gtests: graph: unit: add gtests for verifying fill_local_in_map

oneapi-src · Dec 25, 2024 · 22fdb3a · 22fdb3a
1 parent 880bf7d
commit 22fdb3a
Show file tree

Hide file tree

Showing 2 changed files with 112 additions and 3 deletions.
diff --git a/tests/gtests/graph/unit/utils.hpp b/tests/gtests/graph/unit/utils.hpp
@@ -144,13 +144,14 @@ static inline std::vector<int64_t> compute_dense_strides(
 }
 
 static inline std::vector<dnnl::impl::graph::logical_tensor_t>
-create_logical_tensors(
-        size_t num_lt, impl::data_type_t dtype = impl::data_type::f32) {
+create_logical_tensors(size_t num_lt,
+        impl::data_type_t dtype = impl::data_type::f32,
+        size_t id_start_from = 0) {
     size_t count = 0;
     std::vector<dnnl::impl::graph::logical_tensor_t> lt_vec;
     lt_vec.reserve(num_lt);
     while (count < num_lt) {
-        lt_vec.emplace_back(logical_tensor_init(count, dtype));
+        lt_vec.emplace_back(logical_tensor_init(id_start_from + count, dtype));
         count++;
     }
     return lt_vec;

diff --git a/tests/gtests/graph/unit/utils/test_pattern_matcher_cpu.cpp b/tests/gtests/graph/unit/utils/test_pattern_matcher_cpu.cpp
@@ -1338,6 +1338,114 @@ TEST(test_utils_pattern_matcher, SharedInput) {
     ASSERT_EQ(fusion_ops2.size(), 0U);
 }
 
+TEST(test_utils_pattern_matcher, SharedInputCase2) {
+    /* Pattern that captures shared input to two MatMuls
+
+         Dequant  Dequant
+        IN0  | IN0  |
+          \ /    \ /
+        MatMul  MatMul
+             \ /
+           Multiply
+              |
+    */
+    auto graphp = std::make_shared<pb_graph_t>();
+    auto pdequant1 = graphp->append_op(DynamicDequantize);
+    auto pmm1 = graphp->append_op(MatMul, {in_edge(1, pdequant1, 0)});
+    auto pdequant2 = graphp->append_op(DynamicDequantize);
+    auto pmm2 = graphp->append_op(MatMul, {in_edge(1, pdequant2, 0)});
+    graphp->create_input_port(0, pmm1, 0);
+    graphp->create_input_port(0, pmm2, 0);
+    auto pmul = graphp->append_op(
+            Multiply, {in_edge(0, pmm1, 0), in_edge(1, pmm2, 0)});
+    UNUSED(pmul);
+
+    // test with a graph that has the shared input
+    graph_t agraph;
+    op_t dequant1 {0, DynamicDequantize, "dequant1"};
+    op_t matmul1 {1, MatMul, "matmul1"};
+    op_t dequant2 {2, DynamicDequantize, "dequant2"};
+    op_t matmul2 {3, MatMul, "matmul2"};
+    op_t multiply {4, Multiply, "multiply"};
+
+    std::vector<logical_tensor_t> lt_vec = create_logical_tensors(8);
+    std::vector<logical_tensor_t> lt_vec_s8
+            = create_logical_tensors(4, data_type::s8, 8);
+    dequant1.add_input(lt_vec_s8[0]);
+    dequant1.add_input(lt_vec[0]);
+    dequant1.add_input(lt_vec_s8[1]);
+    dequant1.add_output(lt_vec[1]);
+
+    matmul1.add_input(lt_vec[2]);
+    matmul1.add_input(lt_vec[1]);
+    matmul1.add_output(lt_vec[3]);
+
+    dequant2.add_input(lt_vec_s8[2]);
+    dequant2.add_input(lt_vec[4]);
+    dequant2.add_input(lt_vec_s8[3]);
+    dequant2.add_output(lt_vec[5]);
+
+    matmul2.add_input(lt_vec[2]);
+    matmul2.add_input(lt_vec[5]);
+    matmul2.add_output(lt_vec[6]);
+
+    multiply.add_input(lt_vec[3]);
+    multiply.add_input(lt_vec[6]);
+    multiply.add_output(lt_vec[7]);
+
+    ASSERT_EQ(agraph.add_op(&dequant1), status::success);
+    ASSERT_EQ(agraph.add_op(&matmul1), status::success);
+    ASSERT_EQ(agraph.add_op(&dequant2), status::success);
+    ASSERT_EQ(agraph.add_op(&matmul2), status::success);
+    ASSERT_EQ(agraph.add_op(&multiply), status::success);
+    ASSERT_EQ(agraph.finalize(), status::success);
+
+    std::vector<op_t *> fusion_ops;
+    EXPECT_TRUE(match_pattern(agraph.get_ops()[0].get(), graphp, fusion_ops));
+    ASSERT_EQ(fusion_ops.size(), 5U);
+
+    // test with a graph that does not have the shared input
+    graph_t agraph2;
+    op_t dequant3 {0, DynamicDequantize, "dequant1"};
+    op_t matmul3 {1, MatMul, "matmul1"};
+    op_t dequant4 {2, DynamicDequantize, "dequant2"};
+    op_t matmul4 {3, MatMul, "matmul2"};
+    op_t multiply2 {4, Multiply, "multiply"};
+
+    std::vector<logical_tensor_t> lt_vec2 = create_logical_tensors(9);
+    std::vector<logical_tensor_t> lt_vec2_s8
+            = create_logical_tensors(4, data_type::s8, 9);
+    dequant3.add_input(lt_vec2_s8[0]);
+    dequant3.add_input(lt_vec2[0]);
+    dequant3.add_input(lt_vec2_s8[1]);
+    dequant3.add_output(lt_vec2[1]);
+    matmul3.add_input(lt_vec2[2]);
+    matmul3.add_input(lt_vec2[1]);
+    matmul3.add_output(lt_vec2[3]);
+    dequant4.add_input(lt_vec2_s8[2]);
+    dequant4.add_input(lt_vec2[4]);
+    dequant4.add_input(lt_vec2_s8[3]);
+    dequant4.add_output(lt_vec2[5]);
+    matmul4.add_input(lt_vec2[8]);
+    matmul4.add_input(lt_vec2[5]);
+    matmul4.add_output(lt_vec2[6]);
+    multiply2.add_input(lt_vec2[3]);
+    multiply2.add_input(lt_vec2[6]);
+    multiply2.add_output(lt_vec2[7]);
+
+    ASSERT_EQ(agraph2.add_op(&dequant3), status::success);
+    ASSERT_EQ(agraph2.add_op(&matmul3), status::success);
+    ASSERT_EQ(agraph2.add_op(&dequant4), status::success);
+    ASSERT_EQ(agraph2.add_op(&matmul4), status::success);
+    ASSERT_EQ(agraph2.add_op(&multiply2), status::success);
+    agraph2.finalize();
+
+    std::vector<op_t *> fusion_ops2;
+    EXPECT_FALSE(
+            match_pattern(agraph2.get_ops()[0].get(), graphp, fusion_ops2));
+    ASSERT_EQ(fusion_ops2.size(), 0U);
+}
+
 TEST(test_utils_pattern_matcher, ParallelMatmul) {
     auto graphp = std::make_shared<pb_graph_t>();
     // Pattern that captures shared input to three MatMuls