[Build] fix various build problems (#3117)

include/dgl/runtime/c_runtime_api.h

@@ -542,8 +542,9 @@ DGL_DLL int DGLStreamStreamSynchronize(int device_type,
 
 /*!
  * \brief Load tensor adapter.
+ * \return 0 when success, -1 when failure happens.
  */
-DGL_DLL void DGLLoadTensorAdapter(const char *path);
+DGL_DLL int DGLLoadTensorAdapter(const char *path);
 
 /*!
  * \brief Bug report macro.

include/dgl/runtime/tensordispatch.h

@@ -65,7 +65,7 @@ class TensorDispatcher {
   }
 
   /*! \brief Load symbols from the given tensor adapter library path */
-  void Load(const char *path_cstr);
+  bool Load(const char *path_cstr);
 
   /*!
    * \brief Allocate an empty tensor.

python/dgl/_ffi/base.py

@@ -112,6 +112,7 @@ def decorate(func, fwrapped):
     import decorator
     return decorator.decorate(func, fwrapped)
 
+tensor_adapter_loaded = False
 
 def load_tensor_adapter(backend, version):
     """Tell DGL to load a tensoradapter library for given backend and version.
@@ -123,6 +124,7 @@ def load_tensor_adapter(backend, version):
     version : str
         The version number of the backend.
     """
+    global tensor_adapter_loaded
     version = version.split('+')[0]
     if sys.platform.startswith('linux'):
         basename = 'libtensoradapter_%s_%s.so' % (backend, version)
@@ -133,4 +135,4 @@ def load_tensor_adapter(backend, version):
     else:
         raise NotImplementedError('Unsupported system: %s' % sys.platform)
     path = os.path.join(_DIR_NAME, 'tensoradapter', backend, basename)
-    _LIB.DGLLoadTensorAdapter(path.encode('utf-8'))
+    tensor_adapter_loaded = (_LIB.DGLLoadTensorAdapter(path.encode('utf-8')) == 0)

src/array/cuda/spmm.cu

@@ -526,7 +526,6 @@ void SpMMCsrHetero(const std::string& op, const std::string& reduce,
         (CUDART_VERSION < 11000) &&
         ((op == "copy_lhs" && cusparse_available<bits, IdType>()) ||
          (op == "mul" && is_scalar_efeat && cusparse_available<bits, IdType>()));
-#if CUDART_VERSION < 11000
     // Create temporary output buffer to store non-transposed output
     if (use_legacy_cusparsemm) {
       for (dgl_type_t ntype = 0; ntype < vec_out.size(); ++ntype) {
@@ -539,7 +538,6 @@ void SpMMCsrHetero(const std::string& op, const std::string& reduce,
         trans_out[ntype] = out;
       }
     }
-#endif
 
     // Check shape of ufeat for all relation type and compute feature size
     int64_t x_length = 1;
@@ -634,7 +632,6 @@ void SpMMCsrHetero(const std::string& op, const std::string& reduce,
       }
     }
 
-#if CUDART_VERSION < 11000
     if (use_legacy_cusparsemm) {
       // transpose output
       for (dgl_type_t ntype = 0; ntype < vec_out.size(); ++ntype) {
@@ -646,7 +643,6 @@ void SpMMCsrHetero(const std::string& op, const std::string& reduce,
         device->FreeWorkspace(vec_csr[0].indptr->ctx, trans_out[ntype]);
       }
     }
-#endif
   });
 }
 

src/runtime/c_runtime_api.cc

@@ -379,8 +379,8 @@ int DGLCbArgToReturn(DGLValue* value, int code) {
   API_END();
 }
 
-void DGLLoadTensorAdapter(const char *path) {
-  TensorDispatcher::Global()->Load(path);
+int DGLLoadTensorAdapter(const char *path) {
+  return TensorDispatcher::Global()->Load(path) ? 0 : -1;
 }
 
 // set device api

src/runtime/tensordispatch.cc

@@ -19,18 +19,18 @@ namespace runtime {
 
 constexpr const char *TensorDispatcher::names_[];
 
-void TensorDispatcher::Load(const char *path) {
+bool TensorDispatcher::Load(const char *path) {
   CHECK(!available_) << "The tensor adapter can only load once.";
 
   if (path == nullptr || strlen(path) == 0)
     // does not have dispatcher library; all operators fall back to DGL's implementation
-    return;
+    return false;
 
 #if defined(WIN32) || defined(_WIN32)
   handle_ = LoadLibrary(path);
 
   if (!handle_)
-    return;
+    return false;
 
   for (int i = 0; i < num_entries_; ++i) {
     entrypoints_[i] = reinterpret_cast<void*>(GetProcAddress(handle_, names_[i]));
@@ -40,7 +40,7 @@ void TensorDispatcher::Load(const char *path) {
   handle_ = dlopen(path, RTLD_LAZY);
 
   if (!handle_)
-    return;
+    return false;
 
   for (int i = 0; i < num_entries_; ++i) {
     entrypoints_[i] = dlsym(handle_, names_[i]);
@@ -49,6 +49,7 @@ void TensorDispatcher::Load(const char *path) {
 #endif  // WIN32
 
   available_ = true;
+  return true;
 }
 
 TensorDispatcher::~TensorDispatcher() {