|
MFC
Exascale flow solver
|
Flat per-block field store: dense slot indexing, reservation, alloc/free and the prim/bridge loaders. More...
Functions/Subroutines | |
| impure subroutine, public | s_amr_sync_grid_state_to_device () |
| Push the host global grid state (m/n/p, idwint/idwbuff, coordinate arrays) to its GPU_DECLARE'd device copies after a swap/restore. Coordinates go through amr_gs_buf as one update since each allocatable update costs several synchronous copies. No-op on CPU. | |
| impure subroutine, public | s_amr_copy_fine_fields (k) |
| Device copy amr_cons_st -> amr_stor_st over block k's buffered box, all sys_size fields (RK step-entry backup). | |
| impure subroutine, public | s_amr_alloc_pool () |
| Block-metadata pool for amr_max_blocks slots (geometry, ownership, level, SFC cuts, overlap-list counts; the 2D rank lists are sized in s_amr_build_seam_pairs) and the dense-index maps. | |
| impure subroutine, public | s_amr_init_swap_buffers () |
| Bounce buffers for the copy-based coordinate swap; shaped like the base-level *_alloc arrays, since they are assigned whole to/from x_cb etc. | |
| impure subroutine, public | s_amr_free_pool () |
| Free what s_amr_alloc_pool allocated (plus the per-block caches sized later), and the coordinate swap buffers. | |
| impure subroutine, public | s_amr_set_mbuf () |
| The buffered per-slot array bounds from the max fine extents (collapsed dims 0:0). | |
| impure subroutine, public | s_amr_free_swap_buffers () |
| impure subroutine | s_amr_loc_index_init () |
| Allocate/reset the dense local-index maps (idempotent). | |
| impure subroutine | s_amr_st_move_slot (src, dst) |
| Move one local slot's store data src -> dst on the device, in place (a staging copy would transiently double the store). Ascending-source order (s_amr_compact_store) guarantees dst's old contents are consumed. The host copy goes stale, the store's normal state between rebuilds. | |
| impure subroutine | s_amr_compact_store () |
| Re-densify the local index space after every reconcile so amr_loc_n tracks the live count instead of ratcheting up with SFC window shifts and migration slots (capacity then plateaus at the rebuild transient). Moves are device-side and in place in ascending old-index order: each destination is <= its source and below every pending source. Called at the end of s_amr_reconcile_slots, after the rebuild's overlap carry-forward (which reads amr_stor_st at old indices) has finished. | |
| impure subroutine, public | s_amr_dev_mem (free_b, total_b) |
| Free and total device memory in bytes (hipMemGetInfo on the AMD GPU build); -1 for both where there is no query. | |
| impure subroutine | s_amr_st_reserve (nloc) |
| Size the flat store for at least nloc local slots, growing geometrically and never shrinking. Growth preserves the live columns on the device (no PCIe round trip); the host mirror comes out undefined. | |
| impure subroutine, public | s_amr_br_load (loc) |
| Move block loc's conserved state between the flat store and the bridge over the whole buffered box (all slots share the mbuf extents). | |
| impure subroutine, public | s_amr_br_store (loc) |
| impure subroutine, public | s_amr_br_load_faces (loc, olo, ohi, glo, ghi, woff, w_lo, w_hi) |
| Face-bounded bridge move for the level>=2 reflux apply: moves only the six one-cell planes (outside cell olo(d)/ohi(d) of each active dim, over the caller's transverse window) that s_amr_reflux_apply_faces read-modify-writes, skipping zero-weight planes as the apply does. Bit-identical to a whole-box move. | |
| impure subroutine, public | s_amr_br_store_faces (loc, olo, ohi, glo, ghi, woff, w_lo, w_hi) |
| impure subroutine, public | s_amr_br_load_batch (nb) |
| Batched bridge load: every batch member (amr_bat_blk/amr_bat_loc, extents amr_bat_ext) lands at offset (ibm-1)*amr_bat_w along amr_bat_sd with its ghost shell, in one kernel. Nothing is stored back: s_compute_rhs does not write its cons dummy on the fine path. | |
| impure subroutine, public | s_amr_bat_member_prim (ibm, src, dst) |
| Copy batch member ibm's primitive state from the slab scratch into the block-frame scratch over its buffered extent, for the per-member IB correction. | |
| impure subroutine | s_amr_st_finalize () |
| Free the flat store and dense-index maps (idempotent; called from both finalize paths). | |
| impure subroutine, public | s_amr_alloc_slot (islot) |
| Allocate slot islot's coords and its dense store index (idempotent). | |
| impure subroutine, public | s_amr_free_slot (islot) |
| Free slot islot's per-block field arrays (inverse of s_amr_alloc_slot). Idempotent (no-op if not live). | |
| impure subroutine, public | s_amr_reconcile_slots () |
| Reconcile allocated slots to current ownership: allocate every active block this rank owns, free the rest. Call after ownership is set; regrid must allocate its transient slots before this call. | |
Variables | |
| real(wp), dimension(:), allocatable | amr_gs_buf |
| Staging for the coordinate push of s_amr_sync_grid_state_to_device: the active dims' cb/cc/d arrays back to back. | |
Flat per-block field store: dense slot indexing, reservation, alloc/free and the prim/bridge loaders.
| impure subroutine, public m_amr_store::s_amr_alloc_pool |
Block-metadata pool for amr_max_blocks slots (geometry, ownership, level, SFC cuts, overlap-list counts; the 2D rank lists are sized in s_amr_build_seam_pairs) and the dense-index maps.
Definition at line 1025 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_alloc_slot | ( | integer, intent(in) | islot | ) |
Allocate slot islot's coords and its dense store index (idempotent).
Definition at line 3680 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_bat_member_prim | ( | integer, intent(in) | ibm, |
| type(scalar_field), dimension(sys_size), intent(in) | src, | ||
| type(scalar_field), dimension(sys_size), intent(inout) | dst ) |
Copy batch member ibm's primitive state from the slab scratch into the block-frame scratch over its buffered extent, for the per-member IB correction.
Definition at line 2973 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_br_load | ( | integer, intent(in) | loc | ) |
Move block loc's conserved state between the flat store and the bridge over the whole buffered box (all slots share the mbuf extents).
Definition at line 2338 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_br_load_batch | ( | integer, intent(in) | nb | ) |
Batched bridge load: every batch member (amr_bat_blk/amr_bat_loc, extents amr_bat_ext) lands at offset (ibm-1)*amr_bat_w along amr_bat_sd with its ghost shell, in one kernel. Nothing is stored back: s_compute_rhs does not write its cons dummy on the fine path.
Definition at line 2905 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_br_load_faces | ( | integer, intent(in) | loc, |
| integer, dimension(3), intent(in) | olo, | ||
| integer, dimension(3), intent(in) | ohi, | ||
| integer, dimension(3), intent(in) | glo, | ||
| integer, dimension(3), intent(in) | ghi, | ||
| integer, dimension(3), intent(in) | woff, | ||
| real(wp), dimension(3), intent(in) | w_lo, | ||
| real(wp), dimension(3), intent(in) | w_hi ) |
Face-bounded bridge move for the level>=2 reflux apply: moves only the six one-cell planes (outside cell olo(d)/ohi(d) of each active dim, over the caller's transverse window) that s_amr_reflux_apply_faces read-modify-writes, skipping zero-weight planes as the apply does. Bit-identical to a whole-box move.
Definition at line 2465 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_br_store | ( | integer, intent(in) | loc | ) |
| impure subroutine, public m_amr_store::s_amr_br_store_faces | ( | integer, intent(in) | loc, |
| integer, dimension(3), intent(in) | olo, | ||
| integer, dimension(3), intent(in) | ohi, | ||
| integer, dimension(3), intent(in) | glo, | ||
| integer, dimension(3), intent(in) | ghi, | ||
| integer, dimension(3), intent(in) | woff, | ||
| real(wp), dimension(3), intent(in) | w_lo, | ||
| real(wp), dimension(3), intent(in) | w_hi ) |
|
private |
Re-densify the local index space after every reconcile so amr_loc_n tracks the live count instead of ratcheting up with SFC window shifts and migration slots (capacity then plateaus at the rebuild transient). Moves are device-side and in place in ascending old-index order: each destination is <= its source and below every pending source. Called at the end of s_amr_reconcile_slots, after the rebuild's overlap carry-forward (which reads amr_stor_st at old indices) has finished.
Definition at line 1371 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_copy_fine_fields | ( | integer, intent(in) | k | ) |
Device copy amr_cons_st -> amr_stor_st over block k's buffered box, all sys_size fields (RK step-entry backup).
| [in] | k | block slot (its flat-store column) |
Definition at line 968 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_dev_mem | ( | integer(8), intent(out) | free_b, |
| integer(8), intent(out) | total_b ) |
Free and total device memory in bytes (hipMemGetInfo on the AMD GPU build); -1 for both where there is no query.
Definition at line 1396 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_free_pool |
Free what s_amr_alloc_pool allocated (plus the per-block caches sized later), and the coordinate swap buffers.
Definition at line 1143 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_free_slot | ( | integer, intent(in) | islot | ) |
Free slot islot's per-block field arrays (inverse of s_amr_alloc_slot). Idempotent (no-op if not live).
Definition at line 3717 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_free_swap_buffers |
Definition at line 1162 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_init_swap_buffers |
Bounce buffers for the copy-based coordinate swap; shaped like the base-level *_alloc arrays, since they are assigned whole to/from x_cb etc.
Definition at line 1042 of file m_amr_store.fpp.f90.
|
private |
Allocate/reset the dense local-index maps (idempotent).
Definition at line 1250 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_reconcile_slots |
Reconcile allocated slots to current ownership: allocate every active block this rank owns, free the rest. Call after ownership is set; regrid must allocate its transient slots before this call.
Definition at line 3749 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_set_mbuf |
The buffered per-slot array bounds from the max fine extents (collapsed dims 0:0).
Definition at line 1156 of file m_amr_store.fpp.f90.
|
private |
Free the flat store and dense-index maps (idempotent; called from both finalize paths).
Definition at line 3034 of file m_amr_store.fpp.f90.
|
private |
Move one local slot's store data src -> dst on the device, in place (a staging copy would transiently double the store). Ascending-source order (s_amr_compact_store) guarantees dst's old contents are consumed. The host copy goes stale, the store's normal state between rebuilds.
Definition at line 1261 of file m_amr_store.fpp.f90.
|
private |
Size the flat store for at least nloc local slots, growing geometrically and never shrinking. Growth preserves the live columns on the device (no PCIe round trip); the host mirror comes out undefined.
device bytes kept free through a growth; without a free-memory query, the largest old store staged on the device
Definition at line 1410 of file m_amr_store.fpp.f90.
| impure subroutine, public m_amr_store::s_amr_sync_grid_state_to_device |
Push the host global grid state (m/n/p, idwint/idwbuff, coordinate arrays) to its GPU_DECLARE'd device copies after a swap/restore. Coordinates go through amr_gs_buf as one update since each allocatable update costs several synchronous copies. No-op on CPU.
Definition at line 408 of file m_amr_store.fpp.f90.
| real(wp), dimension(:), allocatable m_amr_store::amr_gs_buf |
Staging for the coordinate push of s_amr_sync_grid_state_to_device: the active dims' cb/cc/d arrays back to back.
Definition at line 384 of file m_amr_store.fpp.f90.