Merge branch 'master' into cyphers/dochow

cmake/external_llvm_prebuilt.cmake

@@ -16,8 +16,11 @@
 
 include(ExternalProject)
 
-if((NGRAPH_CPU_ENABLE OR NGRAPH_GPU_ENABLE) AND (NOT ${CMAKE_SYSTEM_NAME} MATCHES "Darwin") AND
-                         (NOT ${CMAKE_SYSTEM_NAME} MATCHES "Windows"))
+if (NGRAPH_CPU_ENABLE AND (${CMAKE_SYSTEM_NAME} MATCHES "Darwin") OR (${CMAKE_SYSTEM_NAME} MATCHES "Windows"))
+    message(FATAL_ERROR "The NGRAPH_USE_PREBUILT_LLVM option is not supported on this platform.")
+endif()
+
+if (NGRAPH_CPU_ENABLE)
     message(STATUS "Fetching LLVM from llvm.org")
 
     # Override default LLVM binaries

doc/sphinx/source/ops/allreduce.rst

+.. allreduce.rst:
+
+###
+AllReduce
+###
+
+.. code-block:: cpp
+
+   AllReduce // Collective operation
+
+
+Description
+===========
+
+Combines values from all processes or devices and distributes the result back
+to all processes or devices.
+
+
+Inputs
+------
+
++-----------------+-------------------------+--------------------------------+
+| Name            | Element Type            | Shape                          |
++=================+=========================+================================+
+| ``arg``         | ``element::f32``        | Any                            |
+|                 | ``element::f64``        |                                |
++-----------------+-------------------------+--------------------------------+
+
+
+Outputs
+-------
+
++-----------------+-------------------------+--------------------------------+
+| Name            | Element Type            | Shape                          |
++=================+=========================+================================+
+| ``output``      | ``element::f32``        | Same as ``arg``                |
+|                 | ``element::f64``        |                                |
++-----------------+-------------------------+--------------------------------+
+
+
+C++ Interface
+=============
+
+.. doxygenclass:: ngraph::op::AllReduce
+   :project: ngraph
+   :members:

doc/sphinx/source/ops/index.rst

@@ -52,6 +52,7 @@ Not currently a comprehensive list.
    abs.rst
    acos.rst
    add.rst
+   allreduce.rst
    asin.rst
    atan.rst
    avg_pool.rst

src/ngraph/CMakeLists.txt

@@ -34,6 +34,7 @@ set (SRC
     node.cpp
     ops/abs.cpp
     ops/add.cpp
+    ops/allreduce.cpp
     ops/avg_pool.cpp
     ops/batch_norm.cpp
     ops/broadcast.cpp
@@ -215,11 +216,6 @@ if(NGRAPH_DISTRIBUTED_ENABLE AND MPI_CXX_INCLUDE_PATH)
     include_directories(SYSTEM ${MPI_C_INCLUDE_PATH} ${MPI_CXX_INCLUDE_PATH})
     link_directories(${MPI_C_LIBRARIES} ${MPI_CXX_LIBRARIES})
 
-    # Add sources for distributed ngraph
-    # and all its dependencies
-    set(SRC ${SRC}
-        ops/allreduce.cpp
-    )
     set_property(SOURCE codegen/compiler.cpp APPEND PROPERTY COMPILE_DEFINITIONS
         "MPI_HEADER_PATH=\"${MPI_C_INCLUDE_PATH}\";")
 endif()
@@ -353,11 +349,11 @@ install(DIRECTORY
 if (NGRAPH_TBB_ENABLE)
     install(DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/tbb_build/tbb_release/
         DESTINATION ${NGRAPH_INSTALL_LIB}
-        FILES_MATCHING PATTERN "libtbb.so.*"
+        FILES_MATCHING REGEX "/libtbb${CMAKE_SHARED_LIBRARY_SUFFIX}(\\.[0-9]+)*$"
     )
     install(DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/tbb_build/tbb_debug/
         DESTINATION ${NGRAPH_INSTALL_LIB}
-        FILES_MATCHING PATTERN "libtbb_debug.so.*"
+        FILES_MATCHING REGEX "/libtbb_debug${CMAKE_SHARED_LIBRARY_SUFFIX}(\\.[0-9]+)*$"
     )
 endif()
 

src/ngraph/axis_set.hpp

@@ -21,7 +21,6 @@
 
 namespace ngraph
 {
-#ifdef NO_GLOBAL_TYPE_ALIASES
     /// \brief A set of axes.
     class AxisSet : public std::set<size_t>
     {
@@ -54,8 +53,4 @@ namespace ngraph
             return *this;
         }
     };
-#else
-    // Deprecated definition
-    using AxisSet = std::set<size_t>;
-#endif
 }

src/ngraph/axis_vector.hpp

@@ -21,7 +21,6 @@
 
 namespace ngraph
 {
-#ifdef NO_GLOBAL_TYPE_ALIASES
     /// \brief A vector of axes.
     class AxisVector : public std::vector<size_t>
     {
@@ -64,8 +63,4 @@ namespace ngraph
             return *this;
         }
     };
-#else
-    // Deprecated definition
-    using AxisVector = std::vector<size_t>;
-#endif
 }

src/ngraph/coordinate.hpp

@@ -24,7 +24,6 @@
 
 namespace ngraph
 {
-#ifdef NO_GLOBAL_TYPE_ALIASES
     /// \brief Coordinates for a tensor element
     class Coordinate : public std::vector<size_t>
     {
@@ -73,10 +72,6 @@ namespace ngraph
             return *this;
         }
     };
-#else
-    // Deprecated definition
-    using Coordinate = std::vector<size_t>;
-#endif
 
     // Removes some values from a vector of axis values
     template <typename AXIS_VALUES>

src/ngraph/coordinate_diff.hpp

@@ -21,7 +21,6 @@
 
 namespace ngraph
 {
-#ifdef NO_GLOBAL_TYPE_ALIASES
     /// \brief A difference (signed) of tensor element coordinates.
     class CoordinateDiff : public std::vector<std::ptrdiff_t>
     {
@@ -64,8 +63,4 @@ namespace ngraph
             return *this;
         }
     };
-#else
-    // Deprecated definition
-    using CoordinateDiff = std::vector<std::ptrdiff_t>;
-#endif
 }

src/ngraph/ngraph.hpp

@@ -67,6 +67,7 @@
 #include "ngraph/ops/abs.hpp"
 #include "ngraph/ops/acos.hpp"
 #include "ngraph/ops/add.hpp"
+#include "ngraph/ops/allreduce.hpp"
 #include "ngraph/ops/asin.hpp"
 #include "ngraph/ops/atan.hpp"
 #include "ngraph/ops/avg_pool.hpp"
@@ -130,7 +131,3 @@
 #include "ngraph/shape.hpp"
 #include "ngraph/types/element_type.hpp"
 #include "ngraph/types/type.hpp"
-
-#ifdef NGRAPH_DISTRIBUTED
-#include "ngraph/ops/allreduce.hpp"
-#endif

src/ngraph/node_vector.hpp

@@ -23,8 +23,6 @@ namespace ngraph
 {
     class Node;
 
-#ifdef NO_GLOBAL_TYPE_ALIASES
-
     /// \brief Zero or more nodes.
     class NodeVector : public std::vector<std::shared_ptr<Node>>
     {
@@ -46,10 +44,4 @@ namespace ngraph
 
         NodeVector() {}
     };
-
-#else
-    // Deprecated definitions
-    using NodeVector = std::vector<std::shared_ptr<Node>>;
-    using Nodes = NodeVector;
-#endif
 }

src/ngraph/ops/allreduce.cpp

@@ -12,8 +12,6 @@
 // See the License for the specific language governing permissions and
 // ----------------------------------------------------------------------------
 
-#ifdef NGRAPH_DISTRIBUTED
-
 #include "ngraph/ops/allreduce.hpp"
 
 using namespace std;
@@ -31,5 +29,3 @@ op::AllReduce::AllReduce(const std::shared_ptr<Node>& arg)
         throw ngraph_error("Unsupported data type for AllReduce");
     }
 }
-
-#endif

src/ngraph/ops/allreduce.hpp

@@ -14,8 +14,6 @@
 
 #pragma once
 
-#ifdef NGRAPH_DISTRIBUTED
-
 #include <memory>
 #include "ngraph/ops/util/requires_tensor_view_args.hpp"
 
@@ -40,5 +38,3 @@ namespace ngraph
         };
     }
 }
-
-#endif

src/ngraph/ops/parameter_vector.hpp

@@ -24,8 +24,6 @@ namespace ngraph
     {
         class Parameter;
 
-#ifdef NO_GLOBAL_TYPE_ALIASES
-
         /// \brief Zero or more nodes.
         class ParameterVector : public std::vector<std::shared_ptr<op::Parameter>>
         {
@@ -47,11 +45,5 @@ namespace ngraph
 
             ParameterVector() {}
         };
-
-#else
-        // Deprecated definitions
-        using ParameterVector = std::vector<std::shared_ptr<op::Parameter>>;
-        using Parameters = ParameterVector;
-#endif
     }
 }

src/ngraph/runtime/cpu/cpu_call_frame.cpp

@@ -65,7 +65,9 @@ void runtime::cpu::CPU_CallFrame::tensor_call(
 
     if (runtime::cpu::IsTracingEnabled())
     {
-        GenerateTimeline(m_external_function->get_op_attrs(), ctx->op_durations);
+        GenerateTimeline(m_external_function->get_op_attrs(),
+                         ctx->op_durations,
+                         m_external_function->get_function_name() + ".timeline.json");
     }
 }
 

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -245,6 +245,11 @@ runtime::cpu::CPU_ExternalFunction::CPU_ExternalFunction(
     , m_compiled_function(nullptr)
     , m_emit_timing(std::getenv("NGRAPH_CPU_EMIT_TIMING") != nullptr)
     , m_use_tbb(std::getenv("NGRAPH_CPU_USE_TBB") != nullptr)
+    , m_function_name(function->get_name())
+{
+}
+
+runtime::cpu::CPU_ExternalFunction::~CPU_ExternalFunction()
 {
 }
 
@@ -255,16 +260,14 @@ void runtime::cpu::CPU_ExternalFunction::compile()
         return;
     }
 
-    string function_name = m_function->get_name();
-
-    m_mkldnn_emitter.reset(new MKLDNNEmitter(shared_from_this()));
+    m_mkldnn_emitter.reset(new MKLDNNEmitter());
 
     ngraph::pass::Manager pass_manager;
 
     pass_manager.register_pass<ngraph::pass::CoreFusion>();
     pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
-    pass_manager.register_pass<runtime::cpu::pass::CPUAssignment>(shared_from_this());
-    pass_manager.register_pass<runtime::cpu::pass::CPULayout>(shared_from_this());
+    pass_manager.register_pass<runtime::cpu::pass::CPUAssignment>(this);
+    pass_manager.register_pass<runtime::cpu::pass::CPULayout>(this);
     pass_manager.register_pass<ngraph::pass::Liveness>();
     pass_manager.register_pass<ngraph::pass::MemoryLayout>(s_memory_pool_alignment);
 
@@ -537,7 +540,7 @@ using namespace ngraph::runtime;
         }
 
         // Execution tracing support
-        if (runtime::cpu::IsTracingEnabled() && current_function->get_name() == function_name)
+        if (runtime::cpu::IsTracingEnabled() && current_function->get_name() == m_function_name)
         {
             writer << "cpu::Timestamp start_ts;\n"
                    << "int profiler_count = 0;\n\n";
@@ -688,7 +691,7 @@ using namespace ngraph::runtime;
             // Emit operation prologue
             if (!node->is_parameter() && !node->is_constant())
             {
-                if (current_function->get_name() == function_name)
+                if (current_function->get_name() == m_function_name)
                 {
                     m_op_attrs.emplace_back(
                         node->description(), node_output_names, node_input_names);
@@ -706,7 +709,7 @@ using namespace ngraph::runtime;
                     emit_debug_function_entry(writer, node.get(), in, out);
                 }
                 if (runtime::cpu::IsTracingEnabled() &&
-                    current_function->get_name() == function_name)
+                    current_function->get_name() == m_function_name)
                 {
                     writer << "start_ts = cpu::Clock::now();\n";
                 }
@@ -750,7 +753,7 @@ using namespace ngraph::runtime;
                     emit_debug_function_exit(writer, node.get(), in, out);
                 }
                 if (runtime::cpu::IsTracingEnabled() &&
-                    current_function->get_name() == function_name)
+                    current_function->get_name() == m_function_name)
                 {
                     writer << "ctx->op_durations[profiler_count++] = "
                            << "(std::chrono::duration_cast<cpu::Timescale>(cpu::Clock::now() - "
@@ -848,7 +851,7 @@ using namespace ngraph::runtime;
     // TODO: Cleanup and make this a utility function
 
     file_util::make_directory(s_output_dir);
-    string filename = file_util::path_join(s_output_dir, function_name + "_codegen.cpp");
+    string filename = file_util::path_join(s_output_dir, m_function_name + "_codegen.cpp");
     ofstream out(filename);
     string code = writer.get_code();
     out << code;
@@ -867,7 +870,7 @@ using namespace ngraph::runtime;
     }
     m_execution_engine->add_module(codegen_module);
     m_execution_engine->finalize();
-    m_compiled_function = m_execution_engine->find_function<EntryPoint_t>(function_name);
+    m_compiled_function = m_execution_engine->find_function<EntryPoint_t>(m_function_name);
 
     if (m_compiled_function == nullptr)
     {

src/ngraph/runtime/cpu/cpu_external_function.hpp

@@ -75,6 +75,7 @@ namespace ngraph
             public:
                 CPU_ExternalFunction(const std::shared_ptr<ngraph::Function>& function,
                                      bool release_function = true);
+                ~CPU_ExternalFunction();
                 std::shared_ptr<ngraph::runtime::CallFrame> make_call_frame();
 
                 const LayoutDescriptorPtrs& get_parameter_layout_descriptors();
@@ -86,6 +87,7 @@ namespace ngraph
                     return m_mkldnn_emitter;
                 }
 
+                const std::string& get_function_name() const { return m_function_name; }
             protected:
                 void compile();
 
@@ -123,6 +125,8 @@ namespace ngraph
                 std::vector<OpAttributes> m_op_attrs;
 
                 std::unique_ptr<MKLDNNEmitter> m_mkldnn_emitter;
+
+                std::string m_function_name;
             };
         }
     }

src/ngraph/runtime/cpu/cpu_manager.cpp

@@ -30,7 +30,8 @@ std::shared_ptr<ngraph::runtime::Backend> runtime::cpu::CPU_Manager::allocate_ba
 std::shared_ptr<ngraph::runtime::ExternalFunction>
     runtime::cpu::CPU_Manager::compile(const std::shared_ptr<ngraph::Function>& fun)
 {
-    return std::make_shared<CPU_ExternalFunction>(fun);
+    auto rc = std::make_shared<CPU_ExternalFunction>(fun);
+    return rc;
 }
 
 ngraph::runtime::Manager::Factory runtime::cpu::CPU_Manager::factory =

src/ngraph/runtime/cpu/cpu_tracing.cpp

@@ -42,11 +42,12 @@ void ngraph::runtime::cpu::to_json(nlohmann::json& json, const TraceEvent& event
 }
 
 void ngraph::runtime::cpu::GenerateTimeline(const std::vector<OpAttributes>& op_attrs,
-                                            int64_t* op_durations)
+                                            int64_t* op_durations,
+                                            const std::string& file_name)
 {
     nlohmann::json timeline;
     std::list<TraceEvent> trace;
-    std::ofstream out("timeline.json");
+    std::ofstream out(file_name);
 
     int64_t ts = 0;
     for (size_t i = 0; i < op_attrs.size(); i++)

src/ngraph/runtime/cpu/cpu_tracing.hpp

@@ -69,7 +69,9 @@ namespace ngraph
 
             void to_json(nlohmann::json& json, const TraceEvent& event);
 
-            void GenerateTimeline(const std::vector<OpAttributes>& op_attrs, int64_t* op_durations);
+            void GenerateTimeline(const std::vector<OpAttributes>& op_attrs,
+                                  int64_t* op_durations,
+                                  const std::string& file_name);
             bool IsTracingEnabled();
         }
     }

src/ngraph/runtime/cpu/mkldnn_emitter.hpp

@@ -23,6 +23,7 @@
 #include <mkldnn.hpp>
 
 #include "ngraph/coordinate_diff.hpp"
+#include "ngraph/shape.hpp"
 #include "ngraph/strides.hpp"
 
 namespace ngraph
@@ -37,11 +38,7 @@ namespace ngraph
             class MKLDNNEmitter
             {
             public:
-                MKLDNNEmitter(std::shared_ptr<CPU_ExternalFunction> ef)
-                    : external_function(ef)
-                {
-                }
-
+                MKLDNNEmitter() {}
                 const std::vector<mkldnn::primitive*>& get_mkldnn_primitives() const;
 
                 size_t insert_primitive(mkldnn::primitive* primitive);
@@ -69,6 +66,37 @@ namespace ngraph
                                                  const ngraph::CoordinateDiff& padding_below,
                                                  const ngraph::CoordinateDiff& padding_above);
 
+                size_t
+                    build_convolution_backward_weights(const mkldnn::memory::desc& input_desc,
+                                                       const mkldnn::memory::desc& delta_desc,
+                                                       const mkldnn::memory::desc& result_desc,
+                                                       const ngraph::Strides& strides,
+                                                       const ngraph::Strides& dilation_strides,
+                                                       const ngraph::CoordinateDiff& padding_below,
+                                                       const ngraph::CoordinateDiff& padding_above);
+
+                size_t build_convolution_backward_data(const mkldnn::memory::desc& weights_desc,
+                                                       const mkldnn::memory::desc& delta_desc,
+                                                       const mkldnn::memory::desc& result_desc,
+                                                       const ngraph::Strides& strides,
+                                                       const ngraph::Strides& dilation_strides,
+                                                       const ngraph::CoordinateDiff& padding_below,
+                                                       const ngraph::CoordinateDiff& padding_above);
+
+                size_t build_pooling_forward(mkldnn::algorithm pooling_algorithm,
+                                             const mkldnn::memory::desc& input_desc,
+                                             const mkldnn::memory::desc& result_desc,
+                                             const ngraph::Strides& window_strides,
+                                             const ngraph::Shape& window_shape,
+                                             const ngraph::Shape& padding_below,
+                                             const ngraph::Shape& padding_above);
+
+                size_t build_reorder(const mkldnn::memory::desc& input_desc,
+                                     const mkldnn::memory::desc& result_desc);
+
+                size_t build_relu_forward(const mkldnn::memory::desc& input_desc,
+                                          const mkldnn::memory::desc& result_desc);
+
                 size_t build_elementwise_add(
                     const mkldnn::memory::desc& input0_data_desc,
                     const mkldnn::memory::desc& input1_data_desc,
@@ -77,10 +105,9 @@ namespace ngraph
                     const std::vector<mkldnn::memory::primitive_desc>& input_pd);
 
             private:
-                std::shared_ptr<CPU_ExternalFunction> external_function;
-                std::vector<mkldnn::primitive*> mkldnn_primitives;
-                std::vector<mkldnn::stream> mkldnn_streams;
-                std::unordered_map<size_t, std::vector<size_t>> primitive_deps;
+                std::vector<mkldnn::primitive*> m_mkldnn_primitives;
+                std::vector<mkldnn::stream> m_mkldnn_streams;
+                std::unordered_map<size_t, std::vector<size_t>> m_primitive_deps;
             };
         }
     }

src/ngraph/runtime/cpu/pass/cpu_assignment.cpp

@@ -256,7 +256,7 @@ bool runtime::cpu::pass::CPUAssignment::run_on_call_graph(
         auto handler = s_dispatcher.find(TI(n));
         if (handler != s_dispatcher.end())
         {
-            handler->second(m_external_function.get(), node.get());
+            handler->second(m_external_function, node.get());
         }
     }
 

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

@@ -39,7 +39,7 @@ namespace ngraph
                 class CPUAssignment : public ngraph::pass::CallGraphPass
                 {
                 public:
-                    CPUAssignment(std::shared_ptr<CPU_ExternalFunction> external_function)
+                    CPUAssignment(CPU_ExternalFunction* external_function)
                         : m_external_function(external_function)
                     {
                     }
@@ -56,7 +56,7 @@ namespace ngraph
                     }
 
                 private:
-                    std::shared_ptr<CPU_ExternalFunction> m_external_function;
+                    CPU_ExternalFunction* m_external_function;
                 };
             }
         }

src/ngraph/runtime/cpu/pass/cpu_layout.cpp

@@ -719,11 +719,11 @@ bool runtime::cpu::pass::CPULayout::run_on_call_graph(const std::list<std::share
         auto handler = s_dispatcher.find(TI(n));
         if (handler != s_dispatcher.end())
         {
-            handler->second(m_external_function.get(), node);
+            handler->second(m_external_function, node);
         }
         else
         {
-            set_default_layouts(m_external_function.get(), node);
+            set_default_layouts(m_external_function, node);
         }
     }
 

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

@@ -39,7 +39,7 @@ namespace ngraph
                 class CPULayout : public ngraph::pass::CallGraphPass
                 {
                 public:
-                    CPULayout(std::shared_ptr<CPU_ExternalFunction> external_function)
+                    CPULayout(CPU_ExternalFunction* external_function)
                         : m_external_function(external_function)
                     {
                     }
@@ -52,7 +52,7 @@ namespace ngraph
                                std::shared_ptr<ngraph::Node> node);
 
                 private:
-                    std::shared_ptr<CPU_ExternalFunction> m_external_function;
+                    CPU_ExternalFunction* m_external_function;
                     static std::shared_ptr<Node> insert_input_conversions(
                         CPU_ExternalFunction* external_function,
                         std::shared_ptr<Node>& node,

src/ngraph/runtime/gpu/gpu_emitter.cpp

@@ -51,29 +51,6 @@
 using namespace std;
 using namespace ngraph;
 
-#define NVRTC_SAFE_CALL(x)                                                                         \
-    do                                                                                             \
-    {                                                                                              \
-        nvrtcResult result = x;                                                                    \
-        if (result != NVRTC_SUCCESS)                                                               \
-        {                                                                                          \
-            throw std::runtime_error("\nerror: " #x " failed with error " +                        \
-                                     nvrtcGetErrorString(result));                                 \
-        }                                                                                          \
-    } while (0)
-
-#define CUDA_SAFE_CALL(x)                                                                          \
-    do                                                                                             \
-    {                                                                                              \
-        CUresult result = x;                                                                       \
-        if (result != CUDA_SUCCESS)                                                                \
-        {                                                                                          \
-            const char* msg;                                                                       \
-            cuGetErrorName(result, &msg);                                                          \
-            throw std::runtime_error("\nerror: " #x " failed with error " + std::string(msg);      \
-        }                                                                                          \
-    } while (0)
-
 void runtime::gpu::GPU_Emitter::EmitNop(codegen::CodeWriter& writer,
                                         const ngraph::Node* n,
                                         const vector<runtime::gpu::GPU_TensorViewWrapper>& args,
@@ -523,7 +500,6 @@ void runtime::gpu::GPU_Emitter::EmitReshape(codegen::CodeWriter& writer,
     {
         result_shape_product *= i;
     }
-
     // If there is no layout change or we are just going from 1^n to 1^m or a zero-size tensor,
     //  we can just copy.
     if (same_layout || result_shape_product < 2)
@@ -531,7 +507,7 @@ void runtime::gpu::GPU_Emitter::EmitReshape(codegen::CodeWriter& writer,
         writer << "{   // " << n->get_name() << " 1\n";
         writer.indent++;
         writer << "runtime::gpu::cuda_memcpyDtD(" << out[0].get_name() << ", " << args[0].get_name()
-               << ", " << out[0].get_size() << "," << out[0].get_element_type().size() << ");\n";
+               << ", " << out[0].get_size() << " * " << out[0].get_element_type().size() << ");\n";
         writer.indent--;
         writer << "}\n";
     }
@@ -541,8 +517,9 @@ void runtime::gpu::GPU_Emitter::EmitReshape(codegen::CodeWriter& writer,
         // TODO Assert arg0_shape[0] == arg1_shape[0]?
         writer << "{   // " << n->get_name() << "\n";
         writer.indent++;
-        writer << "static const float alpha = 1.0;\n";
-        writer << "static const float beta = 0.0;\n";
+        writer << "const float alpha = 1.0;\n";
+        writer << "const float beta = 0;\n";
+        writer << "cublasSetPointerMode(cublas_handle, CUBLAS_POINTER_MODE_HOST);\n";
         writer << "cublasSgeam("
                << "cublas_handle,"
                << "CUBLAS_OP_T,"
@@ -551,7 +528,8 @@ void runtime::gpu::GPU_Emitter::EmitReshape(codegen::CodeWriter& writer,
                << args[0].get_name() << "," << arg_shape[1] << ","
                << "&beta," // beta
                << args[0].get_name() << "," << arg_shape[1] << "," << out[0].get_name() << ","
-               << out[0].get_shape()[1] << ");\n";
+               << result_shape[1] << ");\n";
+        writer << "cublasSetPointerMode(cublas_handle, CUBLAS_POINTER_MODE_DEVICE);\n";
         writer.indent--;
         writer << "}\n";
     }

src/ngraph/runtime/gpu/gpu_util.cpp

@@ -54,18 +54,17 @@ void* runtime::gpu::create_gpu_buffer(size_t buffer_size)
     return allocated_buffer_pool;
 }
 
-void runtime::gpu::cuda_memcpyDtD(void* d, void* s, size_t element_count, size_t element_size)
+void runtime::gpu::cuda_memcpyDtD(void* dst, void* src, size_t buffer_size)
 {
-    size_t size_in_bytes = element_size * element_count;
-    cudaMemcpy(d, s, size_in_bytes, cudaMemcpyDeviceToDevice);
+    cudaMemcpy(dst, src, buffer_size, cudaMemcpyDeviceToDevice);
 }
 
-void runtime::gpu::cuda_memcpyHtD(void* d, void* s, size_t buffer_size)
+void runtime::gpu::cuda_memcpyHtD(void* dst, void* src, size_t buffer_size)
 {
-    cudaMemcpy(d, s, buffer_size, cudaMemcpyHostToDevice);
+    cudaMemcpy(dst, src, buffer_size, cudaMemcpyHostToDevice);
 }
 
-void runtime::gpu::cuda_memset(void* d, int value, size_t buffer_size)
+void runtime::gpu::cuda_memset(void* dst, int value, size_t buffer_size)
 {
-    cudaMemset(d, value, buffer_size);
+    cudaMemset(dst, value, buffer_size);
 }

src/ngraph/runtime/gpu/gpu_util.hpp

@@ -61,9 +61,9 @@ namespace ngraph
             void print_gpu_f32_tensor(void* p, size_t element_count, size_t element_size);
             void check_cuda_errors(CUresult err);
             void* create_gpu_buffer(size_t buffer_size);
-            void cuda_memcpyDtD(void* d, void* s, size_t element_count, size_t element_size);
-            void cuda_memcpyHtD(void* d, void* s, size_t buffer_size);
-            void cuda_memset(void* d, int value, size_t buffer_size);
+            void cuda_memcpyDtD(void* dst, void* src, size_t buffer_size);
+            void cuda_memcpyHtD(void* dst, void* src, size_t buffer_size);
+            void cuda_memset(void* dst, int value, size_t buffer_size);
         }
     }
 }

src/ngraph/serializer.cpp

@@ -19,6 +19,7 @@
 #include "ngraph/ops/abs.hpp"
 #include "ngraph/ops/acos.hpp"
 #include "ngraph/ops/add.hpp"
+#include "ngraph/ops/allreduce.hpp"
 #include "ngraph/ops/asin.hpp"
 #include "ngraph/ops/atan.hpp"
 #include "ngraph/ops/avg_pool.hpp"
@@ -75,15 +76,27 @@
 #include "ngraph/ops/tan.hpp"
 #include "ngraph/ops/tanh.hpp"
 #include "ngraph/util.hpp"
-
-#ifdef NGRAPH_DISTRIBUTED
-#include "ngraph/ops/allreduce.hpp"
-#endif
+#include "nlohmann/json.hpp"
 
 using namespace ngraph;
 using namespace std;
 using json = nlohmann::json;
 
+template <typename T>
+T get_or_default(nlohmann::json& j, const std::string& key, const T& default_value)
+{
+    T rc;
+    try
+    {
+        rc = j.at(key).get<T>();
+    }
+    catch (...)
+    {
+        rc = default_value;
+    }
+    return rc;
+}
+
 static std::shared_ptr<ngraph::Function>
     read_function(const json&, std::unordered_map<std::string, std::shared_ptr<Function>>&);
 
@@ -263,12 +276,10 @@ static shared_ptr<ngraph::Function>
         {
             node = make_shared<op::Add>(args[0], args[1]);
         }
-#ifdef NGRAPH_DISTRIBUTED
         else if (node_op == "AllReduce")
         {
             node = make_shared<op::AllReduce>(args[0]);
         }
-#endif
         else if (node_op == "Asin")
         {
             node = make_shared<op::Asin>(args[0]);

src/ngraph/serializer.hpp

@@ -21,26 +21,10 @@
 
 #include "ngraph/function.hpp"
 #include "ngraph/node.hpp"
-#include "nlohmann/json.hpp"
 
 namespace ngraph
 {
     std::string serialize(std::shared_ptr<ngraph::Function>, size_t indent = 0);
     std::shared_ptr<ngraph::Function> deserialize(std::istream&);
     std::shared_ptr<ngraph::Function> deserialize(const std::string&);
-
-    template <typename T>
-    T get_or_default(nlohmann::json& j, const std::string& key, const T& default_value)
-    {
-        T rc;
-        try
-        {
-            rc = j.at(key).get<T>();
-        }
-        catch (...)
-        {
-            rc = default_value;
-        }
-        return rc;
-    }
 }

src/ngraph/shape.hpp

@@ -25,7 +25,6 @@
 
 namespace ngraph
 {
-#ifdef NO_GLOBAL_TYPE_ALIASES
     /// \brief Shape for a tensor.
     class Shape : public std::vector<size_t>
     {
@@ -68,10 +67,6 @@ namespace ngraph
             return *this;
         }
     };
-#else
-    // Deprecated definition
-    using Shape = std::vector<size_t>;
-#endif
 
     /// Number of elements in spanned by a shape
     size_t shape_size(const Shape& shape);
@@ -81,5 +76,4 @@ namespace ngraph
 
     inline bool is_scalar(const Shape& shape) { return 0 == shape.size(); }
     inline bool is_vector(const Shape& shape) { return 1 == shape.size(); }
-    Shape project_shape(const Shape& shape, const AxisSet& deleted_axes);
 }

src/ngraph/strides.hpp

@@ -20,7 +20,6 @@
 
 namespace ngraph
 {
-#ifdef NO_GLOBAL_TYPE_ALIASES
     /// \brief Strides for a tensor.
     class Strides : public std::vector<size_t>
     {
@@ -63,8 +62,4 @@ namespace ngraph
             return *this;
         }
     };
-#else
-    // Deprecated definition
-    using Strides = std::vector<size_t>;
-#endif
 }

src/tools/CMakeLists.txt

@@ -14,26 +14,4 @@
 # limitations under the License.
 # ******************************************************************************
 
-if(MKLDNN_INCLUDE_DIR)
-    link_directories(${MKLDNN_LIB_DIR})
-endif()
-
-if (NGRAPH_CPU_ENABLE)
-    set (SRC
-        nbench.cpp
-        ${PROJECT_SOURCE_DIR}/test/util/benchmark.cpp
-    )
-
-    add_executable(nbench ${SRC})
-    add_dependencies(nbench ngraph)
-
-    set(HEADER_SEARCH_DEFINES
-        "NGRAPH_HEADERS_PATH=\"${NGRAPH_INCLUDE_PATH}\""
-    )
-
-    target_link_libraries(nbench ngraph)
-    include_directories(SYSTEM ${JSON_INCLUDE_DIR})
-
-    set_source_files_properties(nbench.cpp PROPERTIES COMPILE_DEFINITIONS "${HEADER_SEARCH_DEFINES}")
-
-endif()
+add_subdirectory(nbench)

src/tools/nbench/CMakeLists.txt

+# ******************************************************************************
+# Copyright 2017-2018 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ******************************************************************************
+
+if(MKLDNN_INCLUDE_DIR)
+    link_directories(${MKLDNN_LIB_DIR})
+endif()
+
+if (NGRAPH_CPU_ENABLE)
+    set (SRC
+        nbench.cpp
+        ${PROJECT_SOURCE_DIR}/test/util/benchmark.cpp
+    )
+
+    add_executable(nbench ${SRC})
+    add_dependencies(nbench ngraph)
+
+    set(HEADER_SEARCH_DEFINES
+        "NGRAPH_HEADERS_PATH=\"${NGRAPH_INCLUDE_PATH}\""
+    )
+
+    target_link_libraries(nbench ngraph)
+    include_directories("${PROJECT_SOURCE_DIR}/test")
+
+    set_source_files_properties(nbench.cpp PROPERTIES COMPILE_DEFINITIONS "${HEADER_SEARCH_DEFINES}")
+
+endif()

src/tools/nbench.cpp → src/tools/nbench/nbench.cpp

@@ -24,8 +24,8 @@
 #include <ngraph/runtime/backend.hpp>
 #include <ngraph/runtime/call_frame.hpp>
 #include <ngraph/runtime/manager.hpp>
-#include "../../test/util/benchmark.hpp"
-#include "../../test/util/test_tools.hpp"
+#include "util/benchmark.hpp"
+#include "util/test_tools.hpp"
 using namespace std;
 
 int main(int argc, char** argv)
@@ -66,16 +66,16 @@ int main(int argc, char** argv)
     if (failed)
     {
         cout << R"###(
-DESCRIPTION                                                         
-    Benchmark ngraph json model with given backend.                 
-                                                                    
-SYNOPSIS                                                            
+DESCRIPTION
+    Benchmark ngraph json model with given backend.
+
+SYNOPSIS
         nbench [-f <filename>] [-b <backend>] [-i <iterations>]
-                                                                    
-OPTIONS                                                             
-        -f          model json file to use (default: model.json)    
-        -b          Backend to use (default: INTERPRETER)           
-        -i          Iterations (default: 10)                        
+
+OPTIONS
+        -f          model json file to use (default: model.json)
+        -b          Backend to use (default: INTERPRETER)
+        -i          Iterations (default: 10)
 )###";
         return 1;
     }

test/backend_test.in.cpp

@@ -43,6 +43,29 @@ static const vector<element::Type> s_known_element_types = {element::from<float>
                                                             element::from<uint32_t>(),
                                                             element::from<uint64_t>()};
 
+TEST(${BACKEND_NAME}, component_cleanup)
+{
+    shared_ptr<runtime::Backend> backend;
+    shared_ptr<runtime::ExternalFunction> external;
+    shared_ptr<runtime::CallFrame> cf;
+    {
+        Shape shape{2, 2};
+        auto A = make_shared<op::Parameter>(element::f32, shape);
+        auto B = make_shared<op::Parameter>(element::f32, shape);
+        auto f = make_shared<Function>(A + B, op::ParameterVector{A, B});
+
+        auto manager = runtime::Manager::get("${BACKEND_NAME}");
+        external = manager->compile(f);
+        backend = manager->allocate_backend();
+        cf = backend->make_call_frame(external);
+    }
+    EXPECT_EQ(cf.use_count(), 1);
+    cf = nullptr;
+    EXPECT_EQ(backend.use_count(), 1);
+    backend = nullptr;
+    EXPECT_EQ(external.use_count(), 1);
+}
+
 TEST(${BACKEND_NAME}, aliased_output)
 {
     Shape shape{2, 2};
@@ -138,7 +161,7 @@ TEST(${BACKEND_NAME}, abc)
     auto f = make_shared<Function>((A + B) * C, op::ParameterVector{A, B, C});
 
     auto manager = runtime::Manager::get("${BACKEND_NAME}");
-    auto external = manager->compile(f);
+    shared_ptr<runtime::ExternalFunction> external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
 
@@ -2373,7 +2396,6 @@ TEST(${BACKEND_NAME}, reshape_s2t)
 
 TEST(${BACKEND_NAME}, reshape_v2m_col)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_a{3};
     auto A = make_shared<op::Parameter>(element::f32, shape_a);
     Shape shape_r{3, 1};
@@ -2396,7 +2418,6 @@ TEST(${BACKEND_NAME}, reshape_v2m_col)
 
 TEST(${BACKEND_NAME}, reshape_v2m_row)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_a{3};
     auto A = make_shared<op::Parameter>(element::f32, shape_a);
     Shape shape_r{1, 3};
@@ -2442,7 +2463,6 @@ TEST(${BACKEND_NAME}, reshape_v2t_middle)
 
 TEST(${BACKEND_NAME}, reshape_m2m_same)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_a{3, 3};
     auto A = make_shared<op::Parameter>(element::f32, shape_a);
     Shape shape_r{3, 3};
@@ -2465,7 +2485,6 @@ TEST(${BACKEND_NAME}, reshape_m2m_same)
 
 TEST(${BACKEND_NAME}, reshape_m2m_transpose)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_a{3, 3};
     auto A = make_shared<op::Parameter>(element::f32, shape_a);
     Shape shape_r{3, 3};
@@ -2488,7 +2507,6 @@ TEST(${BACKEND_NAME}, reshape_m2m_transpose)
 
 TEST(${BACKEND_NAME}, reshape_m2m_dim_change_transpose)
 {
-    SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
     Shape shape_a{3, 2};
     auto A = make_shared<op::Parameter>(element::f32, shape_a);
     Shape shape_r{2, 3};

test/serialize.cpp

@@ -30,6 +30,21 @@ using namespace std;
 using namespace ngraph;
 using json = nlohmann::json;
 
+template <typename T>
+T get_or_default(nlohmann::json& j, const std::string& key, const T& default_value)
+{
+    T rc;
+    try
+    {
+        rc = j.at(key).get<T>();
+    }
+    catch (...)
+    {
+        rc = default_value;
+    }
+    return rc;
+}
+
 TEST(serialize, main)
 {
     // First create "f(A,B,C) = (A+B)*C".

test/util/benchmark.cpp

@@ -13,6 +13,9 @@
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *******************************************************************************/
+
+#include <iomanip>
+
 #include "benchmark.hpp"
 #include "ngraph/runtime/backend.hpp"
 #include "ngraph/runtime/call_frame.hpp"