feat: add forward_type to MNNR_Config and implement backend conversion in Rust

Author: zibo-chen

cpp/include/mnn_wrapper.h

@@ -41,6 +41,7 @@ extern "C"
         int32_t precision_mode; // 0=Normal, 1=Low(faster), 2=High(accurate)
         bool use_cache;         // Whether to use cache file
         int32_t data_format;    // Input/Output data format
+        int32_t forward_type;   // MNNForwardType: 0=CPU, 1=Metal, 2=CUDA, 3=OpenCL, 6=OpenGL, 7=Vulkan, 5=CoreML/NNAPI
     } MNNR_Config;
 
     // ============== Version & Info ==============

cpp/src/mnn_wrapper.cpp

@@ -79,17 +79,17 @@ struct MNN_SessionPool
 
 // ============== Helper Functions ==============
 
-static MNN::ScheduleConfig create_schedule_config(const MNNR_Config *config)
+// Initialize schedule and backend configs from MNNR_Config.
+// Caller must ensure `schedule` and `backend` outlive any use of schedule.backendConfig.
+static void init_schedule_config(MNN::ScheduleConfig &schedule, MNN::BackendConfig &backend, const MNNR_Config *config)
 {
-    MNN::ScheduleConfig schedule;
-    schedule.type = MNN_FORWARD_CPU;
+    schedule.type = (config) ? static_cast<MNNForwardType>(config->forward_type) : MNN_FORWARD_CPU;
     schedule.numThread = config ? config->thread_count : 4;
     if (schedule.numThread <= 0)
     {
         schedule.numThread = 4;
     }
 
-    MNN::BackendConfig backend;
     if (config)
     {
         switch (config->precision_mode)
@@ -106,8 +106,6 @@ static MNN::ScheduleConfig create_schedule_config(const MNNR_Config *config)
         }
     }
     schedule.backendConfig = &backend;
-
-    return schedule;
 }
 
 static bool init_engine_tensors(MNN_InferenceEngine *engine)
@@ -165,7 +163,7 @@ MNN_SharedRuntime *mnnr_create_runtime(const MNNR_Config *config)
 
     runtime->precision_mode = config ? config->precision_mode : 0;
 
-    runtime->schedule_config.type = MNN_FORWARD_CPU;
+    runtime->schedule_config.type = (config) ? static_cast<MNNForwardType>(config->forward_type) : MNN_FORWARD_CPU;
     runtime->schedule_config.numThread = runtime->thread_count;
 
     switch (runtime->precision_mode)
@@ -214,7 +212,9 @@ MNN_InferenceEngine *mnnr_create_engine(
     }
 
     // Create default session
-    MNN::ScheduleConfig schedule = create_schedule_config(config);
+    MNN::ScheduleConfig schedule;
+    MNN::BackendConfig backend;
+    init_schedule_config(schedule, backend, config);
     engine->default_session = engine->interpreter->createSession(schedule);
     if (!engine->default_session)
     {
@@ -405,7 +405,9 @@ MNN_SessionPool *mnnr_create_session_pool(
     auto pool = new MNN_SessionPool();
     pool->engine = engine;
 
-    MNN::ScheduleConfig schedule = create_schedule_config(config);
+    MNN::ScheduleConfig schedule;
+    MNN::BackendConfig backend;
+    init_schedule_config(schedule, backend, config);
 
     // Create sessions
     for (size_t i = 0; i < pool_size; i++)
@@ -418,14 +420,8 @@ MNN_SessionPool *mnnr_create_session_pool(
             {
                 engine->interpreter->releaseSession(s);
             }
-            for (auto t : pool->input_tensors)
-            {
-                delete t;
-            }
-            for (auto t : pool->output_tensors)
-            {
-                delete t;
-            }
+            // Note: input/output tensors are owned by MNN sessions, not by us.
+            // They will be freed when sessions are released above.
             delete pool;
             return nullptr;
         }
@@ -557,7 +553,9 @@ MNN_SingleSession *mnnr_create_session(
     auto session = new MNN_SingleSession();
     session->engine = engine;
 
-    MNN::ScheduleConfig schedule = create_schedule_config(config);
+    MNN::ScheduleConfig schedule;
+    MNN::BackendConfig backend;
+    init_schedule_config(schedule, backend, config);
     session->session = engine->interpreter->createSession(schedule);
 
     if (!session->session)

src/mnn/mod.rs

@@ -117,6 +117,21 @@ mod normal_impl {
         CoreML,
     }
 
+    impl Backend {
+        /// Convert to MNNForwardType integer value
+        fn to_forward_type(self) -> i32 {
+            match self {
+                Backend::CPU => 0,    // MNN_FORWARD_CPU
+                Backend::Metal => 1,  // MNN_FORWARD_METAL
+                Backend::CUDA => 2,   // MNN_FORWARD_CUDA
+                Backend::OpenCL => 3, // MNN_FORWARD_OPENCL
+                Backend::CoreML => 5, // MNN_FORWARD_NN
+                Backend::OpenGL => 6, // MNN_FORWARD_OPENGL
+                Backend::Vulkan => 7, // MNN_FORWARD_VULKAN
+            }
+        }
+    }
+
     /// Inference configuration
     #[derive(Debug, Clone)]
     pub struct InferenceConfig {
@@ -180,6 +195,7 @@ mod normal_impl {
                 precision_mode: self.precision_mode as i32,
                 use_cache: self.use_cache,
                 data_format: self.data_format as i32,
+                forward_type: self.backend.to_forward_type(),
             }
         }
     }