Merge pull request #63 from NervanaSystems/cyphers/mnist

Two shape propagates/checks, bulk of ops.

src/ngraph/element_type.hpp

@@ -29,7 +29,7 @@ namespace ngraph
         {
         public:
             Type(size_t bitwidth, bool is_float, bool is_signed, const std::string& cname);
-        
+
             const std::string& c_type_string() const;
             size_t             size() const;
             size_t             hash() const
@@ -37,23 +37,24 @@ namespace ngraph
                 std::hash<std::string> h;
                 return h(m_cname);
             }
-        
+
             bool operator==(const Type& other) const;
-        
+            bool operator!=(const Type& other) const { return !(*this == other); }
+
         private:
-            static std::map<std::string, Type>        m_element_list;
-            size_t                                    m_bitwidth;
-            bool                                      m_is_float;
-            bool                                      m_is_signed;
-            const std::string                         m_cname;
+            static std::map<std::string, Type> m_element_list;
+            size_t                             m_bitwidth;
+            bool                               m_is_float;
+            bool                               m_is_signed;
+            const std::string                  m_cname;
         };
 
-        const Type float32_t= Type(32, true, true, "float");
-        const Type int8_t   = Type(8, false, true, "int8_t");
-        const Type int32_t  = Type(32, false, true, "int32_t");
-        const Type int64_t  = Type(64, false, true, "int64_t");
-        const Type uint8_t  = Type(8, false, false, "int8_t");
-        const Type uint32_t = Type(32, false, false, "int32_t");
-        const Type uint64_t = Type(64, false, false, "int64_t");
+        const Type float32_t = Type(32, true, true, "float");
+        const Type int8_t    = Type(8, false, true, "int8_t");
+        const Type int32_t   = Type(32, false, true, "int32_t");
+        const Type int64_t   = Type(64, false, true, "int64_t");
+        const Type uint8_t   = Type(8, false, false, "int8_t");
+        const Type uint32_t  = Type(32, false, false, "int32_t");
+        const Type uint64_t  = Type(64, false, false, "int64_t");
     }
 }

src/ngraph/except.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 <stdexcept>
+
+namespace ngraph
+{
+    /// Base error for ngraph runtime errors.
+    struct ngraph_error : std::runtime_error
+    {
+        explicit ngraph_error(const std::string& what_arg)
+            : std::runtime_error(what_arg)
+        {
+        }
+
+        explicit ngraph_error(const char* what_arg)
+            : std::runtime_error(what_arg)
+        {
+        }
+    };
+}

src/ngraph/function.hpp

@@ -35,6 +35,8 @@ namespace ngraph
 
         std::string description() const override { return "Parameter"; }
 
+        virtual void propagate_types() override;
+
     protected:
         Function& m_function;
         size_t    m_index;
@@ -59,7 +61,7 @@ namespace ngraph
     /**
      ** A user-defined function.
      **/
-    class Function : public Op
+    class Function
     {
     public:
         Function(size_t n_parameters);
@@ -68,7 +70,7 @@ namespace ngraph
 
         Parameter::ptr parameter(size_t i) { return m_parameters[i]; }
 
-        std::string name() const override { return m_name; }
+        std::string name() const { return m_name; }
 
     protected:
         std::vector<Parameter::ptr> m_parameters;

src/ngraph/ngraph.hpp

@@ -19,6 +19,7 @@
 #pragma once
 
 #include "ngraph/element_type.hpp"
+#include "ngraph/except.hpp"
 #include "ngraph/function.hpp"
 #include "ngraph/node.hpp"
 #include "ngraph/op.hpp"

src/ngraph/node.hpp

@@ -18,6 +18,8 @@
 #include <string>
 #include <vector>
 
+#include <iostream>
+
 #include "ngraph/type.hpp"
 
 namespace ngraph
@@ -53,17 +55,22 @@ namespace ngraph
         virtual std::string description() const = 0;
 
         /// Propagate types and check arguments for consistency
-        // virtual void propagate_types() = 0;
+        virtual void propagate_types() = 0;
 
-        const std::vector<Node::ptr> arguments() const { return m_arguments; }
-        std::vector<Node::ptr>       arguments() { return m_arguments; }
+        const std::vector<Node::ptr>& arguments() const { return m_arguments; }
 
-        const std::multiset<Node*> users() const { return m_users; }
-        std::multiset<Node*>       users() { return m_users; }
+        const std::multiset<Node*>& users() const { return m_users; }
 
         std::string name() const { return m_name; }
         void        name(const std::string& name) { m_name = name; }
 
+        /**
+         ** Return true if this has the same implementing class as node. This
+         ** will be used by the pattern matcher when comparing a pattern
+         ** graph against the graph.
+         **/
+         bool is_same_op_type(const Node::ptr& node) const { return typeid(*this) == typeid(*node.get()); }
+
     protected:
         std::vector<Node::ptr> m_arguments;
         std::multiset<Node*>   m_users;

src/ngraph/shape.hpp

@@ -32,11 +32,19 @@ namespace ngraph
         {
         }
 
+        Shape(const std::vector<size_t>& sizes)
+            : m_sizes(sizes)
+        {
+        }
+
         /**
          ** Conversion to a vector of sizes.
          **/
         operator const std::vector<size_t>&() const { return m_sizes; }
 
+        bool operator==(const Shape& shape) const { return m_sizes == shape.m_sizes; }
+        bool operator!=(const Shape& shape) const { return m_sizes != shape.m_sizes; }
+
     protected:
         std::vector<size_t> m_sizes;
     };

src/ngraph/type.hpp

@@ -22,6 +22,9 @@
 
 namespace ngraph
 {
+    class TensorViewType;
+    class TupleType;
+
     /**
      ** ValueType is
      **   TensorViewType
@@ -34,6 +37,8 @@ namespace ngraph
          ** Preferred handle
          **/
         using ptr = std::shared_ptr<ValueType>;
+
+        virtual ~ValueType() {}
     };
 
     /**
@@ -57,6 +62,9 @@ namespace ngraph
         {
         }
 
+        const element::Type& element_type() const { return m_element_type; }
+        const Shape&          shape() const { return m_shape; }
+
     protected:
         const element::Type& m_element_type;
         Shape                m_shape;

src/ops/function.cpp

@@ -24,6 +24,14 @@ Parameter::Parameter(Function& function, size_t index)
 {
 }
 
+void Parameter::propagate_types()
+{
+    if (m_type == nullptr)
+    {
+        throw ngraph_error{"Unitialized parameter"};
+    }
+}
+
 Function::Function(size_t n_parameters)
     : m_parameters(n_parameters)
     , m_name("Function")

src/ops/op.cpp

@@ -12,12 +12,22 @@
 // See the License for the specific language governing permissions and
 // ----------------------------------------------------------------------------
 
+#include <algorithm>
+
 #include "ngraph/ngraph.hpp"
 
 using namespace ngraph;
 using namespace std;
 
-std::shared_ptr<BuiltinOp> BroadcastCall::s_op = make_shared<BuiltinOp>("broadcast");
+Node::ptr ngraph::op::abs(const Node::ptr& arg)
+{
+    return make_shared<AbsOp>(arg);
+}
+
+Node::ptr ngraph::op::add(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<AddOp>(arg0, arg1);
+}
 
 /**
  ** /param arg The tensor view to be broadcast.
@@ -25,17 +35,163 @@ std::shared_ptr<BuiltinOp> BroadcastCall::s_op = make_shared<BuiltinOp>("broadca
  ** /param broadcast_axes The axis positions (0-based) in the result that are being broadcast.
  **  the remaining axes in shape must be the same as the shape of arg.
  **/
- shared_ptr<Node> ngraph::op::broadcast(const Node::ptr&      tensor,
-                                       const Shape&          shape,
-                                       const vector<size_t>& broadcast_axes)
+Node::ptr ngraph::op::broadcast(const Node::ptr&      tensor,
+                                const Shape&          shape,
+                                const vector<size_t>& broadcast_axes)
+{
+    return make_shared<BroadcastOp>(tensor, shape, broadcast_axes);
+}
+
+void BroadcastOp::propagate_types()
 {
-    return make_shared<BroadcastCall>(tensor, shape, broadcast_axes);
+    auto arg_type = m_arguments.at(0)->type();
+    if (nullptr == arg_type)
+    {
+        throw ngraph_error("Argument to broadcast is missing type.");
+    }
+    auto arg_tensor_view_type = dynamic_pointer_cast<TensorViewType>(arg_type);
+    if (nullptr == arg_tensor_view_type)
+    {
+        throw ngraph_error("Argument to broadcast is not a tensor view");
+    }
+    vector<size_t> target_shape = m_shape;
+    for (auto i = m_broadcast_axes.rbegin(); i != m_broadcast_axes.rend(); ++i)
+    {
+        target_shape.erase(target_shape.begin() + *i);
+    }
+    if (Shape{target_shape} != arg_tensor_view_type->shape())
+    {
+        throw ngraph_error("Broadcast arg, shape, and axes are incompatible");
+    }
+    // TODO If m_type is already set (by framework), this should verify that the type
+    // we expect is consistent with the type the framework expects.
+    m_type = make_shared<TensorViewType>(arg_tensor_view_type->element_type(), m_shape);
 }
 
-std::shared_ptr<BuiltinOp> DotCall::s_op = make_shared<BuiltinOp>("dot");
+Node::ptr ngraph::op::ceiling(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<CeilingOp>(arg0, arg1);
+}
+
+// 'concatenate',
+// 'constant',
+// 'convert',
+// 'convolution',
+
+Node::ptr ngraph::op::divide(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<DivideOp>(arg0, arg1);
+}
 
 /// TODO: Semantics of arg0 and arg1 axes wrt reduction.
-shared_ptr<Node> ngraph::op::dot(const Node::ptr& arg0, const Node::ptr& arg1)
+Node::ptr ngraph::op::dot(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<DotOp>(arg0, arg1);
+}
+
+void DotOp::propagate_types()
+{
+    auto arg0_tensor_type = dynamic_pointer_cast<TensorViewType>(m_arguments.at(0)->type());
+    auto arg1_tensor_type = dynamic_pointer_cast<TensorViewType>(m_arguments.at(1)->type());
+    if (nullptr == arg0_tensor_type || nullptr == arg1_tensor_type)
+    {
+        throw ngraph_error("Arguments to dot must be tensor views");
+    }
+    if (arg0_tensor_type->element_type() != arg1_tensor_type->element_type())
+    {
+        throw ngraph_error("Arguments to dot must have the same element type");
+    }
+
+    // Use NumPy semantics for now
+    // Last axis of first arg reduces against second to last of second arg if more than one axis, else axis.
+    vector<size_t> arg0_shape     = arg0_tensor_type->shape();
+    vector<size_t> arg1_shape     = arg1_tensor_type->shape();
+    size_t         arg0_reduction = arg0_shape.size() - 1;
+    size_t         arg1_reduction;
+    if (arg1_shape.size() > 1)
+    {
+        arg1_reduction = arg1_shape.size() - 2;
+    }
+    else
+    {
+        arg1_reduction = arg1_shape.size() - 1;
+    }
+    if (arg0_shape.at(arg0_reduction) != arg1_shape.at(arg1_reduction))
+    {
+        throw ngraph_error("Dot reduction axes not compatible");
+    }
+    vector<size_t> result_shape;
+    copy(arg0_shape.begin(), arg0_shape.begin() + arg1_reduction, result_shape.end());
+    copy(arg1_shape.begin(), arg1_shape.begin() + arg1_reduction, result_shape.end());
+    copy(arg1_shape.begin() + arg1_reduction, arg1_shape.end(), result_shape.end());
+    m_type = make_shared<TensorViewType>(arg0_tensor_type->element_type(), result_shape);
+}
+
+Node::ptr ngraph::op::exponential(const Node::ptr& arg0)
+{
+    return make_shared<ExponentialOp>(arg0);
+}
+
+Node::ptr ngraph::op::floor(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<FloorOp>(arg0, arg1);
+}
+
+Node::ptr ngraph::op::log(const Node::ptr& arg0)
 {
-    return make_shared<DotCall>(arg0, arg1);
+    return make_shared<LogOp>(arg0);
 }
+
+Node::ptr ngraph::op::maximum(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<MaximumOp>(arg0, arg1);
+}
+
+Node::ptr ngraph::op::minimum(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<MinimumOp>(arg0, arg1);
+}
+
+Node::ptr ngraph::op::multiply(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<MultiplyOp>(arg0, arg1);
+}
+
+Node::ptr ngraph::op::negate(const Node::ptr& arg0)
+{
+    return make_shared<NegateOp>(arg0);
+}
+
+// 'pad',
+// 'parameter',
+
+Node::ptr ngraph::op::power(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<PowerOp>(arg0, arg1);
+}
+
+//'reduce',
+
+Node::ptr ngraph::op::remainder(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<RemainderOp>(arg0, arg1);
+}
+
+Node::ptr ngraph::op::reshape(const Node::ptr& arg0, const Shape& shape)
+{
+    return make_shared<ReshapeOp>(arg0, shape);
+}
+
+//'reverse',
+//'rng',
+// 'select',
+//'slice',
+
+Node::ptr ngraph::op::subtract(const Node::ptr& arg0, const Node::ptr& arg1)
+{
+    return make_shared<SubtractOp>(arg0, arg1);
+}
+
+// 'transpose',
+//'tuple',
+// 'while'

test/build_graph.cpp

@@ -19,7 +19,7 @@
 using namespace std;
 using namespace ngraph;
 
-TEST(DISABLED_graph, build_simple)
+TEST(build_graph, build_simple)
 {
     // Function with 4 parameters
     auto cluster_0 = make_shared<Function>(4);
@@ -29,11 +29,9 @@ TEST(DISABLED_graph, build_simple)
     cluster_0->parameter(2)->type(element::float32_t, {32, 7});
     cluster_0->parameter(3)->type(element::float32_t, {32, 7});
     auto arg3 = cluster_0->parameter(3);
-    // call broadcast op on arg3, broadcasting on axis 0.
     auto broadcast_1 = op::broadcast(arg3, {10, 32, 7}, {0});
     auto arg2        = cluster_0->parameter(2);
     auto arg0        = cluster_0->parameter(0);
-    // call dot op
     auto dot = op::dot(arg2, arg0);
     ASSERT_EQ(dot->arguments()[0], arg2);
     ASSERT_EQ(dot->arguments()[1], arg0);
@@ -42,3 +40,52 @@ TEST(DISABLED_graph, build_simple)
 
     ASSERT_EQ(cluster_0->result()->value(), dot);
 }
+
+// Check upcasting from ValueType.
+TEST(build_graph, as_type)
+{
+    // Check upcasting a ValueType::ptr that is a TensorViewType to a TensorViewType and Tuple.
+    ValueType::ptr tv_vt = make_shared<TensorViewType>(element::float32_t, Shape{2, 3, 5});
+    auto tv_tv = dynamic_pointer_cast<TensorViewType>(tv_vt);
+    ASSERT_EQ(tv_vt, tv_tv);
+    auto tv_tp = dynamic_pointer_cast<TupleType>(tv_vt);
+    ASSERT_EQ(nullptr, tv_tp);
+
+    // Check upcasting a ValueType::ptr that is a TupleType to a TensorViewType and Tuple.
+    ValueType::ptr tp_vt = make_shared<TupleType>(vector<ValueType::ptr>{tv_vt, tv_vt});
+    auto tp_tv = dynamic_pointer_cast<TensorViewType>(tp_vt);
+    ASSERT_EQ(nullptr, tp_tv);
+    auto tp_tp = dynamic_pointer_cast<TupleType>(tp_vt);
+    ASSERT_EQ(tp_vt, tp_tp);
+}
+
+// Check node comparisons
+TEST(build_graph, node_comparison)
+{
+    auto fun = make_shared<Function>(3);
+    fun->parameter(0)->type(element::float32_t, {32, 3});
+    fun->parameter(1)->type(element::float32_t, {3});
+    fun->parameter(2)->type(element::float32_t, {32});
+    auto arg0 = fun->parameter(0);
+    auto arg1 = fun->parameter(1);
+    auto arg2 = fun->parameter(2);
+
+    auto dot = op::dot(arg0, arg1);
+    auto add = op::add(dot, arg2);
+
+    auto pattern = make_shared<Function>(1);
+    pattern->parameter(0)->type(element::float32_t, {});
+    auto parg = pattern->parameter(0);
+    auto pattern_dot = op::dot(parg, parg);
+    ASSERT_TRUE(pattern_dot->is_same_op_type(dot)); 
+    // TODO This passes because typeid is not behaving as documented.
+    // Need to figure out what's wrong.
+    ASSERT_FALSE(pattern_dot->is_same_op_type(add));
+}
+
+// Check argument inverses
+TEST(build_graph, arg_inverse)
+{
+
+}
+