Generalized dot (#291)

* WIP generalized dot

* Add some multi-axis 3D, 4D, and 5D tests

* Add test on some 'pretty big' tensors

* Reworked dot to have less flexible axis-pairing behavior

* Backprop for dot... and a fix for a dumb bug in CoordinateTransform

* Forgot to commit some stuff in merge

* Disable tests that currently don't work on CPU

* Fix temporarily disabled test that should pass on NGVM and INTERPRETER but wasn't due to new axis-selection convention for dot

* Remove obsolete ScalarTensorProduct kernel/instruction

* Review comment

* s/n_dot_axes/dot_axis_count/

* s/dot_axis_count/reduction_axes_count/

* Adapt CPU emitter dot fallback to new kernel

src/ngraph/common.cpp

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // ----------------------------------------------------------------------------
 
+#include <algorithm>
+
 #include "ngraph/common.hpp"
 
 using namespace ngraph;
@@ -37,31 +39,50 @@ Shape ngraph::project_shape(const Shape& shape, const AxisSet& deleted_axes)
     return project_coordinate(shape, deleted_axes);
 }
 
-// TODO: for the moment, just one axis at a time, please. Later could pass in an std::map from axis positions to axis lengths.
-// TODO: check validity, i.e. that the new axis is < coord_size+1.
-Coordinate
-    ngraph::inject_coordinate(const Coordinate& coord, size_t new_axis_pos, size_t new_axis_val)
+// TODO: check validity, i.e. that the new axis indices are all < coord_size+num_new_axes.
+Coordinate ngraph::inject_coordinate(const Coordinate& coord,
+                                     std::vector<std::pair<size_t, size_t>> new_axis_pos_val_pairs)
 {
     Coordinate result;
 
     size_t original_pos = 0;
 
-    for (size_t result_pos = 0; result_pos < coord.size() + 1; result_pos++)
+    for (size_t result_pos = 0; result_pos < coord.size() + new_axis_pos_val_pairs.size();
+         result_pos++)
     {
-        if (result_pos == new_axis_pos)
+        auto search_it = std::find_if(
+            new_axis_pos_val_pairs.begin(),
+            new_axis_pos_val_pairs.end(),
+            [result_pos](std::pair<size_t, size_t> p) { return p.first == result_pos; });
+
+        if (search_it == new_axis_pos_val_pairs.end())
         {
-            result.push_back(new_axis_val);
+            result.push_back(coord[original_pos++]);
         }
         else
         {
-            result.push_back(coord[original_pos++]);
+            result.push_back(search_it->second);
         }
     }
 
     return result;
 }
 
+Coordinate
+    ngraph::inject_coordinate(const Coordinate& coord, size_t new_axis_pos, size_t new_axis_val)
+{
+    return inject_coordinate(coord,
+                             std::vector<std::pair<size_t, size_t>>{
+                                 std::pair<size_t, size_t>(new_axis_pos, new_axis_val)});
+}
+
 Shape ngraph::inject_shape(const Shape& shape, size_t new_axis_pos, size_t new_axis_length)
 {
     return inject_coordinate(shape, new_axis_pos, new_axis_length);
 }
+
+Shape inject_shape(const Shape& shape,
+                   std::vector<std::pair<size_t, size_t>> new_axis_pos_length_pairs)
+{
+    return inject_coordinate(shape, new_axis_pos_length_pairs);
+}

src/ngraph/common.hpp

@@ -16,6 +16,7 @@
 
 #include <memory>
 #include <set>
+#include <utility>
 #include <vector>
 
 // Names for types that aren't worth giving their own classes
@@ -56,5 +57,9 @@ namespace ngraph
     Shape project_shape(const Shape& shape, const AxisSet& deleted_axes);
 
     Coordinate inject_coordinate(const Coordinate& coord, size_t new_axis_pos, size_t new_axis_val);
+    Coordinate inject_coordinate(const Coordinate& coord,
+                                 std::vector<std::pair<size_t, size_t>> new_axis_pos_val_pairs);
     Shape inject_shape(const Shape& shape, size_t new_axis_pos, size_t new_axis_length);
+    Shape inject_shape(const Shape& shape,
+                       std::vector<std::pair<size_t, size_t>> new_axis_pos_length_pairs);
 }

src/ngraph/coordinate_transform.cpp

@@ -209,8 +209,8 @@ Coordinate CoordinateTransform::to_source_coordinate(const Coordinate& c) const
 
     for (size_t axis = 0; axis < m_n_axes; axis++)
     {
-        result[axis] = c[m_source_axis_order[axis]] * m_source_strides[m_source_axis_order[axis]] +
-                       m_source_start_corner[m_source_axis_order[axis]];
+        result[m_source_axis_order[axis]] =
+            c[axis] * m_source_strides[axis] + m_source_start_corner[axis];
     }
 
     return result;

src/ngraph/runtime/cpu/cpu_emitter.cpp

@@ -24,6 +24,7 @@
 #include "ngraph/ops/broadcast.hpp"
 #include "ngraph/ops/concatenate.hpp"
 #include "ngraph/ops/constant.hpp"
+#include "ngraph/ops/dot.hpp"
 #include "ngraph/ops/function_call.hpp"
 #include "ngraph/ops/get_tuple_element.hpp"
 #include "ngraph/ops/one_hot.hpp"
@@ -137,29 +138,7 @@ void runtime::cpu::CPU_Emitter::EmitDot(const ngraph::Node* n,
     }
     else
     {
-        size_t arg0_dot_axis;
-        size_t arg1_dot_axis;
-        if (arg0_shape.size() == 1 && arg1_shape.size() == 1)
-        {
-            arg0_dot_axis = 0;
-            arg1_dot_axis = 0;
-        }
-
-        // If arg0 is a matrix and arg1 is a vector, dot on axes 1 and 0 respectively.
-        else if (arg0_shape.size() == 2 && arg1_shape.size() == 1)
-        {
-            arg0_dot_axis = 1;
-            arg1_dot_axis = 0;
-        }
-
-        // If arg0 is rank n and arg1 is rank m, dot on axes n-1 and m-2, respectively.
-        //
-        // Note that this happens to handle the vector-matrix and matrix-matrix cases.
-        else
-        {
-            arg0_dot_axis = arg0_shape.size() - 1;
-            arg1_dot_axis = arg1_shape.size() - 2;
-        }
+        const ngraph::op::Dot* dot = static_cast<const ngraph::op::Dot*>(n);
 
         m_out << "kernel::dot(" << args[0].get_name() << ",\n";
         m_out << "            " << args[1].get_name() << ",\n";
@@ -167,8 +146,7 @@ void runtime::cpu::CPU_Emitter::EmitDot(const ngraph::Node* n,
         m_out << "            {" << join(args[0].get_shape()) << "},\n";
         m_out << "            {" << join(args[1].get_shape()) << "},\n";
         m_out << "            {" << join(out[0].get_shape()) << "},\n";
-        m_out << "            " << arg0_dot_axis << ",\n";
-        m_out << "            " << arg1_dot_axis << ");\n";
+        m_out << "            " << dot->get_reduction_axes_count() << ");\n";
     }
 }
 

src/ngraph/runtime/interpreter/int_call_frame.hpp

@@ -23,6 +23,7 @@
 #include "ngraph/ops/broadcast.hpp"
 #include "ngraph/ops/concatenate.hpp"
 #include "ngraph/ops/constant.hpp"
+#include "ngraph/ops/dot.hpp"
 #include "ngraph/ops/get_tuple_element.hpp"
 #include "ngraph/ops/one_hot.hpp"
 #include "ngraph/ops/reduce.hpp"
@@ -66,7 +67,6 @@
 #include "ngraph/runtime/kernel/reduce.hpp"
 #include "ngraph/runtime/kernel/replace_slice.hpp"
 #include "ngraph/runtime/kernel/reshape.hpp"
-#include "ngraph/runtime/kernel/scalar_tensor_product.hpp"
 #include "ngraph/runtime/kernel/select.hpp"
 #include "ngraph/runtime/kernel/sign.hpp"
 #include "ngraph/runtime/kernel/sin.hpp"
@@ -290,54 +290,15 @@ private:
         }
         else if (node_op == "Dot")
         {
-            if (args[0]->get_shape().size() == 0)
-            {
-                kernel::scalar_tensor_product(reinterpret_cast<T*>(args[0]->get_data_ptr()),
-                                              reinterpret_cast<T*>(args[1]->get_data_ptr()),
-                                              reinterpret_cast<T*>(out[0]->get_data_ptr()),
-                                              out[0]->get_element_count());
-            }
-            else if (args[1]->get_shape().size() == 0)
-            {
-                kernel::scalar_tensor_product(reinterpret_cast<T*>(args[1]->get_data_ptr()),
-                                              reinterpret_cast<T*>(args[0]->get_data_ptr()),
-                                              reinterpret_cast<T*>(out[0]->get_data_ptr()),
-                                              out[0]->get_element_count());
-            }
-            else
-            {
-                size_t arg0_dot_axis;
-                size_t arg1_dot_axis;
-                if (args[0]->get_shape().size() == 1 && args[1]->get_shape().size() == 1)
-                {
-                    arg0_dot_axis = 0;
-                    arg1_dot_axis = 0;
-                }
-
-                // If arg0 is a matrix and arg1 is a vector, dot on axes 1 and 0 respectively.
-                else if (args[0]->get_shape().size() == 2 && args[1]->get_shape().size() == 1)
-                {
-                    arg0_dot_axis = 1;
-                    arg1_dot_axis = 0;
-                }
+            ngraph::op::Dot* dot = dynamic_cast<ngraph::op::Dot*>(&node);
 
-                // If arg0 is rank n and arg1 is rank m, dot on axes n-1 and m-2, respectively.
-                //
-                // Note that this happens to handle the vector-matrix and matrix-matrix cases.
-                else
-                {
-                    arg0_dot_axis = args[0]->get_shape().size() - 1;
-                    arg1_dot_axis = args[1]->get_shape().size() - 2;
-                }
             kernel::dot(reinterpret_cast<T*>(args[0]->get_data_ptr()),
                         reinterpret_cast<T*>(args[1]->get_data_ptr()),
                         reinterpret_cast<T*>(out[0]->get_data_ptr()),
                         args[0]->get_shape(),
                         args[1]->get_shape(),
                         out[0]->get_shape(),
-                            arg0_dot_axis,
-                            arg1_dot_axis);
-            }
+                        dot->get_reduction_axes_count());
         }
 
         else if (node_op == "Equal")

src/ngraph/runtime/kernel/dot.hpp

@@ -15,6 +15,7 @@
 #pragma once
 
 #include <cmath>
+#include <utility>
 
 #include "ngraph/common.hpp"
 #include "ngraph/coordinate_transform.hpp"
@@ -32,49 +33,86 @@ namespace ngraph
                      const Shape& arg0_shape,
                      const Shape& arg1_shape,
                      const Shape& out_shape,
-                     size_t arg0_dot_axis,
-                     size_t arg1_dot_axis)
+                     size_t reduction_axes_count)
             {
-                CoordinateTransform output_transform(out_shape);
-
-                for (Coordinate out_coord : output_transform)
-                {
-                    out[output_transform.index(out_coord)] = 0;
-                }
+                // Get the sizes of the dot axes. It's easiest to pull them from arg1 because they're
+                // right up front.
+                Shape dot_axis_sizes(reduction_axes_count);
+                std::copy(arg1_shape.begin(),
+                          arg1_shape.begin() + reduction_axes_count,
+                          dot_axis_sizes.begin());
 
                 CoordinateTransform arg0_transform(arg0_shape);
                 CoordinateTransform arg1_transform(arg1_shape);
+                CoordinateTransform output_transform(out_shape);
 
-                CoordinateTransform arg0_projected_transform(
-                    project_shape(arg0_shape, AxisSet{arg0_dot_axis}));
-                CoordinateTransform arg1_projected_transform(
-                    project_shape(arg1_shape, AxisSet{arg1_dot_axis}));
+                // Create coordinate transforms for arg0 and arg1 that throw away the dotted axes.
+                size_t arg0_projected_rank = arg0_shape.size() - reduction_axes_count;
+                size_t arg1_projected_rank = arg1_shape.size() - reduction_axes_count;
+
+                Shape arg0_projected_shape(arg0_projected_rank);
+                std::copy(arg0_shape.begin(),
+                          arg0_shape.begin() + arg0_projected_rank,
+                          arg0_projected_shape.begin());
+
+                Shape arg1_projected_shape(arg1_projected_rank);
+                std::copy(arg1_shape.begin() + reduction_axes_count,
+                          arg1_shape.end(),
+                          arg1_projected_shape.begin());
+
+                CoordinateTransform arg0_projected_transform(arg0_projected_shape);
+                CoordinateTransform arg1_projected_transform(arg1_projected_shape);
+
+                // Create a coordinate transform that allows us to iterate over all possible values
+                // for the dotted axes.
+                CoordinateTransform dot_axes_transform(dot_axis_sizes);
 
                 for (Coordinate arg0_projected_coord : arg0_projected_transform)
                 {
                     for (Coordinate arg1_projected_coord : arg1_projected_transform)
                     {
-                        for (size_t i = 0; i < arg0_shape[arg0_dot_axis]; i++)
-                        {
-                            Coordinate arg0_coord =
-                                inject_coordinate(arg0_projected_coord, arg0_dot_axis, i);
-                            Coordinate arg1_coord =
-                                inject_coordinate(arg1_projected_coord, arg1_dot_axis, i);
-
+                        // The output coordinate is just the concatenation of the projected coordinates.
                         Coordinate out_coord(arg0_projected_coord.size() +
                                              arg1_projected_coord.size());
 
-                            std::copy(arg0_projected_coord.begin(),
+                        auto out_coord_it = std::copy(arg0_projected_coord.begin(),
                                                       arg0_projected_coord.end(),
                                                       out_coord.begin());
-                            std::copy(arg1_projected_coord.begin(),
-                                      arg1_projected_coord.end(),
-                                      out_coord.begin() + arg0_projected_coord.size());
+                        std::copy(
+                            arg1_projected_coord.begin(), arg1_projected_coord.end(), out_coord_it);
+
+                        // Zero out to start the sum.
+                        T sum = 0;
+
+                        size_t out_index = output_transform.index(out_coord);
+
+                        // Walk along the dotted axes.
+                        for (Coordinate dot_axis_positions : dot_axes_transform)
+                        {
+                            // In order to find the points to multiply together, we need to inject our current
+                            // positions along the dotted axes back into the projected arg0 and arg1 coordinates.
+                            Coordinate arg0_coord(arg0_shape.size());
+                            Coordinate arg1_coord(arg1_shape.size());
 
-                            out[output_transform.index(out_coord)] +=
-                                arg0[arg0_transform.index(arg0_coord)] *
+                            auto arg0_it = std::copy(arg0_projected_coord.begin(),
+                                                     arg0_projected_coord.end(),
+                                                     arg0_coord.begin());
+                            std::copy(
+                                dot_axis_positions.begin(), dot_axis_positions.end(), arg0_it);
+
+                            auto arg1_it = std::copy(dot_axis_positions.begin(),
+                                                     dot_axis_positions.end(),
+                                                     arg1_coord.begin());
+                            std::copy(
+                                arg1_projected_coord.begin(), arg1_projected_coord.end(), arg1_it);
+
+                            // Multiply and add to the sum.
+                            sum += arg0[arg0_transform.index(arg0_coord)] *
                                    arg1[arg1_transform.index(arg1_coord)];
                         }
+
+                        // Write the sum back.
+                        out[out_index] = sum;
                     }
                 }
             }

src/ngraph/runtime/kernel/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 <cmath>
-
-#include "ngraph/common.hpp"
-#include "ngraph/coordinate_transform.hpp"
-
-namespace ngraph
-{
-    namespace runtime
-    {
-        namespace kernel
-        {
-            template <typename T>
-            void scalar_tensor_product(T* arg0, // the scalar (TODO: just pass as T?)
-                                       T* arg1, // the tensor
-                                       T* out,
-                                       size_t count)
-            {
-                for (size_t i = 0; i < count; i++)
-                {
-                    out[i] = (*arg0) * arg1[i];
-                }
-            }
-        }
-    }
-}

src/ngraph/runtime/ngvm/external_function.cpp

@@ -109,7 +109,6 @@
 #include "ngraph/runtime/ngvm/instruction/replace_slice.hpp"
 #include "ngraph/runtime/ngvm/instruction/reshape.hpp"
 #include "ngraph/runtime/ngvm/instruction/return.hpp"
-#include "ngraph/runtime/ngvm/instruction/scalar_tensor_product.hpp"
 #include "ngraph/runtime/ngvm/instruction/select.hpp"
 #include "ngraph/runtime/ngvm/instruction/sign.hpp"
 #include "ngraph/runtime/ngvm/instruction/sin.hpp"
@@ -352,8 +351,6 @@ std::vector<typename ET::type>
     }
 #define PUSH_POLYMORPHIC_INSTRUCTION(et, err_msg, instr, ...)                                      \
     DO_ON_ELEMENT_TYPE(et, err_msg, PUSH_INSTRUCTION, instr, __VA_ARGS__)
-#define PUSH_NUMERIC_POLYMORPHIC_INSTRUCTION(et, err_msg, instr, ...)                              \
-    DO_ON_NUMERIC_TYPE(et, err_msg, PUSH_INSTRUCTION, instr, __VA_ARGS__)
 
 // Turn off complaint suppression (see above)
 #pragma clang diagnostic pop
@@ -550,6 +547,8 @@ ExternalFunction::OpMap& ExternalFunction::get_op_map()
 
         REGISTER_TO_OP_MAP(op::Dot)
         {
+            auto dot = static_cast<const op::Dot*>(n);
+
             auto& arg_nodes = n->get_arguments();
 
             assert(arg_nodes.size() == 2);
@@ -566,70 +565,15 @@ ExternalFunction::OpMap& ExternalFunction::get_op_map()
             auto arg1_shape = arg1_tensor_type->get_shape();
             auto& arg0_element_type = arg0_tensor_type->get_element_type();
 
+            auto reduction_axes_count = dot->get_reduction_axes_count();
+
             auto result_tensor_type =
                 dynamic_pointer_cast<const TensorViewType>(n->get_value_type());
             assert(nullptr != result_tensor_type);
 
             auto result_shape = result_tensor_type->get_shape();
 
-            // If arg0 or arg1 is a scalar, emit a scalar-tensor product.
-            if (arg0_shape.size() == 0)
-            {
-                PUSH_NUMERIC_POLYMORPHIC_INSTRUCTION(arg0_element_type,
-                                                     "Dot has unhandled element type",
-                                                     instruction::ScalarTensorProductInstruction,
-                                                     in[0],
-                                                     in[1],
-                                                     out[0]);
-            }
-            else if (arg1_shape.size() == 0)
-            {
-                PUSH_NUMERIC_POLYMORPHIC_INSTRUCTION(arg0_element_type,
-                                                     "Dot has unhandled element type",
-                                                     instruction::ScalarTensorProductInstruction,
-                                                     in[1],
-                                                     in[0],
-                                                     out[0]);
-            }
-
-            // If arg0 and arg1 are both vectors, dot both on axis 0.
-            else if (arg0_shape.size() == 1 && arg1_shape.size() == 1)
-            {
-                PUSH_NUMERIC_POLYMORPHIC_INSTRUCTION(arg0_element_type,
-                                                     "Dot has unhandled element type",
-                                                     instruction::DotInstruction,
-                                                     in[0],
-                                                     in[1],
-                                                     out[0],
-                                                     arg0_shape,
-                                                     arg1_shape,
-                                                     result_shape,
-                                                     0,
-                                                     0);
-            }
-
-            // If arg0 is a matrix and arg1 is a vector, dot on axes 1 and 0 respectively.
-            else if (arg0_shape.size() == 2 && arg1_shape.size() == 1)
-            {
-                PUSH_NUMERIC_POLYMORPHIC_INSTRUCTION(arg0_element_type,
-                                                     "Dot has unhandled element type",
-                                                     instruction::DotInstruction,
-                                                     in[0],
-                                                     in[1],
-                                                     out[0],
-                                                     arg0_shape,
-                                                     arg1_shape,
-                                                     result_shape,
-                                                     1,
-                                                     0);
-            }
-
-            // If arg0 is rank n and arg1 is rank m, dot on axes n-1 and m-2, respectively.
-            //
-            // Note that this happens to handle the vector-matrix and matrix-matrix cases.
-            else
-            {
-                PUSH_NUMERIC_POLYMORPHIC_INSTRUCTION(arg0_element_type,
+            PUSH_POLYMORPHIC_INSTRUCTION(arg0_element_type,
                                          "Dot has unhandled element type",
                                          instruction::DotInstruction,
                                          in[0],
@@ -638,9 +582,7 @@ ExternalFunction::OpMap& ExternalFunction::get_op_map()
                                          arg0_shape,
                                          arg1_shape,
                                          result_shape,
-                                                     arg0_shape.size() - 1,
-                                                     arg1_shape.size() - 2);
-            }
+                                         reduction_axes_count);
         };
 
         // Parameter is a "runtime no-op" because the output tensor has already been filled.

src/ngraph/runtime/ngvm/instruction/dot.hpp

@@ -38,16 +38,14 @@ namespace ngraph
                                    const Shape& arg0_shape,
                                    const Shape& arg1_shape,
                                    const Shape& out_shape,
-                                   size_t arg0_dot_axis,
-                                   size_t arg1_dot_axis)
+                                   size_t reduction_axes_count)
                         : m_arg0(arg0)
                         , m_arg1(arg1)
                         , m_out(out)
                         , m_arg0_shape(arg0_shape)
                         , m_arg1_shape(arg1_shape)
                         , m_out_shape(out_shape)
-                        , m_arg0_dot_axis(arg0_dot_axis)
-                        , m_arg1_dot_axis(arg1_dot_axis)
+                        , m_reduction_axes_count(reduction_axes_count)
                     {
                     }
 
@@ -63,8 +61,7 @@ namespace ngraph
                                                        m_arg0_shape,
                                                        m_arg1_shape,
                                                        m_out_shape,
-                                                       m_arg0_dot_axis,
-                                                       m_arg1_dot_axis);
+                                                       m_reduction_axes_count);
                     }
 
                 protected:
@@ -74,8 +71,7 @@ namespace ngraph
                     Shape m_arg0_shape;
                     Shape m_arg1_shape;
                     Shape m_out_shape;
-                    size_t m_arg0_dot_axis;
-                    size_t m_arg1_dot_axis;
+                    size_t m_reduction_axes_count;
                 };
             }
         }

src/ngraph/runtime/ngvm/instruction/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/kernel/scalar_tensor_product.hpp"
-#include "ngraph/runtime/ngvm/call_frame.hpp"
-#include "ngraph/runtime/ngvm/instruction.hpp"
-#include "ngraph/runtime/ngvm/utils.hpp"
-#include "ngraph/runtime/tensor_view.hpp"
-
-namespace ngraph
-{
-    namespace runtime
-    {
-        namespace ngvm
-        {
-            namespace instruction
-            {
-                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
-                    {
-                        typename ET::type* arg0 = get_tensor_data_ptr<ET>(call_frame, m_arg0);
-                        typename ET::type* arg1 = get_tensor_data_ptr<ET>(call_frame, m_arg1);
-                        typename ET::type* out = get_tensor_data_ptr<ET>(call_frame, m_out);
-
-                        size_t count = get_tensor_element_count(call_frame, m_arg1);
-
-                        kernel::scalar_tensor_product<typename ET::type>(arg0, arg1, out, count);
-                    }
-
-                protected:
-                    TensorViewInfo m_arg0;
-                    TensorViewInfo m_arg1;
-                    TensorViewInfo m_out;
-                };
-            }
-        }
-    }
-}

test/autodiff.in.cpp

@@ -521,6 +521,28 @@ TEST(${BACKEND_NAME}, backwards_dot_tensor2_tensor2)
         autodiff_numeric_compare<float>(manager, backend, make_graph, {x0, x1}, .01f, .01f));
 }
 
+TEST(${BACKEND_NAME}, backwards_dot_tensor3_tensor3)
+{
+    auto manager = runtime::Manager::get("NGVM");
+    auto backend = manager->allocate_backend();
+
+    test::Uniform<float> rng(-1.0f, 1.0f);
+    auto shape0 = Shape{2, 4, 3};
+    auto shape1 = Shape{4, 3, 3};
+    auto x0 = rng.initialize(backend->make_primary_tensor_view<float>(shape0));
+    auto x1 = rng.initialize(backend->make_primary_tensor_view<float>(shape1));
+
+    auto make_graph = [shape0, shape1]() {
+        auto X0 = make_shared<op::Parameter>(element::Float32::element_type(), shape0);
+        auto X1 = make_shared<op::Parameter>(element::Float32::element_type(), shape1);
+        return make_shared<Function>(make_shared<op::Dot>(X0, X1, 2),
+                                     nullptr,
+                                     std::vector<std::shared_ptr<op::Parameter>>{X0, X1});
+    };
+    EXPECT_TRUE(
+        autodiff_numeric_compare<float>(manager, backend, make_graph, {x0, x1}, .01f, .01f));
+}
+
 TEST(${BACKEND_NAME}, backwards_exp)
 {
     auto manager = runtime::Manager::get("${BACKEND_NAME}");

test/type_prop.cpp

@@ -32,15 +32,6 @@ TEST(type_prop, broadcast_deduce)
     ASSERT_EQ(*bc_vt, TensorViewType(element::Float32::element_type(), Shape{2, 3, 4}));
 }
 
-TEST(type_prop, broadcast_deduce_correct)
-{
-    // Check deduced type against correctly specified type
-    auto param = make_shared<op::Parameter>(element::Float32::element_type(), Shape{2, 4});
-    auto bc = make_shared<op::Broadcast>(param, Shape{2, 3, 4}, AxisSet{1});
-    auto bc_vt = bc->get_value_type();
-    ASSERT_EQ(*bc_vt, TensorViewType(element::Float32::element_type(), Shape{2, 3, 4}));
-}
-
 TEST(type_prop, broadcast_deduce_incorrect)
 {
     // Check deduced type against incorrectly specified type
@@ -228,15 +219,6 @@ TEST(type_prop, convert_deduce)
     ASSERT_EQ(*c_vt, TensorViewType(element::Int32::element_type(), Shape{2, 3, 4}));
 }
 
-TEST(type_prop, convert_deduce_correct)
-{
-    // Check deduced type against incorrectly specified type
-    auto param = make_shared<op::Parameter>(element::Float32::element_type(), Shape{2, 3, 4});
-    auto c = make_shared<op::Convert>(param, element::Int32::element_type());
-    auto c_vt = c->get_value_type();
-    ASSERT_EQ(*c_vt, TensorViewType(element::Int32::element_type(), Shape{2, 3, 4}));
-}
-
 TEST(type_prop, convert_deduce_incorrect)
 {
     // Check deduced type against incorrectly specified type
@@ -322,17 +304,7 @@ TEST(type_prop, dot_deduce_different_rank)
 {
     // Deduce type for different-rank tensor arguments
     auto param1 = make_shared<op::Parameter>(element::Float32::element_type(), Shape{2, 8, 4, 2});
-    auto param2 = make_shared<op::Parameter>(element::Float32::element_type(), Shape{1, 2, 3});
-    auto bc = make_shared<op::Dot>(param1, param2);
-    auto bc_vt = bc->get_value_type();
-    ASSERT_EQ(*bc_vt, TensorViewType(element::Float32::element_type(), Shape{2, 8, 4, 1, 3}));
-}
-
-TEST(type_prop, dot_deduce_different_rank_correct)
-{
-    // Deduced type matches explicitly set type
-    auto param1 = make_shared<op::Parameter>(element::Float32::element_type(), Shape{2, 8, 4, 2});
-    auto param2 = make_shared<op::Parameter>(element::Float32::element_type(), Shape{1, 2, 3});
+    auto param2 = make_shared<op::Parameter>(element::Float32::element_type(), Shape{2, 1, 3});
     auto bc = make_shared<op::Dot>(param1, param2);
     auto bc_vt = bc->get_value_type();
     ASSERT_EQ(*bc_vt, TensorViewType(element::Float32::element_type(), Shape{2, 8, 4, 1, 3}));
@@ -372,7 +344,7 @@ TEST(type_prop, dot_deduce_reduction_axes_size_mismatch)
     }
     catch (const ngraph_error& error)
     {
-        EXPECT_EQ(error.what(), std::string("Dot reduction axes not compatible"));
+        EXPECT_EQ(error.what(), std::string("Dot axes do not have same length"));
     }
     catch (...)
     {
@@ -571,19 +543,6 @@ TEST(type_prop, select_deduce)
     ASSERT_EQ(*bc_vt, TensorViewType(element::Float32::element_type(), Shape{2, 4}));
 }
 
-TEST(type_prop, select_deduce_correct)
-{
-    auto tv0_2_4_param_0 = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Bool::element_type(), Shape{2, 4}));
-    auto tv0_2_4_param_1 = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Float32::element_type(), Shape{2, 4}));
-    auto tv0_2_4_param_2 = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Float32::element_type(), Shape{2, 4}));
-    auto bc = make_shared<op::Select>(tv0_2_4_param_0, tv0_2_4_param_1, tv0_2_4_param_2);
-    auto bc_vt = bc->get_value_type();
-    ASSERT_EQ(*bc_vt, TensorViewType(element::Float32::element_type(), Shape{2, 4}));
-}
-
 TEST(type_prop, select_shape_mismatch_a)
 {
     auto tv0_2_4_param_0 = make_shared<op::Parameter>(
@@ -735,24 +694,6 @@ TEST(type_prop, reduce_deduce)
               TensorViewType(element::Float32::element_type(), Shape{2, 4}));
 }
 
-TEST(type_prop, reduce_deduce_correct)
-{
-    auto param_0 = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Float32::element_type(), Shape{2, 4}));
-    auto param_1 = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Float32::element_type(), Shape{}));
-
-    auto f_param_0 = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Float32::element_type(), Shape{}));
-    auto f_param_1 = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Float32::element_type(), Shape{}));
-    auto rt = make_shared<TensorViewType>(element::Float32::element_type(), Shape{});
-    auto f = make_shared<Function>(f_param_0 + f_param_1, rt, op::Parameters{f_param_0, f_param_1});
-
-    auto r0 = make_shared<op::Reduce>(param_0, param_1, f, AxisSet{0});
-    ASSERT_EQ(*(r0->get_value_type()), TensorViewType(element::Float32::element_type(), Shape{4}));
-}
-
 TEST(type_prop, reduce_nonscalar)
 {
     auto param_0 = make_shared<op::Parameter>(
@@ -1073,14 +1014,6 @@ TEST(type_prop, reshape_deduce_t2v_120)
     ASSERT_EQ(*(r->get_value_type()), TensorViewType(element::Float32::element_type(), Shape{60}));
 }
 
-TEST(type_prop, reshape_deduce_correct_t2v_120)
-{
-    auto param = make_shared<op::Parameter>(
-        make_shared<TensorViewType>(element::Float32::element_type(), Shape{3, 4, 5}));
-    auto r = make_shared<op::Reshape>(param, AxisVector{1, 2, 0}, Shape{60});
-    ASSERT_EQ(*(r->get_value_type()), TensorViewType(element::Float32::element_type(), Shape{60}));
-}
-
 TEST(type_prop, reshape_deduce_not_enough_axes)
 {
     auto param = make_shared<op::Parameter>(