Bugfix plan memory, fully support mxnet executor (#32)

* [PASS] include knullop info in plan memory

* Bugfix plan memory, fully support mxnet

nnvm/include/nnvm/op_attr_types.h

@@ -31,6 +31,18 @@ namespace nnvm {
 using FListInputNames = std::function<std::vector<std::string> (const NodeAttrs& attrs)>;
 
 /*!
+ * \brief Return number of visible outputs by the user.
+ *
+ * \param attrs The attributes of the node.
+ *
+ * \note Register under "FNumVisibleOutputs", default not registered.
+ *  This can be used to hide certain output from the user,
+ *  but the additional outputs can be used to pass information from
+ *  forward to gradient pass.
+ */
+using FNumVisibleOutputs = std::function<uint32_t (const NodeAttrs& attrs)>;
+
+/*!
  * \brief Return list of output arguments names of each operator.
  *
  * \param attrs The attributes of the node.

nnvm/src/core/symbolic.cc

@@ -87,7 +87,7 @@ inline std::vector<std::string> GetKeys(
 
 // whether the symbol is atomic functor
 inline bool IsAtomic(const std::vector<NodeEntry>& outputs) {
-  return outputs.size() == 1 && outputs[0].node->inputs.size() == 0;
+  return outputs[0].node->inputs.size() == 0;
 }
 
 // public functions
@@ -222,6 +222,7 @@ std::vector<std::string> Symbol::ListInputNames(ListInputOption option) const {
 
 std::vector<std::string> Symbol::ListOutputNames() const {
   static auto& flist_ouputs = Op::GetAttr<FListOutputNames>("FListOutputNames");
+
   std::vector<std::string> ret;
   for (auto &head : outputs) {
     if (head.node->is_variable()) {
@@ -256,8 +257,6 @@ void Symbol::Compose(const array_view<const Symbol*>& args,
                      const std::string& name) {
   static auto& flist_inputs = Op::GetAttr<FListInputNames>("FListInputNames");
 
-  CHECK_EQ(outputs.size(), 1)
-      << "Only composition of value function is supported currently";
   CHECK(!outputs[0].node->is_variable()) << "Variable cannot be composed";
   // parameter check.
   for (size_t i = 0; i < args.size(); ++i) {
@@ -400,6 +399,7 @@ void Symbol::AddControlDeps(const Symbol& src) {
 }
 
 Symbol Symbol::GetInternals() const {
+  static auto& fnum_vis_output = Op::GetAttr<FNumVisibleOutputs>("FNumVisibleOutputs");
   Symbol ret;
   DFSVisit(this->outputs, [&ret](const NodePtr& node) {
       Node* n = node.get();
@@ -409,6 +409,9 @@ Symbol Symbol::GetInternals() const {
         ret.outputs.emplace_back(NodeEntry{node, 0, param.version});
       } else {
         uint32_t nout = n->num_outputs();
+        if (fnum_vis_output.count(n->op())) {
+          nout = fnum_vis_output[n->op()](n->attrs);
+        }
         for (uint32_t i = 0; i < nout; ++i) {
           ret.outputs.emplace_back(NodeEntry{node, i, 0});
         }
@@ -467,6 +470,7 @@ std::unordered_map<std::string, std::string> Symbol::ListAttrs(ListAttrOption op
 
 Symbol Symbol::CreateFunctor(const Op* op,
                              std::unordered_map<std::string, std::string> attrs) {
+  static auto& fnum_vis_output = Op::GetAttr<FNumVisibleOutputs>("FNumVisibleOutputs");
   Symbol s;
   NodePtr n = Node::Create();
   n->attrs.op = op;
@@ -474,7 +478,14 @@ Symbol Symbol::CreateFunctor(const Op* op,
   if (n->op()->attr_parser != nullptr) {
     n->op()->attr_parser(&(n->attrs));
   }
-  s.outputs.emplace_back(NodeEntry{std::move(n), 0, 0});
+
+  uint32_t nout = n->num_outputs();
+  if (fnum_vis_output.count(n->op())) {
+    nout = fnum_vis_output[n->op()](n->attrs);
+  }
+  for (uint32_t i = 0; i < nout; ++i) {
+    s.outputs.emplace_back(NodeEntry{n, i, 0});
+  }
   return s;
 }
 

nnvm/src/pass/place_device.cc

@@ -12,7 +12,6 @@ namespace nnvm {
 namespace pass {
 namespace {
 
-
 // simply logic to place device according to device_group hint
 // insert copy node when there is
 Graph PlaceDevice(Graph src) {

nnvm/src/pass/plan_memory.cc

@@ -142,11 +142,11 @@ Graph PlanMemory(Graph ret) {
   // step 1: initialize reference count
   for (uint32_t nid = 0; nid < idx.num_nodes(); ++nid) {
     for (const auto& e : idx[nid].inputs) {
-      ++ref_count[e.node_id];
+      ++ref_count[idx.entry_id(e)];
     }
   }
   for (const auto& e : idx.outputs()) {
-    ++ref_count[e.node_id];
+    ++ref_count[idx.entry_id(e)];
   }
   // step 2: allocate memory.
   StorageVector storage(idx.num_node_entries(), -1);
@@ -202,10 +202,13 @@ Graph PlanMemory(Graph ret) {
       }
     }
     // check if there are outputs that can be freeded immediately
+    // these output are not referenced by any operator.
     for (uint32_t index = 0; index < inode.source->num_outputs(); ++index) {
       uint32_t eid = idx.entry_id(nid, index);
       if (ref_count[eid] == 0 && storage[eid] != GraphAllocator::kBadStorageID) {
         allocator.Release(storage[eid], nid);
+        // use -2 to indicate that the node was never touched.
+        storage_inplace_index[eid] = -2;
       }
       if (storage[eid] == GraphAllocator::kBadStorageID) {
         ++num_not_allocated;