Peter Boyle
c15493218d
Two extra routines to break out SchurRedBlack on many RHS into stages to allow efficient deflation & split grid
...
Split grid solver still to do.
2021-09-15 19:24:39 +01:00
Christoph Lehner
3d0f88e702
A64FX drop mixed precision as well
2021-09-15 18:38:32 +02:00
Christoph Lehner
dd091d0960
consistent pointer offloading instead of views
2021-09-15 16:58:05 +02:00
Christoph Lehner
e2abbf9520
Merge pull request #25 from paboyle/develop
...
Sync
2021-09-15 10:02:43 +02:00
Peter Boyle
86e33c8ab2
Significant GPU perf speed up finished
2021-09-14 16:14:23 +01:00
Peter Boyle
65ef4ec29f
Move tables to device memory
2021-09-14 15:05:01 +01:00
Peter Boyle
d5835c0222
Switch to coalesced stencil face gather
2021-09-14 15:04:14 +01:00
Peter Boyle
a7b943b33e
Remove half prec comms
2021-09-14 05:05:33 +01:00
Peter Boyle
7440cde92f
No half prec comms; coalesced access on GPU
2021-09-14 05:04:56 +01:00
Peter Boyle
0fc662bb24
Dirac cuda 11.4 happy ; force host for functions accessing mult table
...
ET runs these on host BEFORE lodging result in AST for kernel
2021-09-14 05:00:44 +01:00
Peter Boyle
8195890640
Force MPI over NVLINK
2021-09-14 05:00:17 +01:00
Peter Boyle
4c88104a73
Fix compile warns
2021-09-11 23:08:05 +01:00
Peter Boyle
73b944c152
Drop half prec comms for now.
2021-09-11 23:07:18 +01:00
Peter Boyle
d1b0b7f5c6
Half prec comms dropping
2021-09-11 23:05:40 +01:00
Peter Boyle
381d8797d0
Drop half prec comms for now
2021-09-11 23:05:02 +01:00
Peter Boyle
0d588b95f4
Bug fix to Example_Laplacian test
2021-08-23 23:14:26 +01:00
Peter Boyle
5b3c530aa7
Return value
2021-08-23 15:30:45 +01:00
Peter Boyle
c6a5499c8b
Fail on non-apple
2021-08-22 18:40:55 +01:00
Peter Boyle
ffbdd91e0e
Apple happiness
2021-08-20 01:15:00 +01:00
Peter Boyle
5d29e175d8
Typo fix
2021-08-10 18:25:43 +01:00
Peter Boyle
417dbfa257
Fix
2021-08-10 08:55:35 -07:00
peterx.a.boyle
1eda4d8e0b
Merge branch 'develop' of https://github.com/paboyle/Grid into develop
2021-08-10 05:41:18 -07:00
peterx.a.boyle
50181f16e5
Level 0 IPC set up
2021-08-10 05:35:15 -07:00
Peter Boyle
80ac2a73ca
Check is wrong (HtoD / DtoH)
2021-08-05 18:33:20 -04:00
Andrew Yong
770680669d
Whitespace removal.
2021-08-04 09:21:59 +01:00
Andrew Yong
0cdfc5cf22
Merge remote-tracking branch 'upstream/develop' into develop
2021-07-30 14:40:55 +01:00
Ed Bennett
323cf6c038
make message consistent with configure script
2021-06-23 17:00:43 +01:00
Peter Boyle
29a22ae603
Simpler SYCL setup
2021-06-22 17:57:20 +00:00
Peter Boyle
403bff1a47
Force reqd subgroup size fo SYCL
2021-06-22 17:56:10 +00:00
Christoph Lehner
c50f27e68b
Make FFT play nice with split grid
2021-06-20 11:34:38 +02:00
Peter Boyle
6cd9224dd7
SYCL comms buffer allocate
2021-06-16 17:10:55 +00:00
Peter Boyle
4c5440fb06
const happy for sycl
2021-06-15 21:45:07 +00:00
Michael Marshall
e5dbe488a6
Merge branch 'develop' into bugfix/ragged
...
* develop:
Remove synch
2021-06-03 08:25:56 +01:00
Peter Boyle
0e27e3847d
Remove synch
2021-06-03 04:24:19 +00:00
Michael Marshall
2b1fcd78c3
Fixes post review with Peter: a) Correct bug in isRegularShape - detect 3d matrix where 1st slice is 2x2 and second slice is 2x1; b) Synchronisation of EigenResizeCounter done by checking we're the OMP primary thread; c) Move definition of EigenResizeCounter to new file, BaseIO.cc
2021-05-31 22:24:54 +01:00
Michael Marshall
0a4e0b49a0
BaseIO: Added "EigenResizeCounter" to keep track of any allocations/deallocations to Eigen tensors during readback. On read, if the tensor is resized, EigenResizeCounter += delta memory (in bytes)
2021-05-31 12:49:56 +01:00
Michael Marshall
76af169f05
Add global namespace to Writer<T> and Reader<T> inside GRID_SERIALIZABLE_CLASS_MEMBERS (so that "using Grid" not necessary).
...
Fix issue with output of Grid::iMatrix so that M<3>{{148,149,150,} {151,152,153,} {154155156}} becomes M<3>{{148,149,150} {151,152,153} {154,155,156}}
2021-05-31 08:43:02 +01:00
Michael Marshall
7b89232251
Extended HDF5 serialisation of std::vector<T> where T now also includes Grid scalar/vector/matrix
...
Changed VectorUtils element traits to is_flattenable, because: a) contract changed on what it does; and b) no other Grid dependencies on element. Needs review.
Initial tests work ... needs proper regression testing.
2021-05-30 20:27:53 +01:00
Peter Boyle
b5aeae526f
Make Cshift fields static to avoid repeated reallocaate overhead
2021-05-28 16:33:08 +02:00
Michael Marshall
ef0ddd5d04
std::vector serialisation in hdf5 uses a different format if the vector is ragged. When reading back std::vector we need to check which format we're reading (since we don't know a priori) and this involves looking for attributes that may not exist. The c++ API: a) throws; and b) prints voluminous logging. Switched to non-throwing, non-logging, C version of the API after code review.
2021-05-24 18:43:55 +01:00
Michael Marshall
9b73dacf50
First row might still be ragged if multi dimensional. attrExists() doesn't throw, but easier to wrap in try ... catch than to explain in comment.
2021-05-22 04:34:32 +01:00
Michael Marshall
244b4aa07f
Serialise std::vector of numeric types as multidimensional object if size is regular ... or individually if ragged
2021-05-21 20:08:56 +01:00
u61464
8cfc7342cd
staggered hand unroll read coalesce
2021-05-05 14:17:18 -07:00
u61464
15ae317858
Merge branch 'develop' of https://github.com/paboyle/Grid into develop
2021-05-04 08:40:38 -07:00
u61464
834f536b5f
Fastest option on SyCL is now std::complex
2021-05-04 08:40:18 -07:00
cf2923d5dd
Jamie's fix
2021-04-27 16:53:37 +01:00
009ccd581e
bugfix 3D stout smearing
2021-04-26 10:36:33 +01:00
Peter Boyle
d45c868656
Change interface
2021-04-25 10:53:34 -04:00
Peter Boyle
955a8113de
Expose label only to reduce number of parameters
2021-04-25 10:36:38 -04:00
Peter Boyle
dbe210dd53
Open the ens_id
2021-04-25 10:25:59 -04:00
