modify code according to review comments

Author: zhouwg

ggml-qnn.cpp

@@ -4666,6 +4666,10 @@ static ggml_guid_t ggml_backend_qnn_guid() {
 
 
 static ggml_backend_t ggml_backend_qnn_reg_init(const char * params, void * user_data) {
+    if (nullptr == params) {
+        //this path is an Android OS defined path, to make Android permission management subsystem happy
+        params = "/data/local/tmp/";
+    }
     ggml_backend_t qnn_backend = ggml_backend_qnn_init((int) (intptr_t) user_data, params);
 
     return qnn_backend;

llama.cpp

@@ -15403,8 +15403,8 @@ struct llama_context * llama_new_context_with_model(
 #elif defined(GGML_USE_QNN)
         if (model->n_gpu_layers > 0) {
             //the second param is data path of prebuit QNN libs provided by Qualcomm
-            //in Andorid APP mode, can be obtained through JNI from Java layer
-            //in Andorid terminal mode, can be hardcoded to "/data/local/tmp"
+            //can be obtained through JNI from Java layer such as "/data/data/com.ggml.llamacpp/"
+            //or hardcoded to "/data/local/tmp/"
             ggml_backend_t backend = ggml_backend_qnn_init(model->main_gpu, "/data/local/tmp/");
             if (nullptr == backend) {
                 LLAMA_LOG_ERROR("%s: failed to initialize QNN backend\n", __func__);

tests/test-backend-ops.cpp

@@ -2225,11 +2225,7 @@ int main(int argc, char ** argv) {
             continue;
         }
 
-#ifdef GGML_USE_QNN
-        ggml_backend_t backend = ggml_backend_reg_init_backend(i, "/data/local/tmp/");
-#else
         ggml_backend_t backend = ggml_backend_reg_init_backend(i, NULL);
-#endif
         GGML_ASSERT(backend != NULL);
 
         if (backend_filter == NULL && ggml_backend_is_cpu(backend)) {