[GPU] Fix fc_tiled kernel for eltwise fusion and allow fusion for dyn…

…amic shape (openvinotoolkit#26502) ### Details: - Fixed bug in fc_tiled_opt kernel + eltwise fusion - Allow fc (fp precision) + eltwise fusion for dynamic case & unfuse if unsupported kernel is selected ### Tickets: - CVS-151537
mryzhov · Sep 12, 2024 · cbe2edf · cbe2edf
1 parent e8f1167
commit cbe2edf
Show file tree

Hide file tree

Showing 4 changed files with 63 additions and 4 deletions.
diff --git a/src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_primitive_fusing.cpp b/src/plugins/intel_gpu/src/graph/graph_optimizer/prepare_primitive_fusing.cpp
@@ -496,7 +496,7 @@ void prepare_primitive_fusing::fuse_simple_primitives(program &p) {
                 return true;
             } else {
                 auto in_dt = node.get_input_layout(0).data_type;
-                return data_type_traits::is_i8_u8(in_dt);
+                return node.is_dynamic() || data_type_traits::is_i8_u8(in_dt);
             }
         };
 

diff --git a/src/plugins/intel_gpu/src/graph/primitive_inst.cpp b/src/plugins/intel_gpu/src/graph/primitive_inst.cpp
@@ -2382,6 +2382,16 @@ bool primitive_inst::is_valid_fusion() const {
     if (fused_eltwise_prims.empty())
         return true;
 
+    if (_node->is_type<fully_connected>() && _node->get_preferred_impl_type() == impl_types::ocl) {
+        // TODO: Only fc_bf_tiled_kernel & ref kernel are verified for fused eltwise. To support more fc kernels for eltwise fusion
+        if (!_node->get_selected_impl())
+            return false;
+        if ((_node->get_selected_impl()->get_kernel_name().find("fully_connected_gpu_bf_tiled") == std::string::npos)
+            && (_node->get_selected_impl()->get_kernel_name().find("fully_connected_gpu_bfyx_ref") == std::string::npos)) {
+            return false;
+        }
+    }
+
     const auto& out_pshape = _impl_params->get_output_layout().get_partial_shape();
     for (auto& fd : fused_eltwise_prims) {
         auto outer_dep_idx = fd.outer_dep_start_idx;

diff --git a/...intel_gpu/src/kernel_selector/kernels/fully_connected/fully_connected_kernel_bf_tiled.cpp b/...intel_gpu/src/kernel_selector/kernels/fully_connected/fully_connected_kernel_bf_tiled.cpp
@@ -642,10 +642,10 @@ JitConstants FullyConnected_bf_tiled::GetJitConstants(const fully_connected_para
 
     if (!params.fused_ops.empty()) {
         std::vector<std::string> idx_order_scalar = { "(out_b + bi)", "(out_f + sglid)", "0", "0" };
-        std::vector<std::string> idx_order_vec = { "(out_b + bi)", "(out_f + fi + sglid)", "0", "0" };
+        std::vector<std::string> idx_order_vec = { "(out_b + bi)", "(out_f + sglid + fi * SIMD)", "0", "0" };
         if (params.outputs[0].GetLayout() == DataLayout::bfyx) {
-            idx_order_scalar = { "(out_b + bi) / OUTPUT_FEATURE_NUM", "(out_b + bi) % OUTPUT_FEATURE_NUM", "sglid", "0" };
-            idx_order_vec = { "(out_b + bi) / OUTPUT_FEATURE_NUM", "(out_b + bi) % OUTPUT_FEATURE_NUM", "sglid", "0" };
+            idx_order_scalar = { "(out_b + bi) / OUTPUT_FEATURE_NUM", "(out_b + bi) % OUTPUT_FEATURE_NUM", "(out_f + sglid)", "0" };
+            idx_order_vec = { "(out_b + bi) / OUTPUT_FEATURE_NUM", "(out_b + bi) % OUTPUT_FEATURE_NUM", "(out_f + sglid + fi * SIMD)", "0" };
         }
 
         // Simplify fused ops configuration to prevent mixed layout exception in jitter

diff --git a/src/plugins/intel_gpu/tests/unit/fusions/fully_connected_fusion_test.cpp b/src/plugins/intel_gpu/tests/unit/fusions/fully_connected_fusion_test.cpp
@@ -188,6 +188,14 @@ class FullyConnectedFusingTestOneDNN : public BaseFusingTest<fully_connected_tes
 #define CASE_FC_FP16_3D_1 { 2, 32, 3 }, { 2, 32, 16 }, { 16, 3, 1 }, data_types::f16, format::bfyx, data_types::f16, format::oiyx, data_types::f32, format::bfyx
 #define CASE_FC_FP16_3D_2 { 1, 1, 3 }, { 1, 1, 32 }, { 32, 3, 1 }, data_types::f16, format::bfyx, data_types::f16, format::oiyx, data_types::f32, format::bfyx
 
+#define DYN_CASE_FC_FP16_5 { 1, 128, 76 }, { 1, 128, 768 }, { 768, 76 }, data_types::f16, format::bfyx, data_types::f16, format::oiyx, data_types::f32, format::bfyx
+#define DYN_CASE_FC_FP16_6 { 2, 1, 76 }, { 2, 1, 768 }, { 768, 76 }, data_types::f16, format::bfyx, data_types::f16, format::oiyx, data_types::f32, format::bfyx
+#define DYN_CASE_FC_FP16_7 { 2, 128, 76 }, { 2, 128, 768 }, { 768, 76 }, data_types::f16, format::bfyx, data_types::f16, format::oiyx, data_types::f32, format::bfyx
+#define DYN_CASE_FC_FP16_3D_1 { 2, 32, 3 }, { 2, 32, 16 }, { 16, 3 }, data_types::f16, format::bfyx, data_types::f16, format::oiyx, data_types::f32, format::bfyx
+#define DYN_CASE_FC_FP16_3D_2 { 1, 1, 3 }, { 1, 1, 32 }, { 32, 3 }, data_types::f16, format::bfyx, data_types::f16, format::oiyx, data_types::f32, format::bfyx
+
+
+
 #define CASE_FC_FP16_INT4_COMP_1 { 1, 128 }, { 1, 128 }, { 128, 128 }, data_types::f16, format::bfyx, data_types::u4, format::oiyx, data_types::f16, format::bfyx
 
 /* ----------------------------------------------------------------------------------------------------- */
@@ -810,3 +818,44 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, fc_fp32_activation_relu, ::testing::Values
     fully_connected_test_params{ CASE_FC_FP32_1, 2, 3 }
 }));
 #endif
+
+class fc_fp16_eltwise_add_ocl_dynamic : public FullyConnectedFusingTest {
+public:
+    void run_test() {
+        auto p = GetParam();
+        auto test_input_layout = get_input_layout(p);
+        auto dynamic_input_layout = layout{ov::PartialShape::dynamic(test_input_layout.get_partial_shape().size()), test_input_layout.data_type, test_input_layout.format};
+        auto eltwise_data_shape = p.out_shape.size() == 3 ? ov::PartialShape{1, 1, p.out_shape[2]} : ov::PartialShape{1, p.out_shape[1]};
+        auto eltwise_data_layout = layout{eltwise_data_shape, p.default_type, p.default_format};
+        create_topologies(
+            input_layout("input", dynamic_input_layout),
+            data("weights", get_mem(get_weights_layout(p))),
+            data("bias", get_mem(get_bias_layout(p))),
+            data("eltwise_data", get_mem(eltwise_data_layout, 1, 9)),
+            fully_connected("fc_prim", input_info("input"), "weights", "bias", get_output_dim_size(p)),
+            eltwise("eltwise", { input_info("fc_prim"), input_info("eltwise_data") }, eltwise_mode::sum),
+            reorder("reorder_bfyx", input_info("eltwise"), p.default_format, data_types::f32)
+        );
+
+        tolerance = 1e-2f;
+        execute(p, true);
+    }
+};
+
+TEST_P(fc_fp16_eltwise_add_ocl_dynamic, basic) {
+    if (engine.get_device_info().supports_immad)
+        return;
+    run_test();
+}
+
+INSTANTIATE_TEST_SUITE_P(fusings_gpu, fc_fp16_eltwise_add_ocl_dynamic, ::testing::ValuesIn(std::vector<fully_connected_test_params>{
+    fully_connected_test_params{ CASE_FC_FP16_1, 2, 3 },
+    fully_connected_test_params{ CASE_FC_FP16_2, 2, 3 },
+    fully_connected_test_params{ CASE_FC_FP16_3, 2, 3 },
+    fully_connected_test_params{ CASE_FC_FP16_4, 2, 3 },
+    fully_connected_test_params{ DYN_CASE_FC_FP16_5, 2, 3 },
+    fully_connected_test_params{ DYN_CASE_FC_FP16_6, 2, 3 },
+    fully_connected_test_params{ DYN_CASE_FC_FP16_7, 2, 3 },
+    fully_connected_test_params{ DYN_CASE_FC_FP16_3D_1, 2, 3 },
+    fully_connected_test_params{ DYN_CASE_FC_FP16_3D_2, 2, 3 },
+}));