- Introduced double-buffering to reduce copy overhead and overlap BF16 arithmetic with data prefetching.
- Aimed to improve performance of reduction-based collectives by up to 10%.
- Implemented based on recommendations from Guennadi Riguer (AMD)
- Added --force-reduce-pipeline option to install.sh to activate this optimization for BF16 reductions.
- Feature is disabled by default to prevent regressions with large messages until auto-tuning logic is upstreamed.
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
Co-authored-by: Pedram Alizadeh <pmohamma@amd.com>
[ROCm/rccl commit: 0ce20e7e07]
* Support fused all reduce and elementwise operations
Add additional "acc" parameter to RCCL Replayer logs
Add flag which indicates availability of new API
* Fix Recorder json parsing
* Remove unreachable code
* Remove extra acc pointer check
* .
* Revert "[DEVICE] Adding ability to choose unroll factor at runtime (#1734)"
This reverts commit 4cadf3597c.
* Use noinline to reduce kernels linking time
* Don't use noinline for gfx942 and gfx950 to avoid perf regression
---------
Co-authored-by: AtlantaPepsi <timhu102@amd.com>
Co-authored-by: BertanDogancay <bertan.dogancay@gmail.com>
[ROCm/rccl commit: 9a4213356d]
* Reapply "[AG and RS channel tuning] Add thread work threshold to tuning models and precompute reg index in LL128 (#1641)"
This reverts commit 943ad6f7820739385a0b54e81f823d0df1dbf71c.
* Decreasing NCCL_LL128_SHMEM_ELEMS_PER_THREAD from 16 to 8
[ROCm/rccl commit: 3f7c08648f]
* Updated 6.5 release to be 7.0
* Corrected the RCCL version for 6.4.1
* Moved items to the correct releases
* Added NCCL 2.25.1 compatibility item
* Fixed wording
* Added entry for `ManagedMem` and `ManagedMemGraph` test fix
[ROCm/rccl commit: 7b633d5844]
* Update LL128 elems per thread
* Precompute ix[g] in LL128 prim
* Make Threadthreshold part of tuning models
* Ignore channel tuning when channels are env controlled
* Tune LL128 max limit for AG
* Tune LL128 max limit for RS
* Retune AR LL128 limits due to changes
* Update CHANGELOG.md
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
[ROCm/rccl commit: 00c1eb098c]
* Fix when more than 64 channels are used for multi-collective group calls
* Update CHANGELOG.md
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
---------
Co-authored-by: Jeffrey Novotny <jnovotny@amd.com>
[ROCm/rccl commit: 9ef45df8f7]
* Added RCCL env params to control setting the SO_REUSEADDR and SO_LINGER socket options. This can allow control over the number of file descriptors created during bootstrapping.
* Casted the linger value to `int` sooner to avoid a scope of unknown typed-ness.
* Added CHANGELOG entry for this feature.
[ROCm/rccl commit: 2e35417fe5]
- Modifies the ring creation algorithm to be friendlier to rail-optimized topologies (should not affect classic fabric topologies)
[ROCm/rccl commit: 4cb62f999a]
* Reduce AlltoAll port usage when connecting proxy
Reuse socket ports when connecting proxies in AlltoAll.
Existing port usage in AlltoAll is O(n) for recv and O(n) for send,
reusing socket ports in server or client side will make one of them
O(1), reusing both will reduce the total port usage to O(1) and enables
AlltoAll in >64 MI200 nodes.
* Update changelog accordingly
Update changelog accordingly.
[ROCm/rccl commit: 80f53cc171]
* Adding the missing roc:: namespace, effectively changing the value of RCCL_LIBRARY from rccl to roc::rccl.
The important difference is that rccl is treated as a symbolic "-lrccl" by linker (and fail the linking
due to a missing library search path), while roc::rccl is a target name, which can resolve into an absolute
library path.
Co-authored-by: Paul Fultz II <pfultz2@yahoo.com>
* Adding a changelog entry
* minor updates to wording
* missing period
Co-authored-by: Paul Fultz II <pfultz2@yahoo.com>
Co-authored-by: Saad Rahim <44449863+saadrahim@users.noreply.github.com>
[ROCm/rccl commit: d5bea2cfaa]
This is a single commit of the source code changes required to
introduce support for multiple ranks per device.
A new interface (ncclCommRankInitMulti) has to be used to make use of
this new feature.
[ROCm/rccl commit: 0336ffdf70]
* 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
[ROCm/rccl commit: 9d7232c091]
Mustafa Abduljabbar
Wenkai Du
corey-derochie-amd
BertanDogancay
Nilesh M Negi
Pedram Alizadeh
Avinash
Arm Patinyasakdikul
gilbertlee-amd
Istvan Kiss
Jeffrey Novotny
Sandra Polifroni
Bertan Dogancay
Wenkai Du
Yifan Xiong
Dmitry Mikushin
Edgar
Liam Wrubleski
Stanley Tsang