gpu: nvidia: Added missing bf16 threshold for softmax

Author: ShanoToni

tests/benchdnn/softmax/softmax.cpp

@@ -240,10 +240,12 @@ void setup_cmp(compare::compare_t &cmp, const prb_t *prb, data_kind_t kind,
     // minor accuracy drops observed compared to f32
     const float trh = trh_f32;
 #else
-    const bool is_strict_acc
-            = prb->attr.acc_mode == dnnl_accumulation_mode_strict;
+    const bool is_strict_acc = !is_nvidia_gpu()
+            && prb->attr.acc_mode == dnnl_accumulation_mode_strict;
     // Relaxed fp16 computation can get an ulp difference with f32 ref values.
-    const float trh = is_flt_or_dbl || (trh_dt == dnnl_f16 && !is_strict_acc)
+    const float trh = is_flt_or_dbl
+                    || ((trh_dt == dnnl_f16 || trh_dt == dnnl_bf16)
+                            && !is_strict_acc)
             ? trh_f32
             : 0.f;
 #endif