IntelGPU backend: Set of fusion operations (#2809)

* IntelGPU backend: Set of fusion operations

* avoid test failed after pr merge

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -43,7 +43,6 @@
 
 #include "ngraph/pass/algebraic_simplification.hpp"
 #include "ngraph/pass/cse.hpp"
-#include "ngraph/pass/fused_op_decomposition.hpp"
 #include "ngraph/pass/get_output_element_elimination.hpp"
 #include "ngraph/pass/manager.hpp"
 #include "ngraph/pass/nop_elimination.hpp"
@@ -75,6 +74,7 @@
 #include "ngraph/op/embedding_lookup.hpp"
 #include "ngraph/op/equal.hpp"
 #include "ngraph/op/erf.hpp"
+#include "ngraph/op/fused/conv_fused.hpp"
 #include "ngraph/op/get_output_element.hpp"
 #include "ngraph/op/greater.hpp"
 #include "ngraph/op/greater_eq.hpp"
@@ -113,6 +113,7 @@ using intelgpu_space = runtime::intelgpu::IntelGPULayout;
 #define NGRAPH_OP(a, b) a,
 enum class OP_TYPEID
 {
+#include "ngraph/op/fused_op_tbl.hpp"
 #include "ngraph/op/op_tbl.hpp"
 };
 #undef NGRAPH_OP
@@ -125,6 +126,7 @@ static OP_TYPEID get_typeid(const string& s)
 // ...
 #define NGRAPH_OP(a, b) {#a, OP_TYPEID::a},
     static const unordered_map<string, OP_TYPEID> typeid_map{
+#include "ngraph/op/fused_op_tbl.hpp"
 #include "ngraph/op/op_tbl.hpp"
     };
 #undef NGRAPH_OP
@@ -394,7 +396,6 @@ shared_ptr<runtime::Executable>
     {
         ngraph::pass::Manager pass_manager;
 
-        pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>();
         pass_manager.register_pass<ngraph::pass::NopElimination>();
         pass_manager.register_pass<ngraph::pass::AlgebraicSimplification>();
         pass_manager.register_pass<ngraph::pass::CommonSubexpressionElimination>();
@@ -413,13 +414,14 @@ shared_ptr<runtime::Executable>
 
     for (shared_ptr<Node> op : func->get_ops())
     {
+        const OP_TYPEID op_type_id = get_typeid(op->description());
 // We want to check that every OP_TYPEID enumeration is included in the list.
 // These GCC flags enable compile-time checking so that if an enumeration
 // is not in the list an error is generated.
 #pragma GCC diagnostic push
 #pragma GCC diagnostic error "-Wswitch"
 #pragma GCC diagnostic error "-Wswitch-enum"
-        switch (get_typeid(op->description()))
+        switch (op_type_id)
         {
         case OP_TYPEID::Parameter:
         {
@@ -1403,15 +1405,52 @@ shared_ptr<runtime::Executable>
             break;
         }
         case OP_TYPEID::Convolution:
+        case OP_TYPEID::ConvolutionBias:
+        case OP_TYPEID::ConvolutionBiasAdd:
+        {
+            // since bad inheritance design of these classes
+            Strides win_stride;
+            Strides win_dilation;
+            Strides data_dilation;
+            CoordinateDiff pad_below;
+            CoordinateDiff pad_above;
+
+            if (op_type_id == OP_TYPEID::ConvolutionBias)
+            {
+                arguments_check(op, 3, 1);
+
+                const shared_ptr<op::ConvolutionBias> conv_op =
+                    static_pointer_cast<op::ConvolutionBias>(op);
+                win_stride = conv_op->get_window_movement_strides();
+                win_dilation = conv_op->get_window_dilation_strides();
+                data_dilation = conv_op->get_data_dilation_strides();
+                pad_below = conv_op->get_padding_below();
+                pad_above = conv_op->get_padding_above();
+            }
+            else if (op_type_id == OP_TYPEID::ConvolutionBiasAdd)
+            {
+                arguments_check(op, 4, 1);
+
+                const shared_ptr<op::ConvolutionBiasAdd> conv_op =
+                    static_pointer_cast<op::ConvolutionBiasAdd>(op);
+                win_stride = conv_op->get_window_movement_strides();
+                win_dilation = conv_op->get_window_dilation_strides();
+                data_dilation = conv_op->get_data_dilation_strides();
+                pad_below = conv_op->get_padding_below();
+                pad_above = conv_op->get_padding_above();
+            }
+            else
             {
                 arguments_check(op, 2, 1);
 
-            const shared_ptr<op::Convolution> conv_op = static_pointer_cast<op::Convolution>(op);
-            const Strides& win_stride = conv_op->get_window_movement_strides();
-            const Strides& win_dilation = conv_op->get_window_dilation_strides();
-            const Strides& data_dilation = conv_op->get_data_dilation_strides();
-            const CoordinateDiff& pad_below = conv_op->get_padding_below();
-            const CoordinateDiff& pad_above = conv_op->get_padding_above();
+                const shared_ptr<op::Convolution> conv_op =
+                    static_pointer_cast<op::Convolution>(op);
+                win_stride = conv_op->get_window_movement_strides();
+                win_dilation = conv_op->get_window_dilation_strides();
+                data_dilation = conv_op->get_data_dilation_strides();
+                pad_below = conv_op->get_padding_below();
+                pad_above = conv_op->get_padding_above();
+            }
 
             // clDNN has quite limited support for Convolution operation
             // following are the checks to go with workaround
@@ -1420,7 +1459,19 @@ shared_ptr<runtime::Executable>
                 (data_dilation.at(0) != 1) || (data_dilation.at(1) != 1) ||
                 (op->get_output_element_type(0) != element::f32))
             {
-                kern.emit<op::Convolution>(conv_op);
+                if (op_type_id == OP_TYPEID::ConvolutionBias)
+                {
+                    kern.emit<op::ConvolutionBias>(static_pointer_cast<op::ConvolutionBias>(op));
+                }
+                else if (op_type_id == OP_TYPEID::ConvolutionBiasAdd)
+                {
+                    kern.emit<op::ConvolutionBiasAdd>(
+                        static_pointer_cast<op::ConvolutionBiasAdd>(op));
+                }
+                else
+                {
+                    kern.emit<op::Convolution>(static_pointer_cast<op::Convolution>(op));
+                }
             }
             else
             {
@@ -1450,6 +1501,42 @@ shared_ptr<runtime::Executable>
                 const cldnn::tensor strides(1, 1, win_stride.at(1), win_stride.at(0));
                 const cldnn::tensor dilation(1, 1, win_dilation.at(1), win_dilation.at(0));
 
+                if (op_type_id == OP_TYPEID::ConvolutionBias)
+                {
+                    const cldnn::convolution cldnn_conv_bias(op->get_output_tensor_name(0),
+                                                             op_input_name,
+                                                             {op->get_input_tensor_name(1)},
+                                                             {op->get_input_tensor_name(2)},
+                                                             strides,
+                                                             input_offset,
+                                                             dilation);
+                    topology.add(cldnn_conv_bias);
+                }
+                else if (op_type_id == OP_TYPEID::ConvolutionBiasAdd)
+                {
+                    // Do not understand which cldnn::convolution::ctor() should be called
+                    // make it clear by two operations
+                    const string intermediate_name =
+                        op_input_name + op->get_output_tensor_name(0) + "_intermediate";
+
+                    const cldnn::convolution cldnn_conv_bias(intermediate_name,
+                                                             op_input_name,
+                                                             {op->get_input_tensor_name(1)},
+                                                             {op->get_input_tensor_name(2)},
+                                                             strides,
+                                                             input_offset,
+                                                             dilation);
+                    topology.add(cldnn_conv_bias);
+
+                    const cldnn::eltwise cldnn_conv_bias_add(
+                        op->get_output_tensor_name(0),
+                        {intermediate_name, op->get_input_tensor_name(3)},
+                        cldnn::eltwise_mode::sum);
+
+                    topology.add(cldnn_conv_bias_add);
+                }
+                else
+                {
                     const cldnn::convolution cldnn_conv(op->get_output_tensor_name(0),
                                                         op_input_name,
                                                         {op->get_input_tensor_name(1)},
@@ -1458,6 +1545,15 @@ shared_ptr<runtime::Executable>
                                                         dilation);
                     topology.add(cldnn_conv);
                 }
+            }
+            break;
+        }
+        case OP_TYPEID::ConvolutionBiasBackpropFiltersBias:
+        {
+            arguments_check(op, 2, 2);
+
+            kern.emit<op::ConvolutionBiasBackpropFiltersBias>(
+                static_pointer_cast<op::ConvolutionBiasBackpropFiltersBias>(op));
             break;
         }
         case OP_TYPEID::ConvolutionBackpropFilters:
@@ -1839,6 +1935,8 @@ shared_ptr<runtime::Executable>
         case OP_TYPEID::DynBroadcast:
         case OP_TYPEID::Passthrough:
         case OP_TYPEID::DynPad:
+        case OP_TYPEID::PRelu:
+        default:
         {
             throw unsupported_op("Unsupported op '" + op->description() +
                                  "' in IntelGPU back end.");

src/ngraph/runtime/intelgpu/intelgpu_kernels.cpp

@@ -97,6 +97,7 @@ void runtime::intelgpu::CustomKernels::queue_krnl(const krnl_info& krnl_info,
                                                       kr.m_lws);
         stream.add(kernel_item);
 #endif
+        ++m_count_krnls;
     }
 }
 

src/ngraph/runtime/intelgpu/intelgpu_kernels.hpp

@@ -30,6 +30,7 @@
 #include "ngraph/op/broadcast.hpp"
 #include "ngraph/op/convolution.hpp"
 #include "ngraph/op/equal.hpp"
+#include "ngraph/op/fused/conv_fused.hpp"
 #include "ngraph/op/greater.hpp"
 #include "ngraph/op/greater_eq.hpp"
 #include "ngraph/op/less.hpp"
@@ -111,8 +112,6 @@ public:
         krnl_info = build_krnl(op);
 
         queue_krnl(krnl_info, op);
-
-        ++m_count_krnls;
     }
 
     size_t get_custom_kernel_count() const { return m_count_krnls; }
@@ -129,6 +128,9 @@ private:
     krnl_info build_krnl(const std::shared_ptr<op::Convolution>& op) const;
     krnl_info build_krnl(const std::shared_ptr<op::ConvolutionBackpropData>& op) const;
     krnl_info build_krnl(const std::shared_ptr<op::ConvolutionBackpropFilters>& op) const;
+    krnl_info build_krnl(const std::shared_ptr<op::ConvolutionBias>& op) const;
+    krnl_info build_krnl(const std::shared_ptr<op::ConvolutionBiasAdd>& op) const;
+    krnl_info build_krnl(const std::shared_ptr<op::ConvolutionBiasBackpropFiltersBias>& op) const;
     krnl_info build_krnl(const std::shared_ptr<op::Equal>& op) const;
     krnl_info build_krnl(const std::shared_ptr<op::Greater>& op) const;
     krnl_info build_krnl(const std::shared_ptr<op::GreaterEq>& op) const;

src/ngraph/runtime/intelgpu/unit_test.manifest

@@ -25,6 +25,10 @@ shape_of_vector
 floor_int32
 convert_int32_bool
 convert_float32_bool
+prelu
+prelu_shared_slope
+prelu_negative_slope
+group_conv
 
 # Unsupported extra padding modes
 pad_edge_1d

src/ngraph/runtime/intelgpu/visualize_tree.cpp

@@ -55,6 +55,7 @@ using namespace std;
 #define NGRAPH_OP(a, b) a,
 enum class OP_TYPEID
 {
+#include "ngraph/op/fused_op_tbl.hpp"
 #include "ngraph/op/op_tbl.hpp"
     UNDEFINED_OP
 };
@@ -68,6 +69,7 @@ static OP_TYPEID get_typeid(const string& s)
 // ...
 #define NGRAPH_OP(a, b) {#a, OP_TYPEID::a},
     static const unordered_map<string, OP_TYPEID> typeid_map{
+#include "ngraph/op/fused_op_tbl.hpp"
 #include "ngraph/op/op_tbl.hpp"
     };
 #undef NGRAPH_OP
@@ -329,6 +331,8 @@ void print_node_parameters(ostringstream& writer, const shared_ptr<Node>& node)
         break;
     }
     case OP_TYPEID::Convolution:
+    case OP_TYPEID::ConvolutionBias:
+    case OP_TYPEID::ConvolutionBiasAdd:
     {
         const shared_ptr<op::Convolution> conv_op = static_pointer_cast<op::Convolution>(node);
 
@@ -340,6 +344,7 @@ void print_node_parameters(ostringstream& writer, const shared_ptr<Node>& node)
         break;
     }
     case OP_TYPEID::ConvolutionBackpropFilters:
+    case OP_TYPEID::ConvolutionBiasBackpropFiltersBias:
     {
         const shared_ptr<op::ConvolutionBackpropFilters> conv_op_filt =
             static_pointer_cast<op::ConvolutionBackpropFilters>(node);