[BUILD] add target_host to compiler.build (#240)

nnvm/include/nnvm/graph.h

@@ -47,6 +47,12 @@ class Graph {
    */
   template<typename T>
   inline const T& GetAttr(const std::string& attr_name) const;
+  /*!
+   * \brief Check whether has a specific attribute.
+   * \param attr_name the name of the attribute
+   * \return a boolean result
+   */
+  inline bool HasAttr(const std::string& attr_name) const;
   /*!
    * \brief Get a move copy of the attribute, implement copy on write semantics.
    *  The content is moved if the reference counter of shared_ptr is 1.
@@ -226,6 +232,11 @@ inline const T& Graph::GetAttr(const std::string& attr_name) const {
   return nnvm::get<T>(*it->second);
 }
 
+inline bool Graph::HasAttr(const std::string& attr_name) const {
+  auto it = attrs.find(attr_name);
+  return it != attrs.end();
+}
+
 template<typename T>
 inline T Graph::MoveCopyAttr(const std::string& attr_name) {
   auto it = attrs.find(attr_name);

nnvm/python/nnvm/compiler/build_module.py

@@ -112,8 +112,10 @@ def _lower(sch, inputs, func_name, graph):
 
 
 @tvm.register_func("nnvm.compiler.build_target")
-def _build(funcs, target):
-    return tvm.build(funcs, target=target)
+def _build(funcs, target, target_host):
+    if target_host == "":
+        target_host = None
+    return tvm.build(funcs, target=target, target_host=target_host)
 
 
 def _update_shape_dtype(shape, dtype, params):
@@ -161,7 +163,7 @@ def optimize(graph, shape, dtype="float32"):
     return graph
 
 
-def build(graph, target=None, shape=None, dtype="float32", params=None):
+def build(graph, target=None, shape=None, dtype="float32", params=None, target_host=None):
     """Build graph into runtime library.
 
     The build function will optimize the graph and do the compilation.
@@ -189,6 +191,15 @@ def build(graph, target=None, shape=None, dtype="float32", params=None):
         during inference time. Used for pre-compute
         folding optimization.
 
+    target_host : str or :any:`tvm.target.Target` optional
+        Host compilation target, if target is device.
+        When TVM compiles device specific program such as CUDA,
+        we also need host(CPU) side code to interact with the driver
+        setup the dimensions and parameters correctly.
+        target_host is used to specify the host side codegen target.
+        By default, llvm is used if it is enabled,
+        otherwise a stackvm intepreter is used.
+
     Returns
     -------
     graph : Graph
@@ -228,6 +239,8 @@ def build(graph, target=None, shape=None, dtype="float32", params=None):
     graph = graph_attr.set_shape_inputs(graph, shape)
     graph = graph_attr.set_dtype_inputs(graph, dtype)
     graph._set_json_attr("target", str(target), "str")
+    if target_host is not None:
+        graph._set_json_attr("target_host", str(target_host), "str")
     if cfg.pass_enabled("OpFusion"):
         graph._set_json_attr("opt_level", 1, "int")
     else:

nnvm/src/compiler/graph_fuse.cc

@@ -219,6 +219,10 @@ nnvm::Graph GraphFuseCompile(nnvm::Graph g) {
   const std::vector<TOpPattern>& pattern_vec =
       g.GetAttr<std::vector<TOpPattern> >("pattern");
   std::string target = g.GetAttr<std::string>("target");
+  std::string target_host;
+  if (g.HasAttr("target_host"))
+    target_host = g.GetAttr<std::string>("target_host");
+
   std::vector<FuseEntry> fuse_vec(idx.num_nodes());
   // setup inputs and placeholder.
   for (uint32_t nid = 0; nid < idx.num_nodes(); ++nid) {
@@ -398,7 +402,7 @@ nnvm::Graph GraphFuseCompile(nnvm::Graph g) {
   ret.attrs["dltype"] = std::make_shared<any>(std::move(new_dltype_vec));
   // Setup module
   static const PackedFunc& fbuild = GetPackedFunc("nnvm.compiler.build_target");
-  tvm::runtime::Module module = fbuild(func_list, target);
+  tvm::runtime::Module module = fbuild(func_list, target, target_host);
   ret.attrs["module"] = std::make_shared<any>(std::move(module));
   ret = nnvm::ApplyPass(ret, "PlanMemory");
   return ret;