Transformer/Backend (#166)

Add a minimal Backend API and make the interpreter use it.
read/write tensors (for frameworks)

src/ngraph/CMakeLists.txt

@@ -14,9 +14,11 @@
 set (SRC
     descriptor/input.cpp
     descriptor/layout/dense_tensor_view_layout.cpp
+    descriptor/layout/tensor_view_layout.cpp
     descriptor/output.cpp
     descriptor/primary_tensor_view.cpp
     descriptor/tensor.cpp
+    descriptor/tensor_view.cpp
     descriptor/tuple.cpp
     function.cpp
     log.cpp
@@ -50,8 +52,13 @@ set (SRC
     pass/propagate_types.cpp
     pass/topological_sort.cpp
     pass/visualize_tree.cpp
-    runtime/call_frame.cpp
-    runtime/external_function.cpp
+    runtime/backend.cpp
+    runtime/manager.cpp
+    runtime/ngvm/call_frame.cpp
+    runtime/ngvm/external_function.cpp
+    runtime/ngvm/ngvm_backend.cpp
+    runtime/ngvm/ngvm_manager.cpp
+    runtime/tensor_view.cpp
     runtime/tuple.cpp
     runtime/utils.cpp
     shape.cpp

src/ngraph/descriptor/layout/dense_tensor_view_layout.cpp

@@ -15,6 +15,7 @@
 #include "ngraph/descriptor/layout/dense_tensor_view_layout.hpp"
 #include "ngraph/except.hpp"
 #include "ngraph/shape.hpp"
+#include "ngraph/types/element_type.hpp"
 
 using namespace ngraph::descriptor::layout;
 using ngraph::Shape;

src/ngraph/descriptor/layout/dense_tensor_view_layout.hpp

@@ -17,14 +17,14 @@
 #include <cstddef>
 #include <vector>
 
-#include "ngraph/descriptor/buffer.hpp"
 #include "ngraph/descriptor/layout/tensor_view_layout.hpp"
-#include "ngraph/descriptor/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace descriptor
     {
+        class TensorView;
+
         namespace layout
         {
             /// @brief The standard strided layout, used for row-major and column-major, their permutations and slices.

src/ngraph/descriptor/layout/tensor_view_layout.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include "ngraph/descriptor/layout/tensor_view_layout.hpp"
+#include "ngraph/descriptor/tensor_view.hpp"
+#include "ngraph/types/element_type.hpp"
+
+using namespace ngraph::descriptor::layout;
+
+TensorViewLayout::TensorViewLayout(const ngraph::descriptor::TensorView& tensor_view)
+    : m_tensor_view_type(tensor_view.get_tensor_view_type())
+{
+}
+
+const ngraph::element::Type& TensorViewLayout::get_element_type() const
+{
+    return m_tensor_view_type->get_element_type();
+}
+const ngraph::Shape& TensorViewLayout::get_shape() const
+{
+    return m_tensor_view_type->get_shape();
+}

src/ngraph/descriptor/layout/tensor_view_layout.hpp

@@ -23,6 +23,11 @@
 
 namespace ngraph
 {
+    namespace element
+    {
+        class Type;
+    }
+
     namespace descriptor
     {
         class TensorView;
@@ -35,10 +40,7 @@ namespace ngraph
             class TensorViewLayout
             {
             protected:
-                TensorViewLayout(const ngraph::descriptor::TensorView& tensor_view)
-                    : m_tensor_view_type(tensor_view.get_tensor_view_type())
-                {
-                }
+                TensorViewLayout(const ngraph::descriptor::TensorView& tensor_view);
 
             public:
                 virtual ~TensorViewLayout() {}
@@ -52,11 +54,8 @@ namespace ngraph
                 /// With non-linear buffers, this will need to be something other than size_t.
                 virtual size_t get_index_offset(const std::vector<size_t>& indices) = 0;
 
-                const element::Type& get_element_type() const
-                {
-                    return m_tensor_view_type->get_element_type();
-                }
-                const Shape& get_shape() const { return m_tensor_view_type->get_shape(); }
+                const element::Type& get_element_type() const;
+                const Shape& get_shape() const;
                 /// Where this view is located in the buffer.
                 const BufferPos& get_buffer_pos() const { return m_buffer_pos; }
                 BufferPos& get_buffer_pos() { return m_buffer_pos; }

src/ngraph/runtime/tensor_view_index.hpp → src/ngraph/descriptor/tensor_view.cpp

@@ -12,14 +12,12 @@
 // See the License for the specific language governing permissions and
 // ----------------------------------------------------------------------------
 
-#pragma once
-
 #include "ngraph/descriptor/tensor_view.hpp"
+#include "ngraph/types/type.hpp"
+
+using namespace ngraph::descriptor;
 
-namespace ngraph
+std::shared_ptr<const ngraph::ValueType> TensorView::get_value_type() const
 {
-    namespace runtime
-    {
-        using TensorViewIndex = unordered_map<shared_ptr<ngraph::descriptor::TensorView>, size_t>;
-    }
-}
\ No newline at end of file
+    return m_tensor_view_type;
+}

src/ngraph/descriptor/tensor_view.hpp

@@ -16,15 +16,13 @@
 
 #include <memory>
 
-#include "ngraph/descriptor/tensor.hpp"
 #include "ngraph/descriptor/value.hpp"
-#include "ngraph/log.hpp"
 #include "ngraph/shape.hpp"
-#include "ngraph/types/type.hpp"
 
 namespace ngraph
 {
     class Node;
+    class TensorViewType;
 
     namespace descriptor
     {
@@ -34,6 +32,9 @@ namespace ngraph
             class TensorViewLayout;
         }
 
+        class Tensor;
+        class TensorView;
+
         /// @brief Compile-time descriptor of a first-class value that is a view of a tensor.
         class TensorView : public Value
         {
@@ -51,10 +52,7 @@ namespace ngraph
             virtual const Tensor& get_tensor() const = 0;
             virtual Tensor& get_tensor() = 0;
 
-            virtual std::shared_ptr<const ValueType> get_value_type() const override
-            {
-                return m_tensor_view_type;
-            }
+            virtual std::shared_ptr<const ValueType> get_value_type() const override;
 
             const std::string& get_name() const { return m_name; }
             std::shared_ptr<const TensorViewType> get_tensor_view_type() const

src/ngraph/descriptor/value.hpp

@@ -15,11 +15,12 @@
 #pragma once
 
 #include <memory>
-
-#include "ngraph/types/type.hpp"
+#include <vector>
 
 namespace ngraph
 {
+    class ValueType;
+
     namespace descriptor
     {
         class TensorView;

src/ngraph/function.hpp

@@ -25,7 +25,6 @@
 #include "ngraph/node.hpp"
 #include "ngraph/ops/op.hpp"
 #include "ngraph/ops/parameter.hpp"
-#include "ngraph/runtime/instruction.hpp"
 #include "ngraph/types/type.hpp"
 
 namespace ngraph

src/ngraph/ngraph.hpp

@@ -80,9 +80,12 @@
 #include "ngraph/ops/select.hpp"
 #include "ngraph/ops/subtract.hpp"
 #include "ngraph/ops/tuple.hpp"
+#include "ngraph/runtime/backend.hpp"
 #include "ngraph/runtime/call_frame.hpp"
 #include "ngraph/runtime/external_function.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/manager.hpp"
+#include "ngraph/runtime/ngvm/ngvm_backend.hpp"
+#include "ngraph/runtime/ngvm/ngvm_manager.hpp"
 #include "ngraph/runtime/parameterized_tensor_view.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 #include "ngraph/runtime/tuple.hpp"

src/ngraph/runtime/backend.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include <memory>
+
+#include "ngraph/runtime/backend.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
+#include "ngraph/runtime/tuple.hpp"
+#include "ngraph/types/element_type.hpp"
+
+using namespace ngraph::runtime;
+
+std::shared_ptr<TensorView>
+    Backend::make_primary_tensor_view(const ngraph::element::Type& element_type, const Shape& shape)
+{
+    return element_type.make_primary_tensor_view(shape);
+}
+
+std::shared_ptr<ngraph::runtime::Tuple>
+    Backend::make_tuple(const std::vector<std::shared_ptr<ngraph::runtime::Value>>& elements)
+{
+    return std::make_shared<ngraph::runtime::Tuple>(elements);
+}

src/ngraph/runtime/eigen/scalar_tensor_product.hpp → src/ngraph/runtime/backend.hpp

@@ -14,46 +14,57 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
-#include "ngraph/runtime/tensor_view.hpp"
+#include <memory>
+
+#include "ngraph/common.hpp"
 
 namespace ngraph
 {
+    namespace element
+    {
+        class Type;
+    }
+
     namespace runtime
     {
-        namespace eigen
+        class ExternalFunction;
+        class CallFrame;
+        class TensorView;
+        class Tuple;
+        class Value;
+
+        template <typename ET>
+        class ParameterizedTensorView;
+
+        /// @brief Interface to a generic backend.
+        ///
+        /// Backends are responsible for function execution and value allocation.
+        class Backend
         {
+        public:
+            virtual ~Backend() {}
+            /// @brief Make a call frame that can support one concurrent call of an external function.
+            ///
+            /// If more than one concurrent execution is needed, each execution will require its own call frame.
+            virtual std::shared_ptr<ngraph::runtime::CallFrame>
+                make_call_frame(const std::shared_ptr<ExternalFunction>& external_function) = 0;
+
+            /// @brief Return a handle for a tensor on the backend device.
+            virtual std::shared_ptr<ngraph::runtime::TensorView>
+                make_primary_tensor_view(const ngraph::element::Type& element_type,
+                                         const Shape& shape);
+
             template <typename ET>
-            class ScalarTensorProductInstruction : public Instruction
+            std::shared_ptr<ngraph::runtime::ParameterizedTensorView<ET>>
+                make_parameterized_tensor_view(const Shape& shape)
             {
-            public:
-                ScalarTensorProductInstruction(const TensorViewInfo& arg0,
-                                               const TensorViewInfo& arg1,
-                                               const TensorViewInfo& out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
-                {
-                }
-
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    // This is a bit hacky: regardless of the tensor rank we
-                    // pull it out as a vector. This works because of the way
-                    // fmt::V computes sizes---it lumps together any higher
-                    // dimensions---while fmt::M ignores them.
-                    EigenVector<ET>(call_frame, m_out) =
-                        call_frame.get_tensor_view_data<ET>(m_arg0.get_index())[0] *
-                        EigenVector<ET>(call_frame, m_arg1);
-                }
-
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
-        }
+                return std::dynamic_pointer_cast<ngraph::runtime::ParameterizedTensorView<ET>>(
+                    make_primary_tensor_view(ET::element_type(), shape));
+            }
+
+            /// @brief Construct a tuple handle from a sequence of values.
+            virtual std::shared_ptr<ngraph::runtime::Tuple>
+                make_tuple(const std::vector<std::shared_ptr<ngraph::runtime::Value>>& elements);
+        };
     }
 }

src/ngraph/runtime/call_frame.hpp

@@ -25,51 +25,23 @@ namespace ngraph
     namespace runtime
     {
         class PrimaryTensorView;
-        class Instruction;
+        class Value;
 
         // A VM for executing lightly-compiled graph functions.
         class CallFrame
         {
         public:
-            CallFrame(
-                size_t n_inputs,
-                size_t n_outputs,
-                const TensorViewPtrs& temps,
-                size_t initial_pc,
-                const std::shared_ptr<std::vector<std::shared_ptr<Instruction>>>& instructions);
-
+            virtual ~CallFrame() {}
             /// @brief Invoke the function with values matching the signature of the function.
             ///
             /// Tuples will be expanded into their tensor views to build the call frame.
-            void operator()(const std::vector<std::shared_ptr<ngraph::runtime::Value>>& inputs,
-                            const std::vector<std::shared_ptr<ngraph::runtime::Value>>& outputs);
+            virtual void
+                operator()(const std::vector<std::shared_ptr<ngraph::runtime::Value>>& inputs,
+                           const std::vector<std::shared_ptr<ngraph::runtime::Value>>& outputs) = 0;
 
             /// @brief Invoke the function with tuples pre-expanded to their underlying tensor views.
-            void tensor_call(const TensorViewPtrs& inputs, const TensorViewPtrs& outputs);
-
-            void set_return() { m_return = true; }
-            std::shared_ptr<TensorView> get_tensor_view(size_t i) { return m_tensor_views[i]; }
-            template <typename ET>
-            ParameterizedTensorView<ET>* get_parameterized_tensor_view(size_t i)
-            {
-                return m_tensor_views[i]->get_parameterized_tensor_view<ET>();
-            }
-
-            template <typename ET>
-            typename ET::type* get_tensor_view_data(size_t i)
-            {
-                return &get_parameterized_tensor_view<ET>(i)->get_vector()[0];
-            }
-
-        protected:
-            size_t m_n_inputs;
-            size_t m_n_outputs;
-            TensorViewPtrs m_tensor_views;
-            size_t m_initial_pc;
-            std::shared_ptr<std::vector<std::shared_ptr<Instruction>>> m_instructions;
-            size_t m_pc;
-            size_t m_next_pc;
-            bool m_return;
+            virtual void tensor_call(const TensorViewPtrs& inputs,
+                                     const TensorViewPtrs& outputs) = 0;
         };
     }
 }

src/ngraph/runtime/eigen/divide.hpp

-// ----------------------------------------------------------------------------
-// Copyright 2017 Nervana Systems Inc.
-// 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
-// ----------------------------------------------------------------------------
-
-#pragma once
-
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
-#include "ngraph/runtime/tensor_view.hpp"
-
-namespace ngraph
-{
-    namespace runtime
-    {
-        namespace eigen
-        {
-            template <typename ET>
-            class DivideInstruction : public Instruction
-            {
-            public:
-                DivideInstruction(const TensorViewInfo& arg0,
-                                  const TensorViewInfo& arg1,
-                                  const TensorViewInfo& out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
-                {
-                }
-
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) =
-                        EigenArray1d<ET>(call_frame, m_arg0) / EigenArray1d<ET>(call_frame, m_arg1);
-                }
-
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
-        }
-    }
-}

src/ngraph/runtime/eigen/dot.hpp

-// ----------------------------------------------------------------------------
-// Copyright 2017 Nervana Systems Inc.
-// 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
-// ----------------------------------------------------------------------------
-
-#pragma once
-
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
-#include "ngraph/runtime/tensor_view.hpp"
-
-namespace ngraph
-{
-    namespace runtime
-    {
-        namespace eigen
-        {
-            template <typename ET>
-            class DotInstruction : public Instruction
-            {
-            public:
-                DotInstruction(const TensorViewInfo& arg0,
-                               const TensorViewInfo& arg1,
-                               const TensorViewInfo& out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
-                {
-                }
-
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out)
-                        << EigenVector<ET>(call_frame, m_arg0)
-                               .dot(EigenVector<ET>(call_frame, m_arg1));
-                }
-
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
-        }
-    }
-}

src/ngraph/runtime/eigen/maximum.hpp

-// ----------------------------------------------------------------------------
-// Copyright 2017 Nervana Systems Inc.
-// 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
-// ----------------------------------------------------------------------------
-
-#pragma once
-
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
-#include "ngraph/runtime/tensor_view.hpp"
-
-namespace ngraph
-{
-    namespace runtime
-    {
-        namespace eigen
-        {
-            template <typename ET>
-            class MaximumInstruction : public Instruction
-            {
-            public:
-                MaximumInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
-                {
-                }
-
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) =
-                        EigenArray1d<ET>(call_frame, m_arg0)
-                            .max(EigenArray1d<ET>(call_frame, m_arg1));
-                }
-
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
-        }
-    }
-}

src/ngraph/runtime/eigen/utils.hpp

-// ----------------------------------------------------------------------------
-// Copyright 2017 Nervana Systems Inc.
-// 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
-// ----------------------------------------------------------------------------
-
-#pragma once
-
-#include <memory>
-
-#include <Eigen/Dense>
-
-#include "ngraph/descriptor/layout/dense_tensor_view_layout.hpp"
-#include "ngraph/runtime/tensor_view_info.hpp"
-
-namespace ngraph
-{
-    namespace runtime
-    {
-        class TensorViewInfo;
-        class CallFrame;
-
-        namespace eigen
-        {
-            using DynamicStrides = Eigen::Stride<Eigen::Dynamic, Eigen::Dynamic>;
-            using VectorStrides = Eigen::Stride<Eigen::Dynamic, 1>;
-
-            template <typename ET>
-            using DynamicArray = Eigen::Array<typename ET::type, Eigen::Dynamic, Eigen::Dynamic>;
-
-            template <typename ET>
-            using EigenArrayBase = Eigen::Map<DynamicArray<ET>, 0, DynamicStrides>;
-
-            template <typename ET>
-            using DynamicMatrix =
-                Eigen::Matrix<typename ET::type, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
-
-            template <typename ET>
-            using EigenMatrixBase = Eigen::Map<DynamicMatrix<ET>, 0, DynamicStrides>;
-
-            template <typename ET>
-            using DynamicVector = Eigen::Matrix<typename ET::type, Eigen::Dynamic, 1>;
-
-            template <typename ET>
-            using EigenVectorBase = Eigen::Map<DynamicVector<ET>, 0, VectorStrides>;
-
-            namespace fmt
-            {
-                /// @brief vector format for Eigen wrappers.
-                class V
-                {
-                public:
-                    V(const TensorViewInfo& tensor_view_info)
-                        : l0(tensor_view_info
-                                 .get_layout<ngraph::descriptor::layout::DenseTensorViewLayout>()
-                                 ->get_size())
-                    {
-                    }
-
-                public:
-                    size_t l0;
-                    size_t l1{1};
-                    size_t s0{1};
-                    size_t s1{1};
-                };
-
-                class M
-                {
-                    M(const Shape& shape, const Strides& strides)
-                        : l0(shape.at(0))
-                        , l1(shape.at(1))
-                        , s0(strides.at(0))
-                        , s1(strides.at(1))
-                    {
-                    }
-
-                    M(const std::shared_ptr<ngraph::descriptor::layout::DenseTensorViewLayout>&
-                          layout)
-                        : M(layout->get_shape(), layout->get_strides())
-                    {
-                    }
-
-                public:
-                    M(const TensorViewInfo& tensor_view_info)
-                        : M(tensor_view_info
-                                .get_layout<ngraph::descriptor::layout::DenseTensorViewLayout>())
-                    {
-                    }
-
-                public:
-                    size_t l0;
-                    size_t l1;
-                    size_t s0;
-                    size_t s1;
-                };
-            }
-
-            // ET element type
-            // FMT array format (fmt::V for vector, etc.)
-            // BASE select array/matrix
-            template <typename ET, typename FMT, typename BASE, typename STRIDES = DynamicStrides>
-            class EigenWrapper : public BASE
-            {
-                using base = BASE;
-
-            public:
-                EigenWrapper(typename ET::type* t, const FMT& fmt)
-                    : base(t, fmt.l0, fmt.l1, STRIDES(fmt.s0, fmt.s1))
-                {
-                }
-
-                EigenWrapper(
-                    typename ET::type* t,
-                    const std::shared_ptr<ngraph::descriptor::layout::DenseTensorViewLayout>&
-                        layout)
-                    : base(t, layout->get_size(), 1, DynamicStrides(1, 1))
-                {
-                }
-
-                EigenWrapper(CallFrame& call_frame, const TensorViewInfo& tensor_view_info)
-                    : EigenWrapper(
-                          call_frame.get_tensor_view_data<ET>(tensor_view_info.get_index()),
-                          FMT(tensor_view_info))
-                {
-                }
-
-                template <typename U>
-                EigenWrapper& operator=(const U& other)
-                {
-                    this->base::operator=(other);
-                    return *this;
-                }
-            };
-
-            template <typename ET, typename FMT = fmt::V>
-            using EigenArray1d = EigenWrapper<ET, FMT, EigenArrayBase<ET>>;
-
-            template <typename ET, typename FMT = fmt::M>
-            using EigenArray2d = EigenWrapper<ET, FMT, EigenArrayBase<ET>>;
-
-            template <typename ET, typename FMT = fmt::M>
-            using EigenMatrix = EigenWrapper<ET, FMT, EigenMatrixBase<ET>>;
-
-            template <typename ET, typename FMT = fmt::V>
-            using EigenVector = EigenWrapper<ET, FMT, EigenVectorBase<ET>, VectorStrides>;
-        }
-    }
-}

src/ngraph/runtime/external_function.hpp

@@ -26,45 +26,29 @@ namespace ngraph
 {
     namespace runtime
     {
+        class CallFrame;
+
         class ExternalFunction
         {
-            using FunctionMap =
-                std::unordered_map<std::shared_ptr<Function>, std::shared_ptr<ExternalFunction>>;
-
-            using OpFunction = std::function<void(const ngraph::Node*,
-                                                  ExternalFunction*,
-                                                  FunctionMap&,
-                                                  const std::vector<TensorViewInfo>& inputs,
-                                                  const std::vector<TensorViewInfo>& outputs)>;
-            using OpMap = std::unordered_map<std::type_index, OpFunction>;
-
-        public:
+        protected:
             ExternalFunction(const std::shared_ptr<ngraph::Function>& function,
-                             bool release_function = true);
-            std::shared_ptr<ngraph::runtime::CallFrame> make_call_frame();
-            std::shared_ptr<ngraph::runtime::CallFrame> make_call_frame(FunctionMap& function_map);
-            std::shared_ptr<std::vector<std::shared_ptr<ngraph::runtime::Instruction>>>
-                get_instructions()
+                             bool release_function = true)
+                : m_function(function)
+                , m_release_function(release_function)
+                , m_is_compiled(false)
             {
-                return m_instructions;
             }
 
             // Release original function's resources
             void release_function() { m_function = nullptr; }
-        protected:
-            void compile();
-            void compile(FunctionMap& function_map);
+        public:
+            virtual ~ExternalFunction() {}
+            virtual std::shared_ptr<CallFrame> make_call_frame() = 0;
 
+        protected:
             std::shared_ptr<ngraph::Function> m_function;
             bool m_release_function;
             bool m_is_compiled;
-            size_t m_n_inputs;
-            size_t m_n_outputs;
-            std::shared_ptr<std::vector<std::shared_ptr<ngraph::runtime::Instruction>>>
-                m_instructions;
-            ngraph::descriptor::TensorViewPtrs m_temp_views;
-
-            static OpMap& get_op_map();
         };
     }
 }

src/ngraph/runtime/manager.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include "ngraph/runtime/manager.hpp"
+
+using namespace ngraph::runtime;
+
+Manager::FactoryMap& Manager::get_factory_map()
+{
+    static FactoryMap factory_map;
+    return factory_map;
+}
+
+std::shared_ptr<Manager> Manager::get(const std::string& name)
+{
+    return get_factory_map().at(name)(name);
+}
+
+Manager::Factory Manager::register_factory(std::string name, Factory factory)
+{
+    get_factory_map()[name] = factory;
+    return factory;
+}

src/ngraph/runtime/manager.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include <functional>
+#include <map>
+#include <memory>
+#include <string>
+
+namespace ngraph
+{
+    class Function;
+
+    namespace runtime
+    {
+        class Backend;
+        class ExternalFunction;
+
+        /// @brief Interface to a generic manager.
+        ///
+        /// A manager provides access to compilation for a backend, and a means to obtain
+        /// a backed for execution and allocation.
+        class Manager
+        {
+        public:
+            virtual ~Manager() {}
+            /// @brief Allocate a backend for this transformer.
+            ///
+            /// Specific transformers may provide addtional methods for allocating customized backends.
+            virtual std::shared_ptr<Backend> allocate_backend() = 0;
+
+            /// @brief Convert a function to a form that can be run on a backend.
+            virtual std::shared_ptr<ExternalFunction>
+                compile(const std::shared_ptr<ngraph::Function>& fun) = 0;
+
+            using Factory = std::function<std::shared_ptr<Manager>(const std::string&)>;
+            using FactoryMap = std::map<std::string, Factory>;
+
+            static FactoryMap& get_factory_map();
+
+            static std::shared_ptr<Manager> get(const std::string& name);
+
+            static Factory register_factory(std::string name, Factory factory);
+        };
+    }
+}

src/ngraph/runtime/ngvm/call_frame.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include <algorithm>
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
+
+using namespace std;
+using namespace ngraph::runtime::ngvm;
+
+CallFrame::CallFrame(size_t n_inputs,
+                     size_t n_outputs,
+                     const TensorViewPtrs& temps,
+                     size_t initial_pc,
+                     const shared_ptr<vector<shared_ptr<Instruction>>>& instructions)
+
+    : m_n_inputs(n_inputs)
+    , m_n_outputs(n_outputs)
+    , m_tensor_views(n_inputs + n_outputs + temps.size())
+    , m_initial_pc(initial_pc)
+    , m_instructions(instructions)
+{
+    copy(temps.begin(), temps.end(), m_tensor_views.begin() + m_n_inputs + m_n_outputs);
+}
+
+void CallFrame::tensor_call(
+    const std::vector<std::shared_ptr<ngraph::runtime::TensorView>>& inputs,
+    const std::vector<std::shared_ptr<ngraph::runtime::TensorView>>& outputs)
+{
+    copy(inputs.begin(), inputs.end(), m_tensor_views.begin());
+    copy(outputs.begin(), outputs.end(), m_tensor_views.begin() + m_n_inputs);
+    m_next_pc = m_initial_pc;
+    m_return = false;
+    while (!m_return)
+    {
+        m_pc = m_next_pc;
+        m_next_pc = m_pc + 1;
+        m_instructions->at(m_pc)->execute(*this);
+    }
+    // Don't hold onto inputs/outputs
+    fill_n(m_tensor_views.begin(), m_n_inputs + m_n_outputs, nullptr);
+}
+
+void CallFrame::operator()(const std::vector<std::shared_ptr<ngraph::runtime::Value>>& arguments,
+                           const std::vector<std::shared_ptr<ngraph::runtime::Value>>& results)
+{
+    // TODO: Check types of args and result
+    std::vector<std::shared_ptr<ngraph::runtime::TensorView>> inputs;
+    for (auto argument : arguments)
+    {
+        argument->collect_tensor_views(inputs, argument);
+    }
+
+    std::vector<std::shared_ptr<ngraph::runtime::TensorView>> outputs;
+    for (auto result : results)
+    {
+        result->collect_tensor_views(outputs, result);
+    }
+
+    tensor_call(inputs, outputs);
+}

src/ngraph/runtime/ngvm/call_frame.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include <memory>
+#include <vector>
+
+#include "ngraph/function.hpp"
+#include "ngraph/runtime/call_frame.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        class PrimaryTensorView;
+
+        namespace ngvm
+        {
+            class Instruction;
+
+            // A VM for executing lightly-compiled graph functions.
+            class CallFrame : public ngraph::runtime::CallFrame
+            {
+            public:
+                CallFrame(
+                    size_t n_inputs,
+                    size_t n_outputs,
+                    const TensorViewPtrs& temps,
+                    size_t initial_pc,
+                    const std::shared_ptr<std::vector<std::shared_ptr<Instruction>>>& instructions);
+
+                /// @brief Invoke the function with values matching the signature of the function.
+                ///
+                /// Tuples will be expanded into their tensor views to build the call frame.
+                void
+                    operator()(const std::vector<std::shared_ptr<ngraph::runtime::Value>>& inputs,
+                               const std::vector<std::shared_ptr<ngraph::runtime::Value>>& outputs);
+
+                /// @brief Invoke the function with tuples pre-expanded to their underlying tensor views.
+                void tensor_call(const TensorViewPtrs& inputs, const TensorViewPtrs& outputs);
+
+                void set_return() { m_return = true; }
+                std::shared_ptr<TensorView> get_tensor_view(size_t i) { return m_tensor_views[i]; }
+                template <typename ET>
+                ParameterizedTensorView<ET>* get_parameterized_tensor_view(size_t i)
+                {
+                    return m_tensor_views[i]->get_parameterized_tensor_view<ET>();
+                }
+
+                template <typename ET>
+                typename ET::type* get_tensor_view_data(size_t i)
+                {
+                    return &get_parameterized_tensor_view<ET>(i)->get_vector()[0];
+                }
+
+            protected:
+                size_t m_n_inputs;
+                size_t m_n_outputs;
+                TensorViewPtrs m_tensor_views;
+                size_t m_initial_pc;
+                std::shared_ptr<std::vector<std::shared_ptr<Instruction>>> m_instructions;
+                size_t m_pc;
+                size_t m_next_pc;
+                bool m_return;
+            };
+        }
+    }
+}

src/ngraph/runtime/eigen/convert.hpp → src/ngraph/runtime/ngvm/eigen/abs.hpp

@@ -14,9 +14,9 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 #include "ngraph/runtime/tensor_view_info.hpp"
 
@@ -24,28 +24,31 @@ namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ETI, typename ETO>
-            class ConvertInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                ConvertInstruction(const TensorViewInfo& arg, const TensorViewInfo& out)
-                    : m_arg(arg)
-                    , m_out(out)
+                template <typename ET>
+                class AbsInstruction : public Instruction
                 {
-                }
+                public:
+                    AbsInstruction(const TensorViewInfo& arg, const TensorViewInfo& out)
+                        : m_arg(arg)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ETO>(call_frame, m_out) =
-                        EigenArray1d<ETI>(call_frame, m_arg).template cast<typename ETO::type>();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) =
+                            Eigen::abs(EigenArray1d<ET>(call_frame, m_arg));
+                    }
 
-            protected:
-                TensorViewInfo m_arg;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/add.hpp → src/ngraph/runtime/ngvm/eigen/add.hpp

@@ -14,41 +14,44 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class AddInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                AddInstruction(const TensorViewInfo& arg0,
-                               const TensorViewInfo& arg1,
-                               const TensorViewInfo& out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class AddInstruction : public Instruction
                 {
-                }
+                public:
+                    AddInstruction(const TensorViewInfo& arg0,
+                                   const TensorViewInfo& arg1,
+                                   const TensorViewInfo& out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) =
-                        EigenArray1d<ET>(call_frame, m_arg0) + EigenArray1d<ET>(call_frame, m_arg1);
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) = EigenArray1d<ET>(call_frame, m_arg0) +
+                                                              EigenArray1d<ET>(call_frame, m_arg1);
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/ngvm/eigen/broadcast_scalar.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace ngvm
+        {
+            namespace eigen
+            {
+                template <typename ET>
+                class BroadcastScalarInstruction : public Instruction
+                {
+                public:
+                    BroadcastScalarInstruction(const TensorViewInfo& arg, const TensorViewInfo& out)
+                        : m_arg(arg)
+                        , m_out(out)
+                    {
+                    }
+
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        // This is a bit hacky: regardless of the tensor rank we
+                        // pull it out as a vector. This works because of the way
+                        // fmt::V computes sizes---it lumps together any higher
+                        // dimensions---while fmt::M ignores them.
+                        EigenArray1d<ET>(call_frame, m_out) =
+                            EigenArray1d<ET>(call_frame, m_arg)(0, 0);
+                    }
+
+                protected:
+                    TensorViewInfo m_arg;
+                    TensorViewInfo m_out;
+                };
+            }
+        }
+    }
+}

src/ngraph/runtime/eigen/equal.hpp → src/ngraph/runtime/ngvm/eigen/broadcast_vector_colwise.hpp

@@ -14,41 +14,42 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
+
         {
-            template <typename ET>
-            class EqualInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                EqualInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class BroadcastVectorColwiseInstruction : public Instruction
                 {
-                }
+                public:
+                    BroadcastVectorColwiseInstruction(const TensorViewInfo& arg,
+                                                      const TensorViewInfo& out)
+                        : m_arg(arg)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<element::Bool>(call_frame, m_out) =
-                        (EigenArray1d<ET>(call_frame, m_arg0) ==
-                         EigenArray1d<ET>(call_frame, m_arg1))
-                            .template cast<char>();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenMatrix<ET>(call_frame, m_out).colwise() =
+                            EigenVector<ET>(call_frame, m_arg);
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/broadcast_vector_rowwise.hpp → src/ngraph/runtime/ngvm/eigen/broadcast_vector_rowwise.hpp

@@ -14,38 +14,41 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class BroadcastVectorRowwiseInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                BroadcastVectorRowwiseInstruction(const TensorViewInfo& arg,
-                                                  const TensorViewInfo& out)
-                    : m_arg(arg)
-                    , m_out(out)
+                template <typename ET>
+                class BroadcastVectorRowwiseInstruction : public Instruction
                 {
-                }
+                public:
+                    BroadcastVectorRowwiseInstruction(const TensorViewInfo& arg,
+                                                      const TensorViewInfo& out)
+                        : m_arg(arg)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenMatrix<ET>(call_frame, m_out).rowwise() =
-                        EigenVector<ET>(call_frame, m_arg).transpose();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenMatrix<ET>(call_frame, m_out).rowwise() =
+                            EigenVector<ET>(call_frame, m_arg).transpose();
+                    }
 
-            protected:
-                TensorViewInfo m_arg;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/ngvm/eigen/call.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include <memory>
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/external_function.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace ngvm
+        {
+            namespace eigen
+            {
+                class CallInstruction : public Instruction
+                {
+                public:
+                    CallInstruction(std::shared_ptr<ExternalFunction> ef,
+                                    std::vector<TensorViewInfo> in,
+                                    std::vector<TensorViewInfo> out)
+                        : m_external_function(ef)
+                        , m_in(in)
+                        , m_out(out)
+                    {
+                    }
+
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        std::shared_ptr<CallFrame> cf = std::dynamic_pointer_cast<CallFrame>(
+                            m_external_function->make_call_frame());
+
+                        std::vector<std::shared_ptr<ngraph::runtime::Value>> inputs;
+                        std::vector<std::shared_ptr<ngraph::runtime::Value>> outputs;
+
+                        for (auto in : m_in)
+                        {
+                            inputs.push_back(call_frame.get_tensor_view(in.get_index()));
+                        }
+                        for (auto out : m_out)
+                        {
+                            outputs.push_back(call_frame.get_tensor_view(out.get_index()));
+                        }
+                        (*cf)(inputs, outputs);
+                    }
+
+                protected:
+                    std::shared_ptr<ExternalFunction> m_external_function;
+                    std::vector<TensorViewInfo> m_in;
+                    std::vector<TensorViewInfo> m_out;
+                };
+            }
+        }
+    }
+}

src/ngraph/runtime/ngvm/eigen/concat_matrix.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
+#include "ngraph/runtime/tensor_view_info.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace ngvm
+        {
+            namespace eigen
+            {
+                template <typename ET>
+                class ConcatMatrixInstruction : public Instruction
+                {
+                public:
+                    ConcatMatrixInstruction(const std::vector<TensorViewInfo>& args,
+                                            size_t axis,
+                                            const TensorViewInfo& out)
+                        : m_args(args)
+                        , m_axis(axis)
+                        , m_out(out)
+                    {
+                        size_t concat_pos[2]{0, 0};
+                        for (auto arg : args)
+                        {
+                            auto& arg_shape = arg.get_tensor_view_layout()->get_shape();
+                            m_blocks.push_back(
+                                {concat_pos[0], concat_pos[1], arg_shape.at(0), arg_shape.at(1)});
+                            concat_pos[axis] += arg_shape.at(axis);
+                        }
+                    }
+
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenMatrix<ET> out(call_frame, m_out);
+                        for (size_t i = 0; i < m_args.size(); i++)
+                        {
+                            auto& b = m_blocks[i];
+                            out.block(b[0], b[1], b[2], b[3])
+                                << EigenMatrix<ET>(call_frame, m_args.at(i));
+                        }
+                    }
+
+                protected:
+                    std::vector<TensorViewInfo> m_args;
+                    size_t m_axis;
+                    TensorViewInfo m_out;
+                    std::vector<std::vector<size_t>> m_blocks;
+                };
+            }
+        }
+    }
+}

src/ngraph/runtime/eigen/concat_matrix.hpp → src/ngraph/runtime/ngvm/eigen/concat_vector.hpp

@@ -14,9 +14,11 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include <vector>
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 #include "ngraph/runtime/tensor_view_info.hpp"
 
@@ -24,46 +26,45 @@ namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class ConcatMatrixInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                ConcatMatrixInstruction(const std::vector<TensorViewInfo>& args,
-                                        size_t axis,
-                                        const TensorViewInfo& out)
-                    : m_args(args)
-                    , m_axis(axis)
-                    , m_out(out)
+                // Would be better to just generate a sequence of copy into slice of output instructions
+                template <typename ET>
+                class ConcatVectorInstruction : public Instruction
                 {
-                    size_t concat_pos[2]{0, 0};
-                    for (auto arg : args)
+                public:
+                    ConcatVectorInstruction(const std::vector<TensorViewInfo>& args,
+                                            const TensorViewInfo& out)
+                        : m_args(args)
+                        , m_out(out)
                     {
-                        auto& arg_shape = arg.get_tensor_view_layout()->get_shape();
-                        m_blocks.push_back(
-                            {concat_pos[0], concat_pos[1], arg_shape.at(0), arg_shape.at(1)});
-                        concat_pos[axis] += arg_shape.at(axis);
+                        for (auto arg : args)
+                        {
+                            auto& arg_shape = arg.get_tensor_view_layout()->get_shape();
+                            m_sizes.push_back(arg_shape.at(0));
+                        }
                     }
-                }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenMatrix<ET> out(call_frame, m_out);
-                    for (size_t i = 0; i < m_args.size(); i++)
+                    virtual void execute(CallFrame& call_frame) const override
                     {
-                        auto& b = m_blocks[i];
-                        out.block(b[0], b[1], b[2], b[3])
-                            << EigenMatrix<ET>(call_frame, m_args.at(i));
+                        EigenVector<ET> out(call_frame, m_out);
+                        size_t concat_pos = 0;
+                        for (size_t i = 0; i < m_args.size(); i++)
+                        {
+                            out.segment(concat_pos, m_sizes[i])
+                                << EigenVector<ET>(call_frame, m_args.at(i));
+                            concat_pos += m_sizes[i];
+                        }
                     }
-                }
 
-            protected:
-                std::vector<TensorViewInfo> m_args;
-                size_t m_axis;
-                TensorViewInfo m_out;
-                std::vector<std::vector<size_t>> m_blocks;
-            };
+                protected:
+                    std::vector<TensorViewInfo> m_args;
+                    TensorViewInfo m_out;
+                    std::vector<size_t> m_sizes;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/copy.hpp → src/ngraph/runtime/ngvm/eigen/constant.hpp

@@ -14,42 +14,42 @@
 
 #pragma once
 
-#include <cassert>
-
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
+#include "ngraph/runtime/tensor_view_info.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            /// @brief Copies a tensor from in to out.
-            template <typename ET>
-            class CopyInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                /// @param in Index of input tensor in call frame.
-                /// @param out Index of output tensor in call frame.
-                CopyInstruction(size_t in, size_t out)
-                    : m_in(in)
-                    , m_out(out)
+                template <typename ET>
+                class ConstantInstruction : public Instruction
                 {
-                }
+                public:
+                    ConstantInstruction(const std::vector<typename ET::type> value,
+                                        const TensorViewInfo& out)
+                        : m_value(value)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    call_frame.get_parameterized_tensor_view<ET>(m_out)->get_vector() =
-                        call_frame.get_parameterized_tensor_view<ET>(m_in)->get_vector();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        call_frame.get_parameterized_tensor_view<ET>(m_out.get_index())
+                            ->get_vector() = m_value;
+                    }
 
-            protected:
-                size_t m_in;
-                size_t m_out;
-            };
+                protected:
+                    const std::vector<typename ET::type> m_value;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/constant.hpp → src/ngraph/runtime/ngvm/eigen/convert.hpp

@@ -14,9 +14,9 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 #include "ngraph/runtime/tensor_view_info.hpp"
 
@@ -24,29 +24,32 @@ namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class ConstantInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                ConstantInstruction(const std::vector<typename ET::type> value,
-                                    const TensorViewInfo& out)
-                    : m_value(value)
-                    , m_out(out)
+                template <typename ETI, typename ETO>
+                class ConvertInstruction : public Instruction
                 {
-                }
+                public:
+                    ConvertInstruction(const TensorViewInfo& arg, const TensorViewInfo& out)
+                        : m_arg(arg)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    call_frame.get_parameterized_tensor_view<ET>(m_out.get_index())->get_vector() =
-                        m_value;
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ETO>(call_frame, m_out) =
+                            EigenArray1d<ETI>(call_frame, m_arg)
+                                .template cast<typename ETO::type>();
+                    }
 
-            protected:
-                const std::vector<typename ET::type> m_value;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/less_eq.hpp → src/ngraph/runtime/ngvm/eigen/copy.hpp

@@ -14,41 +14,45 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include <cassert>
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class LessEqInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                LessEqInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                /// @brief Copies a tensor from in to out.
+                template <typename ET>
+                class CopyInstruction : public Instruction
                 {
-                }
+                public:
+                    /// @param in Index of input tensor in call frame.
+                    /// @param out Index of output tensor in call frame.
+                    CopyInstruction(size_t in, size_t out)
+                        : m_in(in)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<element::Bool>(call_frame, m_out) =
-                        (EigenArray1d<ET>(call_frame, m_arg0) <=
-                         EigenArray1d<ET>(call_frame, m_arg1))
-                            .template cast<char>();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        call_frame.get_parameterized_tensor_view<ET>(m_out)->get_vector() =
+                            call_frame.get_parameterized_tensor_view<ET>(m_in)->get_vector();
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    size_t m_in;
+                    size_t m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/matrix_mult.hpp → src/ngraph/runtime/ngvm/eigen/divide.hpp

@@ -14,41 +14,44 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class MatrixMultInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                MatrixMultInstruction(const TensorViewInfo& arg0,
+                template <typename ET>
+                class DivideInstruction : public Instruction
+                {
+                public:
+                    DivideInstruction(const TensorViewInfo& arg0,
                                       const TensorViewInfo& arg1,
                                       const TensorViewInfo& out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
-                {
-                }
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenMatrix<ET>(call_frame, m_out) =
-                        EigenMatrix<ET>(call_frame, m_arg0) * EigenMatrix<ET>(call_frame, m_arg1);
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) = EigenArray1d<ET>(call_frame, m_arg0) /
+                                                              EigenArray1d<ET>(call_frame, m_arg1);
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/concat_vector.hpp → src/ngraph/runtime/ngvm/eigen/dot.hpp

@@ -14,51 +14,45 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            // Would be better to just generate a sequence of copy into slice of output instructions
-            template <typename ET>
-            class ConcatVectorInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                ConcatVectorInstruction(const std::vector<TensorViewInfo>& args,
-                                        const TensorViewInfo& out)
-                    : m_args(args)
-                    , m_out(out)
+                template <typename ET>
+                class DotInstruction : public Instruction
                 {
-                    for (auto arg : args)
+                public:
+                    DotInstruction(const TensorViewInfo& arg0,
+                                   const TensorViewInfo& arg1,
+                                   const TensorViewInfo& out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
                     {
-                        auto& arg_shape = arg.get_tensor_view_layout()->get_shape();
-                        m_sizes.push_back(arg_shape.at(0));
                     }
-                }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenVector<ET> out(call_frame, m_out);
-                    size_t concat_pos = 0;
-                    for (size_t i = 0; i < m_args.size(); i++)
+                    virtual void execute(CallFrame& call_frame) const override
                     {
-                        out.segment(concat_pos, m_sizes[i])
-                            << EigenVector<ET>(call_frame, m_args.at(i));
-                        concat_pos += m_sizes[i];
+                        EigenArray1d<ET>(call_frame, m_out)
+                            << EigenVector<ET>(call_frame, m_arg0)
+                                   .dot(EigenVector<ET>(call_frame, m_arg1));
                     }
-                }
 
-            protected:
-                std::vector<TensorViewInfo> m_args;
-                TensorViewInfo m_out;
-                std::vector<size_t> m_sizes;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/select.hpp → src/ngraph/runtime/ngvm/eigen/equal.hpp

@@ -14,46 +14,44 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class SelectInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                SelectInstruction(TensorViewInfo arg0,
-                                  TensorViewInfo arg1,
-                                  TensorViewInfo arg2,
-                                  TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_arg2(arg2)
-                    , m_out(out)
+                template <typename ET>
+                class EqualInstruction : public Instruction
                 {
-                }
+                public:
+                    EqualInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) =
-                        EigenArray1d<element::Bool>(call_frame, m_arg0)
-                            .select(EigenArray1d<ET>(call_frame, m_arg1),
-                                    EigenArray1d<ET>(call_frame, m_arg2));
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<element::Bool>(call_frame, m_out) =
+                            (EigenArray1d<ET>(call_frame, m_arg0) ==
+                             EigenArray1d<ET>(call_frame, m_arg1))
+                                .template cast<char>();
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_arg2;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/call.hpp → src/ngraph/runtime/ngvm/eigen/greater_eq.hpp

@@ -14,53 +14,54 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/external_function.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            class CallInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                CallInstruction(std::shared_ptr<ExternalFunction> ef,
-                                std::vector<TensorViewInfo> in,
-                                std::vector<TensorViewInfo> out)
-                    : m_external_function(ef)
-                    , m_in(in)
-                    , m_out(out)
+                template <typename TI, typename TO>
+                void greater_eq(TI arg0, TI arg1, TO out)
                 {
+                    auto result_as_float = get_map_array(&*arg0) <= get_map_array(&*arg1);
+                    auto result_as_char = result_as_float.template cast<char>();
+                    set_map_array(&*out, result_as_char);
                 }
 
-                virtual void execute(CallFrame& call_frame) const override
+                template <typename ET>
+                class GreaterEqInstruction : public Instruction
                 {
-                    std::shared_ptr<CallFrame> cf = m_external_function->make_call_frame();
-
-                    std::vector<std::shared_ptr<ngraph::runtime::Value>> inputs;
-                    std::vector<std::shared_ptr<ngraph::runtime::Value>> outputs;
-
-                    for (auto in : m_in)
+                public:
+                    GreaterEqInstruction(TensorViewInfo arg0,
+                                         TensorViewInfo arg1,
+                                         TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
                     {
-                        inputs.push_back(call_frame.get_tensor_view(in.get_index()));
                     }
-                    for (auto out : m_out)
+
+                    virtual void execute(CallFrame& call_frame) const override
                     {
-                        outputs.push_back(call_frame.get_tensor_view(out.get_index()));
+                        EigenArray1d<element::Bool>(call_frame, m_out) =
+                            (EigenArray1d<ET>(call_frame, m_arg0) >=
+                             EigenArray1d<ET>(call_frame, m_arg1))
+                                .template cast<char>();
                     }
-                    (*cf)(inputs, outputs);
-                }
 
-            protected:
-                std::shared_ptr<ExternalFunction> m_external_function;
-                std::vector<TensorViewInfo> m_in;
-                std::vector<TensorViewInfo> m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/broadcast_vector_colwise.hpp → src/ngraph/runtime/ngvm/eigen/greater_than.hpp

@@ -14,38 +14,46 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class BroadcastVectorColwiseInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                BroadcastVectorColwiseInstruction(const TensorViewInfo& arg,
-                                                  const TensorViewInfo& out)
-                    : m_arg(arg)
-                    , m_out(out)
+                template <typename ET>
+                class GreaterThanInstruction : public Instruction
                 {
-                }
+                public:
+                    GreaterThanInstruction(TensorViewInfo arg0,
+                                           TensorViewInfo arg1,
+                                           TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenMatrix<ET>(call_frame, m_out).colwise() =
-                        EigenVector<ET>(call_frame, m_arg);
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<element::Bool>(call_frame, m_out) =
+                            (EigenArray1d<ET>(call_frame, m_arg0) >
+                             EigenArray1d<ET>(call_frame, m_arg1))
+                                .template cast<char>();
+                    }
 
-            protected:
-                TensorViewInfo m_arg;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/broadcast_scalar.hpp → src/ngraph/runtime/ngvm/eigen/less_eq.hpp

@@ -14,40 +14,44 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class BroadcastScalarInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                BroadcastScalarInstruction(const TensorViewInfo& arg, const TensorViewInfo& out)
-                    : m_arg(arg)
-                    , m_out(out)
+                template <typename ET>
+                class LessEqInstruction : public Instruction
                 {
-                }
+                public:
+                    LessEqInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    // This is a bit hacky: regardless of the tensor rank we
-                    // pull it out as a vector. This works because of the way
-                    // fmt::V computes sizes---it lumps together any higher
-                    // dimensions---while fmt::M ignores them.
-                    EigenArray1d<ET>(call_frame, m_out) = EigenArray1d<ET>(call_frame, m_arg)(0, 0);
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<element::Bool>(call_frame, m_out) =
+                            (EigenArray1d<ET>(call_frame, m_arg0) <=
+                             EigenArray1d<ET>(call_frame, m_arg1))
+                                .template cast<char>();
+                    }
 
-            protected:
-                TensorViewInfo m_arg;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/less_than.hpp → src/ngraph/runtime/ngvm/eigen/less_than.hpp

@@ -14,41 +14,46 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class LessThanInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                LessThanInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class LessThanInstruction : public Instruction
                 {
-                }
+                public:
+                    LessThanInstruction(TensorViewInfo arg0,
+                                        TensorViewInfo arg1,
+                                        TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<element::Bool>(call_frame, m_out) =
-                        (EigenArray1d<ET>(call_frame, m_arg0) <
-                         EigenArray1d<ET>(call_frame, m_arg1))
-                            .template cast<char>();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<element::Bool>(call_frame, m_out) =
+                            (EigenArray1d<ET>(call_frame, m_arg0) <
+                             EigenArray1d<ET>(call_frame, m_arg1))
+                                .template cast<char>();
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/log.hpp → src/ngraph/runtime/ngvm/eigen/log.hpp

@@ -14,37 +14,40 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class LogInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                LogInstruction(TensorViewInfo arg, TensorViewInfo out)
-                    : m_arg(arg)
-                    , m_out(out)
+                template <typename ET>
+                class LogInstruction : public Instruction
                 {
-                }
+                public:
+                    LogInstruction(TensorViewInfo arg, TensorViewInfo out)
+                        : m_arg(arg)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET, fmt::V>(call_frame, m_out) =
-                        Eigen::log(EigenArray1d<ET, fmt::V>(call_frame, m_arg));
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET, fmt::V>(call_frame, m_out) =
+                            Eigen::log(EigenArray1d<ET, fmt::V>(call_frame, m_arg));
+                    }
 
-            protected:
-                TensorViewInfo m_arg;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/ngvm/eigen/matrix_mult.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace ngvm
+        {
+            namespace eigen
+            {
+                template <typename ET>
+                class MatrixMultInstruction : public Instruction
+                {
+                public:
+                    MatrixMultInstruction(const TensorViewInfo& arg0,
+                                          const TensorViewInfo& arg1,
+                                          const TensorViewInfo& out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
+
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenMatrix<ET>(call_frame, m_out) = EigenMatrix<ET>(call_frame, m_arg0) *
+                                                             EigenMatrix<ET>(call_frame, m_arg1);
+                    }
+
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
+        }
+    }
+}

src/ngraph/runtime/eigen/matrix_vector_product.hpp → src/ngraph/runtime/ngvm/eigen/matrix_vector_product.hpp

@@ -14,41 +14,44 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class MatrixVectorProductInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                MatrixVectorProductInstruction(const TensorViewInfo& arg0,
-                                               const TensorViewInfo& arg1,
-                                               const TensorViewInfo& out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class MatrixVectorProductInstruction : public Instruction
                 {
-                }
+                public:
+                    MatrixVectorProductInstruction(const TensorViewInfo& arg0,
+                                                   const TensorViewInfo& arg1,
+                                                   const TensorViewInfo& out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenVector<ET>(call_frame, m_out) =
-                        EigenMatrix<ET>(call_frame, m_arg0) * EigenVector<ET>(call_frame, m_arg1);
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenVector<ET>(call_frame, m_out) = EigenMatrix<ET>(call_frame, m_arg0) *
+                                                             EigenVector<ET>(call_frame, m_arg1);
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/greater_than.hpp → src/ngraph/runtime/ngvm/eigen/maximum.hpp

@@ -14,41 +14,43 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class GreaterThanInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                GreaterThanInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class MaximumInstruction : public Instruction
                 {
-                }
+                public:
+                    MaximumInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<element::Bool>(call_frame, m_out) =
-                        (EigenArray1d<ET>(call_frame, m_arg0) >
-                         EigenArray1d<ET>(call_frame, m_arg1))
-                            .template cast<char>();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) =
+                            EigenArray1d<ET>(call_frame, m_arg0)
+                                .max(EigenArray1d<ET>(call_frame, m_arg1));
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/multiply.hpp → src/ngraph/runtime/ngvm/eigen/multiply.hpp

@@ -14,38 +14,43 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class MultiplyInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                MultiplyInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class MultiplyInstruction : public Instruction
                 {
-                }
+                public:
+                    MultiplyInstruction(TensorViewInfo arg0,
+                                        TensorViewInfo arg1,
+                                        TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) =
-                        EigenArray1d<ET>(call_frame, m_arg0) * EigenArray1d<ET>(call_frame, m_arg1);
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) = EigenArray1d<ET>(call_frame, m_arg0) *
+                                                              EigenArray1d<ET>(call_frame, m_arg1);
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/abs.hpp → src/ngraph/runtime/ngvm/eigen/negate.hpp

@@ -14,38 +14,39 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
-#include "ngraph/runtime/tensor_view_info.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class AbsInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                AbsInstruction(const TensorViewInfo& arg, const TensorViewInfo& out)
-                    : m_arg(arg)
-                    , m_out(out)
+                template <typename ET>
+                class NegateInstruction : public Instruction
                 {
-                }
+                public:
+                    NegateInstruction(TensorViewInfo arg, TensorViewInfo out)
+                        : m_arg(arg)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) =
-                        Eigen::abs(EigenArray1d<ET>(call_frame, m_arg));
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) = -EigenArray1d<ET>(call_frame, m_arg);
+                    }
 
-            protected:
-                TensorViewInfo m_arg;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/not_equal.hpp → src/ngraph/runtime/ngvm/eigen/not_equal.hpp

@@ -14,41 +14,46 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class NotEqualInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                NotEqualInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class NotEqualInstruction : public Instruction
                 {
-                }
+                public:
+                    NotEqualInstruction(TensorViewInfo arg0,
+                                        TensorViewInfo arg1,
+                                        TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<element::Bool>(call_frame, m_out) =
-                        (EigenArray1d<ET>(call_frame, m_arg0) !=
-                         EigenArray1d<ET>(call_frame, m_arg1))
-                            .template cast<char>();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<element::Bool>(call_frame, m_out) =
+                            (EigenArray1d<ET>(call_frame, m_arg0) !=
+                             EigenArray1d<ET>(call_frame, m_arg1))
+                                .template cast<char>();
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/ngvm/eigen/return.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace ngvm
+        {
+            namespace eigen
+            {
+                class ReturnInstruction : public Instruction
+                {
+                public:
+                    ReturnInstruction() {}
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        call_frame.set_return();
+                    }
+                };
+            }
+        }
+    }
+}

src/ngraph/runtime/ngvm/eigen/scalar_tensor_product.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace ngvm
+        {
+            namespace eigen
+            {
+                template <typename ET>
+                class ScalarTensorProductInstruction : public Instruction
+                {
+                public:
+                    ScalarTensorProductInstruction(const TensorViewInfo& arg0,
+                                                   const TensorViewInfo& arg1,
+                                                   const TensorViewInfo& out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
+
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        // This is a bit hacky: regardless of the tensor rank we
+                        // pull it out as a vector. This works because of the way
+                        // fmt::V computes sizes---it lumps together any higher
+                        // dimensions---while fmt::M ignores them.
+                        EigenVector<ET>(call_frame, m_out) =
+                            call_frame.get_tensor_view_data<ET>(m_arg0.get_index())[0] *
+                            EigenVector<ET>(call_frame, m_arg1);
+                    }
+
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
+        }
+    }
+}

src/ngraph/runtime/eigen/greater_eq.hpp → src/ngraph/runtime/ngvm/eigen/select.hpp

@@ -14,49 +14,49 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename TI, typename TO>
-            void greater_eq(TI arg0, TI arg1, TO out)
+            namespace eigen
             {
-                auto result_as_float = get_map_array(&*arg0) <= get_map_array(&*arg1);
-                auto result_as_char = result_as_float.template cast<char>();
-                set_map_array(&*out, result_as_char);
-            }
-
-            template <typename ET>
-            class GreaterEqInstruction : public Instruction
-            {
-            public:
-                GreaterEqInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class SelectInstruction : public Instruction
                 {
-                }
+                public:
+                    SelectInstruction(TensorViewInfo arg0,
+                                      TensorViewInfo arg1,
+                                      TensorViewInfo arg2,
+                                      TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_arg2(arg2)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<element::Bool>(call_frame, m_out) =
-                        (EigenArray1d<ET>(call_frame, m_arg0) >=
-                         EigenArray1d<ET>(call_frame, m_arg1))
-                            .template cast<char>();
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) =
+                            EigenArray1d<element::Bool>(call_frame, m_arg0)
+                                .select(EigenArray1d<ET>(call_frame, m_arg1),
+                                        EigenArray1d<ET>(call_frame, m_arg2));
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_arg2;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/eigen/subtract.hpp → src/ngraph/runtime/ngvm/eigen/subtract.hpp

@@ -14,39 +14,44 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/ngvm/eigen/utils.hpp"
+#include "ngraph/runtime/ngvm/instruction.hpp"
 #include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class SubtractInstruction : public Instruction
+            namespace eigen
             {
-            public:
-                SubtractInstruction(TensorViewInfo arg0, TensorViewInfo arg1, TensorViewInfo out)
-                    : m_arg0(arg0)
-                    , m_arg1(arg1)
-                    , m_out(out)
+                template <typename ET>
+                class SubtractInstruction : public Instruction
                 {
-                }
+                public:
+                    SubtractInstruction(TensorViewInfo arg0,
+                                        TensorViewInfo arg1,
+                                        TensorViewInfo out)
+                        : m_arg0(arg0)
+                        , m_arg1(arg1)
+                        , m_out(out)
+                    {
+                    }
 
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) =
-                        EigenArray1d<ET>(call_frame, m_arg0) - EigenArray1d<ET>(call_frame, m_arg1);
-                }
+                    virtual void execute(CallFrame& call_frame) const override
+                    {
+                        EigenArray1d<ET>(call_frame, m_out) = EigenArray1d<ET>(call_frame, m_arg0) -
+                                                              EigenArray1d<ET>(call_frame, m_arg1);
+                    }
 
-            protected:
-                TensorViewInfo m_arg0;
-                TensorViewInfo m_arg1;
-                TensorViewInfo m_out;
-            };
+                protected:
+                    TensorViewInfo m_arg0;
+                    TensorViewInfo m_arg1;
+                    TensorViewInfo m_out;
+                };
+            }
         }
     }
 }

src/ngraph/runtime/ngvm/eigen/utils.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include <memory>
+
+#include <Eigen/Dense>
+
+#include "ngraph/descriptor/layout/dense_tensor_view_layout.hpp"
+#include "ngraph/runtime/ngvm/call_frame.hpp"
+#include "ngraph/runtime/tensor_view_info.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        class TensorViewInfo;
+
+        namespace ngvm
+        {
+            class CallFrame;
+
+            namespace eigen
+            {
+                using DynamicStrides = Eigen::Stride<Eigen::Dynamic, Eigen::Dynamic>;
+                using VectorStrides = Eigen::Stride<Eigen::Dynamic, 1>;
+
+                template <typename ET>
+                using DynamicArray =
+                    Eigen::Array<typename ET::type, Eigen::Dynamic, Eigen::Dynamic>;
+
+                template <typename ET>
+                using EigenArrayBase = Eigen::Map<DynamicArray<ET>, 0, DynamicStrides>;
+
+                template <typename ET>
+                using DynamicMatrix = Eigen::
+                    Matrix<typename ET::type, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
+
+                template <typename ET>
+                using EigenMatrixBase = Eigen::Map<DynamicMatrix<ET>, 0, DynamicStrides>;
+
+                template <typename ET>
+                using DynamicVector = Eigen::Matrix<typename ET::type, Eigen::Dynamic, 1>;
+
+                template <typename ET>
+                using EigenVectorBase = Eigen::Map<DynamicVector<ET>, 0, VectorStrides>;
+
+                namespace fmt
+                {
+                    /// @brief vector format for Eigen wrappers.
+                    class V
+                    {
+                    public:
+                        V(const TensorViewInfo& tensor_view_info)
+                            : l0(tensor_view_info
+                                     .get_layout<
+                                         ngraph::descriptor::layout::DenseTensorViewLayout>()
+                                     ->get_size())
+                        {
+                        }
+
+                    public:
+                        size_t l0;
+                        size_t l1{1};
+                        size_t s0{1};
+                        size_t s1{1};
+                    };
+
+                    class M
+                    {
+                        M(const Shape& shape, const Strides& strides)
+                            : l0(shape.at(0))
+                            , l1(shape.at(1))
+                            , s0(strides.at(0))
+                            , s1(strides.at(1))
+                        {
+                        }
+
+                        M(const std::shared_ptr<ngraph::descriptor::layout::DenseTensorViewLayout>&
+                              layout)
+                            : M(layout->get_shape(), layout->get_strides())
+                        {
+                        }
+
+                    public:
+                        M(const TensorViewInfo& tensor_view_info)
+                            : M(tensor_view_info.get_layout<
+                                  ngraph::descriptor::layout::DenseTensorViewLayout>())
+                        {
+                        }
+
+                    public:
+                        size_t l0;
+                        size_t l1;
+                        size_t s0;
+                        size_t s1;
+                    };
+                }
+
+                // ET element type
+                // FMT array format (fmt::V for vector, etc.)
+                // BASE select array/matrix
+                template <typename ET,
+                          typename FMT,
+                          typename BASE,
+                          typename STRIDES = DynamicStrides>
+                class EigenWrapper : public BASE
+                {
+                    using base = BASE;
+
+                public:
+                    EigenWrapper(typename ET::type* t, const FMT& fmt)
+                        : base(t, fmt.l0, fmt.l1, STRIDES(fmt.s0, fmt.s1))
+                    {
+                    }
+
+                    EigenWrapper(
+                        typename ET::type* t,
+                        const std::shared_ptr<ngraph::descriptor::layout::DenseTensorViewLayout>&
+                            layout)
+                        : base(t, layout->get_size(), 1, DynamicStrides(1, 1))
+                    {
+                    }
+
+                    EigenWrapper(CallFrame& call_frame, const TensorViewInfo& tensor_view_info)
+                        : EigenWrapper(
+                              call_frame.get_tensor_view_data<ET>(tensor_view_info.get_index()),
+                              FMT(tensor_view_info))
+                    {
+                    }
+
+                    template <typename U>
+                    EigenWrapper& operator=(const U& other)
+                    {
+                        this->base::operator=(other);
+                        return *this;
+                    }
+                };
+
+                template <typename ET, typename FMT = fmt::V>
+                using EigenArray1d = EigenWrapper<ET, FMT, EigenArrayBase<ET>>;
+
+                template <typename ET, typename FMT = fmt::M>
+                using EigenArray2d = EigenWrapper<ET, FMT, EigenArrayBase<ET>>;
+
+                template <typename ET, typename FMT = fmt::M>
+                using EigenMatrix = EigenWrapper<ET, FMT, EigenMatrixBase<ET>>;
+
+                template <typename ET, typename FMT = fmt::V>
+                using EigenVector = EigenWrapper<ET, FMT, EigenVectorBase<ET>, VectorStrides>;
+            }
+        }
+    }
+}

src/ngraph/runtime/ngvm/external_function.hpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#pragma once
+
+#include <memory>
+#include <typeindex>
+#include <typeinfo>
+#include <unordered_map>
+
+#include "ngraph/function.hpp"
+#include "ngraph/runtime/external_function.hpp"
+#include "ngraph/runtime/tensor_view_info.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace ngvm
+        {
+            class Instruction;
+
+            class ExternalFunction : public ngraph::runtime::ExternalFunction
+            {
+                using FunctionMap = std::unordered_map<std::shared_ptr<Function>,
+                                                       std::shared_ptr<ExternalFunction>>;
+
+                using OpFunction = std::function<void(const ngraph::Node*,
+                                                      ExternalFunction*,
+                                                      FunctionMap&,
+                                                      const std::vector<TensorViewInfo>& inputs,
+                                                      const std::vector<TensorViewInfo>& outputs)>;
+                using OpMap = std::unordered_map<std::type_index, OpFunction>;
+
+            public:
+                ExternalFunction(const std::shared_ptr<ngraph::Function>& function,
+                                 bool release_function = true);
+                std::shared_ptr<ngraph::runtime::CallFrame> make_call_frame();
+                virtual std::shared_ptr<ngraph::runtime::CallFrame>
+                    make_call_frame(FunctionMap& function_map);
+                std::shared_ptr<std::vector<std::shared_ptr<Instruction>>> get_instructions()
+                {
+                    return m_instructions;
+                }
+
+                // Release original function's resources
+                void release_function() { m_function = nullptr; }
+            protected:
+                void compile();
+                void compile(FunctionMap& function_map);
+
+                size_t m_n_inputs;
+                size_t m_n_outputs;
+                std::shared_ptr<std::vector<std::shared_ptr<Instruction>>> m_instructions;
+                ngraph::descriptor::TensorViewPtrs m_temp_views;
+
+                static OpMap& get_op_map();
+            };
+        }
+    }
+}

src/ngraph/runtime/eigen/negate.hpp → src/ngraph/runtime/ngvm/instruction.hpp

@@ -14,35 +14,26 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/eigen/utils.hpp"
-#include "ngraph/runtime/instruction.hpp"
-#include "ngraph/runtime/tensor_view.hpp"
+#include <memory>
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            template <typename ET>
-            class NegateInstruction : public Instruction
+            class CallFrame;
+
+            /// @brief An interpreter for an Op
+            ///
+            /// The call_frame has a vector of instructions and calls execute on each instruction, passing it the call_frame.
+            /// Instructions get argument, result, and intermediate tensor views from the call frame. Instructions may also
+            /// set a flag in the call_frame to end execution, or adjust execution by modifying the position in the instruction vector.
+            class Instruction
             {
             public:
-                NegateInstruction(TensorViewInfo arg, TensorViewInfo out)
-                    : m_arg(arg)
-                    , m_out(out)
-                {
-                }
-
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    EigenArray1d<ET>(call_frame, m_out) = -EigenArray1d<ET>(call_frame, m_arg);
-                }
-
-            protected:
-                TensorViewInfo m_arg;
-                TensorViewInfo m_out;
+                virtual ~Instruction() {}
+                virtual void execute(CallFrame& call_frame) const = 0;
             };
         }
     }

src/ngraph/runtime/ngvm/ngvm_backend.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include "ngraph/runtime/ngvm/ngvm_backend.hpp"
+#include "ngraph/runtime/external_function.hpp"
+
+using namespace ngraph::runtime::ngvm;
+
+std::shared_ptr<ngraph::runtime::CallFrame>
+    NGVMBackend::make_call_frame(const std::shared_ptr<ExternalFunction>& external_function)
+{
+    return external_function->make_call_frame();
+}

src/ngraph/runtime/eigen/return.hpp → src/ngraph/runtime/ngvm/ngvm_backend.hpp

@@ -14,23 +14,20 @@
 
 #pragma once
 
-#include "ngraph/runtime/call_frame.hpp"
-#include "ngraph/runtime/instruction.hpp"
+#include "ngraph/runtime/backend.hpp"
 
 namespace ngraph
 {
     namespace runtime
     {
-        namespace eigen
+        namespace ngvm
         {
-            class ReturnInstruction : public Instruction
+            /// @brief Transformer for the interpreted backend
+            class NGVMBackend : public Backend
             {
             public:
-                ReturnInstruction() {}
-                virtual void execute(CallFrame& call_frame) const override
-                {
-                    call_frame.set_return();
-                }
+                virtual std::shared_ptr<ngraph::runtime::CallFrame> make_call_frame(
+                    const std::shared_ptr<ngraph::runtime::ExternalFunction>& external_function);
             };
         }
     }

src/ngraph/runtime/ngvm/ngvm_manager.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include <memory>
+
+#include "ngraph/runtime/ngvm/external_function.hpp"
+#include "ngraph/runtime/ngvm/ngvm_backend.hpp"
+#include "ngraph/runtime/ngvm/ngvm_manager.hpp"
+
+using namespace ngraph::runtime::ngvm;
+
+std::shared_ptr<ngraph::runtime::Backend> NGVMManager::allocate_backend()
+{
+    return std::make_shared<NGVMBackend>();
+}
+
+std::shared_ptr<ngraph::runtime::ExternalFunction>
+    NGVMManager::compile(const std::shared_ptr<ngraph::Function>& fun)
+{
+    return std::make_shared<ExternalFunction>(fun);
+}
+
+ngraph::runtime::Manager::Factory NGVMManager::factory = ngraph::runtime::Manager::register_factory(
+    "NGVM", [](const std::string& name) -> std::shared_ptr<ngraph::runtime::Manager> {
+        return std::make_shared<NGVMManager>();
+    });

src/ngraph/runtime/instruction.hpp → src/ngraph/runtime/ngvm/ngvm_manager.hpp

@@ -16,24 +16,29 @@
 
 #include <memory>
 
-#include "ngraph/runtime/call_frame.hpp"
+#include "ngraph/runtime/manager.hpp"
 
 namespace ngraph
 {
+    class Function;
+
     namespace runtime
     {
-        class CallFrame;
-
-        /// @brief An interpreter for an Op
-        ///
-        /// The call_frame has a vector of instructions and calls execute on each instruction, passing it the call_frame.
-        /// Instructions get argument, result, and intermediate tensor views from the call frame. Instructions may also
-        /// set a flag in the call_frame to end execution, or adjust execution by modifying the position in the instruction vector.
-        class Instruction
+        class ExternalFunction;
+
+        namespace ngvm
         {
-        public:
-            virtual ~Instruction() {}
-            virtual void execute(CallFrame& call_frame) const = 0;
+            /// @brief Transformer for the interpreted backend
+            class NGVMManager : public Manager
+            {
+            public:
+                virtual std::shared_ptr<Backend> allocate_backend() override;
+
+                virtual std::shared_ptr<ngraph::runtime::ExternalFunction>
+                    compile(const std::shared_ptr<ngraph::Function>& fun) override;
+
+                static Factory factory;
+            };
         };
     }
 }

src/ngraph/runtime/parameterized_tensor_view.hpp

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <cstring>
 #include <memory>
 #include <vector>
 
@@ -61,6 +62,48 @@ namespace ngraph
 
             // For getting the data out
             storage_type& get_vector() { return m_vector; }
+            virtual void write(const void* p, size_t tensor_offset, size_t n) override
+            {
+                size_t elt_offset = tensor_offset / sizeof(typename ET::type);
+                if (elt_offset * sizeof(typename ET::type) != tensor_offset)
+                {
+                    throw ngraph_error("Attempt to write to an address not aligned on an element");
+                }
+                size_t elt_n = n / sizeof(typename ET::type);
+                if (elt_n * sizeof(typename ET::type) != n)
+                {
+                    throw ngraph_error("Attemmpt to write a partial element");
+                }
+                size_t elt_byte_size = sizeof(typename ET::type) * n;
+                if (tensor_offset + n > elt_byte_size)
+                {
+                    throw ngraph_error("Attempt to write beyond the tensor");
+                }
+
+                std::memcpy(&m_vector[elt_offset], p, n);
+            }
+
+            virtual void read(void* p, size_t tensor_offset, size_t n) const override
+            {
+                size_t elt_offset = tensor_offset / sizeof(typename ET::type);
+                if (elt_offset * sizeof(typename ET::type) != tensor_offset)
+                {
+                    throw ngraph_error("Attempt to read from an address not aligned on an element");
+                }
+                size_t elt_n = n / sizeof(typename ET::type);
+                if (elt_n * sizeof(typename ET::type) != n)
+                {
+                    throw ngraph_error("Attemmpt to read a partial element");
+                }
+                size_t elt_byte_size = sizeof(typename ET::type) * n;
+                if (tensor_offset + n > elt_byte_size)
+                {
+                    throw ngraph_error("Attempt to read beyond the tensor");
+                }
+
+                std::memcpy(p, &m_vector[elt_offset], n);
+            }
+
         protected:
             storage_type m_vector;
         };

src/ngraph/runtime/tensor_view.cpp

+// ----------------------------------------------------------------------------
+// Copyright 2017 Nervana Systems Inc.
+// 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
+// ----------------------------------------------------------------------------
+
+#include "ngraph/runtime/tensor_view.hpp"
+#include "ngraph/common.hpp"
+#include "ngraph/types/element_type.hpp"
+
+using namespace ngraph::runtime;
+
+std::shared_ptr<const ngraph::descriptor::TensorView> TensorView::get_tensor_view_descriptor() const
+{
+    return m_descriptor;
+}
+
+std::shared_ptr<ngraph::descriptor::Value> TensorView::get_descriptor() const
+{
+    return m_descriptor;
+}
+
+void TensorView::collect_tensor_views(std::vector<std::shared_ptr<TensorView>>& views,
+                                      const std::shared_ptr<Value>& value) const
+{
+    views.push_back(std::static_pointer_cast<TensorView>(value));
+}
+
+const ngraph::Shape& TensorView::get_shape() const
+{
+    return m_descriptor->get_tensor_view_type()->get_shape();
+}

src/ngraph/runtime/tensor_view.hpp

@@ -20,10 +20,14 @@
 #include "ngraph/descriptor/tensor_view.hpp"
 #include "ngraph/runtime/value.hpp"
 #include "ngraph/shape.hpp"
-#include "ngraph/types/element_type.hpp"
 
 namespace ngraph
 {
+    namespace descriptor
+    {
+        class Value;
+    }
+
     namespace runtime
     {
         template <typename ET>
@@ -46,23 +50,28 @@ namespace ngraph
                 return dynamic_cast<ParameterizedTensorView<ET>*>(this);
             }
 
-            std::shared_ptr<const ngraph::descriptor::TensorView> get_tensor_view_descriptor() const
-            {
-                return m_descriptor;
-            }
+            std::shared_ptr<const ngraph::descriptor::TensorView>
+                get_tensor_view_descriptor() const;
 
-            virtual std::shared_ptr<ngraph::descriptor::Value> get_descriptor() const override
-            {
-                return m_descriptor;
-            }
+            virtual std::shared_ptr<ngraph::descriptor::Value> get_descriptor() const override;
 
             virtual void collect_tensor_views(std::vector<std::shared_ptr<TensorView>>& views,
-                                              const std::shared_ptr<Value>& value) const override
-            {
-                views.push_back(std::static_pointer_cast<TensorView>(value));
-            }
+                                              const std::shared_ptr<Value>& value) const override;
+
+            const ngraph::Shape& get_shape() const;
+
+            /// @brief Write bytes directly into the tensor
+            /// @param p Pointer to source of data
+            /// @param tensor_offset Offset into tensor storage to begin writing. Must be element-aligned.
+            /// @param n Number of bytes to write, must be integral number of elements.
+            virtual void write(const void* p, size_t tensor_offset, size_t n) = 0;
+
+            /// @brief Read bytes directly from the tensor
+            /// @param p Pointer to destination for data
+            /// @param tensor_offset Offset into tensor storage to begin reading. Must be element-aligned.
+            /// @param n Number of bytes to read, must be integral number of elements.
+            virtual void read(void* p, size_t tensor_offset, size_t n) const = 0;
 
-            const Shape& get_shape() { return m_descriptor->get_tensor_view_type()->get_shape(); }
         protected:
             std::shared_ptr<ngraph::descriptor::TensorView> m_descriptor;
         };

src/ngraph/runtime/value.hpp

@@ -21,6 +21,11 @@
 
 namespace ngraph
 {
+    namespace descriptor
+    {
+        class Value;
+    }
+
     namespace runtime
     {
         class TensorView;

src/ngraph/types/element_type.cpp

@@ -21,8 +21,6 @@
 
 using namespace ngraph;
 
-std::map<std::string, ngraph::element::Type> ngraph::element::Type::m_element_list;
-
 ngraph::element::Type::Type(size_t bitwidth,
                             bool is_float,
                             bool is_signed,

src/ngraph/types/element_type.hpp

@@ -19,10 +19,14 @@
 #pragma once
 
 #include <map>
+#include <memory>
 #include <string>
 #include <type_traits>
 
+#include "ngraph/common.hpp"
 #include "ngraph/except.hpp"
+#include "ngraph/runtime/parameterized_tensor_view.hpp"
+#include "ngraph/runtime/tensor_view.hpp"
 
 namespace ngraph
 {
@@ -34,6 +38,7 @@ namespace ngraph
             Type& operator=(const Type&) = delete;
 
         public:
+            virtual ~Type() {}
             Type(size_t bitwidth, bool is_float, bool is_signed, const std::string& cname);
 
             const std::string& c_type_string() const;
@@ -44,6 +49,9 @@ namespace ngraph
                 return h(m_cname);
             }
 
+            virtual std::shared_ptr<ngraph::runtime::TensorView>
+                make_primary_tensor_view(const Shape& shape) const = 0;
+
             bool operator==(const Type& other) const;
             bool operator!=(const Type& other) const { return !(*this == other); }
             friend std::ostream& operator<<(std::ostream&, const Type&);
@@ -102,6 +110,12 @@ namespace ngraph
                 static TraitedType<T> t;
                 return t;
             }
+
+            virtual std::shared_ptr<ngraph::runtime::TensorView>
+                make_primary_tensor_view(const ngraph::Shape& shape) const override
+            {
+                return std::make_shared<runtime::ParameterizedTensorView<TraitedType<T>>>(shape);
+            }
         };
 
         NGRAPH_DEFINE_TRAITED_TYPE_NAME(char)

test/tensor.cpp

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // ----------------------------------------------------------------------------
 
-#include <memory>
+#include <algorithm>
 #include <memory>
 #include <sstream>
 #include <string>
@@ -83,3 +83,37 @@ TEST(tensor, size)
         EXPECT_EQ(1 * 4, output.size());
     }
 }
+
+template <typename ET>
+void test_read_write(const std::vector<typename ET::type>& x)
+{
+    using T = typename ET::type;
+
+    auto manager = ngraph::runtime::Manager::get("NGVM");
+    auto backend = manager->allocate_backend();
+
+    auto a = backend->make_primary_tensor_view(ET::element_type(), Shape{2, x.size()});
+    auto af = a->template get_parameterized_tensor_view<ET>();
+
+    std::vector<T> result(2 * x.size());
+
+    a->write(&x[0], 0, x.size() * sizeof(T));
+    std::copy(x.begin(), x.end(), result.begin());
+    a->write(&x[0], x.size() * sizeof(T), x.size() * sizeof(T));
+    std::copy(x.begin(), x.end(), result.begin() + x.size());
+
+    auto& af_vector = af->get_vector();
+    ASSERT_EQ(af_vector, result);
+
+    std::vector<T> result1(x.size());
+    std::vector<T> result2(x.size());
+    std::copy(result.begin() + 1, result.begin() + 1 + x.size(), result1.begin());
+    a->read(&result2[0], sizeof(T), sizeof(T) * x.size());
+    ASSERT_EQ(result1, result2);
+}
+
+TEST(tensor, read_write)
+{
+    test_read_write<element::Float32>({1.0, 3.0, 5.0});
+    test_read_write<element::Int64>({-1, 2, 4});
+}