87727383fe
Merge remote-tracking branch 'nccl/master' into 2.9.8
2021-05-17 10:15:16 -07:00
0b2bfdd6d8
Multiprocess unit test various fixes ( #367 )
...
* Re-enabling mp unit tests
* Fixing shared memory leak and other bugs related to shared mem for MP unit tests
* Revert 43bfbfc97bf9edbae1f386d461439091618ff8ed
* Further tightening up unlinks
* Moving test check macros to separate header file
* Tightening up shared memory unlinking for clique kernels, add munmap for host barrier for MP unit tests
* Updating new MP unit test
* Fixing mqueue bug
* Fixing memory leak in MP unit tests
2021-05-14 09:38:49 -06:00
caf5c9992a
Allow user to link layer with RCCL_IB_HCA_SKIP_LINK_LAYER ( #361 )
...
To skip Infiniband, set RCCL_IB_HCA_SKIP_LINK_LAYER=1.
To skip Ethernet, RCCL_IB_HCA_SKIP_LINK_LAYER=2.
2021-05-12 14:14:53 -07:00
e796b1645c
Clique tuning upgrade ( #352 ) ( #19 )
...
* Enabling clique for any XGMI-connected topology, adding tuning
* Updating CHANGELOG for clique tuning
* Re-working clique barrier system to work on multi-process / multi-gpu
2021-05-11 08:44:59 -06:00
ca8485b0d0
2.9.8-1
...
Fix memory leaks.
Fix crash in bootstrap error case.
Fix Collnet clean-up issue.
Make PCI switch vendor/device optional for XML injection.
Add support for nvidia-peermem module.
2021-05-10 14:00:03 -07:00
a4ea1fed5b
Merge remote-tracking branch 'nccl/master' into develop
2021-05-05 16:01:01 -07:00
4f8e788a61
Fixing potential race-condition in env var parameter macro ( #359 )
2021-04-28 12:04:41 -06:00
ed237dcaa7
Use better name for kernel collective trace enable ( #357 )
...
"NCCL_DEBUG=INFO NCCL_DEBUG_SUBSYS=INIT,COLL" enables collectives API
trace. Adding "RCCL_KERNEL_COLL_TRACE_ENABLE=1" enables kernel traces.
2021-04-26 08:35:53 -07:00
9cc9c3360b
Control collective trace from kernel separately ( #356 )
2021-04-23 16:36:19 -07:00
70597789d0
Message queue refactor to POSIX implementation and leak fix ( #355 )
...
* Fixing message queue leak.
* Using POSIX implementation of Message Queues
* Adding unlink to msgqueue
* MsgQueue update
* Adding timeout check to msgqueue broadcast; tightening up system checks
* Removing unnecessary code
* Removing extra argument from print
* Adding explicit msg queue close call to all other ranks
2021-04-23 11:33:20 -06:00
415c7cd3d1
Tune number of channels for gfx90a ( #349 )
2021-04-19 15:27:01 -07:00
9c718ce6d6
Use correct WARP_SIZE for gfx1030 ( #348 )
2021-04-14 14:09:52 -07:00
a79f74082e
Limit max channels for ring graph on single node Rome ( #347 )
...
* Limit max channels for ring graph on single node Rome
* Partially revert "Use non-temporal access for streaming data (#341 )"
2021-04-14 10:14:54 -07:00
1fe031402a
Add gfx90a target ( #344 )
...
* Add gfx90a target
* Support gfx90a topology
Co-authored-by: Eiden Yoshida <eiden.yoshida@amd.com >
2021-04-14 09:29:00 -06:00
a46ea10583
2.9.6-1
...
Add support for CUDA graphs.
Fuse BCM Gen4 switches to avoid suboptimal performance on some platforms. Issue #439 .
Fix bootstrap issue caused by connection reordering.
Fix CPU locking block.
Improve CollNet algorithm.
Improve performance on DGX A100 for communicators with only one GPU per node.
2021-04-12 16:00:46 -07:00
87f12cbb86
Add detection of cooperative multi device launch attribute ( #345 )
2021-04-11 13:29:24 -07:00
9dfc2c183e
Use non-temporal access for streaming data ( #341 )
...
* Use non-temporal access for streaming data
* Revert to ulong2 after fixing compiling issue
2021-04-07 17:34:35 -07:00
caba0a63d2
Fixing clique-topology detection ( #342 )
...
* Fixing clique-topology detection
* Fix to enable multi-process clique-based kernels
2021-04-07 11:29:44 -06:00
e26ad2995e
Cleanup number of channels calculation ( #340 )
2021-04-05 17:51:56 -07:00
17491c918e
Fix incorrect net counting ( #339 )
...
* Fix incorrect net counting
* Add comments
2021-04-05 12:21:57 -07:00
1d2946ee4b
Rework network port trimming code ( #338 )
...
* Rework network port trimming code
* Move Rome related changes to separate source files
2021-03-31 10:25:59 -07:00
0c78553ee0
Check fine grained memory on peer GPU before enabling P2P ( #337 )
2021-03-30 09:06:39 -07:00
d87dc7c2e8
collnet: support multiple NICs ( #335 )
2021-03-25 20:59:32 -07:00
289db2a636
Fixing message queue leak. ( #331 )
2021-03-25 19:11:43 -06:00
0fbb9510a5
Remove HDP workaround for ROCm 4.2 HIP ( #334 )
2021-03-23 20:11:37 -07:00
1d6244b18d
Enable collnet in RCCL ( #333 )
...
* Enable CollNet and use different number of channels
* topo_expl: enable collnet
2021-03-19 12:58:13 -07:00
b46260260a
Sort GPUs by HIP device ID ( #329 )
...
* Sort GPUs by HIP device ID
* Remove extra space
2021-03-16 16:51:32 -07:00
f60b76c67a
Add GPU memory usage tracker ( #326 )
2021-03-06 20:32:30 -08:00
8e180cf087
Revert "Port alltoall[v]" ( #325 )
...
This reverts commit f4d5d3d620
2021-03-06 13:59:31 -08:00
c018edf0f2
Enable local sendrecv over network if GDR is available on all GPUs ( #324 )
2021-03-05 19:59:41 -08:00
f4a9b9acba
Adding pthread_join / pthread_detach to clean up pthreads to avoid leaks ( #322 )
2021-02-26 16:29:55 -07:00
e820a943e9
Update tuning parameters for XGMI and NET
2021-02-23 21:41:26 +00:00
ec8d89b1dd
Match NBIO only when GPUs and NICs are directly connected to CPU
2021-02-22 18:52:29 -05:00
45f5255f7c
Fixing cache deletion for CliqueManager; updating copyright
2021-02-19 22:22:46 +00:00
95f178324c
Add support to another Rome model
2021-02-18 02:00:31 +00:00
c985358e11
Merge remote-tracking branch 'nccl/master' into 2.8.3
2021-02-15 18:44:47 -05:00
bf8eb40705
Move HDP flush to CPU
2021-02-12 18:06:19 +00:00
911d61f214
2.8.4-1
...
Fix hang in corner cases of alltoallv using point to point send/recv.
Harmonize error messages.
Fix missing NVTX section in the license.
Update README.
2021-02-09 15:36:48 -08:00
9cc3b56166
Fix GDRDMA read and remove unused files
2021-02-09 01:34:39 +00:00
d00b7d17bd
Update MP UT to support arbitrary # of GPUs; multiple bugfixes ( #16 )
...
* Fixing temp file creation/deletion for Clique kernel mode.
* Refactoring of MP unit tests; include bugfixes and general support for any number of GPUs
* GroupCall MP UT properly quits when too many devices specified
* MP UT will programmatically set NCCL_COMM_ID if not specified; updated install script
2021-02-05 16:49:25 -08:00
ab1e7a0318
Merge remote-tracking branch 'origin/develop' into 2.8.3
2021-02-04 20:02:34 -05:00
1990ffd76a
Tuning some clique-based kernel parameters ( #315 )
2021-02-03 20:00:08 -07:00
5f97122442
Enable GPU direct RDMA read from GPU
2021-02-03 02:48:30 +00:00
3e62ceddc5
Clique kernel support ( #295 ) ( #15 )
...
* Adding experimental clique-based kernels (opt-in only)
Co-authored-by: Stanley Tsang <stanley.tsang@amd.com >
Co-authored-by: Gilbert Lee <gilbert.lee@amd.com >
Co-authored-by: Wenkai Du <43822138+wenkaidu@users.noreply.github.com >
Co-authored-by: Stanley Tsang <stanley.tsang@amd.com >
Co-authored-by: Wenkai Du <43822138+wenkaidu@users.noreply.github.com >
2021-01-28 09:45:01 -07:00
41e47a36e7
Use less unroll for clique kernels ( #313 )
2021-01-15 17:48:10 -08:00
2ddbe6646b
Improve collective trace
2021-01-14 19:28:01 -05:00
f4d5d3d620
Port alltoall[v]
2021-01-14 19:28:01 -05:00
105db19a11
Do not allow GPU as intermediate
2021-01-14 19:28:01 -05:00
e055229e56
Revert "Changes to topology based on XGMI ( #272 )"
...
This reverts commit 01bd2573db
2021-01-14 19:28:01 -05:00
d469947641
Merge remote-tracking branch 'nccl/master' into no-target-id
2021-01-14 19:27:53 -05:00
Wenkai Du
Stanley Tsang
Wenkai Du
gilbertlee-amd
Sylvain Jeaugey
TomSang