-//! This function is documented in the header file
-void gpu_init_atomdata(NbnxmGpu* nb, const nbnxn_atomdata_t* nbat)
-{
- cl_int cl_error;
- int nalloc, natoms;
- bool realloced;
- bool bDoTime = nb->bDoTime;
- Nbnxm::GpuTimers* timers = nb->timers;
- NBAtomData* d_atdat = nb->atdat;
- const DeviceContext& deviceContext = *nb->deviceContext_;
- const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
-
- natoms = nbat->numAtoms();
- realloced = false;
-
- if (bDoTime)
- {
- /* time async copy */
- timers->atdat.openTimingRegion(localStream);
- }
-
- /* need to reallocate if we have to copy more atoms than the amount of space
- available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
- if (natoms > d_atdat->numAtomsAlloc)
- {
- nalloc = over_alloc_small(natoms);
-
- /* free up first if the arrays have already been initialized */
- if (d_atdat->numAtomsAlloc != -1)
- {
- freeDeviceBuffer(&d_atdat->f);
- freeDeviceBuffer(&d_atdat->xq);
- freeDeviceBuffer(&d_atdat->ljComb);
- freeDeviceBuffer(&d_atdat->atomTypes);
- }
-
-
- allocateDeviceBuffer(&d_atdat->f, nalloc, deviceContext);
- allocateDeviceBuffer(&d_atdat->xq, nalloc, deviceContext);
-
- if (useLjCombRule(nb->nbparam->vdwType))
- {
- // Two Lennard-Jones parameters per atom
- allocateDeviceBuffer(&d_atdat->ljComb, nalloc, deviceContext);
- }
- else
- {
- allocateDeviceBuffer(&d_atdat->atomTypes, nalloc, deviceContext);
- }
-
- d_atdat->numAtomsAlloc = nalloc;
- realloced = true;
- }
-
- d_atdat->numAtoms = natoms;
- d_atdat->numAtomsLocal = nbat->natoms_local;
-
- /* need to clear GPU f output if realloc happened */
- if (realloced)
- {
- nbnxn_ocl_clear_f(nb, nalloc);
- }
-
- if (useLjCombRule(nb->nbparam->vdwType))
- {
- static_assert(
- sizeof(Float2) == 2 * sizeof(*nbat->params().lj_comb.data()),
- "Size of a pair of LJ parameters elements should be equal to the size of Float2.");
- copyToDeviceBuffer(&d_atdat->ljComb,
- reinterpret_cast<const Float2*>(nbat->params().lj_comb.data()),
- 0,
- natoms,
- localStream,
- GpuApiCallBehavior::Async,
- bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
- }
- else
- {
- static_assert(sizeof(int) == sizeof(*nbat->params().type.data()),
- "Sizes of host- and device-side atom types should be the same.");
- copyToDeviceBuffer(&d_atdat->atomTypes,
- nbat->params().type.data(),
- 0,
- natoms,
- localStream,
- GpuApiCallBehavior::Async,
- bDoTime ? timers->atdat.fetchNextEvent() : nullptr);
- }
-
- if (bDoTime)
- {
- timers->atdat.closeTimingRegion(localStream);
- }
-
- /* kick off the tasks enqueued above to ensure concurrency with the search */
- cl_error = clFlush(localStream.stream());
- GMX_RELEASE_ASSERT(cl_error == CL_SUCCESS,
- ("clFlush failed: " + ocl_get_error_string(cl_error)).c_str());
-}
-