skipping the prod test for FP8 types in reduce and reduce-scatter (#111)

* skipping the prod test for FP8 types in reduce and reduce-scatter
---------

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

src/all_reduce.cu

@@ -65,8 +65,6 @@ testResult_t AllReduceRunTest(struct threadArgs* args, int root, ncclDataType_t
   ncclRedOp_t *run_ops;
   const char **run_typenames, **run_opnames;
   int type_count, op_count;
-  if((type == ncclFp8E4M3 || type == ncclFp8E5M2) && op == ncclProd)
-    return testSuccess;
 
   if ((int)type != -1) {
     type_count = 1;
@@ -90,8 +88,10 @@ testResult_t AllReduceRunTest(struct threadArgs* args, int root, ncclDataType_t
 
   for (int i=0; i<type_count; i++) {
     for (int j=0; j<op_count; j++) {
-      if((i == ncclFp8E4M3 || i == ncclFp8E5M2) && j == ncclProd)
+#if defined(RCCL_FLOAT8)
+      if((run_types[i] == ncclFp8E4M3 || run_types[i] == ncclFp8E5M2) && run_ops[j] == ncclProd)
         continue;
+#endif
       TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], run_ops[j], run_opnames[j], -1));
     }
   }

src/reduce.cu

@@ -95,6 +95,10 @@ testResult_t ReduceRunTest(struct threadArgs* args, int root, ncclDataType_t typ
 
   for (int i=0; i<type_count; i++) {
     for (int j=0; j<op_count; j++) {
+#if defined(RCCL_FLOAT8)
+      if((run_types[i] == ncclFp8E4M3 || run_types[i] == ncclFp8E5M2) && run_ops[j] == ncclProd)
+        continue;
+#endif
       for (int k=begin_root; k<=end_root; k++) {
         TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], run_ops[j], run_opnames[j], k));
       }

src/reduce_scatter.cu

@@ -92,6 +92,10 @@ testResult_t ReduceScatterRunTest(struct threadArgs* args, int root, ncclDataTyp
 
   for (int i=0; i<type_count; i++) {
     for (int j=0; j<op_count; j++) {
+#if defined(RCCL_FLOAT8)
+      if((run_types[i] == ncclFp8E4M3 || run_types[i] == ncclFp8E5M2) && run_ops[j] == ncclProd)
+        continue;
+#endif
       TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], run_ops[j], run_opnames[j], -1));
     }
   }