Avoid unnecessairly asking cuDNN which algorithms to use, since this is slow in cuDNN 8.0

dlib/cuda/cudnn_dlibapi.cpp

@@ -916,20 +916,24 @@ namespace dlib
         {
             DLIB_CASSERT(data.k() == filters.k());
 
-            // if the last call to setup gave the same exact settings then don't do
-            // anything.
-            if (stride_y_ == stride_y && 
+            const bool non_data_params_unchanged = 
+                stride_y_ == stride_y && 
                 stride_x_ == stride_x &&
                 padding_y_ == padding_y && 
                 padding_x_ == padding_x &&
-                data_num_samples == data.num_samples() &&
-                data_k == data.k() &&
-                data_nr == data.nr() &&
-                data_nc == data.nc() &&
                 filters_num_samples == filters.num_samples() &&
                 filters_k == filters.k() &&
                 filters_nr == filters.nr() &&
-                filters_nc == filters.nc())
+                filters_nc == filters.nc();
+
+            // if the last call to setup gave the same exact settings then don't do
+            // anything.
+            if (non_data_params_unchanged &&
+                data_num_samples == data.num_samples() &&
+                data_k == data.k() &&
+                data_nr == data.nr() &&
+                data_nc == data.nc()
+            )
             {
                 return;
             }
@@ -991,7 +995,16 @@ namespace dlib
                 tensor_descriptor dest_desc;
                 dest_desc.set_size(out_num_samples,out_k,out_nr,out_nc);
 
+                // Ask cuDNN what algorithms are best to use.  We always do this on the first call
+                // to setup().  Then if something other than the size of the input tensor changes we
+                // also ask cuDNN what to use.  Note that in newer versions of cuDNN, asking for the
+                // best algorithm is a relatively slow thing.  So it's important we don't do it
+                // unnecessarily.
+                if (!selected_algos || !non_data_params_unchanged)
+                {
+                    selected_algos = true;
                     select_best_algorithms(data, dest_desc);
+                }
 
                 CHECK_CUDNN(cudnnGetConvolutionForwardWorkspaceSize( 
                         context(),

dlib/cuda/cudnn_dlibapi.h

@@ -233,6 +233,8 @@ namespace dlib
             int forward_algo;
             int backward_data_algo;
             int backward_filters_algo;
+            // true if select_best_algorithms has been called at least once.
+            bool selected_algos = false;
 
             size_t forward_workspace_size_in_bytes;
             size_t backward_data_workspace_size_in_bytes;