#include "types/nbnxn_cuda_types_ext.h"
#include "gpu_utils.h"
-#include "nbnxn_cuda_data_mgmt.h"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.h"
#include "pmalloc_cuda.h"
+#include "nb_verlet.h"
t_forcerec *mk_forcerec(void)
{
fr->excl_load[t] = i;
}
}
+
+/* Frees GPU memory and destroys the CUDA context.
+ *
+ * Note that this function needs to be called even if GPUs are not used
+ * in this run because the PME ranks have no knowledge of whether GPUs
+ * are used or not, but all ranks need to enter the barrier below.
+ */
+void free_gpu_resources(const t_forcerec *fr,
+ const t_commrec *cr)
+{
+ gmx_bool bIsPPrankUsingGPU;
+ char gpu_err_str[STRLEN];
+
+ bIsPPrankUsingGPU = (cr->duty & DUTY_PP) && fr && fr->nbv && fr->nbv->bUseGPU;
+
+ if (bIsPPrankUsingGPU)
+ {
+ /* free nbnxn data in GPU memory */
+ nbnxn_cuda_free(fr->nbv->cu_nbv);
+
+ /* With tMPI we need to wait for all ranks to finish deallocation before
+ * destroying the context in free_gpu() as some ranks may be sharing
+ * GPU and context.
+ * Note: as only PP ranks need to free GPU resources, so it is safe to
+ * not call the barrier on PME ranks.
+ */
+#ifdef GMX_THREAD_MPI
+ if (PAR(cr))
+ {
+ gmx_barrier(cr);
+ }
+#endif /* GMX_THREAD_MPI */
+
+ /* uninitialize GPU (by destroying the context) */
+ if (!free_gpu(gpu_err_str))
+ {
+ gmx_warning("On rank %d failed to free GPU #%d: %s",
+ cr->nodeid, get_current_gpu_device_id(), gpu_err_str);
+ }
+ }
+}