Merge branch 'master' into rearhart/plaidml

cmake/external_mkldnn.cmake

@@ -18,10 +18,10 @@ include(ExternalProject)
 
 # Includes blas 3.8.0 in mkldnn
 set(NGRAPH_MKLDNN_SHORT_VERSION 0)
-set(NGRAPH_MKLDNN_FULL_VERSION 0.19.0.0)
-set(NGRAPH_MKLDNN_VERSION "v0.19")
+set(NGRAPH_MKLDNN_FULL_VERSION 0.20.0.0)
+set(NGRAPH_MKLDNN_VERSION "v0.20")
 set(NGRAPH_MKLDNN_SUB_VERSION "2019.0.5.20190502")
-set(NGRAPH_MKLDNN_GIT_TAG "027de76")
+set(NGRAPH_MKLDNN_GIT_TAG "v0.20")
 
 #------------------------------------------------------------------------------
 # Fetch and install MKL-DNN

cmake/mkldnn.patch

@@ -28,16 +28,3 @@ index f10feb20..05f47961 100644
  set_property(TARGET ${LIB_NAME} PROPERTY PUBLIC_HEADER ${HEADERS})
  
  target_include_directories(${LIB_NAME} PUBLIC
-diff --git a/src/cpu/jit_avx512_common_conv_kernel.cpp b/src/cpu/jit_avx512_common_conv_kernel.cpp
-index 1bb98fa43..b8b54401f 100644
---- a/src/cpu/jit_avx512_common_conv_kernel.cpp
-+++ b/src/cpu/jit_avx512_common_conv_kernel.cpp
-@@ -3055,7 +3055,7 @@ void jit_avx512_common_conv_bwd_weights_kernel_f32::bias_kernel_3d() {
- void jit_avx512_common_conv_bwd_weights_kernel_f32
-     ::compute_oh_loop_common()
- {
--    assert(jcp.harness == harness_mb_reduction);
-+    assert(one_of(jcp.harness, harness_mb_reduction, harness_3d_reduction));
-     int b_pad = jcp.b_pad;
-     int t_pad = jcp.t_pad;
-     bool is_dilated = jcp.dilate_h != 0;

python/test_requirements.txt

 pytest
 tox
+pydocstyle==3.0.0
 flake8
 flake8-commas
 flake8-comprehensions

src/ngraph/runtime/allocator.cpp

@@ -49,7 +49,8 @@ public:
     }
 };
 
-std::unique_ptr<ngraph::runtime::Allocator> ngraph::runtime::create_default_allocator()
+ngraph::runtime::Allocator* ngraph::runtime::get_default_allocator()
 {
-    return std::unique_ptr<DefaultAllocator>(new DefaultAllocator());
+    static std::unique_ptr<DefaultAllocator> allocator(new DefaultAllocator());
+    return allocator.get();
 }

src/ngraph/runtime/allocator.hpp

@@ -30,7 +30,7 @@ namespace ngraph
         class DefaultAllocator;
         /// \brief Create a default allocator that calls into system
         ///        allocation libraries
-        std::unique_ptr<Allocator> create_default_allocator();
+        ngraph::runtime::Allocator* get_default_allocator();
     }
 }
 

src/ngraph/runtime/cpu/cpu_backend.cpp

@@ -185,7 +185,7 @@ runtime::Allocator* runtime::cpu::CPU_Backend::get_host_memory_allocator()
 {
     if (!m_allocator)
     {
-        m_allocator = create_default_allocator();
+        return runtime::get_default_allocator();
     }
     return m_allocator.get();
 }

src/ngraph/runtime/generic_cpu/CMakeLists.txt

@@ -15,10 +15,10 @@
 # ******************************************************************************
 
 if (NGRAPH_GENERIC_CPU_ENABLE)
-    find_package(OpenMP)
-    if (OPENMP_FOUND)
-        add_compile_options(${OpenMP_CXX_FLAGS})
-    endif()
+    # find_package(OpenMP)
+    # if (OPENMP_FOUND)
+    #     add_compile_options(${OpenMP_CXX_FLAGS})
+    # endif()
     add_library(gcpu_backend SHARED gcpu_backend.cpp gcpu_executable.cpp node_wrapper.cpp)
     if(NGRAPH_LIB_VERSIONING_ENABLE)
         set_target_properties(gcpu_backend PROPERTIES

src/ngraph/runtime/generic_cpu/gcpu_backend.cpp

@@ -52,14 +52,14 @@ runtime::gcpu::GCPUBackend::GCPUBackend(const vector<string>& unsupported_op_nam
 shared_ptr<runtime::Tensor> runtime::gcpu::GCPUBackend::create_tensor(const element::Type& type,
                                                                       const Shape& shape)
 {
-    return make_shared<runtime::HostTensor>(type, shape, this);
+    return make_shared<runtime::HostTensor>(type, shape);
 }
 
 shared_ptr<runtime::Tensor> runtime::gcpu::GCPUBackend::create_tensor(const element::Type& type,
                                                                       const Shape& shape,
                                                                       void* memory_pointer)
 {
-    return make_shared<runtime::HostTensor>(type, shape, memory_pointer, this);
+    return make_shared<runtime::HostTensor>(type, shape, memory_pointer);
 }
 
 shared_ptr<runtime::Executable>

src/ngraph/runtime/generic_cpu/gcpu_executable.cpp

@@ -15,17 +15,22 @@
 //*****************************************************************************
 
 #include "ngraph/runtime/generic_cpu/gcpu_executable.hpp"
+#include "ngraph/cpio.hpp"
 #include "ngraph/descriptor/layout/dense_tensor_layout.hpp"
 #include "ngraph/except.hpp"
 #include "ngraph/op/convert.hpp"
 #include "ngraph/op/select.hpp"
 #include "ngraph/op/util/binary_elementwise_comparison.hpp"
 #include "ngraph/pass/assign_layout.hpp"
+#include "ngraph/pass/core_fusion.hpp"
+#include "ngraph/pass/fused_op_decomposition.hpp"
+#include "ngraph/pass/implicit_broadcast_elimination.hpp"
 #include "ngraph/pass/like_replacement.hpp"
 #include "ngraph/pass/liveness.hpp"
 #include "ngraph/pass/manager.hpp"
 #include "ngraph/pass/memory_layout.hpp"
 #include "ngraph/runtime/backend_manager.hpp"
+#include "ngraph/serializer.hpp"
 #include "ngraph/util.hpp"
 
 using namespace std;
@@ -35,21 +40,35 @@ using descriptor::layout::DenseTensorLayout;
 
 runtime::gcpu::GCPUExecutable::GCPUExecutable(const shared_ptr<Function>& function,
                                               bool enable_performance_collection)
+    : m_is_compiled{true}
+    , m_performance_counters_enabled{enable_performance_collection}
 {
+    m_function = clone_function(*function);
+    pass::Manager pass_manager;
+    pass_manager.register_pass<pass::LikeReplacement>();
+    pass_manager.register_pass<pass::FusedOpDecomposition>();
+    pass_manager.register_pass<pass::ImplicitBroadcastElimination>();
+    pass_manager.register_pass<pass::AssignLayout<DenseTensorLayout>>();
+    pass_manager.register_pass<pass::Liveness>();
+    pass_manager.run_passes(m_function);
+
+    for (const shared_ptr<Node>& node : m_function->get_ordered_ops())
     {
-        m_is_compiled = true;
-        pass::Manager pass_manager;
-        pass_manager.register_pass<pass::LikeReplacement>();
-        pass_manager.register_pass<pass::AssignLayout<DenseTensorLayout>>();
-        pass_manager.register_pass<pass::Liveness>();
-        pass_manager.run_passes(function);
+        m_wrapped_nodes.emplace_back(node);
+    }
+    set_parameters_and_results(*m_function);
+}
 
-        for (const shared_ptr<Node>& node : function->get_ordered_ops())
-        {
-            m_wrapped_nodes.emplace_back(node);
-        }
+runtime::gcpu::GCPUExecutable::GCPUExecutable(const std::string& model_string)
+    : m_is_compiled{true}
+    , m_performance_counters_enabled{false}
+{
+    m_function = deserialize(model_string);
+    for (const shared_ptr<Node>& node : m_function->get_ordered_ops())
+    {
+        m_wrapped_nodes.emplace_back(node);
     }
-    set_parameters_and_results(*function);
+    set_parameters_and_results(*m_function);
 }
 
 bool runtime::gcpu::GCPUExecutable::call(const vector<shared_ptr<runtime::Tensor>>& outputs,
@@ -82,7 +101,7 @@ bool runtime::gcpu::GCPUExecutable::call(const vector<shared_ptr<runtime::Tensor
     {
         for (size_t i = 0; i < param->get_output_size(); ++i)
         {
-            descriptor::Tensor* tensor = param->get_output_tensor_ptr(i).get();
+            descriptor::Tensor* tensor = &param->output(i).get_tensor();
             tensor_map.insert({tensor, func_inputs[input_count++]});
         }
     }
@@ -95,14 +114,14 @@ bool runtime::gcpu::GCPUExecutable::call(const vector<shared_ptr<runtime::Tensor
         {
             throw ngraph_error("One of function's outputs isn't op::Result");
         }
-        descriptor::Tensor* tensor = output->get_output_tensor_ptr(0).get();
+        descriptor::Tensor* tensor = &output->output(0).get_tensor();
         tensor_map.insert({tensor, func_outputs[output_count]});
     }
 
     // for each ordered op in the graph
     for (const NodeWrapper& wrapped : m_wrapped_nodes)
     {
-        const Node* op = &wrapped.get_node();
+        auto op = wrapped.get_node();
         auto type_id = wrapped.get_typeid();
         if (type_id == OP_TYPEID::Parameter)
         {
@@ -111,9 +130,9 @@ bool runtime::gcpu::GCPUExecutable::call(const vector<shared_ptr<runtime::Tensor
 
         // get op inputs from map
         vector<shared_ptr<HostTensor>> op_inputs;
-        for (const descriptor::Input& input : op->get_inputs())
+        for (auto input : op->inputs())
         {
-            descriptor::Tensor* tensor = input.get_output().get_tensor_ptr().get();
+            descriptor::Tensor* tensor = &input.get_tensor();
             op_inputs.push_back(tensor_map.at(tensor));
         }
 
@@ -121,14 +140,14 @@ bool runtime::gcpu::GCPUExecutable::call(const vector<shared_ptr<runtime::Tensor
         vector<shared_ptr<HostTensor>> op_outputs;
         for (size_t i = 0; i < op->get_output_size(); ++i)
         {
-            descriptor::Tensor* tensor = op->get_output_tensor_ptr(i).get();
+            descriptor::Tensor* tensor = &op->output(i).get_tensor();
             shared_ptr<HostTensor> host_tensor;
             auto it = tensor_map.find(tensor);
             if (it == tensor_map.end())
             {
                 const Shape& shape = op->get_output_shape(i);
                 const element::Type& type = op->get_output_element_type(i);
-                string name = op->get_output_tensor(i).get_name();
+                string name = op->output(i).get_tensor().get_name();
                 host_tensor = make_shared<runtime::HostTensor>(type, shape, name);
                 tensor_map.insert({tensor, host_tensor});
             }
@@ -177,7 +196,7 @@ bool runtime::gcpu::GCPUExecutable::call(const vector<shared_ptr<runtime::Tensor
         }
         if (m_nan_check_enabled)
         {
-            perform_nan_check(op_outputs, op);
+            perform_nan_check(op_outputs, op.get());
         }
     }
 
@@ -186,19 +205,9 @@ bool runtime::gcpu::GCPUExecutable::call(const vector<shared_ptr<runtime::Tensor
 
 void runtime::gcpu::GCPUExecutable::generate_calls(const element::Type& type,
                                                    const NodeWrapper& op,
-                                                   const vector<shared_ptr<HostTensor>>& outputs,
-                                                   const vector<shared_ptr<HostTensor>>& inputs)
+                                                   const vector<shared_ptr<HostTensor>>& out,
+                                                   const vector<shared_ptr<HostTensor>>& in)
 {
-    vector<void*> out;
-    vector<const void*> in;
-    for (auto t : outputs)
-    {
-        out.push_back(t->get_data_ptr());
-    }
-    for (auto t : inputs)
-    {
-        in.push_back(t->get_data_ptr());
-    }
     stringstream ss;
     switch (type.get_type_enum())
     {
@@ -216,7 +225,8 @@ void runtime::gcpu::GCPUExecutable::generate_calls(const element::Type& type,
     case element::Type_t::undefined:
     case element::Type_t::dynamic:
     case element::Type_t::bf16:
-        ss << "unsupported element type " << type << " op " << op.get_node().get_name();
+    case element::Type_t::f16:
+        ss << "unsupported element type " << type << " op " << op.get_node()->get_name();
         throw ngraph_error(ss.str());
     }
 }
@@ -229,11 +239,9 @@ void runtime::gcpu::GCPUExecutable::set_nan_check(bool enable)
 vector<runtime::PerformanceCounter> runtime::gcpu::GCPUExecutable::get_performance_data() const
 {
     vector<runtime::PerformanceCounter> rc;
-    for (const pair<const Node*, stopwatch> p : m_timer_map)
+    for (const pair<shared_ptr<const Node>, stopwatch> p : m_timer_map)
     {
-        rc.emplace_back(p.first->get_name().c_str(),
-                        p.second.get_total_microseconds(),
-                        p.second.get_call_count());
+        rc.emplace_back(p.first, p.second.get_total_microseconds(), p.second.get_call_count());
     }
     return rc;
 }
@@ -286,3 +294,12 @@ void runtime::gcpu::GCPUExecutable::perform_nan_check(const vector<shared_ptr<Ho
         arg_number++;
     }
 }
+
+void runtime::gcpu::GCPUExecutable::save(ostream& out)
+{
+    cpio::Writer writer(out);
+    string si = "INTERPRETER Save File 1.0";
+    writer.write("save_info", si.data(), si.size());
+    string model = serialize(m_function, 0);
+    writer.write("model", model.data(), model.size());
+}

src/ngraph/runtime/generic_cpu/kernel/broadcast.hpp

@@ -140,6 +140,91 @@ namespace ngraph
                     }
                 }
 
+                template <typename T>
+                void broadcast_5d(const T* in,
+                                  T* out,
+                                  const Shape& in_shape,
+                                  const Shape& out_shape,
+                                  const AxisSet& broadcast_axes)
+                {
+                    size_t index[5];
+                    size_t* out_index = 0;
+                    for (size_t i = 0; i < 5; i++)
+                    {
+                        if (broadcast_axes.count(i) == 0)
+                        {
+                            out_index = &index[i];
+                            break;
+                        }
+                    }
+                    for (index[0] = 0; index[0] < out_shape[0]; ++index[0])
+                    {
+                        for (index[1] = 0; index[1] < out_shape[1]; ++index[1])
+                        {
+                            for (index[2] = 0; index[2] < out_shape[2]; ++index[2])
+                            {
+                                for (index[3] = 0; index[3] < out_shape[3]; ++index[3])
+                                {
+                                    for (index[4] = 0; index[4] < out_shape[4]; ++index[4])
+                                    {
+                                        out[index[0] * out_shape[1] * out_shape[2] * out_shape[3] *
+                                                out_shape[4] +
+                                            index[1] * out_shape[2] * out_shape[3] * out_shape[4] +
+                                            index[2] * out_shape[3] * out_shape[4] +
+                                            index[3] * out_shape[4] + index[4]] = in[*out_index];
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+
+                template <typename T>
+                void broadcast_6d(const T* in,
+                                  T* out,
+                                  const Shape& in_shape,
+                                  const Shape& out_shape,
+                                  const AxisSet& broadcast_axes)
+                {
+                    size_t index[6];
+                    size_t* out_index = 0;
+                    for (size_t i = 0; i < 6; i++)
+                    {
+                        if (broadcast_axes.count(i) == 0)
+                        {
+                            out_index = &index[i];
+                            break;
+                        }
+                    }
+                    for (index[0] = 0; index[0] < out_shape[0]; ++index[0])
+                    {
+                        for (index[1] = 0; index[1] < out_shape[1]; ++index[1])
+                        {
+                            for (index[2] = 0; index[2] < out_shape[2]; ++index[2])
+                            {
+                                for (index[3] = 0; index[3] < out_shape[3]; ++index[3])
+                                {
+                                    for (index[4] = 0; index[4] < out_shape[4]; ++index[4])
+                                    {
+                                        for (index[5] = 0; index[5] < out_shape[5]; ++index[5])
+                                        {
+                                            out[index[0] * out_shape[1] * out_shape[2] *
+                                                    out_shape[3] * out_shape[4] * out_shape[5] +
+                                                index[1] * out_shape[2] * out_shape[3] *
+                                                    out_shape[4] * out_shape[5] +
+                                                index[2] * out_shape[3] * out_shape[4] *
+                                                    out_shape[5] +
+                                                index[3] * out_shape[4] * out_shape[5] +
+                                                index[4] * out_shape[5] + index[5]] =
+                                                in[*out_index];
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+
                 template <typename T>
                 void broadcast(const T* in,
                                T* out,
@@ -167,6 +252,16 @@ namespace ngraph
                         case 4:
                             broadcast_4d<T>(in, out, in_shape, out_shape, broadcast_axes);
                             break;
+                        case 5:
+                            broadcast_5d<T>(in, out, in_shape, out_shape, broadcast_axes);
+                            break;
+                        case 6:
+                            broadcast_6d<T>(in, out, in_shape, out_shape, broadcast_axes);
+                            break;
+                        default:
+                            runtime::reference::broadcast<T>(
+                                in, out, in_shape, out_shape, broadcast_axes);
+                            break;
                         }
                     }
                     else

src/ngraph/runtime/generic_cpu/kernel/reshape.hpp

@@ -244,10 +244,7 @@ namespace ngraph
                     case 4: reshape_in4<T>(in, out, in_shape, in_axis_order, out_shape); break;
                     case 5: reshape_in5<T>(in, out, in_shape, in_axis_order, out_shape); break;
                     case 6: reshape_in6<T>(in, out, in_shape, in_axis_order, out_shape); break;
-                    default:
-                        NGRAPH_INFO << "reference::reshape";
-                        reference::reshape(in, out, in_shape, in_axis_order, out_shape);
-                        break;
+                    default: reference::reshape(in, out, in_shape, in_axis_order, out_shape); break;
                     }
                 }
             }

src/ngraph/runtime/generic_cpu/kernel/result.hpp

-//*****************************************************************************
-// Copyright 2017-2019 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.
-//*****************************************************************************
-
-#pragma once
-
-#include <algorithm>
-#include <cmath>
-#include <numeric>
-#include <vector>
-#include "ngraph/shape.hpp"
-
-namespace ngraph
-{
-    namespace runtime
-    {
-        namespace gcpu
-        {
-            namespace kernel
-            {
-                template <typename T>
-                void result(const T* arg, T* out, size_t count)
-                {
-                    memcpy(out, arg, sizeof(T) * count);
-                }
-            }
-        }
-    }
-}

src/ngraph/runtime/generic_cpu/node_wrapper.hpp

@@ -51,7 +51,7 @@ class ngraph::runtime::gcpu::NodeWrapper
 public:
     NodeWrapper(const std::shared_ptr<const ngraph::Node>& node);
 
-    const Node& get_node() const { return *m_node; }
+    std::shared_ptr<const Node> get_node() const { return m_node; }
     ngraph::runtime::gcpu::OP_TYPEID get_typeid() const { return m_typeid; }
 private:
     std::shared_ptr<const ngraph::Node> m_node;