range work for the given atom locality. More...
 
void Nbnxm::nbnxn_wait_x_on_device (NbnxmGpu *nb)
 sync CPU thread on coordinate copy to device More...
 
void * Nbnxm::getGpuForces (NbnxmGpu *nb)
 Get the pointer to the GPU nonbonded force buffer. More...
 

Enumeration Type Documentation

enum eelType : int

Nbnxm electrostatic GPU kernel flavors.

Types of electrostatics implementations available in the GPU non-bonded force kernels. These represent both the electrostatics types implemented by the kernels (cut-off, RF, and Ewald - a subset of what's defined in enums.h) as well as encode implementation details analytical/tabulated and single or twin cut-off (for Ewald kernels). Note that the cut-off and RF kernels have only analytical flavor and unlike in the CPU kernels, the tabulated kernels are ATM Ewald-only.

The row-order of pointers to different electrostatic kernels defined in nbnxn_cuda.cu by the nb_*_kfunc_ptr function pointer table should match the order of enumerated types below.

enum evdwType : int

Nbnxm VdW GPU kernel flavors.

The enumerates values correspond to the LJ implementations in the GPU non-bonded kernels.

The column-order of pointers to different electrostatic kernels defined in nbnxn_cuda_ocl.cpp/.cu by the nb_*_kfunc_ptr function pointer table should match the order of enumerated types below.