handle redundant slice in SliceOp

modules/dnn/src/cuda4dnn/primitives/slice.hpp

@@ -10,11 +10,13 @@
 #include "../csl/stream.hpp"
 
 #include "../kernels/slice.hpp"
+#include "../kernels/fill_copy.hpp"
 
 #include <opencv2/core.hpp>
 
 #include <cstddef>
 #include <vector>
+#include <algorithm>
 #include <utility>
 
 namespace cv { namespace dnn { namespace cuda4dnn {
@@ -43,6 +45,22 @@ namespace cv { namespace dnn { namespace cuda4dnn {
             auto input_wrapper = inputs[0].dynamicCast<wrapper_type>();
             auto input = input_wrapper->getView();
 
+            CV_Assert(offsets.size() == outputs.size());
+
+            /* one output with the same shape as the input => direct copy */
+            if (outputs.size() == 1)
+            {
+                auto output_wrapper = outputs[0].dynamicCast<wrapper_type>();
+                auto output = output_wrapper->getSpan();
+
+                if (is_shape_same(output, input))
+                {
+                    CV_Assert(std::all_of(std::begin(offsets[0]), std::end(offsets[0]), [] (std::size_t x) { return x == 0; }));
+                    kernels::copy<T>(stream, output, input);
+                    return;
+                }
+            }
+
             for (int i = 0; i < outputs.size(); ++i)
             {
                 auto output_wrapper = outputs[i].dynamicCast<wrapper_type>();