2.8.3-1

Optimization for Tree allreduce on A100.
Improve aggregation performance.
Use shared buffers for inter-node send/recv.
Add NVTX profiling hooks.
Accelerate alltoall connections by merging communication for all
channels.
Add support for one hop communication through NVLink, for faster
send/recv communication on cubemesh topologies like DGX-1.
Improve alltoall scheduling to better balance intra/inter node
communication.
Increase send/recv parallelism by 8x, each warp sending or
receiving to a different peer.
Net: move to v4.
Net: make flush operation asynchronous to accelerate alltoall.
Net: define maximum number of requests.
Fix hang when using LL128 protocol after 2^31 steps.
Fix #379 : topology injection failing when using less GPUs than
described in the XML.
Fix #394 : protocol mismatch causing hangs or crashes when using
one GPU per node.

src/include/devcomm.h

@@ -12,7 +12,7 @@
 #include <stdint.h>
 
 #define NCCL_NUM_FUNCTIONS 5 // SendRecv not included for now
-typedef enum { ncclCollBroadcast, ncclCollReduce, ncclCollAllGather, ncclCollReduceScatter, ncclCollAllReduce, ncclCollSendRecv} ncclFunc_t;
+typedef enum { ncclFuncBroadcast, ncclFuncReduce, ncclFuncAllGather, ncclFuncReduceScatter, ncclFuncAllReduce, ncclFuncSendRecv} ncclFunc_t;
 extern const char* ncclFuncStr[NCCL_NUM_FUNCTIONS];
 
 #define NCCL_NUM_ALGORITHMS 3 // Tree/Ring/CollNet
@@ -47,8 +47,9 @@ union ncclLLFifoLine {
 
 #define WARP_SIZE 32
 #define MAXCHANNELS 32
-#define NCCL_MAX_NTHREADS 512
-#define NCCL_LL_MAX_NTHREADS NCCL_MAX_NTHREADS
+#define NCCL_MAX_NTHREADS 640
+#define NCCL_SIMPLE_MAX_NTHREADS 512
+#define NCCL_LL_MAX_NTHREADS 512
 #define NCCL_LL_LINES_PER_THREAD 8
 #ifdef TEST_LL_CLEANUP
 #define NCCL_LL_CLEAN_MASK 0x078 // Set to 0x100 to disable cleanup
@@ -85,9 +86,11 @@ struct ncclConnInfo {
   uint64_t *head;     // Local for send, remote for recv
 
   int direct;         // Direct communication
+  int shared;         // Buffers are shared
   void **ptrExchange; // Pointer exchange for direct communication
 
-  int *fifo;          // Size fifo for proxy
+  int *sizesFifo;     // Sizes fifo from GPU to proxy
+  void* *ptrsFifo;      // Buffer fifo from proxy to GPU
 
   uint64_t step;      // Keep where we are
   uint64_t llLastCleaning;
@@ -129,60 +132,52 @@ struct ncclPeer {
 
 struct ncclDevComm;
 
-/* CollectiveArgs + ncclColl are to be a power of two, currently 64 bytes, */
-/* to make sure reads to host from the CUDA kernel are aligned. */
-/* Make sure to adjust padding at the end of ncclColl. */
-struct CollectiveArgs {
-  struct ncclDevComm* comm;
+#define NCCL_MAX_WORK_ELEMENTS 8
+#define NCCL_MAX_GROUPS (NCCL_MAX_WORK_ELEMENTS*2)
+
+/* ncclWork is to be a power of two, currently 8x64 bytes, */
+/* to make sure reads to host from the CUDA kernel are aligned. */
+/* Make sure to adjust padding at the end of ncclWorkElem. */
+struct ncclWorkElem {
+  // Header
+  struct ncclDevComm* comm;
+  uint16_t nThreads;
+  uint16_t funcIndex;
+  uint16_t index;
+  uint16_t active;
 
-  // local and remote input, output, and buffer
   const void * sendbuff;
   void * recvbuff;
 
-  // Op-specific fields. Make sure the common part stays the
-  // same on all structs of the union
+  // Op-specific fields.
   union {
     struct {
-      uint16_t nThreads;
-    } common;
-    struct {
-      uint16_t nThreads;
-      uint8_t bid;
-      uint8_t nChannels;
-      uint32_t root;
       size_t count;
       size_t lastChunkSize;
+      uint32_t root;
+      uint8_t bid;
+      uint8_t nChannels;
     } coll;
     struct {
-      uint16_t nThreads;
-      uint16_t unused;
-      int32_t delta;
       size_t sendCount;
       size_t recvCount;
+      int32_t delta;
+      uint16_t nThreads;
     } p2p;
+    uint64_t align[4];
   };
 };
-struct ncclColl {
-  union {
-    struct {
-      struct CollectiveArgs args;
-      uint16_t funcIndex;
-      uint16_t nextIndex;
-      uint8_t  active;
-    };
-    int data[0x10];
-  };
+struct ncclWork {
+  struct ncclWorkElem elems[NCCL_MAX_WORK_ELEMENTS];
 };
-static_assert(sizeof(struct ncclColl) == (0x10*sizeof(int)), "ncclColl must have a pow2 size");
+static_assert(sizeof(struct ncclWorkElem) == (0x10*sizeof(int)), "ncclWorkElem must have a pow2 size");
 
 struct ncclChannel {
   union {
     struct {
       struct ncclRing ring;
-      struct ncclTree treeUp;
-      struct ncclTree treeDn;
-      struct ncclTree collTreeUp;
-      struct ncclTree collTreeDn;
+      struct ncclTree tree;
+      struct ncclTree collTree;
 
       int id;
 
@@ -191,11 +186,9 @@ struct ncclChannel {
       struct ncclPeer* devPeers;
 
       // Operation list for aggregation
-      struct ncclColl* collectives;
-      int collStart;
-      int collCount;
-      int collFifoHead; // Only used by GPU
-      int collFifoTail; // Only used by CPU
+      struct ncclWork* workFifo;
+      int workCount;
+      uint64_t workFifoTail; // Only used by CPU
     };
     int data[0x80];
   };