ggml-qnn: active member variable _device_id in class qnn_instance

Author: zhouwg

ggml/src/ggml-qnn/ggml-qnn-impl.h

@@ -498,6 +498,10 @@ class qnn_instance {
         return _enable_qnn_rpc;
     }
 
+    QNNBackend get_device_id() {
+        return _device_id;
+    }
+
 public:
     std::map<std::string, std::tuple<Qnn_GraphHandle_t, std::vector< Qnn_Tensor_t *>>> _qnn_graph_map;
 

ggml/src/ggml-qnn/ggml-qnn-ops.cpp

@@ -118,6 +118,7 @@ void ggml_qnn_general_node(ggml_backend_qnn_context * ctx, ggml_tensor * op) {
 
     if (!graph_initialized) {
         GGMLQNN_LOG_DEBUG("graph name %s", graph_name.c_str());
+        GGML_ASSERT(instance->get_device_id() == ctx->device);
         error = instance->init_qnn_graph(graph_name, static_cast<QNNBackend>(ctx->device), 8);
         if (QNN_SUCCESS != error) {
             GGMLQNN_LOG_INFO("can't create qnn graph handle with graph name %s, error = %d\n", graph_name.c_str(), error);

ggml/src/ggml-qnn/ggml-qnn.cpp

@@ -1704,6 +1704,17 @@ int qnn_instance::qnn_init(const QnnSaver_Config_t ** saver_config) {
         }
     }
 
+    _device_id = QNN_BACKEND_CPU;
+    if (_backend_name.find("QnnCpu") != std::string::npos) {
+        _device_id = QNN_BACKEND_CPU;
+    }
+    if (_backend_name.find("QnnGpu") != std::string::npos) {
+        _device_id = QNN_BACKEND_GPU;
+    }
+    if (_backend_name.find("QnnHtp") != std::string::npos) {
+        _device_id = QNN_BACKEND_NPU;
+    }
+
     backend_id = _lib_path_to_backend_id[backend_lib_path];
     if (0 == _loaded_backend.count(backend_id) ||
         0 == _loaded_lib_handle.count(backend_id)) {