64-BitBrainstorm_2026/AMSS-NCKU
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
146 Commits 49 Branches 0 Tags
chb-cuda-new
Commit Graph

127 Commits

Author SHA1 Message Date
ianchb
a0dab90bcb Switch to NVIDIA HPC Toolchain 2026-04-29 08:31:49 +08:00
ianchb
c689cc8dc9 [WIP] Add CUDA support for Z4C 
Rewritten done by Codex.
This still has errors, do not pick this one now.
 2026-04-27 11:58:43 +08:00
ianchb
60fee8f1c1 Fix Z4C C++ gauge damping ordering 2026-04-26 15:38:13 +08:00
ianchb
843b116954 Add C++ Z4C RHS path and port some BSSN optimizations 2026-04-25 10:39:01 +08:00
ianchb
c768e1220b Also disable cached sync for Z4C 2026-04-25 10:25:54 +08:00
CGH0S7
02f149e2e3 Disable cached sync for BSSN-EScalar 2026-04-25 10:17:47 +08:00
CGH0S7
422e8ec4dc Fallback BSSN-EScalar restrict/prolong path 2026-04-25 10:10:34 +08:00
ianchb
f521a97563 Fix ABE CPU version build error 2026-04-25 09:39:49 +08:00
ianchb
53c55451b3 Update makefile and scripts for CUDA BSSN configuration and build commands 2026-04-25 09:19:50 +08:00
CGH0S7
768345954f Add optional BSSN kernel profiling switches 
(cherry picked from commit 9c31384b2f)
 2026-04-25 08:39:43 +08:00
CGH0S7
8e9463aa90 Localize chi Ricci intermediates in RHS 
(cherry picked from commit 65e0f95f40)
 2026-04-25 08:37:41 +08:00
CGH0S7
7c6f15002e Elide dead stores in BSSN RHS hot path 
(cherry picked from commit f9fbf97e64)
 2026-04-25 08:37:40 +08:00
CGH0S7
6410c62e3e Add fine-grained step timing and trim BH RHS overhead 
(cherry picked from commit 968522995b)
 2026-04-25 08:37:19 +08:00
CGH0S7
11977eb82f Merge wave and mass extraction interpolation 
(cherry picked from commit f3988ac8ca)
 2026-04-25 08:25:34 +08:00
CGH0S7
cce8a44fc4 Cache wave extraction angular kernels 
(cherry picked from commit e4c25eb21f)
 2026-04-25 08:24:36 +08:00
CGH0S7
c589097618 Reuse mass integrand across detector radii 
(cherry picked from commit 4b10519876)
 2026-04-25 08:24:11 +08:00
CGH0S7
b713e5a9be Batch constraint norm reductions 
(cherry picked from commit 3a58273501)
 2026-04-25 08:22:00 +08:00
CGH0S7
0396701572 Optimize constraint refresh after regrid 
(cherry picked from commit 5c65cea2f0)
 2026-04-25 08:18:51 +08:00
ianchb
bb20c9a876 fix ADM Constrant Violation Analysis 2026-04-15 19:19:16 +08:00
ianchb
8fe60ea703 Add zero matter handling and interpolation for resident state in CUDA BSSN 2026-04-15 00:25:53 +08:00
ianchb
9ab7e7c7f9 Fuse phases 5 and 6 for Gamma_rhs computation and optimize phases 8 and 9 for efficiency 2026-04-14 23:23:04 +08:00
ianchb
f9119e8a2a Add resident-GA mode switch and simplify sync logic 2026-04-14 21:09:27 +08:00
ianchb
726d743376 Fuse Ricci assembly and optimize trK/Aij gauge kernels 2026-04-14 19:20:12 +08:00
ianchb
af344bf1e5 Add Phase-10 Ricci kernels and batch launch flow 2026-04-14 19:00:22 +08:00
ianchb
7191fc0b96 Move resident sync comm buffers into StepAllocation pool 2026-04-13 21:04:44 +08:00
ianchb
b3ec244cf9 Add batched first/second derivative kernels for CUDA RHS 2026-04-13 20:51:08 +08:00
ianchb
e952ee8e91 Batch GA/BH subset sync with indexed GPU pack/unpack buffers 2026-04-13 20:40:09 +08:00
ianchb
c5d1268dd1 Batch patch-boundary copy and gate CPU BC in GPU substeps 2026-04-13 11:52:17 +08:00
ianchb
4bdfc90f22 Pass pointer tables as kernel args and skip redundant symbol uploads 2026-04-13 11:19:00 +08:00
ianchb
c49a4e00c9 Batch symbd_pack/lopsided/kodiss over all state variables 2026-04-13 11:02:55 +08:00
ianchb
1b3c0b80d2 Refactor CUDA step buffers to remove loop-time allocations 2026-04-13 10:33:03 +08:00
ianchb
636e35bfd8 Add direct CUDA resident-state sync path and profiling hooks 2026-04-13 00:57:05 +08:00
ianchb
7f2a391dd2 Cache matter fields in StepContext across RK4 substeps 2026-04-12 22:19:45 +08:00
ianchb
4fa12a2009 Integrate CUDA support into RK4 substep execution 2026-04-12 22:11:44 +08:00
ianchb
86a683de26 Replace legacy ABEGPU stack with ABE_CUDA backend 2026-04-12 21:19:14 +08:00
ianchb
aaf7bf0a26 Merge remote-tracking branch 'origin/main' 2026-04-12 20:55:42 +08:00
CGH0S7
8c1f4d8108 迁移C算子的循环融合和临时量消除 2026-03-03 16:20:15 +08:00
CGH0S7
d310ef918b bssn_rhs(fortran): migrate C kernel loop-fusion optimizations 2026-03-03 16:20:15 +08:00
CGH0S7
b35e1b289f 设置开关关闭内存打印统计 2026-03-03 16:17:47 +08:00
CGH0S7
05851b2c59 关闭静态负载 2026-03-03 16:17:47 +08:00
ianchb
3b39583d67 fix(bssn_rhs) 2026-03-03 16:06:33 +08:00
ianchb
9c44d1c885 fix(bssn_rhs) 2026-03-03 16:00:45 +08:00
ianchb
4b9de28feb 将 Restrict/Prolong 链路里的 coarse-level Sync_cached 改为可选（默认跳过） 
OutBdLow2Hi_cached 读的是 coarse owned 区域（非 coarse ghost/buffer）
回退旧行为：编译时定义 RP_SYNC_COARSE_AFTER_RESTRICT=1
 2026-03-03 14:25:27 +08:00
ianchb
4eb5dc4ddb 删除重复的一次 chi 一阶导计算 2026-03-03 14:23:56 +08:00
CGH0S7
5070134857 perf(transfer_cached): 将 per-call new/delete 的 req_node/req_is_recv/completed 数组移入 SyncCache 复用 
避免 transfer_cached 每次调用分配释放 3 个临时数组，减少堆操作开销。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-02 21:14:35 +08:00
CGH0S7
4012e9d068 perf(RestrictProlong): 用 Restrict_cached/OutBdLow2Hi_cached 替换非缓存版本，Sync_finish 改为渐进式解包 
- RestrictProlong/RestrictProlong_aux 中的 Restrict() 和 OutBdLow2Hi() 替换为 _cached 版本，
  复用 gridseg 列表和 MPI 缓冲区，避免每次调用重新分配
- 新增 sync_cache_restrict/sync_cache_outbd 两组 per-level 缓存
- Sync_finish 从 MPI_Waitall 改为 MPI_Waitsome 渐进式解包，降低尾延迟
- AsyncSyncState 扩展 req_node/req_is_recv/pending_recv 字段支持渐进解包

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
 2026-03-02 20:48:38 +08:00
ianchb
b3c367f15b prolong3 改为先算实际 stencil 窗口；只有窗口触及对称边界时才走全域 symmetry_bd，否则只复制必需窗口。restrict3 同样改成窗口判定，无触边时仅填 ii/jj/kk 必需窗口。 2026-03-02 17:38:56 +08:00
ianchb
e73911f292 perf(restrict3): shrink X-pass ii sweep to required overlap window 
- compute fi_min/fi_max from output i-range and derive ii_lo/ii_hi
 - replace full ii sweep (-1:extf(1)) with windowed sweep in Z/Y precompute passes
 - keep stencil math unchanged; add bounds sanity check for ii window
 2026-03-02 17:37:13 +08:00
ianchb
7543d3e8c7 perf(MPatch): 用空间 bin 索引加速 Interp_Points 的 block 归属查找 
- 为 Patch::Interp_Points 三个重载引入 BlockBinIndex（候选筛选 + 全扫回退）
  - 保持原 point-in-block 判定与后续插值/通信流程不变
  - 将逐点线性扫块从 O(N_points*N_blocks) 降为近似 O(N_points*k)
  - 测试：bin 上限如果太大，会引入不必要的索引构建开销。将 bins 上限设为 16。

Co-authored-by: gpt-5.3-codex
 2026-03-02 17:37:13 +08:00
ianchb
42c69fab24 refactor(Parallel): streamline MPI communication by consolidating request handling and memory management 2026-03-02 17:37:13 +08:00