[IE CLDNN] Disabled fusing into eltwise node for fsv32 layout (#2844)

diff --git a/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp b/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp
index 97b26c7..f316a73 100644 (file)
--- a/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp
+++ b/inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp
@@ -378,6 +378,16 @@ void prepare_primitive_fusing::fuse_simple_primitives(program_impl &p) {
             return false;
         };
 
+        auto eltwise_supports_fusings = [&](eltwise_node& node) -> bool {
+            auto out_layout = node.get_output_layout();
+            if (out_layout.data_type == data_types::f16 && out_layout.size.batch[0] > 1 &&
+                (_lo.get_optimization_attributes().fs_b_yx_fsv32_network || out_layout.format == format::fs_b_yx_fsv32)) {
+                return false;
+            }
+
+            return true;
+        };
+
         auto fuse_activation_f = [&](activation_node& activation_node) {
             auto& input_data = activation_node.get_dependency(0);
             if (input_data.get_users().size() != 1 || activation_node.get_dependencies().size() >= 3)
@@ -425,7 +435,7 @@ void prepare_primitive_fusing::fuse_simple_primitives(program_impl &p) {
 
             should_fuse |= input_data.is_type<scale>();
 
-            should_fuse |= input_data.is_type<eltwise>();
+            should_fuse |= input_data.is_type<eltwise>() && eltwise_supports_fusings(input_data.as<eltwise>());
 
             if (!should_fuse)
                 return;
@@ -484,7 +494,7 @@ void prepare_primitive_fusing::fuse_simple_primitives(program_impl &p) {
 
             should_fuse |= input_data.is_type<scale>();
 
-            should_fuse |= input_data.is_type<eltwise>();
+            should_fuse |= input_data.is_type<eltwise>() && eltwise_supports_fusings(input_data.as<eltwise>());
 
             if (!should_fuse)
                 return;
@@ -574,7 +584,7 @@ void prepare_primitive_fusing::fuse_simple_primitives(program_impl &p) {
                            reduce_supports_fusings(input_data.as<reduce>())
                            && quantize_node.get_scale_shift_opt();
 
-            should_fuse |= input_data.is_type<eltwise>() && quantize_node.get_scale_shift_opt();
+            should_fuse |= input_data.is_type<eltwise>() && eltwise_supports_fusings(input_data.as<eltwise>()) && quantize_node.get_scale_shift_opt();
 
             should_fuse |= input_data.is_type<scale>() && quantize_node.get_scale_shift_opt();
 
@@ -610,7 +620,7 @@ void prepare_primitive_fusing::fuse_simple_primitives(program_impl &p) {
                                       (parents[i]->is_type<gemm>() && gemm_supports_fusings(parents[i]->as<gemm>())) ||
                                       (parents[i]->is_type<batch_to_space>()) ||
                                       (parents[i]->is_type<space_to_batch>()) ||
-                                      (parents[i]->is_type<eltwise>()) ||
+                                      (parents[i]->is_type<eltwise>() && eltwise_supports_fusings(parents[i]->as<eltwise>())) ||
                                       (parents[i]->is_type<scale>()) ||
                                       (parents[i]->is_type<depth_to_space>() && dts_supports_fusings(parents[i]->as<depth_to_space>())) ||
                                       (parents[i]->is_type<reduce>() && reduce_supports_fusings(parents[i]->as<reduce>()));

diff --git a/inference-engine/thirdparty/clDNN/tests/test_cases/fusings_gpu_test.cpp b/inference-engine/thirdparty/clDNN/tests/test_cases/fusings_gpu_test.cpp
index da8847f..e76f85d 100644 (file)
--- a/inference-engine/thirdparty/clDNN/tests/test_cases/fusings_gpu_test.cpp
+++ b/inference-engine/thirdparty/clDNN/tests/test_cases/fusings_gpu_test.cpp
@@ -6231,6 +6231,7 @@ struct eltwise_test_params {
 #define CASE_ELTWISE_FP16_1         {2, 16, 4, 4}, data_types::f16, data_types::f16, format::bfyx,           data_types::f16,  format::bfyx,             eltwise_mode::sum
 #define CASE_ELTWISE_FP16_2         {2, 16, 4, 4}, data_types::f16, data_types::f16, format::bfzyx,          data_types::f16,  format::bfzyx,            eltwise_mode::sum
 #define CASE_ELTWISE_FP16_3         {2, 32, 4, 8}, data_types::f16, data_types::f16, format::b_fs_yx_fsv16,  data_types::f16,  format::b_fs_yx_fsv16,    eltwise_mode::sum
+#define CASE_ELTWISE_FP16_4         {3, 32, 4, 4}, data_types::f16, data_types::f16, format::fs_b_yx_fsv32,  data_types::f16,  format::fs_b_yx_fsv32,    eltwise_mode::sum
 #define CASE_ELTWISE_I8_1           {2, 16, 4, 4}, data_types::i8,  data_types::i8,  format::bfyx,           data_types::f32,  format::bfyx,             eltwise_mode::sum
 #define CASE_ELTWISE_I8_2           {2, 16, 4, 4}, data_types::i8,  data_types::i8,  format::bfzyx,          data_types::f32,  format::bfzyx,            eltwise_mode::sum
 #define CASE_ELTWISE_I8_3           {2, 16, 4, 4}, data_types::i8,  data_types::i8,  format::b_fs_yx_fsv16,  data_types::f32,  format::b_fs_yx_fsv16,    eltwise_mode::sum
@@ -6386,6 +6387,49 @@ INSTANTIATE_TEST_CASE_P(fusings_gpu,
                             eltwise_test_params{CASE_ELTWISE_FP32_3, 3, 5},
                         }), );
 
+class eltwise_fp32_fsv32 : public EltwiseFusingTest {};
+TEST_P(eltwise_fp32_fsv32, add) {
+    auto p = GetParam();
+    create_topologies(input_layout("input", get_input_layout(p)),
+                      input_layout("input2", get_input_layout2(p)),
+                      data("add_data", get_mem(get_per_channel_layout(p), -10, 10)),
+                      eltwise("eltwise", {"input", "input2"}, p.mode, p.default_type),
+                      eltwise("add", {"eltwise", "add_data"}, eltwise_mode::sum),
+                      activation("activation", "add", activation_func::negative),
+                      reorder("out", "activation", p.default_format, data_types::f32));
+
+    implementation_desc eltw_impl = { format::fs_b_yx_fsv32, "eltwise_fs_b_yx_fsv32" };
+    bo_fused.set_option(build_option::force_implementations({ {"eltwise", eltw_impl} }));
+
+    tolerance = 1e-5f;
+    execute(p);
+}
+
+TEST_P(eltwise_fp32_fsv32, add_per_element) {
+    auto p = GetParam();
+    create_topologies(input_layout("input", get_input_layout(p)),
+                      input_layout("input2", get_input_layout2(p)),
+                      data("add_data", get_mem(get_input_layout(p), -10, 10)),
+                      eltwise("eltwise", {"input", "input2"}, p.mode, p.default_type),
+                      eltwise("add", {"eltwise", "add_data"}, eltwise_mode::sum),
+                      activation("activation", "add", activation_func::negative),
+                      reorder("out", "activation", p.default_format, data_types::f32));
+
+    implementation_desc eltw_impl = { format::fs_b_yx_fsv32, "eltwise_fs_b_yx_fsv32" };
+    bo_fused.set_option(build_option::force_implementations({ {"eltwise", eltw_impl} }));
+
+    tolerance = 1e-5f;
+    execute(p);
+}
+
+INSTANTIATE_TEST_CASE_P(fusings_gpu,
+                        eltwise_fp32_fsv32,
+                        ::testing::ValuesIn(std::vector<eltwise_test_params>{
+                            // There's no optimized eltwise kernel yet for fsv32 layout that supports fused_ops
+                            // So only activation is fused via legacy mechanism
+                            eltwise_test_params{CASE_ELTWISE_FP16_4, 4, 5},
+                        }), );
+
 class eltwise_fp32_fused_prims : public EltwiseFusingTest {};
 TEST_P(eltwise_fp32_fused_prims, scale_activation) {
     auto p = GetParam();