|
MFC
Exascale flow solver
|
Per-block fine-grid immersed-boundary state for AMR. A separate module from m_ibm: declaring these derived-type allocatables inside m_ibm corrupts CCE's OpenMP declare-target descriptor table for m_ibm's ghost_points (lib-4425 on its @:ALLOCATE, even for non-AMR runs). Do not fold them back into m_ibm. More...
Data Types | |
| type | ib_fine_state |
| Per-block fine IB state: each AMR slot keeps a host-side copy of its markers, computed at fine resolution and copied into m_ibm's declare-target ib_markers for the fine advance (that array is never reallocated or pointer-swapped, since that corrupts the Cray present table; the swap syncs via GPU_UPDATE). Ghost-point lists park on-device in gp_park; num_gps is the slot's ghost-point count across a swap, off its gp_park offset and cap the entries reserved there (0 = none). More... | |
Variables | |
| type(ib_fine_state), dimension(:), allocatable | ib_fine |
| integer | num_gps_save |
| type(ghost_point), dimension(:), allocatable | gp_park |
| Device-resident park for the coarse and fine ghost-point lists across a fine swap, copied to/from ghost_points by on-device kernels. One packed pool: entries 1..park_coarse hold the coarse list; each slot holding ghost points (a block containing a body) gets a segment after it, so the pool scales with the ghost points parked, not blocks x the longest list. Mapped via move_alloc + GPU_ENTER_DATA (the amr_cg idiom; a direct @:ALLOCATE aborts with lib-4425 on CCE OpenMP). | |
| integer | ib_coarse_slot = 0 |
| The coarse ghost points and markers park (host copies) in an extra ib_fine slot, ib_coarse_slot (= nslots+1): a new module-level derived-type allocatable here would corrupt a sibling's descriptor on CCE OpenMP (lib-4425). | |
| integer(kind=8) | park_coarse = 0_8 |
| gp_park entries reserved for the coarse list | |
| integer(kind=8) | park_top = 0_8 |
| gp_park entries in use (coarse segment + fine segments) | |
| integer(kind=8) | fine_gps_cap = 0_8 |
| Fine-block ghost-point capacity (= buffered fine-block cell count), set by s_ibm_alloc_fine; 0 when AMR-IB is inactive. | |
| integer, dimension(3) | mkr_lo = 0 |
| Bounds of ib_markers, enclosing both coarse and fine blocks so it never needs reallocating. Set by s_ibm_alloc_fine. | |
| integer, dimension(3) | mkr_hi = 0 |
Per-block fine-grid immersed-boundary state for AMR. A separate module from m_ibm: declaring these derived-type allocatables inside m_ibm corrupts CCE's OpenMP declare-target descriptor table for m_ibm's ghost_points (lib-4425 on its @:ALLOCATE, even for non-AMR runs). Do not fold them back into m_ibm.
| integer(kind=8) m_ibm_fine::fine_gps_cap = 0_8 |
Fine-block ghost-point capacity (= buffered fine-block cell count), set by s_ibm_alloc_fine; 0 when AMR-IB is inactive.
Definition at line 403 of file m_ibm.fpp.f90.
| type(ghost_point), dimension(:), allocatable m_ibm_fine::gp_park |
Device-resident park for the coarse and fine ghost-point lists across a fine swap, copied to/from ghost_points by on-device kernels. One packed pool: entries 1..park_coarse hold the coarse list; each slot holding ghost points (a block containing a body) gets a segment after it, so the pool scales with the ghost points parked, not blocks x the longest list. Mapped via move_alloc + GPU_ENTER_DATA (the amr_cg idiom; a direct @:ALLOCATE aborts with lib-4425 on CCE OpenMP).
Definition at line 394 of file m_ibm.fpp.f90.
| integer m_ibm_fine::ib_coarse_slot = 0 |
The coarse ghost points and markers park (host copies) in an extra ib_fine slot, ib_coarse_slot (= nslots+1): a new module-level derived-type allocatable here would corrupt a sibling's descriptor on CCE OpenMP (lib-4425).
Definition at line 398 of file m_ibm.fpp.f90.
| type(ib_fine_state), dimension(:), allocatable m_ibm_fine::ib_fine |
Definition at line 387 of file m_ibm.fpp.f90.
| integer, dimension(3) m_ibm_fine::mkr_hi = 0 |
Definition at line 406 of file m_ibm.fpp.f90.
| integer, dimension(3) m_ibm_fine::mkr_lo = 0 |
Bounds of ib_markers, enclosing both coarse and fine blocks so it never needs reallocating. Set by s_ibm_alloc_fine.
Definition at line 406 of file m_ibm.fpp.f90.
| integer m_ibm_fine::num_gps_save |
Definition at line 388 of file m_ibm.fpp.f90.
| integer(kind=8) m_ibm_fine::park_coarse = 0_8 |
gp_park entries reserved for the coarse list
Definition at line 399 of file m_ibm.fpp.f90.
| integer(kind=8) m_ibm_fine::park_top = 0_8 |
gp_park entries in use (coarse segment + fine segments)
Definition at line 400 of file m_ibm.fpp.f90.