* Added useAcc as a template parameter to address the 2% performance regression in allreduceWithBias
---------
Co-authored-by: Marzieh Berenjkoub <mberenjk@amd.com>
- 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>
* 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 9d72be7b2f.
* 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>
* [SRC] Enable unroll=1 for gfx950
* Fix typo from rebase in generate.py
* Support for unroll=1 and gfx90a when building for all GPU targets
---------
Signed-off-by: nileshnegi <Nilesh.Negi@amd.com>
* Fix collective trace
* Use nontemporal for st_global
* Fix previous commit
* Add HDP flush to data receive path
* Fix previous commit
* Control flushing by NCCL_NET_FORCE_FLUSH and RCCL_NET_HDP_FLUSH
* Introduce RCCL_NET_HDP_FLUSH and RCCL_NET_GDR_FLUSH
Both are on by default. Turn both off will skip all flush will likely
result in data error.
* Enable GDR copy by default
* Remove GDR flush env var because it is disabled by GDC flush
* Output kernel collective trace at comm destroy by default
* Limit kernel timeout messages to 100
* Use system relaxed atomic for loadInt
* Refine timeout messages and use atomic for setting offset from CPU
* Add kernel trace for barrier timeout
* Add backup barrier to avoid race in atomicAdd
* Use different counters for different warps
* Rework barrier implementation
* Fix for other GFX
* Use __hip_atomic_store and __hip_atomic_load
* Fix bug in previous commit
* Don't reset barrier values in running kernel
* Update trace format
* Fix typo
* Switch back to hip_atomic_fetch_add
* Use same barrier implementation for all GFX
* Remove extra threadfence
* Turn off HDP flush by default
Please use RCCL_NET_HDP_FLUSH=1 to switch on HDP flush
* Remove unnecessary changes from alterative barrier implementation
* Added back __threadfence_block
* Revert back to threadfence for gfx other than gfx94x
Network user buffer support for collectives
* Leverage user buffer registration to achieve zero-copy
inter-node communications for Ring, NVLS and Collnet
Add RAS subsystem
* Create a RAS thread keeping track of all NCCL communicators.
* Add a ncclras tool contacting the RAS thread and getting a
report.
Add fp8 support
* Add support for e5m2 and e4m3 8-bit floating point operations.
* Use Tree/PAT algorithms when possible for better numerical
stability.
Add NIC fusion
* Add a NET API to ask the network plugin to fuse a set of
interfaces together.
* Fuse multiple NICs under the same PCI switch as a single,
larger NIC.
Socket connection failure retry
* Retry in case of socket connection failure (unreachable host)
* Avoid "Software caused connection abort" errors on retries
QP connection failure retry
* Retry in case of IB QP connection failure during ibv_modify_qp.
NET API improvements
* Allow plugins to force a flush in case data and completion
ordering is not guaranteed.
* Indicate when completion is not needed (e.g. for the LL128
protocol), allowing plugins to skip generating a completion.
* Allow for full offload of allgather operations when using one
GPU per node.
NCCL_ALGO/NCCL_PROTO strict enforcement
* Extend NCCL_ALGO/NCCL_PROTO syntax to be able to specify
ALGO/PROTO filters for each collective operation.
* Strictly enforce the ALGO/PROTO filters, no longer fall back
on the ring algorithm when the filtering leaves no option and
error out instead.
Enable CUMEM host allocations
* Use cumem functions for host memory allocation by default.
Improved profiler plugin API
* Avoid dependencies with NCCL includes.
* Add information on whether the buffer is registered or not
Adjust PAT tuning
* Improve transition between PAT and ring at scale.
Fix hangs when running with different CPU architectures
* Detect when we use a mix of GPU architectures
* Ensure Algo/Proto decisions are made based on that unified
state.
Fix FD leak in UDS
* Fix a leak when mapping buffers intra-node with cumem IPCs.
Fix crash when mixing buffer registration and graph buffer registration.
* Separate local and graph registration to avoid crashes when we free
buffers.
Fix user buffer registration with dmabuf
* Make ncclSend/ncclRecv communication with buffer registration functional
on network plugins relying on dmabuf for buffer registration.
Fix crash in IB code caused by uninitialized fields.
Fix non-blocking ncclSend/ncclRecv
* Fix case where ncclSend/ncclRecv would return ncclSuccess in non-blocking
mode even though the operation was not enqueued onto the stream.
* Issue #1495
Various compiler tweaks and fixes
* PR #758
Fix typo in ncclTopoPrintGraph
* Issue #1468
Add scalable init API
* Add new ncclCommInitRankScalable to allow for passing multiple
unique IDs to the init function.
* Spreads the load onto multiple bootstrap roots, allowing for
constant bootstrap time.
* Requires multiple ranks to create a unique ID, and the CPU-side
ID exchange code to call allgather[v] instead of broadcast.
Accelerate init bootstrap operations
* Reduce the number of calls to allgather.
* Allow roots to reply early to ranks when information is already
available.
* Add an option to use ncclNet instead of sockets to perform
bootstrap allgather operations.
Add PAT algorithms for Allgather and ReduceScatter
* Parallel Aggregated Trees, variation of Bruck algorithm.
* Logarithmic number of network steps for small sizes at scale.
* Only supports one rank per node at the moment.
Add support for registered buffers for intra-node communication.
* Allow registered user buffers to be accessed directly intra-node
* Avoids extra copies in algorithms which permit it, saving
memory bandwidth and helping with compute overlap.
Add profiler plugin API
* New plugin API for profiling
* Supports various levels of profiling, with a hierarchy.
Asynchronous graph allocation
* Make calls to cudaMalloc and cudaMemcpy during graph allocation
asynchronous.
* Significantly speeds up graph capture.
Use fatal IB asynchronous events to stop network operation
* Avoids many other error messages
* Only fatal errors are affected; potentially transient errors
(e.g. port down) do not cause an immediate stop.
Set P2P level to PXB on AMD CPUs when using more than 2 GPUs per node
* P2P would cause a significant performance degradation when using
many GPUs, and therefore many interleaved data flows.
* Disable P2P through the CPU when we have 3+ GPUs per node; keep it
enabled when we only have 2 GPUs.
Improve the init logs to report the real NCCL function.
* Make the log report ncclCommInitRank or ncclCommSplit, rather than
the generic ncclCommInitRankFunc.
Add a parameter to set the location of the user configuration file.
* Add NCCL_CONF_FILE environment variable to set where the user's
configuration file resides.
Increase default IB timeout
* Increase IB timeout value from 18 to 20.
* Should help avoid fatal errors on large RoCE systems.
Add new check for nvidia peermem
* On linux kernels 6.6+, /sys/kernel/mm/memory_peers is no longer
present; check for /sys/module/nvidia_peermem/version instead.
Fix old performance regression when mixing small and large operations.
* Improves distribution of work on channels.
Fix crash when NUMA IDs are equal to -1.
* Can happen when a NIC is a virtual NIC, or when linux doesn't
know which NUMA node a device is attached to
* Issue NVIDIA/nccl-tests#233
Fix tree graph search when NCCL_CROSS_NIC is set to 1.
* Would force NCCL to use the balanced_tree pattern, thereby
disabling LL128 on platforms with 1 GPU+1 NIC per PCI switch.
* Would also try to use alternate rings even though it was not
needed.
Compiler tweaks and fixes
* PR #1177
* PR #1228
Fix stack smash
* PR #1325
Fixes for multi-node NVLink + IB operation
Coverity fixes and comments.
Add support for alternating rings, allow for cross-nic rings without
cross-rail communication.
Add support for user buffer registration for network send/recv.
Optimize aggregated operations to better utilize all channels.
Add flattening for BCM PCI gen5 switches.
Add support for inter-node NVLink communication
Add support for port fusion in NET/IB.
Add support for ReduceScatter and AllGather using Collnet.
Update net API to v8.
Fix hang during A2A connection.
Add local user buffer registration for NVLink SHARP.
Add tuning plugin support.
Increase net API to v7 to allow for device-side packet reordering;
remove support for v4 plugins.
Add support for RoCE ECE.
Add support for C2C links.
Better detect SHM allocation failures to avoid crash with Bus Error.
Fix missing thread unlocks in bootstrap (Fixes#936).
Disable network flush by default on H100.
Move device code from src/collectives/device to src/device.
Arm Patinyasakdikul
mberenjk
Mustafa Abduljabbar
Wenkai Du
Nilesh M Negi
BertanDogancay
Bertan Dogancay
Sylvain Jeaugey