2.20.3-1

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.


[ROCm/rccl commit: b6475625fb]

projects/rccl/src/include/enqueue.h

@@ -12,8 +12,10 @@
 #include "collectives.h"
 #include "utils.h"
 
-#define NCCL_MIN_CHANNEL_SIZE (NCCL_LL_THREAD_THRESHOLD*64)
-#define NCCL_AGG_CHANNEL_SIZE (1LL << 21) /* 2 MiB, ideal per-channel size to fully utilize bandwidth */
+#define NCCL_LL_ALIGNMENT_PER_THREAD sizeof(uint64_t)
+#define NCCL_LL128_ALIGNMENT_PER_WARP 480
+#define NCCL_SIMPLE_ALIGNMENT (WARP_SIZE * 8LL * 16LL)
+#define NCCL_BYTES_ALIGNMENT 16
 
 ncclResult_t ncclInitKernelsForDevice(int cudaArch, size_t* maxStackSize);
 ncclResult_t ncclEnqueueCheck(struct ncclInfo* info);