cd7a346297
Doubling buffer size to fix NCCL INFO corruption with increased channels ( #1035 )
...
[ROCm/rccl commit: f7e39fced2
2024-01-08 08:14:33 -08:00
626608c172
Increase stack size for gfx906 ( #1034 )
...
Occationally "Memory access fault by GPU node-8 (Agent handle: 0x23a5640) on address 0x7f461ec00000. Reason: Page not present or supervisor privilege" can be seen from gfx906 CI
[ROCm/rccl commit: e5bf56c6d8
2024-01-07 20:25:02 -08:00
1b39fef32a
Fix MSCCL multi-node ( #1032 )
...
1) Move needsProxy initialization before mscclSetupConnections since the latter
will revise it later.
2) Remove mscclAvailable check in net.cc since it's no more required and caused
non-shared buffer allocated for MSCCL which is not expected.
[ROCm/rccl commit: 70bbeb4773
2024-01-05 17:03:43 -08:00
13791d7ee3
Rework barriers and adjust scope of atomics ( #1019 )
...
[ROCm/rccl commit: abf265a911
2024-01-04 08:18:48 -08:00
e3d45f9de4
Improve MSCCL algorithms ( #1023 )
...
[ROCm/rccl commit: 0a53077c9c
2024-01-03 14:51:34 -08:00
0c1f773021
rearranged how the min and max functions are part of msccl ( #1025 )
...
* rearranged how the min and max functions are part of msccl
* added more coverage on in place graph tests
[ROCm/rccl commit: f4858e14b2
2023-12-21 08:58:33 -07:00
401a104c5a
resolved conflicts, fixed the localNetCount/0 bug
...
[ROCm/rccl commit: 0d515f9388
2023-12-18 08:11:34 +00:00
e4b63a8ba0
Fully disable MSCCL when machine is not matched ( #1017 )
...
* Disable MSCCL algorithm meta loading when machine is not matched
* fully disable init
* fix potential segfault
[ROCm/rccl commit: 655742a3a6
2023-12-13 08:36:21 -08:00
918ce6c2e2
msccl: disable on multi-node ( #1018 )
...
[ROCm/rccl commit: 53d807a5b9
2023-12-13 07:41:40 -08:00
48107b18c9
msccl: fix data corruption with MTYPE_RW ( #1014 )
...
[ROCm/rccl commit: 81602814a7
2023-12-11 20:33:15 -08:00
481a35bc59
Fix memory fence and use non-temporal store ( #1007 )
...
* Fix memory fence and use non-temporal store
* Use amdgcn builtin instead of inline asm
* Move threadfence location
* Revert changes to gfx90a
* Rework gfx90a change
* Apply changes to gfx94x
[ROCm/rccl commit: 7965c8b53c
2023-12-09 12:16:08 -08:00
fe250212bf
Fix MSCCL scratch allocation ( #1010 )
...
[ROCm/rccl commit: c002f20029
2023-12-08 17:47:10 -06:00
529f72ceaa
Relax workgroup barrier implementation for MSCCL send/recv ops. ( #997 )
...
* Trim logic.
* Revert "Trim logic."
This reverts commit 8f2dba6c764108acf2bf5428366b9f41d4d206b9.
* Introduce MSCCL template parameters to send / recv.
* Address review feedbacks.
[ROCm/rccl commit: baadda4bd8
2023-12-08 17:46:53 -06:00
22730bdc4d
msccl: build same number of kernels as in ROCm 5.7 ( #1005 )
...
Removed fullOps kernels from build
[ROCm/rccl commit: 12c08fc52a
2023-12-07 13:36:04 -06:00
33aa8b67be
Use a map to host scratch buffers ( #1004 )
...
* Use a map to host scratch buffers
* Address review feedbacks. Deliberately keep mscclSetupScratch function.
[ROCm/rccl commit: 293f0fb752
2023-12-05 13:15:28 -06:00
403a91137c
Fix gcnArch bug in IFC mix build ( #998 ) ( #1002 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
[ROCm/rccl commit: bc44e3faa7
2023-12-04 16:20:22 -06:00
ae0bdad45c
IFC mix build ( #998 )
...
[ROCm/rccl commit: 7c0f49a878
2023-12-02 18:49:52 -07:00
b38b7fa3a2
Increase max channles to 64 ( #993 )
...
[ROCm/rccl commit: 4ba65d1d6a
2023-12-01 16:01:11 -08:00
bf57487384
Fix GPU to NIC mapping for peertopeer ( #994 )
...
[ROCm/rccl commit: 0b53f79196
2023-12-01 08:00:17 -08:00
cef45b8311
Fix mscclAlgoHandle not initialized issue ( #995 )
...
[ROCm/rccl commit: e44e112a17
2023-12-01 07:58:01 -08:00
f0c47d085e
Move MSCCL algorithm loading to initialization to workaround HIP graph conflict ( #982 )
...
* MSCCL: pre-specify channels and pre-load algorithms
* add mutex
* fix bug
* clean include
* disable all-gathers temporarily
[ROCm/rccl commit: 4bb0b4a380
2023-11-30 09:47:20 -08:00
bd982864d5
recreated pr 914 to work with current develop branch ( #979 )
...
[ROCm/rccl commit: 56ce9ef05f
2023-11-28 16:33:47 -07:00
dcf623f2ec
Add special handling of gfx940 ( #976 )
...
* Add special handling of gfx940
* Update ring base
[ROCm/rccl commit: 50b2dd9fd7
2023-11-22 15:07:36 -08:00
382ae08419
Merge remote-tracking branch 'origin/master'
...
[ROCm/rccl commit: b6d7438d31
2023-11-20 05:07:23 -08:00
4c2fa05a23
msccl: allocate scratch as ext-scope fine-grained ( #968 )
...
[ROCm/rccl commit: 569d3f7d59
2023-11-16 09:57:25 -06:00
7c0920cd62
Fix kernel command line warnings ( #961 )
...
* Fix kernel command line warnings
* Remove while loop
[ROCm/rccl commit: bc8661f092
2023-11-15 18:01:12 -08:00
6ce074d92d
Fix MSCCL work FIFO allocation with HIP graph enabled ( #967 )
...
[ROCm/rccl commit: 7fc891bc8d
2023-11-15 16:43:28 -08:00
9e8eb41337
Check to support older ROCm versions ( #963 )
...
[ROCm/rccl commit: 198f14923b
2023-11-15 12:36:31 -07:00
2351578d5b
Optimize MSCCL all-gather algorithms for gfx942 ( #964 )
...
[ROCm/rccl commit: 7ae95db5b8
2023-11-15 08:18:59 -08:00
2c6eededec
Optimize MSCCL reduce primitive switching for gfx942 ( #962 )
...
* Optimize reduce primitive switching for gfx942
* address comment
[ROCm/rccl commit: df128879a6
2023-11-15 08:18:44 -08:00
c78787e089
Fix use of CPUID overwriting registers in use.
...
CPUID writes to EAX, EBX, ECX, and EDX so the inline-asm must state that.
Otherwise currently in-use register might get overwritten which may
cause all kinds of failures like segfaults or wrong results.
Alternatively `__cpuid` can be used which avoids this and related issues.
So do that as suggested in the GCC issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112513
[ROCm/rccl commit: cece6415b0
2023-11-14 12:38:02 +01:00
b16882a024
2.19.4-1
...
Split transport connect phase into multiple steps to avoid port
exhaustion when connecting alltoall at large scale. Defaults to 128
peers per round.
Fix memory leaks on CUDA graph capture.
Fix alltoallv crash on self-sendrecv.
Make topology detection more deterministic when PCI speeds are not
available (fix issue #1020 ).
Properly close shared memory in NVLS resources.
Revert proxy detach after 5 seconds.
Add option to print progress during transport connect.
Add option to set NCCL_DEBUG to INFO on first WARN.
[ROCm/rccl commit: 88d44d777f
2023-11-13 10:36:12 -08:00
534af85d0f
msccl: enable basic collective trace ( #959 )
...
To avoid increasing number of kernels, colltrace is only enabled with
RCCL_MSCCL_FORCE_FULLOPS=1
[ROCm/rccl commit: 5a800e00cd
2023-11-08 20:14:28 -08:00
270aa41f6b
Use send instead of sendWithBarrier. ( #727 )
...
[ROCm/rccl commit: efc42d9045
2023-11-07 13:47:24 -06:00
83a36c65c1
Merge pull request #950 from nusislam/msccl-red2
...
msccl: remove cases from numReduction switch statement
[ROCm/rccl commit: 022735d208
2023-11-04 02:48:03 -05:00
aa02d2b675
Use parallel init of LDS and adjust P2P channels for gfx94x ( #943 )
...
* Use parallel init of LDS and adjust P2P channels for gfx94x
* Move another init to parallel
* Fix NCCL_NCHANNELS_PER_PEER setting
[ROCm/rccl commit: dbcba2923b
2023-11-03 16:06:49 -07:00
b6f47bad7c
msccl: remove cases from numReduction switch statement
...
[ROCm/rccl commit: f545b94d4b
2023-11-03 16:56:51 -05:00
1557a1f258
msccl: use 32-bit LDS access and add RCCL_MSCCL_FORCE_FULLOPS ( #953 )
...
[ROCm/rccl commit: bb84345943
2023-11-03 10:38:02 -07:00
4cd86b185c
MSCCL stream fix ( #948 )
...
[ROCm/rccl commit: 988efe605a
2023-11-03 09:10:52 -06:00
297023a7a6
msccl: add templated kernel ( #945 )
...
* msccl: add templated kernel
* Use defines to improve code readability
* Fix kernel indexing and review feedback
[ROCm/rccl commit: f484ff17b9
2023-11-02 17:21:53 -07:00
88c8bb1495
msccl: remove dereference of reduce args
...
It can be removed because the msccl kernel will never execute this code
according to the current msccl setup.
[ROCm/rccl commit: 6b80a0d0d4
2023-11-02 13:20:00 -05:00
3eeaea3f00
msccl: use atomic to set dependency flags ( #941 )
...
[ROCm/rccl commit: a7400218a2
2023-10-31 14:46:57 -07:00
b736b506c0
NPkit: misc fixes for MSCCL ( #936 )
...
* msccl: add xcc_id to timestamp sync
* NPKit: add timestamp for rrc operator
* NPKit: add timestamp for MSCCL init
[ROCm/rccl commit: a497722894
2023-10-30 10:00:12 -07:00
8b1254a4f1
Fix gcnArchName bug in topology dump ( #937 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
[ROCm/rccl commit: 1e5ca6820b
2023-10-28 12:30:36 -05:00
e1dfb82023
Fix MSCCL work FIFO out-of-bound issue ( #935 )
...
[ROCm/rccl commit: 4c117e5335
2023-10-27 11:24:52 -07:00
b4ba7cc79d
SRC/INIT: fix typo for ENABLE_PROFILING ( #934 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
[ROCm/rccl commit: 96ec3ffe2e
2023-10-26 23:52:46 -05:00
706750597c
remove gcnArch support ( #920 )
...
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com >
[ROCm/rccl commit: f22df90e5c
2023-10-26 12:09:15 -05:00
446c8cbf66
msccl: reduce debug output when using NCCL_DEBUG=INFO ( #932 )
...
[ROCm/rccl commit: fb0eccb57b
2023-10-25 08:05:19 -07:00
cc4de02a86
Add missing gfx942 support ( #927 )
...
[ROCm/rccl commit: c4e65fd382
2023-10-23 12:04:37 -07:00
6f0f614d0b
Remove LDS based software barriers from MSCCL ( #923 )
...
[ROCm/rccl commit: dbb5611a3a
2023-10-19 16:39:41 -05:00
Wenkai Du
Ziyue Yang
akolliasAMD
PedramAlizadeh
Wen-Heng (Jack) Chung
Nilesh M Negi
Bertan Dogancay
pradeep-ramanna
Sylvain Jeaugey
Alexander Grund
Nusrat Islam