recreated pr 914 to work with current develop branch (#979)

src/collectives/device/msccl_kernel_impl.h

@@ -504,6 +504,14 @@ __global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, Simple, fullOps)(struct
   MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, double, fullOps) \
   MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, rccl_bfloat16, fullOps)
 
+#define MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_NOFLOAT(devredop) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, int8_t, fullOps) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, uint8_t, fullOps) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, int32_t, fullOps) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, uint32_t, fullOps) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, int64_t, fullOps) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, uint64_t, fullOps)
+
 #define MSCCL_IMPL_KERNEL_ENTRY_FUNC() \
   MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Sum, false) \
   MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Prod, false) \
@@ -512,6 +520,8 @@ __global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, Simple, fullOps)(struct
   MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Sum, true) \
   MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Prod, true) \
   MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Max, true) \
-  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Min, true)
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(Min, true) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP(PreMulSum, true) \
+  MSCCL_IMPL_KERNEL_ENTRY_FUNC_DEVREDOP_NOFLOAT(SumPostDiv, true)
 
 #endif

src/include/msccl/msccl_kernel.h

@@ -28,6 +28,14 @@ __global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, proto, fullOps)(struct n
   MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, double, fullOps) \
   MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, rccl_bfloat16, fullOps)
 
+#define MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_NOFLOAT(devredop, fullOps) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, int8_t, fullOps) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, uint8_t, fullOps) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, int32_t, fullOps) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, uint32_t, fullOps) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, int64_t, fullOps) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_TYPE(devredop, uint64_t, fullOps)
+
 #define MSCCL_DECL_KERNEL_ENTRY_FUNC() \
   MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(Sum, false) \
   MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(Prod, false) \
@@ -36,7 +44,9 @@ __global__ void MSCCL_KERNEL_ENTRY_NAME(devredop, type, proto, fullOps)(struct n
   MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(Sum, true) \
   MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(Prod, true) \
   MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(Max, true) \
-  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(Min, true)
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(Min, true) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP(PreMulSum, true) \
+  MSCCL_DECL_KERNEL_ENTRY_FUNC_DEVREDOP_NOFLOAT(SumPostDiv, true)
 
 MSCCL_DECL_KERNEL_ENTRY_FUNC()
 

src/misc/msccl/msccl_lifecycle.cc

@@ -215,10 +215,11 @@ static ncclResult_t mscclInternalSchedulerSelectAlgo(struct mscclSchedulerParam*
   mscclStatus& status = mscclGetStatus();
   param->scheduled = false;
 
-  // Current MSCCL doesn't support pre/post op
+  
+  /*// Current MSCCL doesn't support pre/post op
   if (param->op >= ncclAvg) {
     return ncclSuccess;
-  }
+  }*/
 
   // Whether the algorithm is in-place
   bool isInPlace = false;

src/misc/msccl/msccl_setup.cc

@@ -309,6 +309,18 @@ static ncclResult_t hostToDevRedOp(
   MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, double, fullOps), \
   MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, rccl_bfloat16, fullOps)
 
+#define MSCCL_KERNEL_ENTRY_DEVREDOP_NOFLOAT(devredop, fullOps) \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, int8_t, fullOps), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, uint8_t, fullOps), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, int32_t, fullOps), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, uint32_t, fullOps), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, int64_t, fullOps), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, uint64_t, fullOps), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_NULL()
+
 #define MSCCL_KERNEL_ENTRY() \
   MSCCL_KERNEL_ENTRY_DEVREDOP(Sum, false), \
   MSCCL_KERNEL_ENTRY_DEVREDOP(Prod, false), \
@@ -317,10 +329,12 @@ static ncclResult_t hostToDevRedOp(
   MSCCL_KERNEL_ENTRY_DEVREDOP(Sum, true), \
   MSCCL_KERNEL_ENTRY_DEVREDOP(Prod, true), \
   MSCCL_KERNEL_ENTRY_DEVREDOP(Max, true), \
-  MSCCL_KERNEL_ENTRY_DEVREDOP(Min, true)
+  MSCCL_KERNEL_ENTRY_DEVREDOP(Min, true), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP(PreMulSum, true), \
+  MSCCL_KERNEL_ENTRY_DEVREDOP_NOFLOAT(SumPostDiv, true)
 
 // Except for ncclDevPreMulSum and ncclDevSumPostDiv required by ncclAvg
-void* mscclKernelEntries[(ncclNumDevRedOps - 2) * ncclNumTypes * NCCL_NUM_PROTOCOLS * 2] = {
+void* mscclKernelEntries[ncclNumDevRedOps * ncclNumTypes * NCCL_NUM_PROTOCOLS * 2] = {
 #ifdef COMPILE_MSCCL_KERNEL
   MSCCL_KERNEL_ENTRY()
 #endif