|
Gromacs
2027.0-dev-20261113-fb41eee
|
#include <gromacs/nbnxm/cuda/nbnxm_cuda_types.h>
Collaboration diagram for gmx::NbnxmGpu:Main data structure for CUDA nonbonded force calculations.
Public Member Functions | |
| NbnxmGpu (const DeviceStreamManager &deviceStreamManager, std::optional< size_t > nLambda) | |
Public Attributes | |
| const DeviceContext & | deviceContext |
| GPU device context. | |
| const HostAllocationPolicy | hostAllocationPolicy |
| Host allocation policy for buffers for possible GPU transfers. | |
| bool | bUseTwoStreams = false |
| true if doing both local/non-local NB work on GPU | |
| bool | bNonLocalStreamDoneMarked = false |
| true indicates that the nonlocal_done event was marked | |
| NBAtomDataGpu * | atdat = nullptr |
| atom data | |
| int * | atomIndices = nullptr |
| array of atom indices | |
| int | atomIndicesSize = 0 |
| size of atom indices | |
| int | atomIndicesSize_alloc = 0 |
| size of atom indices allocated in device buffer | |
| int * | numAtomsPerCell = nullptr |
| x buf ops num of atoms | |
| int | numAtomsPerCellSize = 0 |
| number of elements in numAtomsPerCell | |
| int | numAtomsPerCellAlloc = 0 |
| number of elements allocated in device buffer | |
| int * | cellToBin = nullptr |
| x buf ops bin index mapping | |
| int | cellToBinSize = 0 |
| number of elements in cellToBin | |
| int | cellToBinAlloc = 0 |
| number of elements allocated in device buffer | |
| NBParamGpu * | nbparam = nullptr |
| parameters required for the non-bonded calc. | |
|
EnumerationArray < InteractionLocality, std::unique_ptr< GpuPairlist > > | plist = { { nullptr } } |
| pair-list data structures (local and non-local) | |
|
EnumerationArray < InteractionLocality, std::unique_ptr< GpuFeplist > > | feplist = { { nullptr } } |
| fep-list data structures (local and non-local) | |
| GpuFepHostData * | fephostdata = nullptr |
| host buffers required for the FEP H2D copies | |
| NBStagingData | nbst |
| staging area where fshift/energies get downloaded | |
|
EnumerationArray < InteractionLocality, const DeviceStream * > | deviceStreams = { { nullptr } } |
| local and non-local GPU streams | |
| GpuEventSynchronizer | nonlocal_done |
| Event triggered when the non-local non-bonded kernel is done (and the local transfer can proceed) | |
| GpuEventSynchronizer | misc_ops_and_local_H2D_done |
| Event triggered when the tasks issued in the local stream that need to precede the non-local force or buffer operation calculations are done (e.g. f buffer 0-ing, local x/q H2D, buffer op initialization in local stream that is required also by nonlocal stream ) | |
| EnumerationArray < InteractionLocality, bool > | haveWork = { { false } } |
| True if there is work for the current domain in the respective locality. More... | |
| bool | bDoTime = false |
| True if event-based timing is enabled. | |
| GpuTimers * | timers = nullptr |
| CUDA event-based timers. | |
|
std::unique_ptr < gmx_wallclock_gpu_nbnxm_t > | timings |
| Timing data. TODO: deprecate this and query timers for accumulated data instead. | |
|
EnumerationArray < InteractionLocality, FusedXToXqLaunchParams > | xToXqLaunchParams |
| Kernel launch parameters per interaction locality, computed once at pair-list setup. | |
| EnumerationArray<InteractionLocality, bool> gmx::NbnxmGpu::haveWork = { { false } } |
True if there is work for the current domain in the respective locality.
This includes local/nonlocal GPU work, either bonded or nonbonded, scheduled to be executed in the current domain. As long as bonded work is not split up into local/nonlocal, if there is bonded GPU work, both flags will be true.
1.8.5