gpu: nvidia: Added missing bf16 threshold for softmax

Author: ShanoToni

tests/benchdnn/softmax/softmax.cpp

@@ -232,7 +232,7 @@ void setup_cmp(compare::compare_t &cmp, const prb_t *prb, data_kind_t kind,
     const float trh_coeff_bwd = (prb->dir & FLAG_FWD) ? 1.f : 4.f;
     const float trh_f32 = trh_coeff_log * trh_coeff_bwd * trh_coeff_f32
             * epsilon_dt(trh_dt);
-#if DNNL_AARCH64 || defined(DNNL_SYCL_HIP)
+#if DNNL_AARCH64 || defined(DNNL_SYCL_HIP) || defined(DNNL_SYCL_CUDA)
     // MIOpen and ACL softmax accumulate in F16, but oneDNN now expects accumulation in
     // F32, this partially reverts 6727bbe8. For more information on ACL softmax, see
     // https://github.com/oneapi-src/oneDNN/issues/1819
@@ -243,7 +243,7 @@ void setup_cmp(compare::compare_t &cmp, const prb_t *prb, data_kind_t kind,
     const bool is_strict_acc
             = 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) && !is_strict_acc)
             ? trh_f32
             : 0.f;
 #endif