diff --git a/dnn/src/cuda/conv_bias/opr_impl.cpp b/dnn/src/cuda/conv_bias/opr_impl.cpp
index 0a606e2ba6b1f9e365fc5780564f2616cdd46789..821f4e1c3d593237d61b0d592d76d0d11d20fd21 100644
--- a/dnn/src/cuda/conv_bias/opr_impl.cpp
+++ b/dnn/src/cuda/conv_bias/opr_impl.cpp
@@ -148,9 +148,9 @@ ConvBiasForward::Algorithm* ConvBiasForwardImpl::get_algorithm_heuristic(
     //! choose for large kernel cases
     size_t fh = args.filter_meta.spatial[0], fw = args.filter_meta.spatial[1];
     size_t hi = src[2], wi = src[3];
-    const bool prefer_dnn_lk_implbmm =
-            hi <= 2 * fh && wi <= 2 * fw && wi < 32 && hi <= 32;
-    const bool prefer_direct_lk = fh > 10 && fw > 10;
+    const bool prefer_dnn_lk_implbmm = hi <= 2 * fh && wi <= 2 * fw;
+    //! filter size > 9, choose large kernel cases
+    const bool prefer_direct_lk = fh > 9 && fw > 9;
     //! avoid bad case in cudnn, check dnn chanwise impl first
     if (is_chanwise) {
         if (prefer_dnn_lk_implbmm) {
diff --git a/dnn/src/cuda/convolution/opr_impl.cpp b/dnn/src/cuda/convolution/opr_impl.cpp
index 5a0109cfe6037ab1ae3267f7d9a4a024af12192c..5ef7404f8b4e1dd08c21ee3bfd3cdc12e4ef6fc1 100644
--- a/dnn/src/cuda/convolution/opr_impl.cpp
+++ b/dnn/src/cuda/convolution/opr_impl.cpp
@@ -119,10 +119,10 @@ ConvolutionBackwardDataImpl::Algorithm* ConvolutionBackwardDataImpl::
     size_t fh = args.filter_meta.spatial[0], fw = args.filter_meta.spatial[1];
     size_t ho = diff[2], wo = diff[3];
     const bool prefer_dnn_lk_implbmm = args.filter_meta.format == Param::Format::NCHW &&
-                                       ho <= 2 * fh && wo <= 2 * fw && ho < 32 &&
-                                       wo < 32;
+                                       ho <= 2 * fh && wo <= 2 * fw;
+    //! filter size > 9, choose large kernel cases
     const bool prefer_direct_lk =
-            args.filter_meta.format == Param::Format::NCHW && fh > 10 && fw > 10;
+            args.filter_meta.format == Param::Format::NCHW && fh > 9 && fw > 9;
     if (prefer_dnn_lk_implbmm) {
 #if CUDA_VERSION >= 10020
         if (sm_algo_pack.implbmm_nchw_hmma[0].is_available_attribute(