Fold Constant + ConvertLayout to reduce memory footprint. (#3465)

* Fold Constant + ConvertLayout.

* Address PR Feedback.

* No folding if data layout is padded.

* Add unit test.

* Fix style error.

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -1244,6 +1244,7 @@ void runtime::cpu::CPU_ExternalFunction::register_common_passes(
     REGISTER_KNOBBED_PASS_WITH_ARGS(
         CommonSubexpressionElimination, true, ngraph::pass, runtime::cpu::get_cse_handlers_map());
     REGISTER_KNOBBED_PASS(CPUPostLayoutOptimizations, true, runtime::cpu::pass);
+    REGISTER_KNOBBED_PASS(CPUConvertLayoutConstantFolding, true, runtime::cpu::pass);
     REGISTER_KNOBBED_PASS(CPUMemoryOptimization, true, runtime::cpu::pass);
     REGISTER_KNOBBED_PASS(GetOutputElementElimination, false, ngraph::pass);
     REGISTER_KNOBBED_PASS_WITH_ARGS(

src/ngraph/runtime/cpu/pass/cpu_post_layout_optimizations.hpp

@@ -27,6 +27,7 @@ namespace ngraph
             namespace pass
             {
                 class CPUPostLayoutOptimizations;
+                class CPUConvertLayoutConstantFolding;
             }
         }
     }
@@ -47,3 +48,11 @@ public:
     void construct_slice_convertLayout_fusion();
     void construct_reshape_convertLayout_fusion();
 };
+
+class CPU_BACKEND_API ngraph::runtime::cpu::pass::CPUConvertLayoutConstantFolding
+    : public ngraph::pass::FunctionPass
+{
+public:
+    CPUConvertLayoutConstantFolding() {}
+    virtual bool run_on_function(std::shared_ptr<ngraph::Function> function) override;
+};

test/cpu_test.cpp

@@ -1051,6 +1051,29 @@ TEST(cpu_test, thread_safe_calls_convolution_2d_2items)
     unset_environment("NGRAPH_CPU_CONCURRENCY");
 }
 
+TEST(cpu_test, constant_convertlayout)
+{
+    Shape data_shape{1, 64, 56, 56};
+    auto data = make_shared<op::Parameter>(element::f32, data_shape);
+    Shape weights_shape{64, 64, 3, 3};
+    test::Uniform<float> rng(-100.0f, 100.0f);
+    vector<float> values_in(shape_size(weights_shape));
+    rng.initialize(values_in);
+    auto weights = make_shared<op::Constant>(element::f32, weights_shape, values_in);
+    Shape bias_shape{64};
+    auto bias = make_shared<op::Parameter>(element::f32, bias_shape);
+
+    auto conv = std::make_shared<op::Convolution>(data, weights, Strides{1, 1}, Strides{1, 1});
+    auto convbias = make_shared<op::ConvolutionBias>(conv, bias);
+
+    auto f = make_shared<Function>(convbias, ParameterVector{data, bias});
+    auto backend = runtime::Backend::create("CPU");
+    auto handle = backend->compile(f);
+
+    size_t convert_layout = count_ops_of_type<runtime::cpu::op::ConvertLayout>(f);
+    ASSERT_EQ(convert_layout, 1);
+}
+
 TEST(cpu_test, constant_reshape)
 {
     Shape shape_in{2, 4};