IntelGPU backend: Cconvolutions operations corner cases fix (#2749)

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -1600,8 +1600,8 @@ shared_ptr<runtime::Executable>
 
             // clDNN has quite limited support for Convolution operation
             // following are the checks to go with workaround
-            if ((win_stride.size() > 2) || (pad_below.size() > 2) || (pad_above.size() > 2) ||
-                (win_dilation.size() > 2) || (data_dilation.size() > 2) ||
+            if ((win_stride.size() != 2) || (pad_below.size() != 2) || (pad_above.size() != 2) ||
+                (win_dilation.size() != 2) || (data_dilation.size() != 2) ||
                 (data_dilation.at(0) != 1) || (data_dilation.at(1) != 1) ||
                 (op->get_output_element_type(0) != element::f32))
             {
@@ -1676,10 +1676,20 @@ shared_ptr<runtime::Executable>
             const Strides& win_dilation = conv_op->get_window_movement_strides_forward();
             const Strides& data_dilation = conv_op->get_data_dilation_strides_forward();
 
-            if ((win_stride.size() > 2) || (win_stride.at(0) != 1) || (win_stride.at(1) != 1) ||
-                (pad_below.size() > 2) || (pad_above.size() > 2) || (data_dilation.size() > 2) ||
+            // workaround to use custom kernel in case of filter output NxCx1x1
+            bool proceed_with_custom_kernel = false;
+            const Shape& output_shape = op->get_output_shape(0);
+            if ((output_shape.size() == 4) && (output_shape.at(2) == 1) &&
+                (output_shape.at(3) == 1))
+            {
+                proceed_with_custom_kernel = true;
+            }
+
+            if ((win_stride.size() != 2) || (win_stride.at(0) != 1) || (win_stride.at(1) != 1) ||
+                (pad_below.size() != 2) || (pad_above.size() != 2) || (data_dilation.size() != 2) ||
                 (data_dilation.at(0) != 1) || (data_dilation.at(1) != 1) ||
-                (win_dilation.size() > 2) || (op->get_output_element_type(0) != element::f32))
+                (win_dilation.size() != 2) || (op->get_output_element_type(0) != element::f32) ||
+                proceed_with_custom_kernel)
             {
                 do_convolution_operation(topology,
                                          op->get_input_tensor_name(0),
@@ -1776,10 +1786,10 @@ shared_ptr<runtime::Executable>
             const Strides& win_dilation = conv_op->get_window_dilation_strides_forward();
             const Strides& data_dilation = conv_op->get_window_movement_strides_forward();
 
-            if ((win_stride.size() > 2) || (win_stride.at(0) != 1) || (win_stride.at(1) != 1) ||
-                (pad_below.size() > 2) || (pad_above.size() > 2) || (data_dilation.size() > 2) ||
+            if ((win_stride.size() != 2) || (win_stride.at(0) != 1) || (win_stride.at(1) != 1) ||
+                (pad_below.size() != 2) || (pad_above.size() != 2) || (data_dilation.size() != 2) ||
                 (data_dilation.at(0) != 1) || (data_dilation.at(1) != 1) ||
-                (win_dilation.size() > 2) || (win_dilation.at(0) != 1) ||
+                (win_dilation.size() != 2) || (win_dilation.at(0) != 1) ||
                 (win_dilation.at(1) != 1) || (op->get_output_element_type(0) != element::f32) ||
                 ((pad_below.at(0) == pad_above.at(0)) && (pad_below.at(1) == pad_above.at(1))))
             {