Minor performance improvement

src/convolutional_kernels.cu

@@ -420,7 +420,7 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
     //if (state.use_mixed_precision) {
     int iteration_num = (*state.net.seen) / (state.net.batch*state.net.subdivisions);
     if (state.index != 0 && state.net.cudnn_half && !l.xnor && (!state.train || iteration_num > 3*state.net.burn_in) &&
-        (l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train && l.groups <= 8)
+        (l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train && l.groups <= 8 && l.size > 1)
     {
         //printf("\n CUDNN_HALF!!! state.index = %d \n", state.index);
 
@@ -671,7 +671,7 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
 //#ifdef CUDNN_HALF
     int iteration_num = (*state.net.seen) / (state.net.batch*state.net.subdivisions);
     if (state.index != 0 && state.net.cudnn_half && !l.xnor && (!state.train || iteration_num > 3*state.net.burn_in) &&
-        (l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train && l.groups <= 8)
+        (l.c / l.groups) % 8 == 0 && l.n % 8 == 0 && !state.train && l.groups <= 8 && l.size > 1)
     {
         const size_t input16_size = l.batch*l.c*l.w*l.h;
         const size_t delta16_size = l.batch*l.n*l.out_w*l.out_h;

src/parser.c

@@ -865,8 +865,8 @@ layer parse_shortcut(list *options, size_params params, network net)
         assert(params.w == net.layers[index].out_w && params.h == net.layers[index].out_h);
 
         if (params.w != net.layers[index].out_w || params.h != net.layers[index].out_h || params.c != net.layers[index].out_c)
-            fprintf(stderr, " w = %d, w2 = %d, h = %d, h2 = %d, c = %d, c2 = %d \n",
+            fprintf(stderr, " (%4d x%4d x%4d) + (%4d x%4d x%4d) \n",
-                params.w, net.layers[index].out_w, params.h, net.layers[index].out_h, params.c, params.net.layers[index].out_c);
+                params.w, params.h, params.c, net.layers[index].out_w, net.layers[index].out_h, params.net.layers[index].out_c);
     }
 
     return s;

src/shortcut_layer.c

@@ -196,12 +196,12 @@ void forward_shortcut_layer_gpu(const layer l, network_state state)
 
     //-----------
     //if (l.outputs == l.input_sizes[0])
-    //if(l.n == 1)
+    if(l.n == 1 && l.nweights == 0)
-    //{
+    {
-    //    input_shortcut_gpu(state.input, l.batch, state.net.layers[l.index].w, state.net.layers[l.index].h, state.net.layers[l.index].c,
+        input_shortcut_gpu(state.input, l.batch, state.net.layers[l.index].w, state.net.layers[l.index].h, state.net.layers[l.index].c,
-    //        state.net.layers[l.index].output_gpu, l.out_w, l.out_h, l.out_c, l.output_gpu);
+            state.net.layers[l.index].output_gpu, l.out_w, l.out_h, l.out_c, l.output_gpu);
-    //}
+    }
-    //else
+    else
     {
         shortcut_multilayer_gpu(l.outputs, l.batch, l.n, l.input_sizes_gpu, l.layers_output_gpu, l.output_gpu, state.input, l.weights_gpu, l.nweights);
     }