IntelGPU backend: Concat operation implementation (#1363)

* IntelGPU backend: Concat operation implementation

* Several remarks were fixed

* Remaining remarks were fixed; List of tests for INTELGPU was updated

* PR1363: Minor Fixes

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -17,6 +17,7 @@
 #include <CPP/activation.hpp>
 #include <CPP/activation_grad.hpp>
 #include <CPP/batch_norm.hpp>
+#include <CPP/concatenation.hpp>
 #include <CPP/convolution.hpp>
 #include <CPP/data.hpp>
 #include <CPP/eltwise.hpp>
@@ -40,6 +41,7 @@
 #include "ngraph/op/avg_pool.hpp"
 #include "ngraph/op/batch_norm.hpp"
 #include "ngraph/op/broadcast.hpp"
+#include "ngraph/op/concat.hpp"
 #include "ngraph/op/constant.hpp"
 #include "ngraph/op/convolution.hpp"
 #include "ngraph/op/dot.hpp"
@@ -324,6 +326,33 @@ bool runtime::intelgpu::IntelGPUBackend::compile(shared_ptr<Function> func)
                                      reversed_axes);
             }
         }
+        else if ("Concat" == op->description())
+        {
+            if (op->get_inputs().empty() || op->get_outputs().size() != 1)
+            {
+                arguments_check(op, 1, 1);
+            }
+            const size_t ngraph_tensor_dims = get_input_shape(op, 0).size();
+            const shared_ptr<op::Concat> concat_op = static_pointer_cast<op::Concat>(op);
+            const size_t ngraph_concat_axis = concat_op->get_concatenation_axis();
+            vector<cldnn::primitive_id> inputs;
+
+            cldnn::concatenation::concatenation_axis cldnn_axis =
+                runtime::intelgpu::IntelGPULayout::get_cldnn_axis(ngraph_tensor_dims,
+                                                                  ngraph_concat_axis);
+
+            for (auto const& input : op->get_inputs())
+            {
+                const Shape& input_shape = input.get_shape();
+                if (shape_size(input_shape))
+                {
+                    inputs.push_back(input.get_tensor().get_name());
+                }
+            }
+
+            const cldnn::concatenation cldnn_concat(get_output_name(op), inputs, cldnn_axis);
+            topology.add(cldnn_concat);
+        }
         else if ("Add" == op->description())
         {
             do_eltwise_operation(topology, op, cldnn::eltwise_mode::sum);

src/ngraph/runtime/intelgpu/intelgpu_layout.cpp

@@ -131,19 +131,23 @@ cldnn::layout runtime::intelgpu::IntelGPULayout::create_cldnn_layout(
 }
 
 cldnn::concatenation::concatenation_axis
-    runtime::intelgpu::IntelGPULayout::get_cldnn_axis(size_t tensor_channel)
+    runtime::intelgpu::IntelGPULayout::get_cldnn_axis(size_t shape_size, size_t axis)
 {
-    switch (tensor_channel)
+    const size_t t_channel = shape_size - axis - 1;
+
+    switch (t_channel)
     {
-    case 0: return cldnn::concatenation::along_b;
-    case 1: return cldnn::concatenation::along_f;
-    case 2: return cldnn::concatenation::along_y;
-    case 3: return cldnn::concatenation::along_x;
+    case 0: return cldnn::concatenation::along_x;
+    case 1: return cldnn::concatenation::along_y;
+    case 2: return cldnn::concatenation::along_f;
+    case 3:
+        if (shape_size < 5)
+        {
+            return cldnn::concatenation::along_b;
+        }
+    /* no break */
     default:
-    {
-        ostringstream os;
-        os << "IntelGPULayout::get_cldnn_axis: wrong tensor channel " << tensor_channel;
-        throw invalid_argument(os.str());
-    }
+        throw invalid_argument("IntelGPULayout::get_cldnn_axis: wrong tensor channel " +
+                               to_string(t_channel));
     }
 }

src/ngraph/runtime/intelgpu/intelgpu_layout.hpp

@@ -53,7 +53,7 @@ public:
     static cldnn::tensor create_cldnn_tensor(const Shape& element_shape);
     static cldnn::tensor create_cldnn_offset(const Shape& pad_below);
     // This function converts Shape dimension_id into cldnn::concatenation id
-    static cldnn::concatenation::concatenation_axis get_cldnn_axis(size_t tensor_channel);
+    static cldnn::concatenation::concatenation_axis get_cldnn_axis(size_t shape_size, size_t axis);
 
 private:
     Strides strides;

src/ngraph/runtime/intelgpu/unit_test.manifest

@@ -17,9 +17,6 @@ backwards_avgpool_n2_c2_hw4x4_numeric
 backwards_avgpool_n2_c2_hw4x4_win_2x2_str_1x1_numeric
 backwards_batch_norm_three_outputs
 backwards_ceiling
-backwards_concat_axis_0
-backwards_concat_axis_1
-backwards_concat_vector
 backwards_cos
 backwards_cosh
 backwards_dot_scalar_tensor
@@ -56,16 +53,7 @@ batch_norm_one_output
 batch_norm_three_outputs
 broadcast_vector_rowwise_int64
 ceiling
-concat_2d_tensor
-concat_4d_tensor
-concat_5d
-concat_matrix_colwise
 concat_matrix_int64
-concat_matrix_rowwise
-concat_vector
-concat_zero_length_1d_last
-concat_zero_length_1d_middle
-concat_zero_length_4d_middle
 constant_multi_use
 convert_float32_bool
 convert_int32_bool