portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-07-17 23:53:27 +01:00
Code Releases Activity
7,602 Commits 95 Branches 14 Tags
070b61f08fb6ec761a8ba9d4719b57ead0e3e9a9
Commit Graph
100 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
Peter Boyle 070b61f08f Simplifying the MultiRHS solver to make it do SRHS *and* MRHS 2024-03-06 14:04:33 -05:00
Peter Boyle ee3b3c4c56 relocate deflation support 2024-02-27 11:52:23 -05:00
Peter Boyle 462d706a63 Move to a blas directory 2024-02-27 11:51:04 -05:00
Peter Boyle ee0d460c8e Blas based block project & deflate for multiRHS 2024-02-27 11:41:44 -05:00
Peter Boyle cd15abe9d1 Mrhs prep 2024-02-27 11:41:13 -05:00
Peter Boyle 9f40467e24 Warning squash 2024-02-27 11:40:36 -05:00
Peter Boyle d0b6593823 More verbose on checksum 2024-02-27 11:40:14 -05:00
Peter Boyle 79fc821d8d reorg headers 2024-02-27 11:39:37 -05:00
Peter Boyle d7fdb9a7e6 Reorg headers 2024-02-27 11:39:06 -05:00
Peter Boyle b74de51c18 Reorder headers 2024-02-27 11:38:52 -05:00
Peter Boyle 44b466e072 Make InsertSliceFast the default at some point in future. 
Should I do this now?
 2024-02-21 14:51:24 -05:00
Peter Boyle 5e5b471bb2 Put/Get and DEviceToDevice 2024-02-21 14:47:06 -05:00
Peter Boyle 9c2565f64e Working and faster version 2024-02-21 14:46:43 -05:00
Peter Boyle e1d0a7cec3 Batched blas 2024-02-21 14:38:20 -05:00
Peter Boyle b19ae8f465 Nbasis method for convenience 2024-02-21 14:36:19 -05:00
Peter Boyle cdff2c8e18 Updated mrhs adef 2024-02-21 14:27:19 -05:00
Peter Boyle eb702f581b Running on 12 rhs on 18 nodes of frontier 2024-01-22 17:44:15 -05:00
Peter Boyle 3d13fd56c5 Precompute phases, save memory in hermitian 2024-01-22 17:43:35 -05:00
Peter Boyle 6f51b49ef8 Use stderr 2024-01-22 17:41:09 -05:00
Peter Boyle addc638856 Fast localCopyRegion, blockProjectFast 2024-01-22 17:40:38 -05:00
Peter Boyle 42ae36bc28 WOrking 2024-01-17 16:39:14 -05:00
Peter Boyle c69f73ff9f Working 2024-01-17 16:38:46 -05:00
Peter Boyle ca5ae8a2e6 Revert to working. 2024-01-17 16:32:05 -05:00
Peter Boyle d967eb53de Working for first time 2024-01-17 16:31:12 -05:00
Peter Boyle 839f9f1bbe Don't log memory by default 2024-01-17 16:25:50 -05:00
Peter Boyle b754a152c6 Flag guard correctly 2024-01-17 16:25:28 -05:00
Peter Boyle e07cb2b9de Accelerator memory 2024-01-17 16:24:31 -05:00
Peter Boyle a1f8bbb078 accelerator memory print 2024-01-17 16:24:09 -05:00
Peter Boyle 7909683f3b MultiRHS 2024-01-17 16:21:07 -05:00
Peter Boyle 25f71913b7 MultiRHS coarse 2024-01-04 12:01:17 -05:00
Peter Boyle 34ddd2b7b1 MultiRHS coarse space 2024-01-04 12:00:53 -05:00
Peter Boyle d5fd90b2f3 Add 48^3 rtest 2024-01-04 12:00:01 -05:00
Peter Boyle b7c7000d0d Don't need the numerical rounding tolerance in multigrid 2023-12-22 18:10:23 -05:00
Peter Boyle 551f6c4edd Synchronise changes 2023-12-22 18:09:11 -05:00
Peter Boyle defd814750 Speed up the coarsened matrix matrix evaluation. 
It is block project limited.
Could be sped up with calls to Batched GEMM and a data layout change.
 2023-12-22 18:07:03 -05:00
Peter Boyle 3d517bbd2a Synchronise decouple from the launch 
Speeds up multileg stencils
 2023-12-22 18:06:13 -05:00
Peter Boyle 78ab955fec Better padded cell exchange 2023-12-22 18:05:41 -05:00
Peter Boyle dd13937bb6 Better opt face gather scatter 2023-12-22 18:03:38 -05:00
Peter Boyle 66a1b63aa9 Faster grid/blas layout change. 
Halo exchange is now the only slow part.
Revisit
 2023-12-21 20:50:18 -05:00
Peter Boyle 22c611bd1a Delete temp file 2023-12-21 18:32:31 -05:00
Peter Boyle c9bb1bf8ea Passing new BLAs based 2023-12-21 18:31:17 -05:00
Peter Boyle 9e489887cf General coarse multiRHS move to BLAS implementation 2023-12-21 15:24:48 -05:00
Peter Boyle 9feb801bb9 Much simpler GPU implementation 2023-12-21 15:24:06 -05:00
Peter Boyle c00b495933 Multigrid 2023-12-21 15:23:31 -05:00
Peter Boyle d22eebe553 BLas options 2023-12-21 15:23:03 -05:00
Peter Boyle 8bcbd82680 BLAS based layout and implementation 2023-12-21 15:21:24 -05:00
Peter Boyle dfa617c439 Batched SGEMM/DGEMM/ZGEMM/CGEMM 
Hip, Cuda version and vanilla CPU
One MKL stub in comments, to be tested as different.
 2023-12-21 14:01:18 -05:00
Peter Boyle 48d1f0df89 Optimised partially, working 2023-12-21 12:33:47 -05:00
Peter Boyle b75cb7a12c Blas batched partial implementation on Frontier only for now 2023-12-21 12:31:33 -05:00
Peter Boyle 332563e037 Debugged, reducing verbose 2023-12-21 12:30:57 -05:00
Peter Boyle 0cce97a4fe verbosity only 2023-12-20 21:30:10 -05:00
Peter Boyle 95a8e4be64 rocblas 2023-12-20 21:27:59 -05:00
Peter Boyle abcd6b8cb6 Faster version 2023-12-19 15:17:46 -05:00
Peter Boyle e8f21c9b6d Memmory verbose control improvement 2023-12-19 15:16:58 -05:00
Peter Boyle e054078b11 Verbose 2023-12-05 16:15:17 -05:00
Peter Boyle 6835a7f208 Better logging, test on 81 point stencil 2023-11-29 19:20:47 -05:00
Peter Boyle f59993b979 Nbasis§ 2023-11-29 09:47:36 -05:00
Peter Boyle 2290b8f680 Verbose 2023-11-29 09:47:04 -05:00
Peter Boyle 2c54be651c Further updates 2023-11-29 09:43:29 -05:00
Peter Boyle e859a199df Reduce volume to interior for coarse stencil -- worth up to 4x gain 2023-11-28 10:23:16 -05:00
Peter Boyle 0a3682ad0b MultiRHS work 2023-11-28 07:43:37 -05:00
Peter Boyle 59abaeb5cd Time stamp 2023-11-24 12:56:45 -05:00
Peter Boyle 3e448435d3 Restrict to interior 2023-11-23 18:23:29 -05:00
Peter Boyle a294bc3c5b Relax constraints for multiRHS 2023-11-23 18:20:42 -05:00
Peter Boyle b302ad3d49 multiRHS test in place, passes Yay! 2023-11-23 18:20:15 -05:00
Peter Boyle 82fc4b1e94 Finalise 2023-11-23 18:19:41 -05:00
Peter Boyle b4f1740380 Finalise message 2023-11-23 18:19:16 -05:00
Peter Boyle 031f85247c multRHS initial support -- needs optimisation for multi project/promote. 
Bug fix in freeing intermediate grids to stop double free
 2023-11-23 18:18:35 -05:00
Peter Boyle 639cc6f73a better support for multiRHS coarse space 
Still to add restriction of domain of last loop to interior of padded cell (expect about 4.5x on test volume on Crusher)
 2023-11-23 18:16:26 -05:00
Peter Boyle 09946cf1ba Improved, works on 48^3 moving to multiRHS optimisations 2023-11-15 18:03:05 -05:00
Peter Boyle f4fa95e7cb Use 5.3.0 2023-11-15 18:01:38 -05:00
Peter Boyle 100e29e35e Allow expression as argument to norm2 2023-11-15 18:00:44 -05:00
Peter Boyle 4cbe471a83 devVector 2023-11-15 18:00:07 -05:00
Peter Boyle 8bece1f861 Faster to transpose the matrix and apply with column major order 2023-11-15 17:58:38 -05:00
Peter Boyle a3ca71ec01 Lots more setup options, still working on them 2023-11-15 17:58:04 -05:00
Peter Boyle e0543e8af5 Implement flexible preconditioned CG 2023-11-15 17:57:39 -05:00
Peter Boyle c1eb80d01a Print which have converged 2023-11-15 17:57:08 -05:00
Peter Boyle a26121d97b Better printing 2023-11-15 17:56:45 -05:00
Peter Boyle 043031a757 Report resid on failed convergence 2023-11-15 17:56:22 -05:00
Peter Boyle 807aeebe4c Resize tol in constructor 2023-11-15 17:55:57 -05:00
Peter Boyle 8aa1a37aad For Mirs preconditioner solver 2023-11-15 17:55:32 -05:00
Peter Boyle 4efa042f50 C++17 change 2023-10-24 10:57:50 -04:00
Peter Boyle c7cb37e970 c++17 accepted 2023-10-24 10:57:24 -04:00
Peter Boyle d34b207eab Avoid HIP warnings 2023-10-24 10:57:04 -04:00
Peter Boyle 0e6fa6f6b8 DOn't need the Cshift for the period optimisation 2023-10-24 10:56:31 -04:00
Peter Boyle 38b87de53f This works around a stacksize limit on AMD GPU 2023-10-24 10:56:07 -04:00
Peter Boyle aa5047a9e4 Faster blockProject blockPromote 2023-10-24 10:49:55 -04:00
Peter Boyle 24b6ee0df9 M4 file 2023-10-24 10:36:48 -04:00
Peter Boyle 1e79cc9cbe Avoid compiler error 2023-10-24 10:36:09 -04:00
Peter Boyle b3925df9c3 Verbose on CPU-GPU xfer, remove performance by default 2023-10-24 10:25:01 -04:00
Peter Boyle 351795ac3a Better messaging 2023-10-20 19:33:04 -04:00
Peter Boyle 9c9c42d0df Tests on frontier with real speed up . 3.5x on 16^3 at mq=0.01 2023-10-20 19:27:13 -04:00
Peter Boyle b6ad1bafc7 Normal memory SendToRecvFrom asynchronous for use in general stencil 
code
 2023-10-20 19:27:13 -04:00
Peter Boyle a5ca40f446 Better verbose -- track CPU GPU motion under --log Memory, others go to 
debug output stream
 2023-10-20 19:27:13 -04:00
Peter Boyle 9ab54c5565 Overlap comms & data copy/buffer assembly in Ghost zone exchange 2023-10-20 19:27:13 -04:00
Peter Boyle 4341d96bde Massively sped up coarse grid mult, comms 
Save 3ms spend (60% of time !) on cudaMalloc !!
 2023-10-20 19:27:13 -04:00
Peter Boyle 5fac47a26d Faster halo exchange 2023-10-20 19:27:13 -04:00
Peter Boyle e064f17346 Faster halo exchange 2023-10-20 19:27:13 -04:00
Peter Boyle afe10ba2a2 More digits 2023-10-20 19:27:13 -04:00
Peter Boyle 7cc3435ba8 Imporved General coarsened matrix 2023-10-20 19:27:13 -04:00