Merge remote-tracking branch 'origin/master' into r0.9

Author: Adam Procter

.ci/onnx/jenkins/prepare_environment.sh

@@ -68,7 +68,7 @@ function build_ngraph() {
     make install || return 1
     cd "${ngraph_directory}/ngraph/python"
     if [ ! -d ./pybind11 ]; then
-        git clone --recursive -b allow-nonconstructible-holders https://github.com/jagerman/pybind11.git
+        git clone --recursive https://github.com/pybind/pybind11.git
     fi
     export PYBIND_HEADERS_PATH="${ngraph_directory}/ngraph/python/pybind11"
     export NGRAPH_CPP_BUILD_PATH="${ngraph_directory}/ngraph_dist"

.ci/travis/ubuntu/Dockerfile

@@ -39,7 +39,7 @@ RUN make install
 
 # Prepare nGraph Python API
 WORKDIR /root/ngraph/python
-RUN git clone --recursive -b allow-nonconstructible-holders https://github.com/jagerman/pybind11.git
+RUN git clone --recursive https://github.com/pybind/pybind11.git
 ENV NGRAPH_CPP_BUILD_PATH /root/ngraph_dist
 ENV LD_LIBRARY_PATH /root/ngraph_dist/lib
 ENV PYBIND_HEADERS_PATH /root/ngraph/python/pybind11

src/ngraph/graph_util.cpp

@@ -558,3 +558,16 @@ bool ngraph::is_strided(const Strides& strides)
 {
     return std::any_of(strides.begin(), strides.end(), [](size_t stride) { return stride != 1; });
 }
+
+bool ngraph::is_valid_rank(const std::shared_ptr<Node>& node, std::vector<size_t> valid_ranks)
+{
+    auto node_rank = node->get_shape().size();
+    for (auto rank : valid_ranks)
+    {
+        if (rank == node_rank)
+        {
+            return true;
+        }
+    }
+    return false;
+}

src/ngraph/graph_util.hpp

@@ -315,4 +315,6 @@ namespace ngraph
     bool possibly_overwritten(Node* node);
 
     bool is_strided(const Strides& strides);
+
+    bool is_valid_rank(const std::shared_ptr<Node>& node, std::vector<size_t> valid_ranks);
 }

src/ngraph/node.hpp

@@ -103,10 +103,11 @@ namespace ngraph
         void validate_and_infer_elementwise_logical();
 
         Node(const std::string& node_type, const NodeVector& arguments, size_t output_size = 1);
-        virtual ~Node();
 
         virtual void generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector& deltas) {}
     public:
+        virtual ~Node();
+        void revalidate_and_infer_types() { validate_and_infer_types(); }
         // Called after transition
         void delayed_validate_and_infer_types();
 

src/ngraph/runtime/cpu/CMakeLists.txt

@@ -110,6 +110,7 @@ set(SRC
     pass/cpu_post_layout_optimizations.cpp
     pass/cpu_rnn_fusion.cpp
     pass/cpu_workspace_insertion.cpp
+    pass/cpu_reshape_sinking.cpp
 )
 
 if (NOT NGRAPH_DEX_ONLY)

src/ngraph/runtime/cpu/builder/softmax.cpp

@@ -19,6 +19,7 @@
 #include "ngraph/runtime/cpu/kernel/softmax.hpp"
 #include "ngraph/runtime/cpu/mkldnn_invoke.hpp"
 #include "ngraph/runtime/cpu/mkldnn_utils.hpp"
+#include "ngraph/runtime/reference/softmax.hpp"
 
 using namespace std;
 using namespace ngraph;
@@ -131,8 +132,35 @@ namespace ngraph
                         };
                         functors.emplace_back(functor);
                     }
+                    else if (arg_shape.size() == 4 && axes.size() == 3)
+                    {
+                        std::function<decltype(runtime::cpu::kernel::softmax_4d_3rd<float>)> kernel;
+
+                        SELECT_KERNEL(kernel,
+                                      args[0].get_element_type(),
+                                      runtime::cpu::kernel::softmax_4d_3rd);
+
+                        auto functor = [&, kernel, arg_shape, axes](CPURuntimeContext* ctx) {
+                            kernel(arg_tensor, out_tensor, arg_shape, axes);
+                        };
+                        functors.emplace_back(functor);
+                    }
+                    else if (softmax->get_element_type() == element::f32)
+                    {
+                        NGRAPH_WARN << "Falling back to refernce kernel for softmax " << arg_shape
+                                    << " over " << axes;
+                        auto functor = [&, arg_shape, axes](CPURuntimeContext* ctx) {
+                            runtime::reference::softmax<float>(static_cast<float*>(arg_tensor),
+                                                               static_cast<float*>(out_tensor),
+                                                               arg_shape,
+                                                               axes);
+                        };
+                        functors.emplace_back(functor);
+                    }
                     else
                     {
+                        NGRAPH_ERR << "Unsupported Softmax " << arg_shape << " over " << axes
+                                   << " in cpu buiilder";
                         throw ngraph_error("Unsupported Softmax");
                     }
                 }

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -126,6 +126,7 @@
 #include "ngraph/pass/manager.hpp"
 #include "ngraph/pass/memory_layout.hpp"
 #include "ngraph/pass/nop_elimination.hpp"
+#include "ngraph/pass/zero_dim_tensor_elimination.hpp"
 #include "ngraph/runtime/aligned_buffer.hpp"
 #include "ngraph/runtime/cpu/cpu_backend.hpp"
 #include "ngraph/runtime/cpu/cpu_builder.hpp"
@@ -1001,6 +1002,7 @@ void runtime::cpu::CPU_ExternalFunction::register_common_passes(ngraph::pass::Ma
 {
     pass_manager.register_pass<ngraph::pass::LikeReplacement>();
     pass_manager.register_pass<ngraph::pass::NopElimination>();
+    pass_manager.register_pass<ngraph::pass::ZeroDimTensorElimination>();
     // TODO (pruthvi): Enable all the disabeled RNN fusion graph pass after fixing
     // failing mxnet unit tests.
     // pass_manager.register_pass<runtime::cpu::pass::LSTMFusion>();
@@ -1013,7 +1015,7 @@ void runtime::cpu::CPU_ExternalFunction::register_common_passes(ngraph::pass::Ma
     pass_manager.register_pass<ngraph::pass::CommonSubexpressionElimination>();
     pass_manager.register_pass<ngraph::pass::CoreFusion>();
     pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
-    pass_manager.register_pass<runtime::cpu::pass::CPUHorizontalFusion>();
+    // pass_manager.register_pass<runtime::cpu::pass::CPUHorizontalFusion>();
     pass_manager.register_pass<runtime::cpu::pass::CPUCollapseDims>();
     NodeVector nv_cwi; // We dont need CPUWorkspaceInsertion to return list of indices
     pass_manager.register_pass<runtime::cpu::pass::CPUWorkspaceInsertion>(nv_cwi, false);

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

@@ -16,6 +16,9 @@
 
 #pragma once
 
+#include <cstddef>
+#include <cstring>
+
 namespace ngraph
 {
     namespace runtime

src/ngraph/runtime/cpu/kernel/softmax.hpp

@@ -21,6 +21,7 @@
 
 #include "ngraph/axis_set.hpp"
 #include "ngraph/runtime/cpu/kernel/eigen_thread_pool.hpp"
+#include "ngraph/shape.hpp"
 
 namespace ngraph
 {
@@ -147,6 +148,15 @@ namespace ngraph
                 {
                     softmax<ElementType, 3, 2>(input, output, input_shape, softmax_axes);
                 }
+
+                template <typename ElementType>
+                void softmax_4d_3rd(void* input,
+                                    void* output,
+                                    const Shape& input_shape,
+                                    const AxisSet& softmax_axes)
+                {
+                    softmax<ElementType, 4, 3>(input, output, input_shape, softmax_axes);
+                }
             }
         }
     }