Merge pull request #24834 from fengyuentau:cuda_naryeltwise_broadcast

dnn (cuda): support broadcasting if a.rank() != b.rank() #24834

Inspired by https://github.com/opencv/opencv/pull/24786. This PR keeps the fusion of `NaryEltwise` and `Concat` while addressed the data missing problem via supporting broadcasting if a.rank() != b.rank().

Resolves https://github.com/opencv/opencv/issues/23977
Resolves https://github.com/opencv/opencv/issues/24606
Resolves https://github.com/opencv/opencv/issues/24635
Resolves https://github.com/opencv/opencv/issues/24721 

### Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake

modules/dnn/test/test_backends.cpp

@@ -102,6 +102,12 @@ public:
     Net net;
 };
 
+TEST_P(DNNTestNetwork, DISABLED_YOLOv8n) {
+    processNet("dnn/onnx/models/yolov8n.onnx", "", Size(640, 640), "output0");
+    expectNoFallbacksFromIE(net);
+    expectNoFallbacksFromCUDA(net);
+}
+
 TEST_P(DNNTestNetwork, AlexNet)
 {
     applyTestTag(CV_TEST_TAG_MEMORY_1GB);
@@ -1518,6 +1524,71 @@ INSTANTIATE_TEST_CASE_P(Layer_Test_Backends, Eltwise, testing::Combine(
                dnnBackendsAndTargets()
 ));
 
+////////////////////////////////////////////////////////////////////////////////
+// Element-wise layers
+////////////////////////////////////////////////////////////////////////////////
+using NaryEltwiseConcat = TestWithParam<tuple<std::vector<int>, tuple<Backend, Target>>>;
+TEST_P(NaryEltwiseConcat, Accuracy) {
+    auto param = GetParam();
+    std::vector<int> input_shape = get<0>(param);
+    auto backend_id = get<0>(get<1>(param));
+    auto target_id = get<1>(get<1>(param));
+
+    /* Build the following net:
+
+           <1x4x84>
+           /
+        [Input] -+-> Mul(B<1x84>) -> Concat(axis=1) -> [Output]
+                 |                     |
+                 +-> Sigmoid ----------+
+
+    */
+    Net net;
+
+    std::vector<int> mul_B_shape(input_shape.size() - 1, 1);
+    mul_B_shape.back() = input_shape.back();
+    Mat mul_B(mul_B_shape, CV_32FC1);
+    randn(mul_B, 0.f, 1.f);
+    LayerParams mul_B_lp;
+    mul_B_lp.name = "mul_B";
+    mul_B_lp.type = "Const";
+    mul_B_lp.blobs.push_back(mul_B);
+    int id_mul_B = net.addLayer(mul_B_lp.name, mul_B_lp.type, mul_B_lp);
+
+    LayerParams mul_lp;
+    mul_lp.name = "mul";
+    mul_lp.type = "NaryEltwise";
+    mul_lp.set("operation", "mul");
+    int id_mul = net.addLayer(mul_lp.name, mul_lp.type, mul_lp);
+    net.connect(0, 0, id_mul, 0);
+    net.connect(id_mul_B, 0, id_mul, 1);
+
+    LayerParams sigmoid_lp;
+    sigmoid_lp.name = "sigmoid";
+    sigmoid_lp.type = "Sigmoid";
+    int id_sigmoid = net.addLayer(sigmoid_lp.name, sigmoid_lp.type, sigmoid_lp);
+    net.connect(0, 0, id_sigmoid, 0);
+
+    LayerParams concat_lp;
+    concat_lp.name = "concat";
+    concat_lp.type = "Concat";
+    concat_lp.set("axis", 1);
+    int id_concat = net.addLayer(concat_lp.name, concat_lp.type, concat_lp);
+    net.connect(id_mul, 0, id_concat, 0);
+    net.connect(id_sigmoid, 0, id_concat, 1);
+
+    // Run test
+    Mat input(input_shape, CV_32FC1);
+    testLayer(input, net, backend_id, target_id, false);
+}
+
+INSTANTIATE_TEST_CASE_P(Layer_Test_Backends, NaryEltwiseConcat, testing::Combine(
+    testing::Values(std::vector<int>{1, 4, 84}),
+    dnnBackendsAndTargets())
+);
+
+
+
 INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_layers_backends, dnnBackendsAndTargets());
 
 }} // namespace