feat(cuda): add large kernel direct conv to heuristic algo chooser

GitOrigin-RevId: bc927b6df736ee807b816d1652d602742c591f8e

dnn/src/cuda/conv_bias/opr_impl.cpp

@@ -148,7 +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;
+    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;
     //! avoid bad case in cudnn, check dnn chanwise impl first
     if (is_chanwise) {
         if (prefer_dnn_lk_implbmm) {
@@ -160,6 +162,11 @@ ConvBiasForward::Algorithm* ConvBiasForwardImpl::get_algorithm_heuristic(
             if (sm_algo_pack.f32_implicit_bmm[0].is_available_attribute(
                         args, positive_attr, negative_attr, workspace_limit_in_bytes))
                 return &sm_algo_pack.f32_implicit_bmm[0];
+        } else if (
+                prefer_direct_lk &&
+                sm_algo_pack.depthwise_large_filter.is_available_attribute(
+                        args, positive_attr, negative_attr, workspace_limit_in_bytes)) {
+            return &sm_algo_pack.depthwise_large_filter;
         } else if (prefer_dnn_chanwise) {
             if (sm_algo_pack.chanwise.is_available_attribute(
                         args, positive_attr, negative_attr, workspace_limit_in_bytes))

dnn/src/cuda/convolution/opr_impl.cpp

@@ -119,7 +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 <= 2 * fh && wo <= 2 * fw && ho < 32 &&
+                                       wo < 32;
+    const bool prefer_direct_lk =
+            args.filter_meta.format == Param::Format::NCHW && fh > 10 && fw > 10;
     if (prefer_dnn_lk_implbmm) {
 #if CUDA_VERSION >= 10020
         if (sm_algo_pack.implbmm_nchw_hmma[0].is_available_attribute(
@@ -131,6 +134,12 @@ ConvolutionBackwardDataImpl::Algorithm* ConvolutionBackwardDataImpl::
             return &sm_algo_pack.implbmm_nchw_fma[0];
     }
 
+    if (prefer_direct_lk &&
+        sm_algo_pack.depthwise_large_filter.is_available_attribute(
+                args, positive_attr, negative_attr, workspace_limit_in_bytes)) {
+        return &sm_algo_pack.depthwise_large_filter;
+    }
+
     if (args.filter_meta.group > 1 &&
         sm_algo_pack.chanwise.is_available_attribute(
                 args, positive_attr, negative_attr, workspace_limit_in_bytes)) {