Optimize Linux shared library modules (*.so files) (#2445)

2019-01-28 21:14:18 -08:00 · 2019-01-28 21:14:18 -08:00 · 75f91c45d7
--- a/3rdparty/HalideIR
+++ b/3rdparty/HalideIR
@ -1 +1 @@
-Subproject commit 6e7c1f046fda536562dc80977e93324fee2324bd
+Subproject commit 97efb11fff13131480fcaa5adc65a0aef4a4cb5d
--- a/3rdparty/dlpack
+++ b/3rdparty/dlpack
@ -1 +1 @@
-Subproject commit bee4d1dd8dc1ee4a1fd8fa6a96476c2f8b7492a3
+Subproject commit 5c792cef3aee54ad8b7000111c9dc1797f327b59
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -84,10 +84,10 @@ else(MSVC)
  include(CheckCXXCompilerFlag)
  check_cxx_compiler_flag("-std=c++11"    SUPPORT_CXX11)
  if ("${CMAKE_BUILD_TYPE}" STREQUAL "Debug")
-    add_compile_options(-O0 -Wall -fPIC -std=c++11)
+    add_compile_options(-O0 -Wall -fPIC -fvisibility=hidden -std=c++11)
  else()
-    set(CMAKE_C_FLAGS "-O2 -Wall -fPIC ${CMAKE_C_FLAGS}")
-    set(CMAKE_CXX_FLAGS "-O2 -Wall -fPIC -std=c++11 ${CMAKE_CXX_FLAGS}")
+    set(CMAKE_C_FLAGS "-O2 -Wall -fPIC -fvisibility=hidden ${CMAKE_C_FLAGS}")
+    set(CMAKE_CXX_FLAGS "-O2 -Wall -fPIC -fvisibility=hidden -std=c++11 ${CMAKE_CXX_FLAGS}")
  endif ()
  if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" AND
      CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 7.0)
--- a/docs/Doxyfile
+++ b/docs/Doxyfile
@ -1974,7 +1974,7 @@ INCLUDE_FILE_PATTERNS  =
 # recursively expanded use the := operator instead of the = operator.
 # This tag requires that the tag ENABLE_PREPROCESSING is set to YES.

-PREDEFINED             = DMLC_USE_CXX11
+PREDEFINED             = DMLC_USE_CXX11 TVM_DLL= NNVM_DLL= __attribute__(x)=

 # If the MACRO_EXPANSION and EXPAND_ONLY_PREDEF tags are set to YES then this
 # tag can be used to specify a list of macro names that should be expanded. The
--- a/include/tvm/ir_operator.h
+++ b/include/tvm/ir_operator.h
@ -332,26 +332,6 @@ TVM_DLL Expr max(Expr a, Expr b);
 *       index types(int32, int64) when possible.
 */
 TVM_DLL Expr min(Expr a, Expr b);
-/*!
- * \brief right shift
- *
- * \param a left operand
- * \param b right operand
- * \return The result expression.
- * \note this function does eager constant folding for
- *       index types(int32, int64) when possible.
- */
-TVM_DLL Expr operator>>(Expr a, Expr b);
-/*!
- * \brief left shift
- *
- * \param a left operand
- * \param b right operand
- * \return The result expression.
- * \note this function does eager constant folding for
- *       index types(int32, int64) when possible.
- */
-TVM_DLL Expr operator<<(Expr a, Expr b);
 /*!
 * \brief take bitwise and of two values
 *
--- a/include/tvm/ir_pass.h
+++ b/include/tvm/ir_pass.h
@ -27,7 +27,7 @@ namespace ir {
 * \param vrange The range information about the variable.
 * \return Canonicalized statement.
 */
-EXPORT Expr Simplify(Expr expr, Map<Var, Range> vrange = Map<Var, Range>());
+TVM_DLL Expr Simplify(Expr expr, Map<Var, Range> vrange = Map<Var, Range>());

 /*!
 * \brief Simplify the statement.
@ -52,7 +52,7 @@ Stmt CanonicalSimplify(Stmt stmt,
 * \param vrange The range information about the variable.
 * \return Canonicalized expression.
 */
-EXPORT Expr CanonicalSimplify(Expr expr,
+TVM_DLL Expr CanonicalSimplify(Expr expr,
                              Map<Var, Range> vrange = Map<Var, Range>());

 /*!
@ -61,7 +61,7 @@ EXPORT Expr CanonicalSimplify(Expr expr,
 * \param rhs The right operand
 * \return The comparison result.
 */
-EXPORT bool Equal(const Expr& lhs, const Expr& rhs);
+TVM_DLL bool Equal(const Expr& lhs, const Expr& rhs);

 /*!
 * \brief Deep compare lhs and rhs
@ -92,13 +92,13 @@ int Compare(const Expr& lhs, const Expr& rhs);
 * \return Whether IR is in SSA form.
 * \note All the passes in this file uses SSA form and outputs SSA form.
 */
-bool VerifySSA(const Stmt& ir);
+TVM_DLL bool VerifySSA(const Stmt& ir);

 /*!
 * \brief Whether the expression have side effect.
 * \return whether expression have side effect
 */
-bool HasSideEffect(const Expr& e);
+TVM_DLL bool HasSideEffect(const Expr& e);

 /*!
 * \brief Whether e expression used var.
@ -121,7 +121,7 @@ bool ExprUseVar(const Expr& e, const std::unordered_set<const Variable*>& vset);
 * \param stmt The source statement to be converted.
 * \return The converted form.
 */
-Stmt ConvertSSA(Stmt stmt);
+TVM_DLL Stmt ConvertSSA(Stmt stmt);

 /*!
 * \brief Substitute the var specified in key->var to be value.
--- a/include/tvm/ir_visitor.h
+++ b/include/tvm/ir_visitor.h
@ -131,7 +131,7 @@ class TVM_DLL IRVisitor {
 * \param node The ir to be visited.
 * \param fvisit The visitor function to be applied.
 */
-void PostOrderVisit(const NodeRef& node, std::function<void(const NodeRef&)> fvisit);
+TVM_DLL void PostOrderVisit(const NodeRef& node, std::function<void(const NodeRef&)> fvisit);

 }  // namespace ir
 }  // namespace tvm
--- a/include/tvm/relay/pass.h
+++ b/include/tvm/relay/pass.h
@ -27,7 +27,7 @@ namespace relay {
 *
 * \return A type checked expression with its checked_type field populated.
 */
-Expr InferType(const Expr& expr, const Module& mod);
+TVM_DLL Expr InferType(const Expr& expr, const Module& mod);

 /*!
 * \brief Infer the type of a function as if it is mapped to var in the mod.
@ -39,7 +39,7 @@ Expr InferType(const Expr& expr, const Module& mod);
 * \return A type checked Function with its checked_type field populated.
 * \note this function mutates mod and is not thread-safe.
 */
-Function InferType(const Function& f, const Module& mod,
+TVM_DLL Function InferType(const Function& f, const Module& mod,
                           const GlobalVar& var);

 /*!
@ -58,7 +58,7 @@ Function InferType(const Function& f, const Module& mod,
 *
 * \return true if the rules are satisified otherwise false
 */
-bool KindCheck(const Type& t, const Module& mod);
+TVM_DLL bool KindCheck(const Type& t, const Module& mod);

 /*! \brief Compare two expressions for structural equivalence.
 *
@ -75,7 +75,7 @@ bool KindCheck(const Type& t, const Module& mod);
 *
 *   \return true if equal, otherwise false
 */
-bool AlphaEqual(const Expr& e1, const Expr& e2);
+TVM_DLL bool AlphaEqual(const Expr& e1, const Expr& e2);

 /*! \brief Compare two types for structural equivalence.
 *
@ -93,7 +93,7 @@ bool AlphaEqual(const Expr& e1, const Expr& e2);
 *
 * \return true if equal, otherwise false
 */
-bool AlphaEqual(const Type& t1, const Type& t2);
+TVM_DLL bool AlphaEqual(const Type& t1, const Type& t2);

 /*! \brief Check that each Var is only bound once.
 *
@ -106,7 +106,7 @@ bool AlphaEqual(const Type& t1, const Type& t2);
 *
  * \return true iff all Var in expr is bound at most once.
 */
-bool WellFormed(const Expr& expr);
+TVM_DLL bool WellFormed(const Expr& expr);

 /*! \brief Get all bound variables from expression expr.
 *
@ -117,7 +117,7 @@ bool WellFormed(const Expr& expr);
 *
 * \return List of bound vars, in the PostDFS order in the expression.
 */
-tvm::Array<Var> BoundVars(const Expr& expr);
+TVM_DLL tvm::Array<Var> BoundVars(const Expr& expr);

 /*! \brief Get free type parameters from expression expr.
 *
@ -128,7 +128,7 @@ tvm::Array<Var> BoundVars(const Expr& expr);
 *
 * \return List of free vars, in the PostDFS order in the expression.
 */
-tvm::Array<Var> FreeVars(const Expr& expr);
+TVM_DLL tvm::Array<Var> FreeVars(const Expr& expr);

 /*! \brief Get all variables from expression expr.
 *
@ -136,7 +136,7 @@ tvm::Array<Var> FreeVars(const Expr& expr);
 *
 * \return List of all vars, in the PostDFS order in the expression.
 */
-tvm::Array<Var> AllVars(const Expr& expr);
+TVM_DLL tvm::Array<Var> AllVars(const Expr& expr);

 /*! \brief Get free TypeVars from expression expr.
 *
@ -147,7 +147,7 @@ tvm::Array<Var> AllVars(const Expr& expr);
 *
 * \return List of free vars, in the PostDFS order visited by expr.
 */
-tvm::Array<TypeVar> FreeTypeVars(const Expr& expr);
+TVM_DLL tvm::Array<TypeVar> FreeTypeVars(const Expr& expr);

 /*! \brief Get free TypeVars from type t.
 *
@ -158,7 +158,7 @@ tvm::Array<TypeVar> FreeTypeVars(const Expr& expr);
 *
 * \return List of free type vars, in the PostDFS order visited by type.
 */
-tvm::Array<TypeVar> FreeTypeVars(const Type& t);
+TVM_DLL tvm::Array<TypeVar> FreeTypeVars(const Type& t);

 /*! \brief Get all bound type variables from expression expr.
 *
@ -169,7 +169,7 @@ tvm::Array<TypeVar> FreeTypeVars(const Type& t);
 *
 * \return List of bound type vars, in the PostDFS order in the expression.
 */
-tvm::Array<TypeVar> BoundTypeVars(const Expr& expr);
+TVM_DLL tvm::Array<TypeVar> BoundTypeVars(const Expr& expr);

 /*! \brief Get all bound type variables from type t.
 *
@ -180,7 +180,7 @@ tvm::Array<TypeVar> BoundTypeVars(const Expr& expr);
 *
 * \return List of bound type vars, in the PostDFS order visited by type.
 */
-tvm::Array<TypeVar> BoundTypeVars(const Type& t);
+TVM_DLL tvm::Array<TypeVar> BoundTypeVars(const Type& t);

 /*! \brief Get all type variables in expression expr.
 *
@ -188,7 +188,7 @@ tvm::Array<TypeVar> BoundTypeVars(const Type& t);
 *
 * \return List of type vars, in the PostDFS order in the expression.
 */
-tvm::Array<TypeVar> AllTypeVars(const Expr& expr);
+TVM_DLL tvm::Array<TypeVar> AllTypeVars(const Expr& expr);

 /*! \brief Get all type variables in type t.
 *
@ -196,7 +196,7 @@ tvm::Array<TypeVar> AllTypeVars(const Expr& expr);
 *
 * \return List of type vars, in the PostDFS order visited by type.
 */
-tvm::Array<TypeVar> AllTypeVars(const Type& t);
+TVM_DLL tvm::Array<TypeVar> AllTypeVars(const Type& t);

 /*! \brief Remove expressions which does not effect the program result.
 *
@ -211,14 +211,14 @@ tvm::Array<TypeVar> AllTypeVars(const Type& t);
 *
 * \return the optimized expression.
 */
-Expr DeadCodeElimination(const Expr& e);
+TVM_DLL Expr DeadCodeElimination(const Expr& e);

 /*!
 * \brief Fold constant expressions.
 * \param expr the expression to be optimized.
 * \return The optimized expression.
 */
-Expr FoldConstant(const Expr& expr);
+TVM_DLL Expr FoldConstant(const Expr& expr);

 /*!
 * \brief Fuse operations into expr into seperate functions.
@ -226,7 +226,7 @@ Expr FoldConstant(const Expr& expr);
 * \param fuse_opt_level Optimization level.
 * \return The optimized expression.
 */
-Expr FuseOps(const Expr& expr, int fuse_opt_level);
+TVM_DLL Expr FuseOps(const Expr& expr, int fuse_opt_level);

 /*!
 * \brief Apply rewrite rules to rewrite the expr in post DFS order.
@ -238,7 +238,7 @@ Expr FuseOps(const Expr& expr, int fuse_opt_level);
 *                           an Expr consumed by multiple callers.
 * \return The rewritten expression.
 */
-Expr ForwardRewrite(const Expr& expr,
+TVM_DLL Expr ForwardRewrite(const Expr& expr,
                    const std::string& rewrite_map_attr_name,
                    std::function<NodeRef(const Call&)> fcontext = nullptr,
                    std::function<Expr(const Expr&)> fmulti_ref_trigger = nullptr);
@ -252,7 +252,7 @@ Expr ForwardRewrite(const Expr& expr,
 *                           an Expr consumed by multiple callers.
 * \return The rewritten expression.
 */
-Expr ForwardRewrite(const Expr& expr,
+TVM_DLL Expr ForwardRewrite(const Expr& expr,
                    const FForwardRewrite& rewrite_func,
                    std::function<NodeRef(const Call&)> fcontext = nullptr,
                    std::function<Expr(const Expr&)> fmulti_ref_trigger = nullptr);
@ -264,14 +264,14 @@ Expr ForwardRewrite(const Expr& expr,
 *                        operators without annotation.
 * \return The updated program.
 */
-Expr RewriteAnnotatedOps(const Expr& expr, int fallback_device);
+TVM_DLL Expr RewriteAnnotatedOps(const Expr& expr, int fallback_device);

 /*!
 * \brief Collect the device mapping information of each expression.
 * \param expr The expression.
 * \return The device mapping.
 */
-Map<Expr, Integer> CollectDeviceInfo(const Expr& expr);
+TVM_DLL Map<Expr, Integer> CollectDeviceInfo(const Expr& expr);

 /*! \brief A hashing structure in the style of std::hash. */
 struct StructuralHash {
--- a/include/tvm/runtime/c_runtime_api.h
+++ b/include/tvm/runtime/c_runtime_api.h
@ -38,7 +38,7 @@
 #define TVM_DLL __declspec(dllimport)
 #endif
 #else
-#define TVM_DLL
+#define TVM_DLL __attribute__((visibility("default")))
 #endif
 #endif

--- a/include/tvm/runtime/device_api.h
+++ b/include/tvm/runtime/device_api.h
@ -40,7 +40,7 @@ constexpr int kMaxStackAlloca = 1024;
 *  \brief TVM Runtime Device API, abstracts the device
 *  specific interface for memory management.
 */
-class DeviceAPI {
+class TVM_DLL DeviceAPI {
 public:
  /*! \brief virtual destructor */
  virtual ~DeviceAPI() {}
@ -103,7 +103,7 @@ class DeviceAPI {
   *
   * \param ctx The context of allocation.
   */
-  TVM_DLL virtual TVMStreamHandle CreateStream(TVMContext ctx);
+  virtual TVMStreamHandle CreateStream(TVMContext ctx);

  /*!
   * \brief Free a stream of execution
@ -111,7 +111,7 @@ class DeviceAPI {
   * \param ctx The context of the stream
   * \param stream The pointer to be freed.
   */
-  TVM_DLL virtual void FreeStream(TVMContext ctx, TVMStreamHandle stream);
+  virtual void FreeStream(TVMContext ctx, TVMStreamHandle stream);

  /*!
   * \brief Synchronize the stream
@ -137,7 +137,7 @@ class DeviceAPI {
   * \param event_src The source stream to synchronize.
   * \param event_dst The destination stream to synchronize.
   */
-  TVM_DLL virtual void SyncStreamFromTo(TVMContext ctx,
+  virtual void SyncStreamFromTo(TVMContext ctx,
                                        TVMStreamHandle event_src,
                                        TVMStreamHandle event_dst);
  /*!
@ -156,7 +156,7 @@ class DeviceAPI {
   * \param type_hint The type of elements. Only needed by certain backends such
   * as OpenGL, as nbytes is sufficient for most backends.
   */
-  TVM_DLL virtual void* AllocWorkspace(TVMContext ctx,
+  virtual void* AllocWorkspace(TVMContext ctx,
                                       size_t nbytes,
                                       TVMType type_hint = {});
  /*!
@ -165,7 +165,7 @@ class DeviceAPI {
   * \param ctx The context of allocation.
   * \param ptr The pointer to be freed.
   */
-  TVM_DLL virtual void FreeWorkspace(TVMContext ctx, void* ptr);
+  virtual void FreeWorkspace(TVMContext ctx, void* ptr);

  /*!
   * \brief Get device API base don context.
--- a/nnvm/include/nnvm/c_api.h
+++ b/nnvm/include/nnvm/c_api.h
@ -16,7 +16,7 @@
 #define NNVM_DLL __declspec(dllimport)
 #endif
 #else
-#define NNVM_DLL
+#define NNVM_DLL __attribute__((visibility("default")))
 #endif

 /*! \brief manually define unsigned int */
--- a/src/relay/backend/interpreter.cc
+++ b/src/relay/backend/interpreter.cc
@ -145,7 +145,7 @@ class InterpreterStateNode : public Node {
    v->Visit("stack", &stack);
  }

-  TVM_DLL static InterpreterState make(Expr current_expr, Stack stack);
+  static InterpreterState make(Expr current_expr, Stack stack);

  static constexpr const char* _type_key = "relay.InterpreterState";
  TVM_DECLARE_NODE_TYPE_INFO(InterpreterStateNode, Node);
--- a/src/runtime/builtin_fp16.cc
+++ b/src/runtime/builtin_fp16.cc
@ -11,11 +11,11 @@ extern "C" {
 // disable under msvc
 #ifndef _MSC_VER

-TVM_WEAK uint16_t __gnu_f2h_ieee(float a) {
+TVM_DLL TVM_WEAK uint16_t __gnu_f2h_ieee(float a) {
  return __truncXfYf2__<float, uint32_t, 23, uint16_t, uint16_t, 10>(a);
 }

-TVM_WEAK float __gnu_h2f_ieee(uint16_t a) {
+TVM_DLL TVM_WEAK float __gnu_h2f_ieee(uint16_t a) {
  return __extendXfYf2__<uint16_t, uint16_t, 10, float, uint32_t, 23>(a);
 }

--- a/src/runtime/workspace_pool.h
+++ b/src/runtime/workspace_pool.h
@ -22,7 +22,7 @@ namespace runtime {
 *  - The release order is usually in reverse order of allocate
 *  - Repeative pattern of same allocations over different runs.
 */
-class WorkspacePool {
+class TVM_DLL WorkspacePool {
 public:
  /*!
   * \brief Create pool with specific device type and device.
--- a/vta/include/vta/runtime.h
+++ b/vta/include/vta/runtime.h
@ -11,6 +11,7 @@
 extern "C" {
 #endif

+#include <tvm/runtime/c_runtime_api.h>
 #include "driver.h"

 #define VTA_MEMCPY_H2D 1
@ -28,13 +29,13 @@ extern "C" {
 * \param size Buffer size.
 * \return A pointer to the allocated buffer.
 */
-void* VTABufferAlloc(size_t size);
+TVM_DLL void* VTABufferAlloc(size_t size);

 /*!
 * \brief Free data buffer.
 * \param buffer The data buffer to be freed.
 */
-void VTABufferFree(void* buffer);
+TVM_DLL void VTABufferFree(void* buffer);

 /*!
 * \brief Copy data buffer from one location to another.
@ -45,7 +46,7 @@ void VTABufferFree(void* buffer);
 * \param size Size of copy.
 * \param kind_mask The memory copy kind.
 */
-void VTABufferCopy(const void* from,
+TVM_DLL void VTABufferCopy(const void* from,
                           size_t from_offset,
                           void* to,
                           size_t to_offset,
@ -56,13 +57,13 @@ void VTABufferCopy(const void* from,
 typedef void* VTACommandHandle;

 /*! \brief Shutdown hook of VTA to cleanup resources */
-void VTARuntimeShutdown();
+TVM_DLL void VTARuntimeShutdown();

 /*!
 * \brief Get thread local command handle.
 * \return A thread local command handle.
 */
-VTACommandHandle VTATLSCommandHandle();
+TVM_DLL VTACommandHandle VTATLSCommandHandle();

 /*!
 * \brief Get the buffer access pointer on CPU.
@ -70,7 +71,7 @@ VTACommandHandle VTATLSCommandHandle();
 * \param buffer The data buffer.
 * \return The pointer that can be accessed by the CPU.
 */
-void* VTABufferCPUPtr(VTACommandHandle cmd, void* buffer);
+TVM_DLL void* VTABufferCPUPtr(VTACommandHandle cmd, void* buffer);

 /*!
 * \brief Perform a write barrier to make a memory region visible to the CPU.
@ -80,7 +81,7 @@ void* VTABufferCPUPtr(VTACommandHandle cmd, void* buffer);
 * \param start The start of the region (in elements).
 * \param extent The end of the region (in elements).
 */
-void VTAWriteBarrier(VTACommandHandle cmd,
+TVM_DLL void VTAWriteBarrier(VTACommandHandle cmd,
                             void* buffer,
                             uint32_t elem_bits,
                             uint32_t start,
@ -93,7 +94,7 @@ void VTAWriteBarrier(VTACommandHandle cmd,
 * \param start The start of the region (in elements).
 * \param extent The end of the region (in elements).
 */
-void VTAReadBarrier(VTACommandHandle cmd,
+TVM_DLL void VTAReadBarrier(VTACommandHandle cmd,
                            void* buffer,
                            uint32_t elem_bits,
                            uint32_t start,
@ -104,7 +105,7 @@ void VTAReadBarrier(VTACommandHandle cmd,
 * \param cmd The VTA command handle.
 * \param debug_flag The debug flag.
 */
-void VTASetDebugMode(VTACommandHandle cmd, int debug_flag);
+TVM_DLL void VTASetDebugMode(VTACommandHandle cmd, int debug_flag);

 /*!
 * \brief Perform a 2D data load from DRAM.
@ -122,7 +123,7 @@ void VTASetDebugMode(VTACommandHandle cmd, int debug_flag);
 * \param dst_sram_index Destination SRAM index.
 * \param dst_memory_type Destination memory type.
 */
-void VTALoadBuffer2D(VTACommandHandle cmd,
+TVM_DLL void VTALoadBuffer2D(VTACommandHandle cmd,
                             void* src_dram_addr,
                             uint32_t src_elem_offset,
                             uint32_t x_size,
@ -147,7 +148,7 @@ void VTALoadBuffer2D(VTACommandHandle cmd,
 * \param y_size The number of rows.
 * \param x_stride The x axis stride.
 */
-void VTAStoreBuffer2D(VTACommandHandle cmd,
+TVM_DLL void VTAStoreBuffer2D(VTACommandHandle cmd,
                              uint32_t src_sram_index,
                              uint32_t src_memory_type,
                              void* dst_dram_addr,
@ -187,7 +188,7 @@ void VTAStoreBuffer2D(VTACommandHandle cmd,
 * \param use_imm Use immediate in ALU mode if set to true.
 * \param imm_val Immediate value in ALU mode.
 */
-void VTAUopPush(uint32_t mode,
+TVM_DLL void VTAUopPush(uint32_t mode,
                        uint32_t reset_out,
                        uint32_t dst_index,
                        uint32_t src_index,
@ -203,7 +204,7 @@ void VTAUopPush(uint32_t mode,
 * \param src_factor The input factor.
 * \param wgt_factor The weight factor.
 */
-void VTAUopLoopBegin(uint32_t extent,
+TVM_DLL void VTAUopLoopBegin(uint32_t extent,
                             uint32_t dst_factor,
                             uint32_t src_factor,
                             uint32_t wgt_factor);
@ -211,7 +212,7 @@ void VTAUopLoopBegin(uint32_t extent,
 /*!
 * \brief Mark end of a micro op loop.
 */
-void VTAUopLoopEnd();
+TVM_DLL void VTAUopLoopEnd();

 /*!
 * \brief Push GEMM uop kernel into the command handle.
@ -221,7 +222,7 @@ void VTAUopLoopEnd();
 * \param nbytes Number of bytes to in the closure arguments.
 * \return 0 if success.
 */
-int VTAPushGEMMOp(void** uop_handle,
+TVM_DLL int VTAPushGEMMOp(void** uop_handle,
                          int (*finit)(void*),
                          void* signature,
                          int nbytes);
@ -234,7 +235,7 @@ int VTAPushGEMMOp(void** uop_handle,
 * \param nbytes Number of bytes to in the closure arguments.
 * \return 0 if success.
 */
-int VTAPushALUOp(void** uop_handle,
+TVM_DLL int VTAPushALUOp(void** uop_handle,
                         int (*finit)(void*),
                         void* signature,
                         int nbytes);
@ -246,7 +247,7 @@ int VTAPushALUOp(void** uop_handle,
 * \param to_qid The destination queue.
 * \return 0 if success.
 */
-int VTADepPush(VTACommandHandle cmd, int from_qid, int to_qid);
+TVM_DLL int VTADepPush(VTACommandHandle cmd, int from_qid, int to_qid);

 /*!
 * \brief Pop dependence signal.
@ -255,7 +256,7 @@ int VTADepPush(VTACommandHandle cmd, int from_qid, int to_qid);
 * \param to_qid The destination queue.
 * \return 0 if success.
 */
-int VTADepPop(VTACommandHandle cmd, int from_qid, int to_qid);
+TVM_DLL int VTADepPop(VTACommandHandle cmd, int from_qid, int to_qid);

 /*!
 * \brief Synchronize the command handle.
@ -266,7 +267,7 @@ int VTADepPop(VTACommandHandle cmd, int from_qid, int to_qid);
 * \param wait_cycles The limit of poll cycles.
 *
 */
-void VTASynchronize(VTACommandHandle cmd, uint32_t wait_cycles);
+TVM_DLL void VTASynchronize(VTACommandHandle cmd, uint32_t wait_cycles);

 #ifdef __cplusplus
 }
--- a/vta/src/runtime.cc
+++ b/vta/src/runtime.cc
@ -10,6 +10,7 @@
 #include <vta/hw_spec.h>
 #include <vta/runtime.h>
 #include <dmlc/logging.h>
+#include <tvm/runtime/c_runtime_api.h>

 #include <cassert>
 #include <cstring>