[GPU] Add left-alignment for broadcastable check of eltwise NUMPY

openvinotoolkit · Jan 17, 2025 · f992609 · f992609
1 parent 4b1f824
commit f992609
Show file tree

Hide file tree

Showing 2 changed files with 64 additions and 1 deletion.
diff --git a/src/plugins/intel_gpu/src/graph/eltwise.cpp b/src/plugins/intel_gpu/src/graph/eltwise.cpp
@@ -410,7 +410,11 @@ eltwise_inst::typed_primitive_inst(network& network, eltwise_node const& node) :
             auto base_pshape = input0_pshape;
             if (prim->broadcast_spec == ov::op::AutoBroadcastType::NUMPY &&
                 base_pshape.size() < input_pshape.size()) {
-                base_pshape.insert(base_pshape.begin(), input_pshape.size() - base_pshape.size(), 1);
+                if (use_new_shape_infer) {
+                    base_pshape.insert(base_pshape.begin(), input_pshape.size() - base_pshape.size(), 1);
+                } else {
+                    base_pshape.insert(base_pshape.end(), input_pshape.size() - base_pshape.size(), 1);
+                }
             }
 
             for (size_t d = 0; d < base_pshape.size(); ++d) {

diff --git a/src/plugins/intel_gpu/tests/unit/test_cases/eltwise_gpu_test.cpp b/src/plugins/intel_gpu/tests/unit/test_cases/eltwise_gpu_test.cpp
@@ -3439,6 +3439,65 @@ TEST(eltwise_gpu_f32, broadcast_test_dim3_dim4) {
     }
 }
 
+TEST(eltwise_gpu_f32, broadcast_test_dim3_dim4_new_shape_infer_false) {
+    auto& engine = get_test_engine();
+
+    ov::Shape in2_shape = {1, 4, 1, 1};
+    auto input2 = engine.allocate_memory({ ov::PartialShape(in2_shape), data_types::f32, format::bfyx });
+
+    std::vector<float> const_input = {
+        1.f,   0.f, 5.f, 1.5f,
+        2.f,   0.f, 6.f, 5.2f,
+        3.f,  0.5f, 7.f, 12.f,
+        4.f, -0.5f, 8.f,  8.f
+    };
+
+    set_values(input2, {
+        0.5f,   2.5f,  0.5f,  2.5f
+    });
+
+    float answers[16] = {
+        1.5, 0.5, 7.5, 4,
+        2.5, 0.5, 8.5, 7.7,
+        3.5, 1, 9.5, 14.5,
+        4.5, 0, 10.5, 10.5
+    };
+
+    ExecutionConfig config = get_test_default_config(engine);
+    config.set_property(ov::intel_gpu::allow_new_shape_infer(false));
+
+    // in1:dim3, int2:dim4
+    {
+        ov::Shape in1_shape = {2, 4, 2};
+
+        auto input = engine.allocate_memory({ ov::PartialShape(in1_shape), data_types::f32, format::bfyx });
+        set_values(input, const_input);
+
+        topology topology;
+        topology.add(input_layout("input", input->get_layout()));
+        topology.add(input_layout("input2", input2->get_layout()));
+        topology.add(eltwise("eltwise", { input_info("input"), input_info("input2") }, eltwise_mode::sum));
+
+        network network(engine, topology, config);
+
+        network.set_input_data("input", input);
+        network.set_input_data("input2", input2);
+        auto outputs = network.execute();
+
+        ASSERT_EQ(outputs.size(), size_t(1));
+        ASSERT_EQ(outputs.begin()->first, "eltwise");
+
+        auto output = outputs.at("eltwise").get_memory();
+
+        cldnn::mem_lock<float> output_ptr(output, get_test_stream());
+
+        for (int i = 0; i < 16; i++)
+        {
+            ASSERT_TRUE(are_equal(answers[i], output_ptr[i]));
+        }
+    }
+}
+
 TEST(eltwise_gpu_f16, fs_b_yx_fsv32_basic)
 {
     // Inputs are 2x2x2x2