Upcast FP8 to Half (FP16) for Sum Operation (#1775)

* adding hadd and hadd2 support using builtin functions.

---------

Co-authored-by: Marzieh Berenjkoub <mberenjk@amd.com>

src/device/reduce_kernel.h

@@ -304,11 +304,13 @@ SPECIALIZE_REDUCE(FuncMinMax, half, 1, half, fn.isMinNotMax ? __hmin(x, y) : __h
   SPECIALIZE_REDUCE(FuncMinMax, __nv_fp8_e5m2, 1, __nv_fp8_e5m2, __nv_fp8_e5m2(fn.isMinNotMax ? __hmin(__half(x), __half(y)) : __hmax(__half(x), __half(y))))
   SPECIALIZE_REDUCE(FuncMinMax, __nv_fp8_e5m2, 2, __nv_fp8x2_e5m2, __nv_fp8x2_e5m2(fn.isMinNotMax ? __hmin2(__half2(x), __half2(y)) : __hmax2(__half2(x), __half2(y))))
 #else
-  SPECIALIZE_REDUCE(FuncSum, rccl_float8, 1, rccl_float8, rccl_float8(float(x) + float(y)))
+  SPECIALIZE_REDUCE(FuncSum, rccl_float8, 1, rccl_float8, hadd(x,y))
+  SPECIALIZE_REDUCE(FuncSum, rccl_float8, 2, fp8x2_storage_t, hadd2(x,y))
   SPECIALIZE_REDUCE(FuncProd, rccl_float8, 1, rccl_float8, rccl_float8(float(x) * float(y)))
   SPECIALIZE_REDUCE(FuncMinMax, rccl_float8, 1, rccl_float8, rccl_float8(fn.isMinNotMax ? fminf(float(x), float(y)) : fmaxf(float(x), float(y))))
   
-  SPECIALIZE_REDUCE(FuncSum, rccl_bfloat8, 1, rccl_bfloat8, rccl_bfloat8(float(x) + float(y)))
+  SPECIALIZE_REDUCE(FuncSum, rccl_bfloat8, 1, rccl_bfloat8, hadd_b(x,y))
+  SPECIALIZE_REDUCE(FuncSum, rccl_bfloat8, 2, fp8x2_storage_t, hadd2_b(x,y))
   SPECIALIZE_REDUCE(FuncProd, rccl_bfloat8, 1, rccl_bfloat8, rccl_bfloat8(float(x) * float(y)))
   SPECIALIZE_REDUCE(FuncMinMax, rccl_bfloat8, 1, rccl_bfloat8, rccl_bfloat8(fn.isMinNotMax ? fminf(float(x), float(y)) : fmaxf(float(x), float(y))))
 #endif