[Runtime] Allow parameter sharing between modules (apache#3489)

As GraphRuntime does not provide control-flow logics, we have to split
our model to two parts. While we need to share parameters between them
to save memory usage.

Solution:
1) add "lazy_init_input" in graph's attributes
   "attrs": {
     ... ...
     "lazy_init_input": [
       "list_str",
       [
         "p0"
       ]
     ]
    }
2) allow un-allocated NDArray entry in SetupStorage
3) utilize "set_input_zero_copy" function to set parameters

include/tvm/runtime/ndarray.h

@@ -99,6 +99,8 @@ class NDArray {
   bool defined() const {
     return data_ != nullptr;
   }
+  /*! \return If NDArray is allocated*/
+  inline bool allocated() const;
   /*! \return If both NDArray reference the same container */
   bool same_as(const NDArray& other) const {
     return data_ == other.data_;
@@ -164,11 +166,13 @@ class NDArray {
    * \param shape The shape of the new array.
    * \param dtype The data type of the new array.
    * \param ctx The context of the Array.
+   * \param allocate Allocate memory if true.
    * \return The created Array
    */
   TVM_DLL static NDArray Empty(std::vector<int64_t> shape,
                                DLDataType dtype,
-                               DLContext ctx);
+                               DLContext ctx,
+                               bool allocate = true);
   /*!
    * \brief Create a NDArray backed by a dlpack tensor.
    *
@@ -354,6 +358,10 @@ inline void NDArray::reset() {
   }
 }
 
+inline bool NDArray::allocated() const {
+  return defined() && data_->dl_tensor.data != nullptr;
+}
+
 /*! \brief return the size of data the DLTensor hold, in term of number of bytes
  *
  *  \param arr the input DLTensor

src/runtime/graph/graph_runtime.cc

@@ -54,7 +54,15 @@ inline size_t GetDataAlignment(const DLTensor& arr) {
 void GraphRuntime::Run() {
   // setup the array and requirements.
   for (size_t i = 0; i < op_execs_.size(); ++i) {
-    if (op_execs_[i]) op_execs_[i]();
+    if (op_execs_[i]) {
+      auto& op_arg = op_args_[i];
+      if (op_arg) {
+        for (auto& arg : op_arg->args) {
+          CHECK(arg.data != nullptr) << "Un-initialized input!";
+        }
+      }
+      op_execs_[i]();
+    }
   }
 }
 /*!
@@ -106,6 +114,8 @@ int GraphRuntime::GetInputIndex(const std::string& name) {
 void GraphRuntime::SetInput(int index, DLTensor* data_in) {
   CHECK_LT(static_cast<size_t>(index), input_nodes_.size());
   uint32_t eid = this->entry_id(input_nodes_[index], 0);
+  CHECK(data_entry_[eid].allocated())
+      << "Invoke 'set_input_zero_copy' for 'lazy_init_input' entry!";
   data_entry_[eid].CopyFrom(data_in);
 }
 /*!
@@ -255,7 +265,14 @@ void GraphRuntime::SetupStorage() {
   for (const std::string& s_type : attrs_.dltype) {
     vtype.push_back(tvm::runtime::String2TVMType(s_type));
   }
-
+  // get the entry id(s) of lazy initialized inputs
+  std::vector<uint32_t> lazy_init_entries;
+  for (auto const& name : attrs_.lazy_init_input) {
+    int in_idx = GetInputIndex(name);
+    CHECK_GE(in_idx, 0) << "input \"" << name << "\" does not exist!";
+    uint32_t eid = this->entry_id(input_nodes_[in_idx], 0);
+    lazy_init_entries.push_back(eid);
+  }
   // Size and device type of each storage pool entry.
   std::vector<PoolEntry> pool_entry;
   // Find the maximum space size.
@@ -286,6 +303,8 @@ void GraphRuntime::SetupStorage() {
     }
     pool_entry[sid].size = std::max(pool_entry[sid].size, bytes);
     pool_entry[sid].device_type = device_type;
+    pool_entry[sid].lazy_init = (std::find(lazy_init_entries.begin(),
+        lazy_init_entries.end(), i) != lazy_init_entries.end());
   }
 
   // Allocate the space.
@@ -300,7 +319,7 @@ void GraphRuntime::SetupStorage() {
     TVMContext ctx = cit == ctxs_.end() ? ctxs_[0] : *cit;
     shape.push_back(static_cast<int64_t>(pit.size + 3) / 4);
     storage_pool_.push_back(
-        NDArray::Empty(shape, DLDataType{kDLFloat, 32, 1}, ctx));
+        NDArray::Empty(shape, DLDataType{kDLFloat, 32, 1}, ctx, !pit.lazy_init));
   }
 
   // Assign the pooled entries. A unified memory pool is used to simplifiy

src/runtime/graph/graph_runtime.h

@@ -188,7 +188,8 @@ class GraphRuntime : public ModuleNode {
   struct PoolEntry {
     size_t size;
     int device_type;
-    PoolEntry(int s, int dev_type) : size(s), device_type(dev_type) {}
+    bool lazy_init;
+    PoolEntry(int s, int dev_type) : size(s), device_type(dev_type), lazy_init(false) {}
   };
   // Node entry
   struct NodeEntry {
@@ -277,6 +278,7 @@ class GraphRuntime : public ModuleNode {
     std::vector<int> device_index;
     std::vector<std::string> dltype;
     std::vector<std::vector<int64_t> > shape;
+    std::vector<std::string> lazy_init_input;
     // The graph attribute fields.
     void Load(dmlc::JSONReader *reader) {
       reader->BeginObject();
@@ -318,6 +320,14 @@ class GraphRuntime : public ModuleNode {
           CHECK(reader->NextArrayItem());
           reader->Read(&device_index);
           CHECK(!reader->NextArrayItem());
+        } else if (key == "lazy_init_input") {
+          reader->BeginArray();
+          CHECK(reader->NextArrayItem());
+          reader->Read(&type);
+          CHECK_EQ(type, "list_str");
+          CHECK(reader->NextArrayItem());
+          reader->Read(&lazy_init_input);
+          CHECK(!reader->NextArrayItem());
         } else {
           reader->BeginArray();
           CHECK(reader->NextArrayItem());

src/runtime/ndarray.cc

@@ -142,14 +142,17 @@ DLManagedTensor* NDArray::ToDLPack() const {
 
 NDArray NDArray::Empty(std::vector<int64_t> shape,
                        DLDataType dtype,
-                       DLContext ctx) {
+                       DLContext ctx,
+                       bool allocate) {
   NDArray ret = Internal::Create(shape, dtype, ctx);
-  // setup memory content
-  size_t size = GetDataSize(ret.data_->dl_tensor);
-  size_t alignment = GetDataAlignment(ret.data_->dl_tensor);
-  ret.data_->dl_tensor.data =
-      DeviceAPI::Get(ret->ctx)->AllocDataSpace(
-          ret->ctx, size, alignment, ret->dtype);
+  if (allocate) {
+    // setup memory content
+    size_t size = GetDataSize(ret.data_->dl_tensor);
+    size_t alignment = GetDataAlignment(ret.data_->dl_tensor);
+    ret.data_->dl_tensor.data =
+        DeviceAPI::Get(ret->ctx)->AllocDataSpace(
+            ret->ctx, size, alignment, ret->dtype);
+  }
   return ret;
 }
 

tests/cpp/build_module_test.cc

@@ -189,6 +189,77 @@ TEST(BuildModule, Heterogeneous) {
   }
 }
 
+TEST(BuildModule, LazyInitInput) {
+  using namespace tvm;
+
+  const int n = 4;
+  Array<Expr> shape{n};
+
+  auto A = placeholder(shape, Float(32), "A");
+  auto B = placeholder(shape, Float(32), "B");
+
+  auto C = compute(A->shape, [&A, &B](Expr i) {
+    return A[i] + B[i];
+  }, "C");
+
+  auto s = create_schedule({ C->op });
+  auto args = Array<Tensor>({ A, B, C });
+  std::unordered_map<Tensor, Buffer> binds;
+
+  auto config = BuildConfig::Create();
+  auto target = target::llvm();
+
+  auto lowered = lower(s, args, "myadd", binds, config);
+  auto module = build(lowered, target, Target(), config);
+
+  std::string json =
+      "{\"nodes\": [{\"op\": \"null\", \"name\": \"x\", \"inputs\": []}, {\"op\": \"null\", \"name\": \"y\", \"inputs\": []}, "
+      "{\"op\": \"tvm_op\", \"name\": \"add\", \"inputs\": [[0, 0, 0], [1, 0, 0]], \"attrs\": {\"func_name\": "
+      "\"myadd\", \"flatten_data\": \"1\", \"num_inputs\": \"2\", \"num_outputs\": \"1\"}}], "
+      "\"arg_nodes\": [0, 1], \"node_row_ptr\": [0, 1, 2, 3], \"heads\": [[2, 0, 0]], "
+      "\"attrs\": {\"shape\": [\"list_shape\", [[4], [4], [4]]], \"dltype\": [\"list_str\", [\"float32\", \"float32\", \"float32\"]], "
+      "\"storage_id\": [\"list_int\", [0, 1, 2]], \"lazy_init_input\": [\"list_str\", [\"y\"]]}}";
+
+  // Setup inputs.
+  auto a_val = runtime::NDArray::Empty({n}, {kDLFloat, 32, 1}, {kDLCPU, 0});
+  auto b_val = runtime::NDArray::Empty({n}, {kDLFloat, 32, 1}, {kDLCPU, 0});
+
+  auto pa = (float*)a_val.ToDLPack()->dl_tensor.data;
+  auto pb = (float*)b_val.ToDLPack()->dl_tensor.data;
+
+  // Assign values.
+  for (int i = 0; i < n; i++) {
+    pa[i] = pb[i] = i;
+  }
+
+  // Initialize graph runtime.
+  int cpu_dev_ty = static_cast<int>(kDLCPU);
+  int cpu_dev_id = 0;
+
+  const runtime::PackedFunc* graph_runtime =
+      tvm::runtime::Registry::Get("tvm.graph_runtime.create");
+  runtime::Module mod = (*graph_runtime)(json, module, cpu_dev_ty, cpu_dev_id);
+
+  PackedFunc get_input = mod.GetFunction("get_input", false);
+  CHECK(((runtime::NDArray)get_input("x")).allocated());
+  CHECK(!((runtime::NDArray)get_input("y")).allocated());
+
+  PackedFunc set_input = mod.GetFunction("set_input", false);
+  PackedFunc set_input_zero_copy = mod.GetFunction("set_input_zero_copy", false);
+  PackedFunc run = mod.GetFunction("run", false);
+  PackedFunc get_output = mod.GetFunction("get_output", false);
+  set_input("x", a_val);
+  set_input_zero_copy("y", b_val);
+  run();
+  tvm::runtime::NDArray out = get_output(0);
+  float* p_out = (float*)out.ToDLPack()->dl_tensor.data;
+
+  // Check correctness.
+  for (int i = 0; i < n; ++i) {
+    CHECK_LT(std::fabs(p_out[i] - i*2), 1e-5);
+  }
+}
+
 int main(int argc, char ** argv) {
   testing::InitGoogleTest(&argc, argv);
   testing::FLAGS_gtest_death_test_style = "threadsafe";