* SWDEV-546178 - hipModuleLoadFatBinary hip header
* SWDEV-546178 - rebase and fix minor comments
* SWDEV-546178 - clean up header function declaration
[ROCm/hip commit: 4a4959b49e]
* SWDEV-546178 - hipModuleLoadFatBinary hip header
* SWDEV-546178 - rebase and fix minor comments
* SWDEV-546178 - clean up header function declaration
* change gfx950 algo/proto selection for multinode allreduce, allgather, reduceScatter
* gfx950 tuning: enable tuning for broadcast, allreduce starts LL128 earlier and switches to ring earlier, change LL128 start for allgather and reduceScatter
* lower LL128 threshold
* update reduceScatter LL128 min to match LL max for consistency
* enable multinode PXN and increase chunksize for gfx950
* change LL128 start to 128KB, adjust ring-start according to node-count
* disable code-path for fused-AR on LL128 for gfx950
* use LL128 starting from 1KB for multinode allgather on gfx950
* start LL128 earlier for multinode reduceScatter on gfx950
* start LL128 earlier for multinode broadcast on gfx950
* set multinode allreduce to start simple on 64MB for gfx950
* start LL128 from 1KB for multinode broadcast on gfx950
* setting multinode AR to use tree instead of ring at 16MB, 64MB, 128MB
* set multinode broadcast to use LL for up to 256KB depending on node-count for gfx950
* adjust algo for 32MB multinode allreduce on gfx950
* make 32MB tree LL128 for multinode AR on gfx950
* make sure ring is not picked on 2N allreduce on small sizes
[ROCm/rccl commit: 44121db890]
* change gfx950 algo/proto selection for multinode allreduce, allgather, reduceScatter
* gfx950 tuning: enable tuning for broadcast, allreduce starts LL128 earlier and switches to ring earlier, change LL128 start for allgather and reduceScatter
* lower LL128 threshold
* update reduceScatter LL128 min to match LL max for consistency
* enable multinode PXN and increase chunksize for gfx950
* change LL128 start to 128KB, adjust ring-start according to node-count
* disable code-path for fused-AR on LL128 for gfx950
* use LL128 starting from 1KB for multinode allgather on gfx950
* start LL128 earlier for multinode reduceScatter on gfx950
* start LL128 earlier for multinode broadcast on gfx950
* set multinode allreduce to start simple on 64MB for gfx950
* start LL128 from 1KB for multinode broadcast on gfx950
* setting multinode AR to use tree instead of ring at 16MB, 64MB, 128MB
* set multinode broadcast to use LL for up to 256KB depending on node-count for gfx950
* adjust algo for 32MB multinode allreduce on gfx950
* make 32MB tree LL128 for multinode AR on gfx950
* make sure ring is not picked on 2N allreduce on small sizes
* SWDEV-546485 Port and clean up for hipPerfBufferCopyRectSpeed
* SWDEV-546485 Port and clean up for hipPerfDevMemReadSpeed
* SWDEV-546485 Port and clean up for hipPerfDevMemWriteSpeed
* SWDEV-546485 Port and clean up for hipPerfHostNumaAlloc
* SWDEV-546485 Port and clean up for hipPerfMemcpy
* SWDEV-546485 Port and clean up for hipPerfMemMallocCpyFree
* SWDEV-546485 Port and clean up for hipPerfMemset
* SWDEV-546485 Port and clean up for hipPerfSampleRate
* SWDEV-546485 Port and clean up for hipPerfSharedMemReadSpeed
* SWDEV-546485 Ported and fixed up segfault for hipPerfMemFill
* SWDEV-545485 Returning to unedited stage
[ROCm/hip-tests commit: 04469c0cde]
* SWDEV-546485 Port and clean up for hipPerfBufferCopyRectSpeed
* SWDEV-546485 Port and clean up for hipPerfDevMemReadSpeed
* SWDEV-546485 Port and clean up for hipPerfDevMemWriteSpeed
* SWDEV-546485 Port and clean up for hipPerfHostNumaAlloc
* SWDEV-546485 Port and clean up for hipPerfMemcpy
* SWDEV-546485 Port and clean up for hipPerfMemMallocCpyFree
* SWDEV-546485 Port and clean up for hipPerfMemset
* SWDEV-546485 Port and clean up for hipPerfSampleRate
* SWDEV-546485 Port and clean up for hipPerfSharedMemReadSpeed
* SWDEV-546485 Ported and fixed up segfault for hipPerfMemFill
* SWDEV-545485 Returning to unedited stage
* SWDEV-543981 new kernel latency test with different timing modes and taking multiple iterations of same test
* SWDEV-543981 cleanup
* SWDEV-543981 removed outdated hit test
* SWDEV-543981 Updated timing kernel
[ROCm/hip-tests commit: d227a8110c]
* SWDEV-543981 new kernel latency test with different timing modes and taking multiple iterations of same test
* SWDEV-543981 cleanup
* SWDEV-543981 removed outdated hit test
* SWDEV-543981 Updated timing kernel
Implemented reference counting to manage init and shutdown processes,
allowing for multiple initializations and shutdowns.
[ROCm/amdsmi commit: c45a53d751]
Leverages the traits of extended-scope fine-grain memory to get rid of a device-scope acquire-release fence. This improves throughput for single node workloads on gfx942 and gfx950 for some input sizes (e.g., ~32 MiB to about 256 MiB) when using the simple protocol. Multinode workloads on MI300X see a smaller but statistically significant uplift for some message sizes. Runtime disablement is supported via setting the environment variable RCCL_GFX942_CHEAP_FENCE_ON to 0.
[ROCm/rccl commit: 1aa2570b48]
Leverages the traits of extended-scope fine-grain memory to get rid of a device-scope acquire-release fence. This improves throughput for single node workloads on gfx942 and gfx950 for some input sizes (e.g., ~32 MiB to about 256 MiB) when using the simple protocol. Multinode workloads on MI300X see a smaller but statistically significant uplift for some message sizes. Runtime disablement is supported via setting the environment variable RCCL_GFX942_CHEAP_FENCE_ON to 0.
Manocha, Rahul
Nilesh M Negi
GunaShekar, Ajay
isaki001
Arandjelovic, Marko
Naeisseh, Hadi
Hernandez, Gerardo
Luo, Phoebe
Kian Cossettini
Ammar ELWazir
Jeff Jiang
Bill Liu
alex-breslow-amd
Daniel Su