Add Inference Engine (IE) backend (#883)

* ie backend and manager with passing unit tests except for select/function

* fix function_call and select

* simplify implemenation by removing support for convert and select

* remove manager

src/ngraph/CMakeLists.txt

@@ -123,6 +123,7 @@ set (SRC
     runtime/aligned_buffer.cpp
     runtime/backend.cpp
     runtime/host_tensor_view.cpp
+    runtime/ie/ie_backend.cpp
     runtime/interpreter/int_backend.cpp
     runtime/interpreter/int_call_frame.cpp
     runtime/interpreter/int_external_function.cpp

src/ngraph/runtime/ie/ie_backend.cpp

+/*******************************************************************************
+* Copyright 2017-2018 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#include "ngraph/runtime/ie/ie_backend.hpp"
+
+#include "ngraph/descriptor/layout/dense_tensor_view_layout.hpp"
+#include "ngraph/pass/assign_layout.hpp"
+#include "ngraph/pass/liveness.hpp"
+#include "ngraph/pass/manager.hpp"
+
+#include "ngraph/util.hpp"
+
+using namespace std;
+using namespace ngraph;
+
+using descriptor::layout::DenseTensorViewLayout;
+
+static bool static_init()
+{
+    runtime::Backend::register_backend("IE", make_shared<runtime::ie::IE_Backend>());
+    return true;
+};
+
+bool runtime::ie::IE_Backend::init = static_init();
+
+shared_ptr<runtime::TensorView> runtime::ie::IE_Backend::create_tensor(const element::Type& type,
+                                                                       const Shape& shape)
+{
+    return make_shared<runtime::HostTensorView>(type, shape, "external");
+}
+
+shared_ptr<runtime::TensorView> runtime::ie::IE_Backend::create_tensor(const element::Type& type,
+                                                                       const Shape& shape,
+                                                                       void* memory_pointer)
+{
+    return make_shared<runtime::HostTensorView>(type, shape, memory_pointer, "external");
+}
+
+bool runtime::ie::IE_Backend::compile(shared_ptr<Function> function)
+{
+    pass::Manager pass_manager;
+    pass_manager.register_pass<pass::AssignLayout<DenseTensorViewLayout>>();
+    pass_manager.register_pass<pass::Liveness>();
+    pass_manager.run_passes(function);
+
+    return true;
+}
+
+bool runtime::ie::IE_Backend::call(shared_ptr<Function> function,
+                                   const vector<shared_ptr<runtime::TensorView>>& outputs,
+                                   const vector<shared_ptr<runtime::TensorView>>& inputs)
+{
+    validate_call(function, outputs, inputs);
+
+    // TODO: check if function already compiled?
+    compile(function);
+
+    // convert inputs to HostTensorView
+    vector<shared_ptr<runtime::HostTensorView>> func_inputs;
+    for (auto tv : inputs)
+    {
+        func_inputs.push_back(static_pointer_cast<runtime::HostTensorView>(tv));
+    }
+
+    // convert outputs to HostTensorView
+    vector<shared_ptr<runtime::HostTensorView>> func_outputs;
+    for (auto tv : outputs)
+    {
+        func_outputs.push_back(static_pointer_cast<runtime::HostTensorView>(tv));
+    }
+
+    // map function params -> HostTensorView
+    unordered_map<descriptor::TensorView*, shared_ptr<runtime::HostTensorView>> tensor_map;
+    size_t input_count = 0;
+    for (auto param : function->get_parameters())
+    {
+        for (size_t i = 0; i < param->get_output_size(); ++i)
+        {
+            descriptor::TensorView* tv = param->get_output_tensor_view(i).get();
+            tensor_map.insert({tv, func_inputs[input_count++]});
+        }
+    }
+
+    // map function outputs -> HostTensorView
+    for (size_t output_count = 0; output_count < function->get_output_size(); ++output_count)
+    {
+        auto output = function->get_output_op(output_count);
+        if (!dynamic_pointer_cast<op::Result>(output))
+        {
+            throw ngraph_error("One of function's outputs isn't op::Result");
+        }
+        descriptor::TensorView* tv = output->get_output_tensor_view(0).get();
+        tensor_map.insert({tv, func_outputs[output_count]});
+    }
+
+    // for each ordered op in the graph
+    for (shared_ptr<Node> op : function->get_ordered_ops())
+    {
+        if (op->description() == "Parameter")
+        {
+            continue;
+        }
+
+        // get op inputs from map
+        vector<shared_ptr<runtime::HostTensorView>> op_inputs;
+        for (const descriptor::Input& input : op->get_inputs())
+        {
+            descriptor::TensorView* tv = input.get_output().get_tensor_view().get();
+            op_inputs.push_back(tensor_map.at(tv));
+        }
+
+        // get op outputs from map or create
+        vector<shared_ptr<runtime::HostTensorView>> op_outputs;
+        for (size_t i = 0; i < op->get_output_size(); ++i)
+        {
+            descriptor::TensorView* tv = op->get_output_tensor_view(i).get();
+            shared_ptr<runtime::HostTensorView> htv;
+            if (!contains_key(tensor_map, tv))
+            {
+                // the output tensor is not in the tensor map so create a new tensor
+                const Shape& shape = op->get_output_shape(i);
+                const element::Type& type = op->get_output_element_type(i);
+                string name = op->get_output_tensor(i).get_name();
+                htv = make_shared<runtime::HostTensorView>(type, shape, name);
+                tensor_map.insert({tv, htv});
+            }
+            else
+            {
+                htv = tensor_map.at(tv);
+            }
+            op_outputs.push_back(htv);
+        }
+
+        // get op type
+        element::Type type = op->get_element_type();
+        if (!op->get_inputs().empty())
+        {
+            type = op->get_inputs().at(0).get_tensor().get_element_type();
+        }
+
+        generate_calls(type, *op, op_outputs, op_inputs);
+
+        // delete any obsolete tensors
+        for (const descriptor::Tensor* t : op->liveness_free_list)
+        {
+            for (auto it = tensor_map.begin(); it != tensor_map.end(); ++it)
+            {
+                if (it->second->get_tensor().get_name() == t->get_name())
+                {
+                    tensor_map.erase(it);
+                    break;
+                }
+            }
+        }
+    }
+
+    return true;
+}
+
+void runtime::ie::IE_Backend::generate_calls(const element::Type& type,
+                                             Node& op,
+                                             const vector<shared_ptr<HostTensorView>>& outputs,
+                                             const vector<shared_ptr<HostTensorView>>& inputs)
+{
+    if (type == element::boolean)
+    {
+        op_engine<char>(op, outputs, inputs);
+    }
+    else if (type == element::f32)
+    {
+        op_engine<float>(op, outputs, inputs);
+    }
+    else if (type == element::f64)
+    {
+        op_engine<double>(op, outputs, inputs);
+    }
+    else if (type == element::i8)
+    {
+        op_engine<int8_t>(op, outputs, inputs);
+    }
+    else if (type == element::i16)
+    {
+        op_engine<int16_t>(op, outputs, inputs);
+    }
+    else if (type == element::i32)
+    {
+        op_engine<int32_t>(op, outputs, inputs);
+    }
+    else if (type == element::i64)
+    {
+        op_engine<int64_t>(op, outputs, inputs);
+    }
+    else if (type == element::u8)
+    {
+        op_engine<uint8_t>(op, outputs, inputs);
+    }
+    else if (type == element::u16)
+    {
+        op_engine<uint16_t>(op, outputs, inputs);
+    }
+    else if (type == element::u32)
+    {
+        op_engine<uint32_t>(op, outputs, inputs);
+    }
+    else if (type == element::u64)
+    {
+        op_engine<uint64_t>(op, outputs, inputs);
+    }
+    else
+    {
+        stringstream ss;
+        ss << "unsupported element type " << type << " op " << op.get_name();
+        throw ngraph_error(ss.str());
+    }
+}

test/CMakeLists.txt

@@ -70,6 +70,7 @@ add_subdirectory(files)
 #================================================================================================
 # TODO add interpreter back to unit tests when it works
 set(BACKEND_NAMES ${BACKEND_NAMES} "INTERPRETER")
+set(BACKEND_NAMES ${BACKEND_NAMES} "IE")
 
 if(MKLDNN_INCLUDE_DIR)
     include_directories(SYSTEM ${MKLDNN_INCLUDE_DIR})

test/autodiff.in.cpp

@@ -829,6 +829,7 @@ TEST(${BACKEND_NAME}, backwards_log)
 
 TEST(${BACKEND_NAME}, backwards_maximum)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}"); // no convert support
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
     test::Uniform<float> rng(-1.0f, 1.0f);
@@ -847,6 +848,7 @@ TEST(${BACKEND_NAME}, backwards_maximum)
 
 TEST(${BACKEND_NAME}, backwards_minimum)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}"); // no convert support
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
 
     test::Uniform<float> rng(-1.0f, 1.0f);
@@ -1017,6 +1019,7 @@ TEST(${BACKEND_NAME}, backwards_reshape)
 
 TEST(${BACKEND_NAME}, backwards_select)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     SKIP_TEST_FOR("NNP_TESTER", "${BACKEND_NAME}");
 
@@ -1046,6 +1049,7 @@ TEST(${BACKEND_NAME}, backwards_select)
 
 TEST(${BACKEND_NAME}, backwards_select_nested)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     SKIP_TEST_FOR("NNP_TESTER", "${BACKEND_NAME}");
 

test/backend_test.in.cpp

@@ -1349,6 +1349,7 @@ TEST(${BACKEND_NAME}, notequal)
 
 TEST(${BACKEND_NAME}, select)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape{2, 2, 2};
     auto A = make_shared<op::Parameter>(element::boolean, shape);
@@ -1766,6 +1767,7 @@ TEST(${BACKEND_NAME}, broadcast_matrix_2)
 
 TEST(${BACKEND_NAME}, convert_int32_float32)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     Shape shape{2, 2};
     auto A = make_shared<op::Parameter>(element::i32, shape);
     auto f =
@@ -1784,6 +1786,7 @@ TEST(${BACKEND_NAME}, convert_int32_float32)
 
 TEST(${BACKEND_NAME}, convert_int32_bool)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     Shape shape{2, 2};
     auto A = make_shared<op::Parameter>(element::i32, shape);
     auto f = make_shared<Function>(make_shared<op::Convert>(A, element::boolean),
@@ -1802,6 +1805,7 @@ TEST(${BACKEND_NAME}, convert_int32_bool)
 
 TEST(${BACKEND_NAME}, convert_float32_bool)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     Shape shape{2, 2};
     auto A = make_shared<op::Parameter>(element::f32, shape);
     auto f = make_shared<Function>(make_shared<op::Convert>(A, element::boolean),
@@ -5144,6 +5148,7 @@ TEST(${BACKEND_NAME}, reduce_window_emulating_max_pool_2d_1channel_1image_stride
 //
 TEST(${BACKEND_NAME}, select_and_scatter_with_overlap)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_sel_a{};
     auto SEL_A = make_shared<op::Parameter>(element::f32, shape_sel_a);
@@ -5198,6 +5203,7 @@ TEST(${BACKEND_NAME}, select_and_scatter_with_overlap)
 //
 TEST(${BACKEND_NAME}, select_and_scatter_without_overlap)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_sel_a{};
     auto SEL_A = make_shared<op::Parameter>(element::f32, shape_sel_a);
@@ -5252,6 +5258,7 @@ TEST(${BACKEND_NAME}, select_and_scatter_without_overlap)
 //
 TEST(${BACKEND_NAME}, select_and_scatter_3d_without_overlap)
 {
+    SKIP_TEST_FOR("IE", "${BACKEND_NAME}");
     SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_sel_a{};
     auto SEL_A = make_shared<op::Parameter>(element::f32, shape_sel_a);