[INIT] Fix fallback for unsupported user-specified runtime unroll factor (#1780)

* [INIT] Fix fallback for unsupported user-specified runtime unroll factor
* Add CollTrace guard
* Move `commSetUnrollFactor()` to rccl_wrap.cc
* Modify comments in the device-code generator script

src/include/enqueue.h

@@ -17,6 +17,8 @@
 #define NCCL_SIMPLE_ALIGNMENT (WARP_SIZE * 8LL * 16LL)
 #define NCCL_BYTES_ALIGNMENT 16
 
+void* rcclGetKernelIndex(int unroll, bool useCollTrace, struct ncclTaskColl* task = NULL);
+
 ncclResult_t ncclInitKernelsForDevice(int cudaArch, int maxSharedMem, size_t* maxStackSize);
 ncclResult_t ncclEnqueueCheck(struct ncclInfo* info);
 ncclResult_t ncclLaunchPrepare(struct ncclComm* comm);

src/include/rccl_common.h

@@ -23,6 +23,8 @@ THE SOFTWARE.
 #define RCCL_COMMON_H_
 #include "nccl_common.h"
 #include "nccl.h"
+#include "param.h"
+
 typedef enum RcclTunableColls {
   RCCL_UNSUPPORTED_TUNABLE = -1,
   RCCL_RS_TUNABLE = 0,    // reduce_scatter index
@@ -78,4 +80,5 @@ ncclResult_t rcclGetAlgoInfo(struct ncclComm* comm, ncclFunc_t coll, uint64_t co
                              int* algo, int* protocol, int* maxChannels);
 
 ncclResult_t rcclFuncMaxSendRecvCount(ncclFunc_t func, int nRanks, size_t count, size_t& maxCount);
-#endif
+ncclResult_t commSetUnrollFactor(struct ncclComm* comm);
+#endif