gpu: nvidia: Added missing bf16 threshold for softmax

oneapi-src · Nov 1, 2024 · e9d0fdb · e9d0fdb
1 parent 188ae7f
commit e9d0fdb
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/tests/benchdnn/softmax/softmax.cpp b/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